{"QuestionId":41241125,"AnswerCount":1,"Tags":"<tensorflow><backpropagation>","CreationDate":"2016-12-20T11:19:30.813","AcceptedAnswerId":41245295.0,"OwnerUserId":3723434.0,"Title":"Siamese Net BackProp, how to effectively update?","Body":"<p>How to most effectively update the shared weights of a Siamese Net, given contrastive loss function in Tensorflow?<\/p>\n","answers":[{"AnswerId":"41245295","CreationDate":"2016-12-20T15:01:10.093","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>Abstract the network that produces the embeddings into a function. Just call the function twice - once for each image you pass in. It's referred to as two networks... that seems like an overstatement to me. It's just that you pass it through the same network. You'll want to <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variable_scope\/\" rel=\"nofollow noreferrer\">reuse variables<\/a> in between the two function calls. My code looks something like this:<\/p>\n\n<pre><code>def pipeline(images, reuse):\n  with tf.variable_scope('pipeline', reuse=reuse) as scope:\n  # ... conv layers etc. That produce the embedding.\n\noutput_a = pipeline(a_images, False)\noutput_b = pipeline(b_images, True)\n<\/code><\/pre>\n"}]}
{"QuestionId":41241491,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-12-20T11:39:41.197","AcceptedAnswerId":null,"OwnerUserId":762517.0,"Title":"Sliding inner product using Tensorflow convolution","Body":"<p>I have two tensors of shape N x D1 and M x D2 where D1 > D2, called X and Y respectively. For my task, X acts as the input and Y acts as the filter.<\/p>\n\n<p>I want to calculate a matrix P of shape N x M x (D1-D2+1) such that:<\/p>\n\n<pre><\/pre>\n\n<p>I can create a for loop and manually slide Y and calculate the dot products. \nHowever I prefer using the correlation operator. \nAs I know, tensorflow has correlation operator implemented (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn\/convolution\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn\/convolution<\/a>) but I don't know how can I use my tensors as inputs and filters. <\/p>\n","answers":[{"AnswerId":"46469931","CreationDate":"2017-09-28T13:04:54.060","ParentId":null,"OwnerUserId":"7704053","Title":null,"Body":"<p>Yes, you can use indeed tf.nn.conv2d(), but you should add both batch and channel dimensions:<\/p>\n\n<pre><code>X = tf.expand_dims(tf.expand_dims(X,0),-1) \n# X.shape [batch=1, in_height, in_width, in_channels=1] \nY = tf.expand_dims(tf.expand_dims(Y,-1),-1)\n# Y.shape = [filter_height, filter_width, in_channels=1, out_channels=1]\n\n# Convolution (actually correlation, see doc of conv2d) \nxcorr = tf.nn.conv2d(X, Y, padding=\"VALID\", strides=[1, 1, 1, 1])\n# Padding should be VALID, since you've already padded your input\n<\/code><\/pre>\n\n<p>CAVEAT: However, you cannot extrapolate this approach for batches of signals, since  tf.nn.conv2d uses always the same filter over the batch dimension, and from my understanding you do want to change it.<\/p>\n"},{"AnswerId":"41252408","CreationDate":"2016-12-20T22:44:53.950","ParentId":null,"OwnerUserId":"5549217","Title":null,"Body":"<p>tf.nn.conv2d(input, filter, strides, padding, use_cudnn_on_gpu=None, data_format=None, name=None)<\/p>\n\n<p>In your case, I'd set strides to 1, and padding to SAME.<\/p>\n\n<p>tf.nn.conv2d(X, Y, strides=1, padding=SAME)<\/p>\n"}]}
{"QuestionId":41241523,"AnswerCount":0,"Tags":"<cmake><gpu><cpu><deep-learning><caffe>","CreationDate":"2016-12-20T11:41:07.540","AcceptedAnswerId":null,"OwnerUserId":3610481.0,"Title":"caffe.set_mode_cpu() error in Caffe","Body":"<p>I have built caffe with only cpu support. Is the command 'caffe.set_mode_cpu() ' only used when we have built with gpu support so that we can switch to cpu when needed? I thought I might need it just to make sure that Caffe is using my cpu but I guess the build takes care of that. Also is this command required even when I have built with cpu support only?<\/p>\n\n<p>Error I get-<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/vP-y8ju92zY\" rel=\"nofollow noreferrer\">Problem posted on caffe users group<\/a><\/p>\n\n<p>This is my output of 'ccmake ..' . It says that CPU_ONLY is off even after removing the comment on CPU flag. How do I make it build with CPU for sure?\n<a href=\"https:\/\/i.stack.imgur.com\/d4MpC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/d4MpC.png\" alt=\"Output of ccmake ..\"><\/a><\/p>\n\n<p>To build Caffe, I used cmake .. instead of make as I got <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2348\" rel=\"nofollow noreferrer\">convert_imageset.bin error<\/a>. So I followed the instructions in the link and I got it to build properly.<\/p>\n\n<p>Now I was looking at my cmake output and realised that the \"CPU_ONLY\" option was set to off. So i followed this <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/ZejVbSP8oLU\" rel=\"nofollow noreferrer\">link<\/a> where i used \"cmake -DCPU_ONLY=ON\" to set it ON. <\/p>\n\n<p>But I'm still getting Cuda error even when cmake option \"CPU_ONLY=ON\" is there. I am not sure why it is still being built with GPU?<\/p>\n\n<p>Looking at my cmake output again, I found this error- <\/p>\n\n<pre><\/pre>\n\n<p>Is this fine since anyways we have to do make pycaffe to build with python?<\/p>\n","answers":[]}
{"QuestionId":41243081,"AnswerCount":1,"Tags":"<tensorflow><keras><training-data>","CreationDate":"2016-12-20T13:05:11.353","AcceptedAnswerId":41252247.0,"OwnerUserId":4663495.0,"Title":"CNN model why the data is too large?","Body":"<p><img src=\"https:\/\/i.stack.imgur.com\/kKdPx.png\" alt=\"the error message\"><\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/09Psh.png\" alt=\"the model structure\"><\/p>\n\n<ul>\n<li>use keras and tensorflow backend <\/li>\n<li>use CPU only , memory 128GB<\/li>\n<li>input data has the shape (45,1024,1024) <\/li>\n<li>the model has only one convolution , one (2,2) max pooling<\/li>\n<li>a 1024 *1024 fully connected.<\/li>\n<\/ul>\n\n<p>I got this error message:<\/p>\n\n<blockquote>\n  <p>Invalid argument: Shape [4194304,1048576] is too large (more than\n  1099511627776 entries)<\/p>\n<\/blockquote>\n\n<p>note that:<\/p>\n\n<pre><\/pre>\n\n<p>How did keras compute this shape? why is it too large?<\/p>\n","answers":[{"AnswerId":"41252247","CreationDate":"2016-12-20T22:28:25.477","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>The shape <code>[4194304, 1048576]<\/code> is computed as follows:<\/p>\n\n<p>Applying 16 convolutions of size <code>[3, 3]<\/code> with the <code>same<\/code> border mode to the inputs of size <code>[1024, 1024, 3]<\/code> gives us output of size <code>[1024, 1024, 16]<\/code>. After max pooling of size 2 it becomes <code>[512, 512, 16]<\/code> which when flattened is <code>512 * 512 * 16 = 4194304<\/code>. <code>1048576<\/code> comes from <code>1024 * 1024<\/code> as you specified in the <code>Dense<\/code> layer constructor. <\/p>\n\n<p>I think you should reconsider the architecture of you model. You can use inputs of smaller size, add several pooling layers, reduce dimensionality applying <code>1 x 1<\/code> convolutions. And I doubt that <code>1024 * 1024<\/code> is a reasonable number of nodes in a fully connected layer.<\/p>\n"}]}
{"QuestionId":41243358,"AnswerCount":0,"Tags":"<image-processing><lua><computer-vision><linear-algebra><torch>","CreationDate":"2016-12-20T13:20:42.370","AcceptedAnswerId":null,"OwnerUserId":7321118.0,"Title":"How to find subgroups of nonzeros inside 2D matrix?","Body":"<p>I'm currently working with Luajit Torch and I ran into a problem.\nI have a Tensor (2D matrix) whose numbers can be any positive integer or zero. I need to group all values that are not zero and find the center of mass of each of these groups. I'm trying to work the problem as it was a greyscale image, (where 0s would be black regions) and trying to find the non-black regions first and afterwards compute the center of mass of these regions, but I couldn't figure out the algorithm yet.<\/p>\n\n<p>I'm still struggling on how to find the indices of the subgroups, thanks in advance!<\/p>\n","answers":[]}
{"QuestionId":41244421,"AnswerCount":4,"Tags":"<neural-network><tensorflow>","CreationDate":"2016-12-20T14:17:17.000","AcceptedAnswerId":41245540.0,"OwnerUserId":7321233.0,"Title":"Linear vs nonlinear neural network?","Body":"<p>I'm new to machine learning and neural networks. I know how to build a nonlinear classification model, but my current problem has a continuous output. I've been searching for information on neural network regression, but all I encounter is information on <strong>linear<\/strong> regression - nothing about <strong>nonlinear<\/strong> cases. Which is odd, because why would someone use neural networks to solve a simple linear regression anyway? Isn't that like killing a fly with a nuclear bomb?<\/p>\n\n<p>So my question is this: what makes a neural network nonlinear? (Hidden layers? Nonlinear activation function?) Or do I have a completely wrong understanding of the word \"linear\" - can a linear regression NN accurately model datasets that are more complex than y=aX+b? Is the word \"linear\" used just as the opposite of \"logistic\"?<\/p>\n\n<p>(I'm planning to use TensorFlow, but the TensorFlow Linear Model Tutorial uses a binary classification problem as an example, so that doesn't help me either.)<\/p>\n","answers":[{"AnswerId":"41246118","CreationDate":"2016-12-20T15:42:32.890","ParentId":null,"OwnerUserId":"7319275","Title":null,"Body":"<p>I had the same struggle, most online courses use ANNs for classification, but you never actually solve a regression problem with them in the courses.<\/p>\n\n<p><strong>What does make an ANN non-linear? The activation function.<\/strong><\/p>\n\n<p>Even if you have an ANN with thousands of perceptrons and hidden units, if all the activations are linear (or not activated at all) you are just training a plain linear regression.<\/p>\n\n<p>But be careful, some activations functions (like sigmoid), have a range of values that act as a linear function and you may get stuck with a linear model even with non-linear activations.<\/p>\n\n<p><strong>How to predict continuous output with an ANN? The same way as when you classify.<\/strong><\/p>\n\n<p>It is the same problem, you just backpropagate the error (label - prediction) and update the weights. But don't forget to C<em>HANGE THE ACTIVATION FUNCTION of the output layer to a continuous function<\/em> (maybe ReLu if all labels are positive or don't activate the output at all), the intermediate hidden layers can be activated however you wish.<\/p>\n\n<p>For small regression problems with ANNs you may need to start with a veeeeeery small learning rate since there will be lots of variance since the error will be \"unbounded\" at first.<\/p>\n\n<p>Hope this helps :)<\/p>\n"},{"AnswerId":"41246088","CreationDate":"2016-12-20T15:40:46.810","ParentId":null,"OwnerUserId":"3325672","Title":null,"Body":"<p>Any non-linearity from the input to output makes the network non-linear. In the way we usually think about and implement neural networks, those non-linearities come from <strong>activation functions<\/strong>.<\/p>\n\n<p>If we are trying to fit non-linear data and only have linear activation functions, our best approximation to the non-linear data will be linear since that's all we can compute. You can see an example of a neural network trying to fit <strong>non-linear data<\/strong> with only <strong>linear activation functions<\/strong> <a href=\"http:\/\/playground.tensorflow.org\/#activation=linear&amp;batchSize=20&amp;dataset=circle&amp;regDataset=reg-gauss&amp;learningRate=0.003&amp;regularizationRate=0.001&amp;noise=0&amp;networkShape=4,2&amp;seed=0.87118&amp;showTestData=false&amp;discretize=true&amp;percTrainData=80&amp;x=true&amp;y=true&amp;xTimesY=false&amp;xSquared=false&amp;ySquared=false&amp;cosX=false&amp;sinX=false&amp;cosY=false&amp;sinY=false&amp;collectStats=false&amp;problem=regression&amp;initZero=false&amp;hideText=false\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>However, if we change the <strong>linear activation function<\/strong> to something <strong>non-linear<\/strong> like <strong>ReLu<\/strong>, then we can see a better non-linear fitting of the data. You can see that <a href=\"http:\/\/playground.tensorflow.org\/#activation=relu&amp;batchSize=20&amp;dataset=circle&amp;regDataset=reg-gauss&amp;learningRate=0.003&amp;regularizationRate=0.001&amp;noise=0&amp;networkShape=4,2&amp;seed=0.97957&amp;showTestData=false&amp;discretize=true&amp;percTrainData=80&amp;x=true&amp;y=true&amp;xTimesY=false&amp;xSquared=false&amp;ySquared=false&amp;cosX=false&amp;sinX=false&amp;cosY=false&amp;sinY=false&amp;collectStats=false&amp;problem=regression&amp;initZero=false&amp;hideText=false\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"41245858","CreationDate":"2016-12-20T15:29:18.543","ParentId":null,"OwnerUserId":"4869038","Title":null,"Body":"<p>When it comes to nonlinear regression, this is referring to how the weights affect the output. If a function is not linear with respect to the weights, then your problem is a nonlinear regression problem. So for example, let's look at a Feedforward Neural Network with one hidden layer where the activation functions in the hidden layer are some function <img src=\"https:\/\/latex.codecogs.com\/gif.latex?g(z)\" title=\"g(z)\" \/> and the output layer has linear activation functions. Given this, the mathematical representation can be:<\/p>\n\n<p><img src=\"https:\/\/latex.codecogs.com\/gif.latex?y&space;=&space;W_2&space;g(W_1&space;x&space;&plus;&space;b_1)&space;&plus;&space;b_2\" title=\"y = W_2 g(W_1 x + b_1) + b_2\" \/><\/p>\n\n<p>where we assume <img src=\"https:\/\/latex.codecogs.com\/gif.latex?g(z)\" title=\"g(z)\" \/> can operator on scalars and vectors with this notation to make it easy. <img src=\"https:\/\/latex.codecogs.com\/gif.latex?W_1\" title=\"W_1\" \/>, <img src=\"https:\/\/latex.codecogs.com\/gif.latex?W_2\" title=\"W_2\" \/>, <img src=\"https:\/\/latex.codecogs.com\/gif.latex?b_1\" title=\"b_1\" \/>, and <img src=\"https:\/\/latex.codecogs.com\/gif.latex?b_2\" title=\"b_2\" \/> are the weight you are aiming to estimate with the regression. If this was linear regression, <img src=\"https:\/\/latex.codecogs.com\/gif.latex?g(z)\" title=\"g(z)\" \/> would equal z, because that would make y linearly dependent on <img src=\"https:\/\/latex.codecogs.com\/gif.latex?W_1\" title=\"W_1\" \/> &amp; <img src=\"https:\/\/latex.codecogs.com\/gif.latex?b_1\" title=\"b_1\" \/>. But if <img src=\"https:\/\/latex.codecogs.com\/gif.latex?g(z)\" title=\"g(z)\" \/> is nonlinear, say like <img src=\"https:\/\/latex.codecogs.com\/gif.latex?g(z)&space;=&space;z%5E2\" title=\"g(z) = z^2\" \/>, then now y is nonlinearly dependent on the weights <img src=\"https:\/\/latex.codecogs.com\/gif.latex?W_2\" title=\"W_2\" \/>. <\/p>\n\n<p>Now provided you understand all that, I am surprised you haven't seen discussion of the nonlinear case because that's pretty much all people talk about in textbooks and research. The use of things like stochastic gradient descent, Nonlinear Conjugate Gradient, RProp, and other methods are to help find local minima (and hopefully good local minima) for these nonlinear regression problems, even though a global optimum is not typically guaranteed.<\/p>\n"},{"AnswerId":"41245540","CreationDate":"2016-12-20T15:13:41.403","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>For starters, a neural network can model any function (not just linear functions) Have a look at this - <a href=\"http:\/\/neuralnetworksanddeeplearning.com\/chap4.html\" rel=\"noreferrer\">http:\/\/neuralnetworksanddeeplearning.com\/chap4.html<\/a>.<\/p>\n\n<p>A Neural Network has got non linear activation layers which is what gives the Neural Network a non linear element.<\/p>\n\n<p>The function for relating the input and the output is decided by the neural network and the amount of training it gets. If you supply two variables having a linear relationship, then your network will learn this as long as you don't overfit. Similarly, a complex enough neural network can learn any function.<\/p>\n"}]}
{"QuestionId":41244472,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-12-20T14:20:39.013","AcceptedAnswerId":41246119.0,"OwnerUserId":308255.0,"Title":"Why is TensorFlow matmul() much slower than NumPy multiply()?","Body":"<p>In the python code below, why is the time for multiplication via numpy much smaller than via tensorflow?<\/p>\n\n<pre><\/pre>\n\n<p>The output is <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41246119","CreationDate":"2016-12-20T15:42:33.287","ParentId":null,"OwnerUserId":"1951176","Title":null,"Body":"<p>Well, citing documentation:<\/p>\n\n<blockquote>\n  <p>numpy.multiply(x1, x2[, out]) =  Multiply arguments\n  element-wise.<\/p>\n<\/blockquote>\n\n<p>and<\/p>\n\n<blockquote>\n  <p>tf.matmul(a, b, transpose_a=False, transpose_b=False,\n  a_is_sparse=False, b_is_sparse=False, name=None)<\/p>\n  \n  <p>Multiplies matrix a by matrix b, producing a * b.<\/p>\n  \n  <p>The inputs must be two-dimensional matrices, with matching inner\n  dimensions, possibly after transposition.<\/p>\n<\/blockquote>\n\n<p>Which suggest that you compare different operations: O(n^2) pointwise multiplication with O(n^3) matrix multiplication. I corrected the test to use matrix multiplication 2 times in both cases:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport time\nsize=2000\nx = tf.placeholder(tf.float32, shape=(size, size))\ny = tf.matmul(x, x)\nz = tf.matmul(y, x)\n\nwith tf.Session() as sess:\n  rand_array = np.random.rand(size, size)\n\n  start_time = time.time()\n  for _ in xrange(10):\n      np.dot(np.dot(rand_array,rand_array), rand_array)\n  print(\"--- %s seconds numpy multiply ---\" % (time.time() - start_time))\n\n  start_time = time.time()\n  for _ in xrange(10):\n      sess.run(z, feed_dict={x: rand_array})\n  print(\"--- %s seconds tensorflow---\" % (time.time() - start_time))\n<\/code><\/pre>\n\n<p>and got the results:<\/p>\n\n<pre><code>--- 2.92911195755 seconds numpy multiply ---\n--- 0.32932305336 seconds tensorflow---\n<\/code><\/pre>\n\n<p>using a fast GPU (gtx 1070).<\/p>\n"}]}
{"QuestionId":41244710,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><keras>","CreationDate":"2016-12-20T14:33:17.800","AcceptedAnswerId":null,"OwnerUserId":4811088.0,"Title":"Keras Binary Classifier Model.Predict() class association?","Body":"<p>A bit of a noob question but ive used Keras to train a neural net as a binary classifier. Now i want to call model.predict_classes() on new unseen data to see how well its generalising.... this is a stupid question but is there any way to understand what class the 0 or 1 correspond to from my training \/ validation data????<\/p>\n\n<p>I suppose it would be good to know how i can identify which is positive \/ negative (i.e. 0 or 1) or A\/B, Cat\/Dog etc etc. I imagine ive overlooked this earlier when building the model <\/p>\n\n<p>Also it would be good to get some feedback on my code....<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vWJLX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vWJLX.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"43471744","CreationDate":"2017-04-18T12:09:46.263","ParentId":null,"OwnerUserId":"7085724","Title":null,"Body":"<p>I think it depends on which class you feed into your classifier first when training. But you could always relabel your training images with corresponding names so that when you classify new images the result will show your new labels instead of numbers.<\/p>\n"}]}
{"QuestionId":41244957,"AnswerCount":0,"Tags":"<python><tensorflow><training-data>","CreationDate":"2016-12-20T14:44:26.570","AcceptedAnswerId":null,"OwnerUserId":7318225.0,"Title":"Tensorflow AlexNet accuracy does not increase during training for ILSVRC2012 data set","Body":"<p>We have been working with a tensorflow AlexNet implementation based on the following model.<\/p>\n\n<p><a href=\"https:\/\/github.com\/SidHard\/tfAlexNet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/SidHard\/tfAlexNet<\/a><\/p>\n\n<p>We have been trying to train the model using the ILSVRC2012 training data which contains 1000 categories of images.  However during this training the accuracy is almost always reported as zero.  The loss decreases and eventually levels out (~7.xx) after about 600 iterations.  Here is the sample output for 1000 iterations:<\/p>\n\n<p>lr 0.001 Iter 20, Minibatch Loss= 32345.199219, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 40, Minibatch Loss= 19276.070312, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 60, Minibatch Loss= 10468.625977, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 80, Minibatch Loss= 8523.987305, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 100, Minibatch Loss= 6895.601562, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 120, Minibatch Loss= 5296.706055, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 140, Minibatch Loss= 4541.700195, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 160, Minibatch Loss= 3658.005371, Training Accuracy= 0.01562<\/p>\n\n<p>lr 0.001 Iter 180, Minibatch Loss= 3368.450195, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 200, Minibatch Loss= 2641.639160, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 220, Minibatch Loss= 2349.733154, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 240, Minibatch Loss= 2258.051270, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 260, Minibatch Loss= 2042.907471, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 280, Minibatch Loss= 1863.766602, Training Accuracy= 0.01562<\/p>\n\n<p>lr 0.001 Iter 300, Minibatch Loss= 1761.209717, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 320, Minibatch Loss= 1295.963623, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 340, Minibatch Loss= 1295.362793, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 360, Minibatch Loss= 1384.312988, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 380, Minibatch Loss= 1054.358154, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 400, Minibatch Loss= 1121.298584, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 420, Minibatch Loss= 874.211853, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 440, Minibatch Loss= 643.881165, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 460, Minibatch Loss= 521.504211, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 480, Minibatch Loss= 312.603027, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 500, Minibatch Loss= 227.839951, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 520, Minibatch Loss= 87.130402, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 540, Minibatch Loss= 8.717880, Training Accuracy= 0.03125<\/p>\n\n<p>lr 0.001 Iter 560, Minibatch Loss= 7.396675, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 580, Minibatch Loss= 7.604724, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 600, Minibatch Loss= 7.333892, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 620, Minibatch Loss= 7.501148, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 640, Minibatch Loss= 7.662533, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 660, Minibatch Loss= 7.267312, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 680, Minibatch Loss= 7.307860, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 700, Minibatch Loss= 7.554455, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 720, Minibatch Loss= 7.164557, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 740, Minibatch Loss= 7.410737, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 760, Minibatch Loss= 7.430292, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 780, Minibatch Loss= 7.378484, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 800, Minibatch Loss= 7.348798, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 820, Minibatch Loss= 7.194911, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 840, Minibatch Loss= 7.180418, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 860, Minibatch Loss= 7.523817, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 880, Minibatch Loss= 7.398925, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 900, Minibatch Loss= 7.427280, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 920, Minibatch Loss= 7.344418, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 940, Minibatch Loss= 7.317279, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.001 Iter 960, Minibatch Loss= 7.290085, Training Accuracy= 0.01562<\/p>\n\n<p>lr 0.001 Iter 980, Minibatch Loss= 7.403413, Training Accuracy= 0.00000<\/p>\n\n<p>lr 0.0001 Iter 1000, Minibatch Loss= 7.329219, Training Accuracy= 0.00000<\/p>\n\n<p>We have run the training for > 10,000 iterations and neither the loss or accuracy change after ~600 iterations.<\/p>\n\n<p>Does anyone have any thoughts\/ideas about this training behavior?<\/p>\n\n<p>Also, has anyone been able to set up and train a tensorflow AlexNet model using the ILSVRC2012 data set?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
{"QuestionId":41245797,"AnswerCount":0,"Tags":"<tensorflow><gradient-descent>","CreationDate":"2016-12-20T15:26:25.730","AcceptedAnswerId":null,"OwnerUserId":2916263.0,"Title":"Increasing and wide spreading cost function with Stochastic Gradient Descent","Body":"<p>I am using Tensorflow in an online learning environment.\nAs cost function is implemented:<\/p>\n\n<pre><\/pre>\n\n<p>Optimization is done like:<\/p>\n\n<pre><\/pre>\n\n<p>And I run Stochastic Gradient Descent like:<\/p>\n\n<pre><\/pre>\n\n<p>Whereby input_batch and label_batch contain only one vector each.<\/p>\n\n<p>So how can I interpret a cost function like:\n<a href=\"https:\/\/i.stack.imgur.com\/4J9tW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4J9tW.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is this a good progress for a stochastic approach?\nWhy the gap gets bigger?<\/p>\n\n<p>I train the network 50'000 times with the same 50 training examples.\nSo each example is used for training 10'000 times every 51th step.<\/p>\n\n<p>I tried already to change the learning rate by factor 10 in both ways.\nThis question is related to my other question from: <a href=\"https:\/\/stackoverflow.com\/questions\/41224983\/does-stochastic-gradient-descent-even-work-with-tensorflow\">Does Stochastic Gradient Descent even work with TensorFlow?<\/a><\/p>\n\n<p>Thanks for any hints.<\/p>\n","answers":[]}
{"QuestionId":41246438,"AnswerCount":4,"Tags":"<numpy><tensorflow>","CreationDate":"2016-12-20T16:00:11.507","AcceptedAnswerId":null,"OwnerUserId":7307680.0,"Title":"tensorflow record with float numpy array","Body":"<p>I want to create tensorflow records to feed my model;\nso far I use the following code to store uint8 numpy array to TFRecord format;<\/p>\n\n<pre><\/pre>\n\n<p>which I read with this example code<\/p>\n\n<pre><\/pre>\n\n<p>and that's working fine. Now I want to do the same with my array \"map\" being a float numpy array, instead of uint8, and I could not find examples on how to do it;\nI tried the function _floats_feature, which works if I pass a scalar to it, but not with arrays;\nwith uint8 the serialization can be done by the method tostring();<\/p>\n\n<p>How can I serialize a float numpy array and how can I read that back?<\/p>\n","answers":[{"AnswerId":"54682037","CreationDate":"2019-02-14T01:41:26.377","ParentId":null,"OwnerUserId":"3595278","Title":null,"Body":"<p>I've stumbled across this while working on a similar problem. Since part of the original question was how to read back the <code>float32<\/code> feature from <code>tfrecords<\/code>, I'll leave this here in case it helps anyone:<\/p>\n\n<p>If <code>map.ravel()<\/code> was used to input <code>map<\/code> of dimensions <code>[x, y, z]<\/code> into <code>_floats_feature<\/code>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>features = {\n    ...\n    'map': tf.FixedLenFeature([x, y, z], dtype=tf.float32)\n    ...\n}\nparsed_example = tf.parse_single_example(serialized=serialized, features=features)\nmap = parsed_example['map']\n<\/code><\/pre>\n"},{"AnswerId":"41251354","CreationDate":"2016-12-20T21:13:42.967","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p><code>FloatList<\/code> and <code>BytesList<\/code> expect an iterable. So you need to pass it a list of floats. Remove the extra brackets in your <code>_float_feature<\/code>, ie<\/p>\n\n<pre><code>def _floats_feature(value):\n  return tf.train.Feature(float_list=tf.train.FloatList(value=value))\n\nnumpy_arr = np.ones((3,)).astype(np.float)\nexample = tf.train.Example(features=tf.train.Features(feature={\"bytes\": _floats_feature(numpy_arr)}))\nprint(example)\n\nfeatures {\n  feature {\n    key: \"bytes\"\n    value {\n      float_list {\n        value: 1.0\n        value: 1.0\n        value: 1.0\n      }\n    }\n  }\n}\n<\/code><\/pre>\n"},{"AnswerId":"45893805","CreationDate":"2017-08-26T09:08:40.483","ParentId":null,"OwnerUserId":"4451484","Title":null,"Body":"<p>Yaroslav's example failed when a nd array was the input:<\/p>\n\n<p>numpy_arr = np.ones((3,3)).astype(np.float)<\/p>\n\n<p>I found that it worked when I used numpy_arr.ravel() as the input. But is there a better way to do it?<\/p>\n"},{"AnswerId":"44870483","CreationDate":"2017-07-02T11:48:45.390","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>I will expand on the Yaroslav's answer. <\/p>\n\n<p>Int64List, BytesList and FloatList expect an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/core\/example\/feature.proto#L65\" rel=\"nofollow noreferrer\">iterator of the underlying elements<\/a> (repeated field). In your case you can use a list as an iterator.<\/p>\n\n<p>You mentioned: <em>it works if I pass a scalar to it, but not with arrays<\/em>. And this is expected, because when you pass a scalar, your <code>_floats_feature<\/code> creates an array of one float element in it (exactly as expected). But when you pass an array you create a list of arrays and pass it to a function which expects a list of floats. <\/p>\n\n<p>So just remove construction of the array from your function: <code>float_list=tf.train.FloatList(value=value)<\/code><\/p>\n"}]}
{"QuestionId":41246621,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-20T16:10:08.870","AcceptedAnswerId":null,"OwnerUserId":7216602.0,"Title":"python-tk package for tensflow on windows","Body":"<p>I installed tensorflow and Docker Tool Box on windows.Tensorflow runs on the virtual machine in docker tool box.Everything went well except when I type 'import matplotlib.pyplot as plt', an error occured that 'no module named _tkinter please install the python-tk package'.I tried to use 'pip install python-tk ' but it doesn't work. However, there is no error when I run it on local host. Plus the default version installed on my laptop is python 3.5 while the default version shown on tensorflow is 2.7. Did anyone know how to solve this?<\/p>\n","answers":[{"AnswerId":"42179994","CreationDate":"2017-02-11T19:01:07.647","ParentId":null,"OwnerUserId":"4058054","Title":null,"Body":"<p>I've got the same problem. <\/p>\n\n<p>Solved it by running next commands in terminal:<\/p>\n\n<pre><code>sudo apt-get update\nsudo apt-get install python-tk\n<\/code><\/pre>\n\n<p>After that import was successfull.<\/p>\n"}]}
{"QuestionId":41246767,"AnswerCount":1,"Tags":"<python><docker><tensorflow><nvidia>","CreationDate":"2016-12-20T16:17:11.150","AcceptedAnswerId":41248156.0,"OwnerUserId":470184.0,"Title":"De-allocating memory after python tensorflow workbook execution","Body":"<p>To limit memory usage I read <a href=\"https:\/\/stackoverflow.com\/questions\/34199233\/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory\">How to prevent tensorflow from allocating the totality of a GPU memory?<\/a> and tried this code : <\/p>\n\n<pre><\/pre>\n\n<p>These commands did free up memory but but memory is not de-allocated after code completion. This issue describes : \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3701\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3701<\/a> a suggested fix is to update the driver\n\u201cAfter upgrading the GPU driver from 352.79 to 367.35 (the newest one), the problem disappeared. \u201c\nUnfortunately I'm not in position to update to latest version of driver. Has this issue been resolved. <\/p>\n\n<p>I also considered limiting the available memory to the docker container. \nReading <a href=\"https:\/\/devblogs.nvidia.com\/parallelforall\/nvidia-docker-gpu-server-application-deployment-made-easy\/\" rel=\"nofollow noreferrer\">https:\/\/devblogs.nvidia.com\/parallelforall\/nvidia-docker-gpu-server-application-deployment-made-easy\/<\/a> states \"Containers can be constrained to a limited set of resources on a system (e.g one CPU core and 1GB of memory)\u201d but kernel does not currently support this, here I try to add 1GB of memory to new docker instance : <\/p>\n\n<pre><\/pre>\n\n<p>But this does not appear possible as receive warning : \nWARNING: Your kernel does not support swap limit capabilities, memory limited without swap.\"<\/p>\n\n<p>Is there a command to free memory after tensorflow python workbook completion ?<\/p>\n\n<pre><\/pre>\n\n<p>After killing \/ restarting the notebook the memory is de-allocated. But how to free memory after completion within the notebook.<\/p>\n","answers":[{"AnswerId":"41248156","CreationDate":"2016-12-20T17:34:32.913","ParentId":null,"OwnerUserId":"6199836","Title":null,"Body":"<p>Ipython and jupyter notebooks will not free memory unless you use <code>del<\/code> or <code>xdel<\/code> on your objects:<\/p>\n\n<p><a href=\"https:\/\/ipython.org\/ipython-doc\/3\/interactive\/magics.html\" rel=\"nofollow noreferrer\">https:\/\/ipython.org\/ipython-doc\/3\/interactive\/magics.html<\/a><\/p>\n\n<blockquote>\n  <p>%xdel:\n  Delete a variable, trying to clear it from anywhere that IPython\u2019s machinery has references to it. By default, this uses the identity of the named object in the user namespace to remove references held under other names. The object is also removed from the output history.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41246969,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-12-20T16:28:51.213","AcceptedAnswerId":null,"OwnerUserId":6770703.0,"Title":"Are there any guidelines on sharding a data set?","Body":"<p>Are there any guidelines on choosing the number of shard files for a data set, or the number of records in each shard?<\/p>\n\n<p>In the examples of using ,<\/p>\n\n<ul>\n<li><p>there are roughly  records in each shard of <em>ImageNet<\/em> data set.(<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"noreferrer\">tensorflow\/models\/inception<\/a>)<\/p><\/li>\n<li><p>there are roughly  records in each shard of <em>flowers<\/em> data set. (<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/datasets\/download_and_convert_flowers.py\" rel=\"noreferrer\">tensorflow\/models\/slim<\/a>)<\/p><\/li>\n<\/ul>\n\n<p>Do the number of shard files and the number of records in each shard have any impact on the training and the performance of the trained model?<\/p>\n\n<p>To my knowledge, if we don't split the data set into multiple shards, it will be not quite random for shuffling data as the  of the  may be less than the size of the data set. <\/p>\n\n<p>Are there any other advantages of using multiple shards?<\/p>\n\n<hr>\n\n<p>Update<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"noreferrer\">The documentation<\/a> says <\/p>\n\n<blockquote>\n  <p>If you have more reading threads than input files, to avoid the risk that you will have two threads reading the same example from the same file near each other.<\/p>\n<\/blockquote>\n\n<p>Why can't we use 50 threads to read from 5 files?<\/p>\n","answers":[]}
{"QuestionId":41247221,"AnswerCount":3,"Tags":"<image-processing><tensorflow><deep-learning>","CreationDate":"2016-12-20T16:42:04.603","AcceptedAnswerId":null,"OwnerUserId":4819881.0,"Title":"Can inception model be used for object counting in an image?","Body":"<p>I have already gone through the image classification part in Inception model, but I require to count the objects in the image. <\/p>\n\n<p>Considering the flowers data-set, one image can have multiple instances of a flower, so how can I get that count?<\/p>\n","answers":[{"AnswerId":"42479298","CreationDate":"2017-02-27T06:39:51.803","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>What you describe is known to research community as <strong>Instance-Level Segmentation<\/strong>. <\/p>\n\n<p>In last year itself there have been a significant spike in papers addressing this problem.<\/p>\n\n<p>Here are some of the papers:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1412.7144v4.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1412.7144v4.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1511.08498v3.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1511.08498v3.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1607.03222v2.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1607.03222v2.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1607.04889v2.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1607.04889v2.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1511.08250v3.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1511.08250v3.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1611.07709v1.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1611.07709v1.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1603.07485v2.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1603.07485v2.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1611.08303v1.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1611.08303v1.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1611.08991v2.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1611.08991v2.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1611.06661v2.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1611.06661v2.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1612.03129v1.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1612.03129v1.pdf<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/pdf\/1605.09410v4.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1605.09410v4.pdf<\/a><\/li>\n<\/ul>\n\n<p>As you see in these papers simple object classification network won't solve the problem. <\/p>\n\n<p>If you search github you will find a few repositories with basic frameworks, you can build on top of them.<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/daijifeng001\/MNC\" rel=\"noreferrer\">https:\/\/github.com\/daijifeng001\/MNC<\/a> (caffe)<\/li>\n<li><a href=\"https:\/\/github.com\/bernard24\/RIS\/blob\/master\/RIS_infer.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/bernard24\/RIS\/blob\/master\/RIS_infer.ipynb<\/a> (torch)<\/li>\n<li><a href=\"https:\/\/github.com\/jr0th\/segmentation\" rel=\"noreferrer\">https:\/\/github.com\/jr0th\/segmentation<\/a> (keras, tensorflow)<\/li>\n<\/ul>\n"},{"AnswerId":"42579491","CreationDate":"2017-03-03T12:57:04.620","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>indraforyou answered the question in how to solve the problem you are having. I want to add something for the inception model specifically. In <a href=\"https:\/\/arxiv.org\/pdf\/1312.6229.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1312.6229.pdf<\/a> they propose a regressor network trained on the output of a model trained on the imagenet dataset like the inception model. This regressor model then is used to propose object boundaries for you to use for counting. The advantage of this approach is that you do not have to annotate any training examples and you can just use the ImageNet dataset for training. <\/p>\n\n<p>If you do not want to train anything I would propose a heuristic in finding object boundaries. Literature in image segmentation <a href=\"https:\/\/en.wikipedia.org\/wiki\/Image_segmentation\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Image_segmentation<\/a> should help you find a suitable heuristic. I do think using a heuristic will decrease your accuracy though.<\/p>\n\n<p>Last but not least this is an open problem in computer vision research. You should not expect to get 100% accuracy or even 95% accuracy on counting. Many very smart people have tried this and reported mixed results. Still some very cool things can be accomplished.<\/p>\n"},{"AnswerId":"42591953","CreationDate":"2017-03-04T04:05:19.510","ParentId":null,"OwnerUserId":"5288679","Title":null,"Body":"<p>Any classification model like <em>inception model<\/em> will identify the object like flower in your case. However, when multiple items are there classifications won't work (<em>get confused in simple language<\/em>).<\/p>\n\n<p>Thus:<\/p>\n\n<p><em>You've to segment main image into child images with one object per image<\/em> and use classification on each segment. This is termed as <strong>image segmentation<\/strong> in image processing.<\/p>\n"}]}
{"QuestionId":41248677,"AnswerCount":2,"Tags":"<numpy><machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2016-12-20T18:07:35.657","AcceptedAnswerId":41248943.0,"OwnerUserId":7017670.0,"Title":"Keras Array Input Error","Body":"<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>in my keras model. <\/p>\n\n<p>I think the model is mistaking something?<\/p>\n\n<p>This happens when I feed input to my model. The same input works perfectly well in another program.<\/p>\n","answers":[{"AnswerId":"48807881","CreationDate":"2018-02-15T12:57:50.843","ParentId":null,"OwnerUserId":"4156298","Title":null,"Body":"<p>The problem really comes from giving the wrong input to the network.<\/p>\n\n<p>In my case the problem was that my custom image generator was passing the entire dataset as input rather than a certain pair of image-label. This is because I thought that generator.flow(x,y, batch_size) of Keras already has a yield structure inside, however the correct generator structure should be as follows(with a separate yield):<\/p>\n\n<pre><code>def generator(batch_size):\n\n(images, labels) = utils.get_data(1000) # gets 1000 samples from dataset\nlabels = to_categorical(labels, 2)\n\ngenerator = ImageDataGenerator(featurewise_center=True,\n                 featurewise_std_normalization=True,\n                 rotation_range=90.,\n                 width_shift_range=0.1,\n                 height_shift_range=0.1,\n                 zoom_range=0.2)\n\ngenerator.fit(images)\n\ngen = generator.flow(images, labels, batch_size=32)\n\nwhile 1:\n    x_batch, y_batch = gen.next()\n    yield ([x_batch, y_batch])\n<\/code><\/pre>\n\n<p>I realize the question is old but it might save some time for someone to find the issue.<\/p>\n"},{"AnswerId":"41248943","CreationDate":"2016-12-20T18:25:14.757","ParentId":null,"OwnerUserId":"3511819","Title":null,"Body":"<p>It's impossible to diagnose your exact problem without more information.<\/p>\n\n<p>I usually specify the <code>input_shape<\/code> parameter of the first layer based on my training data <code>X<\/code>. <\/p>\n\n<p>e.g.<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(32, input_shape=X.shape[0]))\n<\/code><\/pre>\n\n<p>I think you'll want <code>X<\/code> to look something like this:<\/p>\n\n<pre><code>   [\n    [[ 0,  0,  0, ..., 18, 11,  1]],\n    [[ 0,  0,  0, ..., 18,  9,  1]],\n   ....\n   ]\n<\/code><\/pre>\n\n<p>So you could try reshaping it with the following line:<\/p>\n\n<pre><code>X = np.array([[sample] for sample in X])\n<\/code><\/pre>\n"}]}
{"QuestionId":41249293,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2016-12-20T18:52:05.810","AcceptedAnswerId":null,"OwnerUserId":4931135.0,"Title":"How to predict a layer's weights using another model in an end to end fashion?","Body":"<p>I am working on visual question answering task and i want to reproduce the work in this paper <a href=\"https:\/\/arxiv.org\/pdf\/1511.05756v1.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1511.05756v1.pdf<\/a>\nwhere they predict a layer's weights in a network using the question embedding. i wonder can i do that in keras?<br>\ni tried that and i know that for each tensor there is an argument weight\nex: Dense(128 , weight= ) \nbut it expect a real number, and i want to train the network in an end to end.<\/p>\n","answers":[]}
{"QuestionId":41250649,"AnswerCount":1,"Tags":"<deep-learning><caffe><pycaffe>","CreationDate":"2016-12-20T20:21:09.507","AcceptedAnswerId":41250899.0,"OwnerUserId":6786718.0,"Title":"Is it possible to forward the output of a deep-learning network to another network with caffe \/ pycaffe?","Body":"<p>I am using caffe, or more likely pycaffe to train and create my network. I am having a dataset with 5 labels at the end. I had the idea to create one network for each label that can just simply say the score for one class. After having then trained 5 networks I want to compare the outputs of the networks and which one has the highest score. <\/p>\n\n<p>Sadly I do only know how to create one network , but not how to let them interact and moreover how to do something like a max function at the end. I add a picture to describe what I want to do.<\/p>\n\n<p>Moreover, I do not know if this would have a better outcome than just a normal deep neuronal network<a href=\"https:\/\/i.stack.imgur.com\/25ZW6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/25ZW6.png\" alt=\"enter image description here\"><\/a>. <\/p>\n","answers":[{"AnswerId":"41250899","CreationDate":"2016-12-20T20:40:14.803","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>I don't see what you expect to have as the input to this \"max\" function.  Even if you use some sort of <em>is \/ is not<\/em> boundary training, your approach appears to be an inferior version of the <strong>softmax<\/strong> layer available in all popular frameworks.<\/p>\n\n<p>Yes, you can build a multi-channel model, train each channel with a different data set, and then accept the most confident prediction -- but the result will take longer and be less accurate than a cooperative training pass.  Your five channels wind up negotiating their boundaries <em>after<\/em> they've made other parametric assumptions.<\/p>\n\n<p>Feed a single model all the information available from the outset; you'll get faster convergence and more accurate classification.<\/p>\n"}]}
{"QuestionId":41251385,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-12-20T21:16:45.373","AcceptedAnswerId":41251454.0,"OwnerUserId":6121755.0,"Title":"tensorflow saves only the initialized value","Body":"<p>I am trying to save some variables and see if I could restore it later. \nHere is my saving code: <\/p>\n\n<pre><\/pre>\n\n<p>This prints out: <\/p>\n\n<pre><\/pre>\n\n<p>As I restore it: <\/p>\n\n<pre><\/pre>\n\n<p>This gives me : <\/p>\n\n<pre><\/pre>\n\n<p>The restored value of my_c should be 5 since it is the sum of my_a and my_b. However it gives me 4, which is the initialized value of my_c. Could anyone explain why that happens, and how to save the changes to a variable? <\/p>\n","answers":[{"AnswerId":"41251454","CreationDate":"2016-12-20T21:21:46.850","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>In your original code, you have not really assigned the variable named <code>my_c<\/code> (mind you, TensorFlow <code>name<\/code>) to <code>my_a + my_b<\/code>.<\/p>\n\n<p>By writing <code>my_c = tf.add(my_a,my_b)<\/code>, the python variable <code>my_c<\/code> is now different from the <code>tf.Variable<\/code> having <code>name='my_c'<\/code>.<\/p>\n\n<p>When you execute <code>sess.run()<\/code>, you are just executing the operation, not updating that variable.<\/p>\n\n<p>If you want this code to run correctly, use this instead - (see the comments for changes)<\/p>\n\n<pre><code>import tensorflow as tf\nmy_a = tf.Variable(2,name = \"my_a\")\nmy_b = tf.Variable(3,name = \"my_b\")\nmy_c = tf.Variable(4,name=\"my_c\")\n# Use the assign() function to set the new value\nadd = my_c.assign(tf.add(my_a,my_b))\n\nwith tf.Session() as sess:\n    init = tf.initialize_all_variables()\n    sess.run(init)\n    # Execute the add operator\n    sess.run(add)\n    print(\"my_c =  \",sess.run(my_c))\n    saver = tf.train.Saver()\n    saver.save(sess,\"test.ckpt\")\n<\/code><\/pre>\n"}]}
{"QuestionId":41251428,"AnswerCount":1,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2016-12-20T21:20:06.353","AcceptedAnswerId":41251827.0,"OwnerUserId":2712140.0,"Title":"RNN Cell not present in tf.get_collection","Body":"<p>RNN Cells are not showing up when using . What am I missing?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>prints out<\/p>\n\n<pre><\/pre>\n\n<p>Windows 10, Python 3.5<\/p>\n","answers":[{"AnswerId":"41251827","CreationDate":"2016-12-20T21:52:08.933","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>You have not run a <code>__call__<\/code> on the <code>LSTMCell<\/code> which is why you don't see your variables. Try this instead (I'm assuming <code>batch_size=10<\/code> and <code>rnn_size=16<\/code>)<\/p>\n\n<pre><code>import tensorflow as tf\nprint(tf.__version__)\n\nrnn_cell = tf.nn.rnn_cell.LSTMCell(16)\na = tf.placeholder(tf.float32, [10, 16])\nzero = rnn_cell.zero_state(10,tf.float32)\n# The variables are created in the following __call__\nb = rnn_cell(a, zero)\nprint(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES))\n\nother_var = tf.Variable(0)\nprint(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES))\n<\/code><\/pre>\n"}]}
{"QuestionId":41251698,"AnswerCount":2,"Tags":"<python><tensorflow><flops>","CreationDate":"2016-12-20T21:42:16.737","AcceptedAnswerId":43015296.0,"OwnerUserId":562769.0,"Title":"How many FLOPs does tanh need?","Body":"<p>I would like to compute how many flops each layer of LeNet-5 (<a href=\"http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/lecun-01a.pdf\" rel=\"noreferrer\">paper<\/a>) needs. Some papers give FLOPs for other architectures in total (<a href=\"https:\/\/arxiv.org\/pdf\/1512.06473.pdf\" rel=\"noreferrer\">1<\/a>, <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/cUD3IF5NMOk\" rel=\"noreferrer\">2<\/a>, <a href=\"https:\/\/arxiv.org\/pdf\/1511.06530.pdf\" rel=\"noreferrer\">3<\/a>) However, those papers don't give details on how to compute the number of FLOPs and I have no idea how many FLOPs are necessary for the non-linear activation functions. For example, how many FLOPs are necessary to calculate <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/activation_functions_#tanh\" rel=\"noreferrer\"><\/a>?<\/p>\n\n<p>I guess this will be implementation and probably also hardware-specific. However, I am mainly interested in getting an order of magnitude. Are we talking about 10 FLOPs? 100 FLOPs? 1000 FLOPs? So chose any architecture \/ implementation you want for your answer. (Although I'd appreciate answers which are close to \"common\" setups, like an Intel i5 \/ nvidia GPU \/ Tensorflow)<\/p>\n","answers":[{"AnswerId":"43015296","CreationDate":"2017-03-25T10:32:41.570","ParentId":null,"OwnerUserId":"3344139","Title":null,"Body":"<p>Note: This answer is not python specific, but I don't think that something like tanh is fundamentally different across languages.<\/p>\n\n<p>Tanh is usually implemented by defining an upper and lower bound, for which 1 and -1 is returned, respectively. The intermediate part is approximated with different functions as follows:<\/p>\n\n<pre><code> Interval 0  x_small               x_medium               x_large \n  tanh(x) |  x  |  polynomial approx.  |  1-(2\/(1+exp(2x)))  |  1\n<\/code><\/pre>\n\n<p>There exist polynomials that are accurate up to single precisision floating points, and also for double precision.\nThis algorithm is called Cody-Waite algorithm.<\/p>\n\n<p>Citing <a href=\"https:\/\/www.math.utah.edu\/~beebe\/software\/ieee\/tanh.pdf\" rel=\"nofollow noreferrer\">this description<\/a> (you can find more information about the mathematics there as well, e.g. how to determine x_medium),\nCody  and  Waite\u2019s  rational  form  requires  four  multiplications,  three  additions, and one division in single precision, and seven multiplications, six additions, and one division in double precision.<\/p>\n\n<p>For negative x, you can compute |x| and flip the sign.\nSo you need comparisons for which interval x is in, and evaluate the according approximation.\nThat's a total of:<\/p>\n\n<ol>\n<li>Taking the absolute value of x<\/li>\n<li>3 comparisons for the interval<\/li>\n<li>Depending on the interval and the float precision, 0 to a few FLOPS for the exponential, check <a href=\"https:\/\/stackoverflow.com\/questions\/3979942\/c-complexity-of-the-implementation-of-exp-in-cmath-and-real-cost-of-the-cost\">this question<\/a> on how to compute the exponential.<\/li>\n<li>One comparison to decide whether to flip the sign.<\/li>\n<\/ol>\n\n<p>Now, this is a report from 1993, but I don't think much has changed here.<\/p>\n"},{"AnswerId":"43021821","CreationDate":"2017-03-25T20:51:25.103","ParentId":null,"OwnerUserId":"5769463","Title":null,"Body":"<p>If we look at the glibc-implementation of <code>tanh(x)<\/code>, we see:<\/p>\n\n<ol>\n<li>for  <code>x<\/code> values greater 22.0 and double precision, <code>tanh(x)<\/code> can be safely assumed to be 1.0, so there are almost no costs.<\/li>\n<li>for very small <code>x<\/code>, (let's say <code>x&lt;2^(-55)<\/code>) another cheap approximation is possible: <code>tanh(x)=x(1+x)<\/code>, so only two floating point operations are needed.<\/li>\n<li>for the values in beetween, one can rewrite <code>tanh(x)=(1-exp(-2x))\/(1+exp(-2x))<\/code>. However, one must be accurate, because <code>1-exp(t)<\/code> is very problematic for small t-values due to loss of significance, so one uses <code>expm(x)=exp(x)-1<\/code> and calculates <code>tanh(x)=-expm1(-2x)\/(expm1(-2x)+2)<\/code>.<\/li>\n<\/ol>\n\n<p>So basically, the worst case is about 2 times the number of operation needed for <code>expm1<\/code>, which is a pretty complicated function. The best way is probably just to measure the time needed to calculate <code>tanh(x)<\/code> compared with a time needed for a simple multiplication of two doubles.<\/p>\n\n<p>My (sloppy) experiments on an Intel-processor yielded the following result, which gives a rough idea:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LXlRJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LXlRJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So for very small and numbers >22 there are almost no costs, for numbers up to <code>0.1<\/code> we pay  6 FLOPS, then the costs rise to about 20 FLOPS per <code>tanh<\/code>-caclulation.<\/p>\n\n<p>The key takeaway: the costs of calculating <code>tanh(x)<\/code> are dependent on the parameter <code>x<\/code> and maximal costs are somewhere between 10 and 100 FLOPs.<\/p>\n\n<hr>\n\n<p>There is an Intel-instruction called <a href=\"http:\/\/x86.renejeschke.de\/html\/file_module_x86_id_79.html\" rel=\"nofollow noreferrer\"><code>F2XM1<\/code><\/a> which computes <code>2^x-1<\/code> for <code>-1.0&lt;x&lt;1.0<\/code>, which could be used for computing <code>tanh<\/code>, at least for some range. However if <a href=\"http:\/\/www.agner.org\/optimize\/instruction_tables.pdf\" rel=\"nofollow noreferrer\">agner's tables<\/a> are to be believed, this operation's costs are about 60 FLOPs.<\/p>\n\n<hr>\n\n<p>Another problem is the vectorization - the normal glibc-implementation is not vectorized, as far as I can see. So if your program uses vectorization and has to use an unvectorized <code>tanh<\/code> implementation it will slowdown the program even more. For this, the intel compiler has the mkl-library which <a href=\"https:\/\/software.intel.com\/sites\/products\/documentation\/doclib\/mkl\/vm\/functions\/tanh.html\" rel=\"nofollow noreferrer\">vectorizes <code>tanh<\/code><\/a> among the others. <\/p>\n\n<p>As you can see in the tables the maximal costs are about 10 clocks per operation (costs of a float-operation is about 1 clock).<\/p>\n\n<hr>\n\n<p>I guess there is a chance you could win some FLOPs by using <code>-ffast-math<\/code> compiler option, which results in a faster but less precise program (that is an option for Cuda or c\/c++, not sure whether this can be done for python\/numpy).<\/p>\n\n<hr>\n\n<p>The c++ code which produced the data for the figure (compiled with g++ -std=c++11 -O2). Its intend is not to give the exact number, but the first impression about the costs:<\/p>\n\n<pre class=\"lang-cpp prettyprint-override\"><code>#include &lt;chrono&gt;\n#include &lt;iostream&gt;\n#include &lt;vector&gt;\n#include &lt;math.h&gt;\n\nint main(){\n   const std::vector&lt;double&gt; starts={1e-30, 1e-18, 1e-16, 1e-10, 1e-5, 1e-2, 1e-1, 0.5, 0.7, 0.9, 1.0, 2.0, 10, 20, 23, 100,1e3, 1e4};\n   const double FACTOR=1.0+1e-11;\n   const size_t ITER=100000000; \n\n\n   \/\/warm-up:\n   double res=1.0;\n      for(size_t i=0;i&lt;4*ITER;i++){\n      res*=FACTOR;\n   }\n   \/\/overhead:\n   auto begin = std::chrono::high_resolution_clock::now();\n   for(size_t i=0;i&lt;ITER;i++){\n      res*=FACTOR;\n   }\n   auto end = std::chrono::high_resolution_clock::now();\n   auto overhead=std::chrono::duration_cast&lt;std::chrono::nanoseconds&gt;(end-begin).count(); \n   \/\/std::cout&lt;&lt;\"overhead: \"&lt;&lt;overhead&lt;&lt;\"\\n\";\n\n\n   \/\/experiments:\n   for(auto start : starts){\n       begin=std::chrono::high_resolution_clock::now();\n       for(size_t i=0;i&lt;ITER;i++){\n           res*=tanh(start);\n           start*=FACTOR;\n       }\n       auto end = std::chrono::high_resolution_clock::now();\n       auto time_needed=std::chrono::duration_cast&lt;std::chrono::nanoseconds&gt;(end-begin).count();\n       std::cout&lt;&lt;start&lt;&lt;\" \"&lt;&lt;time_needed\/overhead&lt;&lt;\"\\n\"; \n   }\n\n   \/\/overhead check:\n   begin = std::chrono::high_resolution_clock::now();\n   for(size_t i=0;i&lt;ITER;i++){\n      res*=FACTOR;\n   }\n   end = std::chrono::high_resolution_clock::now();\n   auto overhead_new=std::chrono::duration_cast&lt;std::chrono::nanoseconds&gt;(end-begin).count(); \n   std::cerr&lt;&lt;\"overhead check: \"&lt;&lt;overhead\/overhead_new&lt;&lt;\"\\n\";\n   std::cerr&lt;&lt;res;\/\/don't optimize anything out...\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":41252178,"AnswerCount":0,"Tags":"<python><theano><theano.scan>","CreationDate":"2016-12-20T22:22:49.837","AcceptedAnswerId":null,"OwnerUserId":2751847.0,"Title":"using Theano.scan for iterating over ndarray with a condition","Body":"<p>Pardon me for a basic question(I am new to Theano)!<\/p>\n\n<p>I want to get the difference of 2 matrices for only those positions that satisfy a condition. So, suppose we have 2 matrices A and B, this(python equivalent code) is what I want to calculate:<\/p>\n\n<pre><\/pre>\n\n<p>I want a Theano equivalent code to calculate the <em>sum<\/em>.\nI know there is <strong>Theano.scan<\/strong> that can be used to scan an <em>ndarray<\/em>, but I could not get any example that has an if condition.\nThank you in advance :)<\/p>\n","answers":[]}
{"QuestionId":41252184,"AnswerCount":2,"Tags":"<numpy><theano><numpy-broadcasting><softmax>","CreationDate":"2016-12-20T22:23:28.053","AcceptedAnswerId":41256071.0,"OwnerUserId":2320784.0,"Title":"masked softmax in theano","Body":"<p>I am wondering if it possible to apply a  mask before performing theano.tensor.nnet.softmax?<\/p>\n\n<p>This is the behavior I am looking for:<\/p>\n\n<pre><\/pre>\n\n<p>Note that  and  are matrices, so I would like the softmax with work on an entire matrix and perform row-wise masked softmax.<\/p>\n\n<p>Also the output should be the same shape as , so the solution can not do advanced indexing e.g. <\/p>\n","answers":[{"AnswerId":"41252367","CreationDate":"2016-12-20T22:41:25.063","ParentId":null,"OwnerUserId":"2320784","Title":null,"Body":"<p><a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html?highlight=switch#theano.tensor.switch\" rel=\"nofollow noreferrer\">theano.tensor.switch<\/a> is one way to do this.<\/p>\n\n<p>In the computational graph you can do the following:<\/p>\n\n<pre><code>a_mask = theano.tensor.switch(m, a, np.NINF)\nsm = theano.tensor.softmax(a_mask)\n<\/code><\/pre>\n\n<p>hope it helps others.<\/p>\n"},{"AnswerId":"41256071","CreationDate":"2016-12-21T05:56:30.087","ParentId":null,"OwnerUserId":"7324129","Title":null,"Body":"<pre><code>def masked_softmax(a, m, axis):\n    e_a = T.exp(a)\n    masked_e = e_a * m\n    sum_masked_e = T.sum(masked_e, axis, keepdims=True)\n    return masked_e \/ sum_masked_e\n<\/code><\/pre>\n"}]}
{"QuestionId":41252495,"AnswerCount":4,"Tags":"<tensorflow><keras><activation-function>","CreationDate":"2016-12-20T22:54:06.393","AcceptedAnswerId":null,"OwnerUserId":7211678.0,"Title":"Keras How to use max_value in Relu activation function","Body":"<p>Relu function as defined in keras\/activation.py is:<\/p>\n\n<pre><\/pre>\n\n<p>It has a max_value which can be used to clip the value. Now how can this be used\/called in the code?\nI have tried the following:\n(a) <\/p>\n\n<pre><\/pre>\n\n<p>(b) <\/p>\n\n<pre><\/pre>\n\n<p>same error<\/p>\n\n<p>(c) <\/p>\n\n<pre><\/pre>\n\n<p>The problem with this is that it expects the input to be present in the first value. The error is 'relu() takes at least 1 argument (1 given)'<\/p>\n\n<p>So how do I make this a layer?<\/p>\n\n<pre><\/pre>\n\n<p>has the same issue 'relu() takes at least 1 argument (1 given)'<\/p>\n\n<p>If this file cannot be used as layer, then there seems to be no way of specifying a clip value to Relu. This implies that the comment here <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2119\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/2119<\/a> closing a proposed change is wrong...\nAny thoughts? Thanks!<\/p>\n","answers":[{"AnswerId":"53263887","CreationDate":"2018-11-12T14:08:33.063","ParentId":null,"OwnerUserId":"8788960","Title":null,"Body":"<p>That is as easy as one lambda :<\/p>\n\n<pre><code>from keras.activations import relu\nclipped_relu = lambda x: relu(x, max_value=3.14)\n<\/code><\/pre>\n\n<p>Then use it like this:<\/p>\n\n<pre><code>model.add(Conv2D(64, (3, 3)))\nmodel.add(Activation(clipped_relu))\n<\/code><\/pre>\n\n<p>When reading a model saved in <code>hdf5<\/code> use <code>custom_objects<\/code> dictionary:<\/p>\n\n<pre><code>model = load_model(model_file, custom_objects={'&lt;lambda&gt;': clipped_relu})\n<\/code><\/pre>\n"},{"AnswerId":"42954611","CreationDate":"2017-03-22T14:31:47.683","ParentId":null,"OwnerUserId":"7751963","Title":null,"Body":"<p>You can use the ReLU function of the Keras backend. Therefore, first import the backend:<\/p>\n\n<pre><code>from keras import backend as K\n<\/code><\/pre>\n\n<p>Then, you can pass your own function as activation using backend functionality.\nThis would look like<\/p>\n\n<pre><code>def relu_advanced(x):\n    return K.relu(x, max_value=250)\n<\/code><\/pre>\n\n<p>Then you can use it like<\/p>\n\n<pre><code>model.add(Dense(512, input_dim=1, activation=relu_advanced))\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>model.add(Activation(relu_advanced))\n<\/code><\/pre>\n\n<p>Unfortunately, you must hard code additional arguments.\nTherefore, it is better to use a function, that returns your function and passes your custom values:<\/p>\n\n<pre><code>def create_relu_advanced(max_value=1.):        \n    def relu_advanced(x):\n        return K.relu(x, max_value=K.cast_to_floatx(max_value))\n    return relu_advanced\n<\/code><\/pre>\n\n<p>Then you can pass your arguments by either<\/p>\n\n<pre><code>model.add(Dense(512, input_dim=1, activation=create_relu_advanced(max_value=250)))\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>model.add(Activation(create_relu_advanced(max_value=250)))\n<\/code><\/pre>\n"},{"AnswerId":"57062516","CreationDate":"2019-07-16T17:22:02.337","ParentId":null,"OwnerUserId":"7214578","Title":null,"Body":"<p>Tested below, it'd work:<\/p>\n\n<pre><code>import keras\n\ndef clip_relu (x): \n    return keras.activations.relu(x, max_value=1.)\n\npredictions=Dense(num_classes,activation=clip_relu,name='output')\n<\/code><\/pre>\n"},{"AnswerId":"46797071","CreationDate":"2017-10-17T18:41:24.530","ParentId":null,"OwnerUserId":"8791878","Title":null,"Body":"<p>This is what I did using <code>Lambda<\/code> layer to implement clip relu:\nStep 1: define a function to do reluclip:<\/p>\n\n<pre><code>def reluclip(x, max_value = 20):\n    return K.relu(x, max_value = max_value)\n<\/code><\/pre>\n\n<p>Step 2: add <code>Lambda<\/code> layer into model:\n<code>y = Lambda(function = reluclip)(y)<\/code><\/p>\n"}]}
{"QuestionId":41254340,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-21T02:49:56.870","AcceptedAnswerId":null,"OwnerUserId":469679.0,"Title":"The prediction output repeats infinitely","Body":"<p>I'm following the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/input_fn\/\" rel=\"nofollow noreferrer\">tutorial<\/a> for generating input functions on tensorflow.org.<\/p>\n\n<p>Everything works fine until I try to print the predictions (it should be just 6 predictions).<\/p>\n\n<pre><\/pre>\n\n<p>I get this output: .<\/p>\n\n<p>If I try to convert the generator to a list with . The program freezes.<\/p>\n\n<p>If I try to get only the first 30 items (even if there should be only 6):<\/p>\n\n<pre><\/pre>\n\n<p>I get the following:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see it's repeating values ad infinitum.<\/p>\n\n<p>Am I missing something?<\/p>\n\n<p>Tensorflow version: 0.12.0-rc1<\/p>\n\n<p>Python version: 2.7.6<\/p>\n","answers":[{"AnswerId":"41272562","CreationDate":"2016-12-21T21:55:38.243","ParentId":null,"OwnerUserId":"6836618","Title":null,"Body":"<p>We recently revised the \"Building Input Functions with tf.contrib.learn\" tutorial to be up to date with the latest default behavior of predict() in Estimators, which returns a generator. The revised tutorial text is here:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/input_fn\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/input_fn\/<\/a><\/p>\n\n<p>And the latest code is here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/input_fn\/boston.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/input_fn\/boston.py<\/a><\/p>\n\n<p>Here's the relevant section, which uses itertools.islice, as in your last attempt:<\/p>\n\n<pre><code>y = regressor.predict(input_fn=lambda: input_fn(prediction_set))\n# .predict() returns an iterator; convert to a list and print predictions                                                      \npredictions = list(itertools.islice(y, 6))\nprint(\"Predictions: {}\".format(str(predictions)))\n<\/code><\/pre>\n\n<p>I'm not seeing the behavior where the prediction values repeat when the second argument to islice is increased to 30. Could you try pulling the latest code from GitHub at the link above and let me know if you still encounter that?<\/p>\n"}]}
{"QuestionId":41254583,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2016-12-21T03:23:39.167","AcceptedAnswerId":41282455.0,"OwnerUserId":5464787.0,"Title":"Function to access to a tensor in Theano","Body":"<p>I'm working with Theano library.\nI have 1000 object of size 114, defined as a variable of size 1000x114.\nI have a function that takes that variable and produces 1000 numbers between 0 and 113.\nI need to create a function that takes for each one of the 1000 object, the number in the position given by the previous function.\nHow can I do it?<\/p>\n\n<p>This is what works:<\/p>\n\n<pre><\/pre>\n\n<p>This is what doesn't work:<\/p>\n\n<pre><\/pre>\n\n<p>The result of the num_fn is simply the same data I give it as input.<\/p>\n","answers":[{"AnswerId":"41282455","CreationDate":"2016-12-22T11:45:33.537","ParentId":null,"OwnerUserId":"97702","Title":null,"Body":"<p>I'm don't know what the +48 is supposed to do but I'm assuming index + 48 is your vector of indices your function returned. The way theano (and numpy) indexing works is that your second approach, using :, will return the rows with indices in index. Instead, you need to use a range, e.g.:<\/p>\n\n<pre><code>num = input_var[T.arange(T.shape(input_var)[0]), index + 48]\nnum_fn = theano.function([input_var], num, name=\"num function\")\n<\/code><\/pre>\n"}]}
{"QuestionId":41254808,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-21T03:50:49.940","AcceptedAnswerId":null,"OwnerUserId":1610952.0,"Title":"How to split a 1-D tensor by a certain index tensor?","Body":"<p>There's a 1-D tensor of int32. I'd like to replace the elements before the first appearing 1 with 0. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41257555","CreationDate":"2016-12-21T07:43:37.930","ParentId":null,"OwnerUserId":"1610952","Title":null,"Body":"<p>I myself got the answer.<\/p>\n\n<pre><code>import numpy as np\na = np.array([5, 4, 1, 3, 1, 2, 3, 3, 1, 5], np.int32)\nfirst_ind = np.where(a == 1)[0][0] # =&gt; 2\nresult = np.concatenate((np.zeros((first_ind,)), a[first_ind:]))\n# =&gt;[ 0.  0.  1.  3.  1.  2.  3.  3.  1.  5.]\n\nimport tensorflow as tf\n_a = tf.convert_to_tensor(a)\n_first_ind = tf.where(tf.equal(_a, 1))[0]\n\nzero_padding = tf.zeros(tf.to_int32(_first_ind), tf.int32)\n_a_back = tf.slice(_a, _first_ind, [-1])\nout = tf.concat(0, (zero_padding, _a_back))\nwith tf.Session() as sess:\n    print out.eval()\n    #=&gt; [0 0 1 3 1 2 3 3 1 5]\n<\/code><\/pre>\n"}]}
{"QuestionId":41255031,"AnswerCount":1,"Tags":"<memory-leaks><tensorflow><deep-learning><text-classification>","CreationDate":"2016-12-21T04:20:06.947","AcceptedAnswerId":41255301.0,"OwnerUserId":4076922.0,"Title":"Memory leak evaluating CNN model for text clasification","Body":"<p>I've been doing some adaptation to code in this blog about CNN for text clasification: \n<a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"nofollow noreferrer\">http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/<\/a><\/p>\n\n<p>Everything works fine! But when I try to use the model trained to predict new instances it consumes all memory available. It seems that it's not liberating any memory when evaluates and load all the model again and again. As far as I know memory should be liberated after every sess.run command.<\/p>\n\n<p>Here is the part of the code I'm working with:<\/p>\n\n<pre><\/pre>\n\n<p>Any help will be much appreciated<\/p>\n\n<p>Cheers<\/p>\n","answers":[{"AnswerId":"41255301","CreationDate":"2016-12-21T04:51:12.663","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Your training loop creates new TensorFlow operations (<code>tf.reduce_max()<\/code>, <code>tf.nn.softmax()<\/code> and <code>tf.histogram_summary()<\/code>) in each iteration, which will lead to more memory being consumed over time. TensorFlow is most efficient when you run the same graph many times, because it can amortize the cost of optimizing the graph over multiple executions. Therefore,\nto get the best performance, you should revise your program so that you create each of these operations <strong>once<\/strong>, before the <code>for x_test_batch in batches:<\/code> loop, and then re-use the same operations in each iteration.<\/p>\n"}]}
{"QuestionId":41255268,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2016-12-21T04:47:16.447","AcceptedAnswerId":41256212.0,"OwnerUserId":3521339.0,"Title":"Keras LSTM input dimensions with one hot text embedding","Body":"<p>I have 70k samples of text which I have embedded using Keras 'one hot' preprocessing. This gives me an array of  which I then pad for a length of 28 (the longest sample length). All I am trying to do is predict these values to some categorical label (0 to 5 lets say). When I train the model I cannot get anything beyond -.13% accuracy (currently my error is <a href=\"https:\/\/imgur.com\/a\/s0Duw\" rel=\"nofollow noreferrer\">this<\/a> I have tried many ways to pass the input). <\/p>\n\n<p><a href=\"https:\/\/imgur.com\/a\/lvYrJ\" rel=\"nofollow noreferrer\">This is my data<\/a> currently and am just trying to create a simple LSTM. Again my data is X -> [length of 28 integer values, embeddings] and Y -> [1 integer of length 3, (100, 143 etc.)]. Any idea what I am doing wrong?? I have asked many people and no one has been able to help. Here is the code for my model... any ideas? :( <\/p>\n\n<pre><\/pre>\n\n<p><strong>Edit:<\/strong> <\/p>\n\n<p>using  seems to work however still the accuracy never improves, I am at a loss of what to do.<\/p>\n","answers":[{"AnswerId":"41256212","CreationDate":"2016-12-21T06:07:42.210","ParentId":null,"OwnerUserId":"6347629","Title":null,"Body":"<p>I have two suggestions.<\/p>\n\n<ol>\n<li>Use one hot representation for the target variable(y) also. If you give Y as integer, it will become a regression problem. Only if you give a one hot encoding, it becomes a classification problem.<\/li>\n<li>Try word2vec embedding when you have large amount of text, instead of one hot embedding.<\/li>\n<\/ol>\n\n<pre><code>\noptimizer = RMSprop(lr=0.01) \nembedding_vecor_length = 32\nmax_review_length = 28\nnb_classes= 8\nmodel = Sequential()\nmodel.add(Embedding(input_dim=900, output_dim=embedding_vecor_length,\n                    input_length=max_review_length)) \n\nmodel.add(LSTM(150))\n\n#output_dim is a categorical variable with 8 classes\nmodel.add(Dense(output_dim=nb_classes, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])\nprint(model.summary())\n\nmodel.fit(X_train, y_train, nb_epoch=3, batch_size=64)\n\n# Final evaluation of the model\nscores = model.evaluate(X_test, y_test, verbose=0)\nprint(\"Accuracy: %.2f%%\" % (scores[1]*100))\n\n<\/code><\/pre>\n"}]}
{"QuestionId":41256315,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2016-12-21T06:16:17.527","AcceptedAnswerId":null,"OwnerUserId":6313397.0,"Title":"saver.restore() issue in tensorflow 0.12 version","Body":"<p>There is a problem with saver.restore in tensorflow 0.12 version. While i am saving the session to current directory, it is managed to save. But while i am trying to restoring the session, it is not working for same path. When i try with absolute path it is working.<\/p>\n\n<p>Can anyone help me in solving this issue ?<\/p>\n\n<pre><\/pre>\n\n<p>Output :<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41256846,"AnswerCount":1,"Tags":"<graph><machine-learning><computer-vision><tensorflow><deep-learning>","CreationDate":"2016-12-21T06:56:27.113","AcceptedAnswerId":41256895.0,"OwnerUserId":7324309.0,"Title":"How to output the predicted value(label) in tensorflow MNIST tutorial?","Body":"<p>In tensorflow's tutorial of MNIST, the last step is to output the test accuracy of the model using the following code:<\/p>\n\n<pre><\/pre>\n\n<p>However, I was wondering how can I modify this code to output the predicted value(label) of the test set, instead of just printing out the accuracy? <\/p>\n\n<p>Here's the link of the tutorial: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/<\/a><\/p>\n","answers":[{"AnswerId":"41256895","CreationDate":"2016-12-21T06:59:19.453","ParentId":null,"OwnerUserId":"1349525","Title":null,"Body":"<p>Something like this should work<\/p>\n\n<pre><code>print(sess.run(tf.argmax(y, 1), feed_dict={x: mnist.test.images}))\n<\/code><\/pre>\n\n<p>Because <code>y<\/code> in tutorial is tensor where column with index <code>j<\/code> describes how likely image in row <code>i<\/code> is number <code>j<\/code>, so <code>tf.argmax<\/code> just returns index of column with highest probability for each row.<\/p>\n\n<p>PS Sorry for my English<\/p>\n"}]}
{"QuestionId":41258012,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-21T08:12:54.740","AcceptedAnswerId":null,"OwnerUserId":7324628.0,"Title":"Tensorflow inception v3 image retrain unable to test on new .jpg image","Body":"<p>I have successfully retrained tensorflows image_retrain inception v3 model in ubuntu 16.04 for my own image classification. The performance results are good, so I tried with  script to predict new image (.jpg). But I am getting error at <\/p>\n\n<p><\/p>\n\n<p><strong><em>error log<\/em><\/strong> <\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): NodeDef mentions attr 'dct_method' not in Op image:uint8; attr=channels:int,default=0; attr=ratio:int,default=1; attr=fancy_upscaling:bool,default=true; attr=try_recover_truncated:bool,default=false; attr=acceptable_fraction:float,default=1>; NodeDef: DecodeJpeg = DecodeJpegacceptable_fraction=1, channels=3, dct_method=\"\", fancy_upscaling=true, ratio=1, try_recover_truncated=false, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"\n       [[Node: DecodeJpeg = DecodeJpegacceptable_fraction=1, channels=3, dct_method=\"\", fancy_upscaling=true, ratio=1, try_recover_truncated=false, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n<\/blockquote>\n\n<p>Kindly help me to resolve<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"41522739","CreationDate":"2017-01-07T14:53:17.453","ParentId":null,"OwnerUserId":"3590164","Title":null,"Body":"<p>I just ran into this problem myself.<\/p>\n\n<p>As far as I can tell, there's a mismatch between the version of tensorflow you used to generate the output graph and the version you're now using to run the label_image.py script.<\/p>\n\n<p>I spent a lot of time trying to track down the change and figuring out how to run tensorflow in the same version as I previously used to generate the graph.<\/p>\n\n<p>In the end, I found a hacky solution that solved my problem:<\/p>\n\n<p>After loading the graph from your file<\/p>\n\n<pre><code>graph_def.ParseFromString(f.read())\n<\/code><\/pre>\n\n<p>Add<\/p>\n\n<pre><code>del(graph_def.node[1].attr[\"dct_method\"])\n<\/code><\/pre>\n\n<p>This is not a great method, and if possible you should try to retrain the graph and make sure the tensorflow versions are the same.<\/p>\n"}]}
{"QuestionId":41258243,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2016-12-21T08:27:10.257","AcceptedAnswerId":41260477.0,"OwnerUserId":4763311.0,"Title":"How to insert a value at given index or indices ( mutiple index ) into a Tensor?","Body":"<p>I am having a matrix of size 12 x 10 and a list of document_length = [4,3,5] . The first 4 rows belongs to sentences of one document, next 3 to other doc and last 5 to other doc. I need to get a new matrix or Tensor of size 15 x 10 in  the following manner. <\/p>\n\n<p>I need to insert a row of zeros, in such a way that, here max doc length is 5. So, 1 row of zero has to be inserted in 5th index ( because first doc length is 4 ) . Then 2 row of zeros at index 9 and 10 ( second doc length is 3 ) and so on . If the question is confusing please let me know . <\/p>\n\n<p>I have feed 12 x 10 to GRU or LSTM , but it has to be in a 3d . To get a proper 3D, I need to convert 12 X 10 to 15 x 10 . Any help ?<\/p>\n","answers":[{"AnswerId":"41260477","CreationDate":"2016-12-21T10:23:18.143","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Here is a way to do it. You first break down your 12x10 Tensor into a list of size 12. This is done using the <code>tf.unstack()<\/code> function.<\/p>\n\n<pre><code># Assume `words` is my 12x10 tensor\ntensors = tf.unstack(words, 12, axis=0) \n<\/code><\/pre>\n\n<p>After that, you can insert <code>tf.zeros()<\/code> tensors into the list at your indices. Here for simplicity, I will just insert a tensor in index <code>4<\/code>.<\/p>\n\n<pre><code>b = tf.zeros((10))\ntensors.insert(4, b)\n<\/code><\/pre>\n\n<p>Finally, <code>tf.stack()<\/code> your list to re-construct it.<\/p>\n\n<pre><code>words = tf.stack(tensors, axis=0)\n<\/code><\/pre>\n"}]}
{"QuestionId":41258391,"AnswerCount":7,"Tags":"<tensorflow><embedding><tensorboard>","CreationDate":"2016-12-21T08:35:29.703","AcceptedAnswerId":41262360.0,"OwnerUserId":1909965.0,"Title":"TensorBoard Embedding Example?","Body":"<p>I'm looking for a tensorboard embedding example, with iris data for example like the embedding projector <a href=\"http:\/\/projector.tensorflow.org\/\" rel=\"noreferrer\">http:\/\/projector.tensorflow.org\/<\/a><\/p>\n\n<p>But unfortunately i couldn't find one. Just a little bit information about how to do it in <a href=\"https:\/\/www.tensorflow.org\/how_tos\/embedding_viz\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/embedding_viz\/<\/a><\/p>\n\n<p>Does someone knows a basic tutorial for this functionality?<\/p>\n\n<p>Basics:<\/p>\n\n<p>1) Setup a 2D tensor variable(s) that holds your embedding(s).<\/p>\n\n<pre><\/pre>\n\n<p>2) Periodically save your embeddings in a LOG_DIR.<\/p>\n\n<p>3) Associate metadata with your embedding.<\/p>\n","answers":[{"AnswerId":"48260680","CreationDate":"2018-01-15T10:04:14.133","ParentId":null,"OwnerUserId":"2061991","Title":null,"Body":"<p>To take pretrained embeddings and visualize it on tensorboard.<\/p>\n\n<p>embedding -> trained embedding<\/p>\n\n<p>metadata.tsv -> metadata information<\/p>\n\n<p>max_size -> embedding.shape[0]<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.contrib.tensorboard.plugins import projector\n\nsess = tf.InteractiveSession()\n\nwith tf.device(\"\/cpu:0\"):\n    tf_embedding = tf.Variable(embedding, trainable = False, name = \"embedding\")\n\ntf.global_variables_initializer().run()\npath = \"tensorboard\"\nsaver = tf.train.Saver()\nwriter = tf.summary.FileWriter(path, sess.graph)\nconfig = projector.ProjectorConfig()\nembed = config.embeddings.add()\nembed.tensor_name = \"embedding\"\nembed.metadata_path = \"metadata.tsv\"\nprojector.visualize_embeddings(writer, config)\nsaver.save(sess, path+'\/model.ckpt' , global_step=max_size )\n<\/code><\/pre>\n\n<p>$ tensorboard --logdir=\"tensorboard\" --port=8080 <\/p>\n"},{"AnswerId":"41262360","CreationDate":"2016-12-21T12:00:07.150","ParentId":null,"OwnerUserId":"1831325","Title":null,"Body":"<p>It sounds like you want to get the Visualization section with t-SNE running on TensorBoard. As you've described, the API of Tensorflow has only provided the bare essential commands in the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/embedding_viz\/\" rel=\"nofollow noreferrer\">how-to document<\/a>.<\/p>\n\n<p><strong>I\u2019ve uploaded my working solution with the MNIST dataset to <a href=\"https:\/\/github.com\/normanheckscher\/mnist-tensorboard-embeddings\/blob\/master\/mnist_t-sne.py\" rel=\"nofollow noreferrer\">my GitHub repo<\/a>.<\/strong><\/p>\n\n<p>Yes, it is broken down into three general steps:<\/p>\n\n<ol>\n<li>Create metadata for each dimension.<\/li>\n<li>Associate images with each dimension.<\/li>\n<li>Load the data into TensorFlow and save the embeddings in a LOG_DIR. <\/li>\n<\/ol>\n\n<p>Only generic details are inculded with the TensorFlow r0.12 release. There is no full code example that I\u2019m aware of within the official source code. <\/p>\n\n<p>I found that there were two tasks involved that were not documented in the how to.<\/p>\n\n<ol>\n<li>Preparing the data from the source<\/li>\n<li>Loading the data into a <code>tf.Variable<\/code><\/li>\n<\/ol>\n\n<p>While TensorFlow is designed for the use of GPUs, in this situation I opted to generate the t-SNE visualization with the CPU as the process took up more memory than my MacBookPro GPU has access to. API access to the MNIST dataset is included with TensorFlow, so I used that. The MNIST data comes as a structured a numpy array. Using the <code>tf.stack<\/code> function enables this dataset to be stacked into a list of tensors which can be embedded into a visualization. The following code contains is how I extracted the data and setup the TensorFlow embedding variable. <\/p>\n\n<pre><code>with tf.device(\"\/cpu:0\"):\n    embedding = tf.Variable(tf.stack(mnist.test.images[:FLAGS.max_steps], axis=0), trainable=False, name='embedding')\n<\/code><\/pre>\n\n<p>Creating the metadata file was perfomed with the slicing of a numpy array.<\/p>\n\n<pre><code>def save_metadata(file):\n    with open(file, 'w') as f:\n        for i in range(FLAGS.max_steps):\n            c = np.nonzero(mnist.test.labels[::1])[1:][0][i]\n            f.write('{}\\n'.format(c))\n<\/code><\/pre>\n\n<p>Having an image file to associate with is as described in the how-to. I've uploaded a png file of the first 10,000 MNIST images to <a href=\"https:\/\/github.com\/normanheckscher\/mnist-tensorboard-embeddings\/blob\/master\/mnist_t-sne.py\" rel=\"nofollow noreferrer\">my GitHub<\/a>.<\/p>\n\n<p>So far TensorFlow works beautifully for me, it\u2019s computationaly quick, well documented and the API appears to be functionally complete for anything I am about to do for the moment. I look forward to generating some more visualizations with custom datasets over the coming year. This post was edited from <a href=\"http:\/\/www.normansoft.com\/blog\/embedding-tensorboard-visua.html\" rel=\"nofollow noreferrer\">my blog<\/a>. Best of luck to you, please let me know how it goes. :)<\/p>\n"},{"AnswerId":"42678884","CreationDate":"2017-03-08T18:15:56.437","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>An issue has been raised in the TensorFlow to GitHub repository: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6322\" rel=\"nofollow noreferrer\">No real code example for using the tensorboard embedding tab #6322<\/a> (<a href=\"https:\/\/web.archive.org\/web\/20170308181419\/https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6322\" rel=\"nofollow noreferrer\">mirror<\/a>).<\/p>\n\n<p>It contains some interesting pointers.<\/p>\n\n<hr>\n\n<p>If interested, some code that uses TensorBoard embeddings to display character and word embeddings:\n<a href=\"https:\/\/github.com\/Franck-Dernoncourt\/NeuroNER\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Franck-Dernoncourt\/NeuroNER<\/a><\/p>\n\n<p>Example:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SLXk1.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SLXk1.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/U5y9j.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/U5y9j.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>FYI: <a href=\"https:\/\/stackoverflow.com\/q\/42679552\/395857\">How can I select which checkpoint to view in TensorBoard's embeddings tab?<\/a><\/p>\n"},{"AnswerId":"42676076","CreationDate":"2017-03-08T15:59:01.930","ParentId":null,"OwnerUserId":"5231231","Title":null,"Body":"<p>I've used <a href=\"https:\/\/github.com\/facebookresearch\/fastText\/\" rel=\"nofollow noreferrer\">FastText's pre-trained word vectors<\/a> with TensorBoard.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nimport tensorflow as tf\nimport numpy as np\nimport fasttext\nfrom tensorflow.contrib.tensorboard.plugins import projector\n\n# load model\nword2vec = fasttext.load_model('wiki.en.bin')\n\n# create a list of vectors\nembedding = np.empty((len(word2vec.words), word2vec.dim), dtype=np.float32)\nfor i, word in enumerate(word2vec.words):\n    embedding[i] = word2vec[word]\n\n# setup a TensorFlow session\ntf.reset_default_graph()\nsess = tf.InteractiveSession()\nX = tf.Variable([0.0], name='embedding')\nplace = tf.placeholder(tf.float32, shape=embedding.shape)\nset_x = tf.assign(X, place, validate_shape=False)\nsess.run(tf.global_variables_initializer())\nsess.run(set_x, feed_dict={place: embedding})\n\n# write labels\nwith open('log\/metadata.tsv', 'w') as f:\n    for word in word2vec.words:\n        f.write(word + '\\n')\n\n# create a TensorFlow summary writer\nsummary_writer = tf.summary.FileWriter('log', sess.graph)\nconfig = projector.ProjectorConfig()\nembedding_conf = config.embeddings.add()\nembedding_conf.tensor_name = 'embedding:0'\nembedding_conf.metadata_path = os.path.join('log', 'metadata.tsv')\nprojector.visualize_embeddings(summary_writer, config)\n\n# save the model\nsaver = tf.train.Saver()\nsaver.save(sess, os.path.join('log', \"model.ckpt\"))\n<\/code><\/pre>\n\n<p>Then run this command in your terminal:<\/p>\n\n<pre><code>tensorboard --logdir=log\n<\/code><\/pre>\n"},{"AnswerId":"47281641","CreationDate":"2017-11-14T09:10:03.670","ParentId":null,"OwnerUserId":"5223524","Title":null,"Body":"<p>The accepted answer was very helpful to understand the general sequence:<\/p>\n\n<ol>\n<li>Create metadata for each vector (sample)<\/li>\n<li>Associate images (sprites) with each vector <\/li>\n<li>Load the data into TensorFlow and save the embeddings using checkpoint and summary writer (mind that the paths are consistent throughout the process).<\/li>\n<\/ol>\n\n<p>For me, the MNIST-based example still relied too much on pre-trained data and pre-generated sprite &amp; metadata files. To fill this gap I created such an example myself and decided to share it here for anyone interested - the code is on <a href=\"https:\/\/github.com\/altermarkive\/tensorflow-embeddings-minimalistic-example\" rel=\"nofollow noreferrer\">GitHub<\/a>.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/i77UL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/i77UL.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"50961175","CreationDate":"2018-06-21T05:47:37.567","ParentId":null,"OwnerUserId":"7728702","Title":null,"Body":"<p>Here's a link to an official guide. <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r1.1\/get_started\/embedding_viz\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r1.1\/get_started\/embedding_viz<\/a><\/p>\n\n<p>It says it was last updated June 2017.<\/p>\n"},{"AnswerId":"42579467","CreationDate":"2017-03-03T12:55:54.373","ParentId":null,"OwnerUserId":"1969168","Title":null,"Body":"<p>Check out this talk \"Hands-on TensorBoard (TensorFlow Dev Summit 2017)\" <a href=\"https:\/\/www.youtube.com\/watch?v=eBbEDRsCmv4\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=eBbEDRsCmv4<\/a>  It demonstrates TensorBoard embedding on the MNIST dataset.<\/p>\n\n<p>Sample code and slides for the talk can be found here <a href=\"https:\/\/github.com\/mamcgrath\/TensorBoard-TF-Dev-Summit-Tutorial\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mamcgrath\/TensorBoard-TF-Dev-Summit-Tutorial<\/a><\/p>\n"}]}
{"QuestionId":41259213,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2016-12-21T09:26:27.130","AcceptedAnswerId":null,"OwnerUserId":7270518.0,"Title":"How to save and restore partitioned variable in Tensorflow","Body":"<p>I have a large matrix.<\/p>\n\n<p>I use bellow way create this variable as number of shards.<\/p>\n\n<pre><\/pre>\n\n<p>create log:<\/p>\n\n<pre><\/pre>\n\n<p>I can training and save it success. But when I try to restore model I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>I found tensorflow save the variable as a part. The saved parameter just have one . No longer a partitioned variable.<\/p>\n","answers":[{"AnswerId":"46649895","CreationDate":"2017-10-09T15:22:59.760","ParentId":null,"OwnerUserId":"712995","Title":null,"Body":"<p>It happened in tensorflow 0.12 and doesn't happen in 1.3 (the last version as of October 2017). Here a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6551\" rel=\"nofollow noreferrer\">GitHub issue<\/a>, filed by the same author and now fixed. So if you see this error, just upgrade tensorflow.<\/p>\n"}]}
{"QuestionId":41259555,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><loss><cross-entropy>","CreationDate":"2016-12-21T09:41:06.233","AcceptedAnswerId":41274085.0,"OwnerUserId":5242249.0,"Title":"Tensorflow - loss starts high and does not decrease","Body":"<p>i started writing Neuronal Networks with tensorflow and there is one Problem i seem to face in each of my example Projects.<\/p>\n\n<p>My loss allways starts at something like 50 or higher and does not decrease or if it does, it does so slowly that after all my epochs i do not even get near an acceptable loss-rate.<\/p>\n\n<p><strong>Things it already tried<\/strong> (and did not affect the result very much)<\/p>\n\n<ul>\n<li>tested on overfitting, but in the following example\nyou can see that i have 15000 training and 15000 testing-datasets and\nsomething like 900 neurons<\/li>\n<li>tested different optimizers and optimizer-values<\/li>\n<li>tried increasing the traingdata by using the testdata as\ntrainingdata aswell<\/li>\n<li>tried increasing and decreasing the batchsize<\/li>\n<\/ul>\n\n<p>I created the network on knowledge of <a href=\"https:\/\/youtu.be\/vq2nnJ4g6N0\" rel=\"noreferrer\">https:\/\/youtu.be\/vq2nnJ4g6N0<\/a><\/p>\n\n<p>But let us have a look on <strong>one of my testprojects<\/strong>: <\/p>\n\n<p>I have a list of names and wanted to assume the gender so my raw data looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>For feeding it into the network i transform the names into an array of charCodes (expecting a maxlength of 30) and the gender into a bit array<\/p>\n\n<pre><\/pre>\n\n<p>I built the network with 3 hidden layers [30,20],[20,10],[10,10] and [10,2] for the output layer. All hidden layers have a ReLU as activation function. The output layer has a softmax.<\/p>\n\n<pre><\/pre>\n\n<p>Now the calculation for the loss, accuracy and the training operation:<\/p>\n\n<pre><\/pre>\n\n<p>I train the network in batches of 100<\/p>\n\n<pre><\/pre>\n\n<p>And i get the following result:<\/p>\n\n<pre><\/pre>\n\n<p><br\/><\/p>\n\n<p>What am i doing wrong?<\/p>\n\n<p>Why does it start at ~69 in my Project and not lower?<\/p>\n\n<p><br\/><\/p>\n\n<p>Thank you very much guys!<\/p>\n","answers":[{"AnswerId":"41274085","CreationDate":"2016-12-22T00:28:39.353","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>There's nothing wrong with 0.69 <a href=\"https:\/\/en.wikipedia.org\/wiki\/Nat_(unit)\" rel=\"nofollow noreferrer\">nats<\/a> of entropy per samples, as a starting point for a binary classification.<\/p>\n\n<p>If you convert to base 2, <code>0.69\/log(2)<\/code>, you'll see that it's almost exactly 1 bit per sample which is exactly what you would expect if you're unsure about a binary classification.<\/p>\n\n<p>I usually use the mean loss instead of the sum so things are less sensitive to batch size.<\/p>\n\n<p>You should also not calculate the entropy directly yourself, because that method breaks easily. you probably want <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#sigmoid_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.sigmoid_cross_entropy_with_logits<\/code><\/a>.<\/p>\n\n<p>I also like starting with the Adam Optimizer instead of pure gradient descent.<\/p>\n\n<p>Here are two reasons you might be having some trouble with this problem: <\/p>\n\n<p>1) Character codes are ordered, but the order doesn't mean anything. Your inputs would be easier for the network to take as input if they were input as one-hot vectors. So your input would be a 26x30 = 780 element vector. Without that the network has to waste a bunch of capacity learning the boundaries between letters.<\/p>\n\n<p>2) You've only got fully connected layers. This makes it impossible for it to learn a fact independent of it's absolute position in the name. 6 of the top 10 girls names in 2015 ended in 'a', while 0 of the top 10 boys names did. As currently written your network needs to re-learn \"Usually it's a girl's name if it ends in 'a'\" independently for each name length. Using some convolution layers would allow it to learn facts once across all name lengths.<\/p>\n"}]}
{"QuestionId":41260042,"AnswerCount":3,"Tags":"<python><neural-network><deep-learning><keras>","CreationDate":"2016-12-21T10:03:13.040","AcceptedAnswerId":42647196.0,"OwnerUserId":3994824.0,"Title":"Global Weight Decay in Keras","Body":"<p>Is there a way to set a global weight decay in Keras?<\/p>\n\n<p>I know about the layer wise one using regularizers(<a href=\"https:\/\/keras.io\/regularizers\/\" rel=\"noreferrer\">https:\/\/keras.io\/regularizers\/<\/a>), but I could not find any information about a way to set a global weight decay.<\/p>\n","answers":[{"AnswerId":"57550913","CreationDate":"2019-08-19T05:27:02.643","ParentId":null,"OwnerUserId":"11935789","Title":null,"Body":"<p>Posting the full code to apply weight decay on a Keras model (inspierd by the above post):<\/p>\n\n<pre><code># a utility function to add weight decay after the model is defined.\ndef add_weight_decay(model, weight_decay):\n    if (weight_decay is None) or (weight_decay == 0.0):\n        return\n\n    # recursion inside the model\n    def add_decay_loss(m, factor):\n        if isinstance(m, tf.keras.Model):\n            for layer in m.layers:\n                add_decay_loss(layer, factor)\n        else:\n            for param in m.trainable_weights:\n                with tf.keras.backend.name_scope('weight_regularizer'):\n                    regularizer = lambda: tf.keras.regularizers.l2(factor)(param)\n                    m.add_loss(regularizer)\n\n    # weight decay and l2 regularization differs by a factor of 2\n    add_decay_loss(model, weight_decay\/2.0)\n    return\n<\/code><\/pre>\n"},{"AnswerId":"42647196","CreationDate":"2017-03-07T11:30:35.020","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>According to the github repo (<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2717\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/2717<\/a>) there is no way to do global weight decay. I answered it here, so others who have the same problem do not have to look furhter for an answer.<\/p>\n\n<p>To get global weight decay in keras regularizers have to be added to every layer in the model. In my models these layers are batch normalization (beta\/gamma regularizer) and dense\/convolutions (W_regularizer\/b_regularizer) layers.<\/p>\n\n<p>Layer wise regularization is described here: (<a href=\"https:\/\/keras.io\/regularizers\/\" rel=\"noreferrer\">https:\/\/keras.io\/regularizers\/<\/a>).<\/p>\n"},{"AnswerId":"54564848","CreationDate":"2019-02-07T00:56:47.127","ParentId":null,"OwnerUserId":"823874","Title":null,"Body":"<p>There is no way to directly apply a \"global\" weight decay to a whole keras model at once.<\/p>\n\n<p>However, as I describe <a href=\"https:\/\/jricheimer.github.io\/keras\/2019\/02\/06\/keras-hack-1\/\" rel=\"noreferrer\">here<\/a>, you can employ weight decay on a model by looping through its layers and manually applying the regularizers on appropriate layers. Here's the relevant code snippet:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = keras.applications.ResNet50(include_top=True, weights='imagenet')\nalpha = 0.00002  # weight decay coefficient\n\nfor layer in model.layers:\n    if isinstance(layer, keras.layers.Conv2D) or isinstance(layer, keras.layers.Dense):\n        layer.add_loss(keras.regularizers.l2(alpha)(layer.kernel))\n    if hasattr(layer, 'bias_regularizer') and layer.use_bias:\n        layer.add_loss(keras.regularizers.l2(alpha)(layer.bias))\n<\/code><\/pre>\n"}]}
{"QuestionId":41261416,"AnswerCount":2,"Tags":"<machine-learning><neural-network><deep-learning><caffe><lmdb>","CreationDate":"2016-12-21T11:09:58.850","AcceptedAnswerId":null,"OwnerUserId":7325313.0,"Title":"Single Float label , LMDB Format in caffe","Body":"<p>I am using caffe for a regression problem and I want to know how is possible to use <a href=\"\/questions\/tagged\/lmdb\" class=\"post-tag\" title=\"show questions tagged &#39;lmdb&#39;\" rel=\"tag\">lmdb<\/a> format for single float label. <\/p>\n\n<p>Right now, caffe only supports  type label for lmdb data (type of  in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L30-L41\" rel=\"nofollow noreferrer\"><\/a> is ).<br>\nIn order to change this default behavior, I changed some of files in caffe as follows but the problem has stayed and after converting my labels all are zero. <\/p>\n\n<pre><\/pre>\n\n<p>After all I compiled caffe again but it doesn't work. <\/p>\n\n<p>Is there anyone to solve this problem. it is very important to me to solve it as soon as possible. <\/p>\n\n<p>Thanks in advance. <\/p>\n","answers":[{"AnswerId":"41262271","CreationDate":"2016-12-21T11:55:56.833","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you insist to force caffe to support float labels, you might need to change <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/tools\/convert_imageset.cpp#L81\" rel=\"nofollow noreferrer\"><code>convert_imageset.cpp<\/code> line 81<\/a> as well.<br>\nCurrently this line uses <a href=\"http:\/\/www.cplusplus.com\/reference\/cstdlib\/atoi\/\" rel=\"nofollow noreferrer\"><code>atoi<\/code><\/a> to convert string to <strong>int<\/strong>:<\/p>\n\n<pre><code>label = atoi(line.substr(pos + 1).c_str());\n<\/code><\/pre>\n\n<p>You should convert it to user <a href=\"http:\/\/www.cplusplus.com\/reference\/cstdlib\/atof\/\" rel=\"nofollow noreferrer\"><code>atof<\/code><\/a> to convert string to <strong>float<\/strong>:<\/p>\n\n<pre><code>label = atof(line.substr(pos + 1).c_str());\n<\/code><\/pre>\n"},{"AnswerId":"41262150","CreationDate":"2016-12-21T11:49:26.323","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Solving your problem by forcing and tweaking caffe to work with float labels in LMDB dataset seems like not a very good strategy.  <\/p>\n\n<p>I suggest instead to use caffe's <code>\"HDF5Data\"<\/code> layer. Using hdf5 format to store data for caffe is much more flexible and allows you to have float labels.  <\/p>\n\n<p>For more information see <a href=\"https:\/\/stackoverflow.com\/a\/31808324\/1714410\">this answer<\/a>.<\/p>\n"}]}
{"QuestionId":41262284,"AnswerCount":2,"Tags":"<python><docker><theano><nvidia><jupyter-notebook>","CreationDate":"2016-12-21T11:56:35.390","AcceptedAnswerId":41492813.0,"OwnerUserId":470184.0,"Title":"Nvidia Theano docker image not available","Body":"<p>Trying to run docker command : <\/p>\n\n<pre><\/pre>\n\n<p>taken from <a href=\"https:\/\/github.com\/nouiz\/Theano-Docker\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nouiz\/Theano-Docker<\/a><\/p>\n\n<p>returns error : <\/p>\n\n<pre><\/pre>\n\n<p>Appears the theano_secure image is not currently available ? <\/p>\n\n<p>Searching for theano_secure : <\/p>\n\n<pre><\/pre>\n\n<p>The return of this command is empty so image is not available ? \nIf so is there an alternative Theano docker image from nvidia ?<\/p>\n\n<p>Update : <\/p>\n\n<p>building from source :<\/p>\n\n<pre><\/pre>\n\n<p>returns :<\/p>\n\n<pre><\/pre>\n\n<p>and : <\/p>\n\n<pre><\/pre>\n\n<p>Manually checking URL's : <a href=\"http:\/\/developer.download.nvidia.com\" rel=\"nofollow noreferrer\">http:\/\/developer.download.nvidia.com<\/a> &amp; <a href=\"http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/InRelease\" rel=\"nofollow noreferrer\">http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/InRelease<\/a> are both not available. Should I build with alternative docker file ?<\/p>\n\n<p>Update 2 : <\/p>\n\n<p>I think this error is occurring as  <a href=\"http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/InRelease\" rel=\"nofollow noreferrer\">http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/InRelease<\/a> does not exist. However <a href=\"http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/Release\" rel=\"nofollow noreferrer\">http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/Release<\/a> does exist. <\/p>\n\n<p>Can docker be modified to use <a href=\"http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/Release\" rel=\"nofollow noreferrer\">http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/Release<\/a> instead of <a href=\"http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/InRelease\" rel=\"nofollow noreferrer\">http:\/\/archive.ubuntu.com\/ubuntu\/dists\/trusty\/InRelease<\/a> ?<\/p>\n\n<p>OS version : <\/p>\n\n<pre><\/pre>\n\n<p>Update 3 : <\/p>\n\n<p>\"you are supposed to docker build first\", before nvidia-docker run\" I did try <\/p>\n\n<pre><\/pre>\n\n<p>which returns :<\/p>\n\n<pre><\/pre>\n\n<p>I can pull image  but this does not run via Jupyter notebook in same way as  documented at <a href=\"https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/\" rel=\"nofollow noreferrer\">https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/<\/a> . There does not appear to be a docker Jupyter Theano container available.<\/p>\n\n<p>How to expose the docker instance kaixhin\/theano via Jupyter notebook ? <\/p>\n\n<p>I tried :  but receive error : <\/p>\n\n<pre><\/pre>\n\n<p>Modification of  docker container in order expose it via Jupyter notebook ?<\/p>\n","answers":[{"AnswerId":"41401898","CreationDate":"2016-12-30T19:46:54.167","ParentId":null,"OwnerUserId":"6309","Title":null,"Body":"<p>From what I understand of the <a href=\"https:\/\/github.com\/nouiz\/Theano-Docker\" rel=\"nofollow noreferrer\"><code>nouiz\/Theano-Docker README<\/code><\/a>, you are supposed to <code>docker build<\/code> first, before <code>nvidia-docker run<\/code>.<\/p>\n\n<p>But since the build is tricky, I would try instead <a href=\"https:\/\/hub.docker.com\/r\/kaixhin\/theano\/\" rel=\"nofollow noreferrer\"><code>docker pull kaixhin\/theano<\/code><\/a>  (from <a href=\"https:\/\/hub.docker.com\/r\/kaixhin\/cuda-theano\/\" rel=\"nofollow noreferrer\"><code>kaixhin\/cuda-theano\/<\/code><\/a>), much more recent (3 days ago), which is based on <a href=\"https:\/\/github.com\/Kaixhin\/dockerfiles\/blob\/master\/theano\/Dockerfile\" rel=\"nofollow noreferrer\"><code>theano Dockerfile<\/code><\/a>. <\/p>\n\n<p>That image does rely on <a href=\"http:\/\/www.nvidia.com\/object\/cuda_home_new.html\" rel=\"nofollow noreferrer\">CUDA<\/a>and needs to be run on an Ubuntu host OS with NVIDIA Docker installed. The driver requirements can be found on the <a href=\"https:\/\/github.com\/NVIDIA\/nvidia-docker\/wiki\/CUDA#requirements\" rel=\"nofollow noreferrer\">NVIDIA Docker wiki<\/a>.<\/p>\n"},{"AnswerId":"41492813","CreationDate":"2017-01-05T19:09:09.670","ParentId":null,"OwnerUserId":"3378901","Title":null,"Body":"<pre><code>Error: image library\/theano_secure:latest not found\n<\/code><\/pre>\n\n<p>Because <code>theano_secure<\/code> doesn't like <code>ubuntu<\/code>,<code>centos<\/code>, it is not official repository on docker hub, so you need to build it by yourself.<\/p>\n\n<pre><code>Err http:\/\/developer.download.nvidia.com  Release.gpg Unable to connect to developer.download.nvidia.com:http: [IP: 184.24.98.231 80]\n<\/code><\/pre>\n\n<p>Please check your internet connection first, <code>telnet 184.24.98.231 80<\/code>.<\/p>\n\n<p>Maybe you are in a limited network place, try behind a proxy to do this again. You may want take a look about how to <a href=\"https:\/\/stackoverflow.com\/a\/41490826\/3378901\">build image behind a proxy<\/a>.<\/p>\n"}]}
{"QuestionId":41264435,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2016-12-21T13:54:21.347","AcceptedAnswerId":null,"OwnerUserId":7310903.0,"Title":"How to obtain gcloud predictions by passing a base64 image to a retrained inception model?","Body":"<p>I am trying to obtain a prediction with gcloud by passing a base64 encoded image to a retrained inception model, by using a similar approach as the one adopted by <a href=\"https:\/\/stackoverflow.com\/users\/7219743\/davide-biraghi\">Davide Biraghi<\/a> in <a href=\"https:\/\/stackoverflow.com\/questions\/41261701\/how-make-correct-predictions-of-jpeg-image-in-cloud-ml\">this<\/a> post.\nWhen using  as input I also get the same error when trying to get predictions, therefore I adopted a slightly different approach.<\/p>\n\n<p>Following <a href=\"https:\/\/stackoverflow.com\/users\/1399222\/rhaertel80\">rhaertel80<\/a>'s suggestions in his answer to this <a href=\"https:\/\/stackoverflow.com\/questions\/40849195\/how-convert-a-jpeg-image-into-json-file-in-google-machine-learning\">post<\/a>, I have created a graph that takes a jpeg image as input in , preprocess it and feeds it to the inception model in .<\/p>\n\n<p>The prediction returns the values, although the wrong ones (I am trying to find a solution in another post) but at least it doesn't fail. The placeholder I use as input is   <\/p>\n\n<pre><\/pre>\n\n<p>And I add it to the model inputs with<\/p>\n\n<pre><\/pre>\n\n<p>As Davide I am following the suggestions found in these posts: <a href=\"https:\/\/stackoverflow.com\/questions\/40849195\/how-convert-a-jpeg-image-into-json-file-in-google-machine-learning\">here<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/39978526\/deploy-retrained-inception-model-on-google-cloud-machine-learning\">here<\/a> and <a href=\"https:\/\/cloud.google.com\/blog\/big-data\/2016\/12\/how-to-train-and-classify-images-using-google-cloud-machine-learning-and-cloud-dataflow\" rel=\"nofollow noreferrer\">here<\/a> and I am trying to get predictions with<\/p>\n\n<pre><\/pre>\n\n<p>where the file  has been obtained with this code<\/p>\n\n<pre><\/pre>\n\n<p>I would like to know why the prediction fails when I use as input  and it doesn't when I use this different approach, since they look almost identical to me: I use the same script to generate the instances (changing the input_key) and the same command line to request predictions<\/p>\n\n<p>Is there a way to pass the instance fed to  to  in order to get the right predictions?<\/p>\n","answers":[{"AnswerId":"41268108","CreationDate":"2016-12-21T17:01:21.290","ParentId":null,"OwnerUserId":"7310903","Title":null,"Body":"<p>Here I describe more in detail my approach and how I use the images_placeholder.<\/p>\n\n<p>I define a function that resizes the image:<\/p>\n\n<pre><code>  def decode_and_resize(image_str_tensor):\n    \"\"\"Decodes jpeg string, resizes it and returns a uint8 tensor.\"\"\"\n\n    image = tf.image.decode_jpeg(image_str_tensor, channels=MODEL_INPUT_DEPTH)\n\n    # Note resize expects a batch_size, but tf_map supresses that index,\n    # thus we have to expand then squeeze.  Resize returns float32 in the\n    # range [0, uint8_max]\n    image = tf.expand_dims(image, 0)\n    image = tf.image.resize_bilinear(\n        image, [MODEL_INPUT_HEIGHT, MODEL_INPUT_WIDTH], align_corners=False)\n    image = tf.squeeze(image, squeeze_dims=[0])\n    image = tf.cast(image, dtype=tf.uint8)\n    return image\n<\/code><\/pre>\n\n<p>and one that generates the definition of the graph in which the resize takes place and where <code>images_placeholder<\/code> is defined and used<\/p>\n\n<pre><code>def create_b64_graph() :\n  with tf.Graph().as_default() as b64_graph:\n\n    images_placeholder = tf.placeholder(dtype=tf.string, shape=(None,),\n                                     name='B64Connector\/input')\n    decoded_images = tf.map_fn(\n        decode_and_resize, images_placeholder, back_prop=False, dtype=tf.uint8)\n\n    # convert_image_dtype, also scales [0, uint8_max] -&gt; [0, 1).\n    images = tf.image.convert_image_dtype(decoded_images, dtype=tf.float32)\n\n    # Finally, rescale to [-1,1] instead of [0, 1)\n    images = tf.sub(images, 0.5)\n    images = tf.mul(images, 2.0)\n\n    # NOTE: using identity to get a known name for the output tensor.\n    output = tf.identity(images, name='B64Connector\/output')\n\n    b64_graph_def = b64_graph.as_graph_def()\n\n    return b64_graph_def\n<\/code><\/pre>\n\n<p>Moreover I am using the following code to merge the resizing graph with the inception graph. Can I use a similar approach to link <code>images_placeholder<\/code> directly to <code>'DecodeJpeg\/contents:0'<\/code>?<\/p>\n\n<pre><code>def concatenate_to_inception_graph( b64_graph_def ):\n\n  model_dir = INPUT_MODEL_PATH\n  model_filename = os.path.join(\n       model_dir, 'classify_image_graph_def.pb')\n\n  with tf.Session() as sess:\n\n    # Import the b64_graph and get its output tensor\n    resized_b64_tensor, = (tf.import_graph_def(b64_graph_def, name='',\n                             return_elements=['B64Connector\/output:0']))\n\n    with gfile.FastGFile(model_filename, 'rb') as f:\n      inception_graph_def = tf.GraphDef()\n      inception_graph_def.ParseFromString(f.read())\n\n      # Concatenate b64_graph and inception_graph\n      g_1 = tf.import_graph_def(inception_graph_def, name='inception',\n               input_map={'ResizeBilinear:0' : resized_b64_tensor} )\n\n    return sess.graph\n<\/code><\/pre>\n"}]}
{"QuestionId":41265035,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2016-12-21T14:23:58.027","AcceptedAnswerId":41273348.0,"OwnerUserId":5046896.0,"Title":"TensorFlow, why there are 3 files after saving the model?","Body":"<p>Having read the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variables\/#saving_and_restoring\" rel=\"noreferrer\">docs<\/a>, I saved a model in , here is my demo code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>but after that, I found there are 3 files<\/p>\n\n<pre><\/pre>\n\n<p>And I can't restore the model by restore the  file, since there is no such file. Here is my code<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>So, why there are 3 files?<\/p>\n","answers":[{"AnswerId":"48435810","CreationDate":"2018-01-25T04:41:38.653","ParentId":null,"OwnerUserId":"4534873","Title":null,"Body":"<p>If you trained a CNN with dropout, for example, you could do this:<\/p>\n\n<pre><code>def predict(image, model_name):\n    \"\"\"\n    image -&gt; single image, (width, height, channels)\n    model_name -&gt; model file that was saved without any extensions\n    \"\"\"\n    with tf.Session() as sess:\n        saver = tf.train.import_meta_graph('.\/' + model_name + '.meta')\n        saver.restore(sess, '.\/' + model_name)\n        # Substitute 'logits' with your model\n        prediction = tf.argmax(logits, 1)\n        # 'x' is what you defined it to be. In my case it is a batch of RGB images, that's why I add the extra dimension\n        return prediction.eval(feed_dict={x: image[np.newaxis,:,:,:], keep_prob_dnn: 1.0})\n<\/code><\/pre>\n"},{"AnswerId":"46845865","CreationDate":"2017-10-20T09:10:10.103","ParentId":null,"OwnerUserId":"5757263","Title":null,"Body":"<p>I am restoring trained word embeddings from <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/embedding\/word2vec.py\" rel=\"nofollow noreferrer\" title=\"Word2Vec\">Word2Vec<\/a> tensorflow tutorial.<\/p>\n\n<p>In case you have created multiple checkpoints:<\/p>\n\n<p>e.g. files created look like this<\/p>\n\n<blockquote>\n  <p>model.ckpt-55695.data-00000-of-00001<\/p>\n  \n  <p>model.ckpt-55695.index<\/p>\n  \n  <p>model.ckpt-55695.meta<\/p>\n<\/blockquote>\n\n<p>try this<\/p>\n\n<pre><code>def restore_session(self, session):\n   saver = tf.train.import_meta_graph('.\/tmp\/model.ckpt-55695.meta')\n   saver.restore(session, '.\/tmp\/model.ckpt-55695')\n<\/code><\/pre>\n\n<p>when calling restore_session():<\/p>\n\n<pre><code>def test_word2vec():\n   opts = Options()    \n   with tf.Graph().as_default(), tf.Session() as session:\n       with tf.device(\"\/cpu:0\"):            \n           model = Word2Vec(opts, session)\n           model.restore_session(session)\n           model.get_embedding(\"assistance\")\n<\/code><\/pre>\n"},{"AnswerId":"41273348","CreationDate":"2016-12-21T22:58:55.797","ParentId":null,"OwnerUserId":"3821732","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    saver = tf.train.import_meta_graph('\/tmp\/model.ckpt.meta')\n    saver.restore(sess, \"\/tmp\/model.ckpt\")\n<\/code><\/pre>\n\n<p>The TensorFlow save method saves three kinds of files because it stores the <b>graph structure<\/b> separately from the <b>variable values<\/b>. The <code>.meta<\/code> file describes the saved graph structure, so you need to import it before restoring the checkpoint (otherwise it doesn't know what variables the saved checkpoint values correspond to).<\/p>\n\n<p>Alternatively, you could do this:<\/p>\n\n<pre><code># Recreate the EXACT SAME variables\nv1 = tf.Variable(..., name=\"v1\")\nv2 = tf.Variable(..., name=\"v2\")\n\n...\n\n# Now load the checkpoint variable values\nwith tf.Session() as sess:\n    saver = tf.train.Saver()\n    saver.restore(sess, \"\/tmp\/model.ckpt\")\n<\/code><\/pre>\n\n<p>Even though there is no file named <code>model.ckpt<\/code>, you still refer to the saved checkpoint by that name when restoring it. From the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/saver.py\" rel=\"noreferrer\"><code>saver.py<\/code> source code<\/a>: <\/p>\n\n<blockquote>\n  <p>Users only need to interact with the user-specified prefix... instead\n  of any physical pathname.<\/p>\n<\/blockquote>\n"},{"AnswerId":"45033500","CreationDate":"2017-07-11T11:38:45.250","ParentId":null,"OwnerUserId":"8268280","Title":null,"Body":"<ul>\n<li><p><strong>meta file<\/strong>: describes the saved graph structure, includes GraphDef, SaverDef, and so on; then apply <code>tf.train.import_meta_graph('\/tmp\/model.ckpt.meta')<\/code>, will restore <code>Saver<\/code> and <code>Graph<\/code>.<\/p><\/li>\n<li><p><strong>index file<\/strong>: it is a string-string immutable table(tensorflow::table::Table).  Each key is a name of a tensor and its value is a serialized BundleEntryProto.  Each BundleEntryProto describes the metadata of a tensor: which of the \"data\" files contains the content of a tensor, the offset into that file, checksum, some auxiliary data, etc. <\/p><\/li>\n<li><p><strong>data file<\/strong>: it is TensorBundle collection, save the values of all variables. <\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":41265558,"AnswerCount":0,"Tags":"<torch>","CreationDate":"2016-12-21T14:49:42.350","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"THNN.lua:110: input and gradOutput shapes do not match: input [90 x 271], gradOutput [271]","Body":"<p>I have a torch code that can run ok on one machine  but with the error \"THNN.lua:110: input and gradOutput shapes do not match: input [90 x 271], gradOutput [271]\" on another machine. The data is same. I think is the version problem. But i don't know how to fix it. Any help?<\/p>\n\n<p>The problem refers to backward of  logsoftmax module. But when I print the shape, there is no problem.   <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41265720,"AnswerCount":0,"Tags":"<python><autocomplete><tensorflow><pycharm>","CreationDate":"2016-12-21T14:58:27.447","AcceptedAnswerId":null,"OwnerUserId":697053.0,"Title":"PyCharm subclass navigation \"cannot find declaration to go to\"","Body":"<p>When using the TensorFlow library in PyCharm Mac, I can't go to some class implementations. For example,<\/p>\n\n<pre><\/pre>\n\n<p>When I try to go to the implementation of DNNClassifier it can't find it. But when I do<\/p>\n\n<pre><\/pre>\n\n<p>It works fine. I am not sure why PyCharm doesn't see the DNNClassifer they way it was written in the first example above.<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":41267321,"AnswerCount":1,"Tags":"<image><tensorflow><tflearn>","CreationDate":"2016-12-21T16:17:18.590","AcceptedAnswerId":null,"OwnerUserId":2221270.0,"Title":"TFLearn: How to use image preloader for image in -> image out regression problems","Body":"<p>I'm new to tflearn and I'm trying to train a CNN to solve image in -> image out regression problems, where the input images are either MxN grayscale or MxNx3 RGB.<\/p>\n\n<p>I have a very large training set that will not fit in memory, so I would like to train \"on the fly\", pulling images from disk as training is ready to consume them.<\/p>\n\n<p>The image_preloader seems to make the assumption that we will always be solving problems where X is an input image and y is a set of scalar labels corresponding to X. <\/p>\n\n<p>What is the best way within the tflearn framework to do something similar for regression problems where both X and y are images? I feel like I have to be missing something obvious, but I've looked through the Data Utils, Data Augmentation, and Data Preprocessing sections and don't see anything.<\/p>\n\n<p>For the full scope of what I'm trying to do: I have a large set of images on disk. On the fly, I need to be able to load my images, then do some modifications to the image to define the input images X and y. The images on disk need to be augmented to properly define X and y. Given a minibatch X read from disk, I can define X and y a batch at a time.<\/p>\n","answers":[{"AnswerId":"41272472","CreationDate":"2016-12-21T21:50:10.887","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p>In pure TensorFlow you could use something like:<\/p>\n\n<pre><code>def load_image(type=\"feature\")\n  files = '\/path\/to\/my\/images\/%s\/*.jpg' % type\n  filename_queue = tf.train.string_input_producer(tf.gfile.Glob(files))\n  reader = tf.WholeFileReader()\n  filename, data = reader.read()\n  image_tensor = tf.image.decode_jpeg(data, channels=3)\n  return image_tensor\n<\/code><\/pre>\n\n<p>Note the <code>channels=3<\/code> in <code>decode_jpeg<\/code> will ensure your images are all tensors of <code>[h, w, 3]<\/code>, even if they were originally in a different colourspace.<\/p>\n\n<p>At this point you have single image tensors. You can batch them up using background queues like so:<\/p>\n\n<pre><code>features = load_image(\"feature\")\nlabels = load_image(\"labels\")\nbatch_size = 64\ninput_batch, label_batch = tf.train.shuffle_batch([features, labels],\n  batch_size=batch_size, num_threads=4, capacity=1000 + 3 * batch_size,\n  min_after_dequeue=1000)\n<\/code><\/pre>\n\n<p>You'll need to start the queue runners too. This (and the magic numbers in the last code snippet) is explained fairly well in <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/#batching\" rel=\"nofollow noreferrer\">the documentation on batching<\/a>.<\/p>\n"}]}
{"QuestionId":41267650,"AnswerCount":3,"Tags":"<caffe>","CreationDate":"2016-12-21T16:35:57.630","AcceptedAnswerId":42300885.0,"OwnerUserId":2736559.0,"Title":"How are multiple gpus utilized in Caffe?","Body":"<p>I want to know how  utilizes multiple s so that I can decide to upgrade to a new more powerful card or just buy the same card and run on .<br>\nfor example am I better off buying one  , or two  ?<br>\nIf I go SLI the 1080s, will my effective memory get doubled? I mean can I run a network which takes 12 or more GB of vram using them? Or am I left with only 8 GB ?\nAgain how is memory utilized in such scenarios ? \nWhat would happen if two different cards are installed (both NVIDIA) ? Does caffe utilize the memory available the same? (suppose one 980 and one 970!)<\/p>\n","answers":[{"AnswerId":"42300885","CreationDate":"2017-02-17T14:53:33.517","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<blockquote>\n  <p>for example am I better off buying one TitanX 12 GB , or two GTX 1080\n  8 GB ? If I go SLI the 1080s, will my effective memory get doubled? I\n  mean can I run a network which takes 12 or more GB of vram using them?\n  Or am I left with only 8 GB ?<\/p>\n<\/blockquote>\n\n<p>No, <code>effective memory<\/code> size in case of 2 GPU with 8Gb of RAM will be 8Gb, but <code>effective batch size<\/code> will be doubled which will lead to more stable\\fast training.<\/p>\n\n<blockquote>\n  <p>What would happen if two different cards are installed (both NVIDIA) ?\n  Does caffe utilize the memory available the same? (suppose one 980 and\n  one 970!)<\/p>\n<\/blockquote>\n\n<p>I think you will be limited to lower card and may have problem with drivers, so I don't recomend to try this configuration.\nAlso from documentation:<\/p>\n\n<blockquote>\n  <p>Current implementation has a \"soft\" assumption that the devices being\n  used are homogeneous. In practice, any devices of the same general\n  class should work together, but performance and total size is limited\n  by the smallest device being used. e.g. if you combine a TitanX and a\n  GTX980, performance will be limited by the 980. Mixing vastly\n  different levels of boards, e.g. Kepler and Fermi, is not supported.<\/p>\n<\/blockquote>\n\n<p>Summing up: with GPU that have lots of RAM you can train deeper models, with multiply GPUs you can train single model faster and also you can train separate models per GPU. I would choose single GPU with more memory (TitanX) because deep networks nowadays are more RAM bounded(e.g. ResNet-152 or some semantic segmentation network) and more memory will give the opportunity to run deeper networks and with larger batch size, otherwise if you have some task that fit on single GPU (GTX 1080) you can buy 2 or 4 of them just to make things faster.<\/p>\n\n<p>Also here is some info about multi GPU support in Caffe:<\/p>\n\n<blockquote>\n  <p>The current implementation uses a tree reduction strategy. e.g. if\n  there are 4 GPUs in the system, 0:1, 2:3 will exchange gradients, then\n  0:2 (top of the tree) will exchange gradients, 0 will calculate\n  updated model, 0->2, and then 0->1, 2->3.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/docs\/multigpu.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/docs\/multigpu.md<\/a><\/p>\n"},{"AnswerId":"41289087","CreationDate":"2016-12-22T17:49:25.967","ParentId":null,"OwnerUserId":"2736559","Title":null,"Body":"<blockquote>\n  <p>I don't believe Caffe supports SLI mode.  The two GPUs are treated as\n  separate cards.<\/p>\n  \n  <p>When you run Caffe and add the '-gpu' flag (assuming you are using the\n  command line), you can specify which GPU to use (-gpu 0 or -gpu 1 for\n  example).  You can also specify multiple GPUs (-gpu 0,1,3) including\n  using all GPUs (-gpu all). <\/p>\n  \n  <p>When you execute using multiple GPUs, Caffe will execute the training\n  across all of the GPUs and then merge the training updates across the\n  models.  This is effectively doubling (or more if you have more than 2\n  GPUs) the batch size for each iteration.<\/p>\n  \n  <p>In my case, I started with a NVIDIA GTX 970 (4GB card) and then\n  upgraded to a NVIDIA GTX Titan X (Maxwell version with 12 GB) because\n  my models were too large to fit in the GTX 970.  I can run some of the\n  smaller models across both cards (even though they are not the same)\n  as long as the model will fully fit into the 4GB of the smaller card. \n  Using the standard ImageNet model, I could execute across both cards\n  and cut my training time in half.<\/p>\n  \n  <p>If I recall correctly, other frameworks (TensorFlow and maybe the\n  Microsoft CNTK) support splitting a model among different nodes to\n  effectively increase the available GPU memory like what you are\n  describing.  Although I haven't personally tried either one, I\n  understand you can define on a per-layer basis where the layer\n  executes. <\/p>\n  \n  <p>Patrick<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/loJxGcJFo_4\" rel=\"nofollow noreferrer\">Link<\/a><\/p>\n"},{"AnswerId":"41767539","CreationDate":"2017-01-20T15:49:48.037","ParentId":null,"OwnerUserId":"6848625","Title":null,"Body":"<p>Perhaps a late answer, but caffe supports gpu parallelism, which means you can indeed fully utilize both gpu's, but I do recommend getting two gpu's of equal memory size, since I don't think caffe lets you select the batch size per gpu. <\/p>\n\n<p>As for how memory is utilized, using multiple gpu's each gpu gets a batch of batch size as specified in your train_val.prototxt, so if your batch size is for example 16 and you're using 2 gpu's, you'd have an effective batch size 32.<\/p>\n\n<p>Finally, I know that for things such as gaming, SLI seems to be much less effective and often much more problematic than having a single, more powerful GPU. So if you are planning on using the GPU's for more than only Deep Learning, I'd recommend you still go for the Titan X<\/p>\n"}]}
{"QuestionId":41267829,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><lstm>","CreationDate":"2016-12-21T16:45:50.647","AcceptedAnswerId":41273843.0,"OwnerUserId":2368148.0,"Title":"Retrieving last value of LSTM sequence in Tensorflow","Body":"<p>I have sequences of different lengths that I want to classify using LSTMs in Tensorflow. For the classification I just need the LSTM output of the last timestep of each sequence.<\/p>\n\n<pre><\/pre>\n\n<p>I would like to get, in NumPy notation: <\/p>\n\n<p>Is there any way or workaround to get this behaviour in Tensorflow?<\/p>\n\n<p>---UPDATE---<\/p>\n\n<p>Following this post <a href=\"https:\/\/stackoverflow.com\/questions\/36088277\/how-to-select-rows-from-a-3-d-tensor-in-tensorflow\">How to select rows from a 3-D Tensor in TensorFlow?<\/a> it seems that is possible to solve the problem in an efficient manner with  and manipulating the indices. The only requirement is that the batch size must be known in advance. Here is the adaptation of the referred post to this concrete problem:<\/p>\n\n<pre><\/pre>\n\n<p>@petrux option (<a href=\"https:\/\/stackoverflow.com\/questions\/41273361\/get-th-last-output-of-a-dynamic-rnn-in-tensorflow\/41273737#41273737\">Get the last output of a dynamic_rnn in TensorFlow<\/a>) seems also to work but the need of building a list within a for loop may be less optimized, although I did not perform any benchmark to support this statement.<\/p>\n","answers":[{"AnswerId":"43299261","CreationDate":"2017-04-08T19:08:45.050","ParentId":null,"OwnerUserId":"1953800","Title":null,"Body":"<p>Here's a solution, using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"nofollow noreferrer\">gather_nd<\/a>, where batch size does not need to be known ahead of time.<\/p>\n\n<pre><code>def extract_axis_1(data, ind):\n    \"\"\"\n    Get specified elements along the first axis of tensor.\n    :param data: Tensorflow tensor that will be subsetted.\n    :param ind: Indices to take (one for each element along axis 0 of data).\n    :return: Subsetted tensor.\n    \"\"\"\n\n    batch_range = tf.range(tf.shape(data)[0])\n    indices = tf.stack([batch_range, ind], axis=1)\n    res = tf.gather_nd(data, indices)\n\n    return res\n\noutput = extract_axis_1(sequence_outputs, lengths - 1)\n<\/code><\/pre>\n\n<p>Now <code>output<\/code> is a tensor of dimension <code>[batch_size, num_cells]<\/code>.<\/p>\n"},{"AnswerId":"41273843","CreationDate":"2016-12-21T23:58:12.303","ParentId":null,"OwnerUserId":"1861627","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/a\/41273737\/1861627\">This<\/a> could be an answer. I don't think there is anything similar to the NumPy notation you pointed out, but the effect is the same.<\/p>\n"}]}
{"QuestionId":41268447,"AnswerCount":2,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2016-12-21T17:20:41.510","AcceptedAnswerId":41304459.0,"OwnerUserId":2241766.0,"Title":"Dynamic tensor shape for tensorflow RNN","Body":"<p>I'm trying a very simple example for tensorflow RNN. \nIn that example, I use dynamic rnn. The code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Actually, the code is taken from this <a href=\"http:\/\/monik.in\/a-noobs-guide-to-implementing-rnn-lstm-using-tensorflow\/\" rel=\"nofollow noreferrer\">tutorial<\/a>.\nThe input to this RNN network is a sequence of binary numbers. Each number is put into an array. For example, a seuquence has format: <br>\n\nThe shape of the input is [None,10,1] which are batch size, sequence size and embedding size, respectively. Now because dynamic rnn can accept variable input shape, I change the code as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Basically, I want to use variable-length sequences (of course same length for all sequences in the same batch, but different between batches). However, it throws the error:<\/p>\n\n<pre><\/pre>\n\n<p>I understand that the second dimension is , which cannot be used in . However, I believe that there must be a way to overcome this because RNN accepts variable lenth inputs, in general.\nHow can I do it?<\/p>\n","answers":[{"AnswerId":"41272102","CreationDate":"2016-12-21T21:22:07.347","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p><strong>tl;dr<\/strong>: try using <code>tf.batch(..., dynamic_pad=True)<\/code> to batch your data.<\/p>\n\n<hr>\n\n<p>@<a href=\"https:\/\/stackoverflow.com\/users\/2008194\/chris-anderson\">chris_anderson<\/a>'s comment is correct. Ultimately your network needs a dense matrix of numbers to work with and there are a couple of strategies to convert variable length data into hyperrectangles:<\/p>\n\n<ol>\n<li>Pad all batches to a fixed size (e.g. assume a maximum length of say 500 items per input and every item in every batch is padded to 500). There is nothing dynamic about this strategy.<\/li>\n<li>Apply padding per-batch to the length of the longest item in the batch (dynamic padding).<\/li>\n<li>Bucket your input based on length and apply padding per-batch. This is the same as #2, but with less overall padding.<\/li>\n<\/ol>\n\n<p>There are <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.training\/splitting_sequence_inputs_into_minibatches_with_state_saving\" rel=\"nofollow noreferrer\">other strategies<\/a> that you could use too.<\/p>\n\n<p>To do this batching, you use:<\/p>\n\n<ol>\n<li><code>tf.train.batch<\/code> - by default it does no padding, you need to implement it yourself.<\/li>\n<li><code>tf.train.batch(..., dynamic_pad=True)<\/code><\/li>\n<li><code>tf.contrib.training.bucket_by_sequence_length<\/code><\/li>\n<\/ol>\n\n<p>I suspect you're also confused by the use of <code>tf.nn.dynamic_rnn<\/code>. It's important to note that the <code>dynamic<\/code> in <code>dynamic_rnn<\/code> refers to the way that TensorFlow unrolls the recurrent part of the network. in <code>tf.nn.rnn<\/code>, the recurrence is done statically in the graph (there is no internal loop, it's unrolled at graph construction time). In <code>dynamic_rnn<\/code> however, TensorFlow uses <code>tf.while_loop<\/code> to iterate inside the graph at run time. To use dynamic padding, you need to use dynamic unrolling, but it does not do it automatically.<\/p>\n"},{"AnswerId":"41304459","CreationDate":"2016-12-23T15:59:43.990","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p><code>tf.gather<\/code> expects a tensor, so you can use <code>tf.shape(val)<\/code> to get a tensor, calculated at run-time, for the shape of <code>val<\/code> - e.g. <code>tf.gather(val, tf.shape(val)[0] - 1)<\/code><\/p>\n"}]}
{"QuestionId":41269518,"AnswerCount":1,"Tags":"<performance><tensorflow><keras>","CreationDate":"2016-12-21T18:26:50.820","AcceptedAnswerId":null,"OwnerUserId":4313064.0,"Title":"Why is my model twice as fast in Tensorflow than in Keras?","Body":"<p>Am doing the udacity deep learning course which uses tensorflow. I have tried to replicate one of the assignments in keras with tensorflow backend but it is substantially slower taking almost 3 minutes compared with 1.5 minutes for tensorflow. It takes even longer with theano backend - almost 4 minutes. I am not using the GPU.<\/p>\n\n<p>Is there a simple explanation? I was wondering if tensorflow makes better use of multi-processing by putting different nodes of the graph on different processors?<\/p>\n","answers":[{"AnswerId":"41286868","CreationDate":"2016-12-22T15:44:40.403","ParentId":null,"OwnerUserId":"4313064","Title":null,"Body":"<p>Keras beginners mistake. In Keras when you add Dense(20, input_dim=20) this adds a hidden layer and an input layer. The parameter input_dim is the input layer. In tensorflow you add the input layer explicitly. Hence for my Keras model  I accidentally added an extra layer and it took twice as long to complete.<\/p>\n"}]}
{"QuestionId":41269570,"AnswerCount":3,"Tags":"<machine-learning><neural-network><deep-learning><normalization><caffe>","CreationDate":"2016-12-21T18:30:03.507","AcceptedAnswerId":41270097.0,"OwnerUserId":1714410.0,"Title":"What is \"batch normalizaiton\"? why using it? how does it affect prediction?","Body":"<p>Recently, many deep architectures use \"batch normalization\" for training.<\/p>\n\n<p>What is \"batch normalization\"? What does it do mathematically? In what way does it help the training process?<\/p>\n\n<p>How is batch normalization used during training? is it a special layer inserted into the model? Do I need to normalize before each layer, or only once? <\/p>\n\n<p>Suppose I used batched normalization for training. Does this affect my test-time model? Should I replace the batch normalization with some other\/equivalent layer\/operation in my \"deploy\" network?<\/p>\n\n<hr>\n\n<p><a href=\"https:\/\/stackoverflow.com\/q\/29979251\/1714410\">This question<\/a> about batch normalization only covers part of this question, I was aiming and hoping for a more detailed answer. More specifically, I would like to know how training with batch normalization affect test time prediction, i.e., the \"deploy\" network and the TEST phase of the net.<\/p>\n","answers":[{"AnswerId":"41274664","CreationDate":"2016-12-22T01:51:01.233","ParentId":null,"OwnerUserId":"6281477","Title":null,"Body":"<p>As a complement to <a href=\"https:\/\/stackoverflow.com\/a\/41270097\/6281477\">Prune's answer<\/a>, during testing, batch normalization layer will use the average <code>mean\/variance\/scale\/shift<\/code> values from different training iterations to normalize its input(subtract mean and divide by the standard deviation). <\/p>\n\n<p>And the original <a href=\"https:\/\/arxiv.org\/abs\/1502.03167\" rel=\"nofollow noreferrer\">google's batch normalization paper<\/a> only said that it should be a moving average method and no more thorough explanation was provided though. Both caffe and tensorflow use an exponential moving average method.<\/p>\n\n<p>In my experience, a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Moving_average#Simple_moving_average\" rel=\"nofollow noreferrer\">simple moving average method<\/a> usually better than an exponential moving average method, as far as to the validation accuracy(Maybe it need more experiments). For a compare, you can refer to <a href=\"https:\/\/github.com\/knsong\/caffe-mt\" rel=\"nofollow noreferrer\">here<\/a>(I tried the two moving average methods implementations in <code>channel_wise_bn_layer<\/code>, compared with the <code>batch norm<\/code> layer in <a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow noreferrer\">BVLC\/caffe<\/a>). <\/p>\n"},{"AnswerId":"41594572","CreationDate":"2017-01-11T15:21:10.540","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>For what it's worth <a href=\"https:\/\/gist.github.com\/ducha-aiki\/c0d1325f0cebe0b05c36\" rel=\"nofollow noreferrer\">this link<\/a> has an example of using <code>\"BatchNorm\"<\/code> layers in cifar10 classification net.<\/p>\n\n<p>Specifically, it splits the layer between <code>TRAIN<\/code> and <code>TEST<\/code> phases:<\/p>\n\n<pre><code>layer {\n  name: \"bn1\"\n  type: \"BatchNorm\"\n  bottom: \"pool1\"\n  top: \"bn1\"\n  batch_norm_param {\n    use_global_stats: false\n  }\n  param {\n    lr_mult: 0\n  }\n  param {\n    lr_mult: 0\n  }\n  param {\n    lr_mult: 0\n  }\n  include {\n    phase: TRAIN\n  }\n}\nlayer {\n  name: \"bn1\"\n  type: \"BatchNorm\"\n  bottom: \"pool1\"\n  top: \"bn1\"\n  batch_norm_param {\n    use_global_stats: true\n  }\n  param {\n    lr_mult: 0\n  }\n  param {\n    lr_mult: 0\n  }\n  param {\n    lr_mult: 0\n  }\n  include {\n    phase: TEST\n  }\n}\n<\/code><\/pre>\n"},{"AnswerId":"41270097","CreationDate":"2016-12-21T19:02:07.713","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>The batch normalization is for layers that can suffer from deleterious drift.  The math is simple: find the mean and variance of each component, then apply the standard transformation to convert all values to the corresponding Z-scores: subtract the mean and divide by the standard deviation.  This ensures that the component ranges are very similar, so that they'll each have a chance to affect the training deltas (in back-prop).<\/p>\n\n<p>If you're using the network for pure testing (no further training), then simply delete these layers; they've done their job.  If you're training while testing \/ predicting \/ classifying, then leave them in place; the operations won't harm your results at all, and barely slow down the forward computations.<\/p>\n\n<p>As for Caffe specifics, there's really nothing particular to Caffe.  The computation is a basic stats process, and is the same algebra for any framework.  Granted, there will be some optimizations for hardware that supports vector and matrix math, but those consist of simply taking advantage of the chip's built-in operations.<\/p>\n\n<hr>\n\n<p><strong>RESPONSE TO COMMENT<\/strong><\/p>\n\n<p>If you can afford a little extra training time, yes, you'd want to normalize at every layer.  In practice, inserting them less frequently -- say, every 1-3 inceptions -- will work just fine.<\/p>\n\n<p>You can ignore these in deployment because they've already done their job: when there's no back-propagation, there's no drift of weights.  Also, when the model handles only one instance in each batch, the Z-score is always 0: every input is exactly the mean of the batch (being the entire batch).<\/p>\n"}]}
{"QuestionId":41269894,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-21T18:48:40.240","AcceptedAnswerId":null,"OwnerUserId":5691307.0,"Title":"Training on my own image dataset using Tensorflow","Body":"<p>I have some code that I basically cobbled up from different places online and i'm honestly not sure exactly whats going on.<\/p>\n\n<pre><\/pre>\n\n<p>My question is what is happening in the for loop when I call ? Does it queue a batch of 10 images? And what happens when I do it 20 times? I only have 100 images in the directory but it still seems to work? Should I even be using a for loop here? Also, how can I now train a model using these images? I apologize for sounding so confused. I've really been trying to figure this out by looking at other questions and tutorials.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41270181","CreationDate":"2016-12-21T19:07:56.767","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<blockquote>\n  <p>what is happening in the for loop when I call sess.run(images)? <\/p>\n<\/blockquote>\n\n<p>When you're running a bunch of TensorFlow functions, nothing is actually getting computed at first. You're really setting up a computation graph. When you pass something to <code>sess.run<\/code>, that's when your stuff actually gets computed. Maybe you can think of TensorFlow functions as its own language, and <code>sess.run<\/code> as its interpreter.<\/p>\n\n<blockquote>\n  <p>Does it queue a batch of 10 images? And what happens when I do it 20 times? I only have 100 images in the directory but it still seems to work? Should I even be using a for loop here? <\/p>\n<\/blockquote>\n\n<p>None of that code is really necessary to train anything - it's setup stuff that you can do yourself. Loading images myself has worked great so far.<\/p>\n\n<blockquote>\n  <p>Also, how can I now train a model using these images? <\/p>\n<\/blockquote>\n\n<p>That depends what you're trying to train it to do. Generally, you would do something like this: define some computation graph with variables and a loss function, use something like <code>tf.train.AdamOptimizer<\/code> to automatically adjust the values of the variables so your loss function (a measure of how well you're doing) goes down.<\/p>\n\n<blockquote>\n  <p>I apologize for sounding so confused. I've really been trying to figure this out by looking at other questions and tutorials.<\/p>\n<\/blockquote>\n\n<p>No worries! You may want to find some introductory machine learning course or book to figure this out before starting with TensorFlow. <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">This tutorial<\/a> might be a good place to learn about TensorFlow.<\/p>\n"}]}
{"QuestionId":41270243,"AnswerCount":2,"Tags":"<python><python-2.7><numpy><keras>","CreationDate":"2016-12-21T19:12:14.900","AcceptedAnswerId":41271972.0,"OwnerUserId":7327257.0,"Title":"Error installing keras with pip in Python2.7 (Windows 7): numpy not found","Body":"<p>I'm totally new using Python and I have to learn quickly because of my job. I have downloaded Python2.7.13 (and also the pip package. My next step was:<\/p>\n\n<pre><\/pre>\n\n<p>And it finished saying that theano and anaconda2 were installed to and also was fine. \nThe problem came when I tried the following line:<\/p>\n\n<pre><\/pre>\n\n<p>And I get the following message (after others that don't give error):<\/p>\n\n<pre><\/pre>\n\n<p>The fact is that I got out of python and try pip install numpy and it says that I have already have numpy installed and show me in which folder I have it. So I go to that folder and try:<\/p>\n\n<pre><\/pre>\n\n<p>And that returns:<\/p>\n\n<pre><\/pre>\n\n<p>I'm totally new so I'm very lost about what my next step should be and how to fix it, 'cause I'm not familiarized with Python. I would thank any understandable help. \nThanks in advance.<\/p>\n","answers":[{"AnswerId":"41274265","CreationDate":"2016-12-22T00:53:48.630","ParentId":null,"OwnerUserId":"6304774","Title":null,"Body":"<p>You are already using Anaconda Python distribution so you should use conda package manager for installing keras: run <code>conda install keras<\/code> in your terminal. This will install keras and all the dependencies.<\/p>\n"},{"AnswerId":"41271972","CreationDate":"2016-12-21T21:11:31.230","ParentId":null,"OwnerUserId":"1556332","Title":null,"Body":"<p>It sounds like you have vanilla python installed in <code>C:\\Python27\\<\/code> and Anaconda installed in <code>C:\\Users\\Maca\\Anaconda2<\/code>. These are <em>different<\/em>. Anaconda is <em>not an add-on<\/em> to python, it <em>is<\/em> a distribution of python. Anaconda will have the math prerequisites, vanilla will not. You can install Keras on vanilla (it's hard), but it sounds like you want to use Anaconda. Then use <em>only<\/em> anaconda. Run:<\/p>\n\n<pre><code>activate root\n<\/code><\/pre>\n\n<p>You should then be using Anaconda and you can run:<\/p>\n\n<pre><code>pip install keras\n<\/code><\/pre>\n\n<p>Also, keras works on top of a backend so make sure you either have theano (<code>pip install theano<\/code>) or <a href=\"https:\/\/www.tensorflow.org\/\" rel=\"nofollow noreferrer\">tensorflow<\/a>.\nThen, <em>in the same console window<\/em>, using Anaconda, run <code>python<\/code> and type <code>from keras.models import Sequential<\/code>. It should work.<\/p>\n\n<p>From then on, you can access anaconda (as long as it was set up with this command enabled) with <code>activate root<\/code>. This will set <code>python<\/code> and related arguments to refer to <em>Anaconda<\/em> rather than the default <code>C:\\Python27<\/code>.<\/p>\n\n<p>EDIT: If in the future you are planning to use only anaconda, you can delete <code>C:\\Python27<\/code> from your path or even add Anaconda to it. To add anaconda to you path, look up <code>environment variables<\/code> in the start menu, go to the system environment variables, and find <code>PATH<\/code>. If you haven't uninstalled Python27 yet you'll need to remove <code>C:\\Python27\\<\/code> from that path, and then you can add <code>C:\\Users\\Maca\\Anaconda2<\/code> anywhere (probably towards the end depending on how often you use python and how cluttered your path is).<\/p>\n"}]}
{"QuestionId":41271952,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2016-12-21T21:10:09.587","AcceptedAnswerId":45475403.0,"OwnerUserId":6085595.0,"Title":"Effect of class_weight and sample_weight in Keras","Body":"<p>Can someone tell me mathematically how sample_weight and class_weight are used in Keras in the calculation of loss function and metrics? A simple mathematical express will be great.<\/p>\n","answers":[{"AnswerId":"45475403","CreationDate":"2017-08-03T05:18:04.830","ParentId":null,"OwnerUserId":"2848578","Title":null,"Body":"<p>It is a simple multiplication. The loss contributed by the sample is magnified by its sample weight. Assuming <code>i = 1 to n<\/code> samples, a weight vector of sample weights <code>w<\/code> of length <code>n<\/code>, and that the loss for sample <code>i<\/code> is denoted <code>L_i<\/code>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kiHUU.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kiHUU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In Keras in particular, the product of each sample's loss with its weight is divided by the fraction of weights that are not 0 such that the loss per batch is proportional to the number of weight > 0 samples. Let <code>p<\/code> be the proportion of non-zero weights.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LVJ1j.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LVJ1j.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here's the relevant snippet of code from the Keras repo:<\/p>\n\n<pre><code>score_array = loss_fn(y_true, y_pred)\n\nif weights is not None:\n    score_array *= weights\n    score_array \/= K.mean(K.cast(K.not_equal(weights, 0), K.floatx()))\n\nreturn K.mean(score_array)\n<\/code><\/pre>\n\n<p><code>class_weight<\/code> is used in the same way as <code>sample_weight<\/code>; it is just provided as a convenience to specify certain weights across entire classes.<\/p>\n\n<p>The sample weights are currently not applied to metrics, only loss.<\/p>\n"}]}
{"QuestionId":41273054,"AnswerCount":0,"Tags":"<caffe><cudnn>","CreationDate":"2016-12-21T22:34:26.123","AcceptedAnswerId":null,"OwnerUserId":4257226.0,"Title":"Temporarily disable cuDNN in Caffe","Body":"<p>Is it possible to temporarily disable cuDNN in Caffe without having to set engine: Caffe on every single layer in my train_val file? We want to get a deterministic result for refactoring purposes but continue using cuDNN when we're not doing that.<\/p>\n","answers":[]}
{"QuestionId":41273182,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2016-12-21T22:45:13.720","AcceptedAnswerId":42193141.0,"OwnerUserId":753033.0,"Title":"TensorFlow 0.12 tutorials produce warning: \"Rank of input Tensor should be the same as output_rank for column","Body":"<p>I have some experience with writing machine learning programs in python, but I'm new to TensorFlow and am checking it out. My dev environment is a lubuntu 14.04 64-bit virtual machine. I've created a python 3.5 conda environment from miniconda and installed TensorFlow 0.12 and its dependencies. I began trying to run some example code from TensorFlow's tutorials and encountered this warning when calling  in the boston.py example for input functions: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/input_fn\/boston.py\" rel=\"noreferrer\">source<\/a>.<\/p>\n\n<blockquote>\n  <p>WARNING:tensorflow:Rank of input Tensor (1) should be the same as\n  output_rank (2) for column. Will attempt to expand dims. It is highly\n  recommended that you resize your input, as this behavior may change.<\/p>\n<\/blockquote>\n\n<p>After some searching in Google, I found other people encountered this same warning:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6184\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6184<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5098\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5098<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/41068915\/tensorflow-boston-housing-data-tutorial-errors\">Tensorflow - Boston Housing Data Tutorial Errors<\/a><\/li>\n<\/ul>\n\n<p>However, they also experienced errors which prevent code execution from completing. In my case, the code executes with the above warning. Unfortunately, I couldn't find a single answer in those links regarding what caused the warning and how to fix the warning. They all focused on the error. How does one remove the warning? Or is the warning safe to ignore? <\/p>\n\n<p>Cheers!<\/p>\n\n<p>Extra info, I also see the following warnings when running the aforementioned boston.py example.  <\/p>\n\n<blockquote>\n  <p>WARNING:tensorflow:*******************************************************\n  WARNING:tensorflow:TensorFlow's V1 checkpoint format has been\n  deprecated. WARNING:tensorflow:Consider switching to the more\n  efficient V2 format: WARNING:tensorflow:<br>\n  'tf.train.Saver(write_version=tf.train.SaverDef.V2)'\n  WARNING:tensorflow:now on by default.\n  WARNING:tensorflow:*******************************************************<\/p>\n<\/blockquote>\n\n<p>and<\/p>\n\n<blockquote>\n  <p>WARNING:tensorflow:From\n  \/home\/kade\/miniconda3\/envs\/tensorflow\/lib\/python3.5\/site-packages\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn_linear_combined.py:1053\n  in predict.: calling BaseEstimator.predict (from\n  tensorflow.contrib.learn.python.learn.estimators.estimator) with x is\n  deprecated and will be removed after 2016-12-01. Instructions for\n  updating: Estimator is decoupled from Scikit Learn interface by moving\n  into separate class SKCompat. Arguments x, y and batch_size are only\n  available in the SKCompat class, Estimator will only accept input_fn.\n  Example conversion:   est = Estimator(...) -> est =\n  SKCompat(Estimator(...))<\/p>\n<\/blockquote>\n\n<p><strong>UPDATE (2016-12-22):<\/strong>\nI've tracked the warning to this file:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/feature_column_ops.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/feature_column_ops.py<\/a><\/p>\n\n<p>and this code block:<\/p>\n\n<pre><\/pre>\n\n<p>Note the line: <\/p>\n\n<p>The method signature is: <\/p>\n\n<p>The value  is hardcoded as the value of output_rank, but the boston.py example is passing in an  of rank 1. I will continue to investigate.<\/p>\n","answers":[{"AnswerId":"43135755","CreationDate":"2017-03-31T08:14:59.573","ParentId":null,"OwnerUserId":"7796532","Title":null,"Body":"<p>After I specify the shape of the tensor explicitly.<\/p>\n\n<pre><code>continuous_cols = {k: tf.constant(df[k].values, shape=[df[k].size, 1]) for k in CONTINUOUS_COLUMNS}\n<\/code><\/pre>\n\n<p>It works!<\/p>\n"},{"AnswerId":"42193141","CreationDate":"2017-02-12T20:56:00.307","ParentId":null,"OwnerUserId":"5480134","Title":null,"Body":"<p>If you specify the shape of your tensor explicitly:<\/p>\n\n<pre><code>tf.constant(df[k].values, shape=[df[k].size, 1]) \n<\/code><\/pre>\n\n<p>the warning should go away. <\/p>\n"}]}
{"QuestionId":41273361,"AnswerCount":4,"Tags":"<python><tensorflow><neural-network><deep-learning><recurrent-neural-network>","CreationDate":"2016-12-21T23:00:21.377","AcceptedAnswerId":43298689.0,"OwnerUserId":1861627.0,"Title":"Get the last output of a dynamic_rnn in TensorFlow","Body":"<p>I have a 3-D tensor of shape  where the second dimension, i.e. the timesteps, is unknown. I use  to process such input, like in the following snippet:<\/p>\n\n<pre><\/pre>\n\n<p>Actually, running this snipped with some actual numbers, I have some reasonable results:<\/p>\n\n<pre><\/pre>\n\n<p>And the output is:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to get a 3-D tensor of shape  with the <em>last relevant output<\/em> of the dynamic RNN? Thanks!<\/p>\n","answers":[{"AnswerId":"43298689","CreationDate":"2017-04-08T18:15:11.510","ParentId":null,"OwnerUserId":"1953800","Title":null,"Body":"<p>This is what <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"noreferrer\">gather_nd<\/a> is for!<\/p>\n\n<pre><code>def extract_axis_1(data, ind):\n    \"\"\"\n    Get specified elements along the first axis of tensor.\n    :param data: Tensorflow tensor that will be subsetted.\n    :param ind: Indices to take (one for each element along axis 0 of data).\n    :return: Subsetted tensor.\n    \"\"\"\n\n    batch_range = tf.range(tf.shape(data)[0])\n    indices = tf.stack([batch_range, ind], axis=1)\n    res = tf.gather_nd(data, indices)\n\n    return res\n<\/code><\/pre>\n\n<p>In your case:<\/p>\n\n<pre><code>output = extract_axis_1(output, lengths - 1)\n<\/code><\/pre>\n\n<p>Now <code>output<\/code> is a tensor of dimension <code>[batch_size, num_cells]<\/code>.<\/p>\n"},{"AnswerId":"41273737","CreationDate":"2016-12-21T23:44:19.193","ParentId":null,"OwnerUserId":"1861627","Title":null,"Body":"<p>Actually, the solution was not that hard. I implemented the following code:<\/p>\n\n<pre><code>slices = []\nfor index, l in enumerate(tf.unstack(lengths)):\n    slice = tf.slice(rnn_out, begin=[index, l - 1, 0], size=[1, 1, 3])\n    slices.append(slice)\nlast = tf.concat(0, slices)\n<\/code><\/pre>\n\n<p>So, the full snippet would be the following:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nbatch = 2\ndim = 3\nhidden = 4\n\nlengths = tf.placeholder(dtype=tf.int32, shape=[batch])\ninputs = tf.placeholder(dtype=tf.float32, shape=[batch, None, dim])\ncell = tf.nn.rnn_cell.GRUCell(hidden)\ncell_state = cell.zero_state(batch, tf.float32)\noutput, _ = tf.nn.dynamic_rnn(cell, inputs, lengths, initial_state=cell_state)\n\ninputs_ = np.asarray([[[0, 0, 0], [1, 1, 1], [2, 2, 2], [3, 3, 3]],\n                    [[6, 6, 6], [7, 7, 7], [8, 8, 8], [9, 9, 9]]],\n                    dtype=np.int32)\nlengths_ = np.asarray([3, 1], dtype=np.int32)\n\nslices = []\nfor index, l in enumerate(tf.unstack(lengths)):\n    slice = tf.slice(output, begin=[index, l - 1, 0], size=[1, 1, 3])\n    slices.append(slice)\nlast = tf.concat(0, slices)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    outputs = sess.run([output, last], {inputs: inputs_, lengths: lengths_})\n    print 'RNN output:'\n    print(outputs[0])\n    print\n    print 'last relevant output:'\n    print(outputs[1])\n<\/code><\/pre>\n\n<p>And the output:<\/p>\n\n<pre><code>RNN output:\n[[[ 0.          0.          0.          0.        ]\n [-0.06667092 -0.09284072  0.01098599 -0.03676109]\n [-0.09101103 -0.19828682  0.03546784 -0.08721405]\n [ 0.          0.          0.          0.        ]]\n\n[[-0.00025157 -0.05704876  0.05527233 -0.03741353]\n [ 0.          0.          0.          0.        ]\n [ 0.          0.          0.          0.        ]\n [ 0.          0.          0.          0.        ]]]\n\nlast relevant output:\n[[[-0.09101103 -0.19828682  0.03546784]]\n\n [[-0.00025157 -0.05704876  0.05527233]]]\n<\/code><\/pre>\n"},{"AnswerId":"44834855","CreationDate":"2017-06-29T20:55:44.447","ParentId":null,"OwnerUserId":"8139253","Title":null,"Body":"<p>From the following two sources, <\/p>\n\n<p><a href=\"http:\/\/www.wildml.com\/2016\/08\/rnns-in-tensorflow-a-practical-guide-and-undocumented-features\/\" rel=\"noreferrer\">http:\/\/www.wildml.com\/2016\/08\/rnns-in-tensorflow-a-practical-guide-and-undocumented-features\/<\/a><\/p>\n\n<pre><code>outputs, last_states = tf.nn.dynamic_rnn(\ncell=cell,\ndtype=tf.float64,\nsequence_length=X_lengths,\ninputs=X)\n<\/code><\/pre>\n\n<p>Or <a href=\"https:\/\/github.com\/ageron\/handson-ml\/blob\/master\/14_recurrent_neural_networks.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/ageron\/handson-ml\/blob\/master\/14_recurrent_neural_networks.ipynb<\/a>,<\/p>\n\n<p>It is clear the last_states can be directly extracted from the SECOND output of the dynamic_rnn call. It will give you the last_states across <strong>all<\/strong> layers (in LSTM it is compsed from LSTMStateTuple) , while the outputs contains all the states in the <strong>last<\/strong> layer.<\/p>\n"},{"AnswerId":"47894359","CreationDate":"2017-12-19T20:00:36.683","ParentId":null,"OwnerUserId":"5395078","Title":null,"Body":"<p>Okay \u2014 so, looks like there actually <b>is<\/b> an easier solution. As @Shao Tang and @Rahul mentioned, the preferred way to do this would be by accessing the final cell state. Here\u2019s why:<\/p>\n\n<ul>\n<li>If you look at the GRUCell source code (below), you\u2019ll see that the \u201cstate\u201d that the cell maintains is actually the hidden weights themselves. So, when the <code>tf.nn.dynamic_rnn<\/code> returns the final state, it is actually returning the final hidden weights that you are interested in. To prove this, I just tweaked your setup and got the results:<\/li>\n<\/ul>\n\n<p>GRUCell Call (rnn_cell_impl.py):<\/p>\n\n<pre><code>def call(self, inputs, state):\n\"\"\"Gated recurrent unit (GRU) with nunits cells.\"\"\"\nif self._gate_linear is None:\n      bias_ones = self._bias_initializer\nif self._bias_initializer is None:\n        bias_ones = init_ops.constant_initializer(1.0, dtype=inputs.dtype)\nwith vs.variable_scope(\"gates\"):  # Reset gate and update gate.\nself._gate_linear = _Linear(\n            [inputs, state],\n2 * self._num_units,\nTrue,\nbias_initializer=bias_ones,\nkernel_initializer=self._kernel_initializer)\n    value = math_ops.sigmoid(self._gate_linear([inputs, state]))\n    r, u = array_ops.split(value=value, num_or_size_splits=2, axis=1)\n    r_state = r * state\nif self._candidate_linear is None:\nwith vs.variable_scope(\"candidate\"):\nself._candidate_linear = _Linear(\n            [inputs, r_state],\nself._num_units,\nTrue,\nbias_initializer=self._bias_initializer,\nkernel_initializer=self._kernel_initializer)\n    c = self._activation(self._candidate_linear([inputs, r_state]))\n    new_h = u * state + (1 - u) * c\nreturn new_h, new_h\n<\/code><\/pre>\n\n<p>Solution:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nbatch = 2\ndim = 3\nhidden = 4\n\nlengths = tf.placeholder(dtype=tf.int32, shape=[batch])\ninputs = tf.placeholder(dtype=tf.float32, shape=[batch, None, dim])\ncell = tf.nn.rnn_cell.GRUCell(hidden)\ncell_state = cell.zero_state(batch, tf.float32)\noutput, state = tf.nn.dynamic_rnn(cell, inputs, lengths, initial_state=cell_state)\n\ninputs_ = np.asarray([[[0, 0, 0], [1, 1, 1], [2, 2, 2], [3, 3, 3]],\n                    [[6, 6, 6], [7, 7, 7], [8, 8, 8], [9, 9, 9]]],\n                    dtype=np.int32)\nlengths_ = np.asarray([3, 1], dtype=np.int32)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    output_, state_ = sess.run([output, state], {inputs: inputs_, lengths: lengths_})\n    print (output_)\n    print (state_)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[[ 0.          0.          0.          0.        ]\n  [-0.24305521 -0.15512943  0.06614969  0.16873555]\n  [-0.62767833 -0.30741733  0.14819752  0.44313088]\n  [ 0.          0.          0.          0.        ]]\n\n [[-0.99152333 -0.1006391   0.28767768  0.76360202]\n  [ 0.          0.          0.          0.        ]\n  [ 0.          0.          0.          0.        ]\n  [ 0.          0.          0.          0.        ]]]\n[[-0.62767833 -0.30741733  0.14819752  0.44313088]\n [-0.99152333 -0.1006391   0.28767768  0.76360202]]\n<\/code><\/pre>\n\n<ul>\n<li><p>For other readers who are working with the LSTMCell (another popular option), things work a little differently. The LSTMCell maintains the state in a different way - cell state is either a tuple or a concatenated version of the actual cell state and the hidden state. So, to access the final hidden weights, you could set (<code>is_state_tuple<\/code> to <code>True<\/code>) during cell-initialization, and the final state will be a tuple : (final cell state, final hidden weights). So, in this case,<\/p>\n\n<p>_, (_, h) = tf.nn.dynamic_rnn(cell, inputs, lengths, initial_state=cell_state)<\/p><\/li>\n<\/ul>\n\n<p>will give you the final weights.<\/p>\n\n<p>References: \n<a href=\"https:\/\/stackoverflow.com\/questions\/41789133\/c-state-and-m-state-in-tensorflow-lstm\">c_state and m_state in Tensorflow LSTM<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/438604fc885208ee05f9eef2d0f2c630e1360a83\/tensorflow\/python\/ops\/rnn_cell_impl.py#L308\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/438604fc885208ee05f9eef2d0f2c630e1360a83\/tensorflow\/python\/ops\/rnn_cell_impl.py#L308<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/438604fc885208ee05f9eef2d0f2c630e1360a83\/tensorflow\/python\/ops\/rnn_cell_impl.py#L415\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/438604fc885208ee05f9eef2d0f2c630e1360a83\/tensorflow\/python\/ops\/rnn_cell_impl.py#L415<\/a><\/p>\n"}]}
{"QuestionId":41273756,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-21T23:47:09.633","AcceptedAnswerId":null,"OwnerUserId":3150181.0,"Title":"TensorFlow while_loop parallelization TensorArray","Body":"<p>I don't exactly understand how the while_loop parallelization works. Suppose I have a TensorArray having 10 Tensors all of same shape. Now suppose the computations in the loop body for the first 5 Tensors are independent of the computations in the remaining 5 Tensors. Would TensorFlow run these two in parallel? Also if I use a Tensor instead of a TensorArray and made the updates to it using scatter_update, would it pass the gradients properly during backprop?<\/p>\n","answers":[{"AnswerId":"41295384","CreationDate":"2016-12-23T04:55:06.403","ParentId":null,"OwnerUserId":"6410979","Title":null,"Body":"<p>You should probably get the parallel execution of the first 5 iterations and the second 5 iterations. I can say for sure if you provide a code sample. <\/p>\n"}]}
{"QuestionId":41276012,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-22T04:55:14.577","AcceptedAnswerId":41276283.0,"OwnerUserId":6112394.0,"Title":"sess.run(Tensor()) does nothing","Body":"<p>I am attempting to feed my neural net with data stored in a TFRecords. The data is extracted using <strong>inputs()<\/strong> similarly to what is on the tensorflow webpage. The data it returns are two Tensors. When I attempt to <strong>eval()<\/strong> the Tensors so that I can send them through my model with feed_dict, my computer just sits and does nothing. The code does not return an error, my system monitor leads me to believe nothing is going on other than my GPU ram is almost full. <\/p>\n\n<pre><\/pre>\n\n<p>**edit:**My inputs function. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41276283","CreationDate":"2016-12-22T05:24:32.430","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> Try adding the following line after you initialize your variables, and before trying to evaluate <code>input_batch<\/code>:<\/p>\n\n<pre><code>tf.train.start_queue_runners(sess)\n<\/code><\/pre>\n\n<p>It is difficult to be certain without seeing the implementation of <code>inputs()<\/code>, but the tensor name <code>\"input\/shuffle_batch\"<\/code> suggests that the function is building the inputs using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#shuffle_batch\" rel=\"nofollow noreferrer\"><code>tf.train.shuffle_batch()<\/code><\/a> function.<\/p>\n\n<p>Many TensorFlow functions for input processing create prefetching queues internally, including <code>tf.train.shuffle_batch()<\/code>. These prefetching queues initially empty, and your <code>sess.run(input_batch)<\/code> call is probably blocked waiting for elements to be put in those queues. Currently, the way that this typically happens is using a \"queue runner thread\", which is the name for one or more background threads that are started when you call <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/coordinator_and_queuerunner#start_queue_runners\" rel=\"nofollow noreferrer\"><code>tf.train.start_queue_runners()<\/code><\/a>.<\/p>\n\n<p>This is one of the more complicated areas of TensorFlow, and it's something we're working to improve. In the mean time, you may find the documentation on <a href=\"https:\/\/www.tensorflow.org\/how_tos\/threading_and_queues\/\" rel=\"nofollow noreferrer\">threading and queues in TensorFlow<\/a> useful.<\/p>\n"}]}
{"QuestionId":41276085,"AnswerCount":1,"Tags":"<c++><opencv><tensorflow>","CreationDate":"2016-12-22T05:05:19.467","AcceptedAnswerId":null,"OwnerUserId":3994228.0,"Title":"tensorflow + opencv webcam hangs","Body":"<p>Environment: MAC machine, running my code inside virtual machine with guest OS: Ubuntu 14.4 LTS.<\/p>\n\n<p>I am compiling openCV within tensorflow workspace under examples. My WORKSPACE and opencv.BUILD file look similar to the one mentioned <a href=\"https:\/\/stackoverflow.com\/questions\/34984290\/building-opencv-code-using-bazel\">here<\/a><\/p>\n\n<p>My BUILD file for the opencv + tensorflow project looks like following: <\/p>\n\n<pre><\/pre>\n\n<p>If i disable tensorflow dependences (and also comment the tensorflow related code). I can see that the webcam is captures properly. like this:<\/p>\n\n<pre><\/pre>\n\n<p>But if i still keep the code commented\/uncommented and also keep the tensorflow dependences my webcam hangs at VideoCapture::read()<\/p>\n\n<p>By default, opencv use FFMPEG codec and i tried enabling and disabling FFMPEG. Can someone please help me why when tensorflow library is compiled in the project makes my openCV read() hangs?<\/p>\n","answers":[{"AnswerId":"41293174","CreationDate":"2016-12-22T23:33:09.927","ParentId":null,"OwnerUserId":"1653720","Title":null,"Body":"<p>The OpenCV Bazel build configuration you linked above appears to just glob all the .so files that CMake built. Maybe you need to pass the -DWITH_FFMPEG=OFF cflag to CMake? If you pass it to Bazel as you did above, it's only going to apply to the compilation of main.cc.<\/p>\n"}]}
{"QuestionId":41276110,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2016-12-22T05:07:06.190","AcceptedAnswerId":null,"OwnerUserId":7324309.0,"Title":"how to retrieve the predicted labels in TensorFlow's cifar10 example?","Body":"<p>I am new to tensorflow and I am currently trying to use the cifar10 example on another data set which has no labels. In the evaluating part of this example, it just outputs the prediction accuracy. I was wondering how can I modify this code to output the predicted labels of the test case?<\/p>\n\n<p>Here is the tutorial's code:<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_eval.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_eval.py<\/a><\/p>\n","answers":[]}
{"QuestionId":41276935,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><word2vec>","CreationDate":"2016-12-22T06:24:14.073","AcceptedAnswerId":41277032.0,"OwnerUserId":6782093.0,"Title":"TensorFlow : How and where to specify save path in word2vec?","Body":"<p>I am trying to implementing word2vec model of tensorflow and went through the tutorial of tensorflows <a href=\"https:\/\/www.tensorflow.org\/tutorials\/word2vec\/#scaling_up_with_noise-contrastive_training\" rel=\"nofollow noreferrer\">Vector representation of words<\/a>.<\/p>\n\n<p>After that I compiled the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/embedding\/word2vec_optimized.py\" rel=\"nofollow noreferrer\">word2vec code<\/a> in my system and I am getting this output.<\/p>\n\n<pre><\/pre>\n\n<p>I am new to TensorFlow and I don't know where and how should I specify ,  and .<\/p>\n\n<p>Also is there a more detailed tutorial available on TensorFlow vector representation of words to understand the  model better. <\/p>\n","answers":[{"AnswerId":"41277032","CreationDate":"2016-12-22T06:31:03.377","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>When you run the script, you need to provide these as flags. Try running it like this,<\/p>\n\n<pre><code>python word2vec.py --train_data &lt;path to training data&gt; --eval_data &lt;path to eval&gt; --save_path &lt;save directory&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":41276972,"AnswerCount":1,"Tags":"<mysql><tensorflow>","CreationDate":"2016-12-22T06:27:00.233","AcceptedAnswerId":41291796.0,"OwnerUserId":7181306.0,"Title":"Read Mysql Database in Tensorflow","Body":"<p>I can't find the way that using MySQL database in TensorFlow.<\/p>\n\n<p>I made the table and sensor data for reference.<\/p>\n\n<p>Here is the question.<\/p>\n\n<ol>\n<li><p>What should I use to read MySQL database in TensorFlow?<\/p><\/li>\n<li><p>I find the way that making CSV file in MySQL, and read it in TensorFlow. But it is not real-time data. I want to use the data in real-time. <\/p><\/li>\n<\/ol>\n\n<p>Please help me. Thank you.<\/p>\n","answers":[{"AnswerId":"41291796","CreationDate":"2016-12-22T21:19:29.747","ParentId":null,"OwnerUserId":"1327070","Title":null,"Body":"<p>You'd probably need to just use a standard python interface for mysql and feed it into tensorflow. Here are some examples: <\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/35644264\/how-to-read-data-into-tensorflow\">How to read data into Tensorflow?<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/www.quora.com\/What-is-the-best-way-to-read-data-into-Tensorflow\" rel=\"nofollow noreferrer\">https:\/\/www.quora.com\/What-is-the-best-way-to-read-data-into-Tensorflow<\/a><\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":41277913,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-12-22T07:33:12.960","AcceptedAnswerId":null,"OwnerUserId":1279459.0,"Title":"How to display image","Body":"<p>I am trying to display images from the CIFAR-10 TensorFlow tutorial. The images become transformed so that the values  read are floats more less between -1 and 3. I'm not show what kind of transformation has been applied. How can I display them to see the original content?<\/p>\n\n<p>Here is what the part of the image output looks like:<\/p>\n\n<pre><\/pre>\n\n<p>...<\/p>\n\n<pre><\/pre>\n\n<p>This is the link to the tutorial:\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/<\/a><\/p>\n\n<p>Edit:<\/p>\n\n<p>Rescaling does not seem to work for me:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XxXYa.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XxXYa.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41283320","CreationDate":"2016-12-22T12:34:33.557","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>Try scaling the image to be between 0 and 255? Subtract the min and divide by its new max.<\/p>\n"},{"AnswerId":"41287152","CreationDate":"2016-12-22T16:00:13.173","ParentId":null,"OwnerUserId":"7245907","Title":null,"Body":"<p>A couple of ways to do this, for the greyscale MNIST images:<\/p>\n\n<pre><code>tmp = mnist.train.images[0]\ntmp = tmp.reshape((28,28))\n\nplt.imshow(tmp, cmap = cm.Greys)\nplt.show()\n<\/code><\/pre>\n\n<p>Or, for CIFAR-10 images:\nCode below taken from <a href=\"https:\/\/houxianxu.github.io\/implementation\/SoftmaxRegression.html\" rel=\"nofollow noreferrer\">this<\/a> tutorial<\/p>\n\n<pre><code>def visualize_sample(X_train, y_train, classes, samples_per_class=7):\n\"\"\"visualize some samples in the training datasets \"\"\"\nnum_classes = len(classes)\nfor y, cls in enumerate(classes):\n    idxs = np.flatnonzero(y_train == y) # get all the indexes of cls\n    idxs = np.random.choice(idxs, samples_per_class, replace=False)\n    for i, idx in enumerate(idxs): # plot the image one by one\n        plt_idx = i * num_classes + y + 1 # i*num_classes and y+1 determine the row and column respectively\n        plt.subplot(samples_per_class, num_classes, plt_idx)\n        plt.imshow(X_train[idx].astype('uint8'))\n        plt.axis('off')\n        if i == 0:\n            plt.title(cls)\nplt.show()\n<\/code><\/pre>\n"}]}
{"QuestionId":41278270,"AnswerCount":0,"Tags":"<machine-learning><deep-learning><keras><conv-neural-network>","CreationDate":"2016-12-22T07:56:24.773","AcceptedAnswerId":null,"OwnerUserId":2653493.0,"Title":"Using custom filters generated from KMeans as the first set of filters in Keras CNN","Body":"<p>I've generated a bunch of filters from some images using KMeans and I want to initialise the filters of a Keras Convolutional Neural Network. Is that possible and how would I implement that? Would I need to change the source code for my needs or can I use layers.core.Lambda to implement that? <\/p>\n","answers":[]}
{"QuestionId":41278506,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-22T08:10:52.037","AcceptedAnswerId":null,"OwnerUserId":6343552.0,"Title":"Tensorflow - How to batch the dataset","Body":"<p>I am doing a convolution neural network for digit recognition. I want to train an image dataset but I don't know how to <strong>\"batch\"<\/strong> the training data.<\/p>\n\n<p>I get two arrays storing the train_image and train_label:<\/p>\n\n<pre><\/pre>\n\n<p>Now, I want to batch the training data with batch size = 50<\/p>\n\n<pre><\/pre>\n\n<p>When I print out the train_image_batch<\/p>\n\n<pre><\/pre>\n\n<p>I expect the shape should be <\/p>\n\n<p>Am I doing something wrong here?<\/p>\n","answers":[{"AnswerId":"46132809","CreationDate":"2017-09-09T16:35:20.117","ParentId":null,"OwnerUserId":"3152600","Title":null,"Body":"<p><code>shuffle_batch<\/code> expects single sample by default. To enforce it to accept multiple samples pass <code>enqueue_many=True<\/code>. Refer <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.3\/api_docs\/python\/tf\/train\/shuffle_batch\" rel=\"nofollow noreferrer\">doc<\/a><\/p>\n\n<pre><code>train_image_batch, train_label_batch = tf.train.shuffle_batch(\n    [train_image, train_label], batch_size = 50, enqueue_many=True, capacity = 50000, min_after_dequeue = 10000)\n\nprint(train_image_batch.shape)\n\nOutput:\n(50, 1024)\n<\/code><\/pre>\n"}]}
{"QuestionId":41278958,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-22T08:42:50.727","AcceptedAnswerId":null,"OwnerUserId":6408452.0,"Title":"How does tensorflow manage GPU memory for multiple sessions?","Body":"<p>[: ]<\/p>\n\n<p>I have 5 networks with different structure and parameters. I want to deploy them on a server with a GPU. In my understanding, it is more efficient when data is processed in batches. But I don't know how to determine the batch size. <\/p>\n\n<p>So I tried the following:<\/p>\n\n<pre><\/pre>\n\n<p>Running  got warnings like <\/p>\n\n<pre><\/pre>\n\n<p>It seemed that  could be ran with  despite that  is too large.<\/p>\n\n<p>After that the server received requests to use the other four networks. So I loaded the networks with four sessions:<\/p>\n\n<pre><\/pre>\n\n<p>Then I ran the first network again:<\/p>\n\n<pre><\/pre>\n\n<p>This time I got an <\/p>\n\n<pre><\/pre>\n\n<p>So I decided to unload networks,<\/p>\n\n<pre><\/pre>\n\n<p>However I still got the exception  when I called  after closing the other four sessions.<\/p>\n\n<p>My questions are:<\/p>\n\n<ul>\n<li><p>Does  release GPU memory? Why is  not able to forward after I launched and closed ?<\/p><\/li>\n<li><p>Are there any better ways to deploy more than one networks on a server with a GPU?<\/p><\/li>\n<\/ul>\n\n<p>I thought that maybe I could launch 5 processes, one for each network, and called<\/p>\n\n<pre><\/pre>\n\n<p>in each process. However I was afraid that the valid batch size would become smaller compared to .<\/p>\n\n<hr>\n\n<p>[Update]<\/p>\n\n<p>I run the following code<\/p>\n\n<pre><\/pre>\n\n<p>When I called  for the first time, I only got a warning<\/p>\n\n<pre><\/pre>\n\n<p>After I launched the other  sessions, calling  yield an exception . I tried closing those sessions, but it didn't help.<\/p>\n\n<p>A part of the log message:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41292086","CreationDate":"2016-12-22T21:42:59.213","ParentId":null,"OwnerUserId":"6088060","Title":null,"Body":"<p>Tensors in TensorFlow hold most of the memory. Most of them are immutable, and live within session.run() call. Others, such as variable, queue, can live longer. Depending on the type of session you are using, DirectSession releases their variables upon closing the session. <\/p>\n\n<p>So from your code, it is a bit strange that opening other sessions, running the global initializer, and closing them has impact on memory usage. Maybe you can simplify and share the complete repro case so we can understand the full picture.<\/p>\n"}]}
{"QuestionId":41279085,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><neural-network><deep-learning>","CreationDate":"2016-12-22T08:49:41.077","AcceptedAnswerId":41286210.0,"OwnerUserId":6119768.0,"Title":"MNIST classification in tensorflow, RecursionError: maximum recursion depth exceeded","Body":"<p>I ran a Neural Network model for MNIST classification and received error-<\/p>\n\n<pre><\/pre>\n\n<p>I checked some of the issues on stackoverflow and tried to increase the recursion limit to 1500 but did not work. \nHow should I increase the limit? An how do I know what limit will cause stack overflow?<\/p>\n\n<p>I followed the tutorial from <a href=\"https:\/\/www.youtube.com\/watch?v=PwAGxqrXSCs&amp;index=47&amp;list=PLQVvvaa0QuDfKTOs3Keq_kaG2P55YRn5v\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>I have Anaconda 3.5 distribution on my windows 10 machine.<\/p>\n\n<p>The full code is here-<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41286210","CreationDate":"2016-12-22T15:11:05.907","ParentId":null,"OwnerUserId":"7331029","Title":null,"Body":"<p>Looks like you're using code from <a href=\"https:\/\/pythonprogramming.net\/tensorflow-neural-network-session-machine-learning-tutorial\/\" rel=\"nofollow noreferrer\">PythonProgramming.net's deep learning course<\/a><\/p>\n\n<p>In your code, you're doing: <\/p>\n\n<pre><code>def train_neural_network(x):\n    prediction = train_neural_network(x)\n<\/code><\/pre>\n\n<p>You're calling the function within the function to run, which is what is never-ending. <\/p>\n\n<p>Consider your objectives here. Should prediction be the return of train_neural_network, or the return from the neural_network_model?<\/p>\n\n<pre><code>def train_neural_network(x):\n    prediction = neural_network_model(x)\n<\/code><\/pre>\n"},{"AnswerId":"41279196","CreationDate":"2016-12-22T08:57:03.390","ParentId":null,"OwnerUserId":"356528","Title":null,"Body":"<p>I don't know what the exact code is supposed to be, but I'm quite sure the following lines are wrong:<\/p>\n\n<pre><code>def train_neural_network(x):\n    prediction = train_neural_network(x)\n<\/code><\/pre>\n\n<p>This will cause an infinite recursion, and increasing the recursion limit will not solve the problem.<\/p>\n"}]}
{"QuestionId":41279527,"AnswerCount":0,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network><pycaffe>","CreationDate":"2016-12-22T09:16:03.013","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"How to create LMDB files for semantic segmentation?","Body":"<p>I have downloaded <a href=\"http:\/\/www.cs.stanford.edu\/~roozbeh\/pascal-context\/\" rel=\"nofollow noreferrer\">PASCAL trainval.tar.gz and 33_context_labels<\/a> and saved into two folders for running semantic segmentation. The <strong>trainval<\/strong> folder includes images with <strong>.mat<\/strong> format and ground truth images in <strong>33_context_labels<\/strong> folder are in <strong>.png<\/strong> format. <\/p>\n\n<ul>\n<li>How can I divide the training data (trainval folder) and corresponding label images (33_context_labels folder)  into 2 sets: One for training (5\/6 of images) and the other for validation (1\/6 of images)?<\/li>\n<li>How can I create LMDB files for the training and validation in 2 LMDB databases? (train_lmdb for training the model and validation_lmbd for model evaluation)<\/li>\n<\/ul>\n\n<p>Your help is appreciated. <\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":41279836,"AnswerCount":0,"Tags":"<python><neural-network><keras><pre-trained-model>","CreationDate":"2016-12-22T09:31:11.333","AcceptedAnswerId":null,"OwnerUserId":4389065.0,"Title":"Using a pre trained network in Keras","Body":"<p>I am trying to do transfer learning in Keras and couldn't find any good examples. I'm getting an error when trying this:<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting an error: <\/p>\n\n<p>The error is on the line where I have added VGG16<\/p>\n\n<p>Any help will be much appreciated<\/p>\n","answers":[]}
{"QuestionId":41281278,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><keras>","CreationDate":"2016-12-22T10:43:54.667","AcceptedAnswerId":41290058.0,"OwnerUserId":3827807.0,"Title":"Normalization of input in NN for numeric values","Body":"<p>I read a lot of papers concerning the normalization of image in CNN.\nBut I didn't find answers for this case.\nLet's says, the input of my first layer is :\n[1, 4234, 90420394.43, 0]\nDo I have to normalize it ?\nBecause, the gradient descent should be able to recognize itself what are the important parameters and how to handle it with the W's matrix no ?<\/p>\n\n<p>And last question, if I really have to normalize this, will I have a vector of value between 0 and 1 ?\nIsn't the information lost ?\n(And because it just seems to be a rescale, the relative difference stays the same, so, what is the matter in this case ?)<\/p>\n","answers":[{"AnswerId":"41290058","CreationDate":"2016-12-22T19:03:28.730","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p>The answer is not specific to images in CNNs, but rather to neural networks in general. Your intuition that gradient descent should just work it out is correct, the system should learn to scale naturally, but <strong>having inputs on the same<\/strong> scale helps keep gradient steps on the same scale, which will help your network converge faster (without this, it is possible a network may not converge).<\/p>\n\n<p>Additionally, techniques like regularization work with the weight values directly, so may be affected by unbalanced scales.<\/p>\n\n<p>There is also <a href=\"https:\/\/gab41.lab41.org\/batch-normalization-what-the-hey-d480039a9e3b\" rel=\"nofollow noreferrer\">batch normalization<\/a>, which is often used in image CNNs, and does not relate to normalizing input values.<\/p>\n"}]}
{"QuestionId":41281556,"AnswerCount":2,"Tags":"<image-processing><tensorflow><deep-learning><convolution>","CreationDate":"2016-12-22T10:58:08.743","AcceptedAnswerId":41283663.0,"OwnerUserId":7329645.0,"Title":"Why can't I train overfeat network with convolution layer 1x1","Body":"<p>I'm trying to modify tensorflow slim overfeat network to classify small image classes, image size is 60*60 and 3 classes. \nI'm use tensorflow v0.12 on Ubuntu 14.04 with TITAN X GPU.<\/p>\n\n<p>My first network is <\/p>\n\n<pre><\/pre>\n\n<p>I'm use cross entroy loss with tf.nn.sparse_softmax_cross_entropy_with_logits function.<\/p>\n\n<p>And train result is\n<a href=\"https:\/\/i.stack.imgur.com\/CHpCU.jpg\" rel=\"nofollow noreferrer\">Loss And Accuracy with one 1x Conv<\/a><\/p>\n\n<p>This result is passable.\nI'm trying to add one 1x1 conv after fc7 because I think 1x1 conv is same full connected layer and may be improve accuracy.<\/p>\n\n<pre><\/pre>\n\n<p>But I got unreliable result :\n<a href=\"https:\/\/i.stack.imgur.com\/3cObm.jpg\" rel=\"nofollow noreferrer\">Loss And Accuracy with two 1x1 Conv<\/a><\/p>\n\n<p>This network don't be optimized with loss 1. <\/p>\n\n<p>Why I can't add more 1x1 conv or fc layers?<\/p>\n\n<p>And how can I improve this network?<\/p>\n","answers":[{"AnswerId":"41282525","CreationDate":"2016-12-22T11:49:08.637","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>1x1 Convolution is not the same as a fully connected layer. Alone the parameter count is vastly different. A 1x1 Convolution is the weighted sum of all the pixel, which are in the same location in the image, from all the previous filter outputs. <\/p>\n\n<p>A fully connected layer on the other hand takes all the pixels of all the filters into account for every new pixel in the current layer.<\/p>\n\n<p>You should take a look at <a href=\"http:\/\/deeplearning.net\/tutorial\/contents.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/tutorial\/contents.html<\/a> to understand convolutions better.<\/p>\n\n<p>For the final layers fully connected layers are used to combine the features extracted in the previous layers and combine them to the final output. <\/p>\n"},{"AnswerId":"41283663","CreationDate":"2016-12-22T12:55:08.963","ParentId":null,"OwnerUserId":"3109123","Title":null,"Body":"<p>(1,1) convolution layer is not a fully connected layer. if you want to implement fully connected layer as convolution layer you should add last layer kernel size of the layer before. <\/p>\n\n<p>(if the feature map of the layer before is 50x50 your last layer should have a kernel of 50 x 50). convolution layer with (1,1) kernel size is similar ro mlp layer. if tyou want to undertsand more how it works read this paper <a href=\"https:\/\/arxiv.org\/abs\/1312.4400\" rel=\"nofollow noreferrer\">Network in Network<\/a><\/p>\n\n<p>If I understood well, you want to get ride of the fully connected layer. so you have to do two things: <\/p>\n\n<ul>\n<li>assure that you are reducing the last layer to the size of your class by using convolution layer (1,1) with channel equal to your output classes. <\/li>\n<li>use global average pooling to reduce the feature map to 1 and then feed the result to the softmax. <\/li>\n<\/ul>\n"}]}
{"QuestionId":41283115,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-22T12:23:08.520","AcceptedAnswerId":41285054.0,"OwnerUserId":5046896.0,"Title":"Tensorflow, difference between tf.nn.softmax_cross_entropy_with_logits and tf.nn.sparse_softmax_cross_entropy_with_logits","Body":"<p>I have read the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification\" rel=\"nofollow noreferrer\">docs of both functions<\/a>, but as far as I know, for function , the result is the cross entropy loss, in which the dimensions of  and  are the same.<\/p>\n\n<p>But, for function , the dimensions of  and  are not the same?<\/p>\n\n<p>Could you give a more detail example of ?<\/p>\n","answers":[{"AnswerId":"41285054","CreationDate":"2016-12-22T14:08:52.170","ParentId":null,"OwnerUserId":"1951176","Title":null,"Body":"<p>The difference is that <code>tf.nn.softmax_cross_entropy_with_logits<\/code> doesn't assume that the classes are mutually exclusive: <\/p>\n\n<blockquote>\n  <p>Measures the probability error in discrete classification tasks in\n  which each class is independent and not mutually exclusive. For\n  instance, one could perform multilabel classification where a picture\n  can contain both an elephant and a dog at the same time.<\/p>\n<\/blockquote>\n\n<p>Compare with <code>sparse_*<\/code>:<\/p>\n\n<blockquote>\n  <p>Measures the probability error in discrete classification tasks in\n  which the classes are mutually exclusive (each entry is in exactly one\n  class). For example, each CIFAR-10 image is labeled with one and only\n  one label: an image can be a dog or a truck, but not both.<\/p>\n<\/blockquote>\n\n<p>As such, with sparse functions, the dimensions of <code>logits<\/code> and <code>labels<\/code> are not the same: <code>labels<\/code> contain one number per example, whereas <code>logits<\/code> the number of classes per example, denoting probabilities.<\/p>\n"}]}
{"QuestionId":41284752,"AnswerCount":1,"Tags":"<neural-network><theano><lasagne>","CreationDate":"2016-12-22T13:52:22.003","AcceptedAnswerId":41296057.0,"OwnerUserId":7172549.0,"Title":"Theano\/lasagne: how to predict with trained model","Body":"<p>I am new to Theano and Lasagne. I followed <a href=\"https:\/\/lasagne.readthedocs.io\/en\/latest\/user\/tutorial.html\" rel=\"nofollow noreferrer\">the official doc<\/a> and can walk through the tutorial successfully. However, I have few questions.<\/p>\n\n<ol>\n<li><p>What's the format of the saved model? what is .npz file? How to interpret it?<\/p>\n\n<p>np.savez('model.npz', *lasagne.layers.get_all_param_values(network))<\/p><\/li>\n<li><p>I found the following code snippet to reload the neural network model,but have no idea how to use this model to predict new sample? for example, I have a new image with size 28*28. How can I predict the digit number in this image?<\/p>\n\n<p>with np.load('model.npz') as f:\n         param_values = [f['arr_%d' % i] for i in range(len(f.files))]\nlasagne.layers.set_all_param_values(network, param_values)<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"41296057","CreationDate":"2016-12-23T06:09:51.320","ParentId":null,"OwnerUserId":"7172549","Title":null,"Body":"<p>I have figured out the answer, and paste my script for your reference.<\/p>\n\n<pre><code>from __future__ import print_function\n\nimport sys\nimport os\nimport time\nimport pickle\n\nimport numpy as np\nimport theano\nimport theano.tensor as T\n\nimport lasagne\n\ndef load_dataset(sample_data, sample_label):\n    if sys.version_info[0] == 2:\n        from urllib import urlretrieve\n    else:\n        from urllib.request import urlretrieve\n\ndef download(filename, source='http:\/\/yann.lecun.com\/exdb\/mnist'):\n    print(\"Downloading %s\" % filename)\n    urlretrieve(source + filename, filename)\n\nimport gzip\n\n# ##################### Build the neural network model                     #######################\n# This script supports three types of models. For each one, we define a\n# function that takes a Theano variable representing the input and returns\n# the output layer of a neural network model built in Lasagne.\n\ndef build_mlp(input_var=None):\n    # This creates an MLP of two hidden layers of 800 units each, followed by\n    # a softmax output layer of 10 units. It applies 20% dropout to the input\n    # data and 50% dropout to the hidden layers.\n\n    # Input layer, specifying the expected input shape of the network\n    # (unspecified batchsize, 1 channel, 28 rows and 28 columns) and\n    # linking it to the given Theano variable `input_var`, if any:\n    l_in = lasagne.layers.InputLayer(shape=(None, 1, 28, 28),\n                                 input_var=input_var)\n\n    # Apply 20% dropout to the input data:\n    l_in_drop = lasagne.layers.DropoutLayer(l_in, p=0.2)\n\n    # Add a fully-connected layer of 800 units, using the linear rectifier, and\n    # initializing weights with Glorot's scheme (which is the default anyway):\n    l_hid1 = lasagne.layers.DenseLayer(\n        l_in_drop, num_units=800,\n        nonlinearity=lasagne.nonlinearities.rectify,\n        W=lasagne.init.GlorotUniform())\n\n    # We'll now add dropout of 50%:\n    l_hid1_drop = lasagne.layers.DropoutLayer(l_hid1, p=0.5)\n\n    # Another 800-unit layer:\n    l_hid2 = lasagne.layers.DenseLayer(\n        l_hid1_drop, num_units=800,\n        nonlinearity=lasagne.nonlinearities.rectify)\n\n    # 50% dropout again:\n    l_hid2_drop = lasagne.layers.DropoutLayer(l_hid2, p=0.5)\n\n    # Finally, we'll add the fully-connected output layer, of 10 softmax units:\n    l_out = lasagne.layers.DenseLayer(\n        l_hid2_drop, num_units=10,\n        nonlinearity=lasagne.nonlinearities.softmax)\n\n    # Each layer is linked to its incoming layer(s), so we only need to pass\n    # the output layer to give access to a network in Lasagne:\n    return l_out\n\n\ndef build_custom_mlp(input_var=None, depth=2, width=800, drop_input=.2,\n                 drop_hidden=.5):\n# By default, this creates the same network as `build_mlp`, but it can be\n# customized with respect to the number and size of hidden layers. This\n# mostly showcases how creating a network in Python code can be a lot more\n# flexible than a configuration file. Note that to make the code easier,\n# all the layers are just called `network` -- there is no need to give them\n# different names if all we return is the last one we created anyway; we\n# just used different names above for clarity.\n\n# Input layer and dropout (with shortcut `dropout` for `DropoutLayer`):\n    network = lasagne.layers.InputLayer(shape=(None, 1, 28, 28),\n                                    input_var=input_var)\n    if drop_input:\n        network = lasagne.layers.dropout(network, p=drop_input)\n    # Hidden layers and dropout:\n    nonlin = lasagne.nonlinearities.rectify\n    for _ in range(depth):\n        network = lasagne.layers.DenseLayer(\n            network, width, nonlinearity=nonlin)\n    if drop_hidden:\n        network = lasagne.layers.dropout(network, p=drop_hidden)\n    # Output layer:\n    softmax = lasagne.nonlinearities.softmax\n    network = lasagne.layers.DenseLayer(network, 10, nonlinearity=softmax)\n    return network\n\n\ndef build_cnn(input_var=None):\n# As a third model, we'll create a CNN of two convolution + pooling stages\n# and a fully-connected hidden layer in front of the output layer.\n\n    # Input layer, as usual:\n    network = lasagne.layers.InputLayer(shape=(None, 1, 28, 28),\n                                    input_var=input_var)\n# This time we do not apply input dropout, as it tends to work less well\n# for convolutional layers.\n\n# Convolutional layer with 32 kernels of size 5x5. Strided and padded\n# convolutions are supported as well; see the docstring.\n    network = lasagne.layers.Conv2DLayer(\n        network, num_filters=32, filter_size=(5, 5),\n        nonlinearity=lasagne.nonlinearities.rectify,\n        W=lasagne.init.GlorotUniform())\n# Expert note: Lasagne provides alternative convolutional layers that\n# override Theano's choice of which implementation to use; for details\n# please see http:\/\/lasagne.readthedocs.org\/en\/latest\/user\/tutorial.html.\n\n# Max-pooling layer of factor 2 in both dimensions:\n    network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2))\n\n# Another convolution with 32 5x5 kernels, and another 2x2 pooling:\n    network = lasagne.layers.Conv2DLayer(\n        network, num_filters=32, filter_size=(5, 5),\n        nonlinearity=lasagne.nonlinearities.rectify)\n    network = lasagne.layers.MaxPool2DLayer(network, pool_size=(2, 2))\n\n# A fully-connected layer of 256 units with 50% dropout on its inputs:\n    network = lasagne.layers.DenseLayer(\n        lasagne.layers.dropout(network, p=.5),\n        num_units=256,\n        nonlinearity=lasagne.nonlinearities.rectify)\n\n# And, finally, the 10-unit output layer with 50% dropout on its inputs:\n    network = lasagne.layers.DenseLayer(\n        lasagne.layers.dropout(network, p=.5),\n        num_units=10,\n        nonlinearity=lasagne.nonlinearities.softmax)\n\n    return network\ndef load_test_images(filename):\n    if not os.path.exists(filename):\n        download(filename)\n    with gzip.open(filename, 'rb') as f:\n        data = np.frombuffer(f.read(), np.uint8, offset=16)\n\n    data = data.reshape(-1,1,28,28)\n\n    return data \/ np.float32(256)\n\ndef load_test_labels(filename):\n    if not os.path.exists(filename):\n        download(filename)\n\n    with gzip.open(filename, 'rb') as f:\n                    data = np.frombuffer(f.read(), np.uint8, offset=8)\n    return data\n\nX_test = load_test_images(sample_data)\ny_test = load_test_labels(sample_label)\n\nreturn X_test, y_test\n\ndef predict_label(sample, model='model.npz'):\n    input_var = T.tensor4('sample')\n\n    network = build_mlp(input_var)\n\n    with np.load(model) as f:\n        param_values = [f['arr_%d'%i] for i in range(len(f.files))]\n\n    lasagne.layers.set_all_param_values(network, param_values)\n\n    prediction = lasagne.layers.get_output(network, deterministic=True)\n\n    result = T.argmax(prediction, axis=1)\n    predict_fn = theano.function([input_var],result)\n\n    return predict_fn(sample)\n\ndef main(model='mlp'):\n    # load the test dataset\n    print(\"Loading data...\")\n\n    sample_data = 't10k-images-idx3-ubyte.gz'\n    sample_label = 't10k-labels-idx1-ubyte.gz'\n    X_test, y_test = load_dataset(sample_data, sample_label)\n\n    # print(\"build model ...\")\n    # network = build_mlp() \n\n    print(\"Evaluating ...\")\n    label = predict_label(X_test)\n\n# with open(\"predict_label.txt\",\"w\") as text_file:\n#   print(label, file = text_file)\n\n# with open(\"groundtruth_label.txt\",\"w\") as text_file:\n    #        print(y_test, file = text_file)\n\n    print(\"Saving result ...\")\n    out_predict = open('predict.pkl', 'w')\n    pickle.dump(label, out_predict)\n\n    thefile = open('predict.txt','w')\n    for item in label:\n        thefile.write('%s\\n' % item)\n\n    out_groundtruth = open('groundtruth.pkl', 'w')\n    pickle.dump(y_test, out_groundtruth)\n\nif __name__ == '__main__':\n    print(\"This script will predict the lables based on the models leared via CNN, MLP\")\n    print(\"Reference: https:\/\/lasagne.readthedocs.io\/en\/latest\/user\/tutorial.html#run-the-mnist-example\")\n\n    kwargs = {}\n    if len(sys.argv) &gt; 1:\n        kwargs['model'] = sys.argv[1]\n\n    main(**kwargs)\n<\/code><\/pre>\n"}]}
{"QuestionId":41285090,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><word2vec>","CreationDate":"2016-12-22T14:11:34.083","AcceptedAnswerId":null,"OwnerUserId":6121755.0,"Title":"gen_word2vec in tensorflow is not found","Body":"<p>As I ran the code (<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/embedding\/word2vec.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/embedding\/word2vec.py<\/a>) in my laptop(Mac,python3), I received an error:        <\/p>\n\n<pre><\/pre>\n\n<p>tensorflow has been installed and working in my laptop. It seems like \"gen_word2vec\" is missing. Could someone help me? <\/p>\n","answers":[{"AnswerId":"41285613","CreationDate":"2016-12-22T14:39:48.553","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>Try installing the latest version of TensorFlow.<\/p>\n"}]}
{"QuestionId":41285116,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><deep-learning><caffe><imagenet>","CreationDate":"2016-12-22T14:12:46.487","AcceptedAnswerId":null,"OwnerUserId":5498819.0,"Title":"Change the output layer of AlexNet\/GoogleNet\/ImageNet?","Body":"<p>I have a question regarding changing the output layer of the nets (AlexNet\/GoogleNet\/ImageNet). So the standard output is a 1x1000 Vector so one value for each class.<\/p>\n\n<p>I know I can change the output to e.g 5 so I would get 1x5 Vector, if I only have 5 classes.<\/p>\n\n<p>But what is if I do not have classes? Is it possible to change the output to a matrix like 18x18. Because my net should output a density map and not a \"class\". \nAnd is it recommended to use a pre-trained net for my task or should I train from scratch? <\/p>\n\n<p>Thank you for your help :-)<\/p>\n","answers":[{"AnswerId":"41286775","CreationDate":"2016-12-22T15:40:07.750","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<blockquote>\n  <p>But what is if i do not have classes?<\/p>\n<\/blockquote>\n\n<p>The concept of \"class\" is not really connected to the architecture but rather to the loss function itself. In other words if you have 1000 outputs it does not matter whether you want to classify among 1000 disjoint classes, assign 1000 tags, or regress on 1000 dimensional real output - architecture still makes perfect sense.<\/p>\n\n<blockquote>\n  <p>Is it possible to change the output to a matrix like 18x18<\/p>\n<\/blockquote>\n\n<p>The \"naive\" approach would be to output 18*18 = 324 values and treat it as a 2-dim matrix. However, the 2-dim structure suggests there is some characteristics that can be exploited on the architecture side, one typical characteristic is translational invariance, which is exploited in convnets, and if the same is true for your output you might consider deconvolution (of any sort, since there are many) for your model.<\/p>\n\n<blockquote>\n  <p>And is it recommended to use a pretrained net for my task? Or shoud i learn from scratch ?<\/p>\n<\/blockquote>\n\n<p>This does not depend on the architecture but <strong>task<\/strong>. If your task is similar enough to the ones a given net was trained on, you can use the pretrained one as a starting point and just \"fine-tune\" on the new one. In general, using pretrained net as a starting point is a safe thing to do (it should not be worse than training from scratch). Remember to train <strong>the whole network<\/strong>, and not just added parts though (unless you do not have enough data to train the whole structure). <\/p>\n"}]}
{"QuestionId":41285440,"AnswerCount":2,"Tags":"<python><optimization><tensorflow><deep-learning><jupyter-notebook>","CreationDate":"2016-12-22T14:30:27.253","AcceptedAnswerId":null,"OwnerUserId":5711134.0,"Title":"Can't find \"gen_training_ops\" in the tensorflow GitHub","Body":"<p>I'm working on a new optimizer, and I managed to work out most of the process. Only thing I'm stuck on currently is finding .<\/p>\n\n<p>Apparently this file is crucial, because in both implementations of Gradient Descent, and Adagrad optimizers they use functions that are imported out of a wrapper file for  ( in the  folder).\nI can't find this file anywhere, so I suppose I don't understand something and search in the wrong place. Where can I find it? (Or specifically the implementations of  and )<\/p>\n\n<p>Thanks a lot :)<\/p>\n","answers":[{"AnswerId":"54832398","CreationDate":"2019-02-22T17:38:05.150","ParentId":null,"OwnerUserId":"10038330","Title":null,"Body":"<p>If you find it, youll realize it just jumps to pyhon\/framework, where the actual update is just an assign operation and then gets grouped <\/p>\n"},{"AnswerId":"41285726","CreationDate":"2016-12-22T14:45:52.010","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>The file will be generated when you build TensorFlow from source. It is marked as <code>out<\/code> in this file,<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/BUILD#L912\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/BUILD#L912<\/a><\/p>\n"}]}
{"QuestionId":41287289,"AnswerCount":3,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2016-12-22T16:06:50.177","AcceptedAnswerId":null,"OwnerUserId":1158977.0,"Title":"Restoring saved TensorFlow model to evaluate on test set","Body":"<p>I have seen a few <a href=\"https:\/\/stackoverflow.com\/questions\/34982492\/restoring-tensorflow-model\">posts<\/a> on restoring  models and the  doc page on <a href=\"https:\/\/www.tensorflow.org\/how_tos\/meta_graph\/\" rel=\"nofollow noreferrer\">exporting graphs<\/a> but I think I am missing something.<\/p>\n\n<p>I use the code in this <a href=\"https:\/\/gist.github.com\/autojazari\/13e81462cc736c5772277f26249e61f1\" rel=\"nofollow noreferrer\">Gist<\/a> to save the model along with this utils file to which <a href=\"https:\/\/gist.github.com\/autojazari\/8b282d58a38ff6f920e413d82bbefcbd\" rel=\"nofollow noreferrer\">defines<\/a> the model<\/p>\n\n<p>Now I would like to restore it and run in a previously unseen test data as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting a test accuracy of only <em>3%<\/em>.  However If I don't close the Notebook and run the test code immediately after training the model, I get a <em>95%<\/em> accuracy.  <\/p>\n\n<p>This leads me to believe I'm not loading the model correctly?<\/p>\n","answers":[{"AnswerId":"41304642","CreationDate":"2016-12-23T16:13:45.523","ParentId":null,"OwnerUserId":"1158977","Title":null,"Body":"<p>The answer found <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variables\/\" rel=\"nofollow noreferrer\">here<\/a> is what ended up working as follows:<\/p>\n\n<pre><code>save_path = saver.save(sess, '\/home\/ubuntu\/gtsd-12-23-16.chkpt')\nprint(\"Model saved in file: %s\" % save_path)\n## later re-run code that creates the model\n# Image Tensor\nimages_placeholder = tf.placeholder(tf.float32, shape=[None, 32, 32, 3], name='x')\n\ngray = tf.image.rgb_to_grayscale(images_placeholder, name='gray')\n\ngray \/= 255.\n\n# Label Tensor\nlabels_placeholder = tf.placeholder(tf.float32, shape=(None, 43), name='y')\n\n# dropout Tensor\nkeep_prob = tf.placeholder(tf.float32, name='drop')\n\n# construct model\nlogits = inference(gray, keep_prob)\n\n# calculate loss\nloss_value = loss(logits, labels_placeholder)\n\n# training\ntrain_op = training(loss_value, 0.001)\n\n# accuracy\nacc = accuracy(logits, labels_placeholder)\n\nsaver = tf.train.Saver()\n    with tf.Session() as sess:\n        saver.restore(sess, '\/home\/ubuntu\/gtsd-12-23-16.chkpt')\n        print(\"Model restored.\")\n        test_accuracy = evaluate(X_test, y_test)\n        print(\"Test Accuracy = {:.3f}\".format(test_accuracy[0]*100))\n<\/code><\/pre>\n"},{"AnswerId":"41287885","CreationDate":"2016-12-22T16:39:55.200","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem arises from these two lines:<\/p>\n\n<pre><code>loader.restore(sess, tf.train.latest_checkpoint('.\/'))\nsess.run(tf.initialize_all_variables())   \n<\/code><\/pre>\n\n<p>The first line loads the saved model from a checkpoint. The second line <strong>re-initializes<\/strong> all of the variables in the model (such as the weight matrices, convolutional filters, and bias vectors), usually to random numbers, and <strong>overwrites<\/strong> the loaded values.<\/p>\n\n<p>The solution is simple: delete the second line (<code>sess.run(tf.initialize_all_variables())<\/code>) and evaluation will proceed with the trained values loaded from the checkpoint.<\/p>\n\n<hr>\n\n<p>PS. There is a small chance that this change will give you an error about \"uninitialized variables\". In that case, you should execute <code>sess.run(tf.initialize_all_variables())<\/code> to initialize any variables not saved in the checkpoint <strong>before<\/strong> executing <code>loader.restore(sess, tf.train.latest_checkpoint('.\/'))<\/code>.<\/p>\n"},{"AnswerId":"51038448","CreationDate":"2018-06-26T08:37:22.723","ParentId":null,"OwnerUserId":"9926051","Title":null,"Body":"<p>I had a similar problem and for me this worked:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    saver=tf.train.Saver(tf.all_variables())\n    saver=tf.train.import_meta_graph('model.meta')\n    saver.restore(sess,\"model\")\n\n    test_accuracy = evaluate(X_test, y_test)\n<\/code><\/pre>\n"}]}
{"QuestionId":41287867,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2016-12-22T16:38:56.440","AcceptedAnswerId":null,"OwnerUserId":3102085.0,"Title":"How to create a custom pooling function in Tensorflow","Body":"<p>Currently tensor flow provides built-in function to max pooling. I want to customize it. Instead of choosing the max value from the feature vector, I want to get the value by adding all the values of feature vector. For this matter I have to create a custom function. I have come across a solution that you have to define the op in C++ and then use it in python. But is there an easier way to implement it directly in python? And I know this question was asked before on Stack overflow but I couldn't understand the solution to it.<\/p>\n","answers":[]}
{"QuestionId":41288600,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-22T17:19:15.907","AcceptedAnswerId":41290182.0,"OwnerUserId":3102085.0,"Title":"How to extract feature vectors at every convolutional layer?","Body":"<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>I want to extract feature vectors after the convolution layer to pass to my custom function. I am very new to Tensorflow so I don't understand how to do it. <\/p>\n","answers":[{"AnswerId":"41290182","CreationDate":"2016-12-22T19:11:32.230","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p>If your custom function is part of the tensorflow graph, you should be able to refer directly to the tensors you have created. e.g. if you want the activations of <code>conv1_1<\/code>, use the tensor stored in <code>self.conv1_1<\/code>.<\/p>\n\n<p>If your custom function needs the raw numbers, after processing, you can pass the same tensor references into <code>session.run<\/code>, e.g.<\/p>\n\n<pre><code>sess = tf.Session()\nsess.run([self.conv1_1], ...)\n<\/code><\/pre>\n"}]}
{"QuestionId":41288713,"AnswerCount":2,"Tags":"<python><variables><tensorflow><variable-assignment><timing>","CreationDate":"2016-12-22T17:26:12.807","AcceptedAnswerId":41289120.0,"OwnerUserId":7331377.0,"Title":"Tensorflow: When are variable assignments done in sess.run with a list?","Body":"<p>I have thought that variable assignments are done after all operations in a list given to sess.run, but the following code returns different results at different execution. It seems randomly run operations in the list and assign the variable after the run of the operation in the list.<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to know the timing of variable assignments.\nWhich are correct: \"update_x -> assign x -> ... -> udpate_z -> assign z\" or \"update_x -> udpate_y -> udpate_z -> assign a, b, c\"? (where (x, y, z) is a permutation of (a, b, c))\nIn addition, if there is a way that realize the latter assignment (assignment are done after all operations in the list are done), please let me know how to realize it.<\/p>\n","answers":[{"AnswerId":"41309537","CreationDate":"2016-12-24T01:24:44.453","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>Based on this example of yours',<\/p>\n\n<pre><code>v = tf.Variable(0)\nc = tf.constant(3)\nadd = tf.add(v, c)\nupdate = tf.assign(v, add)\nmul = tf.mul(add, update)\n\nwith tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n    res = sess.run([mul, mul])\n    print(res)\n<\/code><\/pre>\n\n<p><strong><em>Output:  [9, 9]<\/em><\/strong><\/p>\n\n<p>You get <code>[9, 9]<\/code> and this is in fact what we've asked it to do. Think of it like this:<\/p>\n\n<p>During the run, once <code>mul<\/code> is taken from the list, it looks for the definition of this and finds <code>tf.mul(add, update)<\/code>. Now, it needs the value of <code>add<\/code> which leads to <code>tf.add(v, c)<\/code>. So, it plugs in the value of <code>v<\/code> and <code>c<\/code>, gets the value of <code>add<\/code> as 3. <\/p>\n\n<p>Ok, now we need the value of <code>update<\/code> which is defined as <code>tf.assign(v, add)<\/code>. We have values of both <code>add<\/code> (which it computed just now as 3) &amp; <code>v<\/code>. So, it updates the value of <code>v<\/code> to be 3 which is also the value for <code>update<\/code>.<\/p>\n\n<p>Now, it has values for both <code>add<\/code> and <code>update<\/code> which are 3. Thus, the multiplication yields 9 in <code>mul<\/code>.<\/p>\n\n<p>Based on the result that we get, I think, for the next item(operation) in the list, it just returns the just computed value of <code>mul<\/code>. I'm not sure whether it does the steps again or just returns the same (cached?) value it just computed for <code>mul<\/code> realizing that we've the result or these operations happen in parallel(for each element in the list). Maybe @mrry or @YaroslavBulatov can comment on this part please?<\/p>\n\n<hr>\n\n<p>Quoting @mrry's comment:<\/p>\n\n<blockquote>\n  <p>When you call <code>sess.run([x, y, z])<\/code> once, TensorFlow executes each op that those tensors depend on <strong>one time only<\/strong> (unless there's a <code>tf.while_loop()<\/code> in your graph). If a tensor appears twice in the list (like <code>mul<\/code> in your example), TensorFlow will execute it once and return <em>two copies<\/em> of the result. To run the assignment more than once, you must either call <code>sess.run()<\/code> multiple times, or use <code>tf.while_loop()<\/code> to put a loop in your graph.<\/p>\n<\/blockquote>\n"},{"AnswerId":"41289120","CreationDate":"2016-12-22T17:51:21.797","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The three operations <code>update_a<\/code>, <code>update_b<\/code>, and <code>update_c<\/code> have no interdependencies in the dataflow graph, so TensorFlow may choose to execute them in any order. (In the current implementation, it is possible that all three of them will be executed in parallel on different threads.) A second nit is that reads of variables are cached by default, so in your program the value assigned in <code>update_b<\/code> (i.e. <code>c + a<\/code>) may use the original or the updated value of <code>a<\/code>, depending on when the variable is first read.<\/p>\n\n<p>If you want to ensure that the operations happen in a particular order, you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/utility_functions#control_dependencies\" rel=\"noreferrer\"><code>with tf.control_dependencies([...]):<\/code><\/a> blocks to enforce that operations created within the block happen after operations named in the list.  You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/variables#Variable.read_value\" rel=\"noreferrer\"><code>tf.Variable.read_value()<\/code><\/a> inside a <code>with tf.control_dependencies([...]):<\/code> block to make the point at which the variable is read explicit.<\/p>\n\n<p>Therefore, to if you want to ensure that <code>update_a<\/code> happens before <code>update_b<\/code> and <code>update_b<\/code> happens before <code>update_c<\/code>, you could do:<\/p>\n\n<pre><code>update_a = tf.assign(a, b + c)\n\nwith tf.control_dependencies([update_a]):\n  update_b = tf.assign(b, c + a.read_value())\n\nwith tf.control_dependencies([update_b]):\n  update_c = tf.assign(c, a.read_value() + b.read_value())\n<\/code><\/pre>\n"}]}
{"QuestionId":41289182,"AnswerCount":0,"Tags":"<tensorflow><gradient><accumulate>","CreationDate":"2016-12-22T17:56:19.037","AcceptedAnswerId":null,"OwnerUserId":4834515.0,"Title":"How to accumulate the gradients and apply them after several iterations in Tensorflow","Body":"<p>I used Caffe for some time but now am using Tensorflow. Caffe has a hyperparameter '<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L154\" rel=\"nofollow noreferrer\">iter_size<\/a>', which accumulates gradients over  x  instances.  is used when GPU memory is limited and there is not enough GPUs. <\/p>\n\n<p>I am wondering whether we can do the same operation in Tensorflow. I have seen this <a href=\"https:\/\/stackoverflow.com\/questions\/37710974\/how-to-accumulate-and-appy-gradients-for-async-n-step-dqnetwork-update-in-tensor#comment69657445_37710974\">question<\/a>. It accumulates the gradients but it does not reset the accumulated gradients to zero after applying gradients on variables.<\/p>\n","answers":[]}
{"QuestionId":41289882,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-22T18:50:01.263","AcceptedAnswerId":41290542.0,"OwnerUserId":7331562.0,"Title":"Do input & output tensors to TensorFlow's OpKernel::Compute() function change their address across multiple function calls?","Body":"<p>I am working on supporting TensorFlow on a new architecture.<\/p>\n\n<p>Consider the following TensorFlow code:<\/p>\n\n<pre><\/pre>\n\n<p>There are two matmul blocks present in the execution graph of the TensorFlow program, and during execution they are represented using two instantiations of the MatMulOp subclass of OpKernel, found in tensorflow\/core\/kernels\/matmul_op.cc.  The first thing done in MatMulOp::Compute() is to grab the addresses of the input tensors:<\/p>\n\n<pre><\/pre>\n\n<p>My understanding of TensorFlow is that during each iteration of sess.run() above, the two instantiations of MatMulOp do not change.  For each MatMul block, can I expect the addresses of the inputs to remain constant over the iterations, or could it be that in the seventh iteration of the sess.run() call, ctx->input(0) will have a different value than it did in the sixth?<\/p>\n\n<p>The Compute() method also invokes ctx->allocate_output(), which eventually wraps our own architecture's allocator.  Is it OK to allocate the output block once and then just keep using the same block over future runs in the same session?<\/p>\n","answers":[{"AnswerId":"41290542","CreationDate":"2016-12-22T19:38:55.733","ParentId":null,"OwnerUserId":"5548115","Title":null,"Body":"<p>Thanks for the question!<\/p>\n\n<p>TL;DR: yes, input and output tensors will change their address across different invocations of the Run() call.<\/p>\n\n<p>Details:<\/p>\n\n<p>If there are two different MatMul ops in the graph, there will be two instantiations of the MatMulOp OpKernel (held cached in an OpSegment, which is looked up \/ created for an Op in the graph and subsequeentally cached for every future call to Session::Run()).<\/p>\n\n<p>The addresses of the inputs are not guaranteed to remain constant over iterations.  In fact, it is almost certain that this is not the case, partially because execution of the dataflow graph is dynamic, not deterministic.<\/p>\n\n<p>ctx->input(0) is memory allocated likely from another call from the input op's \"allocate_output\" function, so the answer for both the input and the output memory locations is the same.  The allocate_output function eventually delegates to the device's \"Allocator\" implementation to allocate the memory of the appropriate size (and alignment).  On CPU, the current implementation delegates to malloc(), so just like malloc(), you may get different memory on every call to Run().  On GPU, we use a custom GPU allocator (the BFCAllocator class) to dynamically allocate memory, and its properties are similar to malloc() in that different memory may be returned based on the allocation\/release order of memory through the Allocator.<\/p>\n\n<p>So in general, the memory returned by the call to allocate_output() is handled by the Allocator implementations for a device.  The CPU and GPU implementations do not provide a stable pointer guarantee across Runs of the same graph.  <\/p>\n\n<p>However, if you are implementing a custom device, you likely have to implement a custom allocator for your device, and it <em>may<\/em> be possible for you to write the allocator in such a way that it returns the same memory for the same output from a graph.  But that would require figuring out how to pass the identifier of the op to the Allocator so that you could return the same memory each time.<\/p>\n\n<p>TensorFlow intentionally does dynamic memory allocation for at least a few reasons:<\/p>\n\n<p>1) The execution order of a dataflow graph may depend on external inputs, so customizing a strict schedule may lead to unnecessary stalls.  Dynamic execution order ensures that ops only execute when all of their inputs are ready.<\/p>\n\n<p>2) Shapes can be dynamic in TensorFlow (e.g., your graph can handle variable batch sizes), which means that an op may need to allocate different amount of memory for the same op's output from Run() to Run()!  This is a big reason why we cannot and do not provide such guarantees.<\/p>\n\n<p>We understand there are cases where a device wants to optimize for one instantiation of a graph (with fixed sizes), so the device can pre-plan for the entire dataflow graph and amortize its execution over several attempts.  These cases are generally better suited by the XLA compiler framework  (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/resources\/xla_prerelease\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/resources\/xla_prerelease<\/a>), but it may be possible to get it to work for limited situations \/ graphs without XLA using the existing device framework.<\/p>\n"}]}
{"QuestionId":41290509,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><caffe>","CreationDate":"2016-12-22T19:36:44.710","AcceptedAnswerId":41292222.0,"OwnerUserId":2475611.0,"Title":"Need Slice layer of Caffe in Tensorflow","Body":"<p>I defined the following slice layer, where I want to split Nx7 into Nx3 and Nx4 two blobs<\/p>\n\n<pre><\/pre>\n\n<p>How can I do the same thing in Tensorflow? <\/p>\n","answers":[{"AnswerId":"41292222","CreationDate":"2016-12-22T21:56:17.300","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>I believe that you can do this with a hand-crafted layer, using the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops\/slicing_and_joining#split\" rel=\"nofollow noreferrer\">split method<\/a>.  You can also do it with separate calls to <strong>slice<\/strong>.  Something like<\/p>\n\n<pre><code>slice3, slice4 = tf.split(input_tensor, [3, 4], 1)\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>Correction per OP comment:<\/strong><\/p>\n\n<pre><code>slice3, slice4 = tf.split_v(labels, [3, 4], 1)\n<\/code><\/pre>\n\n<p>BTW, the spacing follows PEP8 standards.  It's not required, of course, but there is checking software that prefers the spacing this way.<\/p>\n"}]}
{"QuestionId":41290674,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-22T19:50:09.823","AcceptedAnswerId":null,"OwnerUserId":3102085.0,"Title":"TypeError(\"Using a `tf.Tensor` as a Python `bool` is not allowed. \"","Body":"<p>My main task here is to sum the values of tensor generated by conv1_2\nFollowing is my code<\/p>\n\n<pre><\/pre>\n\n<p>When I run this code it gives me the following error<\/p>\n\n<pre><\/pre>\n\n<p>And this error is coming from the line where I am calling the function add_n()\nI can't figure this out. Please tell me what I am doing wrong. <\/p>\n\n<p>Traceback:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41291708","CreationDate":"2016-12-22T21:11:22.230","ParentId":null,"OwnerUserId":"6850023","Title":null,"Body":"<p>The function <em>tf.add_n<\/em> is an element-wise addition, and expects a list of tensors, not a single tensor. I'm not sure what you want to accomplish, but maybe something like <em>tf.reduce_sum<\/em> would be more suitable?<\/p>\n\n<p>A trivial example that triggers the issue:<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensor flow as tf\n&gt;&gt;&gt; x = tf.zeros([10,10])\n&gt;&gt;&gt; y = tf.add_n(x)\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\n  File \"\/usr\/local\/lib\/python3.4\/site-packages\/tensorflow\/python\/ops\/math_ops.py\", line 1826, in add_n\n    if not inputs or not isinstance(inputs, (list, tuple)):\n  File \"\/usr\/local\/lib\/python3.4\/site-packages\/tensorflow\/python\/framework\/ops.py\", line 533, in __bool__\n    raise TypeError(\"Using a `tf.Tensor` as a Python `bool` is not allowed. \"\nTypeError: Using a `tf.Tensor` as a Python `bool` is not allowed. Use `if t is not None:` instead of `if t:` to test if a tensor is defined, and use TensorFlow ops such as tf.cond to execute subgraphs conditioned on the value of a tensor.\n<\/code><\/pre>\n\n<p>Whereas<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; x = tf.zeros([10,10])\n&gt;&gt;&gt; y = tf.add_n([x])\n<\/code><\/pre>\n\n<p>works fine (but doesn't really accomplish anything).<\/p>\n"}]}
{"QuestionId":41291727,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-12-22T21:13:11.593","AcceptedAnswerId":null,"OwnerUserId":7331562.0,"Title":"Will TensorFlow XLA deal with sparse tensors?","Body":"<p>I've read the XLA prerelease document here.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/master\/resources\/xla_prerelease#xla_accelerated_linear_algebra\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/resources\/xla_prerelease#xla_accelerated_linear_algebra<\/a><\/p>\n\n<p>It discusses datatypes of elements, but does not go into much detail about the data organization of the tensors themselves.  How will operations on SparseTensor objects be handled once XLA is available?<\/p>\n","answers":[{"AnswerId":"41645663","CreationDate":"2017-01-14T00:57:13.940","ParentId":null,"OwnerUserId":"822360","Title":null,"Body":"<p>Sparse data is something we'd like to have working, though it has some challenges. E.g. currently XLA depends on knowing the exact size of every buffer statically. We could certainly find a way to deal with that, but have been focusing on dense data so far.<\/p>\n"},{"AnswerId":"41351338","CreationDate":"2016-12-27T19:49:47.600","ParentId":null,"OwnerUserId":"1884421","Title":null,"Body":"<p>No, XLA focuses on dense tensors and doesn't deal with sparse tensors in an efficient way <strong>today<\/strong>. <\/p>\n\n<p>It could be easily extended to allow users to express some sparsity using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/resources\/xla_prerelease#layout\" rel=\"nofollow noreferrer\">layouts<\/a> (e.g. interior padding). <\/p>\n"},{"AnswerId":"41368414","CreationDate":"2016-12-28T19:19:43.003","ParentId":null,"OwnerUserId":"5823391","Title":null,"Body":"<p>The layouts restrict the data organization of input and output tensors and don't include sparse layouts, although as Jingyue suggests, they could be extended in the future. The internal representation of tensors in the AST can in principle be anything a backend wants, and it is expected that the compiler may reorganize the data to different layouts for the convenience of different operators implemented by different backends.<\/p>\n\n<p>I am not aware that anyone has put much thought into how to do this efficiently for sparse tensors. In principle maybe it could be done as a compiler pass to infer sparsity and propagate it, with sparse implementations for all the relevant operators. Nothing like that exists today.<\/p>\n"}]}
{"QuestionId":41291937,"AnswerCount":1,"Tags":"<keras><updatebatchsize>","CreationDate":"2016-12-22T21:31:37.210","AcceptedAnswerId":null,"OwnerUserId":7332326.0,"Title":"How to change batch size of an intermediate layer in Keras?","Body":"<p>My problem is to take all hidden outputs from an LSTM and use them as training examples for a single dense layer. Flattening the output of the hidden layers and feeding them to a dense layer is not what I am looking to do. I have tried the following things:<\/p>\n\n<ol>\n<li><p>I have considered Timedistributed wrapper for the dense layer (<a href=\"https:\/\/keras.io\/layers\/wrappers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/wrappers\/<\/a>). But, this seems to apply the same layer to every time slice, which is not what I want. In other words, the Timedistributed wrapper has input_shape of a 3D tensor (number of samples, number of timesteps, number of features) and produces another 3D tensor of the same type: (number of samples, number of timesteps, number of features). Instead what I want is a 2D tensor as output, which looks like (number of samples*number of timesteps, number of features)<\/p><\/li>\n<li><p>There was a pull request for an AdvancedReshapeLayer: <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/36\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/pull\/36<\/a> on GitHub. This seems to be exactly what I am looking for. Unfortunately, it appears like that pull request was closed with no conclusive outcome. <\/p><\/li>\n<li><p>I tried to build my own lambda layer to accomplish what I want as follows:\n<\/p><\/li>\n<\/ol>\n\n<p>mode.output_shape after (A) prints: (BATCH_SIZE, NUM_TIME_STEPS, NUM_LSTM_UNITS) and model.output_shape after (B) prints: (BATCH_SIZE*NUM_OF_TIMESTEPS, NUM_LSTM_UNITS)<\/p>\n\n<p>Which is exactly what I am trying to achieve.<\/p>\n\n<p>Unfortunately, when I try to run step (C). I get the following error: <\/p>\n\n<blockquote>\n  <p>Input 0 is incompatible with layer dense_1: expected ndim=2, found\n  ndim=3<\/p>\n<\/blockquote>\n\n<p>This is baffling since when I print model.output_shape after (B), I do indeed see (BATCH_SIZE*NUM_OF_TIMESTEPS, NUM_LSTM_UNITS), which is of ndim=2. <\/p>\n\n<p>Really appreciate any help with this. <\/p>\n\n<p>EDIT: When I try to use the functional API instead of a sequential model, I still get the same error on step (C)<\/p>\n","answers":[{"AnswerId":"41413372","CreationDate":"2017-01-01T04:33:06.550","ParentId":null,"OwnerUserId":"3473614","Title":null,"Body":"<p>You can use backend reshape which includes batch_size dimension.<\/p>\n\n<pre><code>def backend_reshape(x):\n    return backend.reshape(x, (-1, NUM_LSTM_UNITS))\n\nmodel.add(Lambda(backend_reshape, output_shape=(NUM_LSTM_UNITS,)))\n<\/code><\/pre>\n\n\n"}]}
{"QuestionId":41292038,"AnswerCount":1,"Tags":"<keras><regularized>","CreationDate":"2016-12-22T21:39:21.543","AcceptedAnswerId":41299554.0,"OwnerUserId":4313064.0,"Title":"In Keras what is the difference between weight_regularizer and activity_ regularizer","Body":"<p>I understand regularization normally adds k*w^2 to the loss to penalize large weights. But in Keras there are two regularizer parameters - weight_regularizer and activity_ regularizer. What is the difference?<\/p>\n","answers":[{"AnswerId":"41299554","CreationDate":"2016-12-23T10:22:00.957","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>The difference is that <code>activity_regularizer<\/code> is applied to the output from an intermediate layer, it penalizes large layer output. <\/p>\n"}]}
{"QuestionId":41293077,"AnswerCount":12,"Tags":"<tensorflow><x86><compiler-optimization><simd><compiler-options>","CreationDate":"2016-12-22T23:21:14.987","AcceptedAnswerId":null,"OwnerUserId":6733064.0,"Title":"How to compile Tensorflow with SSE4.2 and AVX instructions?","Body":"<p>This is the message received from running a script to check if Tensorflow is working:<\/p>\n\n<pre><\/pre>\n\n<p>I noticed that it has mentioned SSE4.2 and AVX,<\/p>\n\n<ol>\n<li>What are SSE4.2 and AVX?<\/li>\n<li>How do these SSE4.2 and AVX improve CPU computations for Tensorflow tasks.<\/li>\n<li>How to make Tensorflow compile using the two libraries?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"59088497","CreationDate":"2019-11-28T12:04:44.037","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<p><strong>2.0 COMPATIBLE SOLUTION:<\/strong><\/p>\n\n<p>Execute the below commands in Terminal (Linux\/MacOS) or in Command Prompt (Windows) to install Tensorflow 2.0 using <strong>Bazel<\/strong>:<\/p>\n\n<pre><code>git clone https:\/\/github.com\/tensorflow\/tensorflow.git\ncd tensorflow\n\n#The repo defaults to the master development branch. You can also checkout a release branch to build:\ngit checkout r2.0\n\n#Configure the Build =&gt; Use the Below line for Windows Machine\npython .\/configure.py \n\n#Configure the Build =&gt; Use the Below line for Linux\/MacOS Machine\n.\/configure\n#This script prompts you for the location of TensorFlow dependencies and asks for additional build configuration options. \n\n#Build Tensorflow package\n\n#CPU support\nbazel build --config=opt \/\/tensorflow\/tools\/pip_package:build_pip_package \n\n#GPU support\nbazel build --config=opt --config=cuda --define=no_tensorflow_py_deps=true \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n"},{"AnswerId":"44509212","CreationDate":"2017-06-12T21:46:59.500","ParentId":null,"OwnerUserId":"7490399","Title":null,"Body":"<p>When building TensorFlow from source, you'll run the <code>configure<\/code> script. One of the questions that the <code>configure<\/code> script asks is as follows:<\/p>\n\n<pre><code>Please specify optimization flags to use during compilation when bazel option \"--config=opt\" is specified [Default is -march=native]\n<\/code><\/pre>\n\n<p>The <code>configure<\/code> script will attach the flag(s) you specify to the bazel command that builds the TensorFlow pip package. Broadly speaking, you can respond to this prompt in one of two ways:<\/p>\n\n<ul>\n<li>If you are building TensorFlow on the same type of CPU type as the one on which you'll run TensorFlow, then you should accept the default (<code>-march=native<\/code>). This option will optimize the generated code for your machine's CPU type.<\/li>\n<li>If you are building TensorFlow on one CPU type but will run TensorFlow on a different CPU type, then consider supplying a more specific optimization flag as described in <a href=\"https:\/\/gcc.gnu.org\/onlinedocs\/gcc-4.5.3\/gcc\/i386-and-x86_002d64-Options.html\" rel=\"nofollow noreferrer\">the gcc \ndocumentation<\/a>.<\/li>\n<\/ul>\n\n<p>After configuring TensorFlow as described in the preceding bulleted list, you should be able to build TensorFlow fully optimized for the target CPU just by adding the <code>--config=opt<\/code> flag to any bazel command you are running.<\/p>\n"},{"AnswerId":"45749548","CreationDate":"2017-08-18T06:04:43.113","ParentId":null,"OwnerUserId":"8476896","Title":null,"Body":"<p>I compiled a small Bash script for Mac (easily can be ported to Linux) to retrieve all CPU features and apply some of them to build TF. Im on TF master and use kinda often (couple times in a month).<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/venik\/9ba962c8b301b0e21f99884cbd35082f\" rel=\"noreferrer\">https:\/\/gist.github.com\/venik\/9ba962c8b301b0e21f99884cbd35082f<\/a><\/p>\n"},{"AnswerId":"45653743","CreationDate":"2017-08-12T18:44:53.030","ParentId":null,"OwnerUserId":"4012144","Title":null,"Body":"<p>To hide those warnings, you could do this before your actual code.<\/p>\n\n<pre><code>import os\nos.environ['TF_CPP_MIN_LOG_LEVEL']='2'\nimport tensorflow as tf\n<\/code><\/pre>\n"},{"AnswerId":"41584791","CreationDate":"2017-01-11T07:13:58.377","ParentId":null,"OwnerUserId":"7403118","Title":null,"Body":"<p>I just ran into this same problem, it seems like Yaroslav Bulatov's suggestion doesn't cover SSE4.2 support, adding <code>--copt=-msse4.2<\/code> would suffice. In the end, I successfully built with<\/p>\n\n<pre><code>bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=cuda -k \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>without getting any warning or errors.<\/p>\n\n<p>Probably the best choice for any system is:<\/p>\n\n<pre><code>bazel build -c opt --copt=-march=native --copt=-mfpmath=both --config=cuda -k \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>(<strong>Update: <a href=\"https:\/\/stackoverflow.com\/questions\/41293077\/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions#comment81255824_47146118\">the build scripts may be eating <code>-march=native<\/code><\/a>, possibly because it contains an <code>=<\/code><\/strong>.)<\/p>\n\n<p><code>-mfpmath=both<\/code> only works with gcc, not clang.  <code>-mfpmath=sse<\/code> is probably just as good, if not better, and is the default for x86-64.  32-bit builds default to <code>-mfpmath=387<\/code>, so changing that will help for 32-bit.  (But if you want high-performance for number crunching, you should build 64-bit binaries.)<\/p>\n\n<p>I'm not sure what TensorFlow's default for <code>-O2<\/code> or <code>-O3<\/code> is.  <code>gcc -O3<\/code> enables full optimization including auto-vectorization, but that sometimes can make code slower.<\/p>\n\n<hr>\n\n<p>What this does: <a href=\"https:\/\/docs.bazel.build\/versions\/master\/bazel-user-manual.html\" rel=\"noreferrer\"><code>--copt<\/code> for <code>bazel build<\/code><\/a> passes an option directly to gcc for compiling C and C++ files (but not linking, so you need a different option for cross-file link-time-optimization)<\/p>\n\n<p>x86-64 gcc defaults to using only SSE2 or older SIMD instructions, so you can run the binaries on <em>any<\/em> x86-64 system.  (See <a href=\"https:\/\/gcc.gnu.org\/onlinedocs\/gcc\/x86-Options.html\" rel=\"noreferrer\">https:\/\/gcc.gnu.org\/onlinedocs\/gcc\/x86-Options.html<\/a>).  That's not what you want.  You want to make a binary that takes advantage of all the instructions your CPU can run, because you're only running this binary on the system where you built it.<\/p>\n\n<p><code>-march=native<\/code> enables all the options your CPU supports, so it makes <code>-mavx512f -mavx2 -mavx -mfma -msse4.2<\/code> redundant.  (Also, <code>-mavx2<\/code> already enables <code>-mavx<\/code> and <code>-msse4.2<\/code>, so Yaroslav's command should have been fine).  Also if you're using a CPU that doesn't support one of these options (like FMA), using <code>-mfma<\/code> would make a binary that faults with illegal instructions.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r1.2\/install\/install_sources\" rel=\"noreferrer\">TensorFlow's <code>.\/configure<\/code> defaults to enabling <code>-march=native<\/code><\/a>, so using that should avoid needing to specify compiler options manually.<\/p>\n\n<p><code>-march=native<\/code> enables <code>-mtune=native<\/code>, so <a href=\"https:\/\/stackoverflow.com\/questions\/44490331\/how-does-mtune-actually-work\">it optimizes for your CPU<\/a> for things like which sequence of AVX instructions is best for unaligned loads.<\/p>\n\n<p>This all applies to gcc, clang, or ICC.  (For ICC, you can use <code>-xHOST<\/code> instead of <code>-march=native<\/code>.)<\/p>\n"},{"AnswerId":"46426188","CreationDate":"2017-09-26T12:10:36.580","ParentId":null,"OwnerUserId":"5774004","Title":null,"Body":"<p>This is the simplest method. Only one step.<\/p>\n\n<p>It has significant impact on speed. In my case, time taken for a training step almost halved.<\/p>\n\n<p>Refer\n<a href=\"https:\/\/github.com\/lakshayg\/tensorflow-build\" rel=\"noreferrer\">custom builds of tensorflow<\/a><\/p>\n"},{"AnswerId":"47146118","CreationDate":"2017-11-06T21:28:28.227","ParentId":null,"OwnerUserId":"5541673","Title":null,"Body":"<p>Thanks to all this replies + some trial and errors, I managed to install it on a Mac with <code>clang<\/code>. So just sharing my solution in case it is useful to someone. <\/p>\n\n<ol>\n<li><p>Follow the instructions on <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"noreferrer\">Documentation - Installing TensorFlow from Sources<\/a>  <\/p><\/li>\n<li><p>When prompted for <\/p>\n\n<p><em>Please specify optimization flags to use during compilation when bazel option \"--config=opt\" is specified [Default is -march=native]<\/em><\/p><\/li>\n<\/ol>\n\n<p>then copy-paste this string: <\/p>\n\n<pre><code>-mavx -mavx2 -mfma -msse4.2\n<\/code><\/pre>\n\n<p>(The default option caused errors, so did some of the other flags. I got no errors with the above flags. BTW I replied <code>n<\/code> to all the other questions)<\/p>\n\n<p>After installing, I verify a ~2x to 2.5x speedup when training deep models with respect to another installation based on the default wheels - <a href=\"https:\/\/www.tensorflow.org\/install\/install_mac\" rel=\"noreferrer\">Installing TensorFlow on macOS<\/a>   <\/p>\n\n<p>Hope it helps<\/p>\n"},{"AnswerId":"41387447","CreationDate":"2016-12-29T21:28:01.753","ParentId":null,"OwnerUserId":"190280","Title":null,"Body":"<p>These are <a href=\"https:\/\/en.wikipedia.org\/wiki\/SIMD\" rel=\"noreferrer\">SIMD<\/a> <a href=\"https:\/\/en.wikipedia.org\/wiki\/Vector_processor\" rel=\"noreferrer\">vector processing instruction sets<\/a>. <\/p>\n\n<p>Using vector instructions is faster for many tasks; machine learning is such a task.<\/p>\n\n<p>Quoting the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup\" rel=\"noreferrer\">tensorflow installation docs<\/a>:<\/p>\n\n<blockquote>\n  <p>To be compatible with as wide a range of machines as possible, TensorFlow defaults to only using SSE4.1 SIMD instructions on x86 machines. Most modern PCs and Macs support more advanced instructions, so if you're building a binary that you'll only be running on your own machine, you can enable these by using <code>--copt=-march=native<\/code> in your bazel build command.<\/p>\n<\/blockquote>\n"},{"AnswerId":"43107977","CreationDate":"2017-03-30T03:27:20.783","ParentId":null,"OwnerUserId":"5848524","Title":null,"Body":"<p><strong>Let me answer your 3rd question first:<\/strong><\/p>\n\n<p>If you want to run a self-compiled version within a conda-env, you can. These are the general instructions I run to get tensorflow to install on my system with additional instructions. Note: This build was for an AMD A10-7850 build (check your CPU for what instructions are supported...it may differ) running Ubuntu 16.04 LTS. I use Python 3.5 within my conda-env. Credit goes to the tensorflow source install page and the answers provided above.<\/p>\n\n<pre><code>git clone https:\/\/github.com\/tensorflow\/tensorflow \n# Install Bazel\n# https:\/\/bazel.build\/versions\/master\/docs\/install.html\nsudo apt-get install python3-numpy python3-dev python3-pip python3-wheel\n# Create your virtual env with conda.\nsource activate YOUR_ENV\npip install six numpy wheel, packaging, appdir\n# Follow the configure instructions at:\n# https:\/\/www.tensorflow.org\/install\/install_sources\n# Build your build like below. Note: Check what instructions your CPU \n# support. Also. If resources are limited consider adding the following \n# tag --local_resources 2048,.5,1.0 . This will limit how much ram many\n# local resources are used but will increase time to compile.\nbazel build -c opt --copt=-mavx --copt=-msse4.1 --copt=-msse4.2  -k \/\/tensorflow\/tools\/pip_package:build_pip_package\n# Create the wheel like so:\nbazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n# Inside your conda env:\npip install \/tmp\/tensorflow_pkg\/NAME_OF_WHEEL.whl\n# Then install the rest of your stack\npip install keras jupyter etc. etc.\n<\/code><\/pre>\n\n<p><strong>As to your 2nd question:<\/strong><\/p>\n\n<p>A self-compiled version with optimizations are well worth the effort in my opinion. On my particular setup, calculations that used to take 560-600 seconds now only take about 300 seconds! Although the exact numbers will vary, I think you can expect about a 35-50% speed increase in general on your particular setup.<\/p>\n\n<p><strong>Lastly your 1st question:<\/strong><\/p>\n\n<p>A lot of the answers have been provided above already. To summarize: <a href=\"https:\/\/en.wikipedia.org\/wiki\/Advanced_Vector_Extensions#AVX2\" rel=\"noreferrer\">AVX<\/a>, <a href=\"https:\/\/en.wikipedia.org\/wiki\/SSE4\" rel=\"noreferrer\">SSE4.1, SSE4.2<\/a>, MFA are different kinds of extended instruction sets on X86 CPUs. Many contain optimized instructions for processing matrix or vector operations. <\/p>\n\n<p>I will highlight my own misconception to hopefully save you some time: It's not that SSE4.2 is a newer version of instructions superseding SSE4.1. SSE4 = SSE4.1 (a set of 47 instructions) + SSE4.2 (a set of 7 instructions).<\/p>\n\n<p>In the context of tensorflow compilation, if you computer supports AVX2 and AVX, and SSE4.1 and SSE4.2, you should put those optimizing flags in for all. Don't do like I did and just go with SSE4.2 thinking that it's newer and should superseed SSE4.1. That's clearly WRONG! I had to recompile because of that which cost me a good 40 minutes.<\/p>\n"},{"AnswerId":"46022170","CreationDate":"2017-09-03T09:50:03.880","ParentId":null,"OwnerUserId":"914874","Title":null,"Body":"<p>I have recently installed it from source and bellow are all the steps needed to install it from source with the mentioned instructions available.<\/p>\n\n<p>Other answers already describe why those messages are shown. My answer gives a step-by-step on how to isnstall, which may help people struglling on the actual installation as I did.<\/p>\n\n<ol>\n<li>Install Bazel<\/li>\n<\/ol>\n\n<p>Download it from one of their available <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/releases\" rel=\"noreferrer\">releases<\/a>, for example <a href=\"https:\/\/github.com\/bazelbuild\/bazel\/releases\/download\/0.5.2\/bazel-0.5.2-dist.zip\" rel=\"noreferrer\">0.5.2<\/a>.\nExtract it, go into the directory and configure it: <code>bash .\/compile.sh<\/code>.\nCopy the executable to <code>\/usr\/local\/bin<\/code>: <code>sudo cp .\/output\/bazel \/usr\/local\/bin<\/code><\/p>\n\n<ol start=\"2\">\n<li>Install Tensorflow<\/li>\n<\/ol>\n\n<p>Clone tensorflow: <code>git clone https:\/\/github.com\/tensorflow\/tensorflow.git<\/code>\nGo to the cloned directory to configure it: <code>.\/configure<\/code><\/p>\n\n<p>It will prompt you with several questions, bellow I have suggested the response to each of the questions, you can, of course, choose your own responses upon as you prefer:<\/p>\n\n<pre><code>Using python library path: \/usr\/local\/lib\/python2.7\/dist-packages\nDo you wish to build TensorFlow with MKL support? [y\/N] y\nMKL support will be enabled for TensorFlow\nDo you wish to download MKL LIB from the web? [Y\/n] Y\nPlease specify optimization flags to use during compilation when bazel option \"--config=opt\" is specified [Default is -march=native]: \nDo you wish to use jemalloc as the malloc implementation? [Y\/n] n\njemalloc disabled\nDo you wish to build TensorFlow with Google Cloud Platform support? [y\/N] N\nNo Google Cloud Platform support will be enabled for TensorFlow\nDo you wish to build TensorFlow with Hadoop File System support? [y\/N] N\nNo Hadoop File System support will be enabled for TensorFlow\nDo you wish to build TensorFlow with the XLA just-in-time compiler (experimental)? [y\/N] N\nNo XLA JIT support will be enabled for TensorFlow\nDo you wish to build TensorFlow with VERBS support? [y\/N] N\nNo VERBS support will be enabled for TensorFlow\nDo you wish to build TensorFlow with OpenCL support? [y\/N] N\nNo OpenCL support will be enabled for TensorFlow\nDo you wish to build TensorFlow with CUDA support? [y\/N] N\nNo CUDA support will be enabled for TensorFlow\n<\/code><\/pre>\n\n<ol start=\"3\">\n<li>The pip package. To build it you have to describe which instructions you want (you know, those Tensorflow informed you are missing).<\/li>\n<\/ol>\n\n<p>Build pip script: <code>bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.1 --copt=-msse4.2 -k \/\/tensorflow\/tools\/pip_package:build_pip_package<\/code><\/p>\n\n<p>Build pip package: <code>bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg<\/code><\/p>\n\n<p>Install Tensorflow pip package you just built: <code>sudo pip install \/tmp\/tensorflow_pkg\/tensorflow-1.2.1-cp27-cp27mu-linux_x86_64.whl<\/code><\/p>\n\n<p>Now next time you start up Tensorflow it will not complain anymore about missing instructions.<\/p>\n"},{"AnswerId":"50886371","CreationDate":"2018-06-16T08:46:23.867","ParentId":null,"OwnerUserId":"8853057","Title":null,"Body":"<p>To compile TensorFlow with SSE4.2 and AVX, you can use directly<\/p>\n\n<blockquote>\n  <p>bazel build --config=mkl \n                  --config=\"opt\" \n                  --copt=\"-march=broadwell\" \n                  --copt=\"-O3\" \n                  \/\/tensorflow\/tools\/pip_package:build_pip_package<\/p>\n<\/blockquote>\n\n<p>Source: \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/docker\/Dockerfile.devel-cpu-mkl\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/docker\/Dockerfile.devel-cpu-mkl<\/a><\/p>\n"},{"AnswerId":"43555840","CreationDate":"2017-04-22T06:04:07.237","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Let's start with the explanation of <strong>why do you see these warnings in the first place<\/strong>. <\/p>\n\n<hr>\n\n<p>Most probably you have not installed TF from source and instead of it used something like <code>pip install tensorflow<\/code>. That means that you installed pre-built (by someone else) binaries which were not optimized for your architecture. And these warnings tell you exactly this: something is available on your architecture, but it will not be used because the binary was not compiled with it. Here is the part from <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide\" rel=\"noreferrer\">documentation<\/a>.<\/p>\n\n<blockquote>\n  <p>TensorFlow checks on startup whether it has been compiled with the\n  optimizations available on the CPU. If the optimizations are not\n  included, TensorFlow will emit warnings, e.g. AVX, AVX2, and FMA\n  instructions not included.<\/p>\n<\/blockquote>\n\n<p>Good thing is that <em>most probably you just want to learn\/experiment with TF so everything will work properly and you should not worry about it<\/em><\/p>\n\n<hr>\n\n<p><strong>What are SSE4.2 and AVX?<\/strong><\/p>\n\n<p>Wikipedia has a good explanation about <a href=\"https:\/\/en.wikipedia.org\/wiki\/SSE4\" rel=\"noreferrer\">SSE4.2<\/a> and <a href=\"https:\/\/en.wikipedia.org\/wiki\/Advanced_Vector_Extensions\" rel=\"noreferrer\">AVX<\/a>. This knowledge is not required to be good at machine-learning. You may think about them as a set of some additional <a href=\"https:\/\/en.wikipedia.org\/wiki\/X86_instruction_listings\" rel=\"noreferrer\">instructions<\/a> for a computer to use multiple data points against a single instruction to perform operations which may be naturally parallelized (for example adding two arrays). <\/p>\n\n<p>Both SSE and AVX are implementation of an abstract idea of <a href=\"https:\/\/en.wikipedia.org\/wiki\/SIMD\" rel=\"noreferrer\">SIMD<\/a> (Single instruction, multiple data), which is<\/p>\n\n<blockquote>\n  <p>a class of parallel computers in Flynn's taxonomy. It describes\n  computers with multiple processing elements that perform the same\n  operation on multiple data points simultaneously. Thus, such machines\n  exploit data level parallelism, but not concurrency: there are\n  simultaneous (parallel) computations, but only a single process\n  (instruction) at a given moment<\/p>\n<\/blockquote>\n\n<p>This is enough to answer your next question.<\/p>\n\n<hr>\n\n<p><strong>How do these SSE4.2 and AVX improve CPU computations for TF tasks<\/strong><\/p>\n\n<p>They allow a more efficient computation of various vector (matrix\/tensor) operations. You can read more in <a href=\"https:\/\/www.polyhedron.com\/web_images\/intel\/productbriefs\/3a_SIMD.pdf\" rel=\"noreferrer\">these slides<\/a><\/p>\n\n<hr>\n\n<p><strong>How to make Tensorflow compile using the two libraries?<\/strong><\/p>\n\n<p>You need to have a binary which was compiled to take advantage of these instructions. The easiest way is to <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"noreferrer\">compile it yourself<\/a>. As Mike and Yaroslav suggested, you can use the following bazel command<\/p>\n\n<p><code>bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.2 --config=cuda -k \/\/tensorflow\/tools\/pip_package:build_pip_package<\/code><\/p>\n"}]}
{"QuestionId":41293276,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2016-12-22T23:44:53.700","AcceptedAnswerId":null,"OwnerUserId":5225453.0,"Title":"Error when checking model input: Found: <keras.preprocessing.image.DirectoryIterator... object> in Keras","Body":"<p>I have batches of data in an iterator, here, using native Keras and nothing fancy:  <\/p>\n\n<p> . <\/p>\n\n<p>which appears fine:  <\/p>\n\n<p><\/p>\n\n<blockquote>\n  <p>keras.preprocessing.image.DirectoryIterator object at 0x7f107c004210\\<\/p>\n<\/blockquote>\n\n<p>But now I've compiled and I'm ready to :  <\/p>\n\n<p> . <\/p>\n\n<p>  <\/p>\n\n<p>but I keep getting:  <\/p>\n\n<p><\/p>\n\n<p>No Keras doesn't like that I'm using an iterator? Why can't I fit using an iterator? I thought that's the whole point -- don't eat up all your memory, instead use some kind of batch iterator.  <\/p>\n","answers":[{"AnswerId":"41296621","CreationDate":"2016-12-23T06:56:08.430","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p><code>fit<\/code> method expects its input to be a Numpy array or list of Numpy array. You should use <a href=\"https:\/\/keras.io\/models\/model\/#methods\" rel=\"nofollow noreferrer\"><code>fit_generator<\/code><\/a> instead, which takes a generator as its argument.<\/p>\n\n<pre><code>model.fit_generator(generator=batches, \n                    validation_data=val_batches, \n                    nb_epoch=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41293576,"AnswerCount":3,"Tags":"<tensorflow><distributed><synchronous>","CreationDate":"2016-12-23T00:23:04.763","AcceptedAnswerId":null,"OwnerUserId":7332650.0,"Title":"Distributed Tensorflow: good example for synchronous training on CPUs","Body":"\n\n<p>I am new to distributed tensorflow and am looking for a good example to perform synchronous training on CPUs.<\/p>\n\n<p>I have already tried the <a href=\"https:\/\/github.com\/ischlag\/distributed-tensorflow-example\" rel=\"nofollow noreferrer\">Distributed Tensorflow Example<\/a> and it can perform the asynchronous training successfully over 1 parameter server (1 machine with 1 CPU) and 3 workers (each worker = 1 machine with 1 CPU). However, when it comes to the synchronous training, I am not able to run it correctly, although I have followed the tutorial of \n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/other_functions_and_classes#usage\" rel=\"nofollow noreferrer\">SyncReplicasOptimizer(V1.0 and V2.0)<\/a>. <\/p>\n\n<p>I have inserted the official SyncReplicasOptimizer code into the working asynchronous training example but the training process is still asynchronous. My detailed code is as follows. Any code relates to synchronous training is within the block of ******.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Anything wrong with my code? Or can I have a good example to follow?<\/p>\n","answers":[{"AnswerId":"43770863","CreationDate":"2017-05-03T22:13:30.793","ParentId":null,"OwnerUserId":"7506316","Title":null,"Body":"<p>I am not sure if you would be interested in user-transparent distributed tensorflow which uses MPI in the backend. We have recently developed one such version with MaTEx: <a href=\"https:\/\/github.com\/matex-org\/matex\" rel=\"nofollow noreferrer\">https:\/\/github.com\/matex-org\/matex<\/a>.<\/p>\n\n<p>Hence, for distributed TensorFlow, you would not need to write a SyncReplicaOptimizer code, since all the changes are abstracted from the user. <\/p>\n\n<p>Hope this helps.<\/p>\n"},{"AnswerId":"43828245","CreationDate":"2017-05-07T05:48:24.873","ParentId":null,"OwnerUserId":"7951227","Title":null,"Body":"<p>I think your question can be answered as the comments in the issue <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/9596\" rel=\"nofollow noreferrer\">#9596<\/a> of the tensorflow.\nThis problem is caused by the bugs of the new version of tf.train.SyncReplicasOptimizer(). You can use old version of this API to avoid this problem.<\/p>\n\n<p>Another solution is from the <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_models\" rel=\"nofollow noreferrer\">Tensorflow Distributed Benchmarks<\/a>. Take a look at the source code, and you can find that they synchronize workers manually through the queue in the tensorflow. Through experiments, this benchmark runs exactly as what you expect.<\/p>\n\n<p>Hope these comments and resources can help you solve your problem. Thanks!<\/p>\n"},{"AnswerId":"44426422","CreationDate":"2017-06-08T04:03:08.660","ParentId":null,"OwnerUserId":"8129008","Title":null,"Body":"<p>One issue is that you need to specify an aggregation_method in the minimize method for it to run synchronously,<\/p>\n\n<pre><code>train = opt.minimize(cost, global_step=global_step, aggregation_method=tf.AggregationMethod.ADD_N)\n<\/code><\/pre>\n"}]}
{"QuestionId":41293588,"AnswerCount":1,"Tags":"<python><tensorflow><bazel>","CreationDate":"2016-12-23T00:24:41.057","AcceptedAnswerId":null,"OwnerUserId":7118229.0,"Title":"Uncertain how to run Bazel: Tensorflow Inception Retrain New Categories Tutorial Python","Body":"<p>I am experiencing difficulty using Bazel. More specifically with the workspace requirements. I am not sure how to interpret the following line of code from the TF <a href=\"https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/\" rel=\"nofollow noreferrer\">Retrain Inception<\/a> tutorial.<\/p>\n\n<pre><\/pre>\n\n<p>I am able to build the \"retrain.py\" file. This is my output from my build:<\/p>\n\n<pre><\/pre>\n\n<p>But I am uncertain how to proceed with the next step. Is \"bazel-bin\" a folder somewhere? is it included in the Tensorflow example folder? Or is this something I have to generate? <\/p>\n\n<p>Also, the way in which \"retrain\" is being referenced in the code makes me think it is a folder and no longer a python file. Did it change through this process? <\/p>\n\n<p>I would greatly appreciate a more detailed breakdown of what Bazel is doing in this line of code and how I adjust it to run the code on my directory of images. <\/p>\n\n<p>Thanks! <\/p>\n","answers":[{"AnswerId":"41294665","CreationDate":"2016-12-23T03:07:24.890","ParentId":null,"OwnerUserId":"3222797","Title":null,"Body":"<p>I believe to run this example you will have had to have built Tensorflow from source.  It appears you're running some Linux distro; you can find the bazel build instructions <a href=\"https:\/\/bazel.build\/versions\/master\/docs\/install.html\" rel=\"nofollow noreferrer\">here<\/a>.  Once you have that installed, you can pull the bleeding edge Tensorflow from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#installing_from_sources\" rel=\"nofollow noreferrer\">here<\/a> are \"building from source\" instructions.  Assuming you've done all of that, if you navigate to the root of the project i.e.<\/p>\n\n<pre><code>cd tensorflow\n<\/code><\/pre>\n\n<p>you should see something like this<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LoEpQ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LoEpQ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>you can see there are 5 <code>bazel<\/code> symlinks, one of which is <code>bazel-bin<\/code>.<\/p>\n\n<p>Now just download the images as directed and build with<\/p>\n\n<pre><code>bazel build -c opt --copt=-mavx tensorflow\/examples\/image_retraining:retrain\n<\/code><\/pre>\n\n<p>Once the build is finished you can run the retraining<\/p>\n\n<pre><code>bazel-bin\/tensorflow\/examples\/image_retraining\/retrain --image_dir ~\/flower_photos\n<\/code><\/pre>\n\n<p>Note that both of the above commands need to be run from the root of the project (<code>tensorflow\/<\/code> in this case) and the tutorial assumes you have put the flower images in your home dir.  <\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":41294064,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2016-12-23T01:35:11.773","AcceptedAnswerId":41295278.0,"OwnerUserId":5737630.0,"Title":"TensorFlow: Getting a Sub-Tensor from a Tensor Using Indexing","Body":"<p> is a  with shape <\/p>\n\n<p>I want to get a new tensor from the following indixes of : <\/p>\n\n<p>I tried all the followings, but none work:<\/p>\n\n<pre><\/pre>\n\n<p>Why are these not working, and what can I do?<\/p>\n","answers":[{"AnswerId":"43007940","CreationDate":"2017-03-24T19:46:09.203","ParentId":null,"OwnerUserId":"3741633","Title":null,"Body":"<p>I think what you are looking for is the gather function. <\/p>\n\n<pre><code>B = tf.constant([10, 20, 30])\ntf.gather(A, B)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather<\/a><\/p>\n"},{"AnswerId":"41294906","CreationDate":"2016-12-23T03:47:02.267","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>I don't think fancy indexing like that is supported yet in TensorFlow. Keep an eye on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206<\/a> for updates (maybe there is somewhere else too). <\/p>\n\n<p>If you'd like to see what <em>is<\/em> available, it looks like they have some documentation about <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Tensor.__getitem__\" rel=\"nofollow noreferrer\"><code>__tensor.__getitem__<\/code><\/a>.<\/p>\n"},{"AnswerId":"41295278","CreationDate":"2016-12-23T04:40:21.230","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<pre><code>C = tf.nn.embedding_lookup(A, B)\n<\/code><\/pre>\n\n<p>where B is a tensor with the values [10,20,30]<\/p>\n\n<p>For reference: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/embeddings\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/embeddings<\/a><\/p>\n"}]}
{"QuestionId":41294094,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2016-12-23T01:39:36.627","AcceptedAnswerId":41294283.0,"OwnerUserId":3993270.0,"Title":"Is there a difference between a placeholder and variable when not building a model?","Body":"<p>I am trying to understand the difference between a placeholder and a variable in TensorFlow:<\/p>\n\n<pre><\/pre>\n\n<p>I also read the Stack Overflow's question below. I understand the difference when they are the input of a model.<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/stackoverflow.com\/q\/38296200\/3924118\">InvalidArgumentError: You must feed a value for placeholder tensor Placeholder<\/a><\/p>\n<\/blockquote>\n\n<p>However, in general, if we are not building a model, is there still a difference between  and ?<\/p>\n","answers":[{"AnswerId":"41294283","CreationDate":"2016-12-23T02:04:32.460","ParentId":null,"OwnerUserId":"2570465","Title":null,"Body":"<h3>Placeholder<\/h3>\n\n<p>A placeholder is used for feeding external data into a Tensorflow computation (stuff outside the graph). Here's some documentation: (<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/how_tos\/reading_data\/#feeding\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.10\/how_tos\/reading_data\/#feeding<\/a>)<\/p>\n\n<blockquote>\n  <p>TensorFlow's feed mechanism lets you inject data into any Tensor in a\n  computation graph. A python computation can thus feed data directly\n  into the graph.<\/p>\n<\/blockquote>\n\n<p>I personally would draw an analogy from placeholders to reading from standard input.<\/p>\n\n<pre><code>x = raw_input()\nX = tf.placeholder(\"float\")\n<\/code><\/pre>\n\n<p>When you read from standard input, you need to \"inject data\" from an external source. Same with a placeholder. It lets you \"inject data\" that's external to the computation graph.<\/p>\n\n<p>If you're training a learning algorithm, the clear use case of placeholder is to feed in your training data. The training data isn't stored in the computation graph. How are you going to get it into the graph? By injecting it through a placeholder. A placeholder is basically you telling the graph \"I don't have this for you yet. But I'll have it for you when I ask you to run.\"<\/p>\n\n<h3>Variable<\/h3>\n\n<p>A variable is used to store state in your graph. It requires an initial value. One use case could be representing weights of a neural network or something similar. Here's documentation: (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable<\/a>)<\/p>\n\n<blockquote>\n  <p>A variable maintains state in the graph across calls to run(). You add\n  a variable to the graph by constructing an instance of the class\n  Variable.<\/p>\n  \n  <p>The Variable() constructor requires an initial value for the variable,\n  which can be a Tensor of any type and shape. The initial value defines\n  the type and shape of the variable. After construction, the type and\n  shape of the variable are fixed. The value can be changed using one of\n  the assign methods.<\/p>\n<\/blockquote>\n\n<p>I personally would draw an analogy between Tensorflow Variables and assigning a variable in Python to anything that is not dependent on external stuff. For example,<\/p>\n\n<pre><code># Tensorflow:\nW = tf.Variable(rng.randn(), name=\"weight\")\n\n# Standard python:\nw = 5\nw = \"hello\"\nw = [1, 2, 3, 4, 5]\n<\/code><\/pre>\n\n<p><code>W<\/code> represents some sort of result of your computation. Just like how you must initialize all your variables in Python (you can't just run a command <code>x<\/code> you have to say <code>x = ...something...<\/code>), you have to initialize all Variable objects in Tensorflow.<\/p>\n\n<h3>Variable vs. Placeholder<\/h3>\n\n<p>There's not much related between <code>tf.Variable<\/code> and <code>tf.placeholder<\/code> in my opinion. You use a <code>Variable<\/code> if you need to store state. You use a <code>placeholder<\/code> if you need to input external data.<\/p>\n\n<p>If you are not building a model, you should still use <code>tf.placeholder<\/code> if you want to insert external data that you don't necessarily have while you're defining the graph. If you are not building a model, you still need <code>tf.Variable<\/code> if you want to store some kind of result of your computation while the graph is being run.<\/p>\n\n<h3>Why have both?<\/h3>\n\n<p>I'm not an expert in Tensorflow, so I can only speculate as to why the design has both. <\/p>\n\n<p>A big difference between placeholders and variables is that placeholders can have variable size, but the shape of a <code>tf.Variable<\/code> must be specified while constructing the graph.<\/p>\n\n<p>Variable size placeholders sense: maybe I only want to input a training batch of size 5 right now, but maybe I want to increase the batch size later on. Maybe I don't know ahead of time how many training examples I'm going to get.<\/p>\n\n<p>Variable size variables don't make sense: <code>tf.Variable<\/code> holds the learned parameters of your model, and the number of parameters shouldn't change. Furthermore, Tensorflow extends to distributed computation. If you had <code>Variables<\/code> whose shape changed throughout the computation, it would be very difficult to keep it properly distributed among 1000 computers.<\/p>\n\n<p>Usually, you build a model and all parameters are known ahead of time, so that's what <code>tf.Variable<\/code> is probably used to represent. <code>tf.placeholder<\/code> is probably for everything else outside of your model (or computation graph) and so that can be more flexible.<\/p>\n"},{"AnswerId":"45003763","CreationDate":"2017-07-10T04:26:41.473","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The most obvious difference between the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"noreferrer\">tf.Variable<\/a> and the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"noreferrer\">tf.placeholder<\/a> is that<\/p>\n\n<hr>\n\n<blockquote>\n  <p>you use variables to hold and update parameters. Variables are\n  in-memory buffers containing tensors. They must be explicitly\n  initialized and can be saved to disk during and after training. You\n  can later restore saved values to exercise or analyze the model.<\/p>\n<\/blockquote>\n\n<p>Initialization of the variables is done with <code>sess.run(tf.global_variables_initializer())<\/code>. Also while creating a variable, you need to pass a Tensor as its initial value to the Variable() constructor and when you create a variable you always know its shape.<\/p>\n\n<hr>\n\n<p>On the other hand you can't update the placeholder. They also should not be initialized, but because they are a promise to have a tensor, you need to feed the value into them <code>sess.run(&lt;op&gt;, {a: &lt;some_val&gt;})<\/code>. And at last, in comparison to a variable, placeholder might not know the shape. You can either provide parts of the dimensions or provide nothing at all.<\/p>\n\n<hr>\n\n<p><strong>There other differences:<\/strong><\/p>\n\n<ul>\n<li>the values inside the variable can be updated during optimizations<\/li>\n<li>variables can be <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variable_scope\" rel=\"noreferrer\">shared<\/a>, and can be <a href=\"https:\/\/stackoverflow.com\/q\/40736859\/1090562\">non-trainable<\/a><\/li>\n<li>the values inside the variable can be stored after training<\/li>\n<li>when the variable is created, <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variables#creation\" rel=\"noreferrer\">3 ops are added to a graph<\/a> (variable op, initializer op, ops for the initial value)<\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/a\/43536220\/1090562\">placeholder is a function, Variable is a class<\/a> (hence an uppercase)<\/li>\n<li>when you use TF in a distributed environment, variables are stored in a special place (<a href=\"https:\/\/www.tensorflow.org\/deploy\/distributed\" rel=\"noreferrer\">parameter server<\/a>) and are shared between the workers.<\/li>\n<\/ul>\n\n<p>Interesting part is that not only placeholders can be fed. You can feed the value to a Variable and even to a constant.<\/p>\n"}]}
{"QuestionId":41294258,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2016-12-23T02:00:59.107","AcceptedAnswerId":41294873.0,"OwnerUserId":3993270.0,"Title":"TensorFlow: Input for a session to compute a variable? FailedPreconditionError: Attempting to use uninitialized value weight_2","Body":"<p>I am trying to run a TensorFlow session to find out the value of a variable W. Below is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Then I got the following outputs &amp; errors:<\/p>\n\n<pre><\/pre>\n\n<p>I am guessing I need a formula of W (or something like that) for the session to run, but not sure exactly what is needed ... What exactly needs to be provided so a session can compute the value of a variable? Thanks!<\/p>\n","answers":[{"AnswerId":"41294873","CreationDate":"2016-12-23T03:41:53.173","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>The function name <code>initialize_all_variables<\/code> is a bit misleading (and has been changed in 0.12). It returns an op you need to run. <code>W<\/code> won't be initialized until you call <code>sess.run(init)<\/code>.<\/p>\n\n<p><code>W.initial_value<\/code> shows the value that <em>it will be initialized with<\/em>, from what I read in the docs (emphasis mine):<\/p>\n\n<blockquote>\n  <p><code>tf.Variable.initial_value<\/code><\/p>\n  \n  <p>Returns the Tensor used as the initial value for the variable.<\/p>\n  \n  <p>Note that this is different from <code>initialized_value()<\/code> which runs the op\n  that initializes the variable before returning its value. <strong>This method\n  returns the tensor that is used by the op that initializes the\n  variable.<\/strong><\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41294542,"AnswerCount":1,"Tags":"<python><tensorflow><distributed-computing><fault-tolerance>","CreationDate":"2016-12-23T02:47:36.793","AcceptedAnswerId":null,"OwnerUserId":6527887.0,"Title":"tensorflow monitoredsession usage","Body":"<p>I have the following code to perform simple arithmetic calculations. I am trying to implement fault tolerance in it by using a Monitored Training session.<\/p>\n\n<pre><\/pre>\n\n<p>But it is throwing the following error:-<\/p>\n\n<pre><\/pre>\n\n<p>which in my opinion is due to incorrect argument given to config. Am I using the parameters right? Please advice.<\/p>\n","answers":[{"AnswerId":"41469423","CreationDate":"2017-01-04T17:03:00.907","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>First issue is in following lines. It uses a local session for a distributed (device assigned) op. Why do you need that?<\/p>\n\n<pre><code>sess = tf.Session()\ntf.train.global_step(sess, global_step_tensor)\n<\/code><\/pre>\n\n<p>Second issue:\nCode uses <code>WorkerSessionCreator<\/code>. One machine should be chief. In this case <code>ChiefSessionCreator<\/code> should be used.\nI recommend to use following <code>tf.train.MonitoredTrainingSession<\/code>.<\/p>\n\n<p>Third issue:\n<code>sess.should_stop()<\/code> should be checked before each <code>run<\/code>.<\/p>\n"}]}
{"QuestionId":41294996,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><deep-learning>","CreationDate":"2016-12-23T04:00:54.940","AcceptedAnswerId":null,"OwnerUserId":4410813.0,"Title":"TensorFlow doesn't learn when input=output (or probably I am missing something)","Body":"<p>I have used TensorFlow before for training a model to recognize images (something similar to this example: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception<\/a>). I was trying to do something similar with inception-v4 from scratch (using the model from <a href=\"https:\/\/github.com\/tensorflow\/models\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models<\/a>). But it was not working, it wasn't learning, so I decided to reduce my code and simplify everything. The model is a hidden layer with the same neurons as inputs and what the model has to learn is that input=output. Seems pretty simple and easy, but the results of my code are far from good.<\/p>\n\n<p>I have tried different things: sigmoid cross entropy instead of softmax cross entropy, other optimizers, different learning rates, models more comples with more neuros and more layers, reduce_sum instead of reduce mean for the loss ... But the only times I got \"good results\" (>90%) it seems because the model falls into a local minimum and I am not able to replicate them. Usually I see 70% precision with a high loss (>=0.7).<\/p>\n\n<p>My conclusion is that there is something wrong in my code that I am not able to spot and that is the reason the learning process is slow and get stuck at high losses. Otherwise the code with inception-v4 for images will at least learn something and don't get the loss stuck at the same value all the time.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/p11tY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/p11tY.png\" alt=\"learning rate and loss graphs\"><\/a><\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I think the issue is related of how backpropagation works and the effects of the relu layer and the softmax layer in the error. With min square error the network converges more often. Also smaller batch sizes helps.\nSo basically online stocastic backpropagation doesn't work well when there are repeated inputs in the batch, the errors seem to cancel somehow. So smaller batch sizes helps.<\/p>\n\n<p>Relu + softmax makes the outputs to have similar and small probabilities and cross entropy makes the network to converge to this. A winner takes all strategy will have solved this problem. Probably this is why sigmoid works beter than softmax, but still have the problem of crossentropy.<\/p>\n\n<p><strong>UPDATE 2:<\/strong><\/p>\n\n<p>The real reason of the bad convergence is the weight initialization and how backpropagation works. The update of the weight when the output of the neuron is 0 or 1 are very different in magnitude. When the output is 1 the input weight with the value 1 and the bias weight are updated with a bigger number than when the output is 0, where the bias is updated in the other direction by a small factor. This makes the network to converge very very slowly, and with it is very dependent of the initialization of the weights.<\/p>\n","answers":[]}
{"QuestionId":41295005,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-23T04:01:20.593","AcceptedAnswerId":41295218.0,"OwnerUserId":5716667.0,"Title":"The objects that have same name within different 'with.name_scope()\" are different?","Body":"<p>I'm making deep CNN layer, so need many variables similar.\nWhen making layer, I want to make variable name simple as much as possible.  <\/p>\n\n<p>Is it okay to use same name within different 'tf.name_scope()' as following code? I'm so wonder whether 'earlier bias(variable) is overlapped by later bias when save model, and earlier bias is not saved'. In my opinion, guessing earlier bias and later bias are managed separately by each name_scope, but I'm not sure...    <\/p>\n\n<pre><\/pre>\n\n<p>Could you let me know the proper and exact concept of this problem? Thank you.<\/p>\n","answers":[{"AnswerId":"41295218","CreationDate":"2016-12-23T04:32:09.453","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>Check this: <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variable_scope\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/variable_scope\/<\/a><\/p>\n\n<p>your variable names are:<\/p>\n\n<pre><code>CNN\/c:0\nCNN\/b:0\nNN\/c:0\nNN\/b:0\n<\/code><\/pre>\n"}]}
{"QuestionId":41295168,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-12-23T04:24:59.310","AcceptedAnswerId":47920224.0,"OwnerUserId":219603.0,"Title":"TensorFlow: How to concatenate tensors by duplicating one of the tensor?","Body":"<p>I would like to concatenate two tensors by duplicating one of the tensor. For example, I have two tensors of shape [2, 2, 3] and [2, 3]. The result should be in the shape of [2, 2, 6].<\/p>\n\n<pre><\/pre>\n\n<p>So If the two tensors are of shape [10, 5, 8] and [10, 3], the result should be of shape [10, 5, 11].<\/p>\n\n<p><strong>UPDATED<\/strong><\/p>\n\n<p>Another example:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41295254","CreationDate":"2016-12-23T04:37:28.353","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>You need to use the function pack <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#pack\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#pack<\/a><\/p>\n\n<p>The code should be something similar to this:<\/p>\n\n<pre><code>t2 = tf.transpose(t2)\ntf.pack([t1, tf.pack(t2,t2, axis=1)], axis=2)\n<\/code><\/pre>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<blockquote>\n  <p>So If the two tensors are of shape [10, 5, 8] and [10, 3], the result\n  should be of shape [10, 5, 11].<\/p>\n<\/blockquote>\n\n<p>I am not sure this is possible. Can you write down how are this tensors and how the result tensor should be? The second dimension of the first and result tensors it is not something you can create from scratch for the second tensor. <\/p>\n"},{"AnswerId":"47920224","CreationDate":"2017-12-21T07:47:27.143","ParentId":null,"OwnerUserId":"2301712","Title":null,"Body":"<p>The function <code>tf.tile<\/code> can help you make it. Click <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/tile\" rel=\"nofollow noreferrer\">here<\/a> to get detail of the function. <\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nt1 = np.reshape(np.arange(3*4*5), [3,4,5])\nt2 = np.reshape(np.arange(3*1*2), [3,2])\n\n# Keep t1 stay\nt1_p = tf.placeholder(tf.float32, [3,4,5])\n\n# Change t2 from shape(3,2) to shape(3,4,2) followed below two steps:\n#    1. copy element of 2rd dimension of t2 as many times as you hope, as the updated example, it is 4\n#    2. reshape the tiled tensor to compatible shape\nt2_p = tf.placeholder(tf.float32, [3,2])\n# copy the element of 2rd dimention of t2 by 4 times\nt2_p_tiled = tf.tile(t2_p, [1, 4])\n# reshape tiled t2 with shape(3,8) to the compatible shape(3,4,2)\nt2_p_reshaped = tf.reshape(t2_p_tiled, [3,4,2])\n\n# Concat t1 and changed t2, then you will get t3 you want\nt3_p = tf.concat([t1_p, t2_p_reshaped], 2)\n\nsess = tf.InteractiveSession()\nt3 = sess.run(t3_p, {t1_p:t1, t2_p:t2})\n\nprint '*' * 20\nprint t1\nprint '*' * 20\nprint t2\nprint '*' * 20\nprint t3\n\n# if you confused what the tf.tile did, you can print t2_p_tiled to see what happend\nt2_tile = sess.run(t2_p_tiled, {t2_p:t2})\nprint '*' * 20\nprint t2_tile\n<\/code><\/pre>\n"}]}
{"QuestionId":41295662,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-12-23T05:29:57.973","AcceptedAnswerId":null,"OwnerUserId":2277321.0,"Title":"How to transfer weightings to tensorflow RNN cell","Body":"<p>I have a set of weightings of the trained model implemented in matlab. I would like to port the weightings to tensorflow. However, tf.rnn.rnn_cell.LSTMCell with 500 cell has weight matrix with shape (1524, 2000). why 1524? why 2000? This does not fit the dimensions of my weightings at all.<\/p>\n\n<p>My model has 3 hidden layers, each has 1000 nodes, and the last hidden layer is temporal layer(Recurrent). the input dimension is 1539. output dimension is 1026.\nThe temporal layer has 1x1000 temporal weighting and 1000x1000 layer weighting and 1x1000 for bias.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41303403","CreationDate":"2016-12-23T14:40:36.947","ParentId":null,"OwnerUserId":"6778091","Title":null,"Body":"<p>first convert your pre-trained weights into tensor, this can be done by reading a file containing your weights and using <code>weights = tf.convert_to_tensor(your_weights)<\/code><\/p>\n\n<p>then find your name of lstm cell weights(this can be done by list all ops in this graph) <\/p>\n\n<p>finally use <code>tf.train.import_meta_graph(meta_graph_def=your_meta_graph_def, input_map={\"lstm-cell-weights-name\": weights})<\/code>(the most important thing is the <code>input_map<\/code> argument)<\/p>\n\n<p><strong>Only for version after 0.12.<\/strong> <\/p>\n\n<p>because older version doesn't support <code>input_map<\/code> argument<\/p>\n"},{"AnswerId":"41297190","CreationDate":"2016-12-23T07:40:05.197","ParentId":null,"OwnerUserId":"2878004","Title":null,"Body":"<p>It looks like you are calling your cell using x as input. The size of the weight matrix will be (500 + 1024) x (4 * 500). The LSTM has four gating functions but for efficiency reasons their individual matrices are concatenated together. That's why the second dimension of the matrix is 4 * 500. The first dimension is the size of the cell plus the size of the input because it needs to be multiplied with the input concatenated with the output from the previous step.<\/p>\n"}]}
{"QuestionId":41296981,"AnswerCount":1,"Tags":"<neural-network><computer-vision><deep-learning><caffe><nvidia-digits>","CreationDate":"2016-12-23T07:25:12.460","AcceptedAnswerId":null,"OwnerUserId":6675900.0,"Title":"face recognition based on Deep Learning (Siamese Architecture)","Body":"<p>I want to used pre trained model of face for the face identification. I want to use Siamese architecture which require a few number of images. Could you give me any trained model which I can change for the Siamese architecture? How can I change the network model which I can put two images for find similarity (I do not want to create image based on the tutorial <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\/tree\/master\/examples\/siamese\" rel=\"nofollow noreferrer\">here<\/a>)? I want to use the system for real time application do you have any recommendation?<\/p>\n","answers":[{"AnswerId":"41297269","CreationDate":"2016-12-23T07:45:58.250","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I suppose you can use <a href=\"https:\/\/github.com\/AlfredXiangWu\/face_verification_experiment\" rel=\"nofollow noreferrer\">this model<\/a>, described in <a href=\"https:\/\/arxiv.org\/abs\/1511.02683\" rel=\"nofollow noreferrer\"><em>Xiang Wu, Ran He, Zhenan Sun, Tieniu Tan<\/em> <strong>A Light CNN for Deep Face Representation with Noisy Labels<\/strong> (arXiv 2015)<\/a> as a a strating point for your experiments.<\/p>\n\n<p>As for the Siamese network, what you are trying to earn is a mapping from a face image into some high dimensional vector space, in which distances between points reflects (dis)similarity between faces.<br>\nTo do so, you only need <em>one<\/em> network that gets a face as an input and produce a high-dim vector as an output.<br>\nHowever, to train this <em>single<\/em> network using the Siamese approach, you are going to duplicate it: creating two instances of the <em>same<\/em> net (you need to explicitly link the weights of the two copies). During training you are going to provide pairs of faces to the nets: one to each copy, then the single loss layer on top of the two copies can compare the high-dimensional vectors representing the two faces and compute a loss according to a \"same\/not same\" label associated with this pair.<br>\nHence, you only need the duplication for the training. In test time (<code>'deploy'<\/code>) you are going to have a <em>single<\/em> net providing you with a semantically meaningful high dimensional representation of faces.<\/p>\n\n<p>For a more advance Siamese architecture and loss see <a href=\"https:\/\/stackoverflow.com\/q\/33330779\/1714410\">this thread<\/a>. <\/p>\n\n<hr>\n\n<p>On the other hand, you might want to consider the approach described in <a href=\"https:\/\/arxiv.org\/abs\/1605.07270\" rel=\"nofollow noreferrer\"><em>Oren Tadmor, Yonatan Wexler, Tal Rosenwein, Shai Shalev-Shwartz, Amnon Shashua<\/em> <strong>Learning a Metric Embedding for Face Recognition using the Multibatch Method<\/strong> (arXiv 2016)<\/a>. This approach is more efficient and easy to implement than pair-wise losses over image pairs. <\/p>\n"}]}
{"QuestionId":41298243,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2016-12-23T09:00:42.753","AcceptedAnswerId":null,"OwnerUserId":5112232.0,"Title":"cnn using tensorflow for own image set - what should be the tfrecord format","Body":"<p>I have an image data set of size 600 x 400 each and I have converted each of the images to TFRecord format. But I am unable to figure out how to use this data? I have seen the imagenet dataset and found only one single binary file (when extracted it form<a href=\"http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz:\/\/\" rel=\"nofollow noreferrer\"> here<\/a>). <\/p>\n\n<p>Is it that for an image dataset there will be only one TFRecord or each individual images will have their own TFRecord files? <\/p>\n","answers":[{"AnswerId":"41300571","CreationDate":"2016-12-23T11:27:28.947","ParentId":null,"OwnerUserId":"3367721","Title":null,"Body":"<p>Tensorflow doesnt look for single tfrecord file. So feel free and point your \"data directory\" and \"train directory\" to the location which has set of tfrecord files.<\/p>\n\n<p>Also, keep in mind files should be in respective directories based on their names like TRAIN-*.tfrecord files in \"train directory\".<\/p>\n\n<p>Answer can be more specific if you mentioned what model of TF you are targeting to run on these TF record files.<\/p>\n\n<p>Hope it helps.<\/p>\n"}]}
{"QuestionId":41298722,"AnswerCount":0,"Tags":"<python><deep-learning><caffe><pycaffe><unet>","CreationDate":"2016-12-23T09:32:46.340","AcceptedAnswerId":null,"OwnerUserId":6362938.0,"Title":"caffe \/ unet: Could not compute map between tops; are they connected by spatial layers?","Body":"<p>I want to generate a U-Net like deep learning architecture with the following python code (for pycaffe):<\/p>\n\n<pre><\/pre>\n\n<p>When executed the following error message is returned:<\/p>\n\n<pre><\/pre>\n\n<p>I have no idea what the problem is. If I reduce the layers such that there is only a single deconvolution+crop+merge step, the network description string is generated without a problem.<\/p>\n","answers":[]}
{"QuestionId":41298791,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-23T09:36:42.673","AcceptedAnswerId":41341175.0,"OwnerUserId":1844810.0,"Title":"Implementation of optical flow using CNN by tensorflow","Body":"<p>I want to implement <em>FlowNetCorr model<\/em>  in Figure 2 on \n<a href=\"https:\/\/arxiv.org\/abs\/1504.06852\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1504.06852<\/a> , by Tensorflow.<\/p>\n\n<p>Although I know the basic implementation of LeNet or Alexnet, it is hard to write more complex models.<\/p>\n\n<p>I don't know how to implement correlation layer and  by Tensorflow. Is there any recommendation material or suggestion?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/uzuTz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/uzuTz.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HBJJD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HBJJD.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41341175","CreationDate":"2016-12-27T07:50:48.973","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>You may consider using this op: <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/nn#Convolution\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/nn#Convolution<\/a><\/p>\n"}]}
{"QuestionId":41299126,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning><google-developer-tools>","CreationDate":"2016-12-23T09:57:05.207","AcceptedAnswerId":null,"OwnerUserId":6761066.0,"Title":"TensorFlow from Google - Data Security","Body":"<p>Does anyone have any idea whether Google collects data that one supplies to Tensorflow? I mean it is open source, but it falls under their licences.<\/p>\n","answers":[{"AnswerId":"41301310","CreationDate":"2016-12-23T12:12:17.547","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>You can read the whole code of TF. There is nothing there that sends any data, thus the answer is <strong>no<\/strong>, there is no data collection. TF is not a service but a library, thus having an open source gives you full guarantee whether anything \"bad\" is happening to your data.<\/p>\n\n<blockquote>\n  <p>falls under their licences<\/p>\n<\/blockquote>\n\n<p>I do not understand what do you mean by that. TensorFlow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/LICENSE\" rel=\"nofollow noreferrer\">is released under Apache 2.0 license<\/a>. There is no such thing as \"their licences\" (you mean Google?).<\/p>\n"},{"AnswerId":"50601700","CreationDate":"2018-05-30T09:56:57.100","ParentId":null,"OwnerUserId":"7243311","Title":null,"Body":"<p>Doesn't TF actually also uses google's models from cloud? I'm pretty sure google uses cloud data to provide better models for TF. <\/p>\n\n<p>I'd recommend you to stay away from it. Only by writing your models from scratch you will learn to do useful stuff with it long term. I can also recommend weka for java, it's open source and you can only look at the code of the models there an implement it yourself adjusting for your needs.<\/p>\n"}]}
{"QuestionId":41300088,"AnswerCount":1,"Tags":"<scikit-learn><tensorflow><neural-network><deep-learning><keras>","CreationDate":"2016-12-23T10:56:42.073","AcceptedAnswerId":41301119.0,"OwnerUserId":3827807.0,"Title":"Normalization produce bad result for mlp","Body":"<p>I tried to normalize my data with Sklearn and MinMaxScaler method (0.1, 1)\nBut the result is very disapointed.\nWithout MinMax normalization, I was at 78% of accuracy on my problem, and with min max normalization, it fall to 71%.\nDo you know what could be the problem ?<\/p>\n\n<p>My data shape is : [n_samples][1D_vector_of_values]<\/p>\n\n<p>Here is how I use sklearn for normalize:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for help!<\/p>\n","answers":[{"AnswerId":"41301119","CreationDate":"2016-12-23T12:01:35.230","ParentId":null,"OwnerUserId":"3109123","Title":null,"Body":"<p>it is not always the case normalization will give you good result as without since you are loosing some data by applying this discrimination method. it all depends on the nature of the data. <\/p>\n\n<p>I'd try to implement standardization instead of range normalization but be careful to use the same same standardization batch for testing and validation. you didn't give us more information about your data but I'd still ask you to implement feature selection after the normalization\/standardization. <\/p>\n"}]}
{"QuestionId":41301430,"AnswerCount":1,"Tags":"<javascript><google-chrome><firefox><tensorflow><tensorboard>","CreationDate":"2016-12-23T12:19:59.830","AcceptedAnswerId":null,"OwnerUserId":5806421.0,"Title":"Tensorboard is not running in Google Chrome","Body":"<p>After launching tensorboard using: <\/p>\n\n<pre><\/pre>\n\n<p>I get: <\/p>\n\n<blockquote>\n  <p>503 Service Unavailable<\/p>\n  \n  <p>Failed to connect to server 127.0.0.1<\/p>\n<\/blockquote>\n\n<p>In Google Chrome (Ubuntu 14.04 LTS, tensorflow 0.11.0).<\/p>\n\n<p>I use proxy settings in my laptop and they are configured correctly.<\/p>\n\n<p>But tensorboard opens in firefox but I get a blank image under events section.<\/p>\n\n<p>How can I get this running in Google Chrome ?<\/p>\n","answers":[{"AnswerId":"41351033","CreationDate":"2016-12-27T19:25:41.620","ParentId":null,"OwnerUserId":"1653720","Title":null,"Body":"<p>Have you configured Chrome to not use your proxy for 127.0.0.1? Possibly by using <code>--proxy-bypass-list<\/code>?<\/p>\n"}]}
{"QuestionId":41302629,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-12-23T13:47:36.140","AcceptedAnswerId":41303178.0,"OwnerUserId":5566910.0,"Title":"How to count the number of the True in boolean matrix by use of Tensorflow?","Body":"<p>How to count the number of the True in boolean matrix by use of Tensorflow? Thanks!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41303178","CreationDate":"2016-12-23T14:24:32.870","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>Cast the boolean values to <code>tf.in32<\/code> and sum them<\/p>\n\n<pre><code>import tensorflow as tf\n\ndata = [\n        [False, True, False,False,False, True, False,False,False,False],\n        [False, True, False,False,False, True, False,False,False,False],\n        [False, True, False,False,False, True, False,False,False,False],\n        [False, True, False,False,False, True, False,False,False,False]\n        ]\n\n\nvar = tf.Variable(data)\nnum_true = tf.reduce_sum(tf.cast(var, tf.int32))\n\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    print(sess.run(num_true))\n<\/code><\/pre>\n\n<blockquote>\n  <p>8<\/p>\n<\/blockquote>\n"},{"AnswerId":"47324423","CreationDate":"2017-11-16T08:02:06.840","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p><strong>Method 1:<\/strong>      <\/p>\n\n<pre><code>In [29]: data = [                                             \n        [False, True, False,False,False, True, False,False,False,False],\n        [False, True, False,False,False, True, False,False,False,False],\n        [False, True, False,False,False, True, False,False,False,False],\n        [False, True, False,False,False, True, False,False,False,False]\n        ]\n\nIn [30]: if_true = tf.where(tf.equal(tf.constant(data), True))\n# or just if_true = tf.where(tf.constant(data))\n\nIn [31]: sess.run(if_true).shape[0] \n# only get the indices where the item is true, then the number of \n# returned indices amounts to the number of True\nOut[31]: 8\n<\/code><\/pre>\n\n<p>This method is better in some way, for example to work out how many Falses are there and etc. <\/p>\n\n<p><strong>Method 2:<\/strong> <\/p>\n\n<pre><code>In [34]: if_true = tf.count_nonzero(tf.constant(data))                                              \n\nIn [35]: sess.run(if_true)\nOut[35]: 8\n<\/code><\/pre>\n"}]}
{"QuestionId":41302986,"AnswerCount":1,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2016-12-23T14:10:13.753","AcceptedAnswerId":null,"OwnerUserId":2006086.0,"Title":"tensorflow cnn error: InvalidArgumentError (see above for traceback): logits and labels must be same size","Body":"<p>I wrote this code by modified from tensorflow official tutorial.\nI had a network as following:<\/p>\n\n<pre><\/pre>\n\n<p>Error is following:<\/p>\n\n<pre><\/pre>\n\n<p>train_feature is (10, 19200) numpy array, and train label is (10,) numpy array. Anyone know why logits_size=[2400,8] ?<\/p>\n","answers":[{"AnswerId":"41307272","CreationDate":"2016-12-23T20:05:41.943","ParentId":null,"OwnerUserId":"777530","Title":null,"Body":"<p>Before passing an image (either it is the network's input or some intermediate output of a convolutional (CONV) layer) to a fully-connected (FC) layer, you have to make sure that (1) you properly reshape the image into a 1D vector as well as (2) set the dimensions of the weights of the network to agree with the 1D vector. In your case, the switching to FC layers happens after the pooling of the second layer. While you make sure that <code>h_pool2_flat<\/code> have a compatible dimension with the weights <code>w_fc1<\/code> of the first FC layer, you are not setting the flat dimension correctly. The hard-coded value 320 is not the correct dimension in this case. And trying to hard-code this may not be the best practice in general and the code may keep breaking whenever you introduce modifications to the size of the input or the convolutional stack of the network (e.g. by adding\/removing pooling layers or adjusting the strides of some layers). <\/p>\n\n<p>Instead, you should add some code to compute the flat dimension automatically and use the computed value to setup the dimensions as in the following example:<\/p>\n\n<pre><code># Here happens conversion from 2\/3D images to 1D vectors.\nh_pool2_shape = h_pool2.get_shape()\n# Don't hard-code the 1D vector dim. Rather, (1) multiply image's height,\n# width and depth to get it.\nh_pool2_dim = h_pool2_shape[1] * h_pool2_shape[2] * h_pool2_shape[3]\n# (2) Use the computed 1D dimension to set the FC1 weight matrix dimensions.\nw_fc1 = weight_variable(tf.stack([h_pool2_dim, 1024]))\nb_fc1 = bias_variable([1024])\n# (3) Use the same 1D dimension to correctly reshape the batch matrix.\nh_pool2_flat = tf.reshape(h_pool2, tf.stack([-1, h_pool2_dim]))\nh_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1) + b_fc1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41303920,"AnswerCount":1,"Tags":"<tensorflow><cloud><platform>","CreationDate":"2016-12-23T15:19:09.797","AcceptedAnswerId":41304111.0,"OwnerUserId":4654190.0,"Title":"how to install tensorflow on google cloud platform","Body":"<p>when I use the command pip install tensorflow the download is only 99% complete and terminated at that point. How can I install tensorflow using google cloud shell. <\/p>\n","answers":[{"AnswerId":"41304111","CreationDate":"2016-12-23T15:35:04.540","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>Instead of installing it by yourself you can use the machine learning api and use TensorFlow for training or inference. Just follow this guidelines: <a href=\"https:\/\/cloud.google.com\/ml\/docs\/quickstarts\/training\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/ml\/docs\/quickstarts\/training<\/a><\/p>\n\n<p>You can submit a TensorFlow job like this:<\/p>\n\n<pre><code>gcloud beta ml jobs submit training ${JOB_NAME} \\\n  --package-path=trainer \\\n  --module-name=trainer.task \\\n  --staging-bucket=\"${TRAIN_BUCKET}\" \\\n  --region=us-central1 \\\n  -- \\\n  --train_dir=\"${TRAIN_PATH}\/train\"\n<\/code><\/pre>\n"}]}
{"QuestionId":41304139,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-12-23T15:36:52.167","AcceptedAnswerId":41305739.0,"OwnerUserId":5718092.0,"Title":"TensorFlow Parallel CPU with Numpy","Body":"<p>I am trying to find a way to speed my code up.<\/p>\n\n<p>In short, I have a trained model which I use something like to obtain predicts, sort them and output a rank. <\/p>\n\n<pre><\/pre>\n\n<p>However, this process is extremely slow. I have about 10K test samples. I believe session doesn't work well with standard multiprocessing libraries in python while the multi-cpu\/gpu support is only available for tensorflow ops. <\/p>\n\n<p>Is there any elegant way to speed this up via multiprocessing? Or do I have to implement it as part of the computational graph in TF.<\/p>\n\n<p>Thanks a lot!   <\/p>\n","answers":[{"AnswerId":"41305739","CreationDate":"2016-12-23T17:45:23.287","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You can translate whole thing into TensorFlow graph:<\/p>\n\n<pre><code>pred_op = tf.constant([1,2,0])\ntruth = [0, 1, 2]\nsess = tf.Session()\nsort = tf.nn.top_k(pred_op)[1] # same as np.argsort(x)[::-1]\nrank = tf.where(tf.equal(sort,truth))[0][0] + 1\nprint(sess.run(rank))  # =&gt; 2\n<\/code><\/pre>\n"}]}
{"QuestionId":41307722,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><tensorflow><deep-learning>","CreationDate":"2016-12-23T20:52:53.010","AcceptedAnswerId":null,"OwnerUserId":3760771.0,"Title":"What is the difference between tf.gradients and tf.train.RMSPropOptimizer(lr_rate).compute_gradients?","Body":"<p>What is the difference between <\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>If there is a difference, say I want to then use , which one would be preferred? <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41308962","CreationDate":"2016-12-23T23:38:00.673","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>No, there's no difference between these two if your only aim is to get the gradients because <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.12\/tensorflow\/python\/training\/rmsprop.py#L50\" rel=\"nofollow noreferrer\"><code>RMSPropOptimizer<\/code><\/a> (or any specific optimizer) has to import the base class <code>Optimizer<\/code> which has the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.12\/tensorflow\/python\/training\/optimizer.py#L281\" rel=\"nofollow noreferrer\"><code>compute_gradients<\/code><\/a> function that in turn uses the <code>tf.gradients<\/code> implementation from here <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.12\/tensorflow\/python\/ops\/gradients_impl.py#L320\" rel=\"nofollow noreferrer\">tf.gradients<\/a> for computing the gradients which is the same as simply using <code>tf.gradients<\/code>.<\/p>\n\n<p>Instead, if you're training a neural net where you need to compute gradients then using <code>tf.train.RMSPropOptimizer(..)<\/code> would be advisable since <code>RMSProp<\/code> optimization is one of the ways for faster learning in a mini-batch setting.<\/p>\n\n<blockquote>\n  <p>RMSProp: Divide the learning rate (lr) for a weight (w) by a running average of the magnitudes of recent gradients for that weight.<\/p>\n<\/blockquote>\n\n<p>The idea behind applying different optimizations like (Momentum, AdaGrad, RMSProp) to the gradients is that after computing the gradients, you want to do some processing on them <em>and then<\/em> apply these processed gradients. This is for better learning.<\/p>\n\n<p>So, in RMSProp, you just don't use a constant learning rate throughout. It is instead dependent on the <em>\"running average of magnitudes of recent gradients\"<\/em><\/p>\n\n<p>Whereas the <code>tf.gradients(..)<\/code> just computes the gradient. If you use one of the optimizers from <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/optimizers\" rel=\"nofollow noreferrer\">tf.train.Optimizer<\/a>, it automatically computes the gradients on your graph.<\/p>\n"}]}
{"QuestionId":41308181,"AnswerCount":0,"Tags":"<image><matrix><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2016-12-23T21:44:33.623","AcceptedAnswerId":null,"OwnerUserId":4896087.0,"Title":"How to create adversarial images for ConvNet?","Body":"<p>Recently <a href=\"https:\/\/stackoverflow.com\/questions\/40966589\/how-to-use-image-and-weight-matrix-to-create-adversarial-images-in-tensorflow\">I asked a question<\/a> about creating adversarial images for a simple Softmax Regressions model. I managed to find the solution by myself. Now, I would like to do the same thing, but instead for a Convnet as outlined in the TensorFlow <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/mnist\/pros\/\" rel=\"nofollow noreferrer\">Deep MNIST for Experts<\/a> tutorial.<\/p>\n\n<p>In the previous question, the situation was very simple, since the weight matrix has exactly the same dimension as an image, so we can just do matrix addition like this:<\/p>\n\n<pre><\/pre>\n\n<p>where images_fool is the adversarial image, x is the original image of digit 2 and w_six is the weight matrix for digit 6, so that we can make an adversarial image of digit 2 that'll fool the classifier which will classify the image as digit 6.<\/p>\n\n<p>But now, with ConvNet, there are many more weight matrices and none of them has the same dimension as the original image. So, my question is, how do we add the weights onto the original image to produce a adversarial image?<\/p>\n\n<p>I appreciate any help. Thanks!<\/p>\n\n<p>The script can be found <a href=\"https:\/\/github.com\/georgeliu1998\/deep-learning\/blob\/master\/adversarial_images\/adversarial_images.ipynb\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n","answers":[]}
{"QuestionId":41308365,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><logistic-regression>","CreationDate":"2016-12-23T22:11:00.073","AcceptedAnswerId":null,"OwnerUserId":4040193.0,"Title":"Tensorflow: low loss resulting in low accuracy","Body":"<p>I am trying to use Tensorflow and build a simple logistic regression classifier. The training set consists of around half a million (480, 000) data points. I am training on batches of size 100. <\/p>\n\n<p>When training the model the loss falls from 0.002 to 0.000000034 within 20 training epochs ( the whole training set is traversed batch by batch in each epoch).<\/p>\n\n<p>However this gives an accuracy of 0.671725 on the training set, which seems very low. To my understanding the accuracy on the training set should be near 1.00 (100%). As I am quite new to Tensorflow I am not sure if I am missing a crucial detail in my calculations.<\/p>\n\n<ol>\n<li><p>Changing the batch size to 1000 does not make much of a difference - from 0.671698 to 0.671725.<\/p><\/li>\n<li><p>Could be overfitting, but I still can't understand why it lowers the accuracy on the training set as well.<\/p><\/li>\n<li><p>Decreasing the epochs to 10 gives the same accuracy.<\/p><\/li>\n<li><p>Increasing the epochs to 100 does not change the accuracy.<\/p><\/li>\n<li><p>Increasing the learning rate from 0.001 to 0.01  yields 0.000000000 after 75 epochs. The value is very small to fit in the precision. The training accuracy stays the same<\/p><\/li>\n<\/ol>\n\n<p>Code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea what could be causing this behaviour? Is it the size that is big, because when going down to 80 training data points the accuracy gets to one as it should be?<\/p>\n","answers":[]}
{"QuestionId":41308369,"AnswerCount":0,"Tags":"<python-2.7><tensorflow>","CreationDate":"2016-12-23T22:11:19.773","AcceptedAnswerId":null,"OwnerUserId":7336133.0,"Title":"Saving weights to memory in tensorflow","Body":"<p>I am using python2.7 and tensorflow and need to save the weights based on the results of the next computation. Therefore I need to temporarily save the previous weights, one way is by using:<\/p>\n\n<pre><\/pre>\n\n<p>But the problem is that the weights get saved to the disk, which is quite slow and adds up over time because i have to temporarily save the weights for every step. Is there a way to temporarily save the weights to the memory?\nAnother way would be by declaring the weight variables as placeholders, saving the weights after each step to other dummy weight variables and then passing the temporary weight variables back to the placeholders for every step. But that is very unsatisfying.<\/p>\n","answers":[]}
{"QuestionId":41308515,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-12-23T22:30:38.857","AcceptedAnswerId":41309577.0,"OwnerUserId":6850023.0,"Title":"Force copy of tensor when enqueuing","Body":"<p>first, I'm not sure if the title is very good, but it was the best I could come up with given my understanding of the situation.<\/p>\n\n<p>The background is that I'm trying to understand how queues work in tensorflow and ran into the following issue which puzzled me.<\/p>\n\n<p>I have a variable <em>n<\/em>, which I enqueue to a <em>tf.FIFOQueue<\/em>, and then I increment the variable. This is repeated several times, and one would expect a result similar to 0, 1, 2, ... However, when emptying the queue all values are the same.<\/p>\n\n<p>More precisely, the code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Which I expect would print:<\/p>\n\n<pre><\/pre>\n\n<p>Instead I get the following result:<\/p>\n\n<pre><\/pre>\n\n<p>It seems like I'm pushing some pointer to n to the queue, instead of the actual value, which is what I want. However, I don't really have any actual understanding of tensorflow internals, so maybe something else is going on?<\/p>\n\n<p>I tried changing<\/p>\n\n<pre><\/pre>\n\n<p>to<\/p>\n\n<pre><\/pre>\n\n<p>since answers to <a href=\"https:\/\/stackoverflow.com\/questions\/33717772\/how-can-i-copy-a-variable-in-tensorflow\">How can I copy a variable in tensorflow<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/34877523\/in-tensorflow-what-is-tf-identity-used-for\">In TensorFlow, what is tf.identity used for?<\/a> gives me the impression that it might help, but it does not change the result. I also tried adding a tf.control_dependencies(), but again, all values are the same when dequeueing.<\/p>\n\n<p>Edit: The output above is from running the code on a computer with a single CPU, when trying to see if there was some difference between different versions of tensorflow, I noticed if I run the code on a computer with CPU and GPU I get the \"expected\" result. Indeed, if I run with CUDA_VISIBLE_DEVICES=\"\" I get the result above, and with CUDA_VISIBLE_DEVICES=\"0\" I get the \"expected\" result.<\/p>\n","answers":[{"AnswerId":"41309577","CreationDate":"2016-12-24T01:34:17.050","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>To force a non-caching read you can do<\/p>\n\n<pre><code>q.enqueue(tf.add(q, 0))\n<\/code><\/pre>\n\n<p>This is what's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/294442996b2aeff00b1bfdc7e7169f7cb35bbf3d\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L625\" rel=\"nofollow noreferrer\">currently done<\/a> by the batch-normalization layer to force a copy.<\/p>\n\n<p>Semantics of how variables get read vs. referenced are in the process of getting revamped so they are temporarily non-intuitive. In particular, I expected <code>q.enqueue(v.read_value())<\/code> to force a non-caching read, but it doesn't fix your example on TF 0.12rc1<\/p>\n\n<p>Using GPU machine puts variable on GPU, while Queue is CPU only, so <code>enqueue<\/code> op forces a GPU->CPU copy.<\/p>\n"},{"AnswerId":"45565575","CreationDate":"2017-08-08T10:14:22.283","ParentId":null,"OwnerUserId":"3176103","Title":null,"Body":"<p>In case it helps, I've found that the other answers despite correct they do not work for all dtypes.<\/p>\n\n<p>For example, this works fine with floats or ints but fails when n is a string tensor:<\/p>\n\n<pre><code>q.enqueue(tf.add(n, 0))\n<\/code><\/pre>\n\n<p>This one fails when the queue uses tuples with heterogeneous types (e.g., ints and floats):<\/p>\n\n<pre><code>q.enqueue_many([[n]])\n<\/code><\/pre>\n\n<p>So, if you see yourself caught in any of these situations try this instead:<\/p>\n\n<pre><code>q.enqueue(tf.add(n, tf.zeros_like(n)))\n<\/code><\/pre>\n\n<p>Or, to enqueue a tuple <em>t<\/em>:<\/p>\n\n<pre><code>q.enqueue([tf.add(n, tf.zeros_like(n)) for n in t])\n<\/code><\/pre>\n\n<p>That works even for string tensors and heterogeneous tuple types.<\/p>\n\n<p>Hope it helps!<\/p>\n\n<p>--<\/p>\n\n<p><strong>Update<\/strong>: it looks like tf.bool types do not work with tf.zeros_like(). For those, an explicit cast to an integer type might be needed.<\/p>\n"}]}
{"QuestionId":41309829,"AnswerCount":1,"Tags":"<image><numpy><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2016-12-24T02:35:30.760","AcceptedAnswerId":null,"OwnerUserId":4896087.0,"Title":"How to compute the gradient of an image based on a certain class?","Body":"<p>In the <a href=\"http:\/\/karpathy.github.io\/2015\/03\/30\/breaking-convnets\/\" rel=\"nofollow noreferrer\">Breaking Linear Classifiers on ImageNet<\/a>, the author proposes the following way to create adversarial images that fool ConvNets:<\/p>\n\n<blockquote>\n  <p>In short, to create a fooling image we start from whatever image we\n  want (an actual image, or even a noise pattern), and then <strong>use\n  backpropagation to compute the gradient of the image pixels on any\n  class score<\/strong>, and nudge it along. We may, but do not have to, repeat\n  the process a few times. You can interpret backpropagation in this\n  setting as using dynamic programming to compute the most damaging\n  local perturbation to the input. Note that this process is very\n  efficient and takes negligible time if you have access to the\n  parameters of the ConvNet (backprop is fast), but it is possible to do\n  this even if you do not have access to the parameters but only to the\n  class scores at the end. In this case, it is possible to compute the\n  data gradient numerically, or to to use other local stochastic search\n  strategies, etc. Note that due to the latter approach, even\n  non-differentiable classifiers (e.g. Random Forests) are not safe (but\n  I haven\u2019t seen anyone empirically confirm this yet).<\/p>\n<\/blockquote>\n\n<p>I know I can calculate the gradient of an image like this:<\/p>\n\n<pre><\/pre>\n\n<p>But how do I compute the gradient of an image relative to another image class using TensorFlow or Numpy? Probably I need to do something similar to the process <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/mnist\/pros\/\" rel=\"nofollow noreferrer\">in this tutorial<\/a>? Such as:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm not sure exactly how...Specifically, I have an image of digit 2 as below:\n<a href=\"https:\/\/i.stack.imgur.com\/lnBnX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lnBnX.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><\/pre>\n\n<p>How do I compute the gradient of this image relative to the the digit 6 image class (with an example shown below)? (I guess I need to compute the gradient for all digit 6 images using back propagation.)\n<a href=\"https:\/\/i.stack.imgur.com\/9qOdq.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9qOdq.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><\/pre>\n\n<p>Thanks in advance for any help!<\/p>\n\n<p>Here're two related questions that I asked:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/40966589\/how-to-use-image-and-weight-matrix-to-create-adversarial-images-in-tensorflow\">How to use image and weight matrix to create adversarial images in TensorFlow?<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/41308181\/how-to-create-adversarial-images-for-convnet\">How to create adversarial images for ConvNet?<\/a><\/p>\n\n<p>And here's <a href=\"https:\/\/github.com\/georgeliu1998\/deep-learning\/blob\/master\/adversarial_images\/adversarial_images.ipynb\" rel=\"nofollow noreferrer\">my script<\/a>.<\/p>\n","answers":[{"AnswerId":"41570330","CreationDate":"2017-01-10T13:43:40.010","ParentId":null,"OwnerUserId":"1983378","Title":null,"Body":"<h2>Class scores only<\/h2>\n\n<p>If you only have access to class scores for any image you suggest there's not much fancy you can do to truly compute a gradient. <\/p>\n\n<p>If what is returned can be seen as a relative score for each category it is a vector <code>v<\/code> that is the result of some function <code>f<\/code> acting on a vector <code>A<\/code> that contains all the information on the image*. The true gradient of the function is given by the matrix <code>D(A)<\/code>, which depends on <code>A<\/code>, such that <code>D(A)*B = (f(A + epsilon*B) -f(A))\/epsilon<\/code> in the limit of small <code>epsilon<\/code> for any <code>B<\/code>. You <strong>could<\/strong> approximate this numerically using some small value for epsilon and a number of test matrices <code>B<\/code> (one for each element of <code>A<\/code> should be enough), but this is likely to be needlessly expensive.<\/p>\n\n<p>What you are trying to do is maximize the difficulty the algorithm has in recognizing the image. That is, for a given algorithm <code>f<\/code> you want to maximize some appropriate measure for how poorly the algorithm recognizes each of your images <code>A<\/code>. There is a plethora of methods for this. I'm not too familiar with them, but a talk I saw recently had some interesting material on this (<a href=\"https:\/\/wsc.project.cwi.nl\/woudschoten-conferences\/2016-woudschoten-conference\/PRtalk1.pdf\" rel=\"nofollow noreferrer\">https:\/\/wsc.project.cwi.nl\/woudschoten-conferences\/2016-woudschoten-conference\/PRtalk1.pdf<\/a>, see page 24 and onwards). Computing the whole gradient is usually way too expensive if you have high dimensional input. Instead you just modify a randomly chosen coordinate and take many (many) small, cheap steps each more or less in the right direction rather than going for somehow optimal large, but expensive steps.<\/p>\n\n<h2>Model available and suitable<\/h2>\n\n<p>If you know the model in full and it is possible to write is explicitly as <code>v = f(A)<\/code> then you can compute the gradient of the function <code>f<\/code>. This would be the case if the algorithm you're trying to beat is a linear regression, possibly with multiple layers. The form of the gradient should be easier for you to figure out than for me to write it down here.<\/p>\n\n<p>With this gradient available and fairly cheap to evaluate its value for different images <code>A<\/code> you can proceed with, for example, a steepest descent (or ascent) approach to making the image less recognizable for the algorithm. <\/p>\n\n<h2>Important note<\/h2>\n\n<p>It's probably best not to forget that your approach should not render the image illegible to humans too, that would make it all rather pointless.<\/p>\n\n<ul>\n<li>I think for the purpose of this discussion it's best to see the image as a vector<\/li>\n<\/ul>\n"}]}
{"QuestionId":41309972,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><dimensions>","CreationDate":"2016-12-24T03:07:20.263","AcceptedAnswerId":41315400.0,"OwnerUserId":3205742.0,"Title":"input dimension reshape in Tensorflow conolutional network","Body":"<p>In the expert mnist tutorial in tensorflow website, it have something like this :<\/p>\n\n<pre><\/pre>\n\n<p>I know that the reshape is like <\/p>\n\n<pre><\/pre>\n\n<p>Q1 : why is the batch_size equals -1? What does the -1 means? <\/p>\n\n<p>And when I go down the code there's one more thing I can not understand<\/p>\n\n<pre><\/pre>\n\n<p>Q2:What does the image_size * 64 means?<\/p>\n","answers":[{"AnswerId":"41310724","CreationDate":"2016-12-24T05:50:29.817","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p>When you pass -1 as a dimension in <code>tf.reshape<\/code>, it preserves the existing dimension. From the docs:<\/p>\n\n<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. In\n  particular, a shape of [-1] flattens into 1-D. At most one component\n  of shape can be -1.<\/p>\n<\/blockquote>\n\n<p>The reference to 7 x 7 x 64 is because the convolutional layer being applied prior to this example has reduced the image to a shape of [7, 7, 64], and the input to the next fully connected layer needs to be a single dimension, so in the next line of the example, the tensor is reshaped from [7,7,64] to [7*7*64] so it can connect to the FC layer.<\/p>\n\n<p>For more info on how convolutions and max pooling works, <a href=\"https:\/\/en.wikipedia.org\/wiki\/Convolutional_neural_network\" rel=\"nofollow noreferrer\">the wikipedia page<\/a> has some helpful graphics:<\/p>\n\n<p>e.g. network architecture:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7XiWr.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7XiWr.png\" alt=\"cnn architecture\"><\/a><\/p>\n\n<p>and pooling:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/N9oGg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/N9oGg.png\" alt=\"cnn pooling\"><\/a><\/p>\n"},{"AnswerId":"41315400","CreationDate":"2016-12-24T17:23:33.863","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<blockquote>\n  <p>Q1 : why is the batch_size equals -1? What does the -1 means?<\/p>\n<\/blockquote>\n\n<p>-1 means \"figure this part out for me\". For example, if I run:<\/p>\n\n<pre><code>reshape([1, 2, 3, 4, 5, 6, 7, 8], [-1, 2])\n<\/code><\/pre>\n\n<p>It creates two columns, and whatever number of rows it needs to get everything to fit:<\/p>\n\n<pre><code>array([[1, 2],\n       [3, 4],\n       [5, 6],\n       [7, 8]])\n<\/code><\/pre>\n\n<blockquote>\n  <p>Q2:What does the image_size * 64 means?<\/p>\n<\/blockquote>\n\n<p>It is the number of filters in that particular filter activation. Shapes of filters in conv layers follow the format <code>[height, width, # of input channels (number of filters in the previous layer), # of filters]<\/code>.<\/p>\n"}]}
{"QuestionId":41310637,"AnswerCount":1,"Tags":"<python><tensorflow><windows-10><anaconda><cudnn>","CreationDate":"2016-12-24T05:33:49.407","AcceptedAnswerId":41314871.0,"OwnerUserId":7336667.0,"Title":"No OpKernel was registered to support Op 'Conv2D' with these attrs","Body":"<p>new to this may be something dumb but cant get conv2d to run<\/p>\n\n<hr>\n\n<p>windows 10<\/p>\n\n<p>anaconda 4.2.13<\/p>\n\n<p>python 3.5.2 <\/p>\n\n<p>C:\\windows\\system32>nvcc --version\nnvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2016 NVIDIA Corporation\nBuilt on Sat_Sep__3_19:05:48_CDT_2016\nCuda compilation tools, release 8.0, V8.0.44<\/p>\n\n<p>cudnn 5.1<\/p>\n\n<p>TensorFlow 0.12<\/p>\n\n<hr>\n\n<pre><\/pre>\n\n<p>results in:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41314871","CreationDate":"2016-12-24T16:16:06.783","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You should change two lines to<\/p>\n\n<pre><code>f=tf.constant(   np.ones(10).reshape(1,1,-1,1).astype(np.float32)   )\ng=tf.constant(   np.ones(3).reshape(1,-1,1,1).astype(np.float32)   )\n<\/code><\/pre>\n\n<p>Otherwise those nodes take on default numpy type of <code>float64<\/code>, which doesn't have <code>Conv2D<\/code> kernel<\/p>\n"}]}
{"QuestionId":41310775,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-24T06:02:11.873","AcceptedAnswerId":null,"OwnerUserId":3102085.0,"Title":"shape must be rank 4 but is rank 5","Body":"<p>I am max pooling over the tensor. But instead of using the built-in tf.nn.max_pool() I have to use tf.reduce_max(). But it gives the error:<\/p>\n\n<pre><\/pre>\n\n<p>And here is the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41311277","CreationDate":"2016-12-24T07:24:22.270","ParentId":null,"OwnerUserId":"6693951","Title":null,"Body":"<p>You may try to remove the <code>[]<\/code> in the first parameter passed into <code>tf.nn.conv2d<\/code>. So instead of <code>conv = tf.nn.conv2d([self.pool1], kernel, [1, 1, 1, 1], padding='SAME')<\/code>, you should try this: <code>conv = tf.nn.conv2d(self.pool1, kernel, [1, 1, 1, 1], padding='SAME')<\/code><\/p>\n"}]}
{"QuestionId":41312030,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-24T09:31:10.027","AcceptedAnswerId":null,"OwnerUserId":7337149.0,"Title":"Value doesn't change using sess.run() in Tensorflow","Body":"<p>I got a problem while doing some simple compute, my code is like below:<\/p>\n\n<pre><\/pre>\n\n<p>and it print out like:<\/p>\n\n<pre><\/pre>\n\n<p>It looks like  is combined of , , .<\/p>\n\n<p>Why isn't : ?<\/p>\n","answers":[{"AnswerId":"41315372","CreationDate":"2016-12-24T17:18:45.120","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>I ran this code:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nsess = tf.Session()\n\nx_k = np.array([[1.0]])\nalpha_T = np.array([[1.60326695]])\n\nZeta = 0.05\nLambda = 2**20\n\nstep1 = sess.run(tf.matmul(x_k,alpha_T))\nstep2 = Lambda*step1\nstep3 = Zeta+step2\nstep1_2_3 = sess.run(Zeta + Lambda*tf.matmul(x_k,alpha_T))\n\nprint('step1', step1)\nprint('step2', step2)\nprint('step3', step3)\nprint('step1_2_3', step1_2_3)\n<\/code><\/pre>\n\n<p>which produced this:<\/p>\n\n<pre><code>('step1', array([[ 1.60326695]]))\n('step2', array([[ 1681147.2453632]]))\n('step3', array([[ 1681147.2953632]]))\n('step1_2_3', array([[ 1681147.2953632]]))\n<\/code><\/pre>\n\n<p>Seems like they're the same for me. Maybe there is something in some code that you didn't include.<\/p>\n"}]}
{"QuestionId":41312197,"AnswerCount":2,"Tags":"<python><computer-vision><tensorflow><conv-neural-network>","CreationDate":"2016-12-24T09:54:04.587","AcceptedAnswerId":41315329.0,"OwnerUserId":6631259.0,"Title":"Number of classes for inception network (Tensorflow)","Body":"<p>I see that a background class is used as a bonus class. So this class is used in case of not classifying an image in the other classes? In my case, I have a binary problem and I want to understand if an image contains food or not. I need to use 2 classes + 1 background class = 3 classes or only 2 classes?<\/p>\n","answers":[{"AnswerId":"41315638","CreationDate":"2016-12-24T17:51:45.373","ParentId":null,"OwnerUserId":"2839624","Title":null,"Body":"<p>Using 2 classes should work for you.<\/p>\n\n<p>In case you are quickly trying to train a neural net and see if it works for your use case, I have created a demo which will require very less number of training examples per class and will train very quickly. I have hosted it <a href=\"http:\/\/demo.nanonets.ai\/\" rel=\"nofollow noreferrer\">here<\/a> <\/p>\n"},{"AnswerId":"41315329","CreationDate":"2016-12-24T17:13:23.373","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>Only two classes. \"Not food\" <em>is<\/em> your background class. If you were trying to detect food or dogs, you could have 3 classes: \"food\", \"dog\", \"neither food nor dog\".<\/p>\n"}]}
{"QuestionId":41312201,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-24T09:54:35.667","AcceptedAnswerId":null,"OwnerUserId":2913150.0,"Title":"Tensorflow feeding in images is extremely slow","Body":"<p>I'm trying to wrap my head around feeding in a custom dataset. The Tensorflow <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\" title=\"tutorial\">tutorial<\/a> gave an object where I could just request another batch of training data and feed that in, so I tried to do something similar. I created a dataset object ds, which exposes the function get_alphabet which uses a WholeFileReader to read a png, and uses tf.image.decode_png to get a Tensor. It returns a list of Tensors as data (train_x), and a list of tensors as one_hot labels (train_y). The problem is that evaluating these images takes a ridiculous amount of time, and it eventually gets to the point where I'm waiting 10 seconds to see the \"2\" printed after the \"1\" gets printed. How do I speed this up?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47968190","CreationDate":"2017-12-25T11:30:27.780","ParentId":null,"OwnerUserId":"7433878","Title":null,"Body":"<h1>Use TFRecord with new Dataset API(released as part of TF 1.4 version) to resolve your problem<br><\/h1>\n\n<h1>TFRecords:<\/h1>\n\n<pre><code>-TensorFlow's own binary format\n-Its format uses a mixture of its Records format and and Protocol Buffers(or Protobuf)\n-Record is simply a binary file that contains tf.train.Example Protobuf Objects\n-Protocol Buffers, it as a way to serialize data structures, given some schema describing what the data is.\n<\/code><\/pre>\n\n<h2>Advantage of TFRecords:<\/h2>\n\n<pre><code>-They make better use of disk cache.\n-They are faster to move around.\n-They can store data of different types (so you can put both images and labels in one place)\n<\/code><\/pre>\n\n<p><br>First convert everything to TFRecords(i.e. Both Image and corresponding labels into TFrecord) file, Then you can use TFRecordDataset API to retrieve the data with batching from TFrecord file<\/p>\n\n<p><strong>Steps to creating TFrecords<\/strong><br><\/p>\n\n<p>1) Use tf.python_io.TFRecordWriter to open the tfrecords file and start writing <br>\n2) Before writing into tfrecords file,the image data and label data should be converted into proper datatype(like float)<br>\n3) Now the datatypes are converted into tf.train.Feature<br>\n4) Finally create an Example Protocol Buffer using tf.Example and use the converted features into it.<br>\n5) Serialize the Example using serialize() function.<br>\n6) Write the serialized Example <br><\/p>\n\n<p>Eg code: <\/p>\n\n<pre><code>from matplotlib.image import imread\ntfrecord_filename = 'TFrecord.tfrecord'\n\nwriter = tf.python_io.TFRecordWriter(tfrecord_filename)\nsess = tf.Session()\ninput_files = _get_file_names() # get list of image file names\nfor input_file in input_files:\n    img = imread(input_file)\n    #tf.image.resize_image_with_crop_or_pad - used to resize image of different size\n    img = sess.run(tf.image.resize_image_with_crop_or_pad(img, \n                  target_height=480, target_width=640))\n    label = &lt;get_correspondinglabels&gt;\n    input_feature = { 'label': _int64_feature(label),\n                  'image': _bytes_feature(img.tostring()) }\n    feature = tf.train.Features(feature=input_feature)\n    example = tf.train.Example(features=feature)\n    writer.write(example.SerializeToString())\nwriter.close()\n\ndef _int64_feature(value):\n     return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\ndef _bytes_feature(value):\n     return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n<\/code><\/pre>\n\n<p><strong>Steps for reading from TFRecords:<\/strong><br><\/p>\n\n<p>1) Read the TFRecord file name <br>\n2) Create TFRecordDataset by providing the TFrecord file name<br>\n3) Create the Parse function for decoding and do any preprocessing work in input data<br>\n4) Create batching, repeat(no of epoch) and shuffling using dataset created in previous steps<br>\n5) Create the Iterator to fetch required inputs as a batch(i.e. mini-batch)<br><\/p>\n\n<p>Eg code: <\/p>\n\n<pre><code>def input_model_function():\n    dataset = tf.data.TFRecordDataset(tfrecord_filename)\n    dataset = dataset.map(_parse_function)\n    dataset = dataset.batch(20)# you can use any number of batching\n    iterator = dataset.make_one_shot_iterator()\n    sess = tf.Session()\n    batch_images, batch_labels = sess.run(iterator.get_next())\nreturn {'x':batch_images}, batch_labels\n\ndef _parse_function(example_proto):\n    parsed_features = tf.parse_single_example(example_proto, tfrecord_features)\n    # Get the image as raw bytes.\n    image_raw = parsed_features['image']\n    # Decode the raw bytes so it becomes a tensor with type.\n    image = tf.decode_raw(image_raw, tf.uint8)\n    # The type is now uint8 but we need it to be float.\n    image = tf.cast(image, tf.float32)\n    #Get the label associated with the image.\n    label = parsed_features['label']\nreturn image, label\n<\/code><\/pre>\n\n<p>Finally input batched Dataset into Model(created using any Pre-Made Estimator or Custom Estimator API)<\/p>\n"}]}
{"QuestionId":41312644,"AnswerCount":0,"Tags":"<python><opencv><tensorflow><histogram><scikit-image>","CreationDate":"2016-12-24T10:56:56.023","AcceptedAnswerId":null,"OwnerUserId":1417053.0,"Title":"Tensorflow - Histogram Loss Function","Body":"<p>I'm new to Tensorflow and I'm trying to design a Loss Function to account for the difference in the 3-Channel Histogram of the Input image and the Output image.<\/p>\n\n<p>I know the L1 loss in Tensorflow is like  but don't know how to write it to compare Histograms.<\/p>\n\n<p>Thanks...<\/p>\n","answers":[]}
{"QuestionId":41312843,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-12-24T11:29:02.647","AcceptedAnswerId":41315307.0,"OwnerUserId":7337373.0,"Title":"How can I Training my data set which composed only True? in Python","Body":"<p>I am a beginner in python and I'm studying machine learning in python using tensorflow.<\/p>\n\n<p>I have two data set 'trainset.csv' 'testset.csv'<\/p>\n\n<p>trainset.csv like(just example):<\/p>\n\n<pre><\/pre>\n\n<p>tesetset.csv like(just example):<\/p>\n\n<pre><\/pre>\n\n<p>How can I train my train set data which only presented normal feature and classify abnormal data in test set?<\/p>\n","answers":[{"AnswerId":"41315307","CreationDate":"2016-12-24T17:09:15.950","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>I think you need abnormal training examples, or else the network will learn to output \"normal\" every single time. If you want to do anomaly detection with no anomalies in the training set, I'm not sure a neural network is the best approach.<\/p>\n"},{"AnswerId":"41313142","CreationDate":"2016-12-24T12:11:31.517","ParentId":null,"OwnerUserId":"1744834","Title":null,"Body":"<p>I'm using <a href=\"http:\/\/pandas.pydata.org\" rel=\"nofollow noreferrer\"><code>pandas<\/code><\/a> library here because it's quite convenient to use with data. Concerning your problem - I think you have to create new train dataset from all possible combinations of values and then train on that.<\/p>\n\n<p>Let's say you have a dictionary with all possible values for all columns. I fake it here by getting all possible combinations from your train dataframe:<\/p>\n\n<pre><code>&gt;&gt;&gt; import pandas as pd\n&gt;&gt;&gt; del df['abnormal']\n&gt;&gt;&gt; cols = [x for x in df.columns if x != 'normal']\n&gt;&gt;&gt; cols\n['1', '2', '3', '4', '5', '6', '7', '8', '9']\n&gt;&gt;&gt; d = df[cols].apply(lambda x: list(x.unique())).to_dict()\n&gt;&gt;&gt; d\n{'1': [0], '3': [2, 1], '2': [1, 2], '5': [1, 4], '4': [3, 1], '7': [1, 3], '6': [2, 1, 4], '9': [6, 1, 2], '8': [4, 2, 1]}\n<\/code><\/pre>\n\n<p>Then you can create full dataframe from this:<\/p>\n\n<pre><code>&gt;&gt;&gt; df_full = pd.DataFrame(list(itertools.product(*d.values())), columns=d.keys())\n<\/code><\/pre>\n\n<p>And now you can create new train dataframe:<\/p>\n\n<pre><code>&gt;&gt;&gt; df_train_new = pd.merge(df_full, df, how='left').fillna(0)\n<\/code><\/pre>\n\n<p>So it will have all possible combinations of values + filled <code>normal<\/code> column<\/p>\n"}]}
{"QuestionId":41313105,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-24T12:06:21.280","AcceptedAnswerId":null,"OwnerUserId":1834160.0,"Title":"Some of tensorflow GPU OpKernel compute by eigen device without stream sync, is that buggy?","Body":"<p>from gpu_device.cc<\/p>\n\n<p>\nand gpu_device only waits when different context. (sync_every_op is false)<\/p>\n\n<p>But in argmax_op.h, for example,<\/p>\n\n<pre><\/pre>\n\n<p>use device compute directly. Is that correct?<\/p>\n","answers":[{"AnswerId":"41322808","CreationDate":"2016-12-25T16:34:39.783","ParentId":null,"OwnerUserId":"1834160","Title":null,"Body":"<p>I missed something. cuda stream is passed to eigen device. so there's no problem<\/p>\n"}]}
{"QuestionId":41313970,"AnswerCount":1,"Tags":"<python><macos><caffe><pycaffe>","CreationDate":"2016-12-24T14:09:57.140","AcceptedAnswerId":null,"OwnerUserId":3550694.0,"Title":"Issues during 'make pycaffe' on OSX","Body":"<p>I am getting the below error while trying to run  on after installing caffe on mac.<\/p>\n\n<pre><\/pre>\n\n<p>Below is the part of my Makefile.config for reference:\n    # We need to be able to find Python.h and numpy\/arrayobject.h. <\/p>\n\n<p>PYTHON_INCLUDE := \/usr\/include\/python2.7 \\\n                    \/usr\/lib\/python2.7\/dist-packages\/numpy\/core\/include<\/p>\n\n<p>Though I am not able to find \/usr\/lib\/python2.7\/dist-packages\/ on my machine, please give suggestions<\/p>\n","answers":[{"AnswerId":"44295722","CreationDate":"2017-05-31T22:35:39.097","ParentId":null,"OwnerUserId":"4242864","Title":null,"Body":"<p>Ran into similar issue while compiling for pycaffe<\/p>\n\n<pre><code>CXX\/LD -o python\/caffe\/_caffe.so python\/caffe\/_caffe.cpp\npython\/caffe\/_caffe.cpp:10:10: \n<\/code><\/pre>\n\n<p>fatal error: <\/p>\n\n<blockquote>\n  <p>numpy\/arrayobject.h file not found<\/p>\n<\/blockquote>\n\n<p>This indicates that numpy is not installed, I used the following command to fix my issue.<\/p>\n\n<pre><code>$ conda install -c anaconda numpy=1.12.1\n<\/code><\/pre>\n"}]}
{"QuestionId":41314316,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-12-24T14:56:40.803","AcceptedAnswerId":41315284.0,"OwnerUserId":7277945.0,"Title":"Tensorflow MNIST: terminate called after throwing an instance of 'std::bad_alloc'","Body":"<p>I am trying to implement a Convolutional Neural Network on Tensorflow, using their default MNIST data set.<\/p>\n\n<pre><\/pre>\n\n<p>On executing, python crashes with this message:\n    terminate called after throwing an instance of 'std::bad_alloc'\n    what():  std::bad_alloc<\/p>\n\n<p>I was able to point out that this happens when I call the  function, or in general, when I load the whole mnist.test iamges and labels.\nAny suggestions on what can be done, given I wish to use this data. I have been able to work with images as a whole, in a different case.<\/p>\n","answers":[{"AnswerId":"47010993","CreationDate":"2017-10-30T08:42:56.227","ParentId":null,"OwnerUserId":"8855190","Title":null,"Body":"<p>I had the same issue. I resolved it by reducing the number of test images to compute accuracy, e.g. I replaced <\/p>\n\n<pre><code>print(compute_accuracy(mnist.test.images, mnist.test.labels))\n<\/code><\/pre>\n\n<p>With something similar to <\/p>\n\n<pre><code>batch_test = mnist.test.next_batch(5000)\nprint(compute_accuracy(batch_test[0], batch_test[1])\n<\/code><\/pre>\n\n<p>I hope this helps.<\/p>\n"},{"AnswerId":"41315284","CreationDate":"2016-12-24T17:05:15.703","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>I think you're running out of memory. It runs okay on my machine (6GB graphics card). Try decreasing the batch size, or using a smaller fully connected layer.<\/p>\n"}]}
{"QuestionId":41314819,"AnswerCount":2,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-12-24T16:08:11.587","AcceptedAnswerId":null,"OwnerUserId":3860898.0,"Title":"Why is embedding_lookup better than one hot encoding with a linear transformation?","Body":"<p>I think I'm missing something obvious here but would love some help figuring this out. <\/p>\n\n<p>Say I have a million words and want to embed them as part of my model. \nWith TF I can do an embedding lookup, though I need to provide a matrix of size [1m*space_size]. So for 50 dimensions that comes out to 50M trainable parameters. \nOn the other hand I can one hot encode the a million words with a vector of dimension 20. I can embed that into a a space of dimension 50 with a [20*50] matrix for 1K parameters. Much cheaper. Since the weights of this matrix are still trainable, I'd expect to learn something about the words and if I need more capacity I can increase the size of the space. <\/p>\n\n<p>That's in theory, in practice I tried and the model didn't learn anything. So my question is, why? \nThanks<\/p>\n","answers":[{"AnswerId":"41335417","CreationDate":"2016-12-26T19:19:35.393","ParentId":null,"OwnerUserId":"2878004","Title":null,"Body":"<p>It is true that you can encode a million words with a 20 digit binary number. But one-hot encoding is not binary. It means that only a single digit in the whole vector can be one and the rest are zero. That digit indicates the index of the word. So you would need a 2 million length vector to encode 2 million words. Then you would need a matrix of 2 million x 50 for the model parameters.<\/p>\n\n<p>Using the 20 digit binary representation is problematic because you are arbitrarily assigning certain words to be closer to each other in the 20-dim space. It's better to let the model learn these associations rather than pre-defining them.<\/p>\n"},{"AnswerId":"41315489","CreationDate":"2016-12-24T17:33:52.760","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Stochastic Gradient Descent seems to require significant overparameterization in order to learn, here's one paper along those lines -- <a href=\"https:\/\/arxiv.org\/abs\/1301.3583\" rel=\"nofollow noreferrer\">\"Big Neural Networks Waste Capacity\"<\/a><\/p>\n"}]}
{"QuestionId":41315240,"AnswerCount":0,"Tags":"<python><c++><tensorflow>","CreationDate":"2016-12-24T16:59:33.143","AcceptedAnswerId":null,"OwnerUserId":5280139.0,"Title":"Using saved Tensorflow Estimator with C++ API","Body":"<p>I have written the Abalone estimator in Python as described in <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/estimators\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/estimators\/<\/a>. I wish to save the state of the estimator, then load it in C++ and use it to make predictions.<\/p>\n\n<p>To save it from Python, I use the  parameter in the  constructor, which creates a (text) protobuf file and several checkpoint files. I then use the  tool (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py<\/a>) to combine the checkpoint and the protobuf file into a standalone GraphDef file.<\/p>\n\n<p>I load this file using the C++ API, load some input values into a Tensor, then run the session. The input node in the protobuf file is called 'input' and the output node 'output', and both are placeholder nodes.<\/p>\n\n<pre><\/pre>\n\n<p>However, since the output node is a placeholder node, this fails since it needs to be fed a value. But you cannot both feed and fetch a node value, so I cannot get access to the output of the estimator. Why is the output node a placeholder node?<\/p>\n\n<p>What is the best way to save a trained estimator from Python and load it for prediction in C++? <\/p>\n","answers":[]}
{"QuestionId":41315365,"AnswerCount":0,"Tags":"<python><theano><keras><loss><resuming-training>","CreationDate":"2016-12-24T17:18:00.940","AcceptedAnswerId":null,"OwnerUserId":1934283.0,"Title":"Keras - Compile model when using theano backend (K.function) to calculate loss","Body":"<p>I am building an NN in Keras and I am using Theano backend.<\/p>\n\n<p>Code:<\/p>\n\n<pre><\/pre>\n\n<p>If I compile this model using the above-mentioned parameters and save it, will the loaded model preserve the training configuration (loss,optim.)?<\/p>\n\n<p>Since I use the K.function(...) method to calculate cost(loss), will this affect the model's configuration in any way?<\/p>\n\n<p>So far I have not been able to figure out a way to check whether the resumed model's config is same as the saved one.\nSo, is there a way to print out the network configuration to check if it has been properly restored i.e. with the exact same config and param it was saved with?<\/p>\n","answers":[]}
{"QuestionId":41316739,"AnswerCount":0,"Tags":"<multithreading><lua><torch>","CreationDate":"2016-12-24T20:48:06.743","AcceptedAnswerId":null,"OwnerUserId":3787518.0,"Title":"Torch Multithreading","Body":"<p>I have a table H_next and I do some calculation (call add()) and store value in H_next, which is shared across all the threads. <\/p>\n\n<pre><\/pre>\n\n<p>This just prints 2 values<\/p>\n\n<p>I am only able to get all the values when I modify the function as follows<\/p>\n\n<pre><\/pre>\n\n<p>The problem with 2nd approach is that it takes a lot of time to communicate between child and main thread, I want to fill the H_next table for that particular thread then transfer those values or copy those values in main thread in temp_H_next using endcallback. \nIs there a better way to do this, where I don't have to send one by one value at a time?<\/p>\n","answers":[]}
{"QuestionId":41317117,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><skflow>","CreationDate":"2016-12-24T21:56:19.133","AcceptedAnswerId":null,"OwnerUserId":4055043.0,"Title":"Cannot achieve repeatability in tensorflow","Body":"<p>I am trying to learn tensorflow with <em>High level API for learning with TensorFlow<\/em> and having problem with not getting repeatable results.<\/p>\n\n<p>TensorFlow 0.12.0-rc0 (CPU Only)\npython 3.5<\/p>\n\n<pre><\/pre>\n\n<p>As you see, I am trying to put random seed MY_SEED wherever possible, but results are different from run to run.<\/p>\n\n<p>What did I miss?<\/p>\n","answers":[{"AnswerId":"47378535","CreationDate":"2017-11-19T15:26:46.983","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Per keveman's comment, this seems to have been a bug in the (now very old) 0.12-RC release candidate release.  There were some bugs related to deterministic random seed setting that have been fixed.<\/p>\n"}]}
{"QuestionId":41317928,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-12-25T00:40:31.543","AcceptedAnswerId":41322847.0,"OwnerUserId":6423473.0,"Title":"Installing TensorFlow with Pip Python on Windows","Body":"<p>Last month they released tensor-flow comparability with windows. Looking at the docs I've installed python 3.6 and run<\/p>\n\n<pre><\/pre>\n\n<p>but it doesn't find it and therefore doesn't install it.<\/p>\n\n<pre><\/pre>\n\n<p>Am I missing something basic? How can I fix or circumvent this?<\/p>\n","answers":[{"AnswerId":"41321493","CreationDate":"2016-12-25T13:26:35.440","ParentId":null,"OwnerUserId":"6802252","Title":null,"Body":"<p>I think you have to install Python 3.5 not 3.6. I'm having the same problem. <\/p>\n"},{"AnswerId":"41322847","CreationDate":"2016-12-25T16:41:31.297","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>A stable release of Python 3.6 for Windows became available on 12\/23\/2016, and we have not yet built TensorFlow packages for that version. (We will look into doing this after the holidays.) For now, your best options are:<\/p>\n\n<ol>\n<li>Downgrade to Python 3.5 (64-bit version), which the pre-built packages support.<\/li>\n<li>Try <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/cmake\" rel=\"noreferrer\">building TensorFlow from source<\/a>.<\/li>\n<\/ol>\n"},{"AnswerId":"42084471","CreationDate":"2017-02-07T07:53:42.517","ParentId":null,"OwnerUserId":"2784721","Title":null,"Body":"<p>Here is what i did to get tensorflow working with windows.<\/p>\n\n<p><a href=\"https:\/\/www.python.org\/ftp\/python\/3.5.0\/python-3.5.0-amd64.exe\" rel=\"noreferrer\">Download<\/a> python 3.5.0 (64 bit from <a href=\"https:\/\/www.python.org\/getit\/windows\/\" rel=\"noreferrer\">Python Releases for Windows<\/a>)\nInstall it and check python version by running below command in cmd:<\/p>\n\n<pre><code>python --version\nPython 3.5.0\n<\/code><\/pre>\n\n<p>Then run below command to upgrade pip to latest<\/p>\n\n<pre><code>python -m pip install --upgrade pip\n<\/code><\/pre>\n\n<p>Now install tensorflow using pip<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>That's it you have installed tensorflow on windows. Below image shows what happens when you type above commands\n<a href=\"https:\/\/i.stack.imgur.com\/Fh5Mp.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Fh5Mp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#test_the_tensorflow_installation\" rel=\"noreferrer\">Example<\/a> to verify tensorflow is working.<\/p>\n\n<pre><code>$ python\n...\nimport tensorflow as tf\nhello = tf.constant('Hello, TensorFlow!')\nsess = tf.Session()\nprint(sess.run(hello))\n<\/code><\/pre>\n"}]}
{"QuestionId":41318471,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-25T03:27:13.890","AcceptedAnswerId":null,"OwnerUserId":5350430.0,"Title":"Too slow for converting to tf.constant when list contains 1000000 elems","Body":"<p>What is the best way to do the training. That's too slow! And I don't know why it is that slow. <\/p>\n\n<pre><\/pre>\n\n<p>The samples_all contains more than 1000000 string elems which represent the image paths. The first two lines in the code take too long to execute.<\/p>\n","answers":[{"AnswerId":"41335240","CreationDate":"2016-12-26T18:58:40.027","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Below an example of loading 2M strings into string variable which takes less than 1 second on my MacBook. Variables are more efficient for large constants than <code>tf.constant<\/code> which are inlined in the graph. Note that there's also <code>tf.string_input_producer<\/code> which can handle large lists of strings without loading them all into TensorFlow memory.<\/p>\n\n<pre><code>n = 2000000\nstring_list = np.asarray([\"string\"]*n);\n\nsess = tf.Session()\nvar_input = tf.placeholder(dtype=tf.string, shape=(n,))\nvar = tf.Variable(var_input)\n\nstart = time.time()\n# var.initializer is equivalent to var.assign(var_input).op\nsess.run(var.initializer, feed_dict={var_input: string_list})\nelapsed = time.time()-start\nrate = n*6.\/elapsed\/10**6    \nprint(\"%.2f MB\/sec\"%(rate,))  # =&gt; 15.13 MB\/sec\n<\/code><\/pre>\n"}]}
{"QuestionId":41320441,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><neural-network><deep-learning>","CreationDate":"2016-12-25T10:41:57.377","AcceptedAnswerId":null,"OwnerUserId":6423822.0,"Title":"how test a model made tensorflow and python following the code","Body":"<pre><\/pre>\n\n<p>If I remove the code related to the test file it works fine. But when I add code to test the model then this is the output of the above code<\/p>\n\n<pre><\/pre>\n\n<p><strong>CONSOLE output:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>the program gets stuck<\/p>\n","answers":[]}
{"QuestionId":41320964,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><theano><keras>","CreationDate":"2016-12-25T12:04:37.167","AcceptedAnswerId":null,"OwnerUserId":1832454.0,"Title":"input_shape parameter mismatch error in Convolution1D in keras","Body":"<p>I want to classify a  dataset using Convulation1D in keras.<\/p>\n\n<p><strong>DataSet Description<\/strong>:<\/p>\n\n<p>train dataset size = [340,30] ; no of sample = 340 , sample dimension = 30<\/p>\n\n<p>test dataset size = [230,30] ; no of sample = 230 , sample dimension = 30<\/p>\n\n<p>label size = 2   <\/p>\n\n<p>Fist I try by the following code  using the information from keras site <a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/convolutional\/<\/a><\/p>\n\n<pre><\/pre>\n\n<p>it gives the following error ,\n<strong>ValueError: Error when checking model input: expected convolution1d_input_1 to have 3 dimensions, but got array with shape (340, 30)<\/strong><\/p>\n\n<p>Then I have transformed the Train and Test data into 3 dimension from 2 dimension by using the following code ,<\/p>\n\n<pre><\/pre>\n\n<p>Then I  run the modified following code ,<\/p>\n\n<pre><\/pre>\n\n<p>But it shows the error ,\n<strong>ValueError: Error when checking model input: expected convolution1d_input_1 to have shape (None, 340, 30) but got array with shape (340, 30, 1)<\/strong><\/p>\n\n<p>I am unable to find the dimension mismatch error here. <\/p>\n","answers":[{"AnswerId":"41344197","CreationDate":"2016-12-27T11:18:47.883","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>Can you try this?<\/p>\n\n<pre><code>X_train = np.reshape(X_train_transformed, (1, X_train_transformed.shape[0], X_train_transformed.shape[1]))\nX_test = np.reshape(X_test_transformed, (1, X_test_transformed.shape[0], X_test_transformed.shape[1]))\n<\/code><\/pre>\n"}]}
{"QuestionId":41321215,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><lstm>","CreationDate":"2016-12-25T12:41:24.087","AcceptedAnswerId":null,"OwnerUserId":4726538.0,"Title":"Why does tensorflow saver takes so much time?","Body":"<p>I am using  to do speech recognition, and all parameters of my model is five millions. When I use  to save my Model, my pc gets stuck and I waited for several hours but it didn't finished. I don't know why?<\/p>\n","answers":[]}
{"QuestionId":41322161,"AnswerCount":0,"Tags":"<python><mingw><theano>","CreationDate":"2016-12-25T14:57:19.773","AcceptedAnswerId":null,"OwnerUserId":6433475.0,"Title":"Can't import Theano on Windows 10 with Anaconda's Python3.5","Body":"<p>After installing mingw libpython via conda and theano via pip, I tried to import theano and got this<\/p>\n\n<pre><\/pre>\n\n<p>This error comes from cutils.py when it tries to run:<\/p>\n\n<pre><\/pre>\n\n<p>So I checked the cutils_ext folder in my AppData\/local\/Theano\/cutils_ext and found cutils_ext.pyd, then renamed it to cutils_ext.dll, but that doesn't seem to fix the problem.\nI also used Dependency Walker and found some missing modules but there were too many. Do I have to download all of those .dll files or there's a faster way to get it done?<\/p>\n","answers":[]}
{"QuestionId":41322243,"AnswerCount":2,"Tags":"<python><theano><keras><recurrent-neural-network>","CreationDate":"2016-12-25T15:08:48.447","AcceptedAnswerId":null,"OwnerUserId":996366.0,"Title":"How to use keras RNN for text classification in a dataset?","Body":"<p>I have coded ANN classifiers using keras and now I am learning myself to code RNN in keras for text and time series prediction. After searching  a while in web I found this <a href=\"http:\/\/machinelearningmastery.com\/sequence-classification-lstm-recurrent-neural-networks-python-keras\/\" rel=\"nofollow noreferrer\">tutorial<\/a> by  Jason Brownlee which is decent for a novice learner in RNN. The original article is using IMDb dataset for text classification with LSTM but because of its large dataset size I changed it to a small sms spam detection dataset.<\/p>\n\n<pre><\/pre>\n\n<p>I have successfully processed the dataset into training and testing set but now how should I model my RNN for this dataset?<\/p>\n","answers":[{"AnswerId":"43720011","CreationDate":"2017-05-01T13:37:15.397","ParentId":null,"OwnerUserId":"4371189","Title":null,"Body":"<p>You need to represent <code>raw text<\/code> data as <code>numeric vector<\/code> before training a neural network model. For this, you can use <code>CountVectorizer<\/code> or <code>TfidfVectorizer<\/code> provided by <code>scikit-learn<\/code>. After converting from raw text format to numeric vector representation, you can train a RNN\/LSTM\/CNN for text classification problem.<\/p>\n"},{"AnswerId":"42453806","CreationDate":"2017-02-25T08:38:20.090","ParentId":null,"OwnerUserId":"6627465","Title":null,"Body":"<p>If you are still stuck on this, <a href=\"http:\/\/machinelearningmastery.com\/sequence-classification-lstm-recurrent-neural-networks-python-keras\/\" rel=\"nofollow noreferrer\">check out this example by Jason Brownlee<\/a>.  Looks like you are most of the way there.  You need to add an LSTM layer and a Dense layer to get a model that should work.<\/p>\n"}]}
{"QuestionId":41322286,"AnswerCount":2,"Tags":"<tensorflow><keras><conv-neural-network>","CreationDate":"2016-12-25T15:16:10.653","AcceptedAnswerId":null,"OwnerUserId":5215538.0,"Title":"Training a fully convolutional neural network with inputs of variable size takes unreasonably long time in Keras\/TensorFlow","Body":"\n\n<p>I am trying to implement a FCNN for image classification that can accept inputs of variable size. The model is built in Keras with TensorFlow backend. <\/p>\n\n<p>Consider the following toy example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This model runs fine but I am running into a performance issue. Training on images of variable size takes unreasonably long time compared to training on the inputs of fixed size. If I resize all images to the maximum size in the data set it still takes far less time to train the model than training on the variable size input. So is  the right way to specify variable size input? And is there any way to mitigate this performance problem? <\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p> for a model with 3 classes and grayscale images:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"57823618","CreationDate":"2019-09-06T14:07:18.563","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>I think @marcin-mo\u017cejko may have the right answer in his comment.\nIt may be related to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/29075\" rel=\"nofollow noreferrer\">this bug<\/a>, which was just fixed. And <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/32258\" rel=\"nofollow noreferrer\">this patch<\/a> may warn you if things are being compiled too often.<\/p>\n\n<p>So upgrading to a tf-nightly-gpu-2.0-preview package may fix this.\nAlso do you get this problem with <code>tf.keras<\/code>.<\/p>\n\n<blockquote>\n  <p>If I resize all images to the maximum size in the data set it still takes far less time to train the model than training on the variable size input<\/p>\n<\/blockquote>\n\n<p>Note that for basic convolutions with \"same\" padding, zero padding should have \"no\" effect on the output, aside from pixel alignment.<\/p>\n\n<p>So one approach would be to train on a fixed list of sizes and zero pad images to those sizes. For example and train on batches of 128x128, 256x256, 512x512. If you can't fix the dynamic compilation thing this at least would only compile it 3 times. This would be a bit like a 2d \"bucket-by-sequence-length\" approach sometimes seen with sequence models.<\/p>\n"},{"AnswerId":"52572567","CreationDate":"2018-09-29T21:09:32.917","ParentId":null,"OwnerUserId":"8670966","Title":null,"Body":"<p>Images of different sizes implies images of similar things at a different scale. If this difference in scale is significant the relative position of the similar things will shift from being in the centre of the frame towards the top left as the image size reduces. The (simple) network architecture shown is spatially aware so it would be consistent for the rate of model convergence to degrade as data of a very different scale would be inconsistent. This architecture is not well suited to finding the same thing in different or multiple places.<\/p>\n\n<p>A certain degree of shearing, rotation, mirroring would help the model generalise, but re-scaled to a consistent size. So, when you re-size you fix the scaling issue and make the input data spatially consistent.<\/p>\n\n<p>In short, I think it\u2019s that this network architecture is not suited \/ capable for the task you are giving it i.e. various scales.<\/p>\n"}]}
{"QuestionId":41323292,"AnswerCount":1,"Tags":"<text><theano><keras><lstm><multilabel-classification>","CreationDate":"2016-12-25T17:49:34.700","AcceptedAnswerId":null,"OwnerUserId":1766682.0,"Title":"Multi-label Text classification with Keras\/Theano and LSTM","Body":"<p>Trying to get runing LSTM multi-label text classification with Keras\/Theano.<\/p>\n\n<p>I have a text\/label csv. Text is pure text, labels are numeric, nine in total, from 1 to 9.<\/p>\n\n<p>I think I am not configuring the model properly for this problem. My code so far:<\/p>\n\n<pre><\/pre>\n\n<p>I am getting this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"55832255","CreationDate":"2019-04-24T14:16:30.457","ParentId":null,"OwnerUserId":"8068733","Title":null,"Body":"<p>Please look at the following example for multi-label classification using LSTM.<\/p>\n\n<pre><code>import numpy as np\nfrom sklearn import preprocessing\nfrom keras.models import Sequential\nfrom keras.layers.core import Dense, Activation\nfrom keras.layers.embeddings import Embedding\nfrom keras.layers.recurrent import LSTM\nfrom keras.preprocessing.text import Tokenizer\nfrom keras.preprocessing.sequence import pad_sequences\nfrom keras.layers import Dropout\nfrom keras.optimizers import Adam\nfrom keras.callbacks import EarlyStopping\nfrom keras.callbacks import ModelCheckpoint\n\nX_train = np.array([\"new york is a hell of a town\",\n                \"new york was originally dutch\",\n                \"the big apple is great\",\n                \"new york is also called the big apple\",\n                \"nyc is nice\",\n                \"people abbreviate new york city as nyc\",\n                \"the capital of great britain is london\",\n                \"london is in the uk\",\n                \"london is in england\",\n                \"london is in great britain\",\n                \"it rains a lot in london\",\n                \"london hosts the british museum\",\n                \"new york is great and so is london\",\n                \"i like london better than new york\"])\ny_train_text = [[\"new york\"],[\"new york\"],[\"new york\"],[\"new york\"],    \n                [\"new york\"],[\"new york\"],[\"london\"],[\"london\"],         \n                [\"london\"],[\"london\"],[\"london\"],[\"london\"],\n                [\"new york\",\"England\"],[\"new york\",\"london\"]]\n\nlb = preprocessing.MultiLabelBinarizer(classes=(\"new york\",\"london\",\"England\"))\nY = lb.fit_transform(y_train_text)\n#Y_test = lb.fit_transform(y_test_text)\n\n#print(Y_test)\n\ntokenizer = Tokenizer(num_words=1000, lower=True)\ntokenizer.fit_on_texts(X_train)\nX_train1 = tokenizer.texts_to_sequences(X_train)\n\nmaxlen = 100\nX_train2 = pad_sequences(X_train1, maxlen=maxlen)\n\nfrom sklearn.model_selection import train_test_split\nx_train, x_test, y_train, y_test = train_test_split(X_train2, Y, test_size=0.1, random_state=42)\n\nmax_features=1000\nemb_dim = 32\n\nmodel = Sequential()\nmodel.add(Embedding(max_features, emb_dim, input_length=X_train2.shape[1]))\nmodel.add(LSTM(128))\nmodel.add(Dense(3, activation='sigmoid'))\n# model.compile(optimizer=Adam(0.015), loss='binary_crossentropy', metrics=['acc'])\nmodel.compile(optimizer=Adam(0.015), loss='categorical_crossentropy', metrics=['acc'])\ncallbacks=[EarlyStopping(monitor='val_loss', patience=2, min_delta=0.0001), ModelCheckpoint(filepath='model-simple.h5', save_best_only=True)]\nhist = model.fit(x_train, y_train, epochs=10, batch_size=32, validation_split=0.1, callbacks=callbacks)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>Train on 10 samples, validate on 2 samples\nEpoch 1\/10\n10\/10 [==============================] - 2s 183ms\/step - loss: 1.3170 - acc: 0.6000 - val_loss: 1.0557 - val_acc: 0.5000\nEpoch 2\/10\n10\/10 [==============================] - 0s 27ms\/step - loss: 1.2505 - acc: 0.7000 - val_loss: 1.0371 - val_acc: 0.5000\nEpoch 3\/10\n10\/10 [==============================] - 0s 27ms\/step - loss: 1.2833 - acc: 0.7000 - val_loss: 0.7377 - val_acc: 0.5000\nEpoch 4\/10\n10\/10 [==============================] - 0s 27ms\/step - loss: 1.0988 - acc: 0.7000 - val_loss: 0.8957 - val_acc: 0.5000\nEpoch 5\/10\n10\/10 [==============================] - 0s 28ms\/step - loss: 1.1339 - acc: 0.7000 - val_loss: 0.8717 - val_acc: 0.5000\n<\/code><\/pre>\n\n<p>Note: The performance of the training model should be fine tuned.<\/p>\n"}]}
{"QuestionId":41323486,"AnswerCount":0,"Tags":"<windows><tensorflow><python-3.5>","CreationDate":"2016-12-25T18:18:21.630","AcceptedAnswerId":null,"OwnerUserId":4884530.0,"Title":"Windows python 3.5 Tensorflow cuda file loading","Body":"<p>Here is my error message first i get when trying to import tensorflow:<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>import tensorflow<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>here is what i have installed:<\/p>\n\n<p>windows 10 pro, python 3.5, then installed\n Microsoft Visual C++ 2015 Redistributable Update 3 ( as suggested on tensorflow page)\nthen i installed tensorflow via pip (the gpu version, and it installed sucessfully)\nCuda 8.0\nCudnn 5.1 ( copied the files into my cuda folder)\nmy Path variable looks like this :<\/p>\n\n<pre><\/pre>\n\n<p>I really am clueless where i went wrong :(<\/p>\n","answers":[]}
{"QuestionId":41323576,"AnswerCount":2,"Tags":"<python><c><tensorflow>","CreationDate":"2016-12-25T18:32:18.457","AcceptedAnswerId":41328203.0,"OwnerUserId":3994228.0,"Title":"AttributeError due to sys.argv[0] for 'import tensorflow as tf' inside c","Body":"<p>I have embedded python2.7\/python3.4m code inside c file running on Ubuntu14.4. The python code is imports tensorflow, when it fails. <\/p>\n\n<pre><\/pre>\n\n<p>This gives following output:<\/p>\n\n<pre><\/pre>\n\n<p>Any help in this would be greatly appreciated. I compiled it useing cmake as well as bazel. <\/p>\n\n<p>This problem is happening even inside virtualenv installation. <\/p>\n","answers":[{"AnswerId":"41328203","CreationDate":"2016-12-26T08:01:03.540","ParentId":null,"OwnerUserId":"3994228","Title":null,"Body":"<p>I am so stupid! well this works, as tensorflow needs sys.argv.<\/p>\n\n<pre><code>int\nmain(int argc, char *argv[])\n{\n  Py_SetProgramName(argv[0]);  \/* optional but recommended *\/\n  Py_Initialize();\n  PyRun_SimpleString(\"import sys\\n\"\n                     \"sys.argv = ['']\");\n  PyRun_SimpleString(\"import tensorflow as tf\");\n  Py_Finalize();\n  return 0;\n} \n<\/code><\/pre>\n"},{"AnswerId":"55330600","CreationDate":"2019-03-25T02:37:54.107","ParentId":null,"OwnerUserId":"3603485","Title":null,"Body":"<p>I found the old version (1.4.0) exist this error, but not included the new version (like <code>tensorflow == 1.13.1<\/code>), if you could update the <code>tensorflow<\/code> to the latest version, that may be a better solution.<\/p>\n"}]}
{"QuestionId":41324801,"AnswerCount":1,"Tags":"<deep-learning><keras>","CreationDate":"2016-12-25T21:56:19.817","AcceptedAnswerId":null,"OwnerUserId":364746.0,"Title":"Keras: ValueError: Error when checking model target: expected dense_1 to have shape (None, 10) but got array with shape (10, 1)","Body":"<p>I'm new to keras and I'm facing this error: ValueError: Error when checking model target: expected dense_1 to have shape (None, 10) but got array with shape (10, 1). My input array shape is (10, 1010)<\/p>\n\n<p>and my model is one layer only<\/p>\n\n<pre><\/pre>\n\n<p>any ideas how can I fix this I tried many combinations nothing works for me?<\/p>\n","answers":[{"AnswerId":"43837317","CreationDate":"2017-05-07T22:20:29.717","ParentId":null,"OwnerUserId":"5225453","Title":null,"Body":"<p>A clue is in the fact that the error is complaining about the dimension of the target. I suspect the target is shape <code>(10, 1)<\/code>, whereas the (final) layer is expecting the target to be <code>(None, 10)<\/code>. The model needs the target to be the same dimension as the last layer in the model. After all, if the two aren't the same dimension, how else would Keras be able to do any kinds of comparisons and ultimately give you an accuracy score?<\/p>\n"}]}
{"QuestionId":41327061,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-26T06:06:41.547","AcceptedAnswerId":null,"OwnerUserId":5350430.0,"Title":"How to use two checkpoint files to init the graph","Body":"<p>I've trained two models. Each of them can produce useful representative features. And I want to use the features to train the third model. That means the first two models can produce the inputs to the third model. And I want to use the tf.contrib.learn.train() to train the third model. The question is how should I import the two checkpoint files to the first two models to init the weights. I know there's init_fn and init_op parameters in the function. However, I don't know how to use more than one checkpoint files. Can you help me out? Thank you.<\/p>\n","answers":[{"AnswerId":"41384949","CreationDate":"2016-12-29T18:02:15.663","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>you can use tf.contrib.framework.init_from_checkpoint.<\/p>\n"}]}
{"QuestionId":41327077,"AnswerCount":1,"Tags":"<python><tensorflow><theano><collaborative-filtering>","CreationDate":"2016-12-26T06:08:30.350","AcceptedAnswerId":41343997.0,"OwnerUserId":5403968.0,"Title":"How to make item based collaborative filtering run faster?","Body":"<p>I am trying to find similarity between each pair of items. Items are in a python dictionary and I find the similarity taking pair at a time. The code is - <\/p>\n\n<pre><\/pre>\n\n<p>The problem is that outer loop is taking 5 seconds for each item. I have ~300,000 items so it takes ~18 days for the whole computation. Is there any way to increase the speed? Can I use packages like Theano, Tensorflow and use GPU for this? Or can take a cloud and parallelize the process?<\/p>\n","answers":[{"AnswerId":"41343997","CreationDate":"2016-12-27T11:06:16.433","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>I don't think a machine learning library would be particularly helpful here if there is no operations or building blocks readily available for this type of all to all similarity comparison. <\/p>\n\n<p>I think you'd have better luck by looking at more generic parallelization solutions: OpenMP, TBB, MapReduce, AVX, CUDA, MPI, map reduce, etc.<\/p>\n\n<p>Also, rewriting the same code in C++ will surely speed things up.<\/p>\n"}]}
{"QuestionId":41327286,"AnswerCount":1,"Tags":"<tensorflow><hexagon-dsp>","CreationDate":"2016-12-26T06:34:17.073","AcceptedAnswerId":null,"OwnerUserId":7341293.0,"Title":"build tensorflow with HEXAGON_LIBS","Body":"<p>I want to build tensorflow with qualcomm hexagon libraries, just as the makefile like this <\/p>\n\n<p><\/p>\n\n<pre><\/pre>\n\n<p>but I cannot find the  in the HEXAGON SDK of Qualcomm.<\/p>\n","answers":[{"AnswerId":"41371327","CreationDate":"2016-12-28T23:51:55.007","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>The Hexagon support is still a work in progress, which is why we haven't announced it widely, so unfortunately that library is not yet ready. We hope to have it available soon.<\/p>\n"}]}
{"QuestionId":41327601,"AnswerCount":2,"Tags":"<machine-learning><keras><neural-network><deep-learning><conv-neural-network>","CreationDate":"2016-12-26T07:02:29.820","AcceptedAnswerId":46004661.0,"OwnerUserId":599184.0,"Title":"Why is binary_crossentropy more accurate than categorical_crossentropy for multiclass classification in Keras?","Body":"<p>I'm learning how to create convolutional neural networks using Keras. I'm trying to get a high accuracy for the MNIST dataset.<\/p>\n\n<p>Apparently  is for more than 2 classes and  is for 2 classes. Since there are 10 digits, I should be using . However, after training and testing dozens of models,  consistently outperforms  significantly.<\/p>\n\n<p>On Kaggle, I got 99+% accuracy using  and 10 epochs. Meanwhile, I can't get above 97% using , even using 30 epochs (which isn't much, but I don't have a GPU, so training takes forever).<\/p>\n\n<p>Here's what my model looks like now:<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"41913968","CreationDate":"2017-01-28T19:34:29.917","ParentId":null,"OwnerUserId":"5133167","Title":null,"Body":"<p>First of all, binary_crossentropy is not when there are two classes.<\/p>\n\n<p>The \"binary\" name is because it is adapted for binary output, and each number of the softmax is aimed at being 0 or 1.\nHere, it checks for each number of the output.<\/p>\n\n<p>It doesn't explain your result, since categorical_entropy exploits the fact that it is a classification problem.<\/p>\n\n<p>Are you sure that when you read your data there is one and only one class per sample? It's the only one explanation I can give.<\/p>\n"},{"AnswerId":"46004661","CreationDate":"2017-09-01T16:41:16.367","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>Short answer: it is <strong>not<\/strong>. <\/p>\n\n<p>To see that, simply try to calculate the accuracy \"by hand\", and you will see that it is different from the one reported by Keras with the <code>model.evaluate<\/code> method:<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code># Keras reported accuracy:\nscore = model.evaluate(x_test, y_test, verbose=0) \nscore[1]\n# 0.99794011611938471\n\n# Actual accuracy calculated manually:\nimport numpy as np\ny_pred = model.predict(x_test)\nacc = sum([np.argmax(y_test[i])==np.argmax(y_pred[i]) for i in range(10000)])\/10000\nacc\n# 0.98999999999999999\n<\/code><\/pre>\n\n<p>The reason it <em>seems<\/em> to be so is a rather subtle issue at how Keras actually <em>guesses<\/em> which accuracy to use, depending on the loss function you have selected, when you include simply <code>metrics=['accuracy']<\/code> in your model compilation.<\/p>\n\n<p>If you check the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/metrics.py\" rel=\"noreferrer\">source code<\/a>, Keras does not define a single accuracy metric, but several different ones, among them <code>binary_accuracy<\/code> and <code>categorical_accuracy<\/code>. What happens <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py#L876\" rel=\"noreferrer\">under the hood<\/a> is that, since you have selected binary cross entropy as your loss function and have not specified a particular accuracy metric, Keras (wrongly...) infers that you are interested in the <code>binary_accuracy<\/code>, and this is what it returns.<\/p>\n\n<p>To avoid that, i.e. to use indeed binary cross entropy as your loss function (nothing wrong with this, in principle) while still getting the <strong>categorical<\/strong> accuracy required by the problem at hand (i.e. MNIST classification), you should ask explicitly for <code>categorical_accuracy<\/code> in the model compilation as follows:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>from keras.metrics import categorical_accuracy\nmodel.compile(loss='binary_crossentropy', optimizer='adamax', metrics=[categorical_accuracy])\n<\/code><\/pre>\n\n<p>And after training, scoring, and predicting the test set as I show above, the two metrics now are the same, as they should be:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>sum([np.argmax(y_test[i])==np.argmax(y_pred[i]) for i in range(10000)])\/10000 == score[1]\n# True\n<\/code><\/pre>\n\n<p>(HT to <a href=\"https:\/\/stackoverflow.com\/questions\/42081257\/keras-binary-crossentropy-vs-categorical-crossentropy-performance\/44498955#44498955\">this great answer<\/a> to a similar problem, which helped me understand the issue...)<\/p>\n\n<p><strong>UPDATE<\/strong>: After my post, I discovered that this issue had already been identified in <a href=\"https:\/\/stackoverflow.com\/questions\/45799474\/keras-model-evaluate-vs-model-predict-accuracy-difference-in-multi-class-nlp-ta\/45834857#45834857\">this answer<\/a>.<\/p>\n"}]}
{"QuestionId":41328089,"AnswerCount":1,"Tags":"<image><tensorflow><classification><python-3.5>","CreationDate":"2016-12-26T07:49:28.740","AcceptedAnswerId":41370340.0,"OwnerUserId":6802252.0,"Title":"Tensorflow image retraining with Inception3; where should new images be stored?","Body":"<p>So, I'm trying to retrain the image classifier model Inception 3 via their retrain.py code from their github (\"<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py<\/a>\")\n(The code is really lengthy, so I figured I shouldn't post it here)<\/p>\n\n<p>However, I can't seem to figure out the proper way to store the images I want to use or in which directory I should store them.  I'm using Windows 7 with Python 3.5 <\/p>\n\n<p>Currently, I have the images in a folder called  that are further broken down in sub-folders which are named according to the labels I want.<\/p>\n\n<p>So as an example, It's something like <\/p>\n\n<p>where is Images is the main folder, and the subfolders are the types of animals.<\/p>\n\n<p>The  folder is at the location  which is the location where the Mnist tutorial autmatically downloaded the the Mnist digit dataset to, when I ran the mnist tutorial.py. So I figured I should store my images around the same location.  <\/p>\n\n<p>When I run the  from the above github link:<\/p>\n\n<p>I get the following error<\/p>\n\n<p><\/p>\n\n<p>I've looked over the code and tried to find the default value for  which is an argument for the function  (which seems to be where the script is looking for images,  but can't find it, or maybe I'm not understanding the code properly. <\/p>\n\n<p>EDIT: I'm having trouble running python 3.5 via command prompt, so I can't use argparse. <\/p>\n\n<p>Thank you. <\/p>\n","answers":[{"AnswerId":"41370340","CreationDate":"2016-12-28T22:04:35.270","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>Are you able to change the default folder path to your path (C:\\tmp\\Images) here?\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py#L903\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py#L903<\/a><\/p>\n"}]}
{"QuestionId":41328129,"AnswerCount":1,"Tags":"<c++><neural-network><deep-learning><normalization><caffe>","CreationDate":"2016-12-26T07:54:11.413","AcceptedAnswerId":41344278.0,"OwnerUserId":3246146.0,"Title":"Min-Max normalization Layer in Caffe","Body":"<p>I am newbie in caffe, I am trying to normalize the convolution output between 0 to 1 with Min-Max Normalization.<\/p>\n\n<p>Out = X - Xmin \/ (Xmax - Xmin)<\/p>\n\n<p>I have checked many layers (Power, Scale, Batch Normalization, MVN) but no one is giving me min-max Normalization output in Layers. Can any one Help me ??<\/p>\n\n<p>*************  my prototxt  *****************<\/p>\n\n<pre><\/pre>\n\n<p>The convolution layer output is not in Normalized form I want Min-Max Normalized output in Layer format. Manually I can do using code but I need in Layers. \nThanks <\/p>\n","answers":[{"AnswerId":"41344278","CreationDate":"2016-12-27T11:24:44.503","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can write your own c++ layer following <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Development\" rel=\"nofollow noreferrer\">these guidelines<\/a>, you'll see how to implement \"forward only\" layers in that page.<\/p>\n\n<p>Alternatively, you can implement the layer in python and execute it in caffe via a <a href=\"https:\/\/stackoverflow.com\/q\/41344168\/1714410\">'\"Python\"' layer<\/a>:<\/p>\n\n<p>First, implement your layer in python, store it in <code>'\/path\/to\/my_min_max_layer.py'<\/code>:<\/p>\n\n<pre><code>import caffe\nimport numpy as np\n\nclass min_max_forward_layer(caffe.Layer):\n  def setup(self, bottom, top):\n    # make sure only one input and one output\n    assert len(bottom)==1 and len(top)==1, \"min_max_layer expects a single input and a single output\"\n\n  def reshape(self, bottom, top):\n    # reshape output to be identical to input\n    top[0].reshape(*bottom[0].data.shape)\n\n  def forward(self, bottom, top):\n    # YOUR IMPLEMENTATION HERE!!\n    in_ = np.array(bottom[0].data)\n    x_min = in_.min()\n    x_max = in_.max()\n    top[0].data[...] = (in_-x_min)\/(x_max-x_min)\n\n  def backward(self, top, propagate_down, bottom):\n    # backward pass is not implemented!\n    pass\n<\/code><\/pre>\n\n<p>Once you have the layer implemented in python, you can simply add it to your net (make sure <code>'\/path\/to'<\/code> is in your <code>$PYTHONPATH<\/code>):<\/p>\n\n<pre><code>layer {\n  name: \"my_min_max_forward_layer\"\n  type: \"Python\"\n  bottom: \"name_your_input_here\"\n  top: \"name_your_output_here\"\n  python_param {\n    module: \"my_min_max_layer\"  # name of python file to be imported\n    layer: \"min_max_forward_layer\" # name of layer class\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":41328769,"AnswerCount":1,"Tags":"<tensorflow><reset><restore><checkpoint><loss>","CreationDate":"2016-12-26T09:03:12.600","AcceptedAnswerId":null,"OwnerUserId":2712125.0,"Title":"Tensorflow: loss resets after successfully restored checkpoint","Body":"<p>There are no errors when saving or restoring. The weights appear to have restored correctly.<\/p>\n\n<p>I am trying to build my own minimal character level RNN by following <a href=\"https:\/\/gist.github.com\/karpathy\/d4dee566867f8291f086\" rel=\"noreferrer\">karpathy\/min-char-rnn.py<\/a>, <a href=\"https:\/\/github.com\/sherjilozair\/char-rnn-tensorflow\" rel=\"noreferrer\">sherjilozair\/char-rnn-tensorflow<\/a>, and the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\/\" rel=\"noreferrer\">Tensorflow RNN tutorial<\/a>. My script seems to work as expected except when I try to restore \/ resume training.<\/p>\n\n<p>If I restart the script and restore from checkpoint and then resume training, the loss would always go back up as if there are no checkpoints (despite the weights having restored correctly). \nHowever, within the script's execution, if I reset the graph, start a new session, and restore, then I am able to continue minimizing the loss as expected.<\/p>\n\n<p>I have tried to run this on my desktop (with GPU) and laptop (CPU only), both on Windows with Tensorflow 0.12.<\/p>\n\n<p>Below is my code, and I have uploaded the code + data + console output here: \n<a href=\"https:\/\/gist.github.com\/dk1027\/777c3da7ba1ff7739b5f5e89491bef73\" rel=\"noreferrer\">https:\/\/gist.github.com\/dk1027\/777c3da7ba1ff7739b5f5e89491bef73<\/a><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"51348261","CreationDate":"2018-07-15T12:12:20.593","ParentId":null,"OwnerUserId":"5634636","Title":null,"Body":"<h2>TL;DR<\/h2>\n\n<p>Please make sure your label is <strong>same<\/strong> each time you run your code, especially for those who use list indices as labels. <\/p>\n\n<p>See <a href=\"https:\/\/stackoverflow.com\/questions\/43646266\/tensorflow-loss-jumps-up-after-restoring-rnn-net\">this question<\/a> for details.<\/p>\n\n<p>If you use list indices as labels, sort data or save indices to disks. Use:<\/p>\n\n<pre><code>labels = sorted(set(data))\n<\/code><\/pre>\n\n<p>instead of<\/p>\n\n<pre><code>labels = set(data))\n<\/code><\/pre>\n\n<h2>General advice<\/h2>\n\n<p>In Python implementation, there are some methods, like <code>set()<\/code> or <code>os.listdir()<\/code>, return a collection which is not sorted. In other words, the index of an item might be different at each run.<\/p>\n\n<p>For <code>set()<\/code>, Python <a href=\"https:\/\/stackoverflow.com\/questions\/43646266\/tensorflow-loss-jumps-up-after-restoring-rnn-net\">use a random method<\/a> to build a <code>set<\/code>. For <code>os.listdir()<\/code>, <a href=\"https:\/\/stackoverflow.com\/questions\/4813061\/non-alphanumeric-list-order-from-os-listdir\">it doesn't promise the order of the returned list<\/a>. So for a robust code, use <code>sorted()<\/code> to your dataset is recommended.<\/p>\n\n<h2>For your question<\/h2>\n\n<pre><code>data_size = len(data)\nself.vocab = set(data)\nself.vocab_size = len(self.vocab)\nself.vocab_to_idx = {v:i for i,v in enumerate(self.vocab)}\nself.idx_to_vocab = {i:v for i,v in enumerate(self.vocab)}\n<\/code><\/pre>\n\n<p>It might be caused by the way you build your label. <code>vocab_to_idx<\/code> might be different each time you run your code.<\/p>\n\n<p>Just add a <code>sorted()<\/code>:<\/p>\n\n<pre><code>data_size = len(data)\nself.vocab = sorted(set(data))\nself.vocab_size = len(self.vocab)\nself.vocab_to_idx = {v:i for i,v in enumerate(self.vocab)}\nself.idx_to_vocab = {i:v for i,v in enumerate(self.vocab)}\n<\/code><\/pre>\n"}]}
{"QuestionId":41330270,"AnswerCount":1,"Tags":"<tensorflow><neural-network><gpu>","CreationDate":"2016-12-26T11:03:50.940","AcceptedAnswerId":null,"OwnerUserId":1468002.0,"Title":"Low Volatile GPU Utilization when running tensor flow","Body":"<p>I'm having trouble making tensorflow to use the Nvidia GeForce GTX 1080 GPU on my system efficiently. I reduced my code to the very simple version shown below; I'm only looping through a session.run() operation that should use the GPU, the data is only fetched once and is reused inside the loop, so this code should only utilize the GPU.<\/p>\n\n<p><\/p>\n\n<p>I noticed that for batch_size = 16, I get a mostly steadily GPU usage at around 8%, as I increase the batch_size to 32 the maximum GPU usage increases to 9-12% but the utilization stays mostly at 0% and from time to time it jumps to 15%-25% and immediately falls back to 0%. This patterns continues for larger batch_sizes, basically any batch size larger than 16 increases the maximum utilization but the utilization stays mostly at 0 and only spikes up from time to time. What am I missing here?<\/p>\n","answers":[{"AnswerId":"46979176","CreationDate":"2017-10-27T16:01:52.343","ParentId":null,"OwnerUserId":"822664","Title":null,"Body":"<p>I've had the same issue. My issue was: My calculation was partially executed on GPU and CPU (so there was a lot of communication between both devices, which lead to a low GPU utilization)\nI've read in a different thread:<\/p>\n\n<ol>\n<li>Don't use the data_dictornaries in loops (use Tensorvariable instead)<\/li>\n<li>There has been an issue with float64 datatype (which HAS to be calculated on CPU only) -> Use float32 datatype (if possible)<\/li>\n<li>use the Profiler suggested by Olivier Moindrot to check, if anything is explicitly run on your CPU. Then try to bring everything to GPU<\/li>\n<\/ol>\n\n<p>By the way, a hint regarding your code:\nYour session (default graph) will grow each iteration until a OutOfMemory exception possible... -> I close the session every x-iteration while resetting the default graph...<\/p>\n"}]}
{"QuestionId":41330846,"AnswerCount":1,"Tags":"<cmake><caffe>","CreationDate":"2016-12-26T11:50:25.207","AcceptedAnswerId":null,"OwnerUserId":7173164.0,"Title":"Caffe installation Error on Windows 10","Body":"<p>Hi I'm trying to install caffe framework by following steps on it's own <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/windows\" rel=\"nofollow noreferrer\"> github page<\/a>. \nWhen I execute \"scripts\\build_win.cmd\". I got some errors:<\/p>\n\n<blockquote>\n  <p>..\\libraries\\bin\\caffehdf5_D.dll : fatal error LNK1107: invalid or corrupt file: cannot read at 0x330 [C:\\Projects2\\caf\n  fe\\build\\tools\\upgrade_solver_proto_text.vcxproj]<\/p>\n<\/blockquote>\n\n<p>Also when I look CMakeError.log:<\/p>\n\n<blockquote>\n  <p>Determining if the include file pthread.h exists failed with the\n  following output: Change Dir:\n  C:\/Projects2\/caffe\/build\/CMakeFiles\/CMakeTmp<\/p>\n  \n  <p>Run Build Command:\"C:\/Program Files\n  (x86)\/MSBuild\/12.0\/bin\/MSBuild.exe\" \"cmTC_5c4c5.vcxproj\"\n  \"\/p:Configuration=Debug\" \"\/p:VisualStudioVersion=12.0\" Microsoft (R)\n  Build Engine version 12.0.31101.0<\/p>\n  \n  <p>[Microsoft .NET Framework, version 4.0.30319.42000]<\/p>\n  \n  <p>Copyright (C) Microsoft Corporation. All rights reserved.<\/p>\n  \n  <p>Build started 26.12.2016 14:34:11.<\/p>\n  \n  <p>Project\n  \"C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\cmTC_5c4c5.vcxproj\" on\n  node 1 (default targets).<\/p>\n  \n  <p>PrepareForBuild:<\/p>\n  \n  <p>Creating directory \"cmTC_5c4c5.dir\\Debug\\\".<\/p>\n  \n  <p>Creating directory\n  \"C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\Debug\\\".<\/p>\n  \n  <p>Creating directory \"cmTC_5c4c5.dir\\Debug\\cmTC_5c4c5.tlog\\\".<\/p>\n  \n  <p>InitializeBuildStatus:<\/p>\n  \n  <p>Creating \"cmTC_5c4c5.dir\\Debug\\cmTC_5c4c5.tlog\\unsuccessfulbuild\"\n  because \"AlwaysCreate\" was specified.<\/p>\n  \n  <p>ClCompile:<\/p>\n  \n  <p>C:\\Program Files (x86)\\Microsoft Visual Studio\n  12.0\\VC\\bin\\x86_amd64\\CL.exe \/c \/Zi \/W3 \/WX- \/Od \/Ob0 \/D \"WIN32\" \/D \"_WINDOWS\" \/D _DEBUG \/D \"CMAKE_INTDIR=\\\"Debug\\\"\" \/D _MBCS \/Gm- \/RTC1\n  \/MDd \/GS \/fp:precise \/Zc:wchar_t \/Zc:forScope\n  \/Fo\"cmTC_5c4c5.dir\\Debug\\\" \/Fd\"cmTC_5c4c5.dir\\Debug\\vc120.pdb\" \/Gd\n  \/TC \/errorReport:queue\n  \"C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\CheckIncludeFile.c\"<\/p>\n  \n  <p>Microsoft (R) C\/C++ Optimizing Compiler Version 18.00.31101 for x64<\/p>\n  \n  <p>Copyright (C) Microsoft Corporation.  All rights reserved.<\/p>\n  \n  <p>cl \/c \/Zi \/W3 \/WX- \/Od \/Ob0 \/D \"WIN32\" \/D \"_WINDOWS\" \/D _DEBUG \/D\n  \"CMAKE_INTDIR=\\\"Debug\\\"\" \/D _MBCS \/Gm- \/RTC1 \/MDd \/GS \/fp:precise\n  \/Zc:wchar_t \/Zc:forScope \/Fo\"cmTC_5c4c5.dir\\Debug\\\"\n  \/Fd\"cmTC_5c4c5.dir\\Debug\\vc120.pdb\" \/Gd \/TC \/errorReport:queue\n  \"C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\CheckIncludeFile.c\"<\/p>\n  \n  <p>CheckIncludeFile.c<\/p>\n  \n  <p>C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\CheckIncludeFile.c(1):\n  fatal error C1083: Cannot open include file: 'pthread.h': No such file\n  or directory\n  [C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\cmTC_5c4c5.vcxproj]<\/p>\n  \n  <p>Done Building Project\n  \"C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\cmTC_5c4c5.vcxproj\"\n  (default targets) -- FAILED.<\/p>\n  \n  <p>Build FAILED.<\/p>\n  \n  <p>\"C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\cmTC_5c4c5.vcxproj\"\n  (default target) (1) -><\/p>\n  \n  <p>(ClCompile target) -> <\/p>\n  \n  <p>C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\CheckIncludeFile.c(1):\n  fatal error C1083: Cannot open include file: 'pthread.h': No such file\n  or directory\n  [C:\\Projects2\\caffe\\build\\CMakeFiles\\CMakeTmp\\cmTC_5c4c5.vcxproj]<\/p>\n\n<pre><\/pre>\n  \n  <p>Time Elapsed 00:00:00.34<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"55545923","CreationDate":"2019-04-06T04:30:14.553","ParentId":null,"OwnerUserId":"6885805","Title":null,"Body":"<p>i have solved this error by changing folder permission to full access. reason is no read write permission.Link<a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/5840\" rel=\"nofollow noreferrer\">1<\/a><\/p>\n"}]}
{"QuestionId":41331019,"AnswerCount":1,"Tags":"<tensorflow><neural-network><one-hot-encoding>","CreationDate":"2016-12-26T12:07:11.940","AcceptedAnswerId":null,"OwnerUserId":4884530.0,"Title":"tensorflow 0.8 one hot encoding","Body":"<p>the data that i wanna encode looks as follows:<\/p>\n\n<pre><\/pre>\n\n<p>[ 0  0  0 ..., 42 42 42]<\/p>\n\n<p>there are 43 classes going from 0-42<\/p>\n\n<p>Now i read that tensorflow in version 0.8 has a new feature for one hot encoding so i tried to use it as following:<\/p>\n\n<pre><\/pre>\n\n<p>only problem is that i think the output is not what i need<\/p>\n\n<pre><\/pre>\n\n<p>Tensor(\"strided_slice:0\", shape=(43,), dtype=int32)<\/p>\n\n<p>Can someone nudge me in the right direction please :) <\/p>\n","answers":[{"AnswerId":"41339631","CreationDate":"2016-12-27T05:42:52.713","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>The output is correct and expected. trainhot[1] is the label of the second (0-based index) training sample, which is of 1D shape (43,). You can play with the code below to better understand tf.one_hot:<\/p>\n\n<pre><code>  onehot = tf.one_hot([0, 0, 41, 42], 43, on_value=1, off_value=0)                  \n  with tf.Session() as sess:                                                        \n    onehot_v = sess.run(onehot)                                                  \n  print(\"v: \", onehot_v)                                                            \n  print(\"v shape: \", onehot_v.shape)                                                \n  print(\"v[1] shape: \", onehot[1])\n\noutput:\nv:  [[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 0 0 0 0 0 0 0 0 0 0 0 0\n  0 0 0 0 0 0]\n [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 0 0 0 0 0 0 0 0 0 0 0 0\n  0 0 0 0 0 0]\n [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0\n  0 0 0 0 1 0]\n [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0\n  0 0 0 0 0 1]]\nv shape:  (4, 43)\nv[1] shape:  Tensor(\"strided_slice:0\", shape=(43,), dtype=int32)\n<\/code><\/pre>\n"}]}
{"QuestionId":41331604,"AnswerCount":1,"Tags":"<machine-learning><keras>","CreationDate":"2016-12-26T12:52:33.907","AcceptedAnswerId":null,"OwnerUserId":6922371.0,"Title":"How does keras(or any other ML framework) calculate the gradient of a lambda function layer for backpropagation?","Body":"<p>Keras enables adding a layer which calculates a user defined lambda function.\nWhat I don't get is how Keras knows to calculate the gradient of this user defined function for the backpropagation.<\/p>\n","answers":[{"AnswerId":"41336239","CreationDate":"2016-12-26T20:59:30.727","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>That one of the benefit of using Theano\/Tensorflow and libraries build on top of them. They can give you automatic gradient calculation of the mathematical functions and operations.<\/p>\n\n<p>Keras gets them by calling:<\/p>\n\n<pre><code># keras\/theano_backend.py\ndef gradients(loss, variables):\n    return T.grad(loss, variables)\n\n# keras\/tensorflow_backend.py\ndef gradients(loss, variables):\n    '''Returns the gradients of `variables` (list of tensor variables)\n    with regard to `loss`.\n    '''\n    return tf.gradients(loss, variables, colocate_gradients_with_ops=True)\n<\/code><\/pre>\n\n<p>which are in turn called by the optimizers(keras\/optimizers.py) <code>grads = self.get_gradients(loss, params)<\/code> to get the gradients which are used to write the update rule for all the <code>params<\/code>. <code>params<\/code> here are the trainable weights of the layers. But layers created by the Lambda functional layers don't have any trainable weights. But they affect the loss function though the forward prob and hence indirectly affect the calculation of the gradients of trainable weights of other layers. <\/p>\n\n<p>The only time you need to write new gradient calculation is when you are defining a new basic mathematical operation\/function. Also, when you write a custom loss function the auto grad almost always takes care of the gradient calculation. But optionally you can optimize training (not always) if you implement analytical gradient of your custom functions. For example softwax function can be expressed in exp, sum and div and auto grad can take care of it, but its analytical\/symbolic grad is usually implemented in Theano\/Tensorflow.<\/p>\n\n<p>For implementing new Ops you can see the below links for that:\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/extending_theano.html\" rel=\"noreferrer\">http:\/\/deeplearning.net\/software\/theano\/extending\/extending_theano.html<\/a>\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/how_tos\/adding_an_op\/index.html\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.12\/how_tos\/adding_an_op\/index.html<\/a><\/p>\n"}]}
{"QuestionId":41331626,"AnswerCount":1,"Tags":"<lua><require><torch>","CreationDate":"2016-12-26T12:54:10.100","AcceptedAnswerId":null,"OwnerUserId":7342278.0,"Title":"Torch 7\/Lua cannot require a .so file, but package.loadlib() can","Body":"<p>I get a problem when playing Torch 7 code. I have a shared library libhashnn.so file, and I want to load the functions in Torch 7 script, so I use this expression: , but the trepl gives an error.<\/p>\n\n<p>However, if I use function, it does work. Here are the results, but I don't know why I can't use , how can I use  to load lib successfully?<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41348231","CreationDate":"2016-12-27T15:55:07.173","ParentId":null,"OwnerUserId":"436275","Title":null,"Body":"<p>Try calling the function that <code>package.loadlib<\/code> gives you. I bet you will then get the same error that you get when you use <code>require<\/code>.<\/p>\n\n<p>To understand why: <code>require<\/code> does the equivalent of <code>package.loadlib<\/code> and then calls that function so that the module that you are loading can initialize itself. With just <code>package.loadlib<\/code>, this initialization is not done and so the error that occurs during initialization does not show up.<\/p>\n\n<p>I don't know what hashnn is and so I cannot tell you why it is broken, but looking at <code>\/home\/dazhen\/torch\/install\/share\/lua\/5.1\/trepl\/init.lua<\/code> line 383 might be a good start to figure out what is going on.<\/p>\n"}]}
{"QuestionId":41331676,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-12-26T12:59:04.717","AcceptedAnswerId":null,"OwnerUserId":806160.0,"Title":"Using tensorflow TFRecords for a dataset with multi task","Body":"<p>I want Use   for <a href=\"http:\/\/mmlab.ie.cuhk.edu.hk\/projects\/CelebA.html\" rel=\"nofollow noreferrer\">celeba face dataset<\/a>. celeba face datase has 5 landmark locations, 40 binary attributes annotations per image.<\/p>\n\n<p>  only can give one lable. but I want save all this labels in TFRecords.\nHow can do this in tensorflow?<\/p>\n\n<hr>\n\n<p>My solution is :\nI use <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/convert_to_records.py#L66\" rel=\"nofollow noreferrer\">this<\/a> to create  file.\nThen I convert celeba attribute file to  and read it by .<\/p>\n\n<p>Then add path to it with . (I used align images the was in  folder).<\/p>\n\n<p>for first then features I used .<\/p>\n\n<p>In the code of :<\/p>\n\n<p><\/p>\n\n<p>label must be , but   has type .<\/p>\n\n<p>So I changed:<\/p>\n\n<p><\/p>\n\n<p>and :<\/p>\n\n<p> <\/p>\n","answers":[{"AnswerId":"41341021","CreationDate":"2016-12-27T07:38:32.073","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>I think each record you write is proto buffer to be defined by you, and you can include whatever you need in the proto buffer. <\/p>\n\n<p>Here is an example proto buffer, which I believe you can reuse to include as many labels as you want with each label as a separate Feature:  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/convert_to_records.py#L66\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/convert_to_records.py#L66<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto#L88\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto#L88<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/feature.proto\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/feature.proto<\/a><\/p>\n"}]}
{"QuestionId":41332551,"AnswerCount":1,"Tags":"<deep-learning><conv-neural-network><caffe><inner-product>","CreationDate":"2016-12-26T14:19:10.317","AcceptedAnswerId":41336214.0,"OwnerUserId":2693257.0,"Title":"How to square weights in inner product layer?","Body":"<p>I started with  and ran it well. I need to square weights in . The  function expresses the , but I don't know how to square it.<\/p>\n\n<p> function is defined as follow:<\/p>\n\n<pre><\/pre>\n\n<p>Note that I use , but  returns an error. When I make my new layer, the warning is:<\/p>\n\n<pre><\/pre>\n\n<p>and after training my model, error is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41336214","CreationDate":"2016-12-26T20:56:01.947","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Note that <code>weight<\/code> is defined as a pointer to <code>Dtype<\/code>, pointers do not have <code>count()<\/code> method.<\/p>\n\n<p>You need the <code>count()<\/code> of the weight blob: <\/p>\n\n<pre><code>this-&gt;blobs_[0]-&gt;count()\n<\/code><\/pre>\n"}]}
{"QuestionId":41333798,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-12-26T16:26:14.330","AcceptedAnswerId":null,"OwnerUserId":5140684.0,"Title":"AttributeError: module 'tensorflow' has no attribute 'InteractiveSession'","Body":"<p>I have installed tensorflow from Source on Ubuntu 16.10 environment. Everything went smooth but now on compiling a program, it shows the following error: <\/p>\n\n<pre><\/pre>\n\n<p>Didn't find any post related to this. Could someone please help?<\/p>\n","answers":[{"AnswerId":"52345296","CreationDate":"2018-09-15T13:47:48.027","ParentId":null,"OwnerUserId":"1805756","Title":null,"Body":"<p>Today I got the same error after instaling <code>tenserflow-serving-api<\/code>. I installed tensorflow again and everything went back to normal.<\/p>\n"},{"AnswerId":"58870041","CreationDate":"2019-11-15T03:44:19.593","ParentId":null,"OwnerUserId":"2550498","Title":null,"Body":"<pre><code>sess=tf.compat.v1.InteractiveSession()\n<\/code><\/pre>\n\n<p>Use above line instead of <code>sess = tf.InteractiveSession()<\/code> line, if you are using tesorflow 2.0.0 version<\/p>\n"},{"AnswerId":"59027965","CreationDate":"2019-11-25T08:45:32.757","ParentId":null,"OwnerUserId":"11734595","Title":null,"Body":"<p>The error causes from Tensorflow version. When you've install v2.x and try to use v1.x, you'll get this error. To avoid this,<\/p>\n\n<pre><code>import tensorflow as tf\nimport tensorflow.compat.v1 as tfc\n<\/code><\/pre>\n\n<p>Use <strong>tfc<\/strong> instead of <strong>tf<\/strong> when you get the same error on the other functions like:<\/p>\n\n<pre><code>sess = tfc.InteractiveSession()\n\nmy_tensor = tfc.random_uniform((4, 4), 0, 1)\nprint(my_tensor)\n<\/code><\/pre>\n"}]}
{"QuestionId":41334173,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2016-12-26T17:05:14.310","AcceptedAnswerId":null,"OwnerUserId":7342861.0,"Title":"Trained keras model much slower making its predictions than in training","Body":"<p>I trained a keras model overnight, and got 75% accuracy which I am happy with right now. It has 60,000 samples, each with a sequence length of 700, and a vocabulary of 30. Each epoch takes about 10 minutes on my gpu. So that's 60,000 \/ 600 seconds which is roughly 100 samples per second, and that has to include back propagation. So I saved my hdf5 file and loaded it again.<\/p>\n\n<pre><\/pre>\n\n<p>When I then make my predictions it is taking more like 1 second per prediction which is 100 times slower than training. The predictions are good, I've looked at small batches and I can use them. The problem is that I need many 100,000s of them. 10ms second per prediction would work, 1 second won't.<\/p>\n\n<p>Can anyone suggest ways of speeding up Keras predictions?<\/p>\n","answers":[{"AnswerId":"53357293","CreationDate":"2018-11-18T02:07:48.573","ParentId":null,"OwnerUserId":"6406288","Title":null,"Body":"<p>I think it's because Keras's default predict behavior is with batch size 32. As a result especially if you're using a GPU, the small batch sizes destroy the performance. If you just change the batch size to predict(X_test, batch_size=128) you'll get significantly faster performance.<\/p>\n"}]}
{"QuestionId":41334603,"AnswerCount":0,"Tags":"<deep-learning><keras><keras-layer>","CreationDate":"2016-12-26T17:48:31.450","AcceptedAnswerId":null,"OwnerUserId":2362514.0,"Title":"Keras pass data through layers explicitly","Body":"<p>I am trying to implement a Pairwise Learning to rank model with keras where features are being computed by deep neural network. <\/p>\n\n<p>In the pairwise L2R model, while training, I am giving the query, one positive and one negative result. And it is trained on the classification loss by difference of feature vector.  <\/p>\n\n<p>I am able to do compile and fit model successfully but the problem is to actually use this model on test data. \nAs in Pairwise L2R model, at testing time I would have only query and sample pair (no separate negative and positives). And I can use the calculated value before softmax to rank samples. <\/p>\n\n<p>Is there any way I can use keras to pass data manually at test time through particular trained layers. (In short I have 3 set of inputs at train time and 2 at testing time.)<\/p>\n","answers":[]}
{"QuestionId":41335431,"AnswerCount":1,"Tags":"<python-3.x><neural-network><keras>","CreationDate":"2016-12-26T19:20:58.897","AcceptedAnswerId":41335683.0,"OwnerUserId":7327361.0,"Title":"How to fit a 3D matrix in keras model?","Body":"<p>I am creating a regression model with keras. I have   matrices of ten digits. I am facing problem to fit the   matrices in keras model.<\/p>\n\n<p> is input matrix<\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p> is target matrix<\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>For each  matrix there will be one value in the target matrix<\/p>\n\n<p><strong>Making the model<\/strong><\/p>\n\n<p><\/p>\n\n<p>Although the previous line is not throwing any error or warning but I am quite sure that it is <strong>not<\/strong> the correct way to fit the model in this case.<\/p>\n\n<p><\/p>\n\n<p>No problem so far. But when I am trying to fit the matrices<\/p>\n\n<p><\/p>\n\n<p>After this line I am getting a long Traceback which ultimately says<\/p>\n\n<p><\/p>\n\n<p>Please help me to correctly fit the matrices in the model. Thanks!<\/p>\n","answers":[{"AnswerId":"41335683","CreationDate":"2016-12-26T19:50:13.663","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Use <code>input_shape<\/code> instead of <code>input_dim<\/code>. Also since the number of dimensions of output is changing you need to use <code>Flatten<\/code> or <code>Reshape<\/code> as one of the dimensions.<\/p>\n\n<pre><code>from keras.layers import Flatten\n\nmodel = Sequential([\n    Dense(32, input_shape=(145,5)),\n    Flatten(),\n    Activation('sigmoid'),\n    Dense(output_dim=10)\n])\n\nmodel.summary()\n<\/code><\/pre>\n\n<p>Use <code>model.summary()<\/code> to check the structure of your model for better understanding.<\/p>\n"}]}
{"QuestionId":41335799,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><deep-learning><recurrent-neural-network>","CreationDate":"2016-12-26T20:05:07.527","AcceptedAnswerId":null,"OwnerUserId":6386637.0,"Title":"What is the most efficient way to implement multi-layer RNNs in TensorFlow?","Body":"<p>I\u2019m trying to figure out if it\u2019s more efficient to run an RNN on the inputs, and then run another RNN on those outputs, repeatedly (one horizontal layer at a time). Or to run one time-step at a time for all layers (one vertical layer at a time).<\/p>\n\n<p>I know tensorflow's MultiCellRNN class does the latter. Why is this method chosen over the former? Is the former equally efficient? Are there cases where going one time-step at a time for all layers is preferable?<\/p>\n\n<p>See <a href=\"http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/\" rel=\"nofollow noreferrer\">http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/<\/a> for reference on multi-layer RNNs.<\/p>\n","answers":[{"AnswerId":"41337420","CreationDate":"2016-12-26T23:58:27.197","ParentId":null,"OwnerUserId":"6610123","Title":null,"Body":"<p>1: How to easily implement an RNN\nUse an lstm cell, they're generally better (no vanishing gradient problems) and tensorflow makes it very easy to implement them through:<\/p>\n\n<p>from tensorflow.python.ops.rnn_cell import BasicLSTMCell\n...\ncell = BasicLSTMCell( state_dim )\nstacked_lstm = tf.nn.rnn_cell.MultiRNNCell([cell]*num_layers, state_is_tuple=True)<\/p>\n\n<p>find out more on the tensorflow website: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/recurrent\/<\/a><\/p>\n\n<p>2:Horizontal or Deep?\nJust like you can have a multi layer neural networks, you can also have a multi layer RNN. Think of the RNN cell as a layer within your neural network, a special layer which allows you to remember sequential inputs. From my experience you will still have linear transforms (or depth) within your network, but the question to have multiple layers of lstm cells depends on your network topology, preference, and computational ability. (the more the merrier) The amount of inputs and outputs depends on your problem, and as far as I can remember there is no such thing as multiple  Horizontal RNN cells, just depth. \nAll computation is done depth wise one input at a time. \nThe multi layer function you referenced is awesome, it handles all computation for you under the hood, just tell it how many cells you want and it does the rest. <\/p>\n\n<p>Good Luck<\/p>\n"}]}
{"QuestionId":41335940,"AnswerCount":0,"Tags":"<deep-learning><caffe><image-segmentation>","CreationDate":"2016-12-26T20:21:25.960","AcceptedAnswerId":null,"OwnerUserId":3425082.0,"Title":"Deeplab v2 error: undefined layer \"IMAGE_SEG_DATA\"","Body":"<p>As I tried to run \"run_pascal.sh\" in deeplapv2, I got the following error:<\/p>\n\n<p>Unknown enumeration value of \"IMAGE_SEG_DATA\" for field \"type\".<\/p>\n\n<p>I found a <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/O8lSpI-DOQY\" rel=\"nofollow noreferrer\">solution<\/a> stating that I should change the type of the layer from \"IMAGE_SEG_DATA\" to \"ImageSegData\", but the solution didn't work.<\/p>\n\n<p>Finally, I found the definition of this layer in \"image_seg_data_layer.cpp\". However, this layer isn't defined in the \"caffe.proto\". <\/p>\n\n<p>I traced back to the \"caffe.proto\" of deeplab v1 and there is actually an enumeration value named IMAGE_SEG_DATA.<\/p>\n\n<p>Please show me how to fix this problem!<\/p>\n","answers":[]}
{"QuestionId":41336576,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-26T21:46:38.753","AcceptedAnswerId":41370997.0,"OwnerUserId":3982224.0,"Title":"What are the input\/output tensors, for the translation(RNN) tutorial?","Body":"<p>As per <a href=\"https:\/\/stackoverflow.com\/questions\/39781946\/unable-to-deploy-a-cloud-ml-model\">Unable to deploy a Cloud ML model<\/a> if I want to deploy my model to the Google Cloud ML I need explicitly set the \"input\"\/\"output\" collections that will store the references to the input\/output tensors, like this:<\/p>\n\n<blockquote>\n  <p>This collection should name all the input tensors for your graph.\n  Similarly, a collection named \u201coutputs\u201d is required to name the output\n  tensors for your graph. Assuming your graph has two input tensors x\n  and y, and one output tensor scores, this can be done as follows:<\/p>\n  \n  <p>tf.add_to_collection(\u201cinputs\u201d, json.dumps({\u201cx\u201d : x.name, \u201cy\u201d:\n  y.name})) \n  tf.add_to_collection(\u201coutputs\u201d, json.dumps({\u201cscores\u201d:\n  scores.name})) <\/p>\n  \n  <p>Here \u201cx\u201d, \u201cy\u201d and \u201cscores\u201d become aliases to the actual\n  tensor names (x.name, y.name and scores.name)<\/p>\n<\/blockquote>\n\n<p>However, I do not know what are the input\/output tensors in the translation(RNN) <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/translate\/translate.py\" rel=\"nofollow noreferrer\">tutorial<\/a>. Without this knowledge, I can't refactor the code and deploy my models to the Google Cloud ML.<\/p>\n","answers":[{"AnswerId":"41370997","CreationDate":"2016-12-28T23:11:37.617","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>According to the code below, the inputs are: encoder_inputs, decoder_inputs, target_weights, and the output is the third element of the output of the return value of step()<\/p>\n\n<p><a href=\"https:\/\/github.com\/petewarden\/tensorflow_makefile\/blob\/master\/tensorflow\/models\/rnn\/translate\/seq2seq_model.py#L170\" rel=\"nofollow noreferrer\">https:\/\/github.com\/petewarden\/tensorflow_makefile\/blob\/master\/tensorflow\/models\/rnn\/translate\/seq2seq_model.py#L170<\/a><\/p>\n"}]}
{"QuestionId":41338352,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2016-12-27T02:38:10.493","AcceptedAnswerId":null,"OwnerUserId":7343939.0,"Title":"CNN for short text classification perform bad in validation set","Body":"<p>Im'using CNN for short text classification (classify the production title).\nThe code is from \n<a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"nofollow noreferrer\">http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/<\/a><\/p>\n\n<p>The accuracy in trainning set, test set, validatino set is blow:\n<img src=\"https:\/\/i.stack.imgur.com\/MWmUB.png\" alt=\"enter image description here\"><\/p>\n\n<p>and loss is different. The loss of validation is double than the loss of trainning set and test set.(I can't upload more than 2 pictures. sorry!)<\/p>\n\n<p>The trainning set and test set are from web by crawler, then split them with 7:3\uff0eAnd the validation is from real app message and tagged by manual marking\uff0e<\/p>\n\n<p>I have tried almost every hyper-parameters. <\/p>\n\n<p>I have tried up-sampling, down-sampling, none-sampling.<\/p>\n\n<p>batch size of 1024, 2048, 5096<\/p>\n\n<p>dropout with 0.3, 0.5, 0.7<\/p>\n\n<p>embedding_size with 30, 50, 75<\/p>\n\n<p>But none of these work!<\/p>\n\n<p>Now I use the param below:<\/p>\n\n<p>batch size is 2048.<\/p>\n\n<p>embedding_size is 30.<\/p>\n\n<p>sentence_length is 15<\/p>\n\n<p>filter_size is 3,4,5<\/p>\n\n<p>dropout_prob is 0.5<\/p>\n\n<p>l2_lambda is 0.005<\/p>\n\n<p>At first I think it is overfit.But the model performs well in test set then trainning set.So I confused!<\/p>\n\n<p>Is it the distribution between test set and trainning set is much different?<\/p>\n\n<p>How can I increase the performance in validation set?<\/p>\n","answers":[{"AnswerId":"45242097","CreationDate":"2017-07-21T16:02:02.963","ParentId":null,"OwnerUserId":"7303170","Title":null,"Body":"<p>I think this difference in loss comes from the fact that the validation dataset was collected from a different domain than the training\/test sets: <\/p>\n\n<blockquote>\n  <p>The training set and test set are from web by crawler, then split them\n  with 7:3\uff0eAnd the validation is from real app message and tagged by manual > marking<\/p>\n<\/blockquote>\n\n<p>The model did not see any real app message data during training, so it unsurprisingly fails to deliver good results on the validation set. Traditionally, all three sets are generated from the same pool of data (say, with a 7-1-2 split). The validation set is used for hyperparameter tuning (batch_size, embedding_length, etc.), while the test set is held-out for an objective measure of model performance.<\/p>\n\n<p>If you are concerned ultimately concerned with performance on the app data, I would split that dataset up 7-1-2 (train-validation-test) and augment the training data with web crawler data.<\/p>\n"},{"AnswerId":"45243964","CreationDate":"2017-07-21T17:59:04.197","ParentId":null,"OwnerUserId":"3698136","Title":null,"Body":"<p>I think the loss on validation set is high because the validation data comes from real app message data which may be more realistic than the training data you obtained from web crawling which may contain noise. Your learning rate is very high and batch size if pretty big than what's recommended. You can try learning rates in [0.1, 0.01, 0.001 and 0.0001], batch size in [32, 64], other hyperparameter values seems to be okay.<\/p>\n\n<p>I would like to comment on the training, validation and test set. Training data is split into training and validation sets for training while test set is the data we don't touch and use only to test our model at last. I think your validation set is the 'test set' and your test set is the 'validation set'. That's how I would refer to them.<\/p>\n"}]}
{"QuestionId":41338509,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-12-27T03:04:36.050","AcceptedAnswerId":41363704.0,"OwnerUserId":2059708.0,"Title":"Tensorflow mean squared error loss function","Body":"<p>I have seen a few different mean squared error loss functions in various posts for regression models in Tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>What are the differences between these?<\/p>\n","answers":[{"AnswerId":"43830040","CreationDate":"2017-05-07T09:45:38.753","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The first and the second loss functions calculate the same thing, but in a slightly different way. The third function calculate something completely different. You can see this by executing this code:<\/p>\n\n<pre><code>import tensorflow as tf\n\nshape_obj = (5, 5)\nshape_obj = (100, 6, 12)\nY1 = tf.random_normal(shape=shape_obj)\nY2 = tf.random_normal(shape=shape_obj)\n\nloss1 = tf.reduce_sum(tf.pow(Y1 - Y2, 2)) \/ (reduce(lambda x, y: x*y, shape_obj))\nloss2 = tf.reduce_mean(tf.squared_difference(Y1, Y2))\nloss3 = tf.nn.l2_loss(Y1 - Y2)\n\nwith tf.Session() as sess:\n    print sess.run([loss1, loss2, loss3])\n# when I run it I got: [2.0291963, 2.0291963, 7305.1069]\n<\/code><\/pre>\n\n<p>Now you can verify that 1-st and 2-nd calculates the same thing (in theory) by noticing that <code>tf.pow(a - b, 2)<\/code> is the same as <code>tf.squared_difference(a - b, 2)<\/code>. Also <code>reduce_mean<\/code> is the same as <code>reduce_sum \/ number_of_element<\/code>. The thing is that computers can't calculate everything exactly. To see what numerical instabilities can do to your calculations take a look at this:<\/p>\n\n<pre><code>import tensorflow as tf\n\nshape_obj = (5000, 5000, 10)\nY1 = tf.zeros(shape=shape_obj)\nY2 = tf.ones(shape=shape_obj)\n\nloss1 = tf.reduce_sum(tf.pow(Y1 - Y2, 2)) \/ (reduce(lambda x, y: x*y, shape_obj))\nloss2 = tf.reduce_mean(tf.squared_difference(Y1, Y2))\n\nwith tf.Session() as sess:\n    print sess.run([loss1, loss2])\n<\/code><\/pre>\n\n<p>It is easy to see that the answer should be 1, but you will get something like this: <code>[1.0, 0.26843545]<\/code>.<\/p>\n\n<p>Regarding your last function, the documentation says that:<\/p>\n\n<blockquote>\n  <p>Computes half the L2 norm of a tensor without the sqrt: output = sum(t\n  ** 2) \/ 2<\/p>\n<\/blockquote>\n\n<p>So if you want it to calculate the same thing (in theory) as the first one you need to scale it appropriately:<\/p>\n\n<pre><code>loss3 = tf.nn.l2_loss(Y1 - Y2) * 2 \/ (reduce(lambda x, y: x*y, shape_obj))\n<\/code><\/pre>\n"},{"AnswerId":"41363704","CreationDate":"2016-12-28T14:04:50.880","ParentId":null,"OwnerUserId":"5711725","Title":null,"Body":"<p>I would say that the third equation is different, while the 1st and 2nd are formally the same but behave differently due to numerical concerns.<\/p>\n\n<p>I think that the 3rd equation (using <code>l2_loss<\/code>) is just returning 1\/2 of the squared Euclidean norm, that is, the sum of the element-wise square of the input, which is <code>x=prediction-Y<\/code>. You are not dividing by the number of samples anywhere. Thus, if you have a very large number of samples, the computation may overflow (returning Inf).<\/p>\n\n<p>The other two are formally the same, computing the mean of the element-wise squared <code>x<\/code> tensor. However, while the documentation does not specify it explicitly, it is very likely that <code>reduce_mean<\/code> uses an algorithm adapted to avoid overflowing with very large number of samples. In other words, it likely does not try to sum everything first and <em>then<\/em> divide by N, but use some kind of rolling mean that can adapt to an arbitrary number of samples without necessarily causing an overflow.<\/p>\n"}]}
{"QuestionId":41338553,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2016-12-27T03:10:37.560","AcceptedAnswerId":null,"OwnerUserId":7344135.0,"Title":"The meaning of batch_size in ptb_word_lm (LSTM model of tensorflow)","Body":"<p>I am new to tensorflow, i am now a little confused about the meaning of . As commonly known that the meaning of  is the number of samples for each batch, but according to the code in , it seems not:<\/p>\n\n<p>reader.py:<\/p>\n\n<pre><\/pre>\n\n<p>What does batch_len mean? The number of batches?<\/p>\n\n<p>ptb_word_lm.py:<\/p>\n\n<pre><\/pre>\n\n<p>What does  mean? The number of sequences in each batch?<\/p>\n\n<p>But if batch_size means the number of batches, then everything make sense. have I misunderstood something?<\/p>\n","answers":[{"AnswerId":"41339126","CreationDate":"2016-12-27T04:45:50.280","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>There are a few different concepts here: epoch, step, batch, and unroll steps for LSTM.<\/p>\n\n<p>At the highest level, you train a network with multiple epochs. In each epoch, you will go through and use all training data (usually in a random order) by steps; In each step, you train a batch of samples. <\/p>\n\n<p>I think the confusion here added by LSTM is that: each step, you will train a sequence of batches, instead of a single batch. The length of sequence is the number of unroll steps (num_steps).<\/p>\n"}]}
{"QuestionId":41338622,"AnswerCount":2,"Tags":"<csv><numpy><tensorflow>","CreationDate":"2016-12-27T03:23:35.403","AcceptedAnswerId":null,"OwnerUserId":7181306.0,"Title":"How to remove extra quotes from csv?","Body":"<p>I'm troubling that read csv file. <\/p>\n\n<p>I tried the replace method. But numpy doesn't support that. <\/p>\n\n<p>The csv file format is like this.<\/p>\n\n<pre><\/pre>\n\n<p>But when I use this file in numpy.loadtxt the result is just like this<\/p>\n\n<p>Source code<\/p>\n\n<pre><\/pre>\n\n<p>Result<\/p>\n\n<pre><\/pre>\n\n<p>I want to remove the \" this one. <\/p>\n\n<p>So I really want this list.<\/p>\n\n<pre><\/pre>\n\n<p>What code shoud I use?<\/p>\n","answers":[{"AnswerId":"41339873","CreationDate":"2016-12-27T06:06:22.677","ParentId":null,"OwnerUserId":"2368818","Title":null,"Body":"<p>Just find an replace double quote (\") to single quote (') from your excel editor.\nAs I don't know what editor you are using, so I am giving you step by step way to replace any character in MS Excel.<\/p>\n\n<p><a href=\"https:\/\/support.office.com\/en-us\/article\/Find-or-replace-text-and-numbers-on-a-worksheet-3a2c910f-01b9-4263-8db2-333dead6ae33\" rel=\"nofollow noreferrer\">https:\/\/support.office.com\/en-us\/article\/Find-or-replace-text-and-numbers-on-a-worksheet-3a2c910f-01b9-4263-8db2-333dead6ae33<\/a><\/p>\n"},{"AnswerId":"41339948","CreationDate":"2016-12-27T06:14:01.287","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>With Pandas I get:<\/p>\n\n<pre><code>In [1278]: pd.read_csv('stack41338622.txt')\nOut[1278]: \n   num phone           sensorID  press  temp  accel  gps_lat  gps_lng  \\\n0    1  null  A0:E6:F8:7B:16:EA      0    17   1.25        0        0   \n1    2  null  A0:E6:F8:7B:16:A9      0    18   1.19        0        0   \n2    3  null  A0:E6:F8:7B:15:A5      0    18   1.19        0        0   \n3    4  null  A0:E6:F8:7B:16:EA      0    17   1.25        0        0   \n4    5  null  A0:E6:F8:7B:16:A9      0    18   1.19        0        0   \n5    6  null  A0:E6:F8:7B:15:A5      0    19   1.38        0        0   \n6    7  null  A0:E6:F8:7B:16:D6      0    18   1.12        0        0   \n7    8  null  A0:E6:F8:7B:16:EA      0    17   1.31        0        0   \n8    9  null  A0:E6:F8:7B:15:A5      0    19   1.38        0        0   \n\n                  time  \n0  2016-12-14 13:34:59  \n1  2016-12-14 13:34:59  \n2  2016-12-14 13:34:59  \n3  2016-12-14 13:35:00  \n4  2016-12-14 13:35:00  \n5  2016-12-14 13:35:00  \n6  2016-12-14 13:35:01  \n7  2016-12-14 13:35:01  \n8  2016-12-14 13:35:01  \n<\/code><\/pre>\n\n<p>With <code>converters<\/code> as described in <a href=\"https:\/\/stackoverflow.com\/questions\/2664790\/reading-csv-files-in-numpy-where-delimiter-is\">Reading CSV files in numpy where delimiter is &quot;,&quot;<\/a> we can strip off the extra quotes.  Unfortunately <code>dtypes=None<\/code> no longer works with converters, so we have to spell that out.  Here's a start:<\/p>\n\n<pre><code>In [1327]: def foo(astr):\n      ...:     return astr[1:-1]\nIn [1328]: convs = dict((col, foo) for col in range(9))\nIn [1329]: dt = ['i','S10','S20','i', 'i','f','i','i','S20']\nIn [1330]: data = np.genfromtxt('stack41338622.txt', dtype=dt, delimiter=',', names=True, converters=convs)\nIn [1331]: data\nOut[1331]: \narray([ (1, b'null', b'A0:E6:F8:7B:16:EA', 0, 17, 1.25, 0, 0, b'2016-12-14 13:34:59'),\n       (2, b'null', b'A0:E6:F8:7B:16:A9', 0, 18, 1.190000057220459, 0, 0, b'2016-12-14 13:34:59'),\n       (3, b'null', b'A0:E6:F8:7B:15:A5', 0, 18, 1.190000057220459, 0, 0, b'2016-12-14 13:34:59'),\n       (4, b'null', b'A0:E6:F8:7B:16:EA', 0, 17, 1.25, 0, 0, b'2016-12-14 13:35:00'),\n       (5, b'null', b'A0:E6:F8:7B:16:A9', 0, 18, 1.190000057220459, 0, 0, b'2016-12-14 13:35:00'),\n       (6, b'null', b'A0:E6:F8:7B:15:A5', 0, 19, 1.3799999952316284, 0, 0, b'2016-12-14 13:35:00'),\n       (7, b'null', b'A0:E6:F8:7B:16:D6', 0, 18, 1.1200000047683716, 0, 0, b'2016-12-14 13:35:01'),\n       (8, b'null', b'A0:E6:F8:7B:16:EA', 0, 17, 1.309999942779541, 0, 0, b'2016-12-14 13:35:01'),\n       (9, b'null', b'A0:E6:F8:7B:15:A5', 0, 19, 1.3799999952316284, 0, 0, b'2016-12-14 13:35:01')], \n      dtype=[('num', '&lt;i4'), ('phone', 'S10'), ('sensorID', 'S20'), ('press', '&lt;i4'), ('temp', '&lt;i4'), ('accel', '&lt;f4'), ('gps_lat', '&lt;i4'), ('gps_lng', '&lt;i4'), ('time', 'S20')])\n<\/code><\/pre>\n\n<p>Given the amount of time I spent on this, I'm inclined to go with the other suggestion - strip out the extra quotes in a text editor.  Those quotes are not needed in a comma delimited file, and more of a nuisance than a help.<\/p>\n\n<p>In an editor I just removed the <code>\"<\/code>:<\/p>\n\n<pre><code>num,phone,sensorID,press,temp,accel,gps_lat,gps_lng,time\n1,null,A0:E6:F8:7B:16:EA,0,17,1.25,0,0,2016-12-14 13:34:59\n2,null,A0:E6:F8:7B:16:A9,0,18,1.19,0,0,2016-12-14 13:34:59\n3,null,A0:E6:F8:7B:15:A5,0,18,1.19,0,0,2016-12-14 13:34:59\n4,null,A0:E6:F8:7B:16:EA,0,17,1.25,0,0,2016-12-14 13:35:00\n5,null,A0:E6:F8:7B:16:A9,0,18,1.19,0,0,2016-12-14 13:35:00\n...\n\nIn [1336]: data = np.genfromtxt('stack41338622_1.txt', dtype=None, delimiter=',', names=True)\nIn [1337]: data\nOut[1337]: \narray([ (1, b'null', b'A0:E6:F8:7B:16:EA', 0, 17, 1.25, 0, 0, b'2016-12-14 13:34:59'),\n       (2, b'null', b'A0:E6:F8:7B:16:A9', 0, 18, 1.19, 0, 0, b'2016-12-14 13:34:59'),\n       (3, b'null', b'A0:E6:F8:7B:15:A5', 0, 18, 1.19, 0, 0, b'2016-12-14 13:34:59'),\n       ..., \n      dtype=[('num', '&lt;i4'), ('phone', 'S4'), ('sensorID', 'S17'), ('press', '&lt;i4'), ('temp', '&lt;i4'), ('accel', '&lt;f8'), ('gps_lat', '&lt;i4'), ('gps_lng', '&lt;i4'), ('time', 'S19')])\n<\/code><\/pre>\n\n<p>The <code>b''<\/code> are the Python3 way of showing bytestrings.  You won't see those in Py2.<\/p>\n"}]}
{"QuestionId":41338859,"AnswerCount":1,"Tags":"<tensorflow><image-segmentation><image-recognition>","CreationDate":"2016-12-27T04:03:57.290","AcceptedAnswerId":41339063.0,"OwnerUserId":3763831.0,"Title":"Running a TensorFlow Image Recognition API to search for an object","Body":"<p>TensorFlow has an api using the inception v3 model for identifying objects. I was wondering, if there was any way to locate smaller images in a larger image. For example, locating all oranges on an orange tree. I tried splitting the larger image into a grid of smaller images and applying tensorflow on each individual smaller image but having a constant grid is extremely error-prone, is there any solution around this?<\/p>\n","answers":[{"AnswerId":"41339063","CreationDate":"2016-12-27T04:38:23.033","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>The term you're looking for is <em>object detection<\/em>. You can use a sliding window at different scales. This is one way, there's probably better ones out there, but I don't know what they are.<\/p>\n\n<p>Let's say some oranges are closer than others. Start with a 10x10 (or something) box in the top left corner, and see if your model classifies it as an orange. Move your box to the right 2 pixels (or something). Try again. Keep moving right, then move down 2 pixels and start a new row, etc. Now resize the image to be smaller (so now you're looking for bigger oranges), and repeat the whole process. You can google things like \"sliding window detection\", and \"image pyramid\" to find out more.<\/p>\n\n<p>Once you've gone through your image, you'll have a bunch of detections - you'll have to figure out some way to perform <em>non-maximum suppression<\/em> on your detections since you might have way too many.<\/p>\n"}]}
{"QuestionId":41339823,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-27T06:02:45.680","AcceptedAnswerId":null,"OwnerUserId":5737630.0,"Title":"Initialization of RMSPropOptimizer","Body":"<p>In TensorFlow (Python), when adding to the graph a , are any additional variables added that need initialization? If yes, how can I get access to them and initialize them manually? (I'd rather not use ). In other words:\n(1) How can I decide which initializer to use?\n(2) How can I add the initilizer op to the graph, specifically for these variables?<\/p>\n\n<p>EDIT 1:<\/p>\n\n<p>I'm referring here to any new  that is added to the graph whenever I add the , and how it is initialized (just like other s). I'm not referring to the arguments in constructor of the , (which are hyper parameters of the model).<\/p>\n","answers":[{"AnswerId":"41340529","CreationDate":"2016-12-27T07:02:41.850","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<blockquote>\n  <p>In TensorFlow (Python), when adding to the graph a tf.train.RMSPropOptimizer, are any additional variables added that need initialization? <\/p>\n<\/blockquote>\n\n<p>Yes.<\/p>\n\n<blockquote>\n  <p>If yes, how can I get access to them and initialize them manually? (I'd rather not use tf.global_variables_initializer). <\/p>\n<\/blockquote>\n\n<p>From the docs:<\/p>\n\n<blockquote>\n  <p><code>tf.train.RMSPropOptimizer.__init__(learning_rate, decay=0.9, momentum=0.0, epsilon=1e-10, use_locking=False, centered=False, name='RMSProp')<\/code><\/p>\n<\/blockquote>\n\n<p>There are at least 3 (<code>decay<\/code>, <code>momentum<\/code>, and <code>epsilon<\/code>). <\/p>\n\n<blockquote>\n  <p>In other words: (1) How can I decide which initializer to use? <\/p>\n<\/blockquote>\n\n<p>This I don't have a good answer to - I've been told in class \"just use Adam\". That might be good advice in general, but I imagine there are cases where others work better. This might be worth searching online for blog posts or survey papers or something.<\/p>\n\n<blockquote>\n  <p>(2) How can I add the initializer op to the graph, specifically for these variables?<\/p>\n<\/blockquote>\n\n<p>You can pass them into the constructor as named parameters.<\/p>\n"}]}
{"QuestionId":41340013,"AnswerCount":0,"Tags":"<machine-learning><deep-learning><keras>","CreationDate":"2016-12-27T06:19:29.113","AcceptedAnswerId":null,"OwnerUserId":6530406.0,"Title":"DNN accuracy drop half way through training","Body":"<p>I am training a simple 2 layer hidden net with 10 dimensional hidden layers.  The accuracy improves until epoch 14 when it drops to be half as good.  What would cause this and how can I fix it? I was thinking it could be something to do with vanishing gradients, but it is such a small net that i'm not sure...  <\/p>\n\n<p>I made the plot bellow to show the relation between depth and width in a DNN.  I dont seem to se a pattern in when it fails.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kmeGp.png\" rel=\"nofollow noreferrer\">depth-vs-width graph<\/a><\/p>\n\n<p><\/p>\n\n<p>I am using keras as a framework with the code<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41340461,"AnswerCount":3,"Tags":"<gpu><deep-learning><keras>","CreationDate":"2016-12-27T06:57:06.223","AcceptedAnswerId":41341085.0,"OwnerUserId":5634912.0,"Title":"Can I implement deep learning models in my laptop with intel hd graphics","Body":"<p>I am currently doing a project on deep learning for my masters degree. I wanted to install keras library, so when I started installing Theano and tensorflow i saw that i have to install CUDA. But my laptop comes with intel hd graphics. So my question is will it work if i install them anyway.\nThanks<\/p>\n","answers":[{"AnswerId":"43902023","CreationDate":"2017-05-10T20:19:24.417","ParentId":null,"OwnerUserId":"961658","Title":null,"Body":"<p>Edit:\nAs of now, you can directly use OpenCL based clDNN  (<a href=\"https:\/\/github.com\/01org\/clDNN\" rel=\"noreferrer\">https:\/\/github.com\/01org\/clDNN<\/a>) instead of using OpenVX, in order to perform Deep Learning inference on Intel Graphics. You will have to do the training on a powerful GPU like Nvidia or AMD and use the pre-trained model and use it in clDNN.<\/p>\n\n<p>You can start using Intel's Computer Vision SDK (<a href=\"https:\/\/software.intel.com\/en-us\/computer-vision-sdk\" rel=\"noreferrer\">https:\/\/software.intel.com\/en-us\/computer-vision-sdk<\/a>) in order to write Deep Learning Applications using OpenCV or OpenVX. <\/p>\n\n<p>OpenVX (<a href=\"https:\/\/www.khronos.org\/openvx\/\" rel=\"noreferrer\">https:\/\/www.khronos.org\/openvx\/<\/a>) programming model allows you to write simple Neural Network pipelines using the following SPEC (<a href=\"https:\/\/www.khronos.org\/registry\/OpenVX\/extensions\/neural_network\/html\/\" rel=\"noreferrer\">https:\/\/www.khronos.org\/registry\/OpenVX\/extensions\/neural_network\/html\/<\/a>)<\/p>\n\n<p>Alternatively you can use Model Optimizer that converts Caffe\/TensorFlow model into OpenVX, and you can accelerate the OpenVX Neural Network graph on Intel Integrated HD Graphics.<\/p>\n\n<p>Hope it helps.<\/p>\n"},{"AnswerId":"55823129","CreationDate":"2019-04-24T05:46:04.553","ParentId":null,"OwnerUserId":"6537788","Title":null,"Body":"<p>Their is a <a href=\"https:\/\/github.com\/plaidml\/plaidml\" rel=\"nofollow noreferrer\">PlaidML<\/a> with that you train deep learning model on Intel and AMD gpu.<\/p>\n"},{"AnswerId":"41341085","CreationDate":"2016-12-27T07:43:45.913","ParentId":null,"OwnerUserId":"7239654","Title":null,"Body":"<p>You can install and use keras without cuda, but you can't get gpu accelerating with intel hd graphics.<\/p>\n\n<p>If you use Theano as keras' backend, first install Theano:<\/p>\n\n<pre><code># for python2\npip install theano\n# for python3\npip3 install theano\n<\/code><\/pre>\n\n<p>Then set ~\/.theanorc file like this:<\/p>\n\n<pre><code>[global]\nfloatX = float32\ndevice = cpu\nallow_gc = True\n\n[blas]\nldflags = -lopenblas\n<\/code><\/pre>\n\n<p>If you use TensorFlow as keras' backend, just install the CPU version of TensorFlow.<\/p>\n\n<pre><code># for python2.7\npip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.12.0-cp27-none-linux_x86_64.whl\n# for python3.4\npip3 install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.12.0-cp34-cp34m-linux_x86_64.whl\n# for python3.5\npip3 install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.12.0-cp35-cp35m-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>Then install keras:<\/p>\n\n<pre><code># for python2\npip install keras\n# for python3\npip3 install keras\n<\/code><\/pre>\n"}]}
{"QuestionId":41340879,"AnswerCount":1,"Tags":"<tensorflow><softmax>","CreationDate":"2016-12-27T07:27:36.360","AcceptedAnswerId":null,"OwnerUserId":1530658.0,"Title":"in tensorflow MNIST softmax tutorial, softmax function is not used","Body":"<p>im following MNIST Softmax tutorials <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/<\/a><\/p>\n\n<p>Followed by the document, the model should be<\/p>\n\n<pre><\/pre>\n\n<p>but in the sample source code, as u can see <\/p>\n\n<pre><\/pre>\n\n<p>softmax is not used. I think it needs to be changed<\/p>\n\n<pre><\/pre>\n\n<p>I assume that, in the testing function it uses argmax so the it doesn't needs to be normalized to 0~1.0 value. But it can bring some confusion to developers.<\/p>\n\n<p>as idea on this?<\/p>\n","answers":[{"AnswerId":"41342296","CreationDate":"2016-12-27T09:19:01.187","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>Softmax is used, row 57:<\/p>\n\n<pre><code># So here we use tf.nn.softmax_cross_entropy_with_logits on the raw\n# outputs of 'y', and then average across the batch.\ncross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y, y_))\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">softmax_cross_entropy_with_logits<\/a> for more details.<\/p>\n"}]}
{"QuestionId":41341529,"AnswerCount":1,"Tags":"<python><caffe><feature-extraction><zero>","CreationDate":"2016-12-27T08:23:28.740","AcceptedAnswerId":null,"OwnerUserId":4473597.0,"Title":"caffe layer zero output-python","Body":"<p>I get the output of the VGG-changed network(here is vectors of 3 float numbers)every thing is OK, but when I just want to get the output of other layers(e.g. last convolution layer conv5_4) instead of the last layer, the output vectors are all zero.\nThis is the python code to get the output:<\/p>\n\n<pre><\/pre>\n\n<p>please help me<\/p>\n","answers":[{"AnswerId":"41518988","CreationDate":"2017-01-07T07:26:39.477","ParentId":null,"OwnerUserId":"4473597","Title":null,"Body":"<p>I changed the model from VGG-changed to LENET. the problem solved. It seemed the changed to vgg causes the above problem.<\/p>\n"}]}
{"QuestionId":41341973,"AnswerCount":1,"Tags":"<python><theano><convolution><sliding-window>","CreationDate":"2016-12-27T08:55:50.977","AcceptedAnswerId":41381982.0,"OwnerUserId":245549.0,"Title":"How to put an \"arbitrary\" operation into a sliding window using Theano?","Body":"<p>I want to define some function on a matrix . For example , where  is another matrix ( is fixed \/ constant). To make it more specific, let's assume that both  and  are  matrices. The result of the operation is a real number.<\/p>\n\n<p>Now I have a big matrix (let's say ). I want to apply the operation defined above to all  sub-matrices of the \"big\" matrix. In other words, I want to slide the  window over the \"big\" matrix. For each location of the window, I should get a real number. So, as a final result, I need to get a real-valued matrix (or 2D tensor) (its shape should be ).<\/p>\n\n<p>What I want to have is close to a convolutional layer but not exactly the same because I want to use a mean squared deviation instead of a linear function represented by a neuron.<\/p>\n","answers":[{"AnswerId":"41381982","CreationDate":"2016-12-29T14:49:45.293","ParentId":null,"OwnerUserId":"1879010","Title":null,"Body":"<p>This is only a Numpy solution, hope it suffices. \nI assume that your function is made up of an operation on the matrix elements and of a mean, i.e. a scaled sum. Hence, it is sufficient to look at <code>Y<\/code> as in<\/p>\n\n<pre><code>Y = np.power(X-X0, 2)\n<\/code><\/pre>\n\n<p>So we only need to deal with determining a windowed mean. Note that for the 1D case the matrix product with an appropriate vector of ones can be determined for calculating the mean, e.g.<\/p>\n\n<pre><code>h = np.array([0, 1, 1, 0])  # same dimension as y\nm1 = np.dot(h, y) \/ 2 \nm2 = (y[1] + y[2]) \/ 2\nprint(m1 == m2)  # True\n<\/code><\/pre>\n\n<p>The 2D case is analogous, but with two matrix multiplications, one for the rows and one for the columns. E.g.<\/p>\n\n<pre><code>m_2 = np.dot(np.dot(h, Y), h) \/ 2**2\n<\/code><\/pre>\n\n<p>To construct a sliding window, we need to build a matrix of shifted windows, e.g.<\/p>\n\n<pre><code>H = [[1, 1, 1, 0, 0, ..., 0],\n     [0, 1, 1, 1, 0, ..., 0],\n            .\n            .\n            .\n     [0, ..., 0, 0, 1, 1, 1]] \n<\/code><\/pre>\n\n<p>to calculate all the sums<\/p>\n\n<pre><code>S = np.dot(np.dot(H, Y), H.T)\n<\/code><\/pre>\n\n<p>A full example for a <code>(n, n)<\/code> matrix with a <code>(m, m)<\/code> window would be<\/p>\n\n<pre><code>import numpy as np\n\nn, m = 500, 10\nX0 = np.ones((n, n))\nX = np.random.rand(n, n)\nY = np.power(X-X0, 2)\n\nh = np.concatenate((np.ones(m), np.zeros(n-m)))  # window at position 0\nH = np.vstack((np.roll(h, k) for k in range(n+1-m)))  # slide the window \nM = np.dot(np.dot(H,Y), H.T) \/ m**2  # calculate the mean\nprint(M.shape)  # (491, 491)\n<\/code><\/pre>\n\n<p>An alternative but probably slightly less efficient way for building <code>H<\/code> is    <\/p>\n\n<pre><code>H = np.sum(np.diag(np.ones(n-k), k)[:-m+1, :] for k in range(m))\n<\/code><\/pre>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>Calculating the mean squared deviation is also possible with that approach. For that, we generalize the vector identity <code>|x-x0|^2 = (x-x0).T (x-x0) = x.T x - 2 x0.T x + x0.T x0<\/code> (a space denotes a scalar or matrix multiplication and <code>.T<\/code> a transposed vector) to the matrix case: <\/p>\n\n<p>We assume <code>W<\/code> is a <code>(m,n)<\/code> matrix containing a block <code>(m.m)<\/code> identity matrix, which is able to extract the <code>(k0,k1)<\/code>-th <code>(m,m)<\/code> sub-matrix by <code>Y = W Z W.T<\/code>, where <code>Z<\/code> is the <code>(n,n)<\/code> matrix containing the data. Calculating the difference<\/p>\n\n<pre><code> D = Y - X0 = Y = W Z W.T - X0 \n<\/code><\/pre>\n\n<p>is straightforward, where <code>X0<\/code> and <code>D<\/code> is a <code>(m,m)<\/code> matrix. The square-root of the squared sum of the elements is called <a href=\"https:\/\/en.wikipedia.org\/wiki\/Matrix_norm#Frobenius_norm\" rel=\"nofollow noreferrer\">Frobenius norm<\/a>. Based on those <a href=\"https:\/\/en.wikipedia.org\/wiki\/Matrix_norm#Frobenius_norm\" rel=\"nofollow noreferrer\">identities<\/a>, we can write the squared sum as  <\/p>\n\n<pre><code>s = sum_{i,j} D_{i,j}^2 = trace(D.T D) = trace((W Z W.T - X0).T (H Z H.T - X0))\n  = trace(W Z.T W.T W Z W.T) - 2 trace(X0.T W Z W.T) + trace(X0.T X0)\n  =: Y0 + Y1 + Y2\n<\/code><\/pre>\n\n<p>The term <code>Y0<\/code> can be interpreted as <code>H Z H.T<\/code> from the method from above.The term <code>Y1<\/code> can be interpreted as a weighted mean on <code>Z<\/code> and <code>Y2<\/code> is a constant, which only needs to be determined once.\nThus, a possible implementation would be:<\/p>\n\n<pre><code>import numpy as np\n\nn, m = 500, 10\nx0 = np.ones(m)\nZ = np.random.rand(n, n)\n\nY0 = Z**2\nh0 = np.concatenate((np.ones(m), np.zeros(n-m)))\nH0 = np.vstack((np.roll(h0, k) for k in range(n+1-m)))\nM0 = np.dot(np.dot(H0, Y0), H0.T)\n\nh1 = np.concatenate((-2*x0, np.zeros(n-m)))\nH1 = np.vstack((np.roll(h1, k) for k in range(n+1-m)))\nM1 = np.dot(np.dot(H1, Z), H0.T)\n\nY2 = np.dot(x0, x0)\nM = (M0 + M1) \/ m**2 + Y2\n<\/code><\/pre>\n"}]}
{"QuestionId":41342098,"AnswerCount":2,"Tags":"<tensorflow><hdf5><keras>","CreationDate":"2016-12-27T09:05:45.377","AcceptedAnswerId":null,"OwnerUserId":4140668.0,"Title":"Keras: Load checkpoint weights HDF5 generated by multiple GPUs","Body":"<p>Checkpoint snippet:<\/p>\n\n<pre><\/pre>\n\n<p>I trained my model on a multiple GPU host which dumps  files in HDF5 format. When I loaded them on a single GPU machine with , an error was raised:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any way to load checkpoint weights generated by multiple GPUs on a single GPU machine? It seems that no issue of Keras discussed this problem thus any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"56048949","CreationDate":"2019-05-08T20:56:27.320","ParentId":null,"OwnerUserId":"11472298","Title":null,"Body":"<p>Try this function:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def keras_model_reassign_weights(model_cpu,model_gpu):\n    weights_temp ={}\n    print('_'*5,'Collecting weights from GPU model','_'*5)\n    for layer in model_gpu.layers:\n        try:\n            for layer_unw in layer.layers:\n                #print('Weights extracted for: ',layer_unw.name)\n                weights_temp[layer_unw.name] = layer_unw.get_weights()\n            break\n        except:\n            print('Skipped: ',layer.name)\n    print('_'*5,'Writing weights to CPU model','_'*5)\n    for layer in model_cpu.layers:\n        try:\n            layer.set_weights(weights_temp[layer.name])\n            #print(layer.name,'Done!')\n        except:\n            print(layer.name,'weights does not set for this layer!')\n    return model_cpu\n<\/code><\/pre>\n\n<p>But you need to load weights to your gpu model first:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>#load or initialize your keras multi-gpu model\nmodel_gpu = None \n#load or initialize your keras model with the same structure, without using keras.multi_gpu function\nmodel_cpu = None \n#load weights into multigpu model\nmodel_gpu.load_weights(r'gpu_model_best_checkpoint.hdf5') \n#execute function\nmodel_cpu = keras_model_reassign_weights(model_cpu,model_gpu)\n#save obtained weights for cpu model\nmodel_cpu.save_weights(r'CPU_model.hdf5')\n<\/code><\/pre>\n\n<p>After transferring you can use weights with a single GPU or CPU model.<\/p>\n"},{"AnswerId":"44062978","CreationDate":"2017-05-19T06:32:39.870","ParentId":null,"OwnerUserId":"6202302","Title":null,"Body":"<p>You can load your model on a single GPU like this:<\/p>\n\n<pre><code>from keras.models import load_model\n\nmulti_gpus_model = load_model('mid')\norigin_model = multi_gpus_model.layers[-2]  # you can use multi_gpus_model.summary() to see the layer of the original model\norigin_model.save_weights('single_gpu_model.hdf5')\n<\/code><\/pre>\n\n<p>'single_gpu_model.hdf5' is the file that you can load to the single GPU machine model.<\/p>\n"}]}
{"QuestionId":41342908,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2016-12-27T09:56:52.503","AcceptedAnswerId":null,"OwnerUserId":6767439.0,"Title":"Error importing Theano in Windows","Body":"<p>I keep getting the error below when I try to import theano. I have mingw, libpython installed. I even tried installing TDM-GCC and adding that to path.<\/p>\n\n<pre><\/pre>\n\n<p>It looks like the command is splitting my username and that is causing the problem but I don't know how to fix that. <\/p>\n","answers":[]}
{"QuestionId":41343827,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-12-27T10:55:56.917","AcceptedAnswerId":41350084.0,"OwnerUserId":482601.0,"Title":"Unused Potiential for Parallelisation","Body":"<p>I use a queue returned from a  call for training in TensorFlow. I had assumed that TF would prefetch the next batch while performing calculations with the current batch on the GPU. I.e., I imagined a flow like this<\/p>\n\n<ul>\n<li>load batch A on CPU<\/li>\n<li>use batch A on GPU, <em>while<\/em> loading batch B on CPU<\/li>\n<li>use batch B on GPU, <em>while<\/em> loading batch C on CPU<\/li>\n<li>...<\/li>\n<\/ul>\n\n<p>However, looking at the tracing output, this does not seem to be the case:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SqlDb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SqlDb.png\" alt=\"Trace\"><\/a><\/p>\n\n<p>For most of the  operation, the GPU is idle.<\/p>\n\n<hr>\n\n<p>Is something like what I envision even possible with TensorFlow and if so: what am I doing wrong?<\/p>\n","answers":[{"AnswerId":"41347623","CreationDate":"2016-12-27T15:12:30.940","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>I don't think TensorFlow is currently doing this optimization. Would be worth filing a feature request (TensorFlow github issue). One potential cost for such optimization is the increased memory footprint for additional buffering for the next batch.<\/p>\n"},{"AnswerId":"41350084","CreationDate":"2016-12-27T18:09:35.837","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>The problem is that FIFOQueue is CPU-only so prefetching only works on CPU.\nHowever you can implement your own prefetching by creating GPU variables and copying stuff from FIFO queue onto GPU variables in parallel, and then relying on those variables instead of the queue for GPU input. Here's an implementation of this idea by Tim Zaman:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5722#issuecomment-266297116\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5722#issuecomment-266297116<\/a><\/p>\n"}]}
{"QuestionId":41344168,"AnswerCount":3,"Tags":"<python><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2016-12-27T11:17:14.207","AcceptedAnswerId":41481539.0,"OwnerUserId":1714410.0,"Title":"What is a `\"Python\"` layer in caffe?","Body":"<p>Caffe has a layer type .  <\/p>\n\n<p>For instance, this layer type can be used as a <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/pycaffe\/layers\/pyloss.py\" rel=\"noreferrer\">loss layer<\/a>.<br>\nOn other occasions it is used as an <a href=\"https:\/\/stackoverflow.com\/a\/34996628\/1714410\">input layer<\/a>.<\/p>\n\n<p>What is this layer type?<br>\nHow can this layer be used?<\/p>\n","answers":[{"AnswerId":"41481539","CreationDate":"2017-01-05T09:37:58.487","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/a\/41350404\/1714410\">Prune<\/a>'s and <a href=\"https:\/\/stackoverflow.com\/a\/41452702\/1714410\">Bharat<\/a>'s answers gives the overall purpose of a <code>\"Python\"<\/code> layer: a general purpose layer which is implemented in python rather than c++.<\/p>\n\n<p>I intend this answer to serve as a tutorial for using <code>\"Python\"<\/code> layer.<\/p>\n\n<hr>\n\n<h2>A Tutorial for <code>\"Python\"<\/code> layer<\/h2>\n\n<h3>what is a <code>\"Python\"<\/code> layer?<\/h3>\n\n<p>Please see the excellent answers of <a href=\"https:\/\/stackoverflow.com\/a\/41350404\/1714410\">Prune<\/a> and <a href=\"https:\/\/stackoverflow.com\/a\/41452702\/1714410\">Bharat<\/a>.<\/p>\n\n<h3>Pre-requisite<\/h3>\n\n<p>In order to use <code>'Python\"<\/code> layer you need to compile caffe with flag<\/p>\n\n<pre><code>WITH_PYTHON_LAYER := 1\n<\/code><\/pre>\n\n<p>set in <code>'Makefile.config'<\/code>.<\/p>\n\n<h3>How to implement a <code>\"Python\"<\/code> layer?<\/h3>\n\n<p>A <code>\"Python\"<\/code> layer should be implemented as a python class derived from <code>caffe.Layer<\/code> base class. This class <strong>must<\/strong> have the following four methods:<\/p>\n\n<pre><code>import caffe\nclass my_py_layer(caffe.Layer):\n  def setup(self, bottom, top):\n    pass\n\n  def reshape(self, bottom, top):\n    pass\n\n  def forward(self, bottom, top):\n    pass\n\n  def backward(self, top, propagate_down, bottom):\n    pass\n<\/code><\/pre>\n\n<p>What are these methods?<\/p>\n\n<p><strong><code>def setup(self, bottom, top)<\/code><\/strong>: This method is called once when caffe builds the net. This function should check that number of inputs (<code>len(bottom)<\/code>) and number of outputs (<code>len(top)<\/code>) is as expected.<br>\nYou should also allocate internal parameters of the net here (i.e., <code>self.add_blobs()<\/code>), see <a href=\"https:\/\/stackoverflow.com\/q\/39458461\/1714410\">this thread<\/a> for more information.<br>\nThis method has access to <code>self.param_str<\/code> - a string passed from the prototxt to the layer. See <a href=\"https:\/\/stackoverflow.com\/a\/34588801\/1714410\">this thread<\/a> for more information.<\/p>\n\n<p><strong><code>def reshape(self, bottom, top)<\/code><\/strong>: This method is called whenever caffe reshapes the net. This function should allocate the outputs (each of the <code>top<\/code> blobs). The outputs' shape is usually related to the <code>bottom<\/code>s' shape.<\/p>\n\n<p><strong><code>def forward(self, bottom, top)<\/code><\/strong>: Implementing the forward pass from <code>bottom<\/code> to <code>top<\/code>.<\/p>\n\n<p><strong><code>def backward(self, top, propagate_down, bottom)<\/code><\/strong>: This method implements the backpropagation, it propagates the gradients from <code>top<\/code> to <code>bottom<\/code>. <code>propagate_down<\/code> is a Boolean vector of <code>len(bottom)<\/code> indicating to which of the <code>bottom<\/code>s the gradient should be propagated.<\/p>\n\n<p>Some more information about <code>bottom<\/code> and <code>top<\/code> inputs you can find in <a href=\"https:\/\/stackoverflow.com\/a\/33797142\/1714410\">this post<\/a>.  <\/p>\n\n<p><strong>Examples<\/strong><br>\nYou can see some examples of simplified python layers <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/test\/test_python_layer.py\" rel=\"noreferrer\">here<\/a>, <a href=\"https:\/\/stackoverflow.com\/a\/34996628\/1714410\">here<\/a> and <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/pycaffe\/layers\/pyloss.py\" rel=\"noreferrer\">here<\/a>.<br>\nExample of \"moving average\" output layer can be found <a href=\"https:\/\/stackoverflow.com\/a\/49986631\/1714410\">here<\/a>.<\/p>\n\n<p><strong>Trainable parameters<\/strong><br>\n<code>\"Python\"<\/code> layer can have trainable parameters (like <code>\"Conv\"<\/code>, <code>\"InnerProduct\"<\/code>, etc.).<br>\nYou can find more information on adding trainable parameters in <a href=\"https:\/\/stackoverflow.com\/q\/41344168\/1714410\">this thread<\/a> and <a href=\"https:\/\/stackoverflow.com\/q\/44418828\/1714410\">this one<\/a>. There's also a very simplified example in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/test\/test_python_layer.py#L31-L45\" rel=\"noreferrer\">caffe git<\/a>.<\/p>\n\n<h3>How to add a <code>\"Python\"<\/code> layer in a prototxt?<\/h3>\n\n<p>See <a href=\"https:\/\/stackoverflow.com\/a\/41452702\/1714410\">Bharat<\/a>'s answer for details.<br>\nYou need to add the following to your prototxt:<\/p>\n\n<pre><code>layer {\n  name: 'rpn-data'\n  type: 'Python'  \n  bottom: 'rpn_cls_score'\n  bottom: 'gt_boxes'\n  bottom: 'im_info'\n  bottom: 'data'\n  top: 'rpn_labels'\n  top: 'rpn_bbox_targets'\n  top: 'rpn_bbox_inside_weights'\n  top: 'rpn_bbox_outside_weights'\n  python_param {\n    module: 'rpn.anchor_target_layer'  # python module name where your implementation is\n    layer: 'AnchorTargetLayer'   # the name of the class implementation\n    param_str: \"'feat_stride': 16\"   # optional parameters to the layer\n  }\n}\n<\/code><\/pre>\n\n<h3>How to add a <code>\"Python\"<\/code> layer using pythonic <code>NetSpec<\/code> interface?<\/h3>\n\n<p>It's very simple:<\/p>\n\n<pre><code>import caffe\nfrom caffe import layers as L\n\nns = caffe.NetSpec()\n# define layers here...\nns.rpn_labels, ns.rpn_bbox_targets, \\\n  ns.rpn_bbox_inside_weights, ns.rpn_bbox_outside_weights = \\\n    L.Python(ns.rpn_cls_score, ns.gt_boxes, ns.im_info, ns.data, \n             name='rpn-data',\n             ntop=4, # tell caffe to expect four output blobs\n             python_param={'module': 'rpn.anchor_target_layer',\n                           'layer': 'AnchorTargetLayer',\n                           'param_str': '\"\\'feat_stride\\': 16\"'})\n<\/code><\/pre>\n\n<h3>How to use a net with a <code>\"Python\"<\/code> layer?<\/h3>\n\n<p>Invoking python code from caffe is nothing you need to worry about. Caffe uses boost API to call python code from compiled c++.<br>\nWhat do you do need to do?<br>\nMake sure the python module implementing your layer is in <code>$PYTHONPATH<\/code> so that when caffe <code>import<\/code>s it - it can be found.<br>\nFor instance, if your module <code>my_python_layer.py<\/code> is in <code>\/path\/to\/my_python_layer.py<\/code> then<\/p>\n\n<pre><code>PYTHONPATH=\/path\/to:$PYTHONPATH $CAFFE_ROOT\/build\/tools\/caffe train -solver my_solver.prototxt\n<\/code><\/pre>\n\n<p>should work just fine.<\/p>\n\n<h3>How to test my layer?<\/h3>\n\n<p><em>You should always test your layer before putting it to use.<\/em><br>\nTesting the <code>forward<\/code> function is entirely up to you, as each layer has a different functionality.<br>\nTesting the <code>backward<\/code> method is <strong>easy<\/strong>, as this method only implements a gradient of <code>forward<\/code> it can be numerically tested automatically!<br>\nCheck out <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/5157\" rel=\"noreferrer\"><code>test_gradient_for_python_layer<\/code><\/a> testing utility:<\/p>\n\n<pre><code>import numpy as np\nfrom test_gradient_for_python_layer import test_gradient_for_python_layer\n\n# set the inputs\ninput_names_and_values = [('in_cont', np.random.randn(3,4)), \n                          ('in_binary', np.random.binomial(1, 0.4, (3,1))]\noutput_names = ['out1', 'out2']\npy_module = 'folder.my_layer_module_name'\npy_layer = 'my_layer_class_name'\nparam_str = 'some params'\npropagate_down = [True, False]\n\n# call the test\ntest_gradient_for_python_layer(input_names_and_values, output_names, \n                               py_module, py_layer, param_str, \n                               propagate_down)\n\n# you are done!\n<\/code><\/pre>\n\n<h3>Special Notice<\/h3>\n\n<p>It is worth while noting that python code runs on CPU only. Thus, if you plan to have a Python layer in the <em>middle<\/em> of your net you will see a <strong>significant<\/strong> degradation in performance if you plan on using GPU. This happens because caffe needs to copy blobs from GPU to CPU before calling python layer and then copy back to GPU to proceed with the forward\/backward pass.<br>\nThis degradation is far less significant if the python layer is either an input layer or the topmost loss layer.<br>\n<em>Update:<\/em> On Sep 19th, 2017 <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/5904\" rel=\"noreferrer\">PR #5904<\/a> was merged into master. This PR exposes GPU pointers of blobs via the python interface.\nYou may access blob._gpu_data_ptr and blob._gpu_diff_ptr directly from python <strong>at your own risk<\/strong>.<\/p>\n"},{"AnswerId":"41350404","CreationDate":"2016-12-27T18:35:20.993","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>Very simply, it's a layer in which <em>you<\/em> provide the implementation code, rather than using one of the pre-defined types -- which are all backed by efficient functions.<\/p>\n\n<p>If you want to define a custom loss function, go ahead: write it yourself, and create the layer with type <strong>Python<\/strong>.  If you have non-standard input needs, perhaps some data-specific pre-processing, no problem: write it yourself, and create the layer with type <strong>Python<\/strong>.<\/p>\n"},{"AnswerId":"41452702","CreationDate":"2017-01-03T21:31:48.030","ParentId":null,"OwnerUserId":"628538","Title":null,"Body":"<p>Python layers are different from C++ layers which need to be compiled, their parameters need to be added to the proto file and finally you need to register the layer in layer_factory. If you write a python layer, you don't need to worry about any of these things. Layer parameters can be defined as a string, which are accessible as a string in python. For example: if you have a parameter in a layer, you can access it using 'self.param_str', if param_str was defined in your prototxt file. Like other layers, you need to define a class with the following functions:<\/p>\n\n<ul>\n<li>Setup - Initialize your layer using parameters obtained from layer variables<\/li>\n<li>Forward - What would be input and output of a layer<\/li>\n<li>Backward - Given the prediction and gradients from the next layer, compute the gradients for the previous layer<\/li>\n<li>Reshape - Reshape your blob if needed<\/li>\n<\/ul>\n\n<p>Prototxt example:<\/p>\n\n<pre><code>layer {\n  name: 'rpn-data'\n  type: 'Python'\n  bottom: 'rpn_cls_score'\n  bottom: 'gt_boxes'\n  bottom: 'im_info'\n  bottom: 'data'\n  top: 'rpn_labels'\n  top: 'rpn_bbox_targets'\n  top: 'rpn_bbox_inside_weights'\n  top: 'rpn_bbox_outside_weights'\n  python_param {\n    module: 'rpn.anchor_target_layer'\n    layer: 'AnchorTargetLayer'\n    param_str: \"'feat_stride': 16\"\n  }\n}\n<\/code><\/pre>\n\n<p>Here, name of the layer is rpn-data, bottom and top are input and output details of the layer respectively. python_param defines what are the parameters of the Python layer. 'module' specifies what is the file name of your layer. If the file called 'anchor_target_layer.py' is located inside a folder called 'rpn', the parameter would be 'rpn.anchor_target_layer'. The 'layer' parameter is the name of your class, in this case it is 'AnchorTargetLayer'. 'param_str' is a parameter for the layer, which contains a value 16 for the key 'feat_stride'.<\/p>\n\n<p>Unlike C++\/CUDA layers, Python layers do not work in a multi-GPU setting in caffe as of now, so that is a disadvantage of using them.<\/p>\n"}]}
{"QuestionId":41344224,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><regression><data-science>","CreationDate":"2016-12-27T11:20:47.037","AcceptedAnswerId":null,"OwnerUserId":2155026.0,"Title":"Using TensorFlow to do regression on a 13 column data set","Body":"<p>I would like to do regression on a 13 column data set. The second column is dependent on the rest of the 12 columns. All column contains real number values.<\/p>\n\n<p>How can I create a neural network using TensorFlow to do the regression? I have tried going through this <a href=\"https:\/\/www.tensorflow.org\/tutorials\/wide\/\" rel=\"nofollow noreferrer\">tutorial<\/a> but it is too advanced for me.<\/p>\n\n<p>Thanks in advance for a MWE.<\/p>\n","answers":[{"AnswerId":"41351418","CreationDate":"2016-12-27T19:57:39.907","ParentId":null,"OwnerUserId":"5538787","Title":null,"Body":"<p>In that tutorial they are using logistic regression, that is a linear binary classifier. They use the class tf.contrib.learn.LinearClassifier as their model.<\/p>\n\n<p>If you use class tf.contrib.learn.LinearRegressor then you can do linear regression instead of classification.<\/p>\n\n<p>In that webpage you have tutorials for other models. If you want to create a neural network you have different tutorial in the left menu, for example:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/<\/a><\/p>\n\n<p>In this repository you have python notebooks with the full code of many different neural networks:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/udacity\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/udacity<\/a><\/p>\n"}]}
{"QuestionId":41344447,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-27T11:35:25.357","AcceptedAnswerId":null,"OwnerUserId":2272798.0,"Title":"What happened to RNN and Seq2seq options in TensorFlow?","Body":"<p>It looks like lots of stuff was recently deprecated or moved to contrib, and it doesn't seem like there are any up-to-date equivalents in core. What was the reasoning behind this? I can't seem to find any GitHub issue or discussion explaining these changes. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/search?utf8=%E2%9C%93&amp;q=embedding_attention_decoder\" rel=\"nofollow noreferrer\">Example GitHub search<\/a><\/p>\n","answers":[{"AnswerId":"47378559","CreationDate":"2017-11-19T15:29:56.750","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>Around the time of the 1.0 release, we made a large effort to carefully choose what should be part of the \"official\" supported API and feature set, because anything included would have to be supported for a long, long time and becomes very hard to change.  The rules for stuff in contrib are less tight in terms of making non-backwards-compatible changes, so components judged likely to change substantially in the near future (i.e., to not be stable yet) were moved to contrib or elsewhere.<\/p>\n"}]}
{"QuestionId":41346052,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2016-12-27T13:25:16.030","AcceptedAnswerId":null,"OwnerUserId":3970726.0,"Title":"Extracting subgraph in tensorflow","Body":"<p>I have pretrained network and I'm trying to get just a part of it (subgraph) tf graph along with variables and saver object.<\/p>\n\n<p>this is how I'm doing it:<\/p>\n\n<pre><\/pre>\n\n<p>This however removes all variables (when I call reset_default_graph). Even If I explicitely add the operation nodes for variables (only the \"variable\" type operations) into the \"list of nodes to preserve\".<\/p>\n\n<p>How can I preserve subgraph of larger graph while preserving values of variables?\nIs it a matter of addition some new nodes to \"preserve list\"?<\/p>\n\n<p>The relation between graph nodes and variables is still unclear to me and tutorial merely mentions that creation of variable creates some operations (nodes) in the graph.<\/p>\n","answers":[{"AnswerId":"41346855","CreationDate":"2016-12-27T14:20:04.597","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>I think what you are doing looks right. As you said, a Variable is a simply an operation (a node in graph) that outputs a tensor of certain values. You should be able to add Variable nodes to the list to preserve them, as you have been already doing. Could you use print(sess.graph_def) to make sure the names you provided are correct? <\/p>\n"}]}
{"QuestionId":41346058,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-12-27T13:25:35.433","AcceptedAnswerId":47378579.0,"OwnerUserId":2920185.0,"Title":"You must feed a value for placeholder tensor 'input_example_tensor' with dtype string and shape [1]","Body":"<p>I am developing a tensorflow serving client\/server application by using <a href=\"https:\/\/github.com\/dennybritz\/chatbot-retrieval\/\" rel=\"nofollow noreferrer\">chatbot-retrieval<\/a> project.<\/p>\n\n<p>My code has two parts, namely serving part and client part.<\/p>\n\n<p>Below is the code snippet for the serving parts.<\/p>\n\n<pre><\/pre>\n\n<p>Below is the code snippet for the client part.<\/p>\n\n<pre><\/pre>\n\n<p>Both of the serving and client parts builds with no error.  After running the serving application and then the client application I get the following strange error propogated to the client application when the rpc call completes.<\/p>\n\n<p>Below is the error I get when rpc call completes<\/p>\n\n<pre><\/pre>\n\n<p>The error is strange since there seems to be no way to feed the placeholder from the client application. <\/p>\n\n<p>How can I provide data for the placeholder 'input_example_tensor' if I am accessing the model through tensorflow serving?<\/p>\n\n<p><strong>ANSWER:<\/strong>\n(I posted my answer here since I couldn't post it as an answer due to lack of StackOverflow badges. Anyone who is volunteer to submit it as his\/her answer to the question is more than welcome. I will approve it as the answer.)<\/p>\n\n<p>I could resolve the problem by using the option use_deprecated_input_fn = False in estimator.export function and change the input signatures accordingly.<\/p>\n\n<p>Below is the final code which is running with no problem.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47378579","CreationDate":"2017-11-19T15:32:31.450","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>OP self-solved but couldn't self-answer, so here's their answer:<\/p>\n\n<p>Problem was fixed by using the option <code>use_deprecated_input_fn = False<\/code> in <code>estimator.export<\/code> function and changing the input signatures accordingly:<\/p>\n\n<pre><code>def my_signature_fn(input_example_tensor, features, predictions):   \n  default_graph_signature = exporter.regression_signature(\n      input_tensor=input_example_tensor,\n      output_tensor=predictions\n      )\n\n  named_graph_signatures = {\n      'inputs':exporter.generic_signature(\n          {\n          'context':features['context'],\n          'utterance':features['utterance']\n          }\n       ),\n      'outputs':exporter.generic_signature(\n          {\n          'scores':predictions\n          }\n       )\n      }\n\n  return default_graph_signature, named_graph_signatures\n\ndef main():\n  ##preliminary codes here##\n\n  estimator.fit(input_fn=input_fn_train, steps=100, monitors=[eval_monitor])\n\n  estimator._targets_info = tf.contrib.learn.estimators.tensor_signature.TensorSignature(tf.constant(0, shape=[1,1]))\n\n  estimator.export(\n      export_dir = FLAGS.export_dir,\n      input_fn = my_input_fn,\n      input_feature_key =\"my_input_example_tensor\",\n      use_deprecated_input_fn = False,\n      signature_fn = my_signature_fn,\n      exports_to_keep = 1\n      )\n<\/code><\/pre>\n"}]}
{"QuestionId":41346155,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-27T13:31:36.750","AcceptedAnswerId":41349665.0,"OwnerUserId":1660762.0,"Title":"TensorFlow: evaluate a single input with sess.run","Body":"<p>I've trained a network in TensorFlow using a session<\/p>\n\n<pre><\/pre>\n\n<p>What's the simplest way to give the session an input of <strong>a single<\/strong> data row and print the output?<\/p>\n\n<p>I've tried (many variations of)<\/p>\n\n<pre><\/pre>\n\n<p>But I get the result<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41349665","CreationDate":"2016-12-27T17:39:32.337","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Assumption - Your first dimension represents the number of inputs (as is the case in most offical examples)<\/p>\n\n<p>I'd suggest making the first dimension of the placeholder as <code>None<\/code> to have the option of passing any number of batches. Have a look at this tutorial for an example - <a href=\"http:\/\/learningtensorflow.com\/lesson4\/\" rel=\"nofollow noreferrer\">http:\/\/learningtensorflow.com\/lesson4\/<\/a><\/p>\n\n<p>Quoting from this tutorial,<\/p>\n\n<blockquote>\n  <p>The first dimension of the placeholder is <code>None<\/code>, meaning we can have any number of rows. The second dimension is fixed at 3, meaning each row needs to have three columns of data.<\/p>\n<\/blockquote>\n\n<p>It's also well documented in the official documentation in <a href=\"https:\/\/www.tensorflow.org\/resources\/faq#how_do_i_build_a_graph_that_works_with_variable_batch_sizes\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/resources\/faq#how_do_i_build_a_graph_that_works_with_variable_batch_sizes<\/a>.<\/p>\n"}]}
{"QuestionId":41347579,"AnswerCount":1,"Tags":"<deep-learning><keras><keras-layer>","CreationDate":"2016-12-27T15:09:55.387","AcceptedAnswerId":41348217.0,"OwnerUserId":3827807.0,"Title":"Keras convolutional autoencoder doesn't work","Body":"<p>I'm testing the convolutional autoencoder from the author of Keras just here :\n<a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-autoencoders-in-keras.html<\/a><\/p>\n\n<p>But I have this problem :<\/p>\n\n<pre><\/pre>\n\n<p>I precise, I already setted the field 'border_mode='same'' in the last conv layer.\nSo I really don't know from where it come from..\nHere is the summary :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41348217","CreationDate":"2016-12-27T15:54:41.430","ParentId":null,"OwnerUserId":"3827807","Title":null,"Body":"<p>Finally found the answer.\nI think the creator of the tutorial tested it with 32x32 MNIST images and not 28x28.<\/p>\n\n<p>Because, when adding to the last conv layer border_mode='same', you get an output shape of (32,32,1)\nSo, to get the good output (28,28,1) you need to add border_mode='valid' to the before last conv layer.<\/p>\n\n<p>In summary :\nCorrect the dimension ordering to 28x28x1 instead of 1x28x28.\nThen add border mode same to the last conv layer\nAnd finally add border mode valid to the last conv layer.<\/p>\n\n<p>Hope this will help.<\/p>\n"}]}
{"QuestionId":41347710,"AnswerCount":2,"Tags":"<python><google-app-engine><tensorflow><google-cloud-storage><google-cloud-platform>","CreationDate":"2016-12-27T15:17:08.940","AcceptedAnswerId":41353098.0,"OwnerUserId":2309451.0,"Title":"GCP Machine Learning job: how to print lines to .csv output file?","Body":"<p>On GCP Machine Learning, I'm trying to run a training job, with intermediate output to a .csv file. I'm using Tensorflow in python.<\/p>\n\n<p>Here's what I've tried so far:<\/p>\n\n<p>1.<\/p>\n\n<pre><\/pre>\n\n<p>... throws exception: <\/p>\n\n<p>2.<\/p>\n\n<pre><\/pre>\n\n<p>... throws exception <\/p>\n\n<p>Any idea?<\/p>\n","answers":[{"AnswerId":"41348099","CreationDate":"2016-12-27T15:45:56.667","ParentId":null,"OwnerUserId":"1900169","Title":null,"Body":"<p>It looks like you are writing code for the App Engine environment, but are you sure that is the correct environment to be writing code for?<\/p>\n\n<p>If you are using regular GCP instances you should probably use <a href=\"https:\/\/cloud.google.com\/storage\/docs\/object-basics#storage-upload-object-python\" rel=\"nofollow noreferrer\">one of the other APIs<\/a>.<\/p>\n\n<pre><code>from google.cloud import storage\n\ndef upload_blob(bucket_name, source_file_name, destination_blob_name):\n    \"\"\"Uploads a file to the bucket.\"\"\"\n    storage_client = storage.Client()\n    bucket = storage_client.get_bucket(bucket_name)\n    blob = bucket.blob(destination_blob_name)\n\n    blob.upload_from_filename(source_file_name)\n\n    print('File {} uploaded to {}.'.format(\n        source_file_name,\n        destination_blob_name))\n<\/code><\/pre>\n\n<p>Edit: replaced API link<\/p>\n"},{"AnswerId":"41353098","CreationDate":"2016-12-27T22:42:15.880","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>For option (1), if you are using TensorFlow, you can use <code>tf.gfile.Open(\"gs:\/\/...\", mode=\"w\")<\/code> to open a file in GCS for writing. <\/p>\n\n<p>The <code>tf.gfile<\/code> module uses TensorFlow's C++ I\/O layer, which includes supports for reading from and writing to GCS. The built-in Python <code>open()<\/code> function will only open files in the local file system.<\/p>\n"}]}
{"QuestionId":41347780,"AnswerCount":1,"Tags":"<python><theano><elementwise-operations>","CreationDate":"2016-12-27T15:22:45.687","AcceptedAnswerId":41379690.0,"OwnerUserId":245549.0,"Title":"How to multiply each column of a matrix by a vector element-wise in Theano?","Body":"<p>I have a Theano  with 100 elements. I also have a matrix with 5 columns and 100 rows (in other words each column contains 100 elements).<\/p>\n\n<p>Now I need to apply an element-wise multiplication of each column by the vector. What is a correct way to do it in Theano?<\/p>\n\n<p>Should I create a new matrix by repeating my vector 5 times and transposing it and then multiply two matrices of the same shape element-wise?<\/p>\n\n<p><strong>ADDED<\/strong><\/p>\n\n<p>I have learned that in numpy, to achieve the desired behavior, I just need to declare my vector as a 2D array with one column. In other words I need to replace a \"row\"-vector by a \"column\"-vectors (or I need to write the values vertically, instead of horizontally). In this case numpy will broadcast the vector (column) as desired (each column of my matrix will be multiplied by my vector element wise). However, it looks like Theano does not inherit this behavior of numpy:<\/p>\n\n<pre><\/pre>\n\n<p>This is the error that I get:<\/p>\n\n<pre><\/pre>\n\n<p>By the way, the code works if I define  in the following way:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41379690","CreationDate":"2016-12-29T12:24:55.570","ParentId":null,"OwnerUserId":"245549","Title":null,"Body":"<p>The \"native\" solution that I have found is to use <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/extra_ops.html#theano.tensor.extra_ops.repeat\" rel=\"nofollow noreferrer\">theano.tensor.extra_ops.repeat<\/a> operation. In more details, I need to use<\/p>\n\n<pre><code>Xr = T.extra_ops.repeat(X, 3, axis=1)\n<\/code><\/pre>\n\n<p>This operation will repeat the column-vector 3 times. So, as a result we will get a matrix with 3 (identical) columns and this matrix can be multiplied (or summed up) with the <code>W<\/code> matrix element wise.<\/p>\n"}]}
{"QuestionId":41348625,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><keras>","CreationDate":"2016-12-27T16:23:05.313","AcceptedAnswerId":null,"OwnerUserId":7345493.0,"Title":"How to calculate the cosine of two tensors in tensorflow?","Body":"<p>How can I apply Lambda function of Keras in Tensorflow? It is mentioned here -- <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/core\/<\/a> <\/p>\n","answers":[]}
{"QuestionId":41348902,"AnswerCount":0,"Tags":"<python><tensorflow><computer-vision><neural-network><conv-neural-network>","CreationDate":"2016-12-27T16:43:15.907","AcceptedAnswerId":null,"OwnerUserId":6631259.0,"Title":"Imbalanced fine-tuning of inception v3 network (tensorflow)","Body":"<p>I am doing fine-tuning on the pretrained on the Image-net Challenge, inception v3 network of tensorflow. In my classification problem I am using 2 classes because I want to understand if an image contains food or not. My dataset is imbalanced because I have a lot more images on the non-food class (which makes sense I think). When I apply fine-tuning, I have high accuracy on the non-food class and very low accuracy on the food class. Then I am cutting the last 2 inception modules and 2 layers are inserted: a new convolutional layer and a global average pooling layer (and then we have a softmax for output). I apply again a short fine-tuning on the newly introduced layers and now the network gives me high accuracy on the food class and very low accuracy on the non-food. <\/p>\n\n<p>Any ideas of what might be happening?<\/p>\n","answers":[]}
{"QuestionId":41349414,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-27T17:21:31.147","AcceptedAnswerId":41349472.0,"OwnerUserId":1876627.0,"Title":"Tensorflow feed_dict with list of floats and InvalidArgumentError","Body":"<p>Can't feed list to Tensorflow feed_dict. The code:<\/p>\n\n<pre><\/pre>\n\n<p>Gives me:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): You must feed a value\n  for placeholder tensor 'x_in' with dtype float and shape [100]<br>\n  [[Node: x_in = Placeholderdtype=DT_FLOAT, shape=[100],\n  _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n<\/blockquote>\n\n<p>What is wrong in the code, for me looks like it passes <\/p>\n\n<blockquote>\n  <p>placeholder tensor 'x_in' with dtype float and shape [100]<\/p>\n<\/blockquote>\n\n<p>? <\/p>\n","answers":[{"AnswerId":"41349472","CreationDate":"2016-12-27T17:24:59.617","ParentId":null,"OwnerUserId":"1951176","Title":null,"Body":"<p>You should pass a tensor of shape <code>[100]<\/code>, so I guess you wanted to do:<\/p>\n\n<pre><code>for curBatchId in range(batchesCount):\n    batchStart = curBatchId * BATCH_SIZE\n    xf = x_train[batchStart: batchStart + BATCH_SIZE]\n    yf = y_train[batchStart: batchStart + BATCH_SIZE]\n<\/code><\/pre>\n"}]}
{"QuestionId":41350464,"AnswerCount":1,"Tags":"<python><machine-learning><keras><conv-neural-network>","CreationDate":"2016-12-27T18:39:57.073","AcceptedAnswerId":41356777.0,"OwnerUserId":7346966.0,"Title":"keras: how to block convolution layer weights","Body":"<p><a href=\"http:\/\/gist.github.com\/joelouismarino\/a2ede9ab3928f999575423b9887abd14\" rel=\"nofollow noreferrer\">Here<\/a> I have a GoogleNet model for Keras. Is there any possibile way to block the changing of the individual layers of network? I want to block the first two layers of pretrained model from changes.<\/p>\n","answers":[{"AnswerId":"41356777","CreationDate":"2016-12-28T06:48:54.623","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>By 'block the changing of the individual layers' I am assuming you don't want to train those layers, that is you don't want to modify the loaded weights(possibly learnt in previous training). <\/p>\n\n<p>if so you can pass <code>trainable=False<\/code> to the layer and the parameters wont be used for the training update rule. <\/p>\n\n<p>Example:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Activation\n\nmodel = Sequential([\n    Dense(32, input_dim=100),\n    Dense(output_dim=10),\n    Activation('sigmoid'),\n])\n\nmodel.summary()\n\nmodel2 = Sequential([\n    Dense(32, input_dim=100,trainable=False),\n    Dense(output_dim=10),\n    Activation('sigmoid'),\n])\n\nmodel2.summary()\n<\/code><\/pre>\n\n<p>You can see in the model summary for the 2nd model the parameters are counted as Non-trainable ones.<\/p>\n\n<pre><code>____________________________________________________________________________________________________\nLayer (type)                     Output Shape          Param #     Connected to                     \n====================================================================================================\ndense_1 (Dense)                  (None, 32)            3232        dense_input_1[0][0]              \n____________________________________________________________________________________________________\ndense_2 (Dense)                  (None, 10)            330         dense_1[0][0]                    \n____________________________________________________________________________________________________\nactivation_1 (Activation)        (None, 10)            0           dense_2[0][0]                    \n====================================================================================================\nTotal params: 3,562\nTrainable params: 3,562\nNon-trainable params: 0\n____________________________________________________________________________________________________\n____________________________________________________________________________________________________\nLayer (type)                     Output Shape          Param #     Connected to                     \n====================================================================================================\ndense_3 (Dense)                  (None, 32)            3232        dense_input_2[0][0]              \n____________________________________________________________________________________________________\ndense_4 (Dense)                  (None, 10)            330         dense_3[0][0]                    \n____________________________________________________________________________________________________\nactivation_2 (Activation)        (None, 10)            0           dense_4[0][0]                    \n====================================================================================================\nTotal params: 3,562\nTrainable params: 330\nNon-trainable params: 3,232 \n<\/code><\/pre>\n"}]}
{"QuestionId":41351057,"AnswerCount":1,"Tags":"<python><keras><embedding><n-gram>","CreationDate":"2016-12-27T19:28:03.527","AcceptedAnswerId":null,"OwnerUserId":5206473.0,"Title":"Keras word embedding in four gram model","Body":"<p>I am following coursera neural network class and I am trying to pass the assignments using python+keras instead of octave.<\/p>\n\n<p>I want to predict the fourth word given the previous three ones. My input documents total 250 unique words. <\/p>\n\n<p>The model should have an embedding layer that maps each word to a 50-d vector space, a hidden layer with 200 neurons with sigmoid activation function and an output layer of 250 units scoring the probability of the forth word to be equal to those in my vocabulary through a softmax activation.<\/p>\n\n<p>I am having troubles with dimensions. Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Yet I never get to compile the model since I am encountering the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any hint will be much appreciated. Thanks in advance<\/p>\n","answers":[{"AnswerId":"42886945","CreationDate":"2017-03-19T13:07:51.637","ParentId":null,"OwnerUserId":"7105094","Title":null,"Body":"<p>From <a href=\"https:\/\/blog.keras.io\/using-pre-trained-word-embeddings-in-a-keras-model.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/using-pre-trained-word-embeddings-in-a-keras-model.html<\/a><\/p>\n\n<p>\"All that the Embedding layer does is to map the integer inputs to the vectors found at the corresponding index in the embedding matrix, i.e. the sequence [1, 2] would be converted to [embeddings[1], embeddings[2]]. This means that the output of the Embedding layer will be a 3D tensor of shape (samples, sequence_length, embedding_dim).\"<\/p>\n\n<p>Your embedding layer outputs 3 dimension vectors while the dense layers expects 2 dim vecs.<\/p>\n\n<p>You can follow the links tutorial and with some mods it will fit your problem.<\/p>\n"}]}
{"QuestionId":41351322,"AnswerCount":0,"Tags":"<windows-10><theano><python-3.5><theano-cuda>","CreationDate":"2016-12-27T19:48:31.540","AcceptedAnswerId":null,"OwnerUserId":7099374.0,"Title":"Troubles with runnig theano on GPU, Windows 10","Body":"<p>I am having trouble running Theano testing code on GPU. I have followed the official Theano installation guides, and after that didn't work i followed this guide:<a href=\"https:\/\/github.com\/philferriere\/dlwin\" rel=\"nofollow noreferrer\">https:\/\/github.com\/philferriere\/dlwin<\/a> . However i am still not able to run the code on GPU plus some strange behaviors are exhibited.<\/p>\n\n<p>Test example i want to run:<\/p>\n\n<pre><\/pre>\n\n<p>When running in Pycharm i get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>But when i just try to import theano from the console i get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I installed latest 64 bit version of cuda (cuda-8.0.44), I have a 64 bit python 3.5.2, theano realease 0.8.2, NVIDIA GTX 680, Microsoft visual studio 2015.<\/p>\n\n<p>I have been bothering for 24 hours with this and i am running out of ideas.\nThank you in advance for the help !<\/p>\n\n<p>edit: Import of theano in PyCharm works fine and i am using the same python in both cases<\/p>\n","answers":[]}
{"QuestionId":41351390,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2016-12-27T19:54:53.310","AcceptedAnswerId":41383119.0,"OwnerUserId":6786718.0,"Title":"Do I have to use a Scale-Layer after every BatchNorm Layer?","Body":"<p>I am using caffe , in detail pycaffe, to create my neuronal network. I noticed that I have to use BatchNormLayer to get a positive result. I am using the Kappa-Score as a result matrix. \nI now have seen several different locations for the BatchNorm-Layers in my network. But I came across the ScaleLayer, too which is not in the Layer Catalogue but gets often mentioned with the BatchNorm Layer<\/p>\n\n<p>Do you always need to put a ScaleLayer after a BatchNorm - Layer and what does it do?<\/p>\n","answers":[{"AnswerId":"41353237","CreationDate":"2016-12-27T22:58:07.503","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>In general, you will get no benefit from a scale layer juxtaposed with batch normalization.  Each is a linear transformation.  Where BatchNorm translates so that the new distribution has a mean of 0 and variance of 1, Scale compresses the entire range into a specified interval, typically [0,1].  Since they're both linear transformations, if you do them in sequence, the second will entirely undo the work of the first.<\/p>\n\n<p>They also deal somewhat differently with outliers.  Consider a set of data: ten values, five each of -1 and +1.  BatchNorm will not change this at all: it already has mean 0 and variance 1.  For consistency, let's specify the same interval for Scale, [-1, 1], which is also a popular choice.<\/p>\n\n<p>Now, add an outlier of, say 99 to the mix.  Scale will transform the set to the range [-1, 1] so that there are now five -1.00 values, one +1.00 value (the former 99), and five values of -0.96 (formerly +1).<\/p>\n\n<p>BatchNorm worries about the mean standard deviation, not the max and min values.  The new mean is +9; the S.D. is 28.48 (rounding everything to 2 decimal places).  The numbers will be scaled to be roughly five values each of -.35 and -.28, and one value of 3.16<\/p>\n\n<p>Whether one scaling works better than the other depends much on the skew and scatter of your distribution.  I prefer BatchNorm, as it tends to differentiate better in dense regions of a distribution.<\/p>\n"},{"AnswerId":"41383119","CreationDate":"2016-12-29T16:00:44.750","ParentId":null,"OwnerUserId":"733313","Title":null,"Body":"<p>From the original <a href=\"https:\/\/arxiv.org\/abs\/1502.03167\" rel=\"nofollow noreferrer\">batch normalization paper<\/a> by Ioffe &amp; Szegedy: \"we make sure that the transformation inserted in the network can represent the identity transform.\" Without the Scale layer after the BatchNorm layer, that would not be the case because the Caffe BatchNorm layer has no learnable parameters.<\/p>\n\n<p>I learned this from the <a href=\"https:\/\/github.com\/KaimingHe\/deep-residual-networks\" rel=\"nofollow noreferrer\">Deep Residual Networks git repo<\/a>; see item 6 under disclaimers and known issues there.<\/p>\n"}]}
{"QuestionId":41351723,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><regression><linear-regression><logistic-regression>","CreationDate":"2016-12-27T20:21:21.527","AcceptedAnswerId":41351963.0,"OwnerUserId":2155026.0,"Title":"Loading Boston housing dataset using TensorFlow","Body":"<p>I am trying to understand the code example <a href=\"http:\/\/Deep%20Neural%20Network%20Regression%20with%20Boston%20Data\" rel=\"nofollow noreferrer\">Deep Neural Network Regression with Boston Data<\/a>.<\/p>\n\n<p>The dataset is described <a href=\"https:\/\/archive.ics.uci.edu\/ml\/datasets\/Housing\" rel=\"nofollow noreferrer\">here<\/a>. It has 14 attributes.<\/p>\n\n<p>The example uses the following code to load the data.<\/p>\n\n<pre><\/pre>\n\n<p>When I want to know more about  and , I have the following.<\/p>\n\n<pre><\/pre>\n\n<p>My questions:<\/p>\n\n<ol>\n<li>Why the dataset has been divided into two objects one with 13 attributes and the other with 1?<\/li>\n<li>On what basis this division has been made?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"41351963","CreationDate":"2016-12-27T20:41:46.217","ParentId":null,"OwnerUserId":"7245907","Title":null,"Body":"<p>The 13 columns in boston.data are your features. The 1 column in boston.target is your target. The reason the split is done is because most of the time, machine learning algorithms require both features and targets as separate data structures. The load_datasets function is just making it easier on you by splitting off the MDEV column, because most of the time, that's the feature that people want to predict on. To put it another way, the designers of load_data sets are assuming you want to try and find the median home prices based on the other 13 features. <\/p>\n\n<p>You don't have to do this. You could choose any of the features as your target. Say you wanted to predict RM, the average number of rooms per dwelling. Just merge the MDEV column back into boston.data and split out RM. Then use RM as your target. <\/p>\n\n<p>BTW, the link you provided was broken, so I google it and came up with this <a href=\"http:\/\/cs.smith.edu\/dftwiki\/index.php\/Tutorial:_Playing_with_the_Boston_Housing_Data\" rel=\"nofollow noreferrer\">Boston Housing price tutorial<\/a>. It looks pretty complete if you want to do regression in tensorflow<\/p>\n"}]}
{"QuestionId":41352332,"AnswerCount":1,"Tags":"<deep-learning><normalization><caffe><pycaffe>","CreationDate":"2016-12-27T21:17:23.323","AcceptedAnswerId":41401763.0,"OwnerUserId":1436964.0,"Title":"Faster-RCNN bbox\/image normalization","Body":"<p>I am playing with py-faster-rcnn on a custom dataset (about 3000 images, 7 different classes, including the background), and following these tutorials:<\/p>\n\n<p><a href=\"https:\/\/github.com\/zeyuanxy\/fast-rcnn\/blob\/master\/help\/train\/README.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/zeyuanxy\/fast-rcnn\/blob\/master\/help\/train\/README.md<\/a> (Fast-RCNN tutorial)\n<a href=\"https:\/\/github.com\/deboc\/py-faster-rcnn\/tree\/master\/help\" rel=\"nofollow noreferrer\">https:\/\/github.com\/deboc\/py-faster-rcnn\/tree\/master\/help<\/a> (Faster-RCNN tutorial)<\/p>\n\n<p>I am using the end2end solution with VGG16 network.\nEverything works fine, expect my results so I have some questions:<\/p>\n\n<ul>\n<li>What kind of normalizations are needed on the images and on the bbox annotations? <\/li>\n<li>It is similar to the previous question: There are two config options: BBOX_NORMALIZE_TARGETS and BBOX_NORMALIZE_TARGETS_PRECOMPUTED. Should I calculate the mean and std before the training and use these options for bbox normalization?<\/li>\n<li>I modified the num_output at the cls_score and bbox_pred layers (according to this thread: <a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\/issues\/1\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rbgirshick\/py-faster-rcnn\/issues\/1<\/a>), but in the end2end solution there are rpn_cls_score and rpn_bbox_pred layers too. Should I modify the num_outputs of these too? If I should then how could I calculate the number of outputs for 7 classes?    <\/li>\n<\/ul>\n","answers":[{"AnswerId":"41401763","CreationDate":"2016-12-30T19:34:39.973","ParentId":null,"OwnerUserId":"628538","Title":null,"Body":"<p>No, you do not need to pre-compute anything. In lib\/roi_data_layer\/roidb.py, it computes the mean and standard deviation for your dataset if you set the BBOX_NORMALIZE_TARGETS_PRECOMPUTED to False, otherwise, it will use the default values which are specified in lib\/fast_rcnn\/config.py. RPN is agnostic to number of classes. It only treats regions which contain any object as positive and everything else as negative.<\/p>\n"}]}
{"QuestionId":41352745,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2016-12-27T22:03:36.803","AcceptedAnswerId":41352964.0,"OwnerUserId":3993270.0,"Title":"TensorFlow: tf.placeholder and tf.Variable - why is the dimension not required?","Body":"<p>I am learning TensorFlow from the example at: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/2_BasicModels\/linear_regression.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/2_BasicModels\/linear_regression.ipynb<\/a><\/p>\n\n<p>I have a few questions in the code below:\nWhen defining the placeholders and variables like X, Y, W, and b, why didn't we need to specify the dimension of them? How would the code allocate memory without knowing the size of these placeholders\/variables? Thanks! <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41352964","CreationDate":"2016-12-27T22:28:19.457","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/placeholders#placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a> tensors do not require you to specify a shape, in order to allow you to <em>feed<\/em> tensors of different shapes in a later <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session.run\" rel=\"noreferrer\"><code>tf.Session.run()<\/code><\/a> call. By default, a placeholder has a completely unconstrained shape, but you can constrain it by passing the optional <code>shape<\/code> argument. For example:<\/p>\n\n<pre><code>w = tf.placeholder(tf.float32)                      # Unconstrained shape\nx = tf.placeholder(tf.float32, shape=[None, None])  # Matrix of unconstrained size\ny = tf.placeholder(tf.float32, shape=[None, 32])    # Matrix with 32 columns\nz = tf.placeholder(tf.float32, shape=[128, 32])     # 128x32-element matrix\n<\/code><\/pre>\n\n<p>When you create a placeholder, TensorFlow does not allocate any memory. Instead, when you feed the placeholder, in the call to <code>tf.Session.run()<\/code>, TensorFlow will allocate appropriately sized memory for the input (and subsequently for any necessary intermediate) tensors.<\/p>\n\n<p>Note that <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/variables#Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a> objects typically do require a shape when you create them, and this shape is inferred from the first argument to the initializer. In your program, <code>rng.randn()<\/code> (an alias for <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.random.randn.html\" rel=\"noreferrer\"><code>numpy.random.randn()<\/code><\/a>) returns a scalar value, and so the variables <code>W<\/code> and <code>b<\/code> will have scalar shape.<\/p>\n\n<p>Although the placeholders in your code (<code>X<\/code> and <code>Y<\/code>) have unconstrained shape, some of the operators, such as <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/arithmetic_operators#add\" rel=\"noreferrer\"><code>tf.add()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/arithmetic_operators#multiply\" rel=\"noreferrer\"><code>tf.mul()<\/code><\/a>, have additional requirements about the shape of their arguments (viz. that they are compatible with the <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"noreferrer\">NumPy broadcasting rules<\/a>). Since TensorFlow doesn't know when you build the graph what the actual shapes of those tensors will be, it trusts that the user knows what they are doing, and performs the check dynamically (during the call to <code>tf.Session.run()<\/code>). If instead you constrain the shapes of the placeholders, you enable TensorFlow to perform some checks earlier, and doing this can help to reduce bugs.<\/p>\n"}]}
{"QuestionId":41352782,"AnswerCount":1,"Tags":"<tensorflow><artificial-intelligence><bazel>","CreationDate":"2016-12-27T22:08:20.650","AcceptedAnswerId":null,"OwnerUserId":505595.0,"Title":"Errors when trying to build label_image neural net with bazel","Body":"<p>Environment info<\/p>\n\n<p>Operating System: El Capitan, 10.11.1<\/p>\n\n<p>I'm doing this tutorial: <a href=\"https:\/\/petewarden.com\/2016\/09\/27\/tensorflow-for-mobile-poets\/\" rel=\"nofollow noreferrer\">https:\/\/petewarden.com\/2016\/09\/27\/tensorflow-for-mobile-poets\/<\/a>\nTrying to classify images using tensorflow on iOS app.\nWhen I try to build my net using bazel:\nbazel build tensorflow\/examples\/label_image:label_image<\/p>\n\n<p>I get these errors:\n<a href=\"https:\/\/gist.github.com\/galharth\/36b8f6eeb12f847ab120b2642083a732\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/galharth\/36b8f6eeb12f847ab120b2642083a732<\/a><\/p>\n","answers":[{"AnswerId":"41371908","CreationDate":"2016-12-29T01:11:28.347","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>From the related github issue <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6487\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6487<\/a> I think we narrowed it down to a lack of resources on the virtual machine. Bazel tends to get flakey with only 2GB of RAM allocated to it.<\/p>\n"}]}
{"QuestionId":41352863,"AnswerCount":0,"Tags":"<tensorflow><neural-network><deep-learning><keras><conv-neural-network>","CreationDate":"2016-12-27T22:17:05.047","AcceptedAnswerId":null,"OwnerUserId":4541216.0,"Title":"keras ImageDataGenerator: size of augmented data?","Body":"<p>I was trying to train some CNN using keras by first augmenting my image dataset (~2GB). When I executed:<\/p>\n\n<pre><\/pre>\n\n<p>There appeared to be a . (When I train my model without data augmentation, everything went smoothly.) But I couldn't figure out how much augmentation is allowed in my case. Could anyone kindly let me know how to calculate the size of the resulting augmented data argument by argument within the ?<\/p>\n\n<p>Thanks so much in advance.<\/p>\n","answers":[]}
{"QuestionId":41352976,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-27T22:29:04.467","AcceptedAnswerId":41530132.0,"OwnerUserId":3606192.0,"Title":"How to install TensorFlow Developer version for both Python2 and Python3","Body":"<p>I was working on the TensorFlow Dev version, and when running the build tests, I can run them only using Python 2 or Python 3 but not both. Currently I am using a soolution where I have two virtual environments, one for 2 and one for 3. This is inconvenient because whenever I want to test my changes to the TF, I have to migrate those changes to the other environment and run them there separately. This is error prone, and really tedious.<\/p>\n\n<p>Another approach I have tried before is running either  or , depending on what I am testing. This is very time consuming, and also not convenient.<\/p>\n\n<p>I know that TF dev files are linked, and not copied.\nThe question is if there is a way to follow the instructions <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#setting_up_tensorflow_for_development\" rel=\"nofollow noreferrer\">here<\/a> but installing the dev for both python2 and python3. <\/p>\n","answers":[{"AnswerId":"41530132","CreationDate":"2017-01-08T06:29:58.787","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>By Dev files, do you mean install the source code?\nYou can always rerun <\/p>\n\n<blockquote>\n  <p>.\/configure<\/p>\n<\/blockquote>\n\n<p>And configure your source to use python3. You cannot tell your source to use both at the same time, but you can switch between different python versions whenever you like using configure. However this will run bazel clean, which will destroy your bazel cache.\nTo avoid that, you may just comment out the call to bazel_clean_and_fetch in configure script. In r0.12 branch it is here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.12\/configure#L146\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.12\/configure#L146<\/a><\/p>\n\n<p>This can work, however you may also see some issues due to reusing bazel cache.<\/p>\n"}]}
{"QuestionId":41352991,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-27T22:30:17.273","AcceptedAnswerId":41353047.0,"OwnerUserId":5605365.0,"Title":"Tensorflow AttributeError: 'DataSet' object has no attribute 'image'","Body":"<p>I am trying to use Tensorflow for the first time<\/p>\n\n<p>Here is my code that I got from a tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>I get this error at the line specified above <\/p>\n\n<p>How do I fix this?<\/p>\n","answers":[{"AnswerId":"41353047","CreationDate":"2016-12-27T22:36:28.677","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The MNIST <code>DataSet<\/code> object (implemented <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/13f9309ccb063a58b0ce34aafc23f93a49e33733\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py\" rel=\"noreferrer\">here<\/a>) does not have an <code>image<\/code> property, but it does have <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/13f9309ccb063a58b0ce34aafc23f93a49e33733\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L145\" rel=\"noreferrer\">an <code>images<\/code> property<\/a>. The following change should fix things: <\/p>\n\n<pre><code>data = mnist.train.images\n<\/code><\/pre>\n"}]}
{"QuestionId":41353079,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-27T22:39:59.490","AcceptedAnswerId":null,"OwnerUserId":6996238.0,"Title":"Tensorflow image.central_crop (mis)behavior","Body":"<p>In the Tensorflow documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/image\/cropping#central_crop\" rel=\"nofollow noreferrer\">tf.image.central_crop<\/a> function:<\/p>\n\n<pre><\/pre>\n\n<p>Consider the following code:<\/p>\n\n<pre><\/pre>\n\n<p>Original image size is 456x450<\/p>\n\n<pre><\/pre>\n\n<p>Crop size is 228x226<\/p>\n\n<p>Which gives area ratio of:<\/p>\n\n<pre><\/pre>\n\n<p>Not 0.5 as I expected. Can someone help to clarify this?<\/p>\n","answers":[{"AnswerId":"41371684","CreationDate":"2016-12-29T00:37:19.940","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>As @Wes Doyle pointed out, it returns 50% of each dimension, which is consistent with the illustration in the comments, the original size is 4x8 (4 \"|\" signs and 8 \"-\" signs), and the cropped size is 2x4 (2 \"X\" vertically, 4 \"X\" horizontally).<\/p>\n"}]}
{"QuestionId":41353169,"AnswerCount":6,"Tags":"<numpy><tensorflow><tensorboard>","CreationDate":"2016-12-27T22:50:18.903","AcceptedAnswerId":null,"OwnerUserId":7071746.0,"Title":"tensorboard with numpy array","Body":"<p>Can someone give a example on how to use tensorboard visualize numpy array value?<\/p>\n\n<p>There is a related question here, I don't really get it.\n<a href=\"https:\/\/stackoverflow.com\/questions\/38060078\/tensorboard-logging-non-tensor-numpy-information-auc\">Tensorboard logging non-tensor (numpy) information (AUC)<\/a><\/p>\n\n<p>For example, \nIf I have <\/p>\n\n<pre><\/pre>\n\n<p>How can I keep tracking the distribution of foo using tensorboard for 100 iterations? Can someone give a code example?\nThanks. <\/p>\n","answers":[{"AnswerId":"50199025","CreationDate":"2018-05-06T11:12:35.870","ParentId":null,"OwnerUserId":"4830338","Title":null,"Body":"<p>You could define a function like this (taken from gyglim's gist):<\/p>\n\n<pre><code>def add_histogram(writer, tag, values, step, bins=1000):\n    \"\"\"\n    Logs the histogram of a list\/vector of values.\n    From: https:\/\/gist.github.com\/gyglim\/1f8dfb1b5c82627ae3efcfbbadb9f514\n    \"\"\"\n\n    # Create histogram using numpy\n    counts, bin_edges = np.histogram(values, bins=bins)\n\n    # Fill fields of histogram proto\n    hist = tf.HistogramProto()\n    hist.min = float(np.min(values))\n    hist.max = float(np.max(values))\n    hist.num = int(np.prod(values.shape))\n    hist.sum = float(np.sum(values))\n    hist.sum_squares = float(np.sum(values ** 2))\n\n    # Requires equal number as bins, where the first goes from -DBL_MAX to bin_edges[1]\n    # See https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/summary.proto#L30\n    # Therefore we drop the start of the first bin\n    bin_edges = bin_edges[1:]\n\n    # Add bin edges and counts\n    for edge in bin_edges:\n        hist.bucket_limit.append(edge)\n    for c in counts:\n        hist.bucket.append(c)\n\n    # Create and write Summary\n    summary = tf.Summary(value=[tf.Summary.Value(tag=tag, histo=hist)])\n    writer.add_summary(summary, step)\n<\/code><\/pre>\n\n<p>And then add to the summary writer like this:<\/p>\n\n<pre><code>add_histogram(summary_writer, \"Histogram_Name\", your_numpy_array, step)\n<\/code><\/pre>\n"},{"AnswerId":"42314953","CreationDate":"2017-02-18T12:11:15.127","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>Another (simplest) way is just using placeholders. First, you can make a placeholder for your numpy array shape.<\/p>\n\n<pre><code># Some place holders for summary\nsummary_reward = tf.placeholder(tf.float32, shape=(), name=\"reward\")\ntf.summary.scalar(\"reward\", summary_reward)\n<\/code><\/pre>\n\n<p>Then, just call session.run the merged summary with the feed_dict.<\/p>\n\n<pre><code># Summary\nsumm = tf.summary.merge_all()\n...\ns = sess.run(summ, feed_dict={summary_reward: reward})\nwriter.add_summary(s, i)\n<\/code><\/pre>\n"},{"AnswerId":"41355215","CreationDate":"2016-12-28T04:10:31.920","ParentId":null,"OwnerUserId":"7071746","Title":null,"Body":"<p>Found a way to work around, create a variable and assign the value of numpy array to the variable, use tensorboard to track the variable<\/p>\n\n<pre><code>mysummary_writer = tf.train.SummaryWriter(\".\/tmp\/test\/\")\na = tf.Variable(tf.zeros([3,2]), name=\"a\")\nsum1 = tf.histogram_summary(\"nparray1\", a)\nsummary_op = tf.merge_all_summaries()\nsess = tf.Session()\n\nsess.run(tf.initialize_all_variables())\n\nfor ii in range(10):\n    foo = np.random.rand(3, 2)\n    assign_op = a.assign(foo)\n    summary, _ = sess.run([summary_op, assign_op])\n    mysummary_writer.add_summary(tf.Summary.FromString(summary), global_step=ii)\n    mysummary_writer.flush()\n<\/code><\/pre>\n"},{"AnswerId":"41356503","CreationDate":"2016-12-28T06:26:27.643","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>For simple values (scalar), you can use this recipe<\/p>\n\n<pre><code>summary_writer = tf.train.SummaryWriter(FLAGS.logdir)\nsummary = tf.Summary()\nsummary.value.add(tag=tagname, simple_value=value)\nsummary_writer.add_summary(summary, global_step)\nsummary_writer.flush()\n<\/code><\/pre>\n\n<p>As far as using array, perhaps you can add 6 values in a sequence, ie<\/p>\n\n<pre><code>for value in foo:\n  summary.value.add(tag=tagname, simple_value=value)\n<\/code><\/pre>\n"},{"AnswerId":"48485773","CreationDate":"2018-01-28T11:22:17.413","ParentId":null,"OwnerUserId":"6540428","Title":null,"Body":"<pre><code>sess = tf.Session()\nwriter = tf.summary.FileWriter('tensorboard_test')\nvar = tf.Variable(0.0,trainable=False,name='loss')\nsess.run(var.initializer)\nsummary_op = tf.summary.scalar('scalar1',var)\n\nfor value in array:    \n    sess.run(var.assign(value))\n    summary = sess.run(summary_op)\n    writer.add_summary(summary,i)\n<\/code><\/pre>\n\n<p>It works, but slow.<\/p>\n"},{"AnswerId":"46396095","CreationDate":"2017-09-25T00:28:05.253","ParentId":null,"OwnerUserId":"4511978","Title":null,"Body":"<p>if you install <a href=\"https:\/\/github.com\/lanpa\/tensorboardX\" rel=\"nofollow noreferrer\">this package<\/a> via <code>pip install tensorboard-pytorch<\/code> it becomes as straightforward as it can get:<\/p>\n\n<pre><code>import numpy as np\nfrom tensorboardX import SummaryWriter\nwriter = SummaryWriter()\nfor i in range(50):\n    writer.add_histogram(\"moving_gauss\", np.random.normal(i, i, 1000), i, bins=\"auto\")\nwriter.close()\n<\/code><\/pre>\n\n<p>Will generate the corresponding histogram data in the <code>runs<\/code> directory:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/p7aBW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/p7aBW.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":41353436,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2016-12-27T23:24:02.227","AcceptedAnswerId":41353669.0,"OwnerUserId":3993270.0,"Title":"TensorFlow: function reduce_sum input","Body":"<p>I am learning TensorFlow from the example at: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/2_BasicModels\/linear_regression.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/2_BasicModels\/linear_regression.ipynb<\/a><\/p>\n\n<p>I have a question in the code below:<\/p>\n\n<pre><\/pre>\n\n<p>The input of tf.reduce_sum is tf.pow(pred-Y, 2)) which seems to be a scalar (or am I wrong?) Then I am wondering why we want to do reduce_sum on a scalar? What did I miss here? Thanks!<\/p>\n","answers":[{"AnswerId":"41353543","CreationDate":"2016-12-27T23:39:49.730","ParentId":null,"OwnerUserId":"5270944","Title":null,"Body":"<p>I'm quite new to Tensorflow too, but it looks like you're right. At the moment the training only runs one sample at a time, so the reduce_sum call would effectively be useless. <\/p>\n\n<p>Try dropping the call to reduce_sum and replacing it with this:<\/p>\n\n<pre><code>cost = tf.pow(pred-Y, 2)\/(2*n_samples)\n<\/code><\/pre>\n\n<p>It should still function the same way, but break if you try training on batches rather than single samples.<\/p>\n"},{"AnswerId":"41353669","CreationDate":"2016-12-27T23:55:58.367","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The tensor <code>pred<\/code> has a <strong>statically<\/strong> unknown shape, but&mdash;due to the <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">broadcasting semantics<\/a> of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/arithmetic_operators#add\" rel=\"nofollow noreferrer\"><code>tf.add()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/arithmetic_operators#multiply\" rel=\"nofollow noreferrer\"><code>tf.mul()<\/code><\/a>&mdash;it will have the same <strong>dynamic<\/strong> shape as the value fed to placeholder <code>X<\/code> (which also has a statically unknown shape).<\/p>\n\n<p>In the tutorial, when the model is being trained, <code>X<\/code> is fed with scalar values, and so <code>pred<\/code> will be a scalar (and the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/reduction#reduce_sum\" rel=\"nofollow noreferrer\"><code>tf.reduce_sum()<\/code><\/a> will have no effect):<\/p>\n\n<pre><code># Fit all training data\nfor epoch in range(training_epochs):\n    for (x, y) in zip(train_X, train_Y):\n        sess.run(optimizer, feed_dict={X: x, Y: y})\n<\/code><\/pre>\n\n<p>When generating the log message, <code>X<\/code> is fed with a vector (containing all of the training examples), so <code>pred<\/code> will also be a vector of the same length, and <code>tf.reduce_sum()<\/code> will aggregate the cost down to a scalar:<\/p>\n\n<pre><code>    #Display logs per epoch step\n    if (epoch+1) % display_step == 0:\n        c = sess.run(cost, feed_dict={X: train_X, Y:train_Y})\n<\/code><\/pre>\n\n<p>This example is a little strange, because we typically train TensorFlow models with a <a href=\"http:\/\/sebastianruder.com\/optimizing-gradient-descent\/index.html#minibatchgradientdescent\" rel=\"nofollow noreferrer\">mini-batch<\/a> of examples, but it demonstrates the usefulness of allowing a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/placeholders#placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a> to have a statically undefined shape.<\/p>\n"}]}
{"QuestionId":41353451,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2016-12-27T23:25:26.267","AcceptedAnswerId":41415749.0,"OwnerUserId":5270944.0,"Title":"What are all my variables duplicated in Tensorboard?","Body":"<p>I'm new to Tensorflow and am running a basic CNN. As a way of visualising the training process, I build a summary with loss and accuracy in order to view later in Tensorboard like this:<\/p>\n\n<pre><\/pre>\n\n<p>I initialise the summaries as below. <em>(get_logdir_string() returns a unique string composed of the given parameter and the current datetime)<\/em><\/p>\n\n<pre><\/pre>\n\n<p>Then for each mini-batch iteration I do this:<\/p>\n\n<pre><\/pre>\n\n<p>Then I run Tensorboard and am presented with something like the following, where only the first of each variable (without the suffix) contains data:\n<a href=\"https:\/\/i.stack.imgur.com\/Ssh8c.png\" rel=\"nofollow noreferrer\">Screenshot<\/a><\/p>\n\n<p>Has anyone encountered this before? Thanks!<\/p>\n","answers":[{"AnswerId":"41415749","CreationDate":"2017-01-01T13:17:29.087","ParentId":null,"OwnerUserId":"5270944","Title":null,"Body":"<p>I found the culprit and thought I'd post it here for future reference.<\/p>\n\n<p>It turns out the I needed to call <code>tf.reset_default_graph()<\/code> before each run.<\/p>\n"}]}
{"QuestionId":41353885,"AnswerCount":1,"Tags":"<python><numpy><graphics><rotation><theano>","CreationDate":"2016-12-28T00:31:56.443","AcceptedAnswerId":41355790.0,"OwnerUserId":3697202.0,"Title":"Theano function using individual elements of input","Body":"<p>I am trying to build a Theano function that takes a  of Euler angles as input and returns a directional vector corresponding to those Euler angles. First, I take the sines and cosines of each element of the vector, then I arrange these into a rotation matrix. Finally, I multiply the directional vector  by this rotation matrix. The problem I am running into is that I can't multiply this NumPy array by the rotation matrix.<\/p>\n\n<p>This is my code:<\/p>\n\n<pre><\/pre>\n\n<p>The second-to-last line throws this error:<\/p>\n\n<pre><\/pre>\n\n<p>I can't think of any way to create this rotation matrix through matrix operations on the Euler angles. How can I create this function in a format that Theano can compile?<\/p>\n","answers":[{"AnswerId":"41355790","CreationDate":"2016-12-28T05:20:38.883","ParentId":null,"OwnerUserId":"772649","Title":null,"Body":"<p>use <code>theano.tensor.stacklists()<\/code>.<\/p>\n\n<pre><code>rot_matrix = T.stacklists([[...], ...])\n<\/code><\/pre>\n"}]}
{"QuestionId":41354258,"AnswerCount":0,"Tags":"<tensorflow><neural-network><conv-neural-network>","CreationDate":"2016-12-28T01:36:33.800","AcceptedAnswerId":null,"OwnerUserId":5270944.0,"Title":"Why do my loss and accuracy seem to be unrelated?","Body":"<p>I have a CNN for predicting CIFAR10 classes, which I've adapted from a working solution for the MNIST data-set. <\/p>\n\n<p>The model is <em>image->conv->pool->conv->pool->FC->dropout->FC->class_prediction<\/em>, which I assume would do substantially better than a fully-connected net (which I was achieving ~50% with when doing the <a href=\"http:\/\/cs231n.github.io\" rel=\"nofollow noreferrer\">CS231 course<\/a>).<\/p>\n\n<p>I'm receiving some strange results, where the accuracy and loss don't really seem to be linked to one-another: <a href=\"https:\/\/i.stack.imgur.com\/ihXR4.png\" rel=\"nofollow noreferrer\">training accuracy and loss graph<\/a><\/p>\n\n<p>As shown, the loss keeps dropping, which is a good thing, but it doesn't seem to really affect the accuracy, even on the training set.<\/p>\n\n<p>Sorry for dumping this all here, but below I've extracted all the code I believe to be relevant. So please let me know if I can add any further information.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41354261,"AnswerCount":1,"Tags":"<python><performance><tensorflow>","CreationDate":"2016-12-28T01:37:13.107","AcceptedAnswerId":null,"OwnerUserId":5566910.0,"Title":"Tensorflow training becomes slower and slower when iteration is more than 10,000. Why?","Body":"<p>I feed the data to the graph with input pipeline methods, and  is implemented to generate batch data. However, as the training progresses, tensorflow becomes slower and slower for later iterations. I am confused about what's the essential reason leading to it? Thanks very much! My code snippet is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"48938773","CreationDate":"2018-02-22T23:13:37.577","ParentId":null,"OwnerUserId":"9161637","Title":null,"Body":"<p>When training you should do sess.run only once.\nRecommend trying something like this, hope it helps:<\/p>\n\n<pre><code>with tf.Session() as sess:\n  sess.run(tf.global_variables_initializer())\n  for i in range(train_set_size * epoch):\n    train_step.run([optimizer, depth_loss, prediction])\n<\/code><\/pre>\n"}]}
{"QuestionId":41354411,"AnswerCount":0,"Tags":"<c++><tensorflow><protocol-buffers>","CreationDate":"2016-12-28T02:01:23.320","AcceptedAnswerId":null,"OwnerUserId":7347887.0,"Title":"How to check node name of tensorflow graph (protocol buffers) by c++","Body":"<p>I want to check the loading graph is correct.<\/p>\n\n<p>I save the learned protocol buffers file by python.\nAnd, I load the protocol buffers file by c++.<\/p>\n\n<p>But I can't get the output tensor when the session run.<\/p>\n\n<p>I want to output and check the graph information.<\/p>\n\n<p><strong>Saveing code by python<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Loading code by C++<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Runing code by c++<\/strong><\/p>\n\n<p>input_node_name = \"input\"<\/p>\n\n<p>output_node_name = \"output_lay\/output_lay\"<\/p>\n\n<pre><\/pre>\n\n<p><strong>result<\/strong> is<\/p>\n\n<p>Invalid argument: Incompatible shapes: <strong>[77,1,513,16] vs. [13780,1,513,16]<\/strong>\n     [[Node: conv1\/dropout\/mul = Mul[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](conv1\/dropout\/Div, conv1\/dropout\/Floor)]]<\/p>\n\n<p>ex) if not use the cnn ( the 3 layer perceptron network)\nInvalid argument: Incompatible shapes: <strong>[77,600] vs. [10242,600]<\/strong>\n     [[Node: hidden1\/dropout\/mul = Mul[T=DT_FLOAT, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](hidden1\/dropout\/Div, hidden1\/dropout\/Floor)]]<\/p>\n\n<p><strong>pickup of the output 1st python code<\/strong><\/p>\n\n<ul>\n<li>input<\/li>\n<li>Placeholder<\/li>\n<li>Reshape\/shape<\/li>\n<li>Reshape<\/li>\n<li>conv1\/weights<\/li>\n<li>conv1\/biases<\/li>\n<li>conv1\/Conv2D<\/li>\n<li>conv1\/MaxPool<\/li>\n<li>conv1\/Add<\/li>\n<li>conv1\/conv1<\/li>\n<li>conv1\/dropout\/Shape<\/li>\n<li>conv1\/dropout\/add<\/li>\n<li>conv1\/dropout\/Floor<\/li>\n<li>conv1\/dropout\/Div<\/li>\n<li>conv1\/dropout\/mul<\/li>\n<li>conv2\/weights<\/li>\n<li>conv2\/biases<\/li>\n<li>conv2\/Conv2D<\/li>\n<li>...<\/li>\n<li>Reshape_1\/shape<\/li>\n<li>Reshape_1<\/li>\n<li>hidden1\/weights<\/li>\n<li>hidden1\/biases<\/li>\n<li>hidden1\/MatMul<\/li>\n<li>hidden1\/Add<\/li>\n<li>hidden1\/hidden1<\/li>\n<li>hidden1\/dropout\/add<\/li>\n<li>hidden1\/dropout\/Floor<\/li>\n<li>hidden1\/dropout\/Div<\/li>\n<li>hidden1\/dropout\/mul<\/li>\n<li>hidden2\/weights<\/li>\n<li>...<\/li>\n<li>output_lay\/weights<\/li>\n<li>output_lay\/biases<\/li>\n<li>output_lay\/MatMul<\/li>\n<li>output_lay\/Add<\/li>\n<li>output_lay\/output_lay<\/li>\n<\/ul>\n","answers":[]}
{"QuestionId":41354691,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><regression><data-science>","CreationDate":"2016-12-28T02:45:24.800","AcceptedAnswerId":null,"OwnerUserId":2155026.0,"Title":"TensorFlow: Invalid argument error in DNNRegressor.fit()","Body":"<p>I am trying to do regression for my own data following the example of  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/boston.py\" rel=\"nofollow noreferrer\">Deep Neural Network Regression with Boston Data<\/a>.<\/p>\n\n<p>Following is my code.<\/p>\n\n<pre><\/pre>\n\n<p>I am getting lots of warning and invalid argument error. The complete console output is available <a href=\"https:\/\/drive.google.com\/file\/d\/0ByGeTvEEJEJJdVUtbkJFUjhBRHc\/view?usp=sharing\" rel=\"nofollow noreferrer\">here<\/a>. I would like to put the highlights of the error messages below.<\/p>\n\n<pre><\/pre>\n\n<p>Is this because I am using some deprecated TensorFlow API?<\/p>\n","answers":[]}
{"QuestionId":41354830,"AnswerCount":2,"Tags":"<arrays><keras>","CreationDate":"2016-12-28T03:03:29.127","AcceptedAnswerId":null,"OwnerUserId":5225453.0,"Title":"Error when checking model target: expected dense_24 to have shape...but got array with shape... in Keras","Body":"<p>I can't understand what this error message is trying to tell me.<\/p>\n\n<p>Looking at the summary of my model, I am looking at the last few layers . <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZZ1sa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZZ1sa.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>But when I fit my model, I get the this:  <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ww6pT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ww6pT.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I don't understand. If I read this closely, Keras seems to be saying, \"I looked at the labels (targets) for the validation set, and sensibly, this made me think the last layer in your model should be shaped (None, 2). But, instead of actually getting a last layer shaped (None, 2) in the model, the layer consisted of an actual array. Which was of some other shape.\" <\/p>\n\n<p>This makes no sense.<\/p>\n\n<p>I think that I suspect that this error really should say, in general, is:  <\/p>\n\n<p> . <\/p>\n\n<p>Does anyone care to agree or disagree?  Thanks.<\/p>\n\n<p>(There's a similar question <a href=\"https:\/\/stackoverflow.com\/questions\/39335434\/exception-error-when-checking-model-target-expected-dense-3-to-have-shape-non\">here<\/a>, but not very helpful.)<\/p>\n","answers":[{"AnswerId":"43850614","CreationDate":"2017-05-08T14:44:48.280","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>I'm not sure if the answer you expect is this, but...<\/p>\n\n<p>First: I agree - the error message seems weird, it should talk about incompatibility between dense_24 and target array.<\/p>\n\n<p>Now, to solve your problem, you should either reshape your target array or create a different Dense at the end to match your array. <\/p>\n\n<p>About your target array, for a classification in two classes, it should be shaped as:<\/p>\n\n<ul>\n<li>(46000,2) if your classification uses two values, one for likelihood of class A and another for likelyhood of class B<\/li>\n<li>(46000,1) if your classification uses a single value being 0 class A and 1 class B (in which case your dense layer should be (None,1))<\/li>\n<\/ul>\n\n<h2>What I think is the easiest solution:<\/h2>\n\n<ul>\n<li>Instead of having <code>Dense(2,...)<\/code> at the end of your model, use <code>Dense(1, activation='sigmoid')<\/code>.   <\/li>\n<\/ul>\n\n<p>Why? Because your target data is shaped like (46000,1), meaning you have only one number for two classes. 0 is one class, 1 is another. <\/p>\n"},{"AnswerId":"43891124","CreationDate":"2017-05-10T11:26:40.353","ParentId":null,"OwnerUserId":"5110359","Title":null,"Body":"<p>Keras is saying the network you constructed outputs arrays of length two, but the your training data contains data with shape (0,1). However, training data shape should match network output shape.<\/p>\n\n<p>The shape (None, 2) simply means that the network accepts a batch with any number of elements that themselves are arrays of two elements (The first element is batch size in Keras shapes). So the correct input data shape would be (2,).<\/p>\n\n<p>From the error message we see you have 4600 pieces of training data with shape (0, 1). This the reason for conflict. I suspect something goes wrong with reading the training data and you are not actually fitting with arrays of two elements, like you have intended.<\/p>\n"}]}
{"QuestionId":41354980,"AnswerCount":1,"Tags":"<tensorflow><gradient><lstm>","CreationDate":"2016-12-28T03:27:34.520","AcceptedAnswerId":null,"OwnerUserId":5694334.0,"Title":"tensorflow lstm,How could I get the gradient of the loss at input data, not the variable weight and bias","Body":"<p>How could I get the gradient of the loss at input data, not the variable weight and bias<\/p>\n\n<pre><\/pre>\n\n<p>i could calculate the gradient at weight and bias,so how could i get the gradient at batch_x directly? <\/p>\n\n<pre><\/pre>\n\n<p>the input_grad value is [None].<\/p>\n","answers":[{"AnswerId":"41373086","CreationDate":"2016-12-29T04:20:50.457","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>The question is resolved in the comment. batch_x should be replaced with x in the line below:<\/p>\n\n<pre><code>input_grad = sess.run(tf.gradients(cost, x), feed_dict=fd) \n<\/code><\/pre>\n"}]}
{"QuestionId":41355097,"AnswerCount":0,"Tags":"<performance><tensorflow><gpu><matrix-multiplication>","CreationDate":"2016-12-28T03:51:47.243","AcceptedAnswerId":null,"OwnerUserId":821806.0,"Title":"Why is my 980 TI outperforming my 1080?","Body":"<p>Trying to make sure my new computer is setup properly, and noticed that its GeForce 1080 GPU is significantly underperforming the 980TI on my old system (when running Tensorflow jobs). Since the systems differ in more than just GPU, I wrote a small benchmark to isolate GPU matrix multiplication performance in TensorFlow. The results confirm that the new GPU is significantly slower. I know this is something to do with the software installed, but I've checked the obvious things: same python3, same cudnn, same numpy. What could be causing this strange performance gap?<\/p>\n\n<p>Benchmarking Script:<\/p>\n\n<pre><\/pre>\n\n<p>Old System (980 Ti):<\/p>\n\n<pre><\/pre>\n\n<p>New System (1080):<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41356093,"AnswerCount":1,"Tags":"<graph><tensorflow><scatter-plot><tensorboard>","CreationDate":"2016-12-28T05:49:34.677","AcceptedAnswerId":null,"OwnerUserId":6324088.0,"Title":"how to make a scatter plots using tensorboard - tensorflow","Body":"<p>now, i'm studying tensorflow.\nbut, i can't draw dot graph using tensorboard.<\/p>\n\n<p>if i have sample data for training, like that<\/p>\n\n<pre><\/pre>\n\n<p>i want to show scatter plots using tensorboard.\ni know \"import matplotlib.pyplot as plt\" can do that.\nbut i can just use console (putty). so can't use this method.<\/p>\n\n<p>can i see dot graph, like scatter plots using tensorboard.<\/p>\n\n<p>can anyone help me?<\/p>\n","answers":[{"AnswerId":"41546648","CreationDate":"2017-01-09T11:08:39.867","ParentId":null,"OwnerUserId":"482601","Title":null,"Body":"<p>Not really a full answer, but what I do is import matplotlib for no display use:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import matplotlib as mpl\nmpl.use('Agg')  # No display\nimport matplotlib.pyplot as plt\n<\/code><\/pre>\n\n<p>Then draw my plots into a buffer and save that as a PNG:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># setting up the necessary tensors:\nplot_buf_ph = tf.placeholder(tf.string)\nimage = tf.image.decode_png(plot_buf_ph, channels=4)\nimage = tf.expand_dims(image, 0)  # make it batched\nplot_image_summary = tf.summary.image('some_name', image, max_outputs=1)\n\n# later, to make the plot:\nplot_buf = get_plot_buf()\nplot_image_summary_ = session.run(\n        plot_image_summary,\n        feed_dict={plot_buf_ph: plot_buf.getvalue()})\nsummary_writer.add_summary(plot_image_summary_, global_step=iteration)\n<\/code><\/pre>\n\n<p>where <code>get_plot_buf<\/code> is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def get_plot_buf(self):\n    plt.figure()\n\n    # ... draw plot here ...\n\n    buf = io.BytesIO()\n    plt.savefig(buf, format='png')\n    plt.close()\n\n    buf.seek(0)\n    return buf\n<\/code><\/pre>\n"}]}
{"QuestionId":41356865,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-28T06:54:43.157","AcceptedAnswerId":41357052.0,"OwnerUserId":5605365.0,"Title":"TensorFlow InvalidArgumentError: Matrix size-compatible: In[0]: [100,784], In[1]: [500,10]","Body":"<p>I'm new to tensorflow and am following a tutorial. I am getting an error that says: <\/p>\n\n<pre><\/pre>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>I have marked the line where the error occurs what am I doing wrong and how can I fix it?<\/p>\n\n<p>Stack overflow wants me to write more, it says that there is not enough details and too much code. I honestly don't understand tensorflow well enough to add any more details. I'm hoping someone can just help me with this. I think the problem is that  and  have different dimensions, but I don't understand why or what I should do about it.<\/p>\n","answers":[{"AnswerId":"41357052","CreationDate":"2016-12-28T07:08:05.633","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>One error lies in this line<\/p>\n\n<pre><code> l2 = tf.add(tf.matmul(data, hidden_2_layer['weights']), hidden_2_layer['biases'])\n<\/code><\/pre>\n\n<p>Your second weights variable has dimensions <code>500 x 500<\/code>, but your <code>data<\/code> variable was fed with data <code>100x784<\/code> so multiplication is incompatible. Make this,<\/p>\n\n<pre><code> l2 = tf.add(tf.matmul(l1, hidden_2_layer['weights']), hidden_2_layer['biases'])\n<\/code><\/pre>\n\n<p>Also make the corresponding change for <code>l3<\/code> and <code>output<\/code>.<\/p>\n\n<hr>\n\n<p>Always specify a shape for the placeholder, like this,<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=(None, 784))\n<\/code><\/pre>\n\n<p>This will allow you to catch such errors while building the graph and TensorFlow will be able to pinpoint these errors.<\/p>\n"}]}
{"QuestionId":41356881,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2016-12-28T06:55:44.813","AcceptedAnswerId":null,"OwnerUserId":3604746.0,"Title":"Q: Tensorflow: tf.random_crop","Body":"<p>I used the  to crop 56*56 in my patch (80, 80). I get the error like follow:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know why the expected size must be in [0, some_num smaller than 56], Could anyone help to solve it?<\/p>\n","answers":[]}
{"QuestionId":41357038,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-12-28T07:07:26.417","AcceptedAnswerId":41386901.0,"OwnerUserId":2059708.0,"Title":"tensorflow tf.contrib.learn DNN Regressor internals","Body":"<p>I have tried to recreate the DNNRegressor model from tf.contrib.learn using tensorflow, but my loss is 6 orders of magnitude higher. Could someone point me in the right direction? I have no idea what is going wrong or what is different : \/ The Data is here if it helps <a href=\"http:\/\/pastebin.com\/BG6r6EF6\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/BG6r6EF6<\/a><\/p>\n\n<blockquote>\n  <p>tf.contrib.learn code:<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>tensorflow code:<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>I changed to <\/p>\n\n<pre><\/pre>\n\n<p>and the loss is about the same for both methods now (~ 200). The tensorflow model is very inaccurate though while the DNNRegressor outputs what I would expect when using validation data. The tensorboard graphs are very different too. <\/p>\n\n<p>DNNRegressor:\n<a href=\"https:\/\/i.stack.imgur.com\/hdIV9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/hdIV9.png\" alt=\"DNNR\"><\/a><\/p>\n\n<p>Tensorflow:\n<a href=\"https:\/\/i.stack.imgur.com\/YRuEz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YRuEz.png\" alt=\"myMLP\"><\/a><\/p>\n","answers":[{"AnswerId":"41386901","CreationDate":"2016-12-29T20:42:24.777","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>I would compare the graphs of two models by using tensorboard. Did you try that?<\/p>\n"}]}
{"QuestionId":41357292,"AnswerCount":1,"Tags":"<python><ubuntu><installation><tensorflow><word2vec>","CreationDate":"2016-12-28T07:24:36.023","AcceptedAnswerId":41357639.0,"OwnerUserId":6782093.0,"Title":"Tensorflow:Unable to visualize embeddings in word2vec_basic.py?","Body":"<p>I am trying to implement tensorflows <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"nofollow noreferrer\">word2vec_basic.py<\/a> model in my system.I went through this <a href=\"https:\/\/www.tensorflow.org\/tutorials\/word2vec\/\" rel=\"nofollow noreferrer\">tutorial<\/a> to understand and implement it.<\/p>\n\n<p>After I ran the word2vec code using python3 in my system it asked me to install sklearn, matplotlib, and scipy to visualize embeddings.<\/p>\n\n<p>So I installed sklearn, matplotlib, and scipy using pip in my system and ran the code a second time but it still asks me to install the above 3 again.<\/p>\n\n<p>No errors were there during installation of sklearn, matplotlib, and scipy. I am using Ubuntu 16.04 LTS.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/yVL4s.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/yVL4s.png\" alt=\"Here is the screen shot\"><\/a><\/p>\n\n<p>UPDATE: Here is the screenshot after removing  and  - \n<a href=\"https:\/\/ibin.co\/377XsooeSdej.png\" rel=\"nofollow noreferrer\">https:\/\/ibin.co\/377XsooeSdej.png<\/a><\/p>\n","answers":[{"AnswerId":"41357639","CreationDate":"2016-12-28T07:47:46.077","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>As evident from the error message, you are missing the <code>python3-tk<\/code> package. Try to install it using, <\/p>\n\n<pre><code>sudo apt-get install python3-tk\n<\/code><\/pre>\n"}]}
{"QuestionId":41357716,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-12-28T07:52:49.250","AcceptedAnswerId":null,"OwnerUserId":2455061.0,"Title":"tensorflow , what does server.join() exactly do?","Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.11\/tensorflow\/tools\/dist_test\/python\/mnist_replica.py#L122\" rel=\"nofollow noreferrer\">here<\/a> shows that  is the parameter server.<\/p>\n\n<p>But <a href=\"https:\/\/github.com\/CheHsuan\/TensorFlowAssignment\/blob\/master\/MapReduce\/create_worker.py#L12\" rel=\"nofollow noreferrer\">here<\/a> shows that  is the worker.<\/p>\n\n<p>It makes me so confused.<\/p>\n","answers":[]}
{"QuestionId":41360096,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-28T10:30:41.580","AcceptedAnswerId":41361871.0,"OwnerUserId":1735003.0,"Title":"Tensorflow: run time test metrics and data queues","Body":"<p>I want to compute and display accuracy on the test set while the network is training.<\/p>\n\n<p>In the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\" rel=\"nofollow noreferrer\">MNIST tutorial<\/a> that uses feeds, one can see that it can be done easily by feeding test data rather than train data. Simple solution to a simple problem.<\/p>\n\n<p>However I am not able to find such an easy example when using queues for batching. AFAICS, the documentation proposes two solutions:<\/p>\n\n<ol>\n<li>Offline testing with saved states. I don't want offline.<\/li>\n<li>Making a second 'test' network that share weights with the network being trained. That doesn't sound simple and I have not seen an example of that.<\/li>\n<\/ol>\n\n<p>Is there a third, easy way to compute test metrics at run time? Or is there an example somewhere of the second, test network with shared weights that proves me wrong by being super simple to implement?<\/p>\n","answers":[{"AnswerId":"41361871","CreationDate":"2016-12-28T12:10:04.530","ParentId":null,"OwnerUserId":"6778091","Title":null,"Body":"<p>If I understand your question correctly, you want to validate your model while training with queue inputs not feed_dict?\nsee my <a href=\"https:\/\/github.com\/zakizhou\/mnist_input_pipeline\/blob\/master\/mnist_train_tfrecords.py\" rel=\"nofollow noreferrer\">program<\/a> that does this.\nHere is a short explanation:<\/p>\n\n<p>First you need to convert you data into train and validation files like 'train.tfreords' and 'valid.tfreocrds'<\/p>\n\n<p>Second in your training program start two queues that parse this two files,\nand use <strong>sharing variables<\/strong> to get the two logits for train and valid<\/p>\n\n<p>In my program this is done by    <\/p>\n\n<pre><code>with tf.variable_scope(\"inference\") as scope:\n        logits = mnist.inference(images)\n        scope.reuse_variables()\n        validation_logits = mnist.inference(validation_images)\n<\/code><\/pre>\n\n<p>then use <code>logits<\/code> to do get train loss and minimize it and use <code>validation_logits<\/code> to get valid accuracy<\/p>\n"}]}
{"QuestionId":41360523,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><gpu>","CreationDate":"2016-12-28T10:55:28.177","AcceptedAnswerId":41399717.0,"OwnerUserId":3862262.0,"Title":"GPU runs during the dev step out of memory","Body":"<p>I'm using the <a href=\"https:\/\/github.com\/dennybritz\/cnn-text-classification-tf\" rel=\"nofollow noreferrer\">Denny Britz implementation<\/a> of <a href=\"https:\/\/arxiv.org\/abs\/1408.5882\" rel=\"nofollow noreferrer\">Yoon Kims CNN<\/a> for sentiment analysis in slightly modified form (I added the word2vec approach, so that the weight matrix is not calculated from scratch).<\/p>\n\n<p>With small datasets (like 10MB) it works fine, but if I try to train on datasets of size >50MB (still not very large) my GPU runs out of memory and throws the following error: <a href=\"http:\/\/pastebin.com\/AMfYkpXZ\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/AMfYkpXZ<\/a><\/p>\n\n<p>The GPU is a GeForce GTX 1080 with 8 gb.<\/p>\n\n<p>I worked out that the error comes from the dev step\/evaluation step:<\/p>\n\n<pre><\/pre>\n\n<p>To be precisely, it comes from the  line.\nIf i comment out the whole dev_step, the training runs without throwing errors. <\/p>\n\n<p>Do you have an idea why this error occurs and how I can fix it? Thanks in advance!<\/p>\n\n<p><strong>Edit:<\/strong>\nThe whole code is available at: <a href=\"https:\/\/gist.github.com\/pexmar\/7b3d074825eeaf5a336fce406d8e9bae\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/pexmar\/7b3d074825eeaf5a336fce406d8e9bae<\/a><\/p>\n","answers":[{"AnswerId":"41399717","CreationDate":"2016-12-30T16:50:45.320","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>Check the size of the batches you're passing in to <code>dev_step<\/code> compared to the size of batches you're running in <code>train_step<\/code>. You may have to evaluate the test set (which I guess is the same as <code>dev<\/code>?) in batches as well.<\/p>\n"}]}
{"QuestionId":41361269,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><caffe><lmdb><nvidia-digits>","CreationDate":"2016-12-28T11:36:06.247","AcceptedAnswerId":null,"OwnerUserId":5498819.0,"Title":"NVIDIA DIGITS - Multiple Input layers","Body":"<p>i have a question regarding NVIDIA DIGITS Framework. \nSo i have been using caffe without DIGITS and have used HDF5 Layers so far. There i could use multiple \"top\" (data_0, data_1, data_2) inputs (see code below). So i could feed the net with more then one input image. But in DIGITS only lmdb input layer works.<\/p>\n\n<p>So is it possible to create a lmdb input layer with multiple input images??<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42212203","CreationDate":"2017-02-13T19:39:57.443","ParentId":null,"OwnerUserId":"2404152","Title":null,"Body":"<p>Sorry, that isn't supported in DIGITS.<\/p>\n\n<p>Since DIGITS manages your datasets for you, it also sets up the data layers in your network for you. This way you don't need to copy+paste the LMDB paths into your network when you want to run a previous network on a new dataset or when you move the location of your jobs on disk. It's a decision which sacrifices flexibility for the sake of making the common case easy.<\/p>\n\n<p>For classification, one LMDB should have two tops: \"data\" and \"label\". For other dataset types, there should be one LMDB with a single \"data\" top, and another LMDB with a single \"label\" top. If you need a more complicated data layer setup, then you'll need to either use Caffe directly or make some changes to the DIGITS source code.<\/p>\n\n<p>DIGITS's HDF5 support is not great because Caffe's HDF5 support is not great.<\/p>\n"}]}
{"QuestionId":41361766,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-28T12:04:38.337","AcceptedAnswerId":41377517.0,"OwnerUserId":2467383.0,"Title":"tensorflow cifar10 tutorial fails","Body":"<p>I have downloaded the CIFAR10 code from the link in the tutorial <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">here<\/a> and am trying to run the tutorial. I run it with the command<\/p>\n\n<pre><\/pre>\n\n<p>It starts ok and downloads the data file as expected. When it tries to open the input file it fails with the following trace:<\/p>\n\n<pre><\/pre>\n\n<p>Sure enough, when I investigate the code there is a call in cifar10_input.py to strided_slice() with only 3 arguments:<\/p>\n\n<pre><\/pre>\n\n<p>Whereas the tensorflow documentation does indeed state that there must be at least 4 arguments.<\/p>\n\n<p>What is going wrong? I have downloaded the latest tensorflow (0.12) and I'm running the master branch of the cifar code.<\/p>\n","answers":[{"AnswerId":"41377517","CreationDate":"2016-12-29T10:13:35.677","ParentId":null,"OwnerUserId":"2467383","Title":null,"Body":"<p>Following some discussion at <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/817\" rel=\"nofollow noreferrer\">github<\/a> I have put the following changes in which seem to make it work:<\/p>\n\n<p>In cifar10_input.py<\/p>\n\n<pre><code>-  result.label = tf.cast(tf.strided_slice(record_bytes, [0], [label_bytes]), tf.int32)\n+  result.label = tf.cast(tf.slice(record_bytes, [0], [label_bytes]), tf.int32)\n\n\n\n-  depth_major = tf.reshape( tf.strided_slice(record_bytes, [label_bytes], [label_bytes + image_bytes]),      [result.depth, result.height, result.width])\n+  depth_major = tf.reshape(tf.slice(record_bytes, [label_bytes], [image_bytes]), [result.depth, result.height, result.width])\n<\/code><\/pre>\n\n<p>Then in both cifar10_input.py and cifar10.py I had to search for \"deprecated\" and wherever I found it, replace it with a valid function based on what I read in the api guide (hopefully correctly). Examples of this:<\/p>\n\n<pre><code>-  tf.contrib.deprecated.image_summary('images', images)\n+  tf.summary.image('images', images)\n<\/code><\/pre>\n\n<p>and<\/p>\n\n<pre><code> - tf.contrib.deprecated.histogram_summary(tensor_name + '\/activations', x)\n - tf.contrib.deprecated.scalar_summary(tensor_name + '\/sparsity',\n + tf.summary.histogram(tensor_name + '\/activations', x)\n + tf.summary.scalar(tensor_name + '\/sparsity',\n<\/code><\/pre>\n\n<p>It seems to be chugging along happily now. I'll see if it completes OK and if the changes I put in above give the desired diagnostic outputs.<\/p>\n\n<p>I'd still like to hear a definitive answer from someone closer to the code.<\/p>\n"}]}
{"QuestionId":41364390,"AnswerCount":1,"Tags":"<memory><tensorflow><gpu><cpu><deep-learning>","CreationDate":"2016-12-28T14:49:44.487","AcceptedAnswerId":null,"OwnerUserId":6260947.0,"Title":"Choice of infrastructure for faster deep learning model training with tensorflow?","Body":"<p>I am a newbie in deep learning with tensorflow.I am trying out a seq2seq model sample code.<\/p>\n\n<p>I wanted to understand:<\/p>\n\n<ol>\n<li><p>What is the minimum values of number of layers, layer size and batch\nsize that I could start off with to be able to test the seq2seq\nmodel with satisfactory accuracy?<\/p><\/li>\n<li><p>Also,the minimum infrastructure setup required in terms of memory\nand cpu capability to train this deep learning model within a max\ntime of a few hours.<\/p><\/li>\n<\/ol>\n\n<p>My experience has been training a seq2seq model to build a neural network with \n2 layers of size 900 and batch size 4 <\/p>\n\n<ul>\n<li>took around 3 days to train on a 4GB RAM,3GHz Intel i5 single core\nprocessor.<\/li>\n<li>took around 1 day to train on a 8GB RAM,3GHz Intel i5 single core\nprocessor.<\/li>\n<\/ul>\n\n<p>Which helps the most for faster training - more RAM capacity, multiple CPU cores or a CPU + GPU combination core?<\/p>\n","answers":[{"AnswerId":"41369053","CreationDate":"2016-12-28T20:11:46.253","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>Disclaimer: I'm also new, and could be wrong on a lot of this. <\/p>\n\n<blockquote>\n  <p>I am a newbie in deep learning with tensorflow.I am trying out a\n  seq2seq model sample code.<\/p>\n  \n  <p>I wanted to understand:<\/p>\n  \n  <p>What is the minimum values of number of layers, layer size and batch\n  size that I could start off with to be able to test the seq2seq model\n  with satisfactory accuracy? <\/p>\n<\/blockquote>\n\n<p>I think that this will just have to be up to your experimentation. Find out what works for your data set. I have heard a few pieces of advice: don't pick your own architecture if you can - find someone else's that is tried and tested. Seems deeper networks are better than wider if you're going to choose between the too. I also think bigger batch sizes are better if you have the memory. I've heard to maximize network size and then regularize so you don't overfit.<\/p>\n\n<p>I have the impression these are big questions that no one really knows the answer to (could be very wrong about this!). We'd all love a smart way of choosing layer size \/ number of layers, but no one knows exactly how changing these things affects training.<\/p>\n\n<blockquote>\n  <p>Also,the minimum infrastructure setup required in terms of memory and cpu capability to train this deep\n  learning model within a max time of a few hours. <\/p>\n<\/blockquote>\n\n<p>Depending on your model, that could be an unreasonable request. Seems like some models train for hundreds if not thousands of hours (on GPUs).<\/p>\n\n<blockquote>\n  <p>My experience has\n  been training a seq2seq model to build a neural network with 2 layers\n  of size 900 and batch size 4\n  took around 3 days to train on a 4GB RAM,3GHz Intel i5 single core\n  processor. took around 1 day to train on a 8GB RAM,3GHz Intel i5\n  single core processor. Which helps the most for faster training - more\n  RAM capacity, multiple CPU cores or a CPU + GPU combination core?<\/p>\n<\/blockquote>\n\n<p>I believe a GPU will help you the most. I have seen some stuff that uses the CPU (asynchronous actor critic or something? They didn't use locking) where it seemed like CPU was better, but I think GPU will give you huge speedups.<\/p>\n"}]}
{"QuestionId":41364745,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2016-12-28T15:10:25.987","AcceptedAnswerId":null,"OwnerUserId":7350296.0,"Title":"Weird Training Issue with keras - sudden huge drop in loss with zeros in FC layer","Body":"<p>I'm getting this odd issue with training a siamese-style CNN with Keras (backend of Tensor Flow, Ubuntu 14.04, Cuda 8, with cudnn). In short, the CNN has a shared set of weights that takes in two images, merges their respective FC layers, and then estimates a regression.  I'm using MSE loss with the Adam optimizer (with default parameters).  I've done this several times with different types of problems and have never seen the following.<\/p>\n\n<p>Essentially what happens is on the first epoch, everything seems to be training fine, and the loss is decreasing slowly, as expected (ends at around an MSE of ~3.3 using a batch size of 32). The regression is estimating a 9-dimensional continuous-valued vector. <\/p>\n\n<p>Then, as soon as the second epoch starts, the loss drops DRAMATICALLY (to ~ 4e-07).  You'd think \"oh yay the loss is really small--I win\", but when I inspect the trained weights by prediction on novel inputs (I'm using the checkpointer to dump out the best set of weights according to the loss), I get odd behavior.  No matter what the inputs are (different images, random noise as inputs, even zeros), I always get the same exact output.  Further inspection shows that the last FC layer in the shared weights are all zeros.  <\/p>\n\n<p>If I look at the weights after the first epoch, when everything seems \"normal\", this doesn't happen--I just don't get optimal results (makes sense--only one epoch has occurred).  This only happens with the second epoch and on.  <\/p>\n\n<p>Has anybody ever seen this?  Any ideas?  You think it's a dumb error on my part, or some weird bug?<\/p>\n\n<p>More details on my network topology here.  Here are the shared weights:<\/p>\n\n<pre><\/pre>\n\n<p>Then I merge them for regression as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Finally, each epoch trains on about 1,000,000 samples, so there should be plenty of data to train.  I've just never seen a FC layer get set to all zeros.  And even at that, I don't understand how that makes for a very low loss when the training data are not all zeros.<\/p>\n","answers":[{"AnswerId":"54521205","CreationDate":"2019-02-04T17:19:50.273","ParentId":null,"OwnerUserId":"9126032","Title":null,"Body":"<p>For the zeroes which are seemingly getting predicted by the last layer what might have happened is the dying ReLU problem. Try LeakyReLU, tweak alpha. This worked for me in eradicating those zeros which I would get in the first epoch itself.<\/p>\n"}]}
{"QuestionId":41365943,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><keras><permission-denied>","CreationDate":"2016-12-28T16:21:46.833","AcceptedAnswerId":41366488.0,"OwnerUserId":5634912.0,"Title":"Getting permission denied error while installing tensorflow and keras","Body":"<p>After following the steps from my previously asked question <a href=\"https:\/\/stackoverflow.com\/questions\/41340461\/can-i-implement-deep-learning-models-in-my-laptop-with-intel-hd-graphics\/41341085#41341085\">Can I implement deep learning models in my laptop with intel hd graphics<\/a>\nI am getting a permission denied error while installing tensorflow<\/p>\n\n<p><img src=\"https:\/\/i.imgur.com\/VJPNdDN.png\" alt=\"error13\"><\/p>\n\n<p>What can I do to install tensorflow and keras without getting this error?<\/p>\n","answers":[{"AnswerId":"41366488","CreationDate":"2016-12-28T16:57:59.870","ParentId":null,"OwnerUserId":"2776376","Title":null,"Body":"<p>pip3 is trying to access <code>\/usr\/bin\/<\/code> for which you need superuser rights. Try<\/p>\n\n<pre><code>sudo pip3 install --upgrade ...\n<\/code><\/pre>\n"},{"AnswerId":"41366779","CreationDate":"2016-12-28T17:17:05.553","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It appears that TensorFlow depends on a newer version of the <code>setuptools<\/code> package than you have installed. As <a href=\"https:\/\/stackoverflow.com\/a\/41366488\/3574081\">Maximilian points out<\/a>, <code>pip<\/code> needs root access to upgrade <code>setuptools<\/code> and install TensorFlow and Keras in your machine's Python distribution. You can either use <code>sudo<\/code> as he suggests, or it may be safer to <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#virtualenv_installation\" rel=\"nofollow noreferrer\">install TensorFlow into a Virtualenv<\/a>. For example, to create a Virtualenv in the directory <code>~\/tensorflow<\/code>, execute the following commands:<\/p>\n\n<pre><code>$ sudo pip install virtualenv\n$ virtualenv --system-site-packages ~\/tensorflow\n$ cd ~\/tensorflow\n$ source bin\/activate\n$ pip install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.12.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n"}]}
{"QuestionId":41366248,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-12-28T16:43:21.460","AcceptedAnswerId":41388489.0,"OwnerUserId":7287271.0,"Title":"Using tf.train.Supervisor's default summary_op on validation data?","Body":"<p>I'm fitting a classifier that has two placeholders, one for my input features and one for my output labels. To get data into the graph I'm placing a ndarray into a feed_dict. How can I use a  to evaluate the model on validation data without having to evaluate the  manually? Is it only possible by using readers instead of feed_dicts?<\/p>\n\n<p>This is my current workaround, where I'm checking elapsed time and manually evaluating the :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"41372292","CreationDate":"2016-12-29T02:18:15.050","ParentId":null,"OwnerUserId":"5549217","Title":null,"Body":"<p>You can use your workaround. That should work.<\/p>\n\n<p>Sherry<\/p>\n"},{"AnswerId":"41388489","CreationDate":"2016-12-29T23:11:39.763","ParentId":null,"OwnerUserId":"7287271","Title":null,"Body":"<p><strong>EDIT: tf.train.Supervisor will probably be deprecated, so nevermind. See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6604#issuecomment-270950324\" rel=\"nofollow noreferrer\">this<\/a> comment.<\/strong><\/p>\n\n<p>One thing that's <em>slightly<\/em> cleaner is to use <code>tf.train.Supervisor.loop<\/code> instead of checking elapsed time manually.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\nsummary_op = tf.summary.merge_all()\nsv = tf.train.Supervisor(logdir=\"tensorboard\/my-run\", summary_op=None)\nwith sv.managed_session() as sess:\n    sv.loop(60, lambda: sv.summary_computed(sess, sess.run(summary_op, feed_dict=next(validation_data))))\n    for data in training_data:\n        sess.run(train_op, feed_dict=data)\n<\/code><\/pre>\n"}]}
{"QuestionId":41366327,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><tflearn>","CreationDate":"2016-12-28T16:48:07.543","AcceptedAnswerId":41366472.0,"OwnerUserId":5140684.0,"Title":"Does the compiler automatically use GPU while compiling or shall I have to use it manually?","Body":"<p>I have been compiling this program for <a href=\"https:\/\/medium.com\/@ageitgey\/machine-learning-is-fun-part-3-deep-learning-and-convolutional-neural-networks-f40359318721#.9fdmwrbmn\" rel=\"nofollow noreferrer\">bird image preprocessing<\/a>  on my pc with 8GB RAM, intel core i7-4th gen. processor but everytime is gets stuck after running for a while showing this error:<\/p>\n\n<pre><\/pre>\n\n<p>Being a newbie to tensorflow, I am not sure whether CUDA (which I've installed successfully) is being used or is it just the CPU still performing the computations. <\/p>\n","answers":[{"AnswerId":"41366472","CreationDate":"2016-12-28T16:57:19.793","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The error you are seeing is raised before any GPU computation (or indeed any TensorFlow computation) has been issued. The <code>MemoryError<\/code> implies that  unpickling the dataset is either leading to running out of memory, or that there is some corruption in the data. Try loading a smaller dataset to see if the problem persists.<\/p>\n\n<p>As to your original question, if you have installed the GPU version of TensorFlow, and a compatible version of CUDA (i.e. CUDA 8.0) and cuDNN, TensorFlow will offload much of the computation to your GPU, including the matrix multiplications and convolutions that will comprise the bulk of the time spent during training.<\/p>\n"}]}
{"QuestionId":41367975,"AnswerCount":0,"Tags":"<python><c++><tensorflow><ipc><reinforcement-learning>","CreationDate":"2016-12-28T18:47:00.470","AcceptedAnswerId":null,"OwnerUserId":1576490.0,"Title":"Connecting Python + Tensorflow to an Emulator in C++","Body":"<p>I am trying to do deep reinforcement learning on the GameBoy platform, similar to what Google DeepMind has done on the Atari platform (<a href=\"https:\/\/deepmind.com\/research\/dqn\/\" rel=\"nofollow noreferrer\">https:\/\/deepmind.com\/research\/dqn\/<\/a>). For this purpose, I found a GameBoy Emulator in C++, and I want to do the reinforcement learning parts using the Python TensorFlow API. The system needs to do at least the following:<\/p>\n\n<ul>\n<li>Send the current video frame from the Emulator to the Neural Network (NN) constructed using the TensorFlow API<\/li>\n<li>Using the NN predict what action the agent should make (which key it should press)<\/li>\n<li>Send this key press back to the emulator<\/li>\n<li>Get back from the emulator the new game score<\/li>\n<li>Retrain the NN<\/li>\n<\/ul>\n\n<p>My questions is how should I do this communication between the emulator written in C++ and the bits doing reinforcement learning using Python plus TensorFlow?<\/p>\n\n<p>So far I have considered the following options:<\/p>\n\n<p>1) Use the C++ TensorFlow API. However, from my understanding this can only run already constructed graphs, so it would be rather cumbersome to construct the graph in Python then run it using the C++ API. Furthermore the documentation for the C++ API is not very good. <\/p>\n\n<p>2) Do some kind of OS piping between the C++ and Python parts. <\/p>\n\n<p>3) Embed Python into C++ like this: <a href=\"https:\/\/docs.python.org\/3.3\/extending\/embedding.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3.3\/extending\/embedding.html<\/a> <\/p>\n\n<p>Which of these would be the best way to do it? Are there other options I haven't considered?<\/p>\n","answers":[]}
{"QuestionId":41368977,"AnswerCount":0,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2016-12-28T20:05:40.043","AcceptedAnswerId":null,"OwnerUserId":6786718.0,"Title":"What's a common solution for an alternating Cohen's kappa score when it comes to deep neuronal networks?","Body":"<p>I am fairly new to deep convolutional networks. I am using a kappa matrix to calculate my score. Right now I am checking my kappa score every 500 iterations and it looks like an increasing saw-signal. <\/p>\n\n<p>I am using an SGD-Solver and a base-learning rate of . \nI have a roughly 15000 images (512x512) per class,evenly distributed, and there are 5 classes. I am using a Batchsize of 4<\/p>\n\n<p>For validation the data set got a different distribution:<\/p>\n\n<ol>\n<li>73%<\/li>\n<li>6%<\/li>\n<li>15%<\/li>\n<li>2%<\/li>\n<li>2%<\/li>\n<\/ol>\n\n<p>I have a net with 24 convolutional layers, which uses BatchNorm-Layers and alternating 3x3 and 1x1 layers between larger Convolutional  - Layers. So the structure is fairly deep. <\/p>\n\n<p>But when it comes to the Cohen's kappa score I am not able to move over the 0.1 score, more likely the Cohen's kappa score is rising and then falling and then rising again.<\/p>\n\n<p>Is there a common solution for this kind of behaviour ? Since I am fairly new I do not really know where to change my behaviour, like: Is my learning rate too high \/ too low ? Is my architecture too small \/ too large ? Am I using an inapropriate sovler? <\/p>\n\n<p>For those interested: The data is from the kaggle competition for diabetic eye retinopathy. I am not taking part in the competition, it's just an interesting project. <\/p>\n\n<p><a href=\"https:\/\/www.kaggle.com\/c\/diabetic-retinopathy-detection\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/diabetic-retinopathy-detection<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/RP0NJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RP0NJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This graph shows how my kappa score operates<\/p>\n","answers":[]}
{"QuestionId":41369062,"AnswerCount":3,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2016-12-28T20:12:46.630","AcceptedAnswerId":41466410.0,"OwnerUserId":4543908.0,"Title":"Tensorflow fail with \"Unable to get element from the feed as bytes.\" when attempting to restore checkpoint","Body":"<p>I am using Tensorflow r0.12.<\/p>\n\n<p>I use google-cloud-ml locally to run 2 different training jobs. In the first job, I find good initial values for my variables. I store them in a V2-checkpoint.<\/p>\n\n<p>When I try to restore my variables for using them in the second job :<\/p>\n\n<pre><\/pre>\n\n<p>I got the following error message :<\/p>\n\n<pre><\/pre>\n\n<p>The checkpoint is created with these lines in the first job :<\/p>\n\n<pre><\/pre>\n\n<p>According to <a href=\"https:\/\/stackoverflow.com\/a\/40340103\/4543908\">this answer<\/a>, it could come from the absence of metadata file but I am loading the metadata file.<\/p>\n\n<p>PS : I use the argument  because the device specifications generated by a launch on google-cloud-ml are quite intricated and I don't need to necessarily get the same dispatch.<\/p>\n","answers":[{"AnswerId":"41466410","CreationDate":"2017-01-04T14:36:39.913","ParentId":null,"OwnerUserId":"4543908","Title":null,"Body":"<p>The error message was due to the absence of the file named \"checkpoint\" by inadvertency.<\/p>\n\n<p>After the reintroduction of this file in the appropriate folder, it appears that the loading of the checkpoint is working.<\/p>\n\n<p>Sorry for having omitted this key point.<\/p>\n"},{"AnswerId":"46905432","CreationDate":"2017-10-24T08:17:49.850","ParentId":null,"OwnerUserId":"7159652","Title":null,"Body":"<p>The error message was also due to the mistakes in the file named \"checkpoint\" by inadvertency.<\/p>\n\n<p>For examples, the folder which contains the models has been moved, but the value of \"model_checkpoint_path:\" in \"checkpoint\" still is old path.<\/p>\n"},{"AnswerId":"41382741","CreationDate":"2016-12-29T15:34:57.333","ParentId":null,"OwnerUserId":"4392784","Title":null,"Body":"<p>I think the problem could be that when you save the model you set write_meta_graph=False. As a result I don't think you are actually saving the graph so when you try to restore there is no graph to restore. Try setting write_meta_graph=True <\/p>\n"}]}
{"QuestionId":41369275,"AnswerCount":2,"Tags":"<terminal><deep-learning><keras><ubuntu-16.04>","CreationDate":"2016-12-28T20:29:39.293","AcceptedAnswerId":null,"OwnerUserId":5634912.0,"Title":"No module named keras","Body":"<p>I have successfully installed keras after much struggle in the directory \/usr\/local\/lib\/python3.5\/dist-packages(by deafult it has been installed there). But whenever i am trying to import or trying to modify the keras backend it says No mudule named keras. What am I doing wrong here and how will I modify the keras.json file in these current conditions?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3UK10.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3UK10.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41394272","CreationDate":"2016-12-30T10:03:33.020","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<blockquote>\n  <p>What am I doing wrong here?<\/p>\n<\/blockquote>\n\n<p>The problem is that you start Python 2, but you installed it for Python 3. Your system has likely both Python versions, but you installed it only for one.<\/p>\n\n<p>If you want to start Python 3, try <code>python3<\/code>:<\/p>\n\n<pre><code>python3 -c \"from keras import backend; print(backend._BACKEND)\"\n<\/code><\/pre>\n\n<blockquote>\n  <p>How do I modify the keras.json file?<\/p>\n<\/blockquote>\n\n<p>Open <code>~\/.keras\/keras.json<\/code> with the editor of your choice. For example:<\/p>\n\n<pre><code>nano ~\/.keras\/keras.json\n<\/code><\/pre>\n"},{"AnswerId":"41371747","CreationDate":"2016-12-29T00:47:25.967","ParentId":null,"OwnerUserId":"4668640","Title":null,"Body":"<p>you should be able to edit keras.json directly. it should be at <code>~\/.keras\/keras.json<\/code> .<\/p>\n\n<p>also it seems you installed keras for python3 but are trying to use it with python(2).\ntry using pip instead of pip3 (or python3 instead of python). :)\nsee details <a href=\"https:\/\/stackoverflow.com\/questions\/11268501\/how-to-use-pip-with-python-3-x-alongside-python-2-x\">here<\/a>.<\/p>\n"}]}
{"QuestionId":41369700,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><neural-network><recurrent-neural-network>","CreationDate":"2016-12-28T21:06:45.670","AcceptedAnswerId":41384913.0,"OwnerUserId":1468002.0,"Title":"How to Initialize LSTMCell with tuple","Body":"<p>I recently upgraded my tesnorflow from Rev8 to Rev12. In Rev8 the default \"state_is_tuple\" flag in rnn_cell.LSTMCell is set to False, so I initialized my LSTM Cell with an list, see code below. <\/p>\n\n<pre><\/pre>\n\n<p>In Rev12, the default \"state_is_tuple\" flag is set to True, in order to make my old code work I had to explicitly turn the flag to False. However, now I got an warning from tensorflow saying:<\/p>\n\n<blockquote>\n  <p>\"Using a concatenated state is slower and will soon be deprecated. \n  Use state_is_tuple=True\"<\/p>\n<\/blockquote>\n\n<p>I tried to initialize LSTM cell with a tuple by changing the placeholder definition for self.init_state to the following:<\/p>\n\n<pre><\/pre>\n\n<p>but now I got an error message saying:<\/p>\n\n<blockquote>\n  <p>\"'Tensor' object is not iterable\"<\/p>\n<\/blockquote>\n\n<p>Does anyone know how to make this work?<\/p>\n","answers":[{"AnswerId":"41384913","CreationDate":"2016-12-29T17:59:09.600","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Feeding a \"zero state\" to an LSTM is much simpler now using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/rnn_cell\/base_interface_for_all_rnn_cells#RNNCell.zero_state\" rel=\"nofollow noreferrer\"><code>cell.zero_state<\/code><\/a>. You do not need to explicitely define the initial state as a placeholder. Define it as a tensor instead and feed it if required. This is how it works,<\/p>\n\n<pre><code>lstm_cell = rnn_cell.LSTMCell(self.config.hidden_dim)\nself.initial_state = lstm_cell.zero_state(self.batch_size, dtype=tf.float32)\noutputs, states = tf.nn.rnn(lstm_cell, data, initial_state=self.init_state)\n<\/code><\/pre>\n\n<p>If you wish to feed some other value as the initial state, Let's say <code>next_state = states[-1]<\/code> for instance, calculate it in your session and pass it in the <code>feed_dict<\/code> like -<\/p>\n\n<pre><code>feed_dict[self.initial_state] = next_state\n<\/code><\/pre>\n\n<p>In the context of your question, <code>lstm_cell.zero_state()<\/code> should suffice.<\/p>\n\n<hr>\n\n<p>Unrelated, but remember that you can pass both Tensors and Placeholders in the feed dictionary! That's how <code>self.initial_state<\/code> is working in the example above. Have a look at the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/ptb\" rel=\"nofollow noreferrer\">PTB Tutorial<\/a> for a working example.<\/p>\n"}]}
{"QuestionId":41370214,"AnswerCount":4,"Tags":"<python><matrix><random><theano>","CreationDate":"2016-12-28T21:51:21.433","AcceptedAnswerId":null,"OwnerUserId":4744673.0,"Title":"How to generate a matrix of two number in python","Body":"<p>Suppose  is a matrix. The following commands generate a matrix with the shape of  in randomly selected the members between  and  ( of the members are  and the rest ).     <\/p>\n\n<pre><\/pre>\n\n<p>The question is how to make the similar random matrix between  and another number likes . In another word, I need a matrix with the shape of  in randomly selected the members between  and  ( of the members are  and the rest )<\/p>\n","answers":[{"AnswerId":"41370523","CreationDate":"2016-12-28T22:20:30.773","ParentId":null,"OwnerUserId":"837534","Title":null,"Body":"<p>You can do this with NumPy:<\/p>\n\n<pre><code>import numpy as np\n\nx = np.zeros((3, 4))\na = np.random.binomial(1, p=0.8, size=x.shape).astype(float)\na[a==0] = 2.5\n<\/code><\/pre>\n\n<p>Now <code>a<\/code> looks like this:<\/p>\n\n<pre><code>array([[ 1. ,  1. ,  2.5,  1. ],\n       [ 1. ,  1. ,  1. ,  2.5],\n       [ 1. ,  1. ,  2.5,  1. ]])\n<\/code><\/pre>\n\n<p>The line:<\/p>\n\n<pre><code>a[a==0] = 2.5\n<\/code><\/pre>\n\n<p>replaces all occurrences of of <code>0<\/code> with <code>2.5<\/code>.<\/p>\n"},{"AnswerId":"41370492","CreationDate":"2016-12-28T22:18:51.190","ParentId":null,"OwnerUserId":"5644961","Title":null,"Body":"<p>the same way you make a line that cross 2 point in the plane<\/p>\n\n<pre><code>y = m*x + a\n<\/code><\/pre>\n\n<p>choose <code>m<\/code> and <code>a<\/code> such that when x=0 is the lower extreme of the interval and when x=1 is the other extreme.<\/p>\n\n<p>So for the interval <code>[a,b]<\/code> it is <code>y = (b-a)x + a<\/code> <\/p>\n\n<p>In this case for [1,2.5] then a=1 and m=1.5, but in your case <code>a<\/code> is a matrix of pure 1.<\/p>\n\n<p>Maybe your library have some other utility to do it, but when everything else fail this always will work<\/p>\n"},{"AnswerId":"41370478","CreationDate":"2016-12-28T22:17:50.333","ParentId":null,"OwnerUserId":"5556887","Title":null,"Body":"<p>If you are not looking to directly generate a matrix with 2.5 values, you can manually replace 0 values :<\/p>\n\n<pre><code>srng = RandomStreams()    \nA = srng.binomial(X.shape, p=0.8)\nA[A == 0]=2.5\n<\/code><\/pre>\n"},{"AnswerId":"41370351","CreationDate":"2016-12-28T22:05:44.413","ParentId":null,"OwnerUserId":"5015569","Title":null,"Body":"<p>How about something like this?<\/p>\n\n<pre><code>import random\n\ndef create_matrix(width, height):\n    mat = []\n    for i in range(height):\n        mat.append([]) # create sublist for each height level\n        for _ in range(width):\n            # generate a random number in inclusive range [1, 5]\n            random_number = random.randint(1, 5)\n            # 80% chance the number is one of 1, 2, 3, or 4 (arbitrary; could be 1, 3, 4, or 5, just not the entire range)\n            if random_number in {1, 2, 3, 4}:\n                mat[i].append(1)\n            # 20% chance it's 5\n            else:\n                mat[i].append(2.5)\n    return mat\n<\/code><\/pre>\n\n<p>One result:<\/p>\n\n<pre><code>&gt;&gt;&gt; print(create_matrix(10, 10))\n[[1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [2.5, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 2.5, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 2.5, 2.5], [2.5, 1, 1, 1, 1, 1, 1, 1, 1, 1], [2.5, 1, 1, 1, 1, 2.5, 1, 1, 2.5, 2.5], [1, 2.5, 1, 1, 1, 2.5, 1, 1, 2.5, 1], [1, 1, 1, 2.5, 2.5, 1, 1, 1, 1, 1], [1, 1, 1, 1, 2.5, 1, 1, 1, 2.5, 1], [1, 2.5, 1, 1, 1, 1, 1, 1, 2.5, 1]]\n<\/code><\/pre>\n\n<p>We can verify the number of <code>1<\/code>s and <code>2.5<\/code>s matches (or gets very close to) the proportion you designated:<\/p>\n\n<pre><code>&gt;&gt;&gt; {number: [val for row in mat for val in row].count(number) for number in {1, 2.5}}\n{1: 82, 2.5: 18}\n<\/code><\/pre>\n\n<p>With many iterations, these proportions will approach 80% and 20%, respectively. But random chance creates some variation around those means.<\/p>\n\n<p>Note that this doesn't ensure at all that each row has at least one <code>2.5<\/code> (don't know if that's relevant to your problem). But this should give you a good start.<\/p>\n"}]}
{"QuestionId":41370591,"AnswerCount":1,"Tags":"<python><gcc><cuda><gpu><theano>","CreationDate":"2016-12-28T22:26:22.713","AcceptedAnswerId":41372768.0,"OwnerUserId":6372052.0,"Title":"Theano errors for GPU use","Body":"<p>I've been attempting to use theano for pthon on my GPU (a 980m) with GCC 6.2.0 and G++ 6.2.0 for Python 3.5.1 on Ubuntu 16.04. I've installed all prerequisites as per the instructions and Theano given no errors on import. However, when I run theano.test() I get the following rather verbose ad cryptic error.<\/p>\n\n<pre><\/pre>\n\n<p>Given the syntax-like errors, I'm guessing it's a compiler issue of some sort, or perhaps Theano not knowing where to find my GCC or G++, but rather than mucking about aimlessly, I was wondering if anyone could shed some light on this.<\/p>\n","answers":[{"AnswerId":"41372768","CreationDate":"2016-12-29T03:31:23.867","ParentId":null,"OwnerUserId":"681865","Title":null,"Body":"<blockquote>\n  <p>I've installed all prerequisites as per the instructions.<\/p>\n<\/blockquote>\n\n<p>Except that you haven't. If you consult the installation documentation for CUDA 8, you will see that GCC 6 is not currently supported. You will need to install a supported compiler version if you wish to compile code with the CUDA toolchain (which is apparently a prerequisite for running Theano).<\/p>\n\n<p>You might also want to consult this <a href=\"https:\/\/stackoverflow.com\/q\/6622454\/681865\">question<\/a>, which outlines a number of different strategies for managing and using an alternate compiler version for the CUDA toolkit.<\/p>\n"}]}
{"QuestionId":41370987,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-12-28T23:09:55.863","AcceptedAnswerId":41371381.0,"OwnerUserId":3610140.0,"Title":"Is it possible to merge multiple TensorFlow graphs into one?","Body":"<p>I have two models trained with Tensorflow Python, exported to binary files named  and . Both files will generate only one output when feeding with input, say output1 and output2.<\/p>\n\n<p>My question is if it is possible to merge two graphs into one big graph so that it will generate output1 and output2 together in one execution.<\/p>\n\n<p>Any comment will be helpful. Thanks in advance!<\/p>\n","answers":[{"AnswerId":"41371381","CreationDate":"2016-12-28T23:58:03.460","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>I kicked this around with my local TF expert, and the brief answer is \"no\"; TF doesn't have a built-in facility for this.  However, you <em>could<\/em> write custom endpoint layers (input and output) with synch operations from Python's process management, so that they'd maintain parallel processing of each input, and concatenate the outputs.<\/p>\n\n<p><strong>Rationale<\/strong><\/p>\n\n<p>I like the way this could be used to get greater accuracy with multiple features, where the features have little or no correlation.  For instance, you could train two character recognition models: one to identify the digit, the other to discriminate between left- and right-handed writers.<\/p>\n\n<p>This would also allow you to examine the internal kernels that evolved for each individual feature, without interdependence with other features: the double-loop of an '8' vs the general slant of right-handed writing.<\/p>\n\n<p>I also expect that the models for individual features will converge measurably faster than one over-arching training session.<\/p>\n\n<p>Finally, it's quite possible that the individual models could be used in mix-and-match feature sets.  For instance, train another model to differentiate letters, while letting your previously-trained left\/right flagger would still have a pretty good guess at the writer's moiety.<\/p>\n"}]}
{"QuestionId":41371045,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2016-12-28T23:17:25.847","AcceptedAnswerId":41372512.0,"OwnerUserId":4015544.0,"Title":"tensorflow restore variables when batch_size in testing is different from the batch_size in training","Body":"<p>I am new to TensorFlow. I have trained the inception_v3 model successfully with my training data; now I want to predict the output of several images, but the number of them is different from the batch_size in training. I did it as follows:<\/p>\n\n<pre><\/pre>\n\n<p>It gave me the errors as follows:<\/p>\n\n<pre><\/pre>\n\n<p>It seems that the model parameters in \"outputs\" are not fed in successfully, but I do not know how to do it. Any ideas? Thanks.<\/p>\n","answers":[{"AnswerId":"41372512","CreationDate":"2016-12-29T02:51:56.137","ParentId":null,"OwnerUserId":"6778091","Title":null,"Body":"<p>Here you explicitly set the first dim of input placeholder x <code>batch_size<\/code>, so each time you need to feed a numpy.array type tensor with the same dim or your program will go wrong.<\/p>\n\n<p>A solution can be setting the first dim of any placeholder(input and label) <code>None<\/code> so that this dim can be any int or different during training and validation<\/p>\n\n<p>UPDATE:<\/p>\n\n<p>if you have already trained your model with a fixed first dim placeholder input(and label), you can change it when restore this graph with <code>tf.train.import_meta_graph(grapg_def=your_graph_def, input_map={'your_train_input_placedholer_name':new_placeholder})<\/code>\nhere <code>new_placeholder<\/code> is a placeholder you newly create with unfixed first dim.<\/p>\n"}]}
{"QuestionId":41371154,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2016-12-28T23:30:48.173","AcceptedAnswerId":null,"OwnerUserId":4392693.0,"Title":"How to debug uncannily accurate TensorFlow model?","Body":"<p>I'm trying to create a model in line with the tutorial at <a href=\"https:\/\/www.tensorflow.org\/tutorials\/wide\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/wide\/<\/a>, but I seem to be corrupting my data in some way, as I'm highly suspicious of the 97.9% accuracy on a very naive model.<\/p>\n\n<p>I'm separating training and evaluation data with the following code, which seems reasonable to me:<\/p>\n\n<pre><\/pre>\n\n<p>However, this is giving me numbers which seem way too high:<\/p>\n\n<pre><\/pre>\n\n<p>The model creation\/evaluation happens here:<\/p>\n\n<pre><\/pre>\n\n<p>I suspect there's a simple error here, but I'm not seeing it.<\/p>\n","answers":[{"AnswerId":"41371957","CreationDate":"2016-12-29T01:17:17.527","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>This depends on your model and your data. You can get 100% accuracy on the training set by overfitting, and <a href=\"http:\/\/cs.nyu.edu\/~wanli\/dropc\/\" rel=\"nofollow noreferrer\">people<\/a> have gotten &lt;1% error rate on the test set of MNIST.<\/p>\n"},{"AnswerId":"44775359","CreationDate":"2017-06-27T08:18:01.633","ParentId":null,"OwnerUserId":"8197469","Title":null,"Body":"<p>You are working with a unbalanced dataset in which the number of examples in one class is significantly higher than the other. For this unbalanced problem, a random predictor can easily achieve a high accuracy by predicting the label of the majority class for all testing data. <\/p>\n\n<p>To correctly validate the performance of this model, you can look at the auc. Here the auc you achieved is only 0.44, which is very poor. A better metric to validate the performance of problem with unbalanced dataset is auc_precision_recall. I suspect the auc_precision_recall for this problem will be fairly small (e.g. 0.02) if you print it out. <\/p>\n\n<p>You can read through this post to have a better understanding of auc:\n<a href=\"https:\/\/classeval.wordpress.com\/simulation-analysis\/roc-and-precision-recall-with-imbalanced-datasets\/\" rel=\"nofollow noreferrer\">https:\/\/classeval.wordpress.com\/simulation-analysis\/roc-and-precision-recall-with-imbalanced-datasets\/<\/a><\/p>\n"}]}
{"QuestionId":41371207,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2016-12-28T23:35:30.910","AcceptedAnswerId":null,"OwnerUserId":3993270.0,"Title":"TensorFlow: information\/variable remembered between each session.run()?","Body":"<p>I am learning TensorFlow from the example at: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/2_BasicModels\/linear_regression.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/2_BasicModels\/linear_regression.ipynb<\/a><\/p>\n\n<p>In the following code, during the training stage, each sess.run() was fed in one data point. <\/p>\n\n<pre><\/pre>\n\n<p>Based on the definition of the optimizer, it is trying to minimize cost. For a cost function J(r), an optimizer should update the parameter r of the cost function using:<\/p>\n\n<p>  where  is the learning rate<\/p>\n\n<p>For each data point fed in, it will update the parameter r once, which means the optimizer remembers the results of the previous fed.<\/p>\n\n<p>Does this mean for the TensorFlow session.run(), it does store the optimizer results from the previous session.run()?<\/p>\n\n<p>So how would a session() defined? And is everything computed in the same session remembered through each run() until a the session() is over ? Thanks!<\/p>\n","answers":[{"AnswerId":"41371918","CreationDate":"2016-12-29T01:12:29.063","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>When you run the optimizer, it adjusts the parameters to minimize the loss function. When you run it multiple times, it keeps adjusting the parameters over time - none of the parameters get reset. <\/p>\n\n<p>You can validate this by using <code>sess.run<\/code> on one of your weight variables after each time you run the optimizer.<\/p>\n"}]}
{"QuestionId":41372375,"AnswerCount":2,"Tags":"<python><machine-learning><scikit-learn><tensorflow>","CreationDate":"2016-12-29T02:31:48.107","AcceptedAnswerId":41380178.0,"OwnerUserId":5441912.0,"Title":"AttributeError: module 'tensorflow.contrib.learn' has no attribute 'TensorFlowDNNClassifier'","Body":"<p><strong>This is the ml tensorflow code I am trying to execute<\/strong> -<\/p>\n\n<pre><\/pre>\n\n<p><strong>It gives the following error<\/strong> -<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):<\/p>\n  \n  <p>File \"C:\\Users\\admin\\test3.py\", line 5, in <\/p>\n  \n  <p>classifier = skflow.TensorFlowDNNClassifier(hidden_units=[10, 20, 10], n_classes=3) AttributeError: module 'tensorflow.contrib.learn'\n  has no attribute 'TensorFlowDNNClassifier'<\/p>\n  \n  <p>[Finished in 69.3s with exit code 1]<\/p>\n  \n  <p>[shell_cmd: python -u \"C:\\Users\\admin\\test3.py\"]<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"41380178","CreationDate":"2016-12-29T12:54:39.373","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>There seems to have been a major refactor in the TensorFlow project, and all the <code>skflow<\/code> code has been moved under the main <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"noreferrer\">tensorflow<\/a> repository.<\/p>\n\n<p>Try to replace <code>TensorFlowDNNClassifier<\/code> with just <code>DNNClassifier<\/code>. The new class can be found out <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/584d4921014db921a8f4722749adff09737826a8\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn.py#L183\" rel=\"noreferrer\">here<\/a>. Your corrected code will look like,<\/p>\n\n<pre><code>import tensorflow.contrib.learn as skflow\nfrom sklearn import datasets, metrics\niris = datasets.load_iris()\n# made a change in the next line\nclassifier = skflow.DNNClassifier(hidden_units=[10, 20, 10], n_classes=3)\nclassifier.fit(iris.data, iris.target)\nscore = metrics.accuracy_score(iris.target, classifier.predict(iris.data))\n\nprint(\"Accuracy: %f\" % score)\n<\/code><\/pre>\n"},{"AnswerId":"56251828","CreationDate":"2019-05-22T07:50:46.583","ParentId":null,"OwnerUserId":"11538013","Title":null,"Body":"<pre class=\"lang-py prettyprint-override\"><code>import tensorflow.contrib.learn.python\n\nfrom tensorflow.contrib.learn.python import learn as learn\n<\/code><\/pre>\n"}]}
{"QuestionId":41372796,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-29T03:35:45.937","AcceptedAnswerId":null,"OwnerUserId":6497581.0,"Title":"What's wrong with my TensorFlow run cifar10_multi_gpu_train.py","Body":"<p>My system is :<\/p>\n\n<pre><\/pre>\n\n<p>and my GPU :<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/x8H1h.png\" alt=\"enter image description here\"><\/p>\n\n<p>It is ok to run sample models like  and  single GPU version.\nso I think my installation of tf is no problem. <\/p>\n\n<p>BUT, when I run :<\/p>\n\n<pre><\/pre>\n\n<p>I got this error :<\/p>\n\n<blockquote>\n  <p>ValueError: Variable conv1\/weights\/ExponentialMovingAverage\/ does not\n  exist, or was not created with tf.get_variable(). Did you mean to set\n  reuse=None in VarScope?<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"41373695","CreationDate":"2016-12-29T05:37:14.900","ParentId":null,"OwnerUserId":"6778091","Title":null,"Body":"<p>Is there any code to see?\nI personally think this problem is caused because you didn't reuse <strong>sharing variables<\/strong> for different inputs in different gpus, check out whether this line <code>tf.get_variable_scope().reuse_variables()<\/code> is in your loop for each gpu<\/p>\n"}]}
{"QuestionId":41372874,"AnswerCount":0,"Tags":"<python><performance><classification><keras>","CreationDate":"2016-12-29T03:48:17.303","AcceptedAnswerId":null,"OwnerUserId":1245951.0,"Title":"Binary classification with Keras: poor performance","Body":"<p>I'm experimenting with Keras to develop a multi-layer perceptron for binary classification purposes and I'm surprised at the (poor) performance I obtain (57% accuracy on training set). A logistic regression correctly classifies 100% of the samples.<\/p>\n\n<p>I created a data set with 2 inputs: input A is a sine function. Input B = input A plus a lag. When input A >= input B, output = 1. Otherwise, output = 0. <\/p>\n\n<pre><\/pre>\n\n<p>The issue I have is similar to what is described in \"Keras low accuracy classification task\". The answer points in the direction of the data set which I do not believe is the problem here. <\/p>\n\n<p>See code below. What am I missing to improve the accuracy of the model? Adding neurons to the layers or changing the activation function of the output layer to softmax hasn't yielded any better results.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41373018,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><word2vec>","CreationDate":"2016-12-29T04:10:59.113","AcceptedAnswerId":null,"OwnerUserId":2530674.0,"Title":"Word2Vec input and output vectors","Body":"<p>Going through word2vec tutorial on udacity, and from the paper it seems to be the case that there are seperate matrices for input word vectors and output. <\/p>\n\n<p>eg. .\nHere the input vectors $w_i$,  will predict the output vector $w_o$ for . It does this via a sampled softmax as shown below where  is the size of the context words (4 in this case).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/n5bhR.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/n5bhR.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Hence once training is done there could potentially be two vectors for  as a input vector, and another for the output vector. <strong>The question is why not have one matrix<\/strong>. This will ensure that the input and output vectors for the same words will be aligned.<\/p>\n\n<p>If it helps, the tensorflow code is attached below <strong>(why not set softmax_weights =  embedding and softmax_biases=0)<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<h2>update:<\/h2>\n\n<p>I implemented it without a seperate output matrix and the results still looks good: <a href=\"https:\/\/github.com\/sachinruk\/word2vec_alternate\" rel=\"nofollow noreferrer\">https:\/\/github.com\/sachinruk\/word2vec_alternate<\/a> . <strong>I suppose the questions now should be is there a mathematical reason as to why the output matrix should be different.<\/strong><\/p>\n","answers":[]}
{"QuestionId":41373079,"AnswerCount":1,"Tags":"<python><gpu><theano><keras><theano-cuda>","CreationDate":"2016-12-29T04:20:07.403","AcceptedAnswerId":null,"OwnerUserId":7351953.0,"Title":"Keras, Theano - Cropping2D introduces \"Input dimension mis-match\" on GPU","Body":"<p>I tried to create an autoencoder using keras.\n(ref: <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-autoencoders-in-keras.html<\/a>)<\/p>\n\n<p>the following test code works just fine on my mac:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I test this piece of code on AWS EC2 g2.2xlarge (my configuration info is on the end of this post), I run into the following Error:<\/p>\n\n<pre><\/pre>\n\n<p>I can run the code on GPU if the Cropping2D function was removed:<\/p>\n\n<pre><\/pre>\n\n<p>In a similar scenario, Cropping3D also raise the similar error... :-(<\/p>\n\n<p>My question is:<\/p>\n\n<p><strong>How can I solve this and run Cropping2D\/Cropping3D on GPU?<\/strong><\/p>\n\n<p>ps. My AWS EC2 g2.2xlarge configuration is as followed:<\/p>\n\n<ul>\n<li>Ubuntu 14.04 <\/li>\n<li>Cuda 7.5 <\/li>\n<li>Cudnn v5 <\/li>\n<li>Keras (1.2.0) <\/li>\n<li>Theano (0.9.0dev4)<\/li>\n<\/ul>\n\n<p><strong>Thank you!<\/strong><\/p>\n","answers":[{"AnswerId":"41390095","CreationDate":"2016-12-30T03:09:43.410","ParentId":null,"OwnerUserId":"7351953","Title":null,"Body":"<p>There appears to be a bug in Keras.\nIt has been taking care of. Case closed!<\/p>\n"}]}
{"QuestionId":41374175,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><tensorboard>","CreationDate":"2016-12-29T06:18:43.940","AcceptedAnswerId":null,"OwnerUserId":6379348.0,"Title":"how to add summary to tensorboard","Body":"<p>I'm doing DNN with tensorflow and want to visualize it with tensorboard. It is a three hidden layer network.  Here is the code -<\/p>\n\n<pre><\/pre>\n\n<p>The code gave me a very confusing error message. \nHowever, when I comment out these two lines (5th and 6th line to the last), the program works. Just tensorboard only show graph, no events and summaries. <\/p>\n\n<pre><\/pre>\n\n<p>I've spent a long time trying to figure it out. What did I do wrong?\nHelp is very much appreciated! thanks a lot. <\/p>\n\n<p>Here is the error message I got.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41374447,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><theano><keras><ubuntu-16.04>","CreationDate":"2016-12-29T06:42:43.003","AcceptedAnswerId":null,"OwnerUserId":5634912.0,"Title":"Getting ImportError while importing keras","Body":"<p>I have followed every step for installing keras but whenever I am trying to import keras it gives me an error No module named tensorflow. But I have already installed tensorflow.\n<a href=\"https:\/\/i.stack.imgur.com\/PVSAJ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PVSAJ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have two questions-\n1. How can I solve this particular ImportError?\n2. If possible can anyone give me a detailed step by step instruction on installing theano, tensorflow and keras? Please don't give any links to install keras because i have tried them all but they keep giving some kind of error or problem.<\/p>\n\n<p>If possible please take a look at my other question since they are related <a href=\"https:\/\/stackoverflow.com\/questions\/41369275\/no-module-named-keras\/41371747#41371747\">No module named keras<\/a><\/p>\n\n<p>Thank You<\/p>\n","answers":[{"AnswerId":"51424466","CreationDate":"2018-07-19T13:53:03.180","ParentId":null,"OwnerUserId":"6409557","Title":null,"Body":"<p>Try sudo -H python -m pip install --upgrade --force-reinstall tensorflow<\/p>\n"}]}
{"QuestionId":41375651,"AnswerCount":1,"Tags":"<python><tensorflow><keras><conv-neural-network>","CreationDate":"2016-12-29T08:16:10.140","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"load_weights Keras Model error","Body":"<p>I am fairly new to CNN. This is my first time using keras, tensorflow etc. I have a problem with load_weights function. I've trained a CNN (cifar100), now I want to test it by loading its weights and evaluating it.<\/p>\n\n<p>This is the stack traceback of the error I get:<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to extend the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py\" rel=\"nofollow noreferrer\">keras cifar10<\/a> code to cifar100 code. I've managed to train it, but I want also to evaluate it. By its evaluation I can determine if my model is good and what is its score.<\/p>\n\n<p>This is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44038864","CreationDate":"2017-05-18T04:55:27.337","ParentId":null,"OwnerUserId":"3068249","Title":null,"Body":"<p>You are saving the model and then loading it back as weights, after that again training your model. <\/p>\n\n<p>First, fix your script to save only the weights, load them back and check if the problem persists.<\/p>\n"}]}
{"QuestionId":41376134,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2016-12-29T08:51:48.137","AcceptedAnswerId":41379745.0,"OwnerUserId":783893.0,"Title":"Issue in Tensor slicing operation","Body":"<p>I'm trying to get  from .<\/p>\n\n<pre><\/pre>\n\n<p>How to get the  from  using TensorFlow.<\/p>\n","answers":[{"AnswerId":"41376917","CreationDate":"2016-12-29T09:39:34.150","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>Use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/shapes_and_shaping#squeeze\" rel=\"nofollow noreferrer\">tf.squeeze<\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\ninput = tf.constant([[2], [3], [3]])\n\nwith tf.Session() as sess:\n    print(sess.run(tf.squeeze(input)))\n<\/code><\/pre>\n"},{"AnswerId":"41379745","CreationDate":"2016-12-29T12:27:37.343","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>In this case you want to remove single dimensional entries from a matrix. In both TensorFlow and Numpy, this operation is termed as <code>squeeze<\/code>.<\/p>\n\n<p>Here is the official documentation for TensorFlow - <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/shapes_and_shaping#squeeze\" rel=\"nofollow noreferrer\"><code>tf.squeeze<\/code><\/a>. Quoting from the documentation, <\/p>\n\n<blockquote>\n  <p>Given a tensor input, this operation returns a tensor of the same type with all dimensions of size 1 removed. If you don't want to remove all size 1 dimensions, you can remove specific size 1 dimensions by specifying axis<\/p>\n<\/blockquote>\n\n<p>Hence to solve your issue, you can either pass <code>None<\/code> to <code>axis<\/code>, your default case, or pass <code>1<\/code>. Here is what the code will look like,<\/p>\n\n<pre><code>expected_out = tf.squeeze(input)\n<\/code><\/pre>\n\n<p>or,<\/p>\n\n<pre><code>expected_out = tf.squeeze(input, 1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41376694,"AnswerCount":2,"Tags":"<python><windows><tensorflow><spyder>","CreationDate":"2016-12-29T09:25:57.750","AcceptedAnswerId":null,"OwnerUserId":7353274.0,"Title":"Tensorflow Error in Windows","Body":"<p>I am getting the following error <\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>The problem is in the second line  The first line doesn't produce any error in the iPython console, the second line produces the above error. I believe that this is the same error that was reported by another user here: <a href=\"https:\/\/stackoverflow.com\/questions\/40866525\/permissionerror-errno-13-permission-denied-when-doing-input-data-read-data-se\">PermissionError: [Errno 13] Permission denied when doing input_data.read_data_sets(..)<\/a><\/p>\n\n<p>It doesn't look like that problem was resolved. Any help will be much appreciated. I am running the code on Spyder that comes with Anaconda. The answer at <a href=\"https:\/\/stackoverflow.com\/questions\/41076047\/windows-tensorflow-with-python-unable-to-read-mnist-data-due-to-permissions\">Windows Tensorflow with Python unable to read mnist data due to permissions<\/a> did not solve my problem. I reinstalled Anaconda and marked the box \"for all users\" but I am still getting the same error. <\/p>\n","answers":[{"AnswerId":"41384338","CreationDate":"2016-12-29T17:18:55.590","ParentId":null,"OwnerUserId":"7353274","Title":null,"Body":"<p>I was able to fix the problem, this is what worked:<\/p>\n\n<p>(i) Uninstall Anaconda and reinstall it for all users.<\/p>\n\n<p>(ii) Open command prompt from windows as an administrator and upgrade pip (upgrading pip from the command prompt within spyder did not work, I got permission errors there).<\/p>\n\n<p>(iii) InstallTensorFlow from the same command prompt using \"pip install tensorflow\". (The full pip command given on the TensorFlow download page did not work, I got permission errors using it.)<\/p>\n"},{"AnswerId":"41509470","CreationDate":"2017-01-06T16:03:48.393","ParentId":null,"OwnerUserId":"1246277","Title":null,"Body":"<p>Here's a simple workaround for this is that doesn't require reinstalling anything. <\/p>\n\n<p>Download the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/input_data.py\" rel=\"nofollow noreferrer\">input_data.py file from here<\/a>, then add the following at the bottom of the file:<\/p>\n\n<pre><code>mnist = read_data_sets('MNIST_data', one_hot=True)\n<\/code><\/pre>\n\n<p>After that, you can use Anaconda's python to run the <code>input_data.py<\/code> file:<\/p>\n\n<pre><code>$ python input_data.py\n<\/code><\/pre>\n\n<p>You only need to do this once. It will download the dataset into the <code>MNIST_data<\/code> folder without issuing the permissions error. After that, you can follow the TensorFlow tutorial and import the MNIST dataset as usual, and it should work correctly.<\/p>\n\n<p><strong>Edit:<\/strong> A possibly even-simpler workaround would be to put the following two lines of code into a <code>.py<\/code> file and run it using the python interpreter:<\/p>\n\n<pre><code>from tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":41377855,"AnswerCount":0,"Tags":"<tensorflow><lstm>","CreationDate":"2016-12-29T10:32:37.133","AcceptedAnswerId":null,"OwnerUserId":3144820.0,"Title":"Implementing __call__ function in MyBasicLSTMCell in tensorflow","Body":"<p>I have a weird problem with my code. I am trying to implement the  function on . Here is the problem. I have the following code.<\/p>\n\n<pre><\/pre>\n\n<p>In the above code, variable  is not being used anywhere. And suprisingly depending on whether I comment it or not, I get different outputs. Weird. What is more weird is that when I move the  variable after the  variable, the difference goes away. This means that irrespective of comment, I get the same output.<\/p>\n","answers":[]}
{"QuestionId":41378461,"AnswerCount":2,"Tags":"<python><deep-learning><keras>","CreationDate":"2016-12-29T11:07:28.970","AcceptedAnswerId":41386444.0,"OwnerUserId":6636290.0,"Title":"How to use models from keras.applications for transfer learnig?","Body":"<p>I want to get pretrained VGG16 model in Keras, remove its output layer, and then put a new output layer with the number of classes suited for my problem, and then to fit it on new data. For this reason, I am trying to use the model here: <a href=\"https:\/\/keras.io\/applications\/#vgg16\" rel=\"noreferrer\">https:\/\/keras.io\/applications\/#vgg16<\/a>, but since it is not Sequential, I cannot just . Popping from layers and adding it also does not work, because in the predictions it still expects the old shape. How would I do that? Is there a way to convert this type of model to ?<\/p>\n","answers":[{"AnswerId":"49419586","CreationDate":"2018-03-22T02:30:46.323","ParentId":null,"OwnerUserId":"3994228","Title":null,"Body":"<p>We can transform VGG model into Sequential as:<\/p>\n\n<pre><code># Create VGG model\nvgg_model = keras.applications.vgg16.VGG16(weights='imagenet')\n\n\n# Created model is of type Model\ntype(vgg_model)\n&gt;&gt; keras.engine.training.Model\n\n\n# Convert it to Sequential\nmodel = Sequential()\nfor layer in vgg_model.layers:\n   model.add(layer)\n\n\n# Now, check the model type, its Sequential! \ntype(model)\n&gt;&gt; keras.models.Sequential\n\n\n# Verify the model details\nmodel.summary()\n&gt;&gt; \n_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ninput_15 (InputLayer)        (None, 224, 224, 3)       0         \n_________________________________________________________________\nblock1_conv1 (Conv2D)        (None, 224, 224, 64)      1792      \n_________________________________________________________________\nblock1_conv2 (Conv2D)        (None, 224, 224, 64)      36928     \n_________________________________________________________________\nblock1_pool (MaxPooling2D)   (None, 112, 112, 64)      0         \n_________________________________________________________________\nblock2_conv1 (Conv2D)        (None, 112, 112, 128)     73856     \n_________________________________________________________________\nblock2_conv2 (Conv2D)        (None, 112, 112, 128)     147584    \n_________________________________________________________________\nblock2_pool (MaxPooling2D)   (None, 56, 56, 128)       0         \n_________________________________________________________________\nblock3_conv1 (Conv2D)        (None, 56, 56, 256)       295168    \n_________________________________________________________________\nblock3_conv2 (Conv2D)        (None, 56, 56, 256)       590080    \n_________________________________________________________________\nblock3_conv3 (Conv2D)        (None, 56, 56, 256)       590080    \n_________________________________________________________________\nblock3_pool (MaxPooling2D)   (None, 28, 28, 256)       0         \n_________________________________________________________________\nblock4_conv1 (Conv2D)        (None, 28, 28, 512)       1180160   \n_________________________________________________________________\nblock4_conv2 (Conv2D)        (None, 28, 28, 512)       2359808   \n_________________________________________________________________\nblock4_conv3 (Conv2D)        (None, 28, 28, 512)       2359808   \n_________________________________________________________________\nblock4_pool (MaxPooling2D)   (None, 14, 14, 512)       0         \n_________________________________________________________________\nblock5_conv1 (Conv2D)        (None, 14, 14, 512)       2359808   \n_________________________________________________________________\nblock5_conv2 (Conv2D)        (None, 14, 14, 512)       2359808   \n_________________________________________________________________\nblock5_conv3 (Conv2D)        (None, 14, 14, 512)       2359808   \n_________________________________________________________________\nblock5_pool (MaxPooling2D)   (None, 7, 7, 512)         0         \n_________________________________________________________________\nflatten (Flatten)            (None, 25088)             0         \n_________________________________________________________________\nfc1 (Dense)                  (None, 4096)              102764544 \n_________________________________________________________________\nfc2 (Dense)                  (None, 4096)              16781312  \n_________________________________________________________________\npredictions (Dense)          (None, 1000)              4097000   \n=================================================================\nTotal params: 138,357,544\nTrainable params: 138,357,544\nNon-trainable params: 0\n_________________________________________________________________\n\n\n# Now, that its sequential, we can perform usual operations.\nmodel.layers.pop()\n\n\n# Freeze the layers \nfor layer in model.layers:\nlayer.trainable = False\n\n\n# Add 'softmax' instead of earlier 'prediction' layer.\nmodel.add(Dense(2, activation='softmax'))\n\n\n# Check the summary, and yes new layer has been added. \nmodel.summary()\n\n    Layer (type)                 Output Shape              Param #   \n=================================================================\ninput_15 (InputLayer)        (None, 224, 224, 3)       0         \n_________________________________________________________________\nblock1_conv1 (Conv2D)        (None, 224, 224, 64)      1792      \n_________________________________________________________________\nblock1_conv2 (Conv2D)        (None, 224, 224, 64)      36928     \n_________________________________________________________________\nblock1_pool (MaxPooling2D)   (None, 112, 112, 64)      0         \n_________________________________________________________________\nblock2_conv1 (Conv2D)        (None, 112, 112, 128)     73856     \n_________________________________________________________________\nblock2_conv2 (Conv2D)        (None, 112, 112, 128)     147584    \n_________________________________________________________________\nblock2_pool (MaxPooling2D)   (None, 56, 56, 128)       0         \n_________________________________________________________________\nblock3_conv1 (Conv2D)        (None, 56, 56, 256)       295168    \n_________________________________________________________________\nblock3_conv2 (Conv2D)        (None, 56, 56, 256)       590080    \n_________________________________________________________________\nblock3_conv3 (Conv2D)        (None, 56, 56, 256)       590080    \n_________________________________________________________________\nblock3_pool (MaxPooling2D)   (None, 28, 28, 256)       0         \n_________________________________________________________________\nblock4_conv1 (Conv2D)        (None, 28, 28, 512)       1180160   \n_________________________________________________________________\nblock4_conv2 (Conv2D)        (None, 28, 28, 512)       2359808   \n_________________________________________________________________\nblock4_conv3 (Conv2D)        (None, 28, 28, 512)       2359808   \n_________________________________________________________________\nblock4_pool (MaxPooling2D)   (None, 14, 14, 512)       0         \n_________________________________________________________________\nblock5_conv1 (Conv2D)        (None, 14, 14, 512)       2359808   \n_________________________________________________________________\nblock5_conv2 (Conv2D)        (None, 14, 14, 512)       2359808   \n_________________________________________________________________\nblock5_conv3 (Conv2D)        (None, 14, 14, 512)       2359808   \n_________________________________________________________________\nblock5_pool (MaxPooling2D)   (None, 7, 7, 512)         0         \n_________________________________________________________________\nflatten (Flatten)            (None, 25088)             0         \n_________________________________________________________________\nfc1 (Dense)                  (None, 4096)              102764544 \n_________________________________________________________________\nfc2 (Dense)                  (None, 4096)              16781312  \n_________________________________________________________________\ndense_4 (Dense)              (None, 2)                 2002      \n=================================================================\nTotal params: 134,262,546\nTrainable params: 2,002\nNon-trainable params: 134,260,544\n_________________________________________________________________\n<\/code><\/pre>\n"},{"AnswerId":"41386444","CreationDate":"2016-12-29T20:02:07.983","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can use <code>pop()<\/code> on <code>model.layers<\/code> and then use <code>model.layers[-1].output<\/code> to create new layers. <\/p>\n\n<p>Example:<\/p>\n\n<pre><code>from keras.models import Model\nfrom keras.layers import Dense,Flatten\nfrom keras.applications import vgg16\nfrom keras import backend as K\n\nmodel = vgg16.VGG16(weights='imagenet', include_top=True)\n\nmodel.input\n\nmodel.summary(line_length=150)\n\nmodel.layers.pop()\nmodel.layers.pop()\n\nmodel.summary(line_length=150)\n\nnew_layer = Dense(10, activation='softmax', name='my_dense')\n\ninp = model.input\nout = new_layer(model.layers[-1].output)\n\nmodel2 = Model(inp, out)\nmodel2.summary(line_length=150)\n<\/code><\/pre>\n\n<p>Alternatively, you can use <code>include_top=False<\/code> option of these models. In this case if you need to use flatten the layer then you need to pass the <code>input_shape<\/code> also.<\/p>\n\n<pre><code>model3 = vgg16.VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 3))\nmodel3.summary(line_length=150)\n\nflatten = Flatten()\nnew_layer2 = Dense(10, activation='softmax', name='my_dense_2')\n\ninp2 = model3.input\nout2 = new_layer2(flatten(model3.output))\n\nmodel4 = Model(inp2, out2)\nmodel4.summary(line_length=150)\n<\/code><\/pre>\n"}]}
{"QuestionId":41378674,"AnswerCount":1,"Tags":"<python><theano><lasagne>","CreationDate":"2016-12-29T11:20:22.780","AcceptedAnswerId":41505017.0,"OwnerUserId":2251132.0,"Title":"Why can't I use dropout in the last layer of a lasagne regression model?","Body":"<p>A toy regression example. With  this works fine and the cost decreases. With  I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>Any thoughts? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41505017","CreationDate":"2017-01-06T11:45:30.063","ParentId":null,"OwnerUserId":"3038791","Title":null,"Body":"<p>Remove the comma after your dropout layer. \nThe code will work with the dropout immediately after the InputLayer or the DenseLayer. The comma creates a tuple with the network variable <em>(network, )<\/em> that causes the error.<\/p>\n\n<pre><code>network = lasagne.layers.InputLayer(shape=(None, num_features), input_var=x)\nif dropout &gt; 0.0:\n    network = lasagne.layers.dropout(network, p=dropout),\nnetwork = lasagne.layers.DenseLayer( network, num_units=1, nonlinearity=None )\n<\/code><\/pre>\n"}]}
{"QuestionId":41379309,"AnswerCount":1,"Tags":"<python><machine-learning><scikit-learn><tensorflow><conv-neural-network>","CreationDate":"2016-12-29T12:00:29.660","AcceptedAnswerId":41382756.0,"OwnerUserId":7353845.0,"Title":"Processing Image To Feed data in to Convolutional Neural Network","Body":"<p>I am new to Machine Learning and recently took a courser by Andrew Ng on Coursera.\nAfter that I shifted to Python and used Pandas, Numpy, Sklearn to implement ML algorithms.\nNow while surfing I came across tensorFLow and found it pretty amazing, and implemented this <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">example<\/a> which takes MNIST data as input. \nNow I want to read my own custom images and use them for training. I am confused as to how should I convert the images to MNIST sort of data. Or some other way to train my Network. \nI took this <a href=\"https:\/\/pythonprogramming.net\/tensorflow-neural-network-session-machine-learning-tutorial\/\" rel=\"nofollow noreferrer\">tutorial<\/a> to create my network.<\/p>\n","answers":[{"AnswerId":"41382756","CreationDate":"2016-12-29T15:35:50.653","ParentId":null,"OwnerUserId":"4899308","Title":null,"Body":"<p>Information on the MNIST dataset can be found on <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">Yann LeCun's website<\/a>.<\/p>\n\n<p>The TensorFlow module <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py\" rel=\"nofollow noreferrer\">tensorflow.examples.tutorials.mnist.mnist_softmax.py<\/a> looks to be acquiring\/preparing the dataset for the train\/test steps.<\/p>\n\n<p>The MNIST dataset contains an image of a handwritten digit and a corresponding label. If you would like to create labels for a new image, you could use <a href=\"https:\/\/docs.scipy.org\/doc\/scipy-0.18.1\/reference\/generated\/scipy.misc.imread.html\" rel=\"nofollow noreferrer\">scipy.misc.imread<\/a>.<\/p>\n"}]}
{"QuestionId":41380037,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-29T12:47:23.227","AcceptedAnswerId":41384322.0,"OwnerUserId":7354006.0,"Title":"Where have the old TensorFlow RNN files gone?","Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow<\/a><\/p>\n\n<p>I found that below files related with RNN are all gone.<\/p>\n\n<pre><\/pre>\n\n<p>Seems new  has come up...<\/p>\n\n<p>I'm a little confused by what happened. Why are these files deleted? Actually I'm analyzing these codes to train a chatbot.<\/p>\n","answers":[{"AnswerId":"41384322","CreationDate":"2016-12-29T17:17:52.537","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>There has been a major refactor in the TensorFlow repository, which is causing a few errors and some inconvinience unfortunately. Coming to your questions,<\/p>\n\n<blockquote>\n  <p>tensorflow\/tensorflow\/models\/...<\/p>\n<\/blockquote>\n\n<p>All the models have moved to <a href=\"https:\/\/github.com\/tensorflow\/models\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models<\/a>.<\/p>\n\n<blockquote>\n  <p>tensorflow\/tensorflow\/python\/ops\/seq2seq.py<\/p>\n<\/blockquote>\n\n<p>This file has been moved to the <code>legacy_seq2seq<\/code>, more specifically to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/legacy_seq2seq\/python\/ops\/seq2seq.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/legacy_seq2seq\/python\/ops\/seq2seq.py<\/a>.<\/p>\n\n<blockquote>\n  <p>tensorflow\/tensorflow\/python\/ops\/rnn_cell.py<\/p>\n<\/blockquote>\n\n<p>The implementations of the RNN cells have moved to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py<\/a>.<\/p>\n"}]}
{"QuestionId":41380126,"AnswerCount":5,"Tags":"<python><tensorflow>","CreationDate":"2016-12-29T12:51:39.880","AcceptedAnswerId":41380540.0,"OwnerUserId":2500650.0,"Title":"What does `tf.strided_slice()` do?","Body":"<p>I am wondering what  operator actually does.<br>\nThe <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#strided_slice\" rel=\"noreferrer\">doc<\/a> says,<\/p>\n\n<blockquote>\n  <p>To a first order, this operation extracts a slice of size end - begin from a tensor input starting at the location specified by begin. The slice continues by adding stride to the begin index until all dimensions are not less than end. Note that components of stride can be negative, which causes a reverse slice.<\/p>\n<\/blockquote>\n\n<p>And in the sample, <\/p>\n\n<pre><\/pre>\n\n<p>So in my understanding of the doc, the first sample (),<\/p>\n\n<ul>\n<li>resulting size is . The sample result shows , that shape is , it seems OK.<\/li>\n<li>the first element of the result is at . The first element of the sample result is , which is , it seems OK.<\/li>\n<li>the slice continues by adding stride to the begin index. So the second element of the result should be , but the sample shows the second element is .<\/li>\n<\/ul>\n\n<p>What  does?<\/p>\n\n<p>(Note: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5761\" rel=\"noreferrer\">method names in the samples and the last example is incorrect<\/a>.)<\/p>\n","answers":[{"AnswerId":"44403719","CreationDate":"2017-06-07T04:48:10.320","ParentId":null,"OwnerUserId":"3152627","Title":null,"Body":"<p>The conceptualization that really helped me understand this was that this function emulates the indexing behavior of numpy arrays.<\/p>\n\n<p>If you're familiar with numpy arrays, you'll know that you can make slices via <code>input[start1:end1:step1, start2:end2:step2, ... startN:endN:stepN]<\/code>. Basically, a very succinct way of writing <code>for<\/code> loops to get certain elements of the array.<\/p>\n\n<p>(If you're familiar with python indexing, you know that you can grab an array slice via <code>input[start:end:step]<\/code>. Numpy arrays, which may be nested, make use of the above tuple of slice objects.)<\/p>\n\n<p>Well, <code>strided_slice<\/code> just allows you to do this fancy indexing without the syntactic sugar. The numpy example from above just becomes<\/p>\n\n<pre><code># input[start1:end1:step1, start2:end2:step2, ... startN:endN:stepN]\ntf.strided_slice(input, [start1, start2, ..., startN],\n    [end1, end2, ..., endN], [step1, step2, ..., stepN])\n<\/code><\/pre>\n\n<p>The documentation is a bit confusing about this in the sense that:<\/p>\n\n<p><strong>a) <code>begin<\/code> - <code>end<\/code> is not strictly the shape of the return value<\/strong>:<\/p>\n\n<p>The documentation claims otherwise, but this is only true if your strides are all ones.\nExamples:<\/p>\n\n<pre><code>rank1 = tf.constant(list(range(10)))\n# The below op is basically:\n# rank1[1:10:2] =&gt; [1, 3, 5, 7, 9]\ntf.strided_slice(rank1, [1], [10], [2])\n\n# [10,10] grid of the numbers from 0 to 99\nrank2 = tf.constant([[i+j*10 for i in range(10)] for j in range(10)])\n# The below op is basically:\n# rank2[3:7:1, 5:10:2] =&gt; numbers 30 - 69, ending in 5, 7, or 9\nsliced = tf.strided_slice(rank2, [3, 5], [7, 10], [1, 2])\n# The below op is basically:\n# rank2[3:7:1] =&gt; numbers 30 - 69\nsliced = tf.strided_slice(rank2, [3], [7], [1]) \n<\/code><\/pre>\n\n<p><strong>b) it states that \"<code>begin<\/code>, <code>end<\/code>, and <code>strides<\/code> will be all length n, where n is in general not the same dimensionality as <code>input<\/code>\"<\/strong><\/p>\n\n<p>It sounds like dimensionality means rank here - but <code>input<\/code> <em>does<\/em> have to be a tensor of at least rank-n; it can't be lower (see rank-2 example above).<\/p>\n\n<p>N.B. I've said nothing\/not really explored the masking feature, but that seems beyond the scope of the question.<\/p>\n"},{"AnswerId":"47500563","CreationDate":"2017-11-26T20:11:13.313","ParentId":null,"OwnerUserId":"5651514","Title":null,"Body":"<p>tf.strided_slice() is used to do numpy style slicing of a tensor variable.\nIt has 4 parameters in general: input, begin, end, strides.The slice continues by adding stride to the begin index until all dimensions are not less than the end. For ex: \nLet us take a tensor constant named \"sample\" of dimensions: [3,2,3]<\/p>\n\n<pre><code>import tensorflow as tf \n\nsample = tf.constant(\n    [[[11, 12, 13], [21, 22, 23]],\n    [[31, 32, 33], [41, 42, 43]],\n    [[51, 52, 53], [61, 62, 63]]])\n\nslice = tf.strided_slice(sample, begin=[0,0,0], end=[3,2,3], strides=[2,2,2])\n\nwith tf.Session() as sess:\n    print(sess.run(slice))\n<\/code><\/pre>\n\n<p>Now, the output will be:<\/p>\n\n<pre><code>[[[11 13]]\n\n [[51 53]]]\n<\/code><\/pre>\n\n<p>This is because the striding starts from [0,0,0] and goes to [2,1,2] discarding any non-existent data like:<\/p>\n\n<pre><code>[[0,0,0], [0,0,2], [0,2,0], [0,2,2],\n[2,0,0], [2,0,2], [2,2,0], [2,2,2]]\n<\/code><\/pre>\n\n<p>If you use [1,1,1] as strides then it will simply print all the values.<\/p>\n"},{"AnswerId":"43440621","CreationDate":"2017-04-16T18:28:26.093","ParentId":null,"OwnerUserId":"5291316","Title":null,"Body":"<p>I experimented a bit with this method, which gave me some insights, which I think might be of some use. let's say we have a tensor. <\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>a = np.array([[[1, 1.2, 1.3], [2, 2.2, 2.3], [7, 7.2, 7.3]],\n              [[3, 3.2, 3.3], [4, 4.2, 4.3], [8, 8.2, 8.3]],\n              [[5, 5.2, 5.3], [6, 6.2, 6.3], [9, 9.2, 9.3]]]) \n# a.shape = (3, 3, 3)\n<\/code><\/pre>\n\n<p><code>strided_slice()<\/code> requires 4 required arguments <code>input_, begin, end, strides<\/code> in which we are giving our <code>a<\/code> as <code>input_<\/code> argument.\n As the case with <code>tf.slice()<\/code> method, the <code>begin<\/code> argument is zero-based and rest of args shape-based.  <strong>However in the docs <code>begin<\/code> and <code>end<\/code> both are zero-based<\/strong>.<\/p>\n\n<p>The functionality of method is quite simple: <br>\nIt works like iterating over a loop, where <code>begin<\/code> is the location of element in the tensor from where the loop initiates and <code>end<\/code> is where it stops. <br><\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.strided_slice(a, [0, 0, 0], [3, 3, 3], [1, 1, 1])\n\n# output =  the tensor itself\n\ntf.strided_slice(a, [0, 0, 0], [3, 3, 3], [2, 2, 2])\n\n# output = [[[ 1.   1.3]\n#            [ 7.   7.3]]\n#           [[ 5.   5.3]\n#            [ 9.   9.3]]]\n<\/code><\/pre>\n\n<p><code>strides<\/code> are like steps over which the loop iterates, here the <code>[2,2,2]<\/code> makes method to produce values starting at (0,0,0), (0,0,2), (0,2,0), (0,2,2), (2,0,0), (2,0,2) ..... in the <code>a<\/code> tensor.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tf.strided_slice(input3, [1, 1, 0], [2, -1, 3], [1, 1, 1]) \n<\/code><\/pre>\n\n<p>will produce output similar to <code>tf.strided_slice(input3, [1, 1, 0], [2, 2, 3], [1, 1, 1])<\/code> as the tensor<code>a<\/code> has <code>shape = (3,3,3)<\/code>. <\/p>\n"},{"AnswerId":"54857808","CreationDate":"2019-02-25T00:00:17.960","ParentId":null,"OwnerUserId":"7552040","Title":null,"Body":"<p>I find this technique useful to debug the solution.\nrule: always omit recurring patterns and try to keep step at <code>(end-1)<\/code>.<\/p>\n\n<pre><code>t = tf.constant([[[1, 1, 1], [2, 2, 2]],\n             [[3, 3, 3], [4, 4, 4]],\n             [[5, 5, 5], [6, 6, 6]]])  \n\n# ex 1:\ntf.strided_slice(t, [1, 0, 0], [2, 1, 3], [1, 1, 1])\n\n# 3rd position:\n1,0,0 &gt; 3     \n1,0,1 &gt; 3\n1,0,2 &gt; 3\n# 2nd and 1st position:satisfies the rule listed above, skipping these.\n\n# ex 2:\ntf.strided_slice(t, [1, 0, 0], [2, 2, 3], [1, 1, 1])\n\n# 3rd position:\n1,0,0 &gt; 3     \n1,0,1 &gt; 3\n1,0,2 &gt; 3\n# 2nd positon:\n1,1,0 &gt; 4\n1,1,1 &gt; 4\n1,1,2 &gt; 4\n# 1st position: satisfies the rule listed above, skipping.\n\n# Ex 3:\ntf.strided_slice(t, [1, -1, 0], [2, -3, 3], [1, -1, 1])\n\n# 3rd position:\n1,-1,0 &gt; 4\n1,-1,1 &gt; 4\n1,-1,2 &gt; 4\n# 2nd position:\n1,-2,0 &gt; 3\n1,-2,1 &gt; 3\n1,-2,2 &gt; 3\n# 1st position:satisfies the rule listed above, skipping.\n<\/code><\/pre>\n"},{"AnswerId":"41380540","CreationDate":"2016-12-29T13:15:55.257","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>The mistake in your argument is the fact that you are directly adding the lists <code>strides<\/code> and <code>begin<\/code> element by element. This will make the function a lot less useful. Instead, it increments the <code>begin<\/code> list one dimension at a time, starting from the last dimension.<\/p>\n\n<p>Let's solve the first example part by part. <code>begin = [1, 0, 0]<\/code> and <code>end = [2, 1, 3]<\/code>. Also, all the <code>strides<\/code> are <code>1<\/code>. Work your way backwards, from the last dimension.<\/p>\n\n<p>Start with element <code>[1,0,0]<\/code>. Now <strong>increase the last dimension only by <em>its<\/em> stride amount<\/strong>, giving you <code>[1,0,1]<\/code>. Keep doing this until you reach the limit. Something like <code>[1,0,2]<\/code>, <code>[1,0,3]<\/code> (end of the loop). Now in your next iteration, start by incrementing the second to last dimension and resetting the last dimension, <code>[1,1,0]<\/code>. Here the second to last dimension is equal to <code>end[1]<\/code>, so move to the first dimension (third to last) and reset the rest, giving you <code>[2,0,0]<\/code>. Again you are at the first dimension's limit, so quit the loop.<\/p>\n\n<p>The following code is a recursive implementation of what I described above,<\/p>\n\n<pre><code># Assume global `begin`, `end` and `stride`\ndef iterate(active, dim):\n    if dim == len(begin):\n        # last dimension incremented, work on the new matrix\n        # Note that `active` and `begin` are lists\n        new_matrix[active - begin] = old_matrix[active]\n    else:\n        for i in range(begin[dim], end[dim], stride[dim]):\n            new_active = copy(active)\n            new_active[dim] = i\n            iterate(new_active, dim + 1)\n\niterate(begin, 0)\n<\/code><\/pre>\n"}]}
{"QuestionId":41381596,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-29T14:25:45.477","AcceptedAnswerId":41396174.0,"OwnerUserId":1150961.0,"Title":"How to calculate gradients in a numerically stable fashion","Body":"<p>I would like to compute derivatives of a ratio  in a numerically stable fashion using tensorflow but am running into problems when  and  are small ( when using 32-bit floating point representation). Of course, the derivative of  is  but because of the operator precedence, the square in the denominator is computed first, underflows, and leads to an undefined gradient.<\/p>\n\n<p>If the derivative was calculated as , the underflow would not occur and the gradient would be well-defined as illustrated in the figure below which shows the gradient as a function of . The blue line corresponds to the domain in which tensorflow can calculate the derivative. The orange line corresponds to the domain in which the denominator underflows  yielding an infinite gradient. The green line corresponds to the domain in which the denominator overflows  yielding a zero gradient. In both problematic domains, the gradient can be calculated using the modified expression above.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/NqBuI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NqBuI.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I've been able to get a more numerically stable expression by using the ugly hack<\/p>\n\n<pre><\/pre>\n\n<p>which is equivalent to  but yields a different tensorflow graph. But I run into the same problem if I want to calculate a higher-order derivative. The code to reproduce the problem can be found <a href=\"https:\/\/gist.github.com\/tillahoffmann\/cff5ce8cad89b5a61092ed2d4b88c6fc\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Is there a recommended approach to alleviate such problems? Is it possible to explicitly define a derivative of an expression in tensorflow if one doesn't want to rely on autodifferentiation?<\/p>\n","answers":[{"AnswerId":"41396174","CreationDate":"2016-12-30T12:20:48.857","ParentId":null,"OwnerUserId":"1150961","Title":null,"Body":"<p>Thanks to Yaroslav Bulatov's pointer, I was able to implement a custom function with the desired gradient.<\/p>\n\n<pre><code># Define the division function and its gradient\n@function.Defun(tf.float32, tf.float32, tf.float32)\ndef newDivGrad(x, y, grad):\n    return tf.reciprocal(y) * grad, - tf.div(tf.div(x, y), y) * grad\n\n\n@function.Defun(tf.float32, tf.float32, grad_func=newDivGrad)\ndef newDiv(x, y):\n    return tf.div(x, y)\n<\/code><\/pre>\n\n<p>Full notebook is <a href=\"https:\/\/gist.github.com\/tillahoffmann\/e2e639335f89f2e5e270bb18edd11ded\" rel=\"nofollow noreferrer\">here<\/a>. PR is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/6562\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Xgdjv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Xgdjv.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":41383244,"AnswerCount":1,"Tags":"<neural-network><deep-learning><keras><autoencoder>","CreationDate":"2016-12-29T16:07:07.760","AcceptedAnswerId":null,"OwnerUserId":7354709.0,"Title":"Using hidden activations in loss function","Body":"<p>I want to create a custom loss function for a double-input double-output model in Keras that:<\/p>\n\n<ol>\n<li>minimizes the reconstruction error of two autoencoders;<\/li>\n<li>maximizes the correlation of the bottleneck features of the autoencoders.<\/li>\n<\/ol>\n\n<p>For this I need to pass to the loss function:<\/p>\n\n<ul>\n<li>both inputs;<\/li>\n<li>both outputs \/ reconstructions;<\/li>\n<li>output of intermediate layers for both (hidden activations).<\/li>\n<\/ul>\n\n<p>I know I can pass both inputs and outputs to Model, but am struggling to find a way to pass the hidden activations.<\/p>\n\n<p>I could create two new Models that have the output of the intermediate layers and pass that to loss, like:<\/p>\n\n<pre><\/pre>\n\n<p>But still, I would need to find a way to match the  in loss to the correct intermediate model.<\/p>\n\n<p>What is the right way to approach this?<\/p>\n\n<p><strong>Edit<\/strong><br>\nHere's an approach that I think should work. Simplified:  <\/p>\n\n<pre><\/pre>\n\n<p>Then in  I can split  and do the calculations.<\/p>\n","answers":[{"AnswerId":"41387177","CreationDate":"2016-12-29T21:05:42.143","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>How about:<\/p>\n\n<ol>\n<li><p>Defining a single model with a multiple outputs (be sure that you named a coding and reconstruction layer properly):<\/p>\n\n<pre><code>duo_model = Model(input=input, output=[coding_layer, reconstruction_layer])\n<\/code><\/pre><\/li>\n<li><p>Compiling your model with two different losses (or even performing a loss reweighting):<\/p>\n\n<pre><code>duo_model.compile(optimizer='rmsprop',\n                   loss={'coding_layer': correlation_loss, \n                         'reconstruction_layer': 'mse'})\n<\/code><\/pre><\/li>\n<li><p>Taking your final model as a:<\/p>\n\n<pre><code>encoder = Model(input=input, output=[coding_layer])\nautoencoder = Model(input=input, output=[reconstruction_layer])\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>After proper compilation this should do the job.<\/p>\n\n<p>When it comes to defining a proper correlation loss function there are two ways:<\/p>\n\n<ol>\n<li>when coding layer and your output layer have the same dimension -\nyou could easly use predefinied <code>cosine_proximity<\/code> function from\nKeras library.<\/li>\n<li>when coding layer has different dimensonality -\nyou shoud first find embedding of coding vector and reconstruction vector to the same space and then - compute correlation there. Remember that this embedding should either be a Keras layer \/ function or Theano \/ Tensor flow operation (depending on which backend you are using). Of course you can compute both embedding and correlation function as a part of one loss function.<\/li>\n<\/ol>\n"}]}
{"QuestionId":41383300,"AnswerCount":1,"Tags":"<python><numpy><neural-network><theano>","CreationDate":"2016-12-29T16:10:36.693","AcceptedAnswerId":null,"OwnerUserId":3687629.0,"Title":"TypeError: only length-1 arrays can be converted to Python scalars Neural network","Body":"<p>When training my Neural Network, It gives me this error at , but when I remove the float casting to the X_Train and X_Test the error is removed but I get another error instead  at this part of code <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41385116","CreationDate":"2016-12-29T18:15:46.007","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>Make sure you understand what <\/p>\n\n<pre><code>X_Test = x_actual[int(ceil((lines*(percentage\/100)))+1):]\n<\/code><\/pre>\n\n<p>produces (the outer layer of () is unnecessary).  It probably is an array - and judging from the error, 2d dimensional.  So iterating on it produces its rows, 1d arrays.  <\/p>\n\n<p>If it is too large to print whole, print its <code>shape<\/code> and <code>dtype<\/code>.  <code>type(X_Test)<\/code> might also help to verify that it is an <code>ndarray<\/code>.<\/p>\n\n<p>Even if it worked, this expression would not change <code>X_Test<\/code> to floats.  It just produces a new list of something or other.<\/p>\n\n<pre><code> [float(i) for i in X_Test]\n<\/code><\/pre>\n\n<p><code>X_Test = X_Test.astype(float)<\/code> is required to produce a new array of floats.<\/p>\n\n<p>I think your 2nd error has nothing to do with the first one.  You just never go to it.  That looks like a <code>Theano<\/code> error, which is outside my area of expertise.<\/p>\n"}]}
{"QuestionId":41384167,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-29T17:06:56.920","AcceptedAnswerId":null,"OwnerUserId":6227294.0,"Title":"No gradient defined for operation MaxPoolWithArgmax in tensorflow","Body":"<p>I wanna use  operation in TensorFlow, but I got following error:<\/p>\n\n<blockquote>\n  <p>LookupError: No gradient defined for operation 'MaxPoolWithArgmax_1' (op type: MaxPoolWithArgmax)<\/p>\n<\/blockquote>\n\n<p>Here is a piece of my code that I am using :<\/p>\n\n<pre><\/pre>\n\n<p>and here is  module:<\/p>\n\n<pre><\/pre>\n\n<p>and after I built my network I am using  to train my network. <\/p>\n\n<p>By the way, when I am using  (instead of ) everything is working fine. <\/p>\n\n<p>I am running my code on GPU and I am using python 2.7 on ubuntu14.<\/p>\n\n<p>Thanks, Ali <\/p>\n","answers":[{"AnswerId":"41403158","CreationDate":"2016-12-30T21:56:31.080","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>The gradient op is actually implemented <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/ops\/nn_ops.cc#L1316\" rel=\"nofollow noreferrer\">here<\/a>, but somehow not registered <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/nn_grad.py#L399\" rel=\"nofollow noreferrer\">here<\/a>.  Would be worth filing a github issue on it.<\/p>\n\n<p>At the same time, you can register the op by yourself following this <a href=\"https:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/\" rel=\"nofollow noreferrer\">tutorial<\/a> (only the gradient registration part).<\/p>\n"}]}
{"QuestionId":41384796,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2016-12-29T17:50:08.493","AcceptedAnswerId":null,"OwnerUserId":7355100.0,"Title":"Run CIFAR10 eval script with one GPU","Body":"<p>I'm trying to run the CIFAR10 training + eval scripts from the TensorFlow tutorial.\nI have only one GPU.\nIs it possible to do so? Maybe make the eval script run somehow on the CPU?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
{"QuestionId":41385137,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-12-29T18:17:55.863","AcceptedAnswerId":null,"OwnerUserId":4993221.0,"Title":"No OpKernel was registered to support Op 'L2Loss' with these attrs [[Node: L2Loss = L2Loss[T=DT_INT64](L2Loss\/t)]]","Body":"<p>I am getting the above-mentioned error while using this tensorflow code:-<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41385450","CreationDate":"2016-12-29T18:42:53.997","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You should see something like this<\/p>\n\n<pre><code>InvalidArgumentError: No OpKernel was registered to support Op 'L2Loss' with these attrs.  Registered kernels:\n  device='CPU'; T in [DT_FLOAT]\n  device='CPU'; T in [DT_DOUBLE]\n  device='CPU'; T in [DT_HALF]\n\n     [[Node: L2Loss = L2Loss[T=DT_INT32](Const)]]\n<\/code><\/pre>\n\n<p>This means that <code>T<\/code> attribute(dtype) is <code>int32<\/code>, but ther are only <code>float16<\/code>, <code>float32<\/code>, <code>float64<\/code> implementations. So you can do this instead<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\ntf.reset_default_graph()\na = tf.constant([[1,1,1],[2,2,2]], dtype=tf.float32)\nnorm = tf.nn.l2_loss(a)\nsess = tf.Session()\nprint(sess.run(norm))\n<\/code><\/pre>\n"}]}
{"QuestionId":41385808,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network><lstm>","CreationDate":"2016-12-29T19:08:30.220","AcceptedAnswerId":null,"OwnerUserId":6095482.0,"Title":"Confused about weight and bias dependencies affecting learning","Body":"<p>I had a working LSTM model that had one weight\/bias layer from its recurrent state to the output.  I then also coded up the same system, but with two layers. This means that I would have the LSTM, then a hidden layer, and then the output.  I wrote the lines to define this double layer model, but did not use them a single time.  But, now that those layers exist but are not used at all, it wouldn't learn! So my weights and biases are defined like so:<\/p>\n\n<pre><\/pre>\n\n<p>So I had the double layer weights and biases defined, but they were never used a single time in training or testing.  <\/p>\n\n<p>My incorporation of the weights\/biases was a single line:<\/p>\n\n<pre><\/pre>\n\n<p>where relevant is the LSTM output.  So I am only using the 'out' variables in the weight and bias dictionaries.<\/p>\n\n<p>It wouldn't learn anything.  Then, once I commented the double layer variables out like this:<\/p>\n\n<pre><\/pre>\n\n<p>...it started working again.  How does the existence of those variables impede learning? I initialize them, but no gradients should be running through them, and backprop shouldn't have any association with those unused variables. Or am I misunderstanding something?<\/p>\n","answers":[]}
{"QuestionId":41386167,"AnswerCount":0,"Tags":"<python><tensorflow><keras><lstm>","CreationDate":"2016-12-29T19:38:58.643","AcceptedAnswerId":null,"OwnerUserId":6028884.0,"Title":"Keras Shape with unknown rank","Body":"<p>I am using the code from <a href=\"http:\/\/machinelearningmastery.com\/text-generation-lstm-recurrent-neural-networks-python-keras\/\" rel=\"nofollow noreferrer\" title=\"here\">here<\/a> in order to test out LSTM networks on text generation. The only line that I changed is <\/p>\n\n<pre><\/pre>\n\n<p>instead of <\/p>\n\n<pre><\/pre>\n\n<p>And I get this error <\/p>\n\n<pre><\/pre>\n\n<p>I've been looking around, I cannot understand why it just won't take the input dimensions. Any idea is more than welcome.<\/p>\n\n<p>Thanks in advance!<\/p>\n\n<p>Later Edit: It doesn't work even with the original file.<\/p>\n","answers":[]}
{"QuestionId":41386795,"AnswerCount":1,"Tags":"<python><tensorflow><ubuntu-14.04><jupyter-notebook>","CreationDate":"2016-12-29T20:33:13.503","AcceptedAnswerId":41386946.0,"OwnerUserId":980526.0,"Title":"Tensorflow saver: Kernel appears to have died","Body":"<p>I have a lot of troubles with saving \/ restoring tensorflow models, either my \"Kernels seems to have died\" or I get errors (\"Variable ... already exits\").<\/p>\n\n<p>When my Kernel dies, I get this error log in my console:<\/p>\n\n<pre><\/pre>\n\n<p>My code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>To restore the model, I use:<\/p>\n\n<pre><\/pre>\n\n<p>I tried a lot of different ways, sometimes it works (but not always!?), sometimes it crashes and sometime I get the Variable error. Do I have to use the saving\/restoring in an other way?<\/p>\n\n<p>I am using:\nUbuntu 14.04\nAnaconda 3\nPython 3.5.2\nTensorflow 0.12<\/p>\n\n<p>inside jupyter notebook<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"41386946","CreationDate":"2016-12-29T20:46:14.983","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>This can happen when you run out of memory, solution is to try smaller batch sizes. I see that you are feeding your test set into a single <code>run<\/code> call, which needs enough memory to do all examples at once. You can do something like <a href=\"https:\/\/github.com\/petewarden\/tensorflow_makefile\/blob\/master\/tensorflow\/models\/image\/mnist\/convolutional.py#L255\" rel=\"nofollow noreferrer\">eval_in_batches<\/a> to aggregate accuracy over several smaller run calls<\/p>\n"}]}
{"QuestionId":41386878,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2016-12-29T20:40:36.357","AcceptedAnswerId":null,"OwnerUserId":5213649.0,"Title":"Multi-dimension dynamic rnn with tensorflow","Body":"<p>In tensorflow's dynamic_rnn function, I was surprised by the output shape and I was hoping someone could help improve my understanding of the RNN cells. <\/p>\n\n<p>For example, if the input is defined as: <\/p>\n\n<pre><\/pre>\n\n<p>where seq_size = 5 and input_dim = 2 (ie. two time series) and 110 is the batch size; and the cell is defined as,<\/p>\n\n<pre><\/pre>\n\n<p>where hidden_dim = 6<\/p>\n\n<p>When I create a dynamic_rnn <\/p>\n\n<pre><\/pre>\n\n<p>and check the size of output it is [110 5 6]. These dimensions are batch size by seq_size by hidden_dim. <\/p>\n\n<p>Questions: <\/p>\n\n<p>1: These dimensions imply there are 6 hidden nodes per time step in the time series for a total of 30 hidden nodes (5 x 6) rather than 6 hidden nodes total. Is this the correct interpretation? <\/p>\n\n<p>2: Since my input dimension is 5 x 2 (5 steps in each sequence and 2 sequences) how is tensorflow connecting the inputs to the hidden nodes at each time step? Is the tensorflow assuming a fully connected graph with 2 inputs, 12 weights, and 6 biases before each hidden cell? Or something else?<\/p>\n","answers":[{"AnswerId":"41387354","CreationDate":"2016-12-29T21:21:14.437","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>You can think of a sequence as a sentence and an input as a word. The sequence length is the number of words in the sentence, which is also the number of hidden nodes in LSTM; each input\/word is corresponding to one hidden node, which maps the input to one output. This is why the number of output is seq_size (5).<\/p>\n\n<p>A word is a vector, that is positioned in a multi-dimensional space, whose number of dimensions is input_dim. In LSTM, a word is mapped from this input space to a higher dimensional space whose number of dimensions is hidden_dim. This is why the size of each output is hidden_dim (6).<\/p>\n\n<p>I believe epoch is an irrelevant concept for this discussion. Please see  <a href=\"https:\/\/stackoverflow.com\/questions\/41338553\/the-meaning-of-batch-size-in-ptb-word-lm-lstm-model-of-tensorflow\/41339126#41339126\">The meaning of batch_size in ptb_word_lm (LSTM model of tensorflow)<\/a><\/p>\n"}]}
{"QuestionId":41387511,"AnswerCount":1,"Tags":"<python-2.7><theano><theano-cuda>","CreationDate":"2016-12-29T21:34:41.257","AcceptedAnswerId":null,"OwnerUserId":6058173.0,"Title":"Theano Windows 7 import error","Body":"<p>Today I tried to install Theano on W7 x64. I guess I installed it but I also made a mistake in somewhere and I don't know where. I actually want to use it with CUDA. <\/p>\n\n<p>Here is the programs I installed:\nAnaconda 4.2.0 x64 (Python 2.7.12),\nTDM GCC,\nCUDA 8.0\nVisual Studio 13\nTheano via Git<\/p>\n\n<p>Here is the <em>theanorc<\/em> file:<\/p>\n\n<pre><\/pre>\n\n<p>Like I said I've no idea where I made a mistake. Here is the output when I write \"import theano\"<\/p>\n\n<p>Thanks in advance guys!<\/p>\n\n<pre><\/pre>\n\n<p>ps: one more error -> <a href=\"http:\/\/pastebin.com\/V59Pm9Qa\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/V59Pm9Qa<\/a><\/p>\n","answers":[{"AnswerId":"41406758","CreationDate":"2016-12-31T08:35:38.027","ParentId":null,"OwnerUserId":"7349628","Title":null,"Body":"<p>Have you added system variables? If no, try this:<\/p>\n\n<p>Right click Computer -> properties -> advanced system settings -> environment variables<\/p>\n\n<p>Add a new system variable<\/p>\n\n<p>Name = THEANO_FLAGS<\/p>\n\n<p>Value = floatX=float32,device=gpu,nvcc.fastmath=True<\/p>\n\n<p>Also add Visual Studio's c++ compiler to the path<\/p>\n\n<p>Add ;pathToYourVSInstallation\\VC\\bin\\<\/p>\n\n<p>If this wont work, please go there and do it again, from the top:\n<a href=\"https:\/\/stackoverflow.com\/a\/41401271\/7349628\">https:\/\/stackoverflow.com\/a\/41401271\/7349628<\/a><\/p>\n"}]}
{"QuestionId":41388235,"AnswerCount":2,"Tags":"<build><configuration><tensorflow>","CreationDate":"2016-12-29T22:45:43.287","AcceptedAnswerId":41388591.0,"OwnerUserId":3713926.0,"Title":"What do I put in the \"compute capability list\" field when configuring Tensorflow?","Body":"<p>I'm building Tensorflow from source, and it asks me to <\/p>\n\n<pre><\/pre>\n\n<p>I figured out that it has to do with my cards compute capability (which is 5.0, btw), but clearly that's not what it wants, because it's a number, not a list. I couldn't find any explanations, nor am I skilled enough to understand it from the source code. What should I put in there?<\/p>\n","answers":[{"AnswerId":"41388591","CreationDate":"2016-12-29T23:23:14.497","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>Summarizing the comments as an answer:<\/p>\n\n<p>You can put 5.0 there, which is the compute capability that your card supports and should be your best choice. 3.5 is too low (doesn't use the full features of your card) and 5.2 is too high (not supported by your card); in either case, I believe the binary code would be re-compiled with 5.0 just in time before the execution on the GPU. If you directly use 5.0, it won't trigger just in time compilation and will use more optimized offline compilation. <\/p>\n\n<p>If you provide a list, say 3.5,5.0,5.2, it will compile for all three compute capabilities offline and allow the offline compiled (more optimized) code to run on GPU cards that support either 3.5, or 5.0, or 5.2, without triggering less optimized just-in-time compilation. The cost of putting more compute capabilities in the list is the increased offline compilation time and binary size.<\/p>\n"},{"AnswerId":"49971927","CreationDate":"2018-04-23T00:21:15.050","ParentId":null,"OwnerUserId":"1596433","Title":null,"Body":"<p>Go to <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-gpus<\/a><\/p>\n\n<p>and find out your GPU. You will find the compute capability, like mine was 6.1<\/p>\n"}]}
{"QuestionId":41388673,"AnswerCount":2,"Tags":"<tensorflow><ipython><jupyter><graph-visualization><tensorboard>","CreationDate":"2016-12-29T23:35:10.250","AcceptedAnswerId":41463991.0,"OwnerUserId":3055351.0,"Title":"Visualizing a TensorFlow graph in Jupyter doesn't work","Body":"<p>I saw this <a href=\"https:\/\/stackoverflow.com\/questions\/38189119\/simple-way-to-visualize-a-tensorflow-graph-in-jupyter\">question<\/a> about how to visualise a tensorflow graph in Jupyter notebook. I found that this answer comes from <a href=\"https:\/\/stackoverflow.com\/questions\/38189119\/simple-way-to-visualize-a-tensorflow-graph-in-jupyter\">this<\/a> example with just one modification in ( is replaced by ). However if I try to rerun it on  the visualisation doesn't work: the iframe is white and there is no nodes displayed.<\/p>\n\n<p>I would highly appreciate if you explain me what I am doing wrong. Here there is a simple example to reproduce the problem.<\/p>\n\n<p>Import:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Create graph:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The output:<\/p>\n\n<pre><\/pre>\n\n<p>The visualisation functions:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The result:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/P3wg1.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/P3wg1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>UPD<\/strong> I tried a simpler example:<\/p>\n\n<pre><\/pre>\n\n<p>But it still doesn't work. I suspect the problem is related to Javascript\/HTML code that is generated:<\/p>\n\n<pre><\/pre>\n\n<p>Maybe something with  and ?<\/p>\n","answers":[{"AnswerId":"41389623","CreationDate":"2016-12-30T01:53:49.487","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's the version I'm using <a href=\"https:\/\/gist.github.com\/yaroslavvb\/97504b8221a8529e7a51a50915206d68\" rel=\"nofollow noreferrer\">right now<\/a><\/p>\n\n<p>You should be able to do something like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wbMje.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wbMje.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"41463991","CreationDate":"2017-01-04T12:35:53.480","ParentId":null,"OwnerUserId":"1082571","Title":null,"Body":"<p>The reason behind the failure is imports (<code>&lt;link rel=\"import\" ...<\/code>) are <a href=\"http:\/\/caniuse.com\/#feat=imports\" rel=\"noreferrer\">only supported under Chrome<\/a> failing in <a href=\"https:\/\/hacks.mozilla.org\/2014\/12\/mozilla-and-web-components\/\" rel=\"noreferrer\">Firefox<\/a> and Safari, and there's no sight to be adopted by others until WebComponents definition arrives. So, you'd better run Jupyter in Chrome.<\/p>\n\n<p>If you're against Chrome, there's good news. You can use a Polyfill (a piece of code that implements a feature on web browsers that do not support the feature) to make it work:<\/p>\n\n<pre><code>&lt;script src=\"\/\/cdnjs.cloudflare.com\/ajax\/libs\/polymer\/0.3.3\/platform.js\"&gt;&lt;\/script&gt;\n<\/code><\/pre>\n\n<p>I've tested it in Firefox and Safari and it works, but not perfectly fine. It is a little bit slower to load the Polypill and the graph canvas is reduced to one inch wide (I don't know why, TensorBoard internals). Then I realised <code>platform.js<\/code> has been <a href=\"http:\/\/webcomponents.org\/articles\/platformjs-becomes-webcomponentsjs\/\" rel=\"noreferrer\">deprecated<\/a> but new implementations incorporates new errors (unhandled events and XML parsing).<\/p>\n\n<p>Following is the modified code:<\/p>\n\n<pre><code># TensorFlow Graph visualizer code\nimport numpy as np\nfrom IPython.display import clear_output, Image, display, HTML\n\ndef strip_consts(graph_def, max_const_size=32):\n    \"\"\"Strip large constant values from graph_def.\"\"\"\n    strip_def = tf.GraphDef()\n    for n0 in graph_def.node:\n        n = strip_def.node.add() \n        n.MergeFrom(n0)\n        if n.op == 'Const':\n            tensor = n.attr['value'].tensor\n            size = len(tensor.tensor_content)\n            if size &gt; max_const_size:\n                tensor.tensor_content = \"&lt;stripped %d bytes&gt;\"%size\n    return strip_def\n\ndef show_graph(graph_def, max_const_size=32):\n    \"\"\"Visualize TensorFlow graph.\"\"\"\n    if hasattr(graph_def, 'as_graph_def'):\n        graph_def = graph_def.as_graph_def()\n    strip_def = strip_consts(graph_def, max_const_size=max_const_size)\n    code = \"\"\"\n        &lt;script src=\"\/\/cdnjs.cloudflare.com\/ajax\/libs\/polymer\/0.3.3\/platform.js\"&gt;&lt;\/script&gt;\n        &lt;script&gt;\n          function load() {{\n            document.getElementById(\"{id}\").pbtxt = {data};\n          }}\n        &lt;\/script&gt;\n        &lt;link rel=\"import\" href=\"https:\/\/tensorboard.appspot.com\/tf-graph-basic.build.html\" onload=load()&gt;\n        &lt;div style=\"height:600px\"&gt;\n          &lt;tf-graph-basic id=\"{id}\"&gt;&lt;\/tf-graph-basic&gt;\n        &lt;\/div&gt;\n    \"\"\".format(data=repr(str(strip_def)), id='graph'+str(np.random.rand()))\n\n    iframe = \"\"\"\n        &lt;iframe seamless style=\"width:1200px;height:620px;border:0\" srcdoc=\"{}\"&gt;&lt;\/iframe&gt;\n    \"\"\".format(code.replace('\"', '&amp;quot;'))\n    display(HTML(iframe))\n<\/code><\/pre>\n\n<p>Notice there's only one line added at the beginning of the <code>code = \"\"\"<\/code> block. It must be there because it's required by the Polyfill.<\/p>\n\n<p>Original source code can be found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/deepdream\/deepdream.ipynb\" rel=\"noreferrer\">here<\/a>. You can ask him to evolve it within Google to cover other browsers than Chrome, but I don't think it will happen.<\/p>\n"}]}
{"QuestionId":41389387,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-30T01:17:37.297","AcceptedAnswerId":41418934.0,"OwnerUserId":3084800.0,"Title":"tensorflow loss is nan while training an RNN","Body":"<p>Running an RNN with a single GRU cell, I'm running into situations where I get the following stacktrace<\/p>\n\n<pre><\/pre>\n\n<p>The step for calculating loss appears to be the issue<\/p>\n\n<pre><\/pre>\n\n<p>Here is the code for <\/p>\n\n<p>Any place with loss comes back <\/p>\n\n<pre><\/pre>\n\n<p>This is arising when I test my code by using the following for my data\n which is designed to mimic using word ids to pass around as inputs. The rest of my code can be found in the following <a href=\"https:\/\/gist.github.com\/IoDmitri\/499b8273d4504bec0ace61f39ce1a7c3\" rel=\"nofollow noreferrer\">gist<\/a><\/p>\n\n<p><strong>EDIT<\/strong> Here is the way that I run the test suit where this issue arises:<\/p>\n\n<pre><\/pre>\n\n<p>If I substitute  into other language models, they will also output  for the cost. My understanding is that tensorflow, via passing in the feed-dict, should be taking the numeric values and retrieving the appropriate embedding vector corresponding to the id, I'm not understand why the  is causing 's for other models. <\/p>\n","answers":[{"AnswerId":"41418934","CreationDate":"2017-01-01T20:53:47.177","ParentId":null,"OwnerUserId":"3084800","Title":null,"Body":"<p>There are several problems with the code above<\/p>\n\n<p>lets start with this line <\/p>\n\n<pre><code>random_data = np.random.normal(0, 100, size=[42068, 46])\n<\/code><\/pre>\n\n<p><code>np.random.normal(...)<\/code> doesn't produce distinct values, it rather produces float values, lets try the following example above, but with a manageable size.<\/p>\n\n<pre><code>&gt;&gt;&gt; np.random.normal(0, 100, size=[5])\narray([-53.12407229,  39.57335574, -98.25406749,  90.81471139, -41.05069646])\n<\/code><\/pre>\n\n<p>There's no way a machine learning algorithm could learn those since those are meant to be inputs to an embedding model and we've got negative values along with floating point values. <\/p>\n\n<p>What is actually wanted is the following code:<\/p>\n\n<pre><code>random_data = np.random.randint(0, 101, size=...)\n<\/code><\/pre>\n\n<p>Checking its output we get<\/p>\n\n<pre><code>&gt;&gt;&gt; np.random.randint(0, 100, size=[5])\narray([27, 47, 33, 12, 24])\n<\/code><\/pre>\n\n<p>Next, the following line is actually creating a subtle problem.<\/p>\n\n<pre><code>def _run_epoch(self, data, session, inputs, rnn_ouputs, loss, train, verbose=10):\n    with session.as_default() as sess:\n        total_steps = sum(1 for x in data_iterator(data, self._batch_size, self._max_steps))\n        train_loss = []\n        for step, (x,y, l) in enumerate(data_iterator(data, self._batch_size, self._max_steps)):\n            print \"step - {0}\".format(step)\n            feed = {\n                self.input_placeholder: x,\n                self.label_placeholder: y,\n                self.sequence_length: l,\n                self._dropout_placeholder: self._dropout,\n            }\n            loss, _= sess.run([loss, train], feed_dict=feed)\n            print \"loss - {0}\".format(loss)\n            train_loss.append(loss)\n            if verbose and step % verbose == 0:\n                sys.stdout.write('\\r{} \/ {} : pp = {}'. format(step, total_steps, np.exp(np.mean(train_loss))))\n                sys.stdout.flush()\n            if verbose:\n                sys.stdout.write('\\r')\n\n        return np.exp(np.mean(train_loss))\n<\/code><\/pre>\n\n<p><code>loss<\/code> is both a parameter argument and a variable, so the first time its run, it will no longer be a tensor and so we can't actually call it in a session. <\/p>\n"}]}
{"QuestionId":41389754,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2016-12-30T02:16:28.770","AcceptedAnswerId":null,"OwnerUserId":958266.0,"Title":"Rename Keras tensor name for loss function","Body":"<p>Are there simple renaming for keras tensor? I needs it for giving tensor more appropriate name in certain context. <\/p>\n\n<p>When i compile model that has multiple output or input, the model needs the name of the input or output tensor. For example, below model needs name of 'advc_out_tn' and 'atoz.output'. <\/p>\n\n<pre><\/pre>\n\n<p>But it gives me error, because They are not their real name. <\/p>\n\n<pre><\/pre>\n\n<p>But the real name of the tensors are somewhat clumsy. Especially if i use given model(in this case, atoz) and access their output tensor by '.out', it gives last layer name('concat_decoded_img'), not the name of the model. <\/p>\n\n<p>How can i solve this?<\/p>\n","answers":[{"AnswerId":"44000996","CreationDate":"2017-05-16T12:06:13.017","ParentId":null,"OwnerUserId":"8019423","Title":null,"Body":"<p>Had a similar problem. For me the following worked to assign a name to an output of the tensor type. I used a dummy Lambda layer (identity function) and assign the new name to this lambda layer (the result is no longer of the tensor type though, but for my purpose this didn't matter).<\/p>\n\n<pre><code>from keras.layers import Lambda\n\nnaming_layer = Lambda(lambda x: x, name='your_name')\nnewly_named_output = naming_layer (tensor_output)\n<\/code><\/pre>\n"}]}
{"QuestionId":41391718,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-12-30T06:33:39.300","AcceptedAnswerId":null,"OwnerUserId":7021851.0,"Title":"Tensorflow's gradient_override_map function","Body":"<p>Can someone explain me  function in TensorFlow?\nI couldn't understand its usage precisely.<\/p>\n\n<p>I see code usage as: <\/p>\n\n<pre><\/pre>\n\n<p>What exactly is happening here? What is ?<\/p>\n","answers":[{"AnswerId":"46865977","CreationDate":"2017-10-21T17:24:13.553","ParentId":null,"OwnerUserId":"8652817","Title":null,"Body":"<p>Both \"Floor\" and \"Identity\" are type strings of operations, the former is corresponding to <em>tf.floor<\/em> while the latter <em>tf.identity<\/em>. <strong>So the function of your code, I guess, is to substitute <em>tf.identity<\/em>'s back-propagated gradient(BPG for short) calculation mechanism for BPG calculation mechanism of <em>tf.floor<\/em> operations within graph G while passing forward output of <em>tf.reduce_mean<\/em><\/strong>. It seems a little weird since in all applications of <code>gradient_override_map<\/code> I've found so far, the key of op_type_map is always identical to the type string of the operation used to produce an output in the context. By this I mean I'm more familiar with scenarios with <code>tf.floor(SomeVals)<\/code> returned, instead of <code>tf.reduce_mean(SomeVals)<\/code>.<\/p>\n\n<p>What <code>gradient_override_map({op_A_type: op_B_type})<\/code> does is to replace op_A's BPG calculation mechanism with op_B's while remaining op_A_type's forward propagation calculation mechanism. A common application of gradient_override_map is shown in lahwran's answer.<\/p>\n\n<pre><code>@tf.RegisterGradient(\"CustomGrad\")\ndef _const_mul_grad(unused_op, grad):\n    return 5.0 * grad\n\ng = tf.get_default_graph()\nwith g.gradient_override_map({\"Identity\": \"CustomGrad\"}):\n    output = tf.identity(input, name=\"Identity\")\n<\/code><\/pre>\n\n<p>by<\/p>\n\n<pre><code>@tf.RegisterGradient(\"CustomGrad\")\ndef _const_mul_grad(unused_op, grad):\n    return 5.0 * grad\n<\/code><\/pre>\n\n<p>the decorator, <code>tf.RegisterGradient(\"CustomGrad\")<\/code> registers the gradient function defined by <code>_const_mul_grad(unused_op, grad)<\/code> for a customized op type -- \"CustomGrad\",<\/p>\n\n<p>while<\/p>\n\n<pre><code>g = tf.get_default_graph()\nwith g.gradient_override_map({\"Identity\": \"CustomGrad\"}):\n    output = tf.identity(input, name=\"Identity\") \n<\/code><\/pre>\n\n<p>assures outputs of all operations (in graph g) with string type \"Identity\" (<em>tf.identity<\/em>) are as they were whereas BPG calculation mechanism of <em>tf.identity<\/em>s replaced by BPG calculation mechanism of operation with string type \"CustomGrad\".<\/p>\n\n<p>P.S. <\/p>\n\n<ol>\n<li><p>The type string of an op corresponds to the <code>OpDef.name<\/code> field for the proto that defines the operation. To find an op's <code>OpDef.name<\/code> , please refer to MingXing's answer under <a href=\"https:\/\/stackoverflow.com\/questions\/47013530\/tensorflow-gradient-override-map-cannot-override-op-tf-stack-s-backward-gradie\">this question<\/a><\/p><\/li>\n<li><p>It is not necessary to declare the name of <em>tf.identity<\/em> operation since the arg 'name' in <em>tf.identity<\/em> is optional.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"44853315","CreationDate":"2017-06-30T19:06:27.840","ParentId":null,"OwnerUserId":"1102705","Title":null,"Body":"<p>As best as I can tell, gradient_override_map allows you to say \"in this context, any time you would use the gradient of X, instead use the gradient of Y\". which means you still need the <em>gradient<\/em> of Y to be the gradient you want to use.<\/p>\n\n<p>This is an example I've seen floating around while looking for how this works:<\/p>\n\n<pre><code>@tf.RegisterGradient(\"CustomGrad\")\ndef _const_mul_grad(unused_op, grad):\n    return 5.0 * grad\n\ng = tf.get_default_graph()\nwith g.gradient_override_map({\"Identity\": \"CustomGrad\"}):\n    output = tf.identity(input, name=\"Identity\")\n<\/code><\/pre>\n\n<p>cite: <a href=\"https:\/\/stackoverflow.com\/a\/43948872\/1102705\">https:\/\/stackoverflow.com\/a\/43948872\/1102705<\/a><\/p>\n\n<p><code>RegisterGradient()<\/code> allows you to register the gradient of a new op you're defining, thereby allowing you to have an op that has the gradient you wanted, and then you can use that op in the gradient override map. It's kind of clunky - you're defining an op with no forward pass. <\/p>\n\n<p>Something I'm not clear on is whether the name=\"Identity\" is actually necessary.<\/p>\n"}]}
{"QuestionId":41392201,"AnswerCount":1,"Tags":"<keras><recurrent-neural-network><panel-data><mxnet>","CreationDate":"2016-12-30T07:11:19.847","AcceptedAnswerId":49100075.0,"OwnerUserId":1190427.0,"Title":"Recurrent Neural Networks for Panel Data","Body":"<p>There are 2 parts to this question. Suppose we are looking at sales S of a product across $> 1000$ stores where a it sells. For each of these 1000 stores we have 24 months recorded data. <\/p>\n\n<ol>\n<li>We want to be able to predict S_t &lt;- f(S_{t-1}). We could build a RNN for each of the store time series, calculate test RMSE and then take an average after taking care of normalizing values etc. But the problem is there are very few samples per time series. If we were to segment stores into groups (by say Dynamic Time Warping) then could we create a monologue of text sentiment mining where like in text two sentences are separated by a dot here we would have two time series separated by a special symbol (let's say). In that case, we would generate a RNN model on <\/li>\n<\/ol>\n\n<p>Train_1 | Train_2 |...|Train_t  <\/p>\n\n<p>data and predict on <\/p>\n\n<p>Test_1 | Test_2 |...|Test_t<\/p>\n\n<ol start=\"2\">\n<li>After this, we would like to set it up as a panel data problem where S_t &lt;- f(x_{t1},x_{t2},...,x_{tn}). In that case should I build a separate neural network for each t and then connect the hidden layers from t -> t+1 -> t+2 .... <\/li>\n<\/ol>\n\n<p>How should I implement these through packages like Keras\/Theano\/Mxnet etc.? Any help would be great! <\/p>\n","answers":[{"AnswerId":"49100075","CreationDate":"2018-03-04T20:28:01.197","ParentId":null,"OwnerUserId":"9015998","Title":null,"Body":"<p>For your first question, implementing this in MXNet Gluon is very straight forward. You can formulate your problem as an auto regression problem so that it doesn't have any dependency on the sequence length during or you can formulate it as single prediction and require a specific sequence length for S in order to predict S_t. Either way, <a href=\"http:\/\/gluon.mxnet.io\/chapter05_recurrent-neural-networks\/rnns-gluon.html\" rel=\"nofollow noreferrer\">this gluon tutorial<\/a> can help get you started.<\/p>\n"}]}
{"QuestionId":41393096,"AnswerCount":3,"Tags":"<theano><keras>","CreationDate":"2016-12-30T08:32:02.537","AcceptedAnswerId":null,"OwnerUserId":6947837.0,"Title":"what is the meaning of border_mode in keras?","Body":"<p>I'm confused what to use, valid,same or full. I also don't know what does it do. I can't find it in the Docs. And  for  layer does not make sense to me. (It does make sense to me for the Convolution layers though).<\/p>\n","answers":[{"AnswerId":"56494392","CreationDate":"2019-06-07T12:30:46.237","ParentId":null,"OwnerUserId":"2459114","Title":null,"Body":"<p>Border_mode full, same, and valid are explained very well above.<\/p>\n\n<p>Yes you are right <code>MaxPooling<\/code> is used to reduce the channels if you use <code>border_mode = full<\/code> or <code>border_mode = same<\/code> it doesn't make any sense.<\/p>\n"},{"AnswerId":"41421390","CreationDate":"2017-01-02T04:34:50.143","ParentId":null,"OwnerUserId":"7363938","Title":null,"Body":"<p>When you use a two-dimentional image with <code>m rows and n cols<\/code>, and a <code>a*b size kernel<\/code> to convolute the input image, this is what happens:<br><\/p>\n\n<p>If border_mode is <em>'full'<\/em>, returns a <code>(m+a-1)x(n+b-1)<\/code> image;<br>\n  if border_mode is <em>'same'<\/em>, returns the same dimention as the input image;<br>if border_mode is <em>'valid'<\/em>,returns a <code>(m-a+1)x(n-b+1)<\/code> image.\n    for example, <\/p>\n\n<p><strong>Input<\/strong>:\nIn following 4x4 image <\/p>\n\n<p><code>A = [12 13 14 15;1 2 3 4;16 17 18 19;5 6 7 8], and a 3x3 kernel B = [1 2 3;4 5 6;7 8 9]<\/code>,<\/p>\n\n<p>if border_mode is 'full', then returns a <code>6x6<\/code> matrix;<br>\nif border_mode is 'same', then returns a <code>4x4<\/code> matrix;<br>\nif border_mode is 'valid', then returns a <code>2x2<\/code> matrix.<br><\/p>\n\n<p>you can also use function conv2(A,B,border_mode) in MATLAB to test the output matrix.<\/p>\n\n<p>Hope this answer could help.<\/p>\n"},{"AnswerId":"58870913","CreationDate":"2019-11-15T05:36:07.463","ParentId":null,"OwnerUserId":"1009508","Title":null,"Body":"<p>This is for Keras 2+ as they replaced the Border_mode with padding\nAnd it can be used for down and upsampling in a network. <\/p>\n"}]}
{"QuestionId":41394468,"AnswerCount":1,"Tags":"<python><shell><ubuntu><keras>","CreationDate":"2016-12-30T10:17:42.263","AcceptedAnswerId":null,"OwnerUserId":2479950.0,"Title":"Long-running batch process being terminated prematurely in Ubuntu","Body":"<p>I'm trying to batch some python scripts that train deep neural networks using Keras and I'm encountering the problem that when the first python script finishes, the whole batch is stopped.<\/p>\n\n<p>I tried writing a python wrapper, but still nothing changes. I wonder if this problem is OS related or it has to do with Keras framework or python.<\/p>\n\n<p>The first attempt was to write a shell script:<\/p>\n\n<pre><\/pre>\n\n<p>However, after running  once, the whole batch stops.<\/p>\n\n<p>I tried using Python:<\/p>\n\n<pre><\/pre>\n\n<p>But I still get the same result.<\/p>\n\n<p>Has anyone encountered this problem before? If not I would welcome ideas to try and figure out what's going on.<\/p>\n\n<p>Additionally, I'm running this process in a remote machine via SSH, in the following way:<\/p>\n\n<pre><\/pre>\n\n<p>Or in the case of python:<\/p>\n\n<pre><\/pre>\n\n<p>Cheers<\/p>\n","answers":[{"AnswerId":"41394723","CreationDate":"2016-12-30T10:33:31.247","ParentId":null,"OwnerUserId":"1763614","Title":null,"Body":"<p>This might happen if some processes still have standard input or output linked to the tty, or if some sub-processes are not properly disowned.<\/p>\n\n<p>A possible workaround is to use a terminal multiplexer such as <a href=\"https:\/\/www.gnu.org\/software\/screen\/\" rel=\"nofollow noreferrer\">screen<\/a> or <a href=\"http:\/\/tmux.github.io\" rel=\"nofollow noreferrer\">tmux<\/a> to solve the problem of leaving a background process running after an SSH logout. It is much easier to use than disowning \/ nohuping a process.<\/p>\n\n<p>Typically you will run <code>screen<\/code> or <code>tmux<\/code>, and you will get a new terminal session. You will be able to 'detach' from that session and leave it running. After you log out and log in again, you will be able to 'reattach' to the running sessions and see the output of the process as if you never logged out.<\/p>\n"}]}
{"QuestionId":41396196,"AnswerCount":0,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2016-12-30T12:21:40.133","AcceptedAnswerId":null,"OwnerUserId":611611.0,"Title":"exporting pb graph for tensorflow serve","Body":"<p>I'm trying to export my retrained inception graph (pb file) for tensorflow serve since I was not able to find any snippets for exporting pb file I had to create my own but obviously I'm doing something wrong since I'm getting  \"No variables to save\" error, I would appreciate any help<\/p>\n\n<p>Update: after researching more I think I need meta graph to supply variables but retrain.py doesn't give me meta graph, any ideas ?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41396767,"AnswerCount":1,"Tags":"<python><neural-network><keras>","CreationDate":"2016-12-30T13:07:44.160","AcceptedAnswerId":41401263.0,"OwnerUserId":5726763.0,"Title":"Python Keras ValueError NeuralNetwork FlappyBird Game","Body":"<p>I cloned the following  project (<a href=\"https:\/\/github.com\/yanpanlau\/Keras-FlappyBird\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yanpanlau\/Keras-FlappyBird<\/a>) involving  locally in my pc. I installed the dependencies, but when I run it (), I received the following error:<\/p>\n\n<pre><\/pre>\n\n<p>By default, Keras uses  as backend, so maybe the order of paramters can imply this error, but I couldn't figure it out.<\/p>\n","answers":[{"AnswerId":"41401263","CreationDate":"2016-12-30T18:54:56.960","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>The saved model is trained with <code>\"image_dim_ordering\": \"th\"<\/code>. I am pretty sure on your system currently its configured to <code>tf<\/code>. Just open <code>~\/.keras\/keras.json<\/code> and change it to <code>th<\/code> and the model should load properly.<\/p>\n"}]}
{"QuestionId":41396970,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2016-12-30T13:20:08.497","AcceptedAnswerId":41399856.0,"OwnerUserId":859559.0,"Title":"Tensorflow 3 channel order of color inputs","Body":"<p>I'm using tensor flow to process color images with a convolutional neural network. A code snippet is below.<\/p>\n\n<p>My code runs so I think I got the number of channels right. My question is, how do I correctly order the rgb data? Is it in the form rgbrgbrgb or would it be rrrgggbbb? Presently I am using the latter. Thanks. Any help would be appreciated.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41399856","CreationDate":"2016-12-30T17:01:52.177","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> With your current program, the in-memory layout of the data should be should be R-G-B-R-G-B-R-G-B-R-G-B...<\/p>\n\n<p>I assume from this line that you are passing in RGB images with 28x28 pixels:<\/p>\n\n<pre><code>self.x_image = tf.reshape(self.c_x, [-1, 28, 28, 3])\n<\/code><\/pre>\n\n<p>We can call the dimensions of <code>self.x_image<\/code> are \"batch\", \"height\", \"width\", and \"channel\". This matches the default data format for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/convolution#conv2d\" rel=\"nofollow noreferrer\"><code>tf.nn.conv_2d()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/pooling#max_pool\" rel=\"nofollow noreferrer\"><code>tf.nn.max_pool()<\/code><\/a>.<\/p>\n\n<p>In TensorFlow, the in-memory representation of a tensor is <a href=\"https:\/\/en.wikipedia.org\/wiki\/Row-_and_column-major_order\" rel=\"nofollow noreferrer\">row-major order<\/a> (or \"C\" ordering, because that is the representation of arrays in the C programming language). Essentially this means that the rightmost dimension is the fastest changing, and the elements of the tensor are packed together in memory in the following order (where <code>?<\/code> stands for the unknown batch size, minus 1):<\/p>\n\n<pre><code>[0,  0,  0,  0]\n[0,  0,  0,  1]\n[0,  0,  0,  2]\n[0,  0,  1,  0]\n...\n[?, 27, 27,  1]\n[?, 27, 27,  2]\n<\/code><\/pre>\n\n<p>Therefore your program probably isn't interpreting the image data correctly. There are at least two options:<\/p>\n\n<ol>\n<li><p>Reshape your data to match its true order (\"batch\", \"channels\", \"height\", \"width\"):<\/p>\n\n<pre><code>self.x_image = tf.reshape(self.c_x, [-1, 3, 28, 28])\n<\/code><\/pre>\n\n<p>In fact, this format is sometimes more efficient for convolutions. You can instruct <code>tf.nn.conv2d()<\/code> and <code>tf.nn.max_pool()<\/code> to use it without transposing by passing the optional argument <code>data_format=\"NCHW\"<\/code>, but you will also need to change the shape of your bias variables to match.<\/p><\/li>\n<li><p>Transpose your image data to match the result of your program using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#transpose\" rel=\"nofollow noreferrer\"><code>tf.transpose()<\/code><\/a>:<\/p>\n\n<pre><code>self.x_image = tf.transpose(tf.reshape(self.c_x, [-1, 3, 28, 28]), [0, 2, 3, 1])\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":41397071,"AnswerCount":2,"Tags":"<deep-learning><caffe><pycaffe>","CreationDate":"2016-12-30T13:27:32.380","AcceptedAnswerId":41743008.0,"OwnerUserId":6786718.0,"Title":"He-initialization for Caffe?","Body":"<p>I know that I can either initialize the weights of a convolutional network in caffe with  or . Sadly I can not find any other initialization methods implemented in caffe. Especially, I would like to use the PReLU initialization as suggested in this paper: <a href=\"https:\/\/arxiv.org\/pdf\/1502.01852.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1502.01852.pdf<\/a> <\/p>\n\n<p>Is this shomewho implemented in caffe and pycaffe?<\/p>\n","answers":[{"AnswerId":"41401477","CreationDate":"2016-12-30T19:11:25.527","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>He et al use an initialization scheme which samples each weight w of a given layer independantly identically distributed (iid):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8TLcb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8TLcb.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However, you might know that Var(aX) = a^2 Var(X). Hence:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AORum.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AORum.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So you can simply initialize the weights with <code>gaussian<\/code> and then multiply them with <code>2 \/ n_l<\/code>.<\/p>\n"},{"AnswerId":"41743008","CreationDate":"2017-01-19T13:30:20.610","ParentId":null,"OwnerUserId":"7441262","Title":null,"Body":"<p>The <strong>MSRA<\/strong> (for Microsoft Research Asia) filler was implemented in Caffe a few months after the paper was released. You can use it with the <code>msra<\/code> parameter.<\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1MSRAFiller.html#details\" rel=\"nofollow noreferrer\">See the documentation for more information.<\/a><\/p>\n"}]}
{"QuestionId":41397297,"AnswerCount":0,"Tags":"<python><model><save><tensorflow><restore>","CreationDate":"2016-12-30T13:44:36.827","AcceptedAnswerId":null,"OwnerUserId":4172154.0,"Title":"how does tensorflow save and restore model?","Body":"<p>First, I know how to use tensorflow to save and restore a model, just like the tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, it works well. Then I wrote two model in single file  just like this:<\/p>\n\n<pre><\/pre>\n\n<p>It works well too! Then I wrote a  file in which I wrapped  and  in two models, which are called  and , just like this:<\/p>\n\n<pre><\/pre>\n\n<p>Then I directly use model saved by , it works well again. \n<a href=\"https:\/\/i.stack.imgur.com\/OTkLU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OTkLU.png\" alt=\"1\"><\/a>\nNow I will throw my question, I split  and  in two files:<\/p>\n\n<p>ModelV.py<\/p>\n\n<pre><\/pre>\n\n<p>ModelP.py<\/p>\n\n<pre><\/pre>\n\n<p>Then i run following, first ModelV.py and then ModelP.py,then model.py; \n<a href=\"https:\/\/i.stack.imgur.com\/Uqy7M.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Uqy7M.png\" alt=\"2\"><\/a>\nsee what happend, Error happend! Following is my error:<\/p>\n\n<pre><\/pre>\n\n<p>This is just an example which I use to test because I have another project in which I wrote two models in one file, and each model is an object, each provides a predict and train interface, same error happened. It sames tensorflow just search first model's variables in another model's checkpoint file. It mix up, but they are in two models! Thanks for giving me some help!<\/p>\n","answers":[]}
{"QuestionId":41397526,"AnswerCount":1,"Tags":"<image><tensorflow><classification><pixel><tflearn>","CreationDate":"2016-12-30T14:03:38.387","AcceptedAnswerId":41398817.0,"OwnerUserId":2148425.0,"Title":"Tensorflow architecture for pixel based image classification","Body":"<p>I have an RGBI image including 4 bands and want to be able to classify image pixels using tensorflow and deep learning into two classes.\nIn the training data, each pixel is regarded an observation with 4 values\/features as image intensity. I used the following function to create the network <\/p>\n\n<pre><\/pre>\n\n<p>but I got the following error <\/p>\n\n<blockquote>\n  <p>ValueError: Cannot feed value of shape (64, 4) for Tensor 'InputData\/X:0', which has shape '(?, 1, 4)'<\/p>\n<\/blockquote>\n\n<p>I don't know where the problem is. \nIs there any benefit in using Deep Neural Networks versus Random Forest for pixel-based classification?\nand if yes, how can I do this using the above function.<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"41398817","CreationDate":"2016-12-30T15:41:12.247","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>You need to expand the dimensions of variable <code>X<\/code>, to account for the time step in the LSTM. Instead of directly passing <code>X<\/code>, use <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.expand_dims.html\" rel=\"nofollow noreferrer\"><code>np.expand_dims<\/code><\/a> like -<\/p>\n\n<pre><code>X = np.expand_dims(X, axis=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41398541,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><tensorflow>","CreationDate":"2016-12-30T15:21:06.813","AcceptedAnswerId":41412408.0,"OwnerUserId":7123245.0,"Title":"Tensorflow logistic regression different output","Body":"<p>hello I am trying to preform logistic regression using tensor-flow ( sorry if my code looks dumb) and I have written the cost function once in numpy and once in tensor-flow , I am getting different results for the same starting weights , could some one help me?<\/p>\n\n<pre><\/pre>\n\n<p>Cost sample of the cost functions:<\/p>\n\n<p>('cost in numpy', array([ 2.37780175])) ('cost in tensorflow', 0.073667422)<\/p>\n","answers":[{"AnswerId":"41412408","CreationDate":"2016-12-31T23:48:02.037","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>You are initializing the weights to random values in this line:<\/p>\n\n<pre><code>session.run(tf.initialize_all_variables())\n<\/code><\/pre>\n\n<p>After that line you can set the values with something like this:<\/p>\n\n<pre><code>session.run(tf.assign(b,tf.ones((1,2))))\n<\/code><\/pre>\n"}]}
{"QuestionId":41398547,"AnswerCount":0,"Tags":"<tensorflow><neural-network>","CreationDate":"2016-12-30T15:21:39.463","AcceptedAnswerId":null,"OwnerUserId":6201311.0,"Title":"Non-gradient methods in TensorFlow","Body":"<p>Good day everyone!\nI have an idea (prototyped on Python) of non-gradient training method for specific neural networks. I want to re-write it on Tensorflow to make computations on GPU. But all documentation of TF is about gradient computation, propagation and applying. Is it possible to run any non-gradient training method?<\/p>\n","answers":[]}
{"QuestionId":41399382,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2016-12-30T16:24:25.940","AcceptedAnswerId":null,"OwnerUserId":2977382.0,"Title":"LSTM with variable sequences & return full sequences","Body":"<p>How can I set up a keras model such that the final LSTM layer outputs a prediction for each time step while having variable sequence lengths as input?<\/p>\n\n<p>I'd then like to provide labels for each of the timesteps after a dense layer with linear activation.<\/p>\n\n<p>When I try to add a reshape or a dense layer to the LSTM model that is returning the full sequence and has a masking layer to take care of variable sequence lengths, it says:<\/p>\n\n<blockquote>\n  <p>The reshape and the dense layers do not support masking.<\/p>\n<\/blockquote>\n\n<p>Would this be possible to do?<\/p>\n","answers":[{"AnswerId":"41511331","CreationDate":"2017-01-06T17:46:13.180","ParentId":null,"OwnerUserId":"3697202","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/keras.io\/layers\/wrappers\/#timedistributed\" rel=\"nofollow noreferrer\"><code>TimeDistributed<\/code><\/a> layer wrapper for this. This applies the layer you want to each timestep. In your case, you could also just use <code>TimeDistributedDense<\/code>.<\/p>\n"}]}
{"QuestionId":41399481,"AnswerCount":3,"Tags":"<python><tensorflow><one-hot-encoding>","CreationDate":"2016-12-30T16:30:18.190","AcceptedAnswerId":41399559.0,"OwnerUserId":6112394.0,"Title":"How do you decode one-hot labels in Tensorflow?","Body":"<p>Been looking, but can't seem to find any examples of how to decode or convert back to a single integer from a one-hot value in TensorFlow.<\/p>\n\n<p>I used  and was able to train my model but am a bit confused on how to make sense of the label after my classification. My data is being fed in via a  file that I created. I thought about storing a text label in the file but wasn't able to get it to work. It appeared as if  couldn't store text string or maybe I was mistaken.<\/p>\n","answers":[{"AnswerId":"51947542","CreationDate":"2018-08-21T11:19:37.650","ParentId":null,"OwnerUserId":"7017416","Title":null,"Body":"<pre><code>data = np.array([1, 5, 3, 8])\nprint(data)\n\n\ndef encode(data):\n    print('Shape of data (BEFORE encode): %s' % str(data.shape))\n    encoded = to_categorical(data)\n    print('Shape of data (AFTER  encode): %s\\n' % str(encoded.shape))\n    return encoded\n\n\nencoded_data = encode(data)\nprint(encoded_data)\n\ndef decode(datum):\n    return np.argmax(datum)\n\ndecoded_Y = []\nprint(\"****************************************\")\nfor i in range(encoded_data.shape[0]):\n    datum = encoded_data[i]\n    print('index: %d' % i)\n    print('encoded datum: %s' % datum)\n    decoded_datum = decode(encoded_data[i])\n    print('decoded datum: %s' % decoded_datum)\n    decoded_Y.append(decoded_datum)\n\n\nprint(\"****************************************\")\n\nprint(decoded_Y)\n<\/code><\/pre>\n"},{"AnswerId":"41399559","CreationDate":"2016-12-30T16:37:44.117","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>You can find out the index of the largest element in the matrix using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/sequence_comparison_and_indexing#argmax\" rel=\"noreferrer\"><code>tf.argmax<\/code><\/a>. Since your one hot vector will be one dimensional and will have just one <code>1<\/code> and other <code>0<\/code>s, This will work assuming you are dealing with a single vector.<\/p>\n\n<pre><code>index = tf.argmax(one_hot_vector, axis=0)\n<\/code><\/pre>\n\n<p>For the more standard matrix of <code>batch_size * num_classes<\/code>, use <code>axis=1<\/code> to get a result of size <code>batch_size * 1<\/code>.<\/p>\n"},{"AnswerId":"41399582","CreationDate":"2016-12-30T16:39:28.137","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Since a one-hot encoding is typically just a matrix with <code>batch_size<\/code> rows and <code>num_classes<\/code> columns, and each row is all zero with a single non-zero corresponding to the chosen class, you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/sequence_comparison_and_indexing#argmax\" rel=\"noreferrer\"><code>tf.argmax()<\/code><\/a> to recover a vector of integer labels:<\/p>\n\n<pre><code>BATCH_SIZE = 3\nNUM_CLASSES = 4\none_hot_encoded = tf.constant([[0, 1, 0, 0],\n                               [1, 0, 0, 0],\n                               [0, 0, 0, 1]])\n\n# Compute the argmax across the columns.\ndecoded = tf.argmax(one_hot_encoded, axis=1)\n\n# ...\nprint sess.run(decoded)  # ==&gt; array([1, 0, 3])\n<\/code><\/pre>\n"}]}
{"QuestionId":41399528,"AnswerCount":1,"Tags":"<resize><theano>","CreationDate":"2016-12-30T16:35:07.373","AcceptedAnswerId":41633399.0,"OwnerUserId":852592.0,"Title":"How to do fast nearest-neighbor upsampling in Theano?","Body":"<p>Currently I am using this code, but it is really slow:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any way to speed it up?<\/p>\n","answers":[{"AnswerId":"41633399","CreationDate":"2017-01-13T11:17:31.007","ParentId":null,"OwnerUserId":"1950580","Title":null,"Body":"<p>This code is more vectorized. I'm not sure how fast this is due to reshape.<\/p>\n\n<pre><code>def resizeNN(inp, scale):\n    inp_shp = T.shape(inp)\n    return T.tile(inp.dimshuffle(0,1,2,3,'x','x'),\n        (scale,scale)).transpose(\n            0,1,2,4,3,5).reshape(\n                inp_shp[0], inp_shp[1],\n                inp_shp[2]*scale, inp_shp[3]*scale)\n<\/code><\/pre>\n"}]}
{"QuestionId":41399752,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network><classification><kaggle>","CreationDate":"2016-12-30T16:53:34.840","AcceptedAnswerId":null,"OwnerUserId":1764528.0,"Title":"Really bad Accuracy with Neural Network","Body":"<p>I was trying to solve the <a href=\"https:\/\/www.kaggle.com\/c\/dogs-vs-cats-redux-kernels-edition\" rel=\"nofollow noreferrer\">Dogs vs. Cats Redux: Kernels Edition<\/a> problem on Kaggle. It is a simple image classification problem. However, I am doing worse than a random predictor with a score of 17+. Does anyone know why this might be?<\/p>\n\n<p><strong>Neural Network Model<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The the output has no activation function.<\/p>\n\n<p><strong>Prediction, Optimizer and Loss Function<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The images are converted into an numpy array of size <strong>128<\/strong>x<strong>128<\/strong>x<strong>3<\/strong> and fed into the neural network with a batch size of 64.<\/p>\n\n<p><a href=\"https:\/\/github.com\/DollarAkshay\/Python-Programs\/blob\/master\/Kaggle\/Kaggle_Dogs_vs_Cats.py\" rel=\"nofollow noreferrer\">Full Code Here<\/a> <\/p>\n\n<p><strong>Edit :<\/strong>  Ran the same code for 200 epochs. No improvement. I did Slightly worse. <\/p>\n","answers":[{"AnswerId":"41409054","CreationDate":"2016-12-31T14:20:03.780","ParentId":null,"OwnerUserId":"7360707","Title":null,"Body":"<p>Improving accuracy is an art than one task solution, you can try some of these methods: <\/p>\n\n<ul>\n<li>try different gradient optimization, SGD, momentum, nestrov, adap, ...<\/li>\n<li>Try adaptive learning rate <\/li>\n<li>improve regularization methods L1,L2,dropout, drop connect, ...<\/li>\n<li>Augment your training data (have more data). <\/li>\n<li>change your network hyper parameters<\/li>\n<li>finally if nothing helped change the network structure. <\/li>\n<\/ul>\n"},{"AnswerId":"41602910","CreationDate":"2017-01-11T23:56:30.357","ParentId":null,"OwnerUserId":"2759654","Title":null,"Body":"<p>This is more of a comment but not enough privilege points for that:<\/p>\n\n<p>Did you normalize your data (i.e. divide the pixels values by 255)? I can't you see doing that in the script.<\/p>\n\n<p>When you get terrible results like 17 logloss that means your model is always predicting one class with 100% confidence. Usually in this case it's not the architecture or learning rate or number of epochs but rather some silly mistake like forgetting to normalize or mixing up your labels. For this particular problem and given your architecture you should see an accuracy of about 80% and 0.4 logloss within 40 number of epochs. No need for thousands of epochs :)<\/p>\n"}]}
{"QuestionId":41400391,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2016-12-30T17:44:41.393","AcceptedAnswerId":null,"OwnerUserId":163173.0,"Title":"Tensorflow: saving and resoring session - multiple variables","Body":"<p>Given the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I get an error at the line  saying .<\/p>\n\n<p>How can I get around this? Is there anyway to <em>overwrite<\/em> the compute graph when importing a metagraph?<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I have arrived at the following code:<\/p>\n\n<pre><\/pre>\n\n<p>This prints the correct value for , but when I come to save the graph a second time, I get the same original error: <\/p>\n","answers":[{"AnswerId":"41402945","CreationDate":"2016-12-30T21:32:52.357","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>In this case you are loading the variable in the default graph where you already have the variable defined. Hence\nyou will need to reset the TensorFlow graph before importing it.<\/p>\n\n<p>Do this using <code>tf.reset_default_graph()<\/code>. before you import. Have a look at the \"Import within the default graph\" section under <a href=\"https:\/\/www.tensorflow.org\/how_tos\/meta_graph\/\" rel=\"nofollow noreferrer\">Exporting and Importing a MetaGraph<\/a>.<\/p>\n\n<p>Of course, you will have to redefine the variable <code>var<\/code> using <code>tf.get_variable()<\/code>. Try out this code,<\/p>\n\n<pre><code>import tensorflow as tf\n\nwith tf.Session() as sess:\n    var = tf.Variable(42, name='var')\n    sess.run(tf.global_variables_initializer())\n    tf.train.export_meta_graph('file.meta')\ntf.reset_default_graph()\nwith tf.Session() as sess:\n    saver = tf.train.import_meta_graph('file.meta')\n    var = tf.global_variables()[0]\n    sess.run(tf.initialize_all_variables())\n    print sess.run(var)\n<\/code><\/pre>\n\n<hr>\n\n<p>The reason your intermediate code wasn't working was that <code>tf.get_variable()<\/code> was creating a new variable which was being randomly initialized. Make sure you first do <code>tf.get_variable_scope().reuse_variables()<\/code> first.\nHave a look at <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variable_scope\/#understanding_tfget_variable\" rel=\"nofollow noreferrer\">Understanding <code>tf.get_variable()<\/code><\/a>.<\/p>\n\n<p>Unfortunately, your variable created using <code>tf.Variable()<\/code> cannot be reused with <code>tf.get_variable()<\/code> directly. Have a look at this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1325#issuecomment-190428758\" rel=\"nofollow noreferrer\">comment<\/a> and this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1325#issuecomment-259040409\" rel=\"nofollow noreferrer\">comment<\/a> to know exactly why. Hence you will need to use <code>tf.get_variable()<\/code> to <em>create<\/em> the variable if you wish to reuse it in the future.<\/p>\n"}]}
{"QuestionId":41400873,"AnswerCount":4,"Tags":"<android><tensorflow>","CreationDate":"2016-12-30T18:24:08.990","AcceptedAnswerId":null,"OwnerUserId":3836902.0,"Title":"How to reduce tensorflow size for android?","Body":"<p>I am trying to use TensorFlow in my android app. But the .so file is little bit larger than needed, is there a way to reduce libtensorflow_inference.so size?<\/p>\n\n<p>Link <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/android\/cmake\" rel=\"nofollow noreferrer\">TensorFlow-Android-Inference<\/a><\/p>\n","answers":[{"AnswerId":"41491019","CreationDate":"2017-01-05T17:20:50.243","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>In addition to selective registration, you can try passing <code>-c opt<\/code> to bazel. For me this takes the uncompressed .so size from 24mb -> 13mb, and compressed from 5.5mb -> 4.4mb.<\/p>\n"},{"AnswerId":"41448478","CreationDate":"2017-01-03T16:53:35.840","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>There's no good guide to doing this yet, but you can start by removing unneeded ops using the print_selective_registration_header script:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/print_selective_registration_header.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/print_selective_registration_header.py<\/a><\/p>\n\n<p>This needs to be documented more extensively.<\/p>\n"},{"AnswerId":"42615201","CreationDate":"2017-03-05T23:14:29.413","ParentId":null,"OwnerUserId":"112705","Title":null,"Body":"<p>I don't know how to reduce the size of the .so files, but there is a way to save huge amounts of space by making it possible to compress the TensorFlow model.  See <a href=\"https:\/\/petewarden.com\/2016\/09\/27\/tensorflow-for-mobile-poets\/\" rel=\"nofollow noreferrer\">this excellent blog post by Pete Warden<\/a> for more details.<\/p>\n\n<p>In summary, you can use <code>optimize_for_inference<\/code> and <code>quantize_graph<\/code> to make it possible to compress a 87MB model to 24MB (you'll see this space saving when building the APK).<\/p>\n"},{"AnswerId":"41402171","CreationDate":"2016-12-30T20:15:38.093","ParentId":null,"OwnerUserId":"3623636","Title":null,"Body":"<p>You can try to divide your apk to reduce apk size. <a href=\"https:\/\/developer.android.com\/google\/play\/publishing\/multiple-apks.html\" rel=\"nofollow noreferrer\">Check this<\/a>.<\/p>\n"}]}
{"QuestionId":41400931,"AnswerCount":1,"Tags":"<tensorflow><bazel>","CreationDate":"2016-12-30T18:28:04.797","AcceptedAnswerId":null,"OwnerUserId":7358090.0,"Title":"tensorflow model im2txt: error when using Bazel build \"couldn't determine target from filename\"","Body":"<p>In trying to use the im2txt Tensorflow model, I get this error:<\/p>\n\n<p>couldn't determine target from filename 'im2txt\/download_and_preprocess_mscoco'<\/p>\n\n<p>Here is the link to the Github repo where the model is located:\n<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/im2txt\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/im2txt<\/a><\/p>\n\n<p>I'm following the directions from the tutorial\/readme.  Any help would be much appreciated. <\/p>\n","answers":[{"AnswerId":"41447516","CreationDate":"2017-01-03T16:03:39.500","ParentId":null,"OwnerUserId":"7358090","Title":null,"Body":"<p>I figured it out. It had to do with the path to the directory with the IM2TXT models.  As long as you are running the command from the correct directory that you downloaded the models to, the error goes away.<\/p>\n"}]}
{"QuestionId":41402332,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2016-12-30T20:30:56.767","AcceptedAnswerId":41465631.0,"OwnerUserId":2059708.0,"Title":"Tensorflow create a tfrecords file from csv","Body":"<p>I am trying to write a csv file (all columns are floats) to a tfrecords file then read them back out. All the examples I have seen pack the csv columns then feed it to sess.run() directly but I can't figure out how to write the feature columns and label column to a tfrecord instead. How could I do this?<\/p>\n","answers":[{"AnswerId":"48560347","CreationDate":"2018-02-01T10:21:29.883","ParentId":null,"OwnerUserId":"8732893","Title":null,"Body":"<pre><code>def convert_to():\nfilename = os.path.join(wdir, 'ml-100k' + '.tfrecords')\nprint('Writing', filename)\nwith tf.python_io.TFRecordWriter(filename) as writer:\n    with open(\"\/Users\/shishir\/Documents\/botconnect_Playground\/tfRecords\/ml-100k.train.rating\", \"r\") as f:\n        line = f.readline()\n        while line != None and line != \"\":\n            arr = line.split(\"\\t\")\n            u, i, l  = int(arr[0]), int(arr[1]), int(arr[2])\n            u_arr = np.reshape(u,[1]).astype('int64')\n            i_arr = np.reshape(i,[1]).astype('int64')\n            l_arr = np.reshape(l,[1]).astype('int64')\n            example = tf.train.Example()\n            example.features.feature[\"user\"].int64_list.value.extend(u_arr)\n            example.features.feature[\"item\"].int64_list.value.extend(i_arr)\n            example.features.feature[\"label\"].int64_list.value.append(int(l_arr))\n            writer.write(example.SerializeToString())\n            line = f.readline()\n<\/code><\/pre>\n\n<p>So that is my Solution and it works! Hope this helps<\/p>\n\n<p>Cheers.<\/p>\n"},{"AnswerId":"41465631","CreationDate":"2017-01-04T13:59:44.543","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p>You will need a separate script to convert your csv file to TFRecords.<\/p>\n\n<p>Imagine you have a CSV with the following header:<\/p>\n\n<pre><code>feature_1, feature_2, ..., feature_n, label\n<\/code><\/pre>\n\n<p>You need to read your CSV with something like <code>pandas<\/code>, construct <code>tf.train.Example<\/code> manually and then write it to file with <code>TFRecordWriter<\/code><\/p>\n\n<pre><code>csv = pandas.read_csv(\"your.csv\").values\nwith tf.python_io.TFRecordWriter(\"csv.tfrecords\") as writer:\n    for row in csv:\n        features, label = row[:-1], row[-1]\n        example = tf.train.Example()\n        example.features.feature[\"features\"].float_list.value.extend(features)\n        example.features.feature[\"label\"].int64_list.value.append(label)\n        writer.write(example.SerializeToString())\n<\/code><\/pre>\n"},{"AnswerId":"55947944","CreationDate":"2019-05-02T07:55:44.503","ParentId":null,"OwnerUserId":"8932540","Title":null,"Body":"<p>The above solution not worked in my case.Another way to read csv file and create tfRecord is shown below:<\/p>\n\n<p>The feature set column names are :Sl.No:,Time,Height,   Width,Mean,Std, Variance, Non-homogeneity, PixelCount, contourCount, Class.<\/p>\n\n<p>Sample features that we get from dataset.csv:<\/p>\n\n<p>Features= [5, 'D', 268, 497, 13.706, 863.4939, 29.385, 0.0427, 39675, 10]<\/p>\n\n<p>label : medium<\/p>\n\n<pre><code>import pandas as pd\nimport tensorflow as tf\n\ndef create_tf_example(features, label):\n\n    tf_example = tf.train.Example(features=tf.train.Features(feature={\n        'Time': tf.train.Feature(bytes_list=tf.train.BytesList(value=[features[1].encode('utf-8')])),\n        'Height':tf.train.Feature(int64_list=tf.train.Int64List(value=[features[2]])),\n        'Width':tf.train.Feature(int64_list=tf.train.Int64List(value=[features[3]])),\n        'Mean':tf.train.Feature(float_list=tf.train.FloatList(value=[features[4]])),\n        'Std':tf.train.Feature(float_list=tf.train.FloatList(value=[features[5]])),\n        'Variance':tf.train.Feature(float_list=tf.train.FloatList(value=[features[6]])),\n        'Non-homogeneity':tf.train.Feature(float_list=tf.train.FloatList(value=[features[7]])),\n        'PixelCount':tf.train.Feature(int64_list=tf.train.Int64List(value=[features[8]])),\n        'contourCount':tf.train.Feature(int64_list=tf.train.Int64List(value=[features[9]])),\n        'Class':tf.train.Feature(bytes_list=tf.train.BytesList(value=[label.encode('utf-8')])),\n    }))\n    return tf_example\n\ncsv = pd.read_csv(\"dataset.csv\").values\nwith tf.python_io.TFRecordWriter(\"dataset.tfrecords\") as writer:\n  for row in csv:\n     features, label = row[:-1], row[-1]\n     print features, label\n     example = create_tf_example(features, label)\n     writer.write(example.SerializeToString())\nwriter.close()\n<\/code><\/pre>\n\n<p>For more details <a href=\"https:\/\/medium.com\/mostly-ai\/tensorflow-records-what-they-are-and-how-to-use-them-c46bc4bbb564\" rel=\"nofollow noreferrer\">click here<\/a>.This works for me, hope it works.<\/p>\n"}]}
{"QuestionId":41402409,"AnswerCount":5,"Tags":"<tensorflow>","CreationDate":"2016-12-30T20:37:30.350","AcceptedAnswerId":41530102.0,"OwnerUserId":7091875.0,"Title":"Tensorflow doesn't seem to see my gpu","Body":"<p>I've tried tensorflow on both cuda 7.5 and 8.0, w\/o cudnn (my GPU is old, cudnn doesn't support it). <\/p>\n\n<p>When I execute , there is no gpu in the output. Theano sees my gpu, and works fine with it, and examples in \/usr\/share\/cuda\/samples work fine as well. <\/p>\n\n<p>I installed tensorflow through pip install. Is my gpu too old for tf to support it? gtx 460<\/p>\n","answers":[{"AnswerId":"44990513","CreationDate":"2017-07-08T20:33:36.927","ParentId":null,"OwnerUserId":"6031995","Title":null,"Body":"<p>I came across this same issue in jupyter notebooks. This could be an easy fix.<\/p>\n\n<pre><code>$ pip uninstall tensorflow\n$ pip install tensorflow-gpu\n<\/code><\/pre>\n\n<p>You can check if it worked with:<\/p>\n\n<pre><code>tf.test.gpu_device_name()\n<\/code><\/pre>\n"},{"AnswerId":"51373397","CreationDate":"2018-07-17T04:50:10.590","ParentId":null,"OwnerUserId":"10003652","Title":null,"Body":"<p>If you are using conda, you might have installed the cpu version of the tensorflow. Check package list (<code>conda list<\/code>) of the environment to see if this is the case . If so, remove the package by using <code>conda remove tensorflow<\/code> and install keras-gpu instead (<code>conda install -c anaconda keras-gpu<\/code>. This will install everything you need to run your machine learning codes in GPU. Cheers!<\/p>\n\n<p>P.S. You should check first if you have installed the drivers correctly using <code>nvidia-smi<\/code>. By default, this is not in your PATH so you might as well need to add the folder to your path. The .exe file can be found at <code>C:\\Program Files\\NVIDIA Corporation\\NVSMI<\/code><\/p>\n"},{"AnswerId":"55891075","CreationDate":"2019-04-28T14:08:19.627","ParentId":null,"OwnerUserId":"10820672","Title":null,"Body":"<p>Summary:<\/p>\n\n<ol>\n<li>check if tensorflow sees your GPU (optional)<\/li>\n<li>check if your videocard can work with tensorflow (optional)<\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/install\/source#linux\" rel=\"nofollow noreferrer\">find versions of CUDA Toolkit and cuDNN SDK, compatible with your tf version<\/a><\/li>\n<li><a href=\"https:\/\/developer.nvidia.com\/cuda-toolkit-archive\" rel=\"nofollow noreferrer\">install CUDA Toolkit<\/a><\/li>\n<li><a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-archive\" rel=\"nofollow noreferrer\">install cuDNN SDK<\/a><\/li>\n<li>pip uninstall tensorflow; pip install tensorflow-gpu <\/li>\n<li>check if tensorflow sees your GPU<\/li>\n<\/ol>\n\n<p><code>*<\/code> source - <a href=\"https:\/\/www.tensorflow.org\/install\/gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu<\/a><\/p>\n\n<p>Detailed instruction:<\/p>\n\n<ol>\n<li><p>check if tensorflow sees your GPU (optional)<\/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]\nprint(get_available_devices()) \n# my output was =&gt; ['\/device:CPU:0']\n# good output must be =&gt; ['\/device:CPU:0', '\/device:GPU:0']\n<\/code><\/pre><\/li>\n<li><p>check if your card can work with tensorflow (optional)<\/p>\n\n<ul>\n<li>my PC: GeForce GTX 1060 notebook (driver version - 419.35), windows 10, jupyter notebook<\/li>\n<li><p>tensorflow needs Compute Capability 3.5 or higher. (<a href=\"https:\/\/www.tensorflow.org\/install\/gpu#hardware_requirements\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/gpu#hardware_requirements<\/a>)<\/p><\/li>\n<li><p><a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-gpus<\/a><\/p><\/li>\n<li>select \"CUDA-Enabled GeForce Products\"<\/li>\n<li>result - \"GeForce GTX 1060    Compute Capability = 6.1\"<\/li>\n<li>my card can work with tf!<\/li>\n<\/ul><\/li>\n<li><p>find versions of CUDA Toolkit and cuDNN SDK, that you need<\/p>\n\n<p>a) find your tf version<\/p>\n\n<pre><code>import sys\nprint (sys.version)\n# 3.6.4 |Anaconda custom (64-bit)| (default, Jan 16 2018, 10:22:32) [MSC v.1900 64 bit (AMD64)]\nimport tensorflow as tf\nprint(tf.__version__)\n# my output was =&gt; 1.13.1\n<\/code><\/pre>\n\n<p>b) find right versions of CUDA Toolkit and cuDNN SDK for your tf version<\/p>\n\n<pre><code>https:\/\/www.tensorflow.org\/install\/source#linux\n* it is written for linux, but worked in my case\nsee, that tensorflow_gpu-1.13.1 needs: CUDA Toolkit v10.0, cuDNN SDK v7.4\n<\/code><\/pre><\/li>\n<li><p>install CUDA Toolkit<\/p>\n\n<p>a) install CUDA Toolkit 10.0<\/p>\n\n<pre><code>https:\/\/developer.nvidia.com\/cuda-toolkit-archive\nselect: CUDA Toolkit 10.0 and download base installer (2 GB)\ninstallation settings: select only CUDA\n    (my installation path was: D:\\Programs\\x64\\Nvidia\\Cuda_v_10_0\\Development)\n<\/code><\/pre>\n\n<p>b) add environment variables:<\/p>\n\n<pre><code>system variables \/ path must have:\n    D:\\Programs\\x64\\Nvidia\\Cuda_v_10_0\\Development\\bin\n    D:\\Programs\\x64\\Nvidia\\Cuda_v_10_0\\Development\\libnvvp\n    D:\\Programs\\x64\\Nvidia\\Cuda_v_10_0\\Development\\extras\\CUPTI\\libx64\n    D:\\Programs\\x64\\Nvidia\\Cuda_v_10_0\\Development\\include\n<\/code><\/pre><\/li>\n<li><p>install cuDNN SDK<\/p>\n\n<p>a) download cuDNN SDK v7.4<\/p>\n\n<pre><code>https:\/\/developer.nvidia.com\/rdp\/cudnn-archive (needs registration, but it is simple)\nselect \"Download cuDNN v7.4.2 (Dec 14, 2018), for CUDA 10.0\"\n<\/code><\/pre>\n\n<p>b) add path to 'bin' folder into \"environment variables \/ system variables \/ path\":<\/p>\n\n<pre><code>D:\\Programs\\x64\\Nvidia\\cudnn_for_cuda_10_0\\bin\n<\/code><\/pre><\/li>\n<li><p>pip uninstall tensorflow\npip install tensorflow-gpu <\/p><\/li>\n<li><p>check if tensorflow sees your GPU<\/p>\n\n<pre><code>- restart your PC\n- print(get_available_devices()) \n- # now this code should return =&gt; ['\/device:CPU:0', '\/device:GPU:0']\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"46912222","CreationDate":"2017-10-24T13:46:28.650","ParentId":null,"OwnerUserId":"8825550","Title":null,"Body":"<p>The following worked for me, hp laptop. I have a Cuda Compute capability\n(version) 3.0 compatible Nvidia card. Windows 7.<\/p>\n\n<pre><code>pip3.6.exe uninstall tensorflow-gpu\npip3.6.exe uninstall tensorflow-gpu\npip3.6.exe install tensorflow-gpu\n<\/code><\/pre>\n"},{"AnswerId":"41530102","CreationDate":"2017-01-08T06:24:37.803","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>When I look up your GPU, I see that it only supports CUDA Compute Capability 2.1. (Can be checked through <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"noreferrer\">https:\/\/developer.nvidia.com\/cuda-gpus<\/a>) Unfortunately, TensorFlow needs a GPU with minimum CUDA Compute Capability 3.0.\n<a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#optional_install_cuda_gpus_on_linux\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/get_started\/os_setup#optional_install_cuda_gpus_on_linux<\/a><\/p>\n\n<p>You might see some logs from TensorFlow checking your GPU, but ultimately the library will avoid using an unsupported GPU. <\/p>\n"}]}
{"QuestionId":41402903,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2016-12-30T21:27:37.833","AcceptedAnswerId":null,"OwnerUserId":497600.0,"Title":"Tensorflow: use pretrained inception model","Body":"<p>I would like to use one of these pretrained tensorflow models: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim<\/a><\/p>\n\n<p>After loading the inceptionv4 model, I've had problems with some test predictions.\nThere is a similar question: <a href=\"https:\/\/stackoverflow.com\/questions\/39582703\/using-pre-trained-inception-resnet-v2-with-tensorflow\">Using pre-trained inception_resnet_v2 with Tensorflow<\/a><\/p>\n\n<p>In that question, the solution was to fix the image preprocessing.\nI tried using ranges for the color channels from 0 to 1 and from -1 to 1.<\/p>\n\n<p>Here is my code (I've imported everything from the inceptionv4 source file):<\/p>\n\n<pre><\/pre>\n\n<p>In the code, I'm testing the network with a horse. This is the picture.\n<a href=\"https:\/\/i.stack.imgur.com\/N91h1.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/N91h1.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>With the current preprocessing, color channels from -1 to 1, the network thinks that this horse is a bathing cap.\nFor scaling from 0 to 1, it becomes a bittern, apparently a small bird.\nI've used this table to find out the predicted classes: <a href=\"https:\/\/gist.github.com\/aaronpolhamus\/964a4411c0906315deb9f4a3723aac57\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/aaronpolhamus\/964a4411c0906315deb9f4a3723aac57<\/a><\/p>\n\n<p>I've also checked more than one image. The network is consistently off.<\/p>\n\n<p>What is going wrong ?<\/p>\n","answers":[{"AnswerId":"56867769","CreationDate":"2019-07-03T10:11:38.730","ParentId":null,"OwnerUserId":"11733761","Title":null,"Body":"<p>when you import the inceptions, <\/p>\n\n<pre><code>net, logits, end_points = inception_v4(inputs)\n<\/code><\/pre>\n\n<p>should be <\/p>\n\n<pre><code>logits, end_points = inception_v4(inputs, is_training=False, dropout_keep_prob=1.0) \n<\/code><\/pre>\n\n<p>for inference<\/p>\n"},{"AnswerId":"43616239","CreationDate":"2017-04-25T16:29:55.757","ParentId":null,"OwnerUserId":"7847475","Title":null,"Body":"<p>I think you used the wrong synsets for Imagenet. To be specific, the one you used is 2012 version. You may try these two: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/data\/imagenet_lsvrc_2015_synsets.txt\" rel=\"nofollow noreferrer\">imagenet_lsvrc_2015_synsets.txt<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/example\/imagenet_metadata.txt\" rel=\"nofollow noreferrer\">imagenet_metadata<\/a>.<\/p>\n\n<p>For example, if your output is 340, then 340->n02389026-> sorrel<\/p>\n"},{"AnswerId":"44229065","CreationDate":"2017-05-28T15:36:08.317","ParentId":null,"OwnerUserId":"6451244","Title":null,"Body":"<p>I would agree for the wrong synset, it can be automatically downloaded with the imagenet file, that way your are sure to have the correct one:<\/p>\n\n<pre><code>from datasets import imagenet\nnames = imagenet.create_readable_names_for_imagenet_labels()\n\nprint(names[0])\n<\/code><\/pre>\n"}]}
{"QuestionId":41403146,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-30T21:55:34.933","AcceptedAnswerId":41404075.0,"OwnerUserId":2193429.0,"Title":"Tensorflow can't save model","Body":"<p>I encountered this weird problem...I use this code to construct tensorflow saver: <\/p>\n\n<pre><\/pre>\n\n<p>which is supposed to be very standard. However, when I point the saving directory to my custom directory (under my username) instead of \"\/tmp\", all of a sudden, the saved models are files like <\/p>\n\n<p><strong>translate.ckpt-329.data-00000-of-00001<\/strong><\/p>\n\n<p><strong>translate.ckpt-329.index<\/strong><\/p>\n\n<p><strong>translate.ckpt-329.meta<\/strong><\/p>\n\n<p>I can't find the file \"translate.ckpt-329\".<\/p>\n\n<p>The generated checkpoint file is pointing to:<\/p>\n\n<pre><\/pre>\n\n<p>while this file does not exist and create problems for me restoring my model.<\/p>\n\n<p>Can someone shed any light on this?? What could possibly be the problem?<\/p>\n\n<hr>\n\n<p>Thanks for the first answer! I guess my bigger problem is the  method:<\/p>\n\n<p>The original code uses this way to restore a session:<\/p>\n\n<pre><\/pre>\n\n<p>Which failed with V2 saving :(<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41404075","CreationDate":"2016-12-30T23:42:57.617","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> In the new checkpoint format, the \"filename\" that you pass to the saver is actually used as the prefix of several filenames, and no file with that exact name is written. You can use the old checkpoint format by constructing your <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.__init__\" rel=\"nofollow noreferrer\"><code>tf.train.Saver<\/code><\/a> with the optional argument <code>write_version=tf.train.SaverDef.V1<\/code>.<\/p>\n\n<p>From the names of the saved files, it appears that you are using the \"V2\" checkpoint format, which became the default in TensorFlow 0.12. This format stores the checkpoint data in multiple files: one or more data files (e.g. <code>translate.ckpt-329.data-00000-of-00001<\/code> in your case) and an index file (<code>translate.ckpt-329.index<\/code>) that tells TensorFlow where each saved variable is located in the data files. The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver\" rel=\"nofollow noreferrer\"><code>tf.train.Saver<\/code><\/a> uses the \"filename\" that you pass as the prefix for these files' names, but doesn't produce a file with that exact name.<\/p>\n\n<p>Although there is no file with the exact name you gave, you can use the value returned from <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.save\" rel=\"nofollow noreferrer\"><code>saver.save()<\/code><\/a> as the argument to a subsequent <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.restore\" rel=\"nofollow noreferrer\"><code>saver.restore()<\/code><\/a>, and the other checkpoint locating mechanisms should continue to work as before.<\/p>\n"}]}
{"QuestionId":41403694,"AnswerCount":0,"Tags":"<python><tensorflow><division>","CreationDate":"2016-12-30T22:52:32.557","AcceptedAnswerId":null,"OwnerUserId":7359223.0,"Title":"Tensorflow division evaluates to infinity","Body":"<p>i am new in Tensorflow, so sorry for obvious mistakes.<\/p>\n\n<p>I try to make a division wich evaluates to inf. The following line produces the problem :<\/p>\n\n<pre><\/pre>\n\n<p>I know that the i only multiply numbers, but if this works, i want to replace them by Tensorflow variables. So i need this multiplication.\nThe thing which makes me crazy is, that when i change this line to this one : <\/p>\n\n<pre><\/pre>\n\n<p>everything works fine and i get the result : 7.41678e+19<\/p>\n\n<p>I would be thankful for any help :)<\/p>\n","answers":[]}
{"QuestionId":41403727,"AnswerCount":1,"Tags":"<lua><gpu><torch>","CreationDate":"2016-12-30T22:55:31.623","AcceptedAnswerId":41403873.0,"OwnerUserId":7241218.0,"Title":"Set GPU usage in Torch Machine Learning Framework","Body":"<p>currently I'm using Theano for machine learning, now I wanted to try out Torch.<\/p>\n\n<p>In Theano there is an option to set flags for GPU Memory usage:<\/p>\n\n<p><\/p>\n\n<p>So Theano uses the set capacity of the GPU, here 90%. In Torch however a similar network uses around 30% GPU load.<\/p>\n\n<p>Is there any way to set a higher GPU load in Torch similar to Theano?<\/p>\n","answers":[{"AnswerId":"41403873","CreationDate":"2016-12-30T23:15:21.697","ParentId":null,"OwnerUserId":"117844","Title":null,"Body":"<p>Torch will use as much GPU memory as it needs based on it's standard allocator. <\/p>\n\n<p>The amount of memory torch uses does not need to be pre-specified like the example you do in Theano.<\/p>\n"}]}
{"QuestionId":41403838,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2016-12-30T23:11:23.580","AcceptedAnswerId":null,"OwnerUserId":288609.0,"Title":"Regarding setting up the target tensor shape for sparse_categorical_crossentropy","Body":"<p>I am trying to experiment with a multi-layer encoder-decoder type of network. The screenshot of the last several layers of network architecture is as follows.  This is how I setup model compiling and training process.<\/p>\n\n<pre><\/pre>\n\n<p> and  are of shape .  represent the raw image and  represents the mask image. I am trying to solve a semantic segmentation problem. However, running the program generates the following error message, (I only keep the main related part). <\/p>\n\n<pre><\/pre>\n\n<p>It seems to me that the loss function of  causes the problem for the current (imgs_train, imgs_mask_train) shape setting. The Keras API does not include the detail about how to setup the target tensor. Any suggestions are highly appreciated! <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/f12s4.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/f12s4.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42070861","CreationDate":"2017-02-06T15:01:32.837","ParentId":null,"OwnerUserId":"2115967","Title":null,"Body":"<p>I am currently trying to figure the same problem and as far as I can tell it takes a sparse representation of the target category. That means integers as the target label instead of the one-hot encoded binary class matrix.<\/p>\n\n<p>Concerning your problem, do you have categories in your masking or do you just have information about the outline of an object? With outline information it becomes a pixel wise binary loss instead of a categorical one. If you have categories, the output of your decoder should have dimensionality (None, number_of_classes, 128, 128). On that you should be able to use a sparse target mask but I haven't tried this myself...\nHope that helps<\/p>\n"}]}
{"QuestionId":41403999,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><sequence-to-sequence>","CreationDate":"2016-12-30T23:33:17.173","AcceptedAnswerId":null,"OwnerUserId":2827214.0,"Title":"Do I need the Sequence Length parameter for RNNCell in Tensorflow","Body":"<p>If my input is variable length and I pad it, for example:<\/p>\n\n<pre><\/pre>\n\n<p>And I pass to Tensorflow's RNN cell that:<\/p>\n\n<pre><\/pre>\n\n<p>Computation is stopped before the PAD symbol is evaluated, and you get a sequence of output states (one for each input):<\/p>\n\n<pre><\/pre>\n\n<p>state4 and state5 are all zeros:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Is it correct that the gradient calculations of my model will not change if I do not pass the sequence_length parameter and instead manually replace state4 and state5 with zero vectors?<\/strong> <\/p>\n","answers":[]}
{"QuestionId":41404497,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-31T00:55:49.780","AcceptedAnswerId":41404617.0,"OwnerUserId":6108836.0,"Title":"Exiting current scope","Body":"<p>I have some function that I repeatedly call to generate summaries. It really clutters the graph and it makes debugging it really hard.<\/p>\n\n<p>Is there a way to exit the scope for the particular function?\nLike creating a with block that goes to the root.<\/p>\n","answers":[{"AnswerId":"41404617","CreationDate":"2016-12-31T01:22:21.517","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/utility_functions#name_scope\" rel=\"nofollow noreferrer\"><code>with tf.name_scope(None):<\/code><\/a> to define a block of operations that will be created in the root scope.<\/p>\n\n<p>(Note that this functionality is not documented as part of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/utility_functions#name_scope\" rel=\"nofollow noreferrer\"><code>tf.name_scope()<\/code><\/a>, but it is documented in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Graph.name_scope\" rel=\"nofollow noreferrer\"><code>tf.Graph.name_scope()<\/code><\/a>, which <code>tf.name_scope()<\/code> wraps.)<\/p>\n"}]}
{"QuestionId":41404580,"AnswerCount":1,"Tags":"<python><machine-learning><keras><embedding><flatten>","CreationDate":"2016-12-31T01:14:57.047","AcceptedAnswerId":null,"OwnerUserId":1127644.0,"Title":"Using Keras to predict whether two numbers have the same \"oddness\" using an embedding, am I on the right track?","Body":"<p>I'm new to machine learning, and I'm working on a simple example as practice. I generate pairs of numbers, and assign the label '1' if both values are even or odd, and the label '0' if one is even and the other is odd.<\/p>\n\n<p>My model sometimes gets to around ~75%, but I'm fairly sure it's simply memorizing which pairs of numbers lead to 1 and which lead to 0. I want the model to learn that there are two categories of numbers, for example, knowing that [1, 4] = 0 and [1, 6] = 0, therefor 4 and 6 are in the same category.<\/p>\n\n<p>Am I on the right track? Is this even a reasonable problem to solve with ML?<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41506779","CreationDate":"2017-01-06T13:27:40.920","ParentId":null,"OwnerUserId":"6938290","Title":null,"Body":"<p>This model won't be able to generalize to unseen numbers since the embedding layer assigns for each number a feature vector. if a number is not in any training example, it will never get updated and will keep its initial random value. I think supplying the numbers in binary form (e.g. supply '9' as 8 inputs of value 00001001) will help. anyway this is a dirty hack in my opinion since all the network is required to do now is to learn to detect when the last bit of each input are equal.<\/p>\n"}]}
{"QuestionId":41406518,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2016-12-31T07:51:45.453","AcceptedAnswerId":null,"OwnerUserId":7360109.0,"Title":"tensorflow placeholder of single dimension","Body":"<p>I am having trouble getting this very simple tensorflow code off the ground. I am trying to do a simple line fitting of the form y = theta1 * x + theta2<\/p>\n\n<p>I created the data for x and y as a numpy float32 array of shape [10], I created their corresponding placeholders as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I feed them like follows:<\/p>\n\n<pre><\/pre>\n\n<p>the full code is bit long so I created a gist:\n<a href=\"https:\/\/gist.github.com\/meowmiau\/369393f41b679dd95f4ac4e2e16b0782\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/meowmiau\/369393f41b679dd95f4ac4e2e16b0782<\/a><\/p>\n\n<p>The issue I am getting is this:<\/p>\n\n<pre><\/pre>\n\n<p>However, as far as I know there is no mismatch.<\/p>\n","answers":[{"AnswerId":"41406890","CreationDate":"2016-12-31T08:57:13.477","ParentId":null,"OwnerUserId":"7277946","Title":null,"Body":"<p>Try this modification in your code:<\/p>\n\n<pre><code>for i in range(1000):\n  x_data, y_data = gen_data()\n  _, e = sess.run([train, err], feed_dict={tf_x: x_data, tf_y: y_data})\n  print e\n<\/code><\/pre>\n\n<p>@Meng Sun:<\/p>\n\n<p>I think to pass the list [train, err] to sess.run() is a possible solution. \nThe following snippet works the same way:<\/p>\n\n<pre><code>for i in range(1000):\n  x_data, y_data = gen_data()\n  feed_dict={tf_x: x_data, tf_y: y_data}\n  print(sess.run(err, feed_dict=feed_dict))\n  sess.run(train, feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>In your code the two placeholders throw the error \"You must feed a value for placeholder tensor\" \nbecause sess.run(err) was executed without a feed.<\/p>\n"}]}
{"QuestionId":41406633,"AnswerCount":0,"Tags":"<python><keras><keras-layer>","CreationDate":"2016-12-31T08:10:34.390","AcceptedAnswerId":null,"OwnerUserId":7360138.0,"Title":"something wrong with my keras's embedding's test","Body":"<p>Here is my python's code runned on pyspark with distkeras.<\/p>\n\n<pre><\/pre>\n\n<p>But it print error like this:<\/p>\n\n<pre><\/pre>\n\n<p>My 11.csv is like this :    <\/p>\n\n<pre><\/pre>\n\n<p>It's C0 column is label,and C1 to C50 are features.\nHope for your answer.Thanks a lots.   <\/p>\n\n<p>Here is dataset.printSchema():<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41406697,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-12-31T08:21:04.813","AcceptedAnswerId":41413058.0,"OwnerUserId":5350430.0,"Title":"TensorFlow: running training_op hangs forever","Body":"<p>I can't run my the code even one training loop, it stucks at sess.run or training_op.run function (run the code forever...). I don't know where's the bug. <\/p>\n\n<pre><\/pre>\n\n<p>The code above loads the dataset.\nThe samples_all is a list which contains images' path. The size of images is 240 * 320 * 3. labels_all is a list which contains dense classes. There're 101 classes.\n   I step in the sess.run function,  and find that it go into the _do_call function, and exec fn(*args). However, it never returns, and no exception is captured.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44824300","CreationDate":"2017-06-29T11:47:49.820","ParentId":null,"OwnerUserId":"3323084","Title":null,"Body":"<p>Try using <code>tf.train.MonitoredTrainingSession<\/code> which will take of all the initialization and starting of queue runners.<\/p>\n"},{"AnswerId":"41413058","CreationDate":"2017-01-01T02:55:53.123","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> You need to add <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/coordinator_and_queuerunner#start_queue_runners\" rel=\"noreferrer\"><code>tf.train.start_queue_runners(sess)<\/code><\/a> after you run the <code>init_op<\/code>, and before you start the training loop.<\/p>\n\n<p>The <code>tf.train.batch()<\/code> function uses TensorFlow queues to accumulate input data into batches. These queues are filled by background threads, which are created when you call <code>tf.train.start_queue_runners()<\/code>. If you don't call this method, the background threads will not start, the queues will remain empty, and the training op will block indefinitely waiting for input.<\/p>\n"}]}
{"QuestionId":41407124,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2016-12-31T09:34:27.867","AcceptedAnswerId":null,"OwnerUserId":2542516.0,"Title":"Retraining the last layer of Inception-ResNet-v2","Body":"<p>I am trying to retrain the last layer of inception-resnet-v2. Here's what I came up with:<\/p>\n\n<ol>\n<li>Get names of variables in the final layer<\/li>\n<li>Create a  to minimise only these variables wrt loss<\/li>\n<li>Restore the whole graph except the final layer while initialising only the last layer randomly.<\/li>\n<\/ol>\n\n<p>And I implemented that as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This doesn't seem to work (training loss, error don't improve much from initial values). Is there a better\/elegant way to do this? It would be good learning for me if you can also tell me what's going wrong here.<\/p>\n","answers":[{"AnswerId":"41514168","CreationDate":"2017-01-06T20:54:01.763","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>There are several things:<\/p>\n\n<ul>\n<li>how is the learning rate? a too high value can mess with everything (probably not the reason)<\/li>\n<li>try to use stochastic gradient descent, you should have less problems<\/li>\n<li><p>is the scope correctly set? if you don't use L2 regularization and batch normalization of the gradients you might fall into a local minimum very soon and the network is unable to learn<\/p>\n\n<pre><code>from nets import inception_resnet_v2 as net\nwith net.inception_resnet_v2_arg_scope():\n    logits, end_points = net.inception_resnet_v2(images_ph, num_classes=num_classes,\n                                                 is_training=True)\n<\/code><\/pre><\/li>\n<li><p>you should add the regularization variables to the loss (or at least the ones of the last layer):<\/p>\n\n<pre><code>regularization_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)\nall_losses = [loss] + regularization_losses\ntotal_loss = tf.add_n(all_losses, name='total_loss')\n<\/code><\/pre><\/li>\n<li><p>training only the full connected layer might not be a good idea, I would train all the network as the features you need for your class aren't necessarily defined in the last layer but few layers before and you need to change them.<\/p><\/li>\n<li><p>double check the train_op runs after the loss:<\/p>\n\n<pre><code>with ops.name_scope('train_op'):\n    train_op = control_flow_ops.with_dependencies([train_op], total_loss)\n<\/code><\/pre><\/li>\n<\/ul>\n"}]}
{"QuestionId":41407241,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2016-12-31T09:53:03.037","AcceptedAnswerId":null,"OwnerUserId":5629538.0,"Title":"Does TensorFlow execute entire computation graph with sess.run()?","Body":"<p>For example, when we compute a variable  as , does TF only compute the inputs required for computing  or updates all the variables of the complete computational graph?<\/p>\n\n<p>Also, I don't seem to be able to do this:\n\nas I am stuck with  error even after initializing  as . Any suggestions?<\/p>\n","answers":[{"AnswerId":"45947287","CreationDate":"2017-08-29T19:50:10.067","ParentId":null,"OwnerUserId":"651945","Title":null,"Body":"<p>Since Python code of TF only setups the graph, which is actually executed by native implementation of all <code>ops<\/code>, your variables need to be executed in this underlying environment. This happens by executing two ops - for local and global variables initialization:<\/p>\n\n<p><code>\nsession.run(tf.global_variables_initializer(), tf.local_variables_initializer())\n<\/code><\/p>\n\n<p>On the original question - as far as I know - YES, it computes all the graph, and it requires you to feed placeholders, even if the executed op (in the session) is not dependent on them.<\/p>\n"},{"AnswerId":"53330013","CreationDate":"2018-11-16T01:04:11.280","ParentId":null,"OwnerUserId":"2730161","Title":null,"Body":"<p>This question is fairly old but I'll answer it just in case anyone comes across this again. <\/p>\n\n<p>To answer the first question, no. Tensorflow only executes what is necessary to execute the operation you passed it. From tensorflow.org:<\/p>\n\n<blockquote>\n  <p>tf.Session.run requires you to specify a list of fetches, which determine the return values, and may be a tf.Operation, a tf.Tensor, or a tensor-like type such as tf.Variable. These fetches determine what subgraph of the overall tf.Graph must be executed to produce the result<\/p>\n<\/blockquote>\n\n<p>To answer the second part of the question, variables need to be initialized before you can use them (see link below). the easiest way to do this is to run tensorflow.global_variables_initializer()<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/guide\/variables#initializing_variables\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/guide\/variables#initializing_variables<\/a><\/p>\n"}]}
{"QuestionId":41407270,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-31T09:56:31.930","AcceptedAnswerId":null,"OwnerUserId":7360291.0,"Title":"how can I scale a tensor in tensorflow?","Body":"<p>For example, I have a tensor shape is [128,128,3] and it's range is random,then I want to scale all nums in this tensor to [0,255],what function in tensorflow should I use to do it?\nThanks<\/p>\n","answers":[{"AnswerId":"41407314","CreationDate":"2016-12-31T10:03:33.083","ParentId":null,"OwnerUserId":"4396113","Title":null,"Body":"<p>You can do the scaling by yourself.<\/p>\n\n<pre><code>\/\/ x is your tensor\ncurrent_min = tf.reduce_min(x)\ncurrent_max = tf.reduce_max(x)\ntarget_min = 0\ntarget_max = 255\n\n\/\/ scale to [0; 1]\nx = (x - current_min) \/ (current_max - current_min)\n\n\/\/ scale to [target_min; target_max]\nx = x * (target_max - target_min) + target_min\n<\/code><\/pre>\n\n<p>You just have to handle the edge case when all the values are equal.<\/p>\n"}]}
{"QuestionId":41408204,"AnswerCount":0,"Tags":"<lua><torch>","CreationDate":"2016-12-31T12:14:56.667","AcceptedAnswerId":null,"OwnerUserId":4513683.0,"Title":"Got `internal error in __sub: no metatable` error in Lua and Torch","Body":"<p>I tried to create a image-to-image model by  (<a href=\"https:\/\/github.com\/phillipi\/pix2pix\" rel=\"nofollow noreferrer\">https:\/\/github.com\/phillipi\/pix2pix<\/a>) and followed Getting Started instruction.\nThen, downloaded facades dataset by bash \nand run <\/p>\n\n<p>but I got the following error.<\/p>\n\n<blockquote>\n  <p>~\/torch-cl\/install\/bin\/luajit: .\/models.lua:69: internal error in\n  __sub: no metatable stack traceback:  [C]: in function '__sub'    .\/models.lua:69: in function 'defineG_unet'     train.lua:110: in\n  function 'defineG'    train.lua:146: in main chunk    [C]: in function\n  'dofile'\n    ...i\/torch-cl\/install\/lib\/luarocks\/rocks\/trepl\/scm-1\/bin\/th:145: in\n  main chunk    [C]: at 0x010ec81ce0<\/p>\n<\/blockquote>\n\n<p>Does anyone know how I can fix this?\nI'm new to Lua and Torch, so have no idea how to fix this.<\/p>\n\n<p>Thank you<\/p>\n\n<p>this is run on Mac OSX\npython2.7<\/p>\n","answers":[]}
{"QuestionId":41408561,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-31T13:08:54.980","AcceptedAnswerId":41412747.0,"OwnerUserId":6718080.0,"Title":"dynamic_partition with dynamic num_partitions","Body":"<p>The variable  in the method  is not a , but an . \nTherefore, if the number of partitions isn't known in advance, it can't be inferred from the data by e.g., counting the number of unique values nor can't be given by a .\nHow can one use this method in such a dynamic scenario?<\/p>\n\n<p>If it's impossible, a plausible workaround is to set the value of this parameter to some upper bound. Then at running time there will be some empty lists. The question is how those empty lists can be eliminated?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41412747","CreationDate":"2017-01-01T01:20:56.100","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>To have fully dynamic partition you could have an op that returns a single Tensor with dynamic shape rather than Python-time fixed number of tensors, but the problem is that tensor dimensions must be rectangular and your partitions could have different lengths. To work-around it you might encode your variable size lists into strings which have shape () or use <code>TensorArray<\/code>. Here's a way to do it by using png encoding for arrays&lt;=>string conversion <\/p>\n\n<pre><code>def dynamic_partition_png(vals, idx, max_partitions):\n    \"\"\"Encodes output of dynamic partition as a Tensor of png-encoded strings.\"\"\"\n    max_idx = tf.reduce_max(idx)\n    max_vals = tf.reduce_max(idx)\n    with tf.control_dependencies([tf.Assert(max_vals&lt;256, [\"vals must be &lt;256\"])]):\n        outputs = tf.dynamic_partition(vals, idx, num_partitions=max_partitions)\n    png_outputs = []\n    dummy_png = tf.image.encode_png(([[[2]]]))\n    not_empty_ops = [] # ops that detect empty lists that aren't at the end\n    for i, o in enumerate(outputs):\n        reshaped_o = tf.reshape(tf.cast(o, tf.uint8), [-1, 1, 1])\n        png_output = tf.cond(tf.size(reshaped_o)&gt;0, lambda: tf.image.encode_png(reshaped_o), lambda: dummy_png)\n        png_outputs.append(png_output)\n        not_empty_ops.append(tf.logical_or(i&gt;max_idx, tf.size(reshaped_o)&gt;0))\n    packed_tensor = tf.pack(png_outputs)\n    no_illegal_empty_lists = tf.reduce_all(tf.pack(not_empty_ops))\n    with tf.control_dependencies([tf.Assert(no_illegal_empty_lists, [\"empty lists must be last\"])]):\n        result = packed_tensor[:max_idx+1]\n    return result\n\ndef decode(p):\n    return tf.image.decode_png(p)[:, 0, 0]\n\nsess = tf.Session()\nvals = tf.constant([1,2,3,4,5])\nidx = [0, 1, 1, 1, 1]\ntf_vals = dynamic_partition_png(vals, idx, 3)\nprint(sess.run(decode(tf_vals[0]))) # =&gt; [1 2]\nprint(sess.run(decode(tf_vals[1]))) # =&gt; [3 4 5]\nprint(sess.run(decode(tf_vals[2]))) # =&gt; slice index 2 of dimension 0 out of bounds\n<\/code><\/pre>\n"}]}
{"QuestionId":41409248,"AnswerCount":4,"Tags":"<tensorflow><computer-vision><deep-learning><theano><keras>","CreationDate":"2016-12-31T14:46:29.013","AcceptedAnswerId":41409315.0,"OwnerUserId":288609.0,"Title":"softmax and sigmoid function for the output layer","Body":"<p>In the deep learning implementations related to object detection and semantic segmentation, I have seen the output layers using either sigmoid or softmax. I am not very clear when to use which? It seems to me both of them can support these tasks. Are there any guidelines for this choice?<\/p>\n","answers":[{"AnswerId":"56310324","CreationDate":"2019-05-26T03:02:36.470","ParentId":null,"OwnerUserId":"9078508","Title":null,"Body":"<p>The sigmoid and the softmax function have different purposes. For a detailed explanation of when to use sigmoid vs. softmax in neural network design, you can look at this article: <a href=\"https:\/\/glassboxmedicine.com\/2019\/05\/26\/classification-sigmoid-vs-softmax\/\" rel=\"nofollow noreferrer\">\"Classification: Sigmoid vs. Softmax.\"<\/a><\/p>\n\n<p>Short summary: <\/p>\n\n<p>If you have a multi-label classification problem where there is more than one \"right answer\" (the outputs are NOT mutually exclusive) then you can use a sigmoid function on each raw output independently. The sigmoid will allow you to have high probability for all of your classes, some of them, or none of them. <\/p>\n\n<p>If you instead have a multi-class classification problem where there is only one \"right answer\" (the outputs are mutually exclusive), then use a softmax function. The softmax will enforce that the sum of the probabilities of your output classes are equal to one, so in order to increase the probability of a particular class, your model must correspondingly decrease the probability of at least one of the other classes.<\/p>\n"},{"AnswerId":"44160266","CreationDate":"2017-05-24T13:46:09.997","ParentId":null,"OwnerUserId":"6712021","Title":null,"Body":"<p>In general Softmax is used (Softmax Classifier) when \u2018n\u2019 number of classes are there. Sigmoid or softmax both can be used for binary (n=2) classification.<\/p>\n\n<p>Sigmoid:\n            S(x) = 1\/  ( 1+ (  e^(-x) ))<\/p>\n\n<p>Softmax:<\/p>\n\n<pre><code>         \u03c3(x)j = e    \/  **\u03a3**{k=1 to K} e^zk    for(j=1.....K)\n<\/code><\/pre>\n\n<p>Softmax is kind of Multi Class Sigmoid, but if you see the function of Softmax, the sum of all softmax units are supposed to be 1. In sigmoid it\u2019s not really necessary.<\/p>\n\n<p>Digging deep, you can also use sigmoid for multi-class classification. When you use a softmax, basically you get a probability of each class, (join distribution and a multinomial likelihood) whose sum is bound to be one. In case you use sigmoid for multi class classification, it\u2019d be like a marginal distribution and a Bernoulli likelihood, p(y0\/x) , p(y1\/x) etc<\/p>\n"},{"AnswerId":"41409315","CreationDate":"2016-12-31T14:57:01.520","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p><code>softmax()<\/code> helps when you want a probability distribution, which sums up to 1. <code>sigmoid<\/code> is used when you want the output to be ranging from 0 to 1, but need not sum to 1.<\/p>\n\n<p>In your case, you wish to classify and choose between two alternatives. I would recommend using <code>softmax()<\/code> as you will get a probability distribution which you can apply cross entropy loss function on.<\/p>\n"},{"AnswerId":"41409397","CreationDate":"2016-12-31T15:06:30.680","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Object detection is object classification used on a sliding window in the image. In classification it is important to find the correct output in some class space. E.g. you detect 10 different objects and you want to know which object is the most likely one in there. Then softmax is good because of its proberty that the whole layer sums up to 1. <\/p>\n\n<p>Semantic segmentation on the other hand segments the image in some way. I have done semantic medical segmentation and there the output is a binary image. This means you can have sigmoid as output to predict if this pixel belongs to this specific class, because sigmoid values are between 0 and 1 for each output class.<\/p>\n"}]}
{"QuestionId":41410007,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><keras>","CreationDate":"2016-12-31T16:22:18.097","AcceptedAnswerId":41415104.0,"OwnerUserId":297850.0,"Title":"The loss plot and the criteria for determining whether it converges","Body":"<p>I have been running a keras u-net implementation with tensorflow as backend. In the experiment, I run 60 epochs. The loss plot is shown as following, can I assume that the program has converged? Looks to me that the loss does not change significantly.<\/p>\n\n<p>This is the code related to defining optimizer and launching the training process.<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/npIn6.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/npIn6.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41415104","CreationDate":"2017-01-01T11:27:14.933","ParentId":null,"OwnerUserId":"1828289","Title":null,"Body":"<p>It is converging - the initial convergence is fast, then it slows down - that's normal, it can slow down a lot and still improve meaningfully. Sixty epochs may or may not be a long time depending on the number of parameters in your network and the amount of training data.  You want essentially no downward trend over a significant amount of epochs before you decide it is converged. <\/p>\n\n<p>A pretty good rule of thumb may be something like \"number of epochs since lowest loss was seen > total number of epochs \/ 4\".  <\/p>\n\n<p>A separate question is whether it is overtraining.  You need to look at validation loss for that.  Hold back 1\/10 or 1\/5 of the data from training and use that for validation, when validation loss starts increasing consistently then you're overtraining.<\/p>\n"}]}
{"QuestionId":41410014,"AnswerCount":0,"Tags":"<caching><tensorflow>","CreationDate":"2016-12-31T16:23:21.817","AcceptedAnswerId":null,"OwnerUserId":3276106.0,"Title":"Tensorflow avoid multiple computation of same variable","Body":"<p>I'm pretty new using tensorflow and I'd like to know how can I avoid that a value is computed twice when actually I want to use the same value.<\/p>\n\n<p>I'm trying to implement k-means by scratch and here's the code that creates the initial points (centroids + samples):<\/p>\n\n<pre><\/pre>\n\n<p>I've added the s for debugging purpose. And here's the code that calls this function:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the output:<\/p>\n\n<pre><\/pre>\n\n<p>Before adding the  and by checking only the output of the function I was really confused on what was happening. As you can see the value of the variable  is computed twice, I suppose the first time to create the  variable and the second time when it is summed up to . As you can imagine it is not what I want, how can I avoid that  is computed twice for each iteration in the  loop? Are there more methods to obtain this behaviour? What is the best practice?<\/p>\n\n<p>System info:<\/p>\n\n<p>I'm running Python 3.5.2 with tensorflow 0.11 installed through conda () on OSX 10.11.<\/p>\n","answers":[]}
{"QuestionId":41410438,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-31T17:28:18.540","AcceptedAnswerId":null,"OwnerUserId":7361100.0,"Title":"Tensorflow startup time?","Body":"<p>I've been working with a GPU version of Tensorflow 0.9.0 on my University's cluster. When I submit the job, it begins running and outputs a message such as:<\/p>\n\n<p>(stuff that says CUDA found the device...)<br>\nI tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:808] Creating TensorFlow device (\/gpu:0) -> (device: 0, name: GeForce GTX TITAN X, pci bus id: 0000:04:00.0)<\/p>\n\n<p>However after this, it doesn't begin actually processing anything for a long time. It seems like it just hangs there for a while... For the record, I'm using Imagenet data formatted as in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/data\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/data<\/a>, and creating all of my Queues, etc. on a CPU and running all variables\/operations on the GPU.<\/p>\n\n<p>I have tried not explicitly calling for the CPU\/GPU splits, and permitting the soft_device_placement to do its thing, but this results in the same hang-ups too.<\/p>\n\n<p>Edit: Should also mention that even when working with the raw .JPEG files (ie: not using the processing techniques above) this still happens. So, I don't think it's much of an issue with that?<\/p>\n\n<p>Has anybody else experienced this, and is there anyway around it?<\/p>\n\n<p>Thank you.<\/p>\n\n<p>Edit: Code snippet<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"52472462","CreationDate":"2018-09-24T03:50:57.380","ParentId":null,"OwnerUserId":"1499598","Title":null,"Body":"<p>Nvidia drivers take time to wake up for some machines. Run the following command before running the script. <\/p>\n\n<pre><code>sudo nvidia-persistenced --persistence-mode\n<\/code><\/pre>\n"}]}
{"QuestionId":41410879,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2016-12-31T18:40:37.160","AcceptedAnswerId":41410965.0,"OwnerUserId":4943253.0,"Title":"When calling tf.Session.run, is it important for performance to minimize what is fetched?","Body":"<p>In Tensorflow, the  argument to  serves two purposes: it ensures that certain ops are run, and it returns values to python.  My question is how concerned I need to be to avoid returning values I don't actually need.  For instance, suppose that op  depends on op  so that if  is evaluated  will be too.  Could it be faster to do<\/p>\n\n<pre><\/pre>\n\n<p>than to do<\/p>\n\n<pre><\/pre>\n\n<p>because the latter would copy the output of  from the GPU to the CPU only for it to be discarded?  If only a fraction of the information in  is needed on the python side, is it generally a good practice to make a tensor depending on  that only has exactly what is needed and to fetch that instead of ?  Or am I misunderstanding something about how these GPU-CPU transfers work?  I think it is possible to profile how much time is spent in transfers like these (e.g., see graphs here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4526\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4526<\/a>), but I don't understand how to do it yet, and I would like to know if I am at least conceptualizing things correctly regardless of whether my particular application happens to be bottlenecked here.<\/p>\n","answers":[{"AnswerId":"41410965","CreationDate":"2016-12-31T18:55:22.607","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Is relatively important, because the python wrapper has to convert from the C++ tensor memory to the python numpy array, thus doing a copy operation. If you are fetching big Variables the time could be not trivial, think of 1GB having to be converted each time.<\/p>\n"}]}
{"QuestionId":41411846,"AnswerCount":0,"Tags":"<python><json><python-2.7><tensorflow>","CreationDate":"2016-12-31T21:32:06.927","AcceptedAnswerId":null,"OwnerUserId":6687172.0,"Title":"How to solve ImportError: cannot import name json_format in Python","Body":"<p>was running a python\/tensorflow program that doesn't even require json, when this message popped up - i tried doing pip install json_format, which worked fine, but did not solve the issue. This started after I implemented \"import tensorflow as tf\" in my code to fix another issue I was having where tf wasn't being recognized. I'm running on python 2.7 on a linux-ubuntu system.<\/p>\n\n<p>Any help is appreciated. Below is the full error text:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41412290,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning><supervised-learning>","CreationDate":"2016-12-31T23:08:59.230","AcceptedAnswerId":null,"OwnerUserId":7361548.0,"Title":"create a dataset that has the same format as the cifar-10 data set","Body":"<p>I want to create a dataset that has the same format as the cifar-10 data set to use with Tensorflow. It should have images and labels. Basically, I'd like to be able to take the cifar-10 code but different images and labels, and run that code. I haven't found any information on how to do this online, and am completely new to machine learning.<\/p>\n","answers":[{"AnswerId":"41412542","CreationDate":"2017-01-01T00:24:32.127","ParentId":null,"OwnerUserId":"1164106","Title":null,"Body":"<p>The CIFAR-10 is a subset of a much larger <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow noreferrer\">dataset<\/a>. The images you need are scaled color images that have a height and width of 32 pixels with three color channels.  One approach toward your goal would be to start by selecting 10 different labels from the CIFAR-100 dataset, saving your  and running the existing code.  For example, you may want to select the vehicles 1 and vehicles 2 superclasses.  This would give you 6000 labeled images covering: bicycle, bus, motorcycle, pickup truck, train, lawn-mower, rocket, streetcar, tank, and tractor classes. You could then build a predictor of vehicle type - a pretty cool way to get more familiar with machine learning. :- )<\/p>\n\n<p>Inside the cifar10.py file you can see the directories that are used for the training file downloaded from '<a href=\"http:\/\/www.cs.toronto.edu\/~kriz\/cifar-10-binary.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/www.cs.toronto.edu\/~kriz\/cifar-10-binary.tar.gz<\/a>'.  Without changing any code you can simply update these inflated training files with your data.  Take a look in the \/tmp\/cifar10_data\/cifar-10-batches-bin directory.  E.g. The batches.meta.txt file contains the labels as described under the section 'Binary version' here: <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow noreferrer\">https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html<\/a><\/p>\n"},{"AnswerId":"41412551","CreationDate":"2017-01-01T00:28:18.903","ParentId":null,"OwnerUserId":"7352138","Title":null,"Body":"<p>I had to do this also, and made a bunch of functions to format images and a text file into a readable format for tensorflow. Here is the modifications I made to use a group of images in a folder called images (I used glob to iterate through them) and a text file with the information about the images encoded (I had a series of numbers for each image, where the numbers described where the user was directing the robot at the time each image was taken). I made a function to generate minibatches, and to create a training and test data set. I also converted the numbers I associated with each image into one-hot vectors to fit in (you can use this if you want, but may not be useful). <\/p>\n\n<pre><code>#!\/usr\/bin\/python\nimport cv2\nimport numpy as np\nimport tensorflow as tf\nimport glob\nimport re\nimport random\n\n\n# Parameters\nlearning_rate = 0.001\ntraining_iters = 20000\nbatch_size = 120\ndisplay_step = 10\n\n# Network Parameters\nn_input = 784 # MNIST data input (img shape: 28*28)\nn_classes = 1 # MNIST total classes (0-9 digits)\ndropout = 0.75 # Dropout, probability to keep units\n\nimage = np.reshape(np.asarray(mnist.train.images[0]), (28,28))\n\n#Process Images\n\ncv_img = []\nfor img in glob.glob(\".\/images\/*.jpeg\"):\n    n  = cv2.cvtColor(cv2.resize(cv2.imread(img), (28,28)), cv2.COLOR_BGR2GRAY)\n    n = np.asarray(n)\n    n = np.reshape(n, n_input)\n    cv_img.append(n)\n\n#Process File for angle, here we read the text line by line and make a list\nwith open(\".\/images\/allinfo.txt\") as f:\n    content = f.readlines()\n\n#Initialize arrays to unpack data file\nangle = []\nimage_number = []\n\n\n#Iterate through the text list and split each one by the comma separating the values. \n#Turn the text into floats for use in the network\nfor i in range(len(content)):\n    content[i] = content[i][:-1].split(',')\n    image_number.append(float(content[i][1]))\n    angle.append(float(content[i][7]))\n\n#Divide both angle and image number into test and train data sets\nangle = np.atleast_2d(angle).T\n\n\n##Encode angle into 10 classes (it ranges -1 to 1)\nfor i in range(len(angle)):\n    angle[i] = random.uniform(-1,1)\n    angle[i] = int((angle[i]+1.0)*n_classes\/2.)\n\n\n#Create a one-hot version of angle\nangle_one_hot = np.zeros((len(angle),n_classes))\n\nfor c in range(len(angle)):\n    one_hot = np.zeros(n_classes)\n    one_hot[int(angle[c])] = 1\n    angle_one_hot[c] = one_hot\n\n\nimage_number = np.atleast_2d(image_number).T\ntest_data =  np.hstack((image_number, angle))\n#print test_data\ntrain_percent = .8\ntrain_number = int(len(test_data)*train_percent)\ntrain_data = np.zeros((train_number, 2))\nfor i in range(train_number):\n    rand = random.randrange(0,len(test_data))\n    train_data[i] = test_data[rand]\n    test_data = np.delete(test_data, rand, 0)\ntest_data_images = test_data[:,0]\ntest_data_angles = test_data[:,1]\ntrain_data_images, train_data_angles = train_data[:,0], train_data[:,1]\n\n\n\ndef gen_batch(angles, images, batch_size, image_array=cv_img):\n    indices = random.sample(xrange(0,len(images)), batch_size)\n    batch_images = []\n    batch_angles = []\n #   print angles\n    for i in range(batch_size):\n        batch_images.append(image_array[int(images[indices[i]])][:])\n        batch_angles.append(angles[indices[i]])\n    batch_images = np.asarray(batch_images)\n    batch_angles = np.asarray(batch_angles)\n\n    return batch_images, batch_angles\n\n\n# tf Graph input\nx = tf.placeholder(tf.float32, [None, n_input])\ny = tf.placeholder(tf.float32)\nkeep_prob = tf.placeholder(tf.float32) #dropout (keep probability)\n\n# Create some wrappers for simplicity\ndef conv2d(x, W, b, strides=1):\n    # Conv2D wrapper, with bias and relu activation\n    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')\n    x = tf.nn.bias_add(x, b)\n    return tf.nn.relu(x)\n\n\ndef maxpool2d(x, k=2):\n    # MaxPool2D wrapper\n    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],\n                          padding='SAME')\n\n\n# Create model\ndef conv_net(x, weights, biases, dropout):\n    # Reshape input picture\n    x = tf.reshape(x, shape=[-1, 28, 28, 1])\n\n    # Convolution Layer\n    conv1 = conv2d(x, weights['wc1'], biases['bc1'])\n    # Max Pooling (down-sampling)\n    conv1 = maxpool2d(conv1, k=2)\n\n    # Convolution Layer\n    conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])\n    # Max Pooling (down-sampling)\n    conv2 = maxpool2d(conv2, k=2)\n\n    # Fully connected layer\n    # Reshape conv2 output to fit fully connected layer input\n    fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])\n    fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])\n    fc1 = tf.nn.relu(fc1)\n    # Apply Dropout\n    fc1 = tf.nn.dropout(fc1, dropout)\n\n    # Output, class prediction\n    out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])\n    return out\n\n# Store layers weight &amp; bias\nweights = {\n    # 5x5 conv, 1 input, 32 outputs\n    'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32])),\n    # 5x5 conv, 32 inputs, 64 outputs\n    'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64])),\n    # fully connected, 7*7*64 inputs, 1024 outputs\n    'wd1': tf.Variable(tf.random_normal([7*7*64, 1024])),\n    # 1024 inputs, 10 outputs (class prediction)\n    'out': tf.Variable(tf.random_normal([1024, n_classes]))\n}\n\nbiases = {\n    'bc1': tf.Variable(tf.random_normal([32])),\n    'bc2': tf.Variable(tf.random_normal([64])),\n    'bd1': tf.Variable(tf.random_normal([1024])),\n    'out': tf.Variable(tf.random_normal([n_classes]))\n}\n\n# Construct model\npred = conv_net(x, weights, biases, keep_prob)\n\n# Define loss and optimizer\n#cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y))\ncost = tf.reduce_mean(pred)\noptimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize((pred-y)**2)\n\n# Evaluate model\ncorrect_pred = y\naccuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\n# Initializing the variables\ninit = tf.initialize_all_variables()\n\n# Launch the graph\nwith tf.Session() as sess:\n    sess.run(init)\n    step = 1\n    print(y)\n    # Keep training until reach max iterations\n    while step * batch_size &lt; training_iters:\n        batch_x, batch_y = gen_batch(train_data_angles, train_data_images, batch_size)\n        #cv2.imshow('trash', batch_x[0,:].reshape((28,28)))\n        #cv2.waitKey(0)\n        #print(batch_y)\n        # Run optimization op (backprop)\n        sess.run(optimizer, feed_dict={x: batch_x, y: batch_y,\n                                       keep_prob: dropout})\n        if step % display_step == 0:\n            # Calculate batch loss and accuracy\n            loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x,\n                                                              y: batch_y,\n                                                              keep_prob: 1.})\n            print \"Iter \" + str(step*batch_size) + \", Minibatch Loss= \" + \\\n                  \"{:.6f}\".format(loss) + \", Training Accuracy= \" + \\\n                  \"{:.5f}\".format(acc)\n        step += 1\n    print \"Optimization Finished!\"\n\n    # Calculate accuracy for all test images\n    img, lbls = gen_batch(test_data_angles, test_data_images, len(test_data_angles))\n    print \"Testing Accuracy:\", \\\n        sess.run(accuracy, feed_dict={x: img,\n                                      y: lbls,\n                                      keep_prob: 1.})\n<\/code><\/pre>\n\n<p>This doesn't work as a good nn (the data isn't normalized, the learning rate is two high, and the training accuracy is not yet programmed) but the image processing code works. <\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":41412335,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2016-12-31T23:22:52.520","AcceptedAnswerId":41413215.0,"OwnerUserId":5605365.0,"Title":"tf.nn.softmax_cross_entropy_with_logits() error: logits and labels must be same size","Body":"<p>I am new to TensorFlow and am trying to write an algorithm to classify images in the CIFAR-10 dataset. I am getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>I'm pretty sure this means that in line 48 (shown above)  and  are not the same shape, but I don't understand TensorFlow well enough to know how to fix it. I don't even really understand where  is being set, I got most of this code from a tutorial and fiddled with it to apply it to a different dataset. How can I fix this error?<\/p>\n","answers":[{"AnswerId":"41413215","CreationDate":"2017-01-01T03:50:13.987","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.softmax_cross_entropy_with_logits(logits, labels)<\/code><\/a> op expects its <code>logits<\/code> and <code>labels<\/code> arguments to be tensors with the same shape. Furthermore, the <code>logits<\/code> and <code>labels<\/code> arguments should be 2-D tensors (matrices) with <code>batch_size<\/code> rows, and <code>num_classes<\/code> columns.<\/p>\n\n<p>From the error message and the size of <code>logits<\/code>, I'm guessing that <code>batch_size<\/code> is 10000, and <code>num_classes<\/code> is 10. From the size of <code>labels<\/code>, I'm guessing that your labels are encoded as a list of integers, where the integer represent the index of the class for the corresponding input example. (I'd have expected this to be a <code>tf.int32<\/code> value, rather than <code>tf.float32<\/code> as it appears to be in your program, but perhaps there is some automatic conversion going on.)<\/p>\n\n<p>In TensorFlow, you can use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#sparse_softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.sparse_softmax_cross_entropy_with_logits()<\/code><\/a> to compute cross-entropy on data in this form. In your program, you could do this by replacing the <code>cost<\/code> calculation with:<\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(\n    prediction, tf.squeeze(y)))\n<\/code><\/pre>\n\n<p>Note that the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/shapes_and_shaping#squeeze\" rel=\"nofollow noreferrer\"><code>tf.squeeze()<\/code><\/a> op is needed to convert <code>y<\/code> into a vector of length <code>batch_size<\/code> (in order to be a valid argument to <code>tf.nn.sparse_softmax_cross_entropy_with_logits()<\/code>.<\/p>\n"},{"AnswerId":"44350769","CreationDate":"2017-06-04T04:29:06.737","ParentId":null,"OwnerUserId":"3902230","Title":null,"Body":"<p>Here are some updates to the code to support TensorFlow version 1.0: <\/p>\n\n<pre><code>def train_neural_network(x): &lt;br&gt;\n&amp;emsp;prediction = neural_network_model(x) &lt;br&gt;\n&amp;emsp;# OLD VERSION: &lt;br&gt;\n&amp;emsp;cost = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits(prediction,y) ) &lt;br&gt;\n&amp;emsp;# NEW:&lt;br&gt; \n&amp;emsp;cost = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y) )&lt;br&gt;\n&amp;emsp;optimizer = tf.train.AdamOptimizer().minimize(cost) &lt;br&gt;\n&amp;emsp;hm_epochs = 10 &lt;br&gt;\n&amp;emsp;with tf.Session() as sess: &lt;br&gt;\n&amp;emsp;&amp;emsp;#OLD: #sess.run(tf.initialize_all_variables()) &lt;br&gt;\n&amp;emsp;&amp;emsp;#NEW:&lt;br&gt; \n&amp;emsp;&amp;emsp;sess.run(tf.global_variables_initializer())\n<\/code><\/pre>\n"}]}
{"QuestionId":41412514,"AnswerCount":0,"Tags":"<python><neural-network><keras><lstm>","CreationDate":"2017-01-01T00:18:04.793","AcceptedAnswerId":null,"OwnerUserId":7361586.0,"Title":"keras nb_samples, timesteps, input_dim","Body":"<p>I am working on my first LSTM in keras. I intend to train a NN to generate piano music. I have input data (notes) which I encode with integers. The input data after reading from the text files and encoding looks like this for a single piece of music.<\/p>\n\n<p><strong>TimeStep'\\t'Voice1'\\t'Voice2'\\t'Voice3'\\t'Voice4<\/strong><\/p>\n\n<p>1'\\t'33'\\t'55'\\t'97'\\t'12<\/p>\n\n<p>2'\\t'71'\\t'26'\\t'83'\\t'42<\/p>\n\n<p>3\n....<\/p>\n\n<p>I have several pieces of music with different number of TimeStep (usually hundreds of time steps in each piece of music). Each piece can have different number of voices (columns). I was thinking of just combining all files together into one giant sequence of music (time step) and padding missing voices with 0 in order to train the network (all pieces are from the same composer, so I hope to capture the essence of this particular composer with this approach). \nAny thoughts on this approach are highly appreciated :)<\/p>\n\n<p>How does this data best translate into the keras input format (nb_samples, timesteps, input_dim) and output format?<\/p>\n\n<p>Thanks a lot for your help!<\/p>\n\n<p>Michael<\/p>\n","answers":[]}
{"QuestionId":41412641,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-01T00:54:41.037","AcceptedAnswerId":null,"OwnerUserId":6636753.0,"Title":"tensorflow program getting error saying KeyError: 'enc_vocab_size'","Body":"<p>I was working on a Tensorflow program when I got an error saying\nI tensorflow\/core\/common_runtime\/direct_session.cc:255] Device mapping:<\/p>\n\n<pre><\/pre>\n\n<p>I could not figure out what to do to fix this The code is bellow please tell me.\n<a href=\"https:\/\/raw.githubusercontent.com\/llSourcell\/tensorflow_chatbot\/master\/execute.py\" rel=\"nofollow noreferrer\">Like of code I was using<\/a><\/p>\n","answers":[]}
{"QuestionId":41412860,"AnswerCount":0,"Tags":"<debugging><machine-learning><keras>","CreationDate":"2017-01-01T01:54:29.540","AcceptedAnswerId":null,"OwnerUserId":56349.0,"Title":"Keras: Low Accuracy on Linear Regression","Body":"<p>I'm grid searching and yielding very low accuracy, and looking to understand what it is I need to fix, to suss out a correlation.  My gut tells me the either the data itself isn't highly correlated based on the features I've selected, or that my inputs are garbage, or that my choice in model is incorrect.<\/p>\n\n<p>I thought I'd outline what I have here in case anyone spots something I might've missed. Let me know if more information is required.<\/p>\n\n<h2>Model<\/h2>\n\n<pre><\/pre>\n\n<h2>Grid Search<\/h2>\n\n<p>And I'm yielding accuracy of around 0.0105 despite grid searching across a variety of options:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41413582,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-01T05:32:02.567","AcceptedAnswerId":41413923.0,"OwnerUserId":5605365.0,"Title":"Tensorflow InvalidArgumentError (see above for traceback): Minimum tensor rank: 2 but got: 1","Body":"<p>I am new to TensorFlow and am trying to write an algorithm to classify images in the CIFAR-10 dataset. I am getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the traceback:<\/p>\n\n<pre><\/pre>\n\n<p>I've marked above where the error is occurring. It is on the line that says . Why am I getting this and how can I fix it? <\/p>\n","answers":[{"AnswerId":"41413923","CreationDate":"2017-01-01T07:04:01.933","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Quoting from the official TensorFlow documentation on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/sequence_comparison_and_indexing#argmax\" rel=\"nofollow noreferrer\"><code>tf.argmax<\/code><\/a>,<\/p>\n\n<blockquote>\n  <p>axis: A Tensor. Must be one of the following types: int32, int64. int32, 0 &lt;= axis &lt; rank(input). Describes which axis of the input Tensor to reduce across. For vectors, use axis = 0.<\/p>\n<\/blockquote>\n\n<p>You are getting this error because one of the <code>argmax()<\/code> have an <code>input<\/code> with rank <code>&lt;=1<\/code>. Since you are passing <code>axis=1<\/code>, you need to pass tensors having rank > 1 to get a valid output.<\/p>\n\n<hr>\n\n<p>Looking  more closely at your code, it seems like this is becuase of <code>tf.argmax(y, 1)<\/code>. Try to pass <code>0<\/code> or <code>None<\/code> instead of <code>1<\/code>.<\/p>\n"}]}
{"QuestionId":41413589,"AnswerCount":0,"Tags":"<python-2.7><numpy><keras>","CreationDate":"2017-01-01T05:35:43.817","AcceptedAnswerId":null,"OwnerUserId":6318662.0,"Title":"How to train a RNN with variable length MFCC Data?","Body":"<p>Im new to Python and Keras.<\/p>\n\n<p>Im trying to train a Recurrent network with MFCC data for each audio file having variable length of features. Meaning first MFCC file will have a MFCC matrix of (1454,13) and second MFCC file will have a MFCC matrix of (2512,13).<\/p>\n\n<p>I have written following script to load data required for training.<\/p>\n\n<pre><\/pre>\n\n<p>I have loaded all the MFCC data into 3D lists of list because I cannot find a way to create a 3D numpy array.<\/p>\n\n<p>Following Script was used for building the RNN and Train it. <\/p>\n\n<pre><\/pre>\n\n<p>Once I run the script for Training the RNN. I get the following error: <\/p>\n\n<pre><\/pre>\n\n<p>I am assuming that this is a issue of loading data to x_train. Please can some one help me with solving this issue. <\/p>\n\n<p>Thank you in Advance :) <\/p>\n","answers":[]}
{"QuestionId":41414305,"AnswerCount":1,"Tags":"<python><tensorflow><classification><pixel><tflearn>","CreationDate":"2017-01-01T08:38:50.123","AcceptedAnswerId":null,"OwnerUserId":2148425.0,"Title":"Wrong output of prediction function in tensorflow","Body":"<p>I am going to perform pixel-based classification on an image. Here is the code I used for training the NN <\/p>\n\n<pre><\/pre>\n\n<p>The problem is that after training the model, the result of p.array(model.predict(x_test)) is 1 only although I expected this to be either 2 or 3. In one example where I have had 4 classes of objects and I expected the result of that command to be a label between 2 and 5 (note:y_train has int values between 2 and 5) but again the output of the prediction function is 1. Could that be a problem of training phase?<\/p>\n","answers":[{"AnswerId":"41414983","CreationDate":"2017-01-01T11:03:57.333","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>The <code>None<\/code> parameter is used to denote <em>different<\/em> training examples. In your case, each image has a total of <code>28*28*4<\/code> parameters, due to the custom four channel dataset you are using.<\/p>\n\n<p>To make this LSTM work, you should try to do the following -<\/p>\n\n<pre><code>X = np.reshape(X, (-1, 28, 28, 4))\ntestX = np.reshape(testX, (-1, 28, 28, 4))\n\nnet = tflearn.input_data(shape=[None, 28, 28, 4])\n<\/code><\/pre>\n\n<p>Of course, <em>(this is very important)<\/em>, make sure that <code>reshape()<\/code> puts the four different channels corresponding to a single pixel in the last dimension of the numpy array, and the <code>28, 28<\/code> correspond to pixels in a single image.<\/p>\n\n<hr>\n\n<p>In case your images don't have dimension <code>28*28<\/code>, adjust those parameters accordingly.<\/p>\n"}]}
{"QuestionId":41414359,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><computer-vision><convolution><cross-validation>","CreationDate":"2017-01-01T08:55:54.337","AcceptedAnswerId":null,"OwnerUserId":3996266.0,"Title":"How to use cross-validation method in Tensor-Flow","Body":"<p>I'm trying to use cross-validation in Tensor-flow . i write this semi code as my algorithm <\/p>\n\n<pre><\/pre>\n\n<p>but i don't know how to combine Cross-validation and tensor-flow training parameters in one algorithm.<\/p>\n\n<p>is method is correct ? or\ncan anyone correct this method ?<\/p>\n","answers":[]}
{"QuestionId":41414489,"AnswerCount":9,"Tags":"<python><theano><keras>","CreationDate":"2017-01-01T09:23:22.720","AcceptedAnswerId":null,"OwnerUserId":996366.0,"Title":"How to uninstall Keras?","Body":"<p>I have installed Keras using this command:<\/p>\n\n<pre><\/pre>\n\n<p>It installed properly and worked fine until I tried to import application modules:<\/p>\n\n<pre><\/pre>\n\n<p>I came across <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/3OpfQHzCk64\" rel=\"noreferrer\">this link which recommends<\/a> to uninstall Keras and directly install Keras from GitHub:<\/p>\n\n<pre><\/pre>\n\n<p>Before reinstalling Keras from GitHub, I tried to unistall Keras using this command but it throws this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"55676595","CreationDate":"2019-04-14T14:52:27.783","ParentId":null,"OwnerUserId":"10911678","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>sudo pip uninstall git+https:\/\/github.com\/fchollet\/keras.git\n<\/code><\/pre>\n\n<p>If that does not work, what you can see after <code>sudo pip list<\/code>?<\/p>\n"},{"AnswerId":"54851408","CreationDate":"2019-02-24T11:26:59.713","ParentId":null,"OwnerUserId":"9239113","Title":null,"Body":"<p>In Anaconda prompt type<\/p>\n\n<pre><code>conda uninstall keras\n<\/code><\/pre>\n"},{"AnswerId":"55272559","CreationDate":"2019-03-21T01:36:14.580","ParentId":null,"OwnerUserId":"10964645","Title":null,"Body":"<p>you can  uninstall from conda by using this command.\n<code>conda uninstall keras<\/code><\/p>\n"},{"AnswerId":"49510477","CreationDate":"2018-03-27T10:27:48.613","ParentId":null,"OwnerUserId":"243072","Title":null,"Body":"<p>uninstall keras:<\/p>\n\n<pre><code>pip uninstall keras -y                                                                                                                                                                            \n<\/code><\/pre>\n\n<p>reinstall with specific version<\/p>\n\n<pre><code>pip install -Iv keras==2.1.4 \n<\/code><\/pre>\n"},{"AnswerId":"53486423","CreationDate":"2018-11-26T17:43:36.587","ParentId":null,"OwnerUserId":"4970438","Title":null,"Body":"<p>Please also check the Anaconda environment if you are using any<\/p>\n"},{"AnswerId":"41435526","CreationDate":"2017-01-03T02:15:03.103","ParentId":null,"OwnerUserId":"996366","Title":null,"Body":"<p>I followed <a href=\"https:\/\/stackoverflow.com\/a\/5959706\/996366\">this method<\/a> to solve my problem<\/p>\n\n<pre><code>&gt;&gt; import keras\n&gt;&gt; keras.__path__\n['\/usr\/local\/lib\/python2.7\/dist-packages\/keras']\n\n#copy the path\n&gt;&gt; sudo  rm -r \/usr\/local\/lib\/python2.7\/dist-packages\/keras\n\n#re installation\n&gt;&gt; sudo pip install git+https:\/\/github.com\/fchollet\/keras.git\n<\/code><\/pre>\n"},{"AnswerId":"44491027","CreationDate":"2017-06-12T03:29:44.833","ParentId":null,"OwnerUserId":"6596664","Title":null,"Body":"<p>You can simply do that with\n<code>\nsudo pip -U install keras\n<\/code><\/p>\n"},{"AnswerId":"44536160","CreationDate":"2017-06-14T05:31:19.813","ParentId":null,"OwnerUserId":"1914543","Title":null,"Body":"<p>I think you really want to do is not to uninstall the keras, but use the <code>keras.applications.vgg16<\/code>.<br>\nThe following checklist will help you to clarify the issue.<\/p>\n\n<ul>\n<li><strong>Keras version<\/strong>: confirm the version of the keras is latest (now 2.0.5)<\/li>\n<li><strong>Backend<\/strong>: Theano\/Tensorflow or the other<\/li>\n<li><strong>Device<\/strong>: GPU or CPU<\/li>\n<li><strong>Python version<\/strong>: 2 or 3 and use Anaconda or not<\/li>\n<li><strong>Operating system<\/strong>: Mac, Windows, Linux, and so on<\/li>\n<\/ul>\n\n<p>I recommend to check <strong>latest keras, TensorFlow backend(CPU mode), Python3<\/strong> setting at first. It will solve various problems.<\/p>\n"},{"AnswerId":"46440188","CreationDate":"2017-09-27T05:36:13.387","ParentId":null,"OwnerUserId":"5424617","Title":null,"Body":"<p>You can simply try from the following command:<\/p>\n\n<pre><code>pip uninstall keras\n<\/code><\/pre>\n"}]}
{"QuestionId":41414867,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-01T10:43:16.013","AcceptedAnswerId":41415193.0,"OwnerUserId":5605365.0,"Title":"Tensorflow: You must feed a value for placeholder tensor 'Placeholder' with dtype float [but the value is a float]","Body":"<p>I am going through a tensorflow tutorial and keep getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the traceback:<\/p>\n\n<pre><\/pre>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see here, both x_data and y_data are floats, so I don't get why I'm getting this error.<\/p>\n\n<p>When I add a print statement  just before . I get this:  So clearly they are both floats<\/p>\n\n<p>What am I missing here? <\/p>\n","answers":[{"AnswerId":"41415193","CreationDate":"2017-01-01T11:44:09.037","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>You have an error in the <code>print()<\/code> statement. Luckily, it was picked up for me as a <code>SyntaxError<\/code> as I use TensorFlow r0.11.<\/p>\n\n<p>Replace,<\/p>\n\n<pre><code>print(sess.run(loss), feed_dict={xs:x_data, ys:y_data})\n<\/code><\/pre>\n\n<p>with,<\/p>\n\n<pre><code>print(sess.run(loss, feed_dict={xs:x_data, ys:y_data}))\n<\/code><\/pre>\n"}]}
{"QuestionId":41415054,"AnswerCount":0,"Tags":"<python><csv><tensorflow><deep-learning><lstm>","CreationDate":"2017-01-01T11:18:36.827","AcceptedAnswerId":null,"OwnerUserId":5174951.0,"Title":"Error in loading files from a csv while training an LSTM","Body":"<p>I am relatively new to tensorflow and am trying to train a two layer LSTM in tensorflow for some data I crawled, and saved into a CSV. But, after I use the methods shows on the tensorflow website, I keep getting the following error: <\/p>\n\n<blockquote>\n  <p>TypeError: inputs must be a sequence<\/p>\n<\/blockquote>\n\n<p>The original code is:<\/p>\n\n<pre><\/pre>\n\n<p>The last line is what causes the error. I think I understand what the problem is, but I am not sure how I would go about fixing it<\/p>\n","answers":[]}
{"QuestionId":41415195,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-01T11:44:30.733","AcceptedAnswerId":null,"OwnerUserId":5350430.0,"Title":"python tensorflow: nan loss using softmax_cross_entropy","Body":"<p>I use the pretrained parameter to my vgg net(except last three layers). And I only train the last three layers of my vgg net.<\/p>\n\n<p>I noticed that the preprocess of image of vgg net should minus the mean value of RGB. However, it seems needn't to convert the pixels value to [-1,1].<\/p>\n\n<p>I use softmax_cross_entropy to get my loss. The logit is the last output of vgg net. The last layer of vgg is conv(actually fc). And the label is one hot label.<\/p>\n\n<p>I've tried a lot of configure of the learning rate, eg 0.1 0.03 0.01 0.003 0.001. However, the first three learning rate lead to the nan loss. The last two configurations lead to unstable loss. Should I still decrease the learning rate? I noticed that for many examples values such as 0.1 or 0.01 are good to go.\n   Please help me. Here's my code.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41415268,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-01T11:57:28.573","AcceptedAnswerId":41415367.0,"OwnerUserId":6467567.0,"Title":"Operations on python datatypes","Body":"<p>Im new to python and tensorflow. How can I check what operations can be done on a variable ? Im trying to store images into a list. Id like to find out its size just to confirm.<\/p>\n\n<p>This is the code taken from the website.<\/p>\n\n<pre><\/pre>\n\n<p>I got an error below. This is the website for the function string_input_producer <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#string_input_producer\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#string_input_producer<\/a> it does say it returns a list so im not sure what im doing wrong<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41415367","CreationDate":"2017-01-01T12:14:10.873","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>As mentioned in the comments, you should use <code>dir(filename_queue)<\/code> to check the set of functions in an object. This outputs the following -<\/p>\n\n<pre><code>['__class__', '__delattr__', '__dict__', '__doc__', '__format__', \n '__getattribute__', '__hash__', '__init__', '__module__', '__new__',\n '__reduce__', '__reduce_ex__', '__repr__', '__setattr__', \n '__sizeof__', '__str__', '__subclasshook__', '__weakref__', \n '_check_enqueue_dtypes', '_dequeue_return_value', '_dtypes',\n '_name', '_names', '_queue_ref', '_scope_vals', '_shapes',\n 'close', 'dequeue', 'dequeue_many', 'dequeue_up_to', 'dtypes',\n 'enqueue', 'enqueue_many', 'from_list', 'name', 'names',\n 'queue_ref', 'shapes', 'size']\n<\/code><\/pre>\n\n<p>Make use of the <code>size()<\/code> function to get your result. You can read more about <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/queues#FIFOQueue\" rel=\"nofollow noreferrer\"><code>FIFOQueue<\/code><\/a>.<\/p>\n"}]}
{"QuestionId":41415527,"AnswerCount":2,"Tags":"<python><anaconda><python-import><keras>","CreationDate":"2017-01-01T12:39:37.120","AcceptedAnswerId":null,"OwnerUserId":6310830.0,"Title":"Installing Keras package with conda install","Body":"<p>I have installed  package on a server as a user account, then I installed  by ,but after installation, when I run , it raised ,anyone can help? thanks very much!<\/p>\n","answers":[{"AnswerId":"41415759","CreationDate":"2017-01-01T13:18:49.097","ParentId":null,"OwnerUserId":"837534","Title":null,"Body":"<p>One solution could be creating a conda environment:<\/p>\n\n<pre><code>conda create -n keras python=3.5\n<\/code><\/pre>\n\n<p>Now activate it:<\/p>\n\n<pre><code>conda activate keras\n<\/code><\/pre>\n\n<p>and install keras:<\/p>\n\n<pre><code>(keras)$ conda install keras\n<\/code><\/pre>\n\n<p>Try if it works:<\/p>\n\n<pre><code>(keras)$ python\n&gt;&gt;&gt; import keras\n<\/code><\/pre>\n"},{"AnswerId":"45798962","CreationDate":"2017-08-21T13:56:27.640","ParentId":null,"OwnerUserId":"1936946","Title":null,"Body":"<p>Once creating the Conda environment, use the command below to list available environment:<\/p>\n\n<pre><code>conda info -e\n<\/code><\/pre>\n\n<p>Once activate your conda environment, you can try to install Keras by <\/p>\n\n<pre><code>pip install keras==version_your_desired.\n<\/code><\/pre>\n\n<p>Then use <\/p>\n\n<pre><code>pip freeze\n<\/code><\/pre>\n\n<p>to check the version. <\/p>\n"}]}
{"QuestionId":41415629,"AnswerCount":4,"Tags":"<python><tensorflow><keras><neural-network>","CreationDate":"2017-01-01T12:58:24.367","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"ImportError: No module named 'tensorflow.python'","Body":"<p>here i wanna run this code for try neural network with python :<\/p>\n\n<pre><\/pre>\n\n<p>at the beginning it must install keras, and success. but when try to run the code at the first the error is :<\/p>\n\n<blockquote>\n  <p>ImportError : No Moduled Name \"tensorflow\"<\/p>\n<\/blockquote>\n\n<p>then i install using pip :<\/p>\n\n<blockquote>\n  <p>pip install tensorflow<\/p>\n<\/blockquote>\n\n<p>after installation i try to run code again, got another message like this :<\/p>\n\n<blockquote>\n  <p>ImportError : No Moduled Name \"tensorflow.python\"<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/tYgfR.png\" rel=\"noreferrer\">Message Error<\/a>\ni dont have any idea with the error<\/p>\n","answers":[{"AnswerId":"45767511","CreationDate":"2017-08-19T04:00:04.307","ParentId":null,"OwnerUserId":"8486447","Title":null,"Body":"<p>Uninstall tensorflow:<\/p>\n\n<pre><code>pip uninstall tensorflow\n<\/code><\/pre>\n\n<p>Then reinstall it:<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"53421008","CreationDate":"2018-11-21T21:55:25.407","ParentId":null,"OwnerUserId":"7870598","Title":null,"Body":"<p>I have the same problem in Windows 10. Until now I don't know why.<\/p>\n\n<p>But if I create an virtual environment<\/p>\n\n<p><code>cd &lt;your project path&gt;<\/code><\/p>\n\n<p>Install virtualenv<\/p>\n\n<p><code>pip install virtualenv<\/code><\/p>\n\n<p>Create the virtual environment<\/p>\n\n<p><code>virtualenv &lt;envname&gt;<\/code><\/p>\n\n<p>Activate the env<\/p>\n\n<ul>\n<li>Windows Powershell: <code>.\\&lt;envname&gt;\\Scripts\\activate<\/code><\/li>\n<li>Unix with Bash or zsh: <code>source &lt;envname&gt;\/bin\/activate<\/code><\/li>\n<\/ul>\n\n<p>Then now you install tensorflow<\/p>\n\n<p><code>(&lt;envname&gt;) $ pip install tensorflow<\/code><\/p>\n\n<p>And then run Hello World successfully.<\/p>\n\n<p>*Don't forget that you need to activate or configure everytime the virtual environment jupyter, command-line, etc.<\/p>\n"},{"AnswerId":"41418198","CreationDate":"2017-01-01T19:10:40.910","ParentId":null,"OwnerUserId":"6906747","Title":null,"Body":"<p>Open a python shell and type:   <\/p>\n\n<pre><code>help('modules')   \n<\/code><\/pre>\n\n<p>This will gather a list of all available modules. tensor flow should not show up, as it is not installed correctly (according to the <code>traceback<\/code>).<br>\nThen:<\/p>\n\n<pre><code>import sys\nsys.path()\n<\/code><\/pre>\n\n<p>This will give you a list of system paths where modules can be installed. If there is a known issue with installing a module, I recommend moving the files manually to the right system path.<br>\nThe system path depends on the OS you are using, so without knowing that I can't tell you where to move it.<br>\nBut <code>sys.path()<\/code> can!<\/p>\n\n<p>Hope I could help,\nNarusan<\/p>\n"},{"AnswerId":"45698541","CreationDate":"2017-08-15T17:48:48.117","ParentId":null,"OwnerUserId":"7625684","Title":null,"Body":"<p>try to change the actual running python directory.\nand make sure that running python directory is not where you downloaded tensorflow. else go to any other directory and you're fine.\ni hope that solves your probleme.<\/p>\n"}]}
{"QuestionId":41415926,"AnswerCount":3,"Tags":"<python><python-2.7><ubuntu><numpy><tensorflow>","CreationDate":"2017-01-01T13:45:39.720","AcceptedAnswerId":41417733.0,"OwnerUserId":null,"Title":"Tensorflow (CPU only) installation Error | Ubuntu 14.10 | numpy 1.8.2","Body":"<p>I'm in the process of installing Tensorflow in my Ubuntu 14.10. I've previously installed <strong>numpy<\/strong>, <strong>scipy<\/strong>, <strong>sklearn<\/strong> and also <strong>ipython-notebook<\/strong>. I followed the official documentation at Tensorflow.org (<a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup\" rel=\"nofollow noreferrer\">pip installation<\/a>).The location of my tensorflow installation directory is <strong>\/usr\/local\/lib\/python2.7\/dist-packages<\/strong>. I'm getting following error when I execute  in my Ubuntu terminal. <\/p>\n\n<pre><\/pre>\n\n<p>According to the error, it has to be with numpy. I tried to upgrade numpy using command  but this doesn't seem to upgrade my existing numpy version and the problem still remains. <\/p>\n\n<p>Any suggestions in solving this issue ?<\/p>\n","answers":[{"AnswerId":"41417592","CreationDate":"2017-01-01T17:48:25.957","ParentId":null,"OwnerUserId":"4313064","Title":null,"Body":"<p>Did you follow the instructions?<\/p>\n\n<blockquote>\n  <p>please exit the tensorflow source tree, and relaunch your python\n  interpreter from there.<\/p>\n<\/blockquote>\n"},{"AnswerId":"42593665","CreationDate":"2017-03-04T08:00:40.677","ParentId":null,"OwnerUserId":"3091286","Title":null,"Body":"<p>I ran into the same problem on OS X (10.12).  Like you, I got caught in a loop where two versions of python didn't match but reinstalling wasn't fixing the mismatch.  Eventually, a more thorough removal and reinstallation worked: <\/p>\n\n<ol>\n<li>upgrade pip (<code>sudo pip install --upgrade pip<\/code>)<\/li>\n<li>delete the ~\/tensorflow directory and its contents<\/li>\n<li>reinstall virtualenv and activate it as per instructions<\/li>\n<li>create a virtualenv environment (<code>virtualenv --system-site-packages ~\/tensorflow<\/code>)<\/li>\n<li>reinstall tensorflow, using sudo to deal with a permissions issue<br>\n(<code>sudo pip install --upgrade tensorflow<\/code>)<\/li>\n<\/ol>\n"},{"AnswerId":"41417733","CreationDate":"2017-01-01T18:06:00.167","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>As I mentioned in the description, I had previously installed numpy. For TensorFlow to run properly numpy has to be there. There has been two numpy versions in my machine. <\/p>\n\n<p>When I typed<\/p>\n\n<pre><code>&gt;&gt; import numpy\n&gt;&gt; print numpy.__version__\n<\/code><\/pre>\n\n<p>I got 1.8.something as my numpy version. According to the threads I read, this one is somewhat old and doesn't agree with Tensorflow, so I tried to upgrade the numpy version via<\/p>\n\n<pre><code>pip install numpy --upgrade\n<\/code><\/pre>\n\n<p>command. Still I got 1.8.* as my numpy version by executing above commands in python console. What I did next was to locate my numpy in python and removed numpy from there. These are the commands I used to locate and remove numpy.<\/p>\n\n<pre><code>numpy.__path__ #to locate numpy location\/path\n<\/code><\/pre>\n\n<p>Then I exit from python console and executed following in Ubuntu terminal to remove numpy.<\/p>\n\n<pre><code>sudo rm -rf \/path_to_numpy\/numpy  #this removed numpy 1.8.* version \n<\/code><\/pre>\n\n<p>Then I ran a <code>sudo apt-get update<\/code> command and checked the numpy version again and now python console reflected that I'm referring to the newer version of numpy (<strong>1.11.3<\/strong>).<\/p>\n\n<p>The problem solved.<\/p>\n\n<p>(I again encountered another issue while installing Tensorflow after successfully solving this issue. In that case I had to upgrade <strong>python <em>six<\/em><\/strong> package)<\/p>\n\n<p>Now Tensorflow works well in my Ubuntu 14.10 :)<\/p>\n\n<p>Thanks for your concern :)<\/p>\n"}]}
{"QuestionId":41416152,"AnswerCount":1,"Tags":"<theano><theano.scan>","CreationDate":"2017-01-01T14:21:15.920","AcceptedAnswerId":41456654.0,"OwnerUserId":969289.0,"Title":"TypeError when accessing RNN results","Body":"<p>I am trying to extract the decisions of an RNN for all tokens in a sequence but receive the error mentioned at the end of the post.<\/p>\n\n<p>Any help will be appreciated!<\/p>\n\n<p>Code:<\/p>\n\n<pre><\/pre>\n\n<p>The results should be the \"all_tokens_output_vector\", but I get the following error when trying to output it:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41456654","CreationDate":"2017-01-04T05:16:55.863","ParentId":null,"OwnerUserId":"3737792","Title":null,"Body":"<p>What are you doing with \"all_tokens_output_vector\" after that? Your print it directly?<\/p>\n\n<p>There should be a theano.function to compile the graph<\/p>\n"}]}
{"QuestionId":41417294,"AnswerCount":0,"Tags":"<python><numpy><machine-learning><tensorflow><kaggle>","CreationDate":"2017-01-01T17:09:09.423","AcceptedAnswerId":null,"OwnerUserId":3541828.0,"Title":"Tensorflow accuracy is 1.0 but the network isn't learning","Body":"<p>I'm trying to use mlp on titanic dataset from kaggle, but when i train it it always give accuracy of 1.0 even if i lower the epochs to 1 or the nodes to very little( 4,1,4 ), i have testing the data type of the inputs all is alright.\nThis is the code<\/p>\n\n<pre><\/pre>\n\n<p>and this is the patch function <\/p>\n\n<pre><\/pre>\n\n<p>and this is the results <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41417692,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-01T18:02:02.867","AcceptedAnswerId":41418053.0,"OwnerUserId":3319713.0,"Title":"Concatenate 3d tensors by cloning one tensor?","Body":"<p>I have two tensors:<\/p>\n\n<pre><\/pre>\n\n<p>I want to append  to , to create  such as  has a shape of .<\/p>\n\n<p>This seems rather simple but I haven't figured out how to do this with :<\/p>\n\n<pre><\/pre>\n\n<p>Do I need to reshape  and  first then concat? <\/p>\n","answers":[{"AnswerId":"41418053","CreationDate":"2017-01-01T18:47:43.840","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>This can be done using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#tile\" rel=\"nofollow noreferrer\"><code>tf.tile<\/code><\/a>. You will need to <em>clone<\/em> the tensor along dimension 1, <code>20<\/code> times to make it compatible with <code>a<\/code>. Then a simple concatenation along dimension 2 will give you the result.<\/p>\n\n<p>Here is the complete code,<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.placeholder(tf.float32, [None, 20, 100])\nb = tf.placeholder(tf.float32, [None, 1, 100])\nc = tf.tile(b, [1,20,1])\nprint c.get_shape()\n# Output - (?, 20, 100)\nd = tf.concat(2, [a,c])\nprint d.get_shape()\n# Output - (?, 20, 200)\n<\/code><\/pre>\n"}]}
{"QuestionId":41417886,"AnswerCount":0,"Tags":"<python><performance><tensorflow><neural-network>","CreationDate":"2017-01-01T18:25:41.350","AcceptedAnswerId":null,"OwnerUserId":7361100.0,"Title":"Tensorflow running slower than Benchmarks?","Body":"<p>So I'm just experimenting with an average AlexNet. I have the ImageNet data processed as in <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception<\/a>, and I am loading as such:<\/p>\n\n<pre><\/pre>\n\n<p>Each batch takes about 15 seconds to run.... When I look at the benchmarks (<a href=\"https:\/\/github.com\/soumith\/convnet-benchmarks\/blob\/master\/README.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/soumith\/convnet-benchmarks\/blob\/master\/README.md<\/a>) for similar GPUs, I see that they take only 81 ms per batch. Is there something I'm doing dreadfully wrong here?<\/p>\n","answers":[]}
{"QuestionId":41420642,"AnswerCount":2,"Tags":"<keras>","CreationDate":"2017-01-02T01:58:24.753","AcceptedAnswerId":null,"OwnerUserId":7134968.0,"Title":"Keras Tutorial Example cifar10_cnn.py gives acc: 0.34 val_acc: 0.40 in 200 epochs on windows 10?","Body":"<p>I am running the tutorial example cifar10_cnn.py from <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Here is the environment\/configuration for the test:<\/p>\n\n<ul>\n<li>Windows 10<\/li>\n<li>Keras 1.2.0<\/li>\n<li>Theano 0.8.2<\/li>\n<li>Numpy 1.11.2<\/li>\n<li>Enthought\/Canopy\/MKL(2017.0.1-1)<\/li>\n<li>.theanorc [blas] ldflags -L... -lmk2_rt<\/li>\n<\/ul>\n\n<p>The program took about one and half day to finish 200 epochs. The following figure shows the header and the last three epochs.<\/p>\n\n<blockquote>\n  <p>Using Theano backend<\/p>\n  \n  <p>X_train shape: (50000L, 32L, 32L, 3L) ...<\/p>\n  \n  <p>Epoch 198\/200\n  50000\/50000 [==============================] - 639s - loss: 1.7894 - acc: 0.3497 - val_loss: 1.5930 - val_acc: 0.3968<\/p>\n  \n  <p>Epoch 199\/200\n  50000\/50000 [==============================] - 617s - loss: 1.8111 - acc: 0.3446 - val_loss: 1.6960 - val_acc: 0.3824<\/p>\n  \n  <p>Epoch 200\/200\n  50000\/50000 [==============================] - 612s - loss: 1.8005 - acc: 0.3497 - val_loss: 1.6164 - val_acc: 0.4041<\/p>\n<\/blockquote>\n\n<p>I have two questions here:<\/p>\n\n<ol>\n<li>The accuracy <strong>(0.34\/0.40)<\/strong> is too bad. Does anyone have similar problems?<\/li>\n<li>The shape of X_train as <strong>(50000, 32, 32, 3)<\/strong> seems strange to me because\nother keras\/cifar examples give the shape of X_train as <strong>(50000, 3, 32, 32)<\/strong>. If I add <strong>\"set_image_dim_ordering('th')\"<\/strong> to the code, the shape will become <strong>(50000, 3, 32, 32)<\/strong> but the program will give even lower accuracy as \"<strong>Epoch 80\/200<\/strong> 50000\/50000 - 635s - loss: 14.5010 - <strong>acc: 0.1003<\/strong> - val_loss: 14.5063 - <strong>val_acc: 0.1000<\/strong>\" How to explain the effect of set dim here? <\/li>\n<\/ol>\n\n<p>Thanks for any comments.<\/p>\n","answers":[{"AnswerId":"44995765","CreationDate":"2017-07-09T11:13:21.830","ParentId":null,"OwnerUserId":"914874","Title":null,"Body":"<p>That may not be the answer for the whole subject, but in order to get the correct image shape, you should just verify whether the image channel is the first or the last item of the shape. You can do it the following way, considering Keras 2:<\/p>\n\n<pre><code>from keras import backend as K\n\ndef image_input_shape(img_width, img_height):\n    if K.image_data_format() == 'channels_first':\n        input_shape = (3, img_width, img_height)\n    else:\n        input_shape = (img_width, img_height, 3)\n    return input_shape\n<\/code><\/pre>\n\n<p>Just pass the image width and height and it will give back the correct ordering of the shape.<\/p>\n\n<p>From there you can discard or validate you hypothesis for the \"strange\" shape.<\/p>\n"},{"AnswerId":"45048653","CreationDate":"2017-07-12T04:46:56.217","ParentId":null,"OwnerUserId":"7134968","Title":null,"Body":"<p>The problem I posted might be related to Keras 1.2.0 only. Since Keras has upgraded to 2.x and the official file cifar10_cnn.py has also changed, the problem I posted might not happen in new Keras. As a result, I would like to close this issue. Bytheway, one more piece info is that set_image_dim_ordering will create a different network architecture. The following output show the difference with this function and without this function.<\/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>Using Theano backend.\r\n with K.set_image_dim_ordering('th')\r\n\r\nX_train shape: (50000L, 3L, 32L, 32L)\r\n50000 train samples\r\n10000 test samples\r\n          InputLayer (None, 3L, 32L, 32L)\r\n       Convolution2D (None, 32, 32L, 32L)\r\n                Relu (None, 32, 32L, 32L)\r\n       Convolution2D (None, 32, 30L, 30L)\r\n                Relu (None, 32, 30L, 30L)\r\n        MaxPooling2D (None, 32, 15L, 15L)\r\n             Dropout (None, 32, 15L, 15L)\r\n       Convolution2D (None, 64, 15L, 15L)\r\n                Relu (None, 64, 15L, 15L)\r\n       Convolution2D (None, 64, 13L, 13L)\r\n                Relu (None, 64, 13L, 13L)\r\n        MaxPooling2D (None, 64, 6L, 6L)\r\n             Dropout (None, 64, 6L, 6L)\r\n             Flatten (None, 2304)\r\n               Dense (None, 512)\r\n                Relu (None, 512)\r\n             Dropout (None, 512)\r\n               Dense (None, 10)\r\n             Softmax (None, 10)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>Using Theano backend.\r\n without K.set_image_dim_ordering('th')\r\n\r\nX_train shape: (50000L, 32L, 32L, 3L)\r\n50000 train samples\r\n10000 test samples\r\n          InputLayer (None, 32L, 32L, 3L)\r\n       Convolution2D (None, 32L, 32L, 32)\r\n                Relu (None, 32L, 32L, 32)\r\n       Convolution2D (None, 30L, 30L, 32)\r\n                Relu (None, 30L, 30L, 32)\r\n        MaxPooling2D (None, 15L, 15L, 32)\r\n             Dropout (None, 15L, 15L, 32)\r\n       Convolution2D (None, 15L, 15L, 64)\r\n                Relu (None, 15L, 15L, 64)\r\n       Convolution2D (None, 13L, 13L, 64)\r\n                Relu (None, 13L, 13L, 64)\r\n        MaxPooling2D (None, 6L, 6L, 64)\r\n             Dropout (None, 6L, 6L, 64)\r\n             Flatten (None, 2304)\r\n               Dense (None, 512)\r\n                Relu (None, 512)\r\n             Dropout (None, 512)\r\n               Dense (None, 10)\r\n             Softmax (None, 10)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
{"QuestionId":41421029,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-02T03:27:45.200","AcceptedAnswerId":41422326.0,"OwnerUserId":2878004.0,"Title":"What is the default strategy for device placement in Tensorflow?","Body":"<p>I am trying to set up distributed training. Right now I have one parameter server and two workers. If I add another parameter server how will Tensorflow split up the parameters between the two servers? Is it done randomly or do I need to manually specify it?<\/p>\n","answers":[{"AnswerId":"41422326","CreationDate":"2017-01-02T06:33:48.563","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>They get placed round-robin on available ps tasks, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/27711108b5fce2e1692f9440631a183b3808fa01\/tensorflow\/python\/training\/device_setter_test.py\" rel=\"nofollow noreferrer\">device_setter_test.py<\/a><\/p>\n"}]}
{"QuestionId":41421117,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-02T03:44:56.083","AcceptedAnswerId":41580430.0,"OwnerUserId":2975464.0,"Title":"Storing tensorflow models in memory","Body":"<p>The program I'm writing involves switching between models during run-time.<\/p>\n\n<p>I am currently using Saver to save\/load models from the disk as specified here: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver\">https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver<\/a>.<\/p>\n\n<p>The models are fairly small and can be stored in memory, so I was wondering if anyone knows of a way to store and restore these models in-memory instead of saving them to disk.<\/p>\n\n<p>I tried to modify the tensorflow source to save the model to memory however  seems to be generated during compile time. Another possible way is to use memory mapped files. Does anyone know of an easier way?<\/p>\n","answers":[{"AnswerId":"41580430","CreationDate":"2017-01-10T23:42:00.767","ParentId":null,"OwnerUserId":"3893224","Title":null,"Body":"<p>I would just have two different sessions with their own computation graphs.  Alternatively, you could just duplicate the computation graph (two copies of the variables, operations, etc) in the same session.  Then you would call <code>sess.run(comp1 if useCompOne else comp2)<\/code>, however you'd like to set it up.<\/p>\n"}]}
{"QuestionId":41421174,"AnswerCount":2,"Tags":"<python><docker><tensorflow><jupyter-notebook>","CreationDate":"2017-01-02T03:56:45.983","AcceptedAnswerId":41834429.0,"OwnerUserId":4896087.0,"Title":"How to run a new Jupyter Notebook file that's not part of a pre-built docker image in docker?","Body":"<p>I am new to Docker. In order to take the Udacity Deep Learning course, I had to set up TensorFlow on my Windows machine using Docker. (Although TensorFlow is now available on Windows, it only supports Python 3.5, however the Udacity course material requires Python 2.7. Therefore, I have to stick with the Docker way of using TensorFlow.)<\/p>\n\n<p>To work on the assignments, I followed <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/README.md\" rel=\"nofollow noreferrer\">the instructions here<\/a> as detailed below:<\/p>\n\n<ul>\n<li>First, I <a href=\"https:\/\/www.docker.com\/products\/docker-toolbox\" rel=\"nofollow noreferrer\">installed docker toolbox<\/a>.<\/li>\n<li>Then, I launch Docker using the Docker Quickstart Terminal. For the first time, I ran: <\/li>\n<\/ul>\n\n<p>. <\/p>\n\n<p>Each time after, I just run this in my docker terminal: <\/p>\n\n<p><\/p>\n\n<ul>\n<li>Finally, in the address bar, with  I get the assignment Jupyter notebooks up and running (see image below).<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4qDQz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4qDQz.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However, what I want now is <strong>to run the final project of the course which is not part of the pre-built Udacity docker image<\/strong>. How can I do that? <a href=\"https:\/\/github.com\/udacity\/machine-learning\/tree\/master\/projects\/digit_recognition\" rel=\"nofollow noreferrer\">The final project can be found here<\/a>, with the \"digit_recognition.ipynb\" specifically being the file to run in docker.<\/p>\n\n<p>Any guidance is much appreciated.<\/p>\n","answers":[{"AnswerId":"41427299","CreationDate":"2017-01-02T12:57:47.820","ParentId":null,"OwnerUserId":"3050140","Title":null,"Body":"<p>First of all, you need a way to get this Jupyter notebook (final project) on your Docker instance.<\/p>\n\n<p>What is an easy way to copy a file inside of a Docker container? Well, not a lot.<\/p>\n\n<ul>\n<li>We could <a href=\"https:\/\/jpetazzo.github.io\/2015\/01\/13\/docker-mount-dynamic-volumes\/\" rel=\"nofollow noreferrer\">attach a volume<\/a>.<\/li>\n<li>We could rewrite the Dockerfile to include the final project.<\/li>\n<li>We could also enter in the Docker container and download the file.<\/li>\n<\/ul>\n\n<p>I am going to detail the last one, but, don't forget that there are many solutions to one problem.<\/p>\n\n<p>How do we enter in the Docker container?<\/p>\n\n<blockquote>\n  <p><code>docker exec -it [container-id] bash<\/code><\/p>\n<\/blockquote>\n\n<p>How can we get the <code>[container-id]<\/code> ?<\/p>\n\n<blockquote>\n  <p><code>docker ps<\/code><\/p>\n<\/blockquote>\n\n<p>It will show you a list of containers, match the one you want to enter in.<\/p>\n\n<p>Once you're in your container. How can we download the file we want?<\/p>\n\n<p>We should try to figure out if we have <code>wget<\/code> or <code>curl<\/code> utilities to download a file. If we don't, we have to install them from any package manager available (try <code>apt-get<\/code>, if it works, do: <code>apt-get install wget<\/code>).<\/p>\n\n<p>Once we have something to download files from the Internet, we have to find out where the notebooks are stored. That is the difficult part.<\/p>\n\n<p>Look for any folder which might contain, there could also be some kind of magic one liner to type using <code>find<\/code>, unfortunately, I am no magic wizard anymore.<\/p>\n\n<p>Let's assume you are in the good folder.<\/p>\n\n<blockquote>\n  <p>wget <a href=\"https:\/\/raw.githubusercontent.com\/udacity\/machine-learning\/master\/projects\/digit_recognition\/digit_recognition.ipynb\" rel=\"nofollow noreferrer\">https:\/\/raw.githubusercontent.com\/udacity\/machine-learning\/master\/projects\/digit_recognition\/digit_recognition.ipynb<\/a><\/p>\n<\/blockquote>\n\n<p>That's all! Reload your page and you should see the notebook displayed.<\/p>\n\n<p>Side-note: You might also need to install extra dependencies in the container.<\/p>\n"},{"AnswerId":"41834429","CreationDate":"2017-01-24T17:20:11.903","ParentId":null,"OwnerUserId":"4896087","Title":null,"Body":"<p>An alternative and much easier way is this:<\/p>\n\n<ul>\n<li>Just start up your container like this: $ docker start -ai\ntensorflow-udacity <\/li>\n<li>Then, click the upload button and locate the final\nproject iPython Notebook file and upload it.<\/li>\n<\/ul>\n\n<p>That's it. Whatever changes you make will be retained and you'll be able to see the new file in the container going forward!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dWzgs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dWzgs.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":41422689,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-02T07:11:21.630","AcceptedAnswerId":null,"OwnerUserId":6467567.0,"Title":"tensorflow get size of string_input_producer","Body":"<p>How do I check the number of filenames string_input_producer has read? Different operations will be performed depending on size in input data so I need to know how many images will be read or have been read.<\/p>\n\n<p>Code below not telling me how much images I have read or am about to read.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44882163","CreationDate":"2017-07-03T09:31:52.613","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p><code>string_input_producer<\/code> returns you back a standard <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/FIFOQue\" rel=\"nofollow noreferrer\">FIFOQueue<\/a> (it returns you an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/python\/training\/input.py#L226\" rel=\"nofollow noreferrer\"><code>input_producer<\/code><\/a> and it returns you a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/python\/training\/input.py#L172\" rel=\"nofollow noreferrer\">queue<\/a>. <\/p>\n\n<p>A FIFOQueue does not have information about the number of elements it has read, only the number of elements are currently in a queue (<code>q.size()<\/code>). If you want to know how many element has been read you need to manually add a counter which you will increment each time you read an element.<\/p>\n"},{"AnswerId":"41427835","CreationDate":"2017-01-02T13:32:05.827","ParentId":null,"OwnerUserId":"6778091","Title":null,"Body":"<p>Functions like <code>string_input_producer<\/code> will add a queue to current graph which can only <strong>dequeue only one example<\/strong> each time. Usually the output tensors will be feed to functions like <code>tf.train.shuffle_batch<\/code> which is what you want. The argument <code>batch_size<\/code> of this function can control how many examples each time dequeued as the input of of your model<\/p>\n\n<p>UPDATE:<\/p>\n\n<p>if you want to check whether your input data is correct, you can run it out with <code>sess.run(my_img)<\/code> which will give you a numpy.array tensor. You can directly look at the element of this tensor or just plot it with <code>matplotlib<\/code>.<\/p>\n\n<p>make sure you have already started queue runners before <code>sess.run<\/code> or your program will hang forever<\/p>\n"}]}
{"QuestionId":41422800,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-01-02T07:21:31.567","AcceptedAnswerId":null,"OwnerUserId":5605365.0,"Title":"Why is my simple neural network not learning?","Body":"<p>I am new to TensorFlow and neural networks. I am trying to build a neural network that can classify images in the CIFAR-10 dataset.<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>I am using two hidden layers with 50 nodes in each layer. I am running 1,000 cycles, where in each cycle I shuffle data in the dataset and pick the first 100 images of that shuffle to train on.<\/p>\n\n<p>I am consistently getting ~0.1 accuracy, the machine is not learning at all.<\/p>\n\n<p>When I modify the code to use the MNIST dataset instead of the CIFAR-10 dataset I get ~0.87 accuracy.<\/p>\n\n<p>I took code from an MNIST tutorial and am trying to modify it to classify CIFAR-10 data.<\/p>\n\n<p>I can't figure out what's wrong here. How do I get my algorithm to learn?<\/p>\n","answers":[]}
{"QuestionId":41422910,"AnswerCount":1,"Tags":"<regression><caffe><hdf5>","CreationDate":"2017-01-02T07:33:18.627","AcceptedAnswerId":null,"OwnerUserId":7325313.0,"Title":"how split regression data to convert to hdf5 (Caffe)","Body":"<p>I am using @shai's code to convert data to hdf5, but after converting the size of data is too large, larger than limit size of caffe (2GB)\nso , my question is how we should split the data? \nwe only need convert data separately depends on what we want ? <\/p>\n","answers":[{"AnswerId":"41447228","CreationDate":"2017-01-03T15:48:45.113","ParentId":null,"OwnerUserId":"249226","Title":null,"Body":"<p>It would be helpful to have more information to give a better answer. Please give the actual error that Caffe gives you. Also, what exactly is larger than 2GB? Is it the txt file?<\/p>\n\n<p>HDF5 is not very efficient. I would recommend either using the database layer (<a href=\"http:\/\/caffe.help\/manual\/layers\/data.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.help\/manual\/layers\/data.html<\/a>) or writing your own customized Python layer (<a href=\"http:\/\/caffe.help\/manual\/layers\/python.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.help\/manual\/layers\/python.html<\/a>).<\/p>\n"}]}
{"QuestionId":41424358,"AnswerCount":1,"Tags":"<python><visual-studio><tensorflow>","CreationDate":"2017-01-02T09:41:29.283","AcceptedAnswerId":null,"OwnerUserId":6822876.0,"Title":"No output when running TensorFlow through Visual Studios with Python Tools","Body":"<p>I have my Visual Studio's setup with Python Tools and CPython installed.<\/p>\n\n<p>Running a sample \"Hello World!\" program shows that Python is setup correctly in Visual Studios and the messages prints out.<\/p>\n\n<p>However, I am looking to use Visual Studio as my IDE of choice for working with TensorFlow. I have imported all necessary packages. yet when running a sample TensorFlow program as shown below, the running window simply displays a black screen.<\/p>\n\n<pre><\/pre>\n\n<p>There are no errors shown, no messages hence I am unsure what exactly is missing or not setup correctly.<\/p>\n","answers":[{"AnswerId":"44493189","CreationDate":"2017-06-12T07:01:54.610","ParentId":null,"OwnerUserId":"7072306","Title":null,"Body":"<p>I had the same problem and turns out it was not really a problem! It actually works but it takes a noticeably long time to show you something so you might think its not working! Just give it a few minutes and you will see the output. Still I have the problem of very slow processing time which I don't know how to approach. Could be a VS problem.<\/p>\n"}]}
{"QuestionId":41424792,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-02T10:13:27.633","AcceptedAnswerId":null,"OwnerUserId":3138322.0,"Title":"How to extract Tensor from Tensorflow Graph without creating a Session?","Body":"<p>I am trying to modify tensorflow graph object to perform graph level modifications in the framework<\/p>\n\n<p><strong>Python Version:<\/strong> 3.5<br>\n<strong>Tensorflow Version:<\/strong> 0.12rc<\/p>\n\n<p><strong>Goal:<\/strong> Extract tensor from the graph<\/p>\n\n<pre><\/pre>\n\n<p>The first statement creates an empty graph object and the second statement adds a node into the graph.<\/p>\n\n<p>The next statement gets the default graph as graph definition and should return the Tensor.<\/p>\n\n<p><strong>Expected Result:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Error:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Update 1:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The above statement gives:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Update 2:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The above statement returns <\/p>\n\n<pre><\/pre>\n\n<p>Somehow due to extraction process, the name is changed from <strong>'abcd:0'<\/strong> to <strong>'import\/abcd:0'<\/strong>. <strong>How to get back the original name?<\/strong><\/p>\n","answers":[]}
{"QuestionId":41426846,"AnswerCount":1,"Tags":"<docker><tensorflow><jupyter-notebook><docker-machine>","CreationDate":"2017-01-02T12:27:38.057","AcceptedAnswerId":41426942.0,"OwnerUserId":5094187.0,"Title":"docker local host url not opening","Body":"<p>I installed docker and with tensorflow image I am unable to open in browser with jupyter notebook.<br>\nWhat am I missing??  <\/p>\n\n<p>command used:  <strong>docker run -it -v \/home\/$USER_NAME\/tf_files:\/tf_files  gcr.io\/tensorflow\/tensorflow<\/strong><br>\nwhere \"gcr.io\/tensorflow\/tensorflow\"  is the tensorflow image and \"\/home\/surya\" is $HOME.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YKpIz.png\" rel=\"nofollow noreferrer\">in terminal<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Qvlbx.png\" rel=\"nofollow noreferrer\">output in browser<\/a><\/p>\n\n<p>PS: docker installation is correct as \"docker run hello-world\" gives required message.<\/p>\n","answers":[{"AnswerId":"41426942","CreationDate":"2017-01-02T12:33:35.767","ParentId":null,"OwnerUserId":"2417204","Title":null,"Body":"<p>You missed to bind some ports. The official documentation of tensorflow provides the exposed ports with this command:<\/p>\n\n<pre><code>docker run -it -p 8888:8888 -v \/home\/surya\/tf_files:\/tf_files gcr.io\/tensorflow\/tensorflow\n<\/code><\/pre>\n\n<p>where <code>-p 8888:8888<\/code> means: link the port 8888 of my local machine with the service in the container, which is also 8888. Then you can access the service at <a href=\"http:\/\/localhost:8888\" rel=\"nofollow noreferrer\">http:\/\/localhost:8888<\/a><\/p>\n\n<h3>Why do I have to map a port?<\/h3>\n\n<p>Your container shows the following:<\/p>\n\n<pre><code>Copy\/paste this URL into your browser when you connect for the first time,\nto login with a token:\n    http:\/\/localhost:8888\/?token=1b3ec72ff1ed67f77a09beaee1dc4b9ad4e7aee26401b6f0\n<\/code><\/pre>\n\n<p>which means that you have to connect to the running process inside the container with the port 8888. To make the port of the container accessible from your local machine, you have to add <code>-p 8888:8888<\/code> to your command. Then accessing the URL given to you from your container makes it possible to access the container's notebook via your local browser.<\/p>\n"}]}
{"QuestionId":41427951,"AnswerCount":1,"Tags":"<python><visual-studio><tensorflow>","CreationDate":"2017-01-02T13:41:18.327","AcceptedAnswerId":null,"OwnerUserId":6822876.0,"Title":"Importing TensorFlow stops python program from running","Body":"<p>I have Python Tools setup in Visual Studios with CPython installed.<\/p>\n\n<p>In Visual Studios, if i run the following code:<\/p>\n\n<pre><\/pre>\n\n<p>The program runs fine, prints two 'hello', and exits normally.<\/p>\n\n<p>However, if I run the following code:<\/p>\n\n<pre><\/pre>\n\n<p>The program hangs, prints one 'hello', and refuses to continue.<\/p>\n\n<p>All packages should be correctly installed - using the TensorFlow in the Python interactive window prints the correct output and works perfectly.<\/p>\n\n<p>Why does the program hang in the second scenario?<\/p>\n","answers":[{"AnswerId":"41428394","CreationDate":"2017-01-02T14:10:31.223","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>Once you import tensorflow it automatically tries to load cuda, it prints something like this:<\/p>\n\n<pre><code>I tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcublas.so locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcudnn.so locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcufft.so locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcuda.so.1 locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcurand.so locally\n<\/code><\/pre>\n\n<p>So I think what is happening is that you don't have cuda installed correctly and it is failing because of it. You can try to install the CPU version which doesn't use the GPU and doesn't load those libraries.<\/p>\n"}]}
{"QuestionId":41428358,"AnswerCount":1,"Tags":"<python-2.7><machine-learning><deep-learning><keras><conv-neural-network>","CreationDate":"2017-01-02T14:07:07.297","AcceptedAnswerId":null,"OwnerUserId":5438836.0,"Title":"How to show the result of prediction in percentage in keras?","Body":"<p>The result of output always in one hot format in keras look like:<\/p>\n\n<pre><\/pre>\n\n<p>But I want to have a probability percent for each classification, for example<\/p>\n\n<pre><\/pre>\n\n<p>My simple code is given below  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41428451","CreationDate":"2017-01-02T14:14:02.000","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>Use <a href=\"https:\/\/keras.io\/models\/sequential\/#sequential-model-methods\" rel=\"nofollow noreferrer\"><code>predict_proba<\/code><\/a>. From the documentation:<\/p>\n\n<blockquote>\n  <p><code>predict_proba(self, x, batch_size=32, verbose=1)<\/code><\/p>\n  \n  <p>Generates class probability predictions for the input samples batch by batch.<\/p>\n  \n  <p>Returns a Numpy array of probability predictions.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41428471,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-02T14:15:32.510","AcceptedAnswerId":null,"OwnerUserId":7287271.0,"Title":"How to migrate NumPy preprocessing into a TensorFlow graph?","Body":"<p>I'm using TensorFlow for machine learning on Ogg and MIDI data, but a lot of preprocessing is done in NumPy (with feed_dict:s), and I'd like to migrate as much of it as possible into the computational graph in order to simplify production deployment (Google Cloud ML, or maybe self-hosted TensorFlow Serving). How would I go about this? Are there ways of converting NumPy to TensorFlow operations automatically?<\/p>\n","answers":[{"AnswerId":"41431373","CreationDate":"2017-01-02T18:05:43.907","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Most of the Numpy functions have their TensorFlow equivalent documented in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/\" rel=\"nofollow noreferrer\"><code>array_ops<\/code><\/a>.\nFor the more mathematical operations, have a look at <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/\" rel=\"nofollow noreferrer\"><code>math_ops<\/code><\/a>.<\/p>\n\n<p>Finally, if you have more specific queries or are unable to convert some Numpy code to TensorFlow, you should always try to search StackOverflow Q\/A or ask a question here. (Have a look at <a href=\"https:\/\/stackoverflow.com\/questions\/41417692\/concatenate-3d-tensors-by-cloning-one-tensor\">this<\/a> for a good example of such a question).<\/p>\n\n<hr>\n\n<p>Unrelated - If you face difficulties carrying out some matrix manipulation, try to look at the existing <a href=\"https:\/\/stackoverflow.com\/questions\/tagged\/numpy\">Numpy Q\/A<\/a> on StackOverflow. They can easily be applied to TensorFlow using the APIs above.<\/p>\n"}]}
{"QuestionId":41428756,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-01-02T14:35:47.393","AcceptedAnswerId":null,"OwnerUserId":133374.0,"Title":"how to call sgemm in a custom TensorFlow C++ op","Body":"<p>I'm following the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/adding_an_op\/\" rel=\"nofollow noreferrer\">tutorial on how to define my own op for TensorFlow in C++<\/a>.<\/p>\n\n<p>I want to call  in my custom TensorFlow C++ op. I am writing two kernels, one for CUDA and one for CPU. How would the  call in each case look like? Or is there a generic method which works for both cases?<\/p>\n\n<p>I tried with this code snippet but I cannot get it to work because of missing include files (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6602\" rel=\"nofollow noreferrer\">here<\/a>):<\/p>\n\n<pre><\/pre>\n\n<p>Also, not sure if this is generic or if this would be for CUDA only.<\/p>\n\n<p>Is this anywhere documented?<\/p>\n\n<p>How can I call  in my GPU\/CUDA implementation?\nOr more precisely, how to get the ?<\/p>\n\n<p>I searched a bit through the TF code and there is the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/stream_executor\/cuda\/cuda_blas.h\" rel=\"nofollow noreferrer\">class <\/a> which seems to provide wrappers around the cuBLAS functions. Do I need to use this or can I use  directly?\nI guess I need to use the wrapper as this will make sure that the CUDA stream executer stays in a sane state? How do I use the wrapper?<\/p>\n\n<p>I also found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/kernels\/blas_gemm.cc\" rel=\"nofollow noreferrer\">contrib\/rnn\/kernels\/blas_gemm.cc<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/matmul_op.cc\" rel=\"nofollow noreferrer\">core\/kernels\/matmul_op.cc<\/a> which seems to do what I want. The code looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>I.e. in my , I would call<\/p>\n\n<pre><\/pre>\n\n<p>I tried that but it seems that some include headers are missing for me (TensorFlow 0.12.0 with GPU for Linux). I get . I reported that upstream <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6602\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Is there any documentation about all this, i.e. how to handle cuBLAS, or this  interface, the stream executer logic, etc?<\/p>\n\n<p>My current solution was kind of hacky. For CPU, I try to link against some available Blas library on the system and use  from there. For CUDA, I link to  because in there I have found which I can use. See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/kernels\/blas_gemm.cc\" rel=\"nofollow noreferrer\">here<\/a>. Basically, in that contrib op, they face the same issue and have come up with <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/kernels\/blas_gemm.h\" rel=\"nofollow noreferrer\">this<\/a>. But that partly depends on include files which are not available in general, see the issue above.<\/p>\n","answers":[{"AnswerId":"51230296","CreationDate":"2018-07-08T08:49:09.723","ParentId":null,"OwnerUserId":"1328391","Title":null,"Body":"<p>You could try the following which worked for me today:\nIn your *.cu.cc file at the beginning:<\/p>\n\n<pre><code>#include &lt;cublas_v2.h&gt;\ncublasHandle_t cublas_handle = NULL;\n<\/code><\/pre>\n\n<p>In the same *.cu.cc file in the <strong>functor<\/strong> implementation:<\/p>\n\n<pre><code>if (cublas_handle == NULL)\n{\n    assert(cublasCreate(&amp;cublas_handle) == CUBLAS_STATUS_SUCCESS);\n    assert(cublasSetStream(cublas_handle, d.stream()) == CUBLAS_STATUS_SUCCESS);\n}\n<\/code><\/pre>\n\n<p>where the <code>d<\/code> is passed into the functor from the *.cc file as a parameter whose value is <code>ctx-&gt;eigen_device&lt;Eigen::GpuDevice&gt;()<\/code><\/p>\n\n<p>Hope this helps, cheers!<\/p>\n"}]}
{"QuestionId":41429479,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-02T15:31:29.270","AcceptedAnswerId":41430205.0,"OwnerUserId":6467567.0,"Title":"tensorflow resize image changes color","Body":"<p>In my algorithm below, resize_images changes the color of my image to something wrong. why? my image is 375 row 1242 colum 3 channel<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41430205","CreationDate":"2017-01-02T16:23:38.243","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Quoting from the documentation of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/image\/resizing#resize_images\" rel=\"nofollow noreferrer\"><code>tf.resize_images<\/code><\/a>,<\/p>\n\n<blockquote>\n  <p>Resized images will be distorted if their original aspect ratio is not the same as size. To avoid distortions see <code>resize_image_with_crop_or_pad<\/code>.<\/p>\n<\/blockquote>\n\n<p>Your aspect ratio is not correct, so due to the distorting the color of image appears to be changing. Use <code>resize_image_with_crop_or_pad<\/code> if you don't want this to happen.<\/p>\n\n<hr>\n\n<p>Please start reading the documentations. They will answer more than half your questions.<\/p>\n"}]}
{"QuestionId":41429665,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-01-02T15:46:18.463","AcceptedAnswerId":null,"OwnerUserId":457290.0,"Title":"Why do I get \"You must feed a value for placeholder tensor 'output' with dtype int64\"?","Body":"<p>I'm trying to build a C++ program that reads a pre-trained model and uses it. I took the code <a href=\"https:\/\/medium.com\/jim-fleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f#.rzskqmq5z\" rel=\"nofollow noreferrer\">from here<\/a> and modified it a little bit.\nWhat I have now is:<\/p>\n\n<pre><\/pre>\n\n<p>But when I run it, I get<\/p>\n\n<pre><\/pre>\n\n<p>The graph I'm reading in  has 14K lines, so I am not copying it here. I'll put the relevant parts:<\/p>\n\n<pre><\/pre>\n\n<p>So reads the question: what is this error message telling me?<\/p>\n","answers":[{"AnswerId":"41448622","CreationDate":"2017-01-03T17:01:13.950","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In the original graph, there is a node named <code>\"output\"<\/code> that is a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/placeholders#placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a>, i.e. a symbolic tensor that must be <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/#feeding\" rel=\"nofollow noreferrer\">fed<\/a> when you run any operations that depend on it. <\/p>\n\n<p>In the following line, which calls <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/session#classtensorflow_1_1_session_1a8e3dd1ea08fe770eb78b86638d75bb5d\" rel=\"nofollow noreferrer\"><code>session-&gt;Run()<\/code><\/a>, you are telling TensorFlow to evaluate and fetch the result of a node named <code>\"output\"<\/code>:<\/p>\n\n<pre><code>status = session-&gt;Run({{\"input\", inputs}}, {\"output\"}, {}, &amp;outputs);\n<\/code><\/pre>\n\n<p>This doesn't seem to make sense: why fetch the value of a placeholder that must be fed on the same line?<\/p>\n\n<p>I suspect that the node named <code>\"output\"<\/code> does not actually correspond to the output of your model (e.g. a prediction), but rather is a placeholder for feeding in the expected output (e.g. a known label for the corresponding value fed to <code>\"input\"<\/code>). There is probably some other node in the graph that you can evaluate to get the predictions, but its name will depend on how you originally constructed the graph.<\/p>\n"}]}
{"QuestionId":41432476,"AnswerCount":1,"Tags":"<thread-safety><tensorflow>","CreationDate":"2017-01-02T19:42:28.720","AcceptedAnswerId":41432608.0,"OwnerUserId":4394807.0,"Title":"TensorFlow: Thread safety of class member when adding a new Op","Body":"<p>I'm confused about when it is necessary to consider the issue of thread safety when writing a new Op. <\/p>\n\n<p>In TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/#implement_the_kernel_for_the_op\" rel=\"nofollow noreferrer\">How-to add a new op guide,<\/a> I read below sentence which reminds users to add a mutex to prevent data race.<\/p>\n\n<blockquote>\n  <p>Important note: Instances of your OpKernel may be accessed concurrently. Your Compute method must be thread-safe. Guard any\n  access to class members with a mutex (Or better yet, don't share state\n  via class members! Consider using a ResourceMgr to keep track of Op\n  state).<\/p>\n<\/blockquote>\n\n<p>However, in the followed text in the same web page about <a href=\"https:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/#op_registration\" rel=\"nofollow noreferrer\">how to add attr to op,<\/a> I found the class member  is not protected by mutex in the given code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>So, is there any implicit mechanism in TensorFlow to protect the variable named \"preserve_index_\" from data racing? If yes, could you please tell me the location of corresponding codes which ensures the thread safety of this variable? <\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"41432608","CreationDate":"2017-01-02T19:54:33.043","ParentId":null,"OwnerUserId":"5548115","Title":null,"Body":"<p>Note that the documentation says that the \"Compute\" method must be thread-safe.  A mutex would be needed if in the Compute function of the OpKernel, the variable was being modified, since multiple threads can call Compute() on the same OpKernel object. <\/p>\n\n<p>In this example, preserve_index_ is a member variable that is set during construction of the OpKernel, not the Compute method. An object's constructor will only be accessed from one thread (the thread creating that unique object), so there is no need to protect that variable.  The object is then read-only afterwards in Compute(), so no mutex is needed.<\/p>\n"}]}
{"QuestionId":41432502,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-02T19:44:47.337","AcceptedAnswerId":41478220.0,"OwnerUserId":7366631.0,"Title":"Tensorflow not converging with nx1 integer input (column vector)","Body":"<p>I am new to Tensorflow and machine learning. I tried to modify the basic <a href=\"https:\/\/www.tensorflow.org\/get_started\/\" rel=\"nofollow noreferrer\">Tensorflow example<\/a> in order to mimic batched input but could not get it to converge.<\/p>\n\n<p>If I change x_data to be within [0,1) it is able to correctly compute W though.<\/p>\n\n<pre><\/pre>\n\n<p>Is there something wrong with my code? Here is a copy:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41478220","CreationDate":"2017-01-05T06:01:37.683","ParentId":null,"OwnerUserId":"7366631","Title":null,"Body":"<p>My friend helped me figure out that my Gradient Descent training rate is too high. Using tips in this <a href=\"https:\/\/stackoverflow.com\/questions\/33833818\/printing-the-loss-during-tensorflow-training\">post<\/a> I can clearly see that the loss is getting bigger and bigger and started overflowing eventually.<\/p>\n\n<p>I changed the learning rate to 0.005 and it started converging.<\/p>\n"}]}
{"QuestionId":41433225,"AnswerCount":0,"Tags":"<python><ios><docker><machine-learning><tensorflow>","CreationDate":"2017-01-02T21:01:58.703","AcceptedAnswerId":null,"OwnerUserId":3938893.0,"Title":"Tensorflow docker export weights","Body":"<p>Hi I've followed the Retrain tutorial of Tensorflow in order to retrain the last layer of the inception v3 machine.<\/p>\n\n<p>The tutorial is here:\n<a href=\"https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/<\/a> \nand here:\n<a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/#0\" rel=\"nofollow noreferrer\">https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/#0<\/a><\/p>\n\n<p>after I've completed the retrain process I want to export all the network's weights in order to replicate the network on iOS using BNNS API.<\/p>\n\n<p>I've found a tutorial here:\n<a href=\"https:\/\/github.com\/bignerdranch\/bnns-cocoa-example\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bignerdranch\/bnns-cocoa-example<\/a><\/p>\n\n<p>but I'm having some trouble since I'm not comfortable with python.<\/p>\n\n<p>Can someone help me?<\/p>\n\n<p>Since now I've tried to run this script from docker<\/p>\n\n<pre><\/pre>\n\n<p>but I've received multiple errors.<\/p>\n","answers":[]}
{"QuestionId":41433244,"AnswerCount":0,"Tags":"<tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-01-02T21:04:23.633","AcceptedAnswerId":null,"OwnerUserId":5631412.0,"Title":"tf.contrib.learn yields error message \"module has no attribute 'learn' \"","Body":"<p>Here is a snippet of my code taken directly from the tf.contrib.learn tutorial on tensorflow.org:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the error message:<\/p>\n\n<pre><\/pre>\n\n<p>Clearly the module has the attribute learn since tensorflow has a section on learning tf.contrib.learn. What am I doing wrong? All guidance is appreciated.<\/p>\n","answers":[]}
{"QuestionId":41433252,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-02T21:04:52.993","AcceptedAnswerId":41475540.0,"OwnerUserId":2046918.0,"Title":"Number of input channels for convolutional network","Body":"<p>I'm following the \"Deep MNIST for Experts\" tutorial for TensorFlow: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/<\/a><\/p>\n\n<p>The second convolutional layer has the shape [5, 5, 32, 64]; that is, it has 32 inputs whereas the first convolutional layer had 1 input (that input being I understand the grayscale values of the original image).<\/p>\n\n<p>What does it mean that the second convolutional layer has 32 input channels? Does it mean the 64 filters that are learned in the second layer will all be applied (shifted around) to a \"virtual\" image having 32 points per pixel (this \"virtual\" image being composed of the original image to which each filter learned in the first step has been applied)? How do you apply a 2D 5x5 filter to an image having 32 points\/values per pixel if what I said previously is correct?<\/p>\n","answers":[{"AnswerId":"41475540","CreationDate":"2017-01-05T00:39:15.427","ParentId":null,"OwnerUserId":"2082009","Title":null,"Body":"<p>The first convolution layer has the following weights:<\/p>\n\n<pre><code>W_conv1 = weight_variable([5, 5, 1, 32])\n<\/code><\/pre>\n\n<p>Here <code>5x5<\/code> is the patch size <code>1<\/code> is the number of input channels and <code>32<\/code> is the number of output channels. So after the first convolution the output has 32 channels, hence the shape of the weight matrix for second convolution layer has <code>32<\/code> input channels.<\/p>\n"}]}
{"QuestionId":41433282,"AnswerCount":3,"Tags":"<python><windows><machine-learning><tensorflow>","CreationDate":"2017-01-02T21:07:37.873","AcceptedAnswerId":41475720.0,"OwnerUserId":7085724.0,"Title":"Tensorflow Retrain on Windows","Body":"<p>When I follow the tutorials of \"<a href=\"https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/\" rel=\"noreferrer\">How to Retrain Inception's Final Layer for New Categories<\/a>\", I need to build the retainer like this <\/p>\n\n<pre><\/pre>\n\n<p>However, my tensorflow on windows does not have such directory. I am wondering why and how can I solve the problem?\n<a href=\"https:\/\/i.stack.imgur.com\/R4X8g.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/R4X8g.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Thank you in advance<\/p>\n","answers":[{"AnswerId":"41475720","CreationDate":"2017-01-05T01:02:05.830","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>From the screenshot, it appears that you have installed the TensorFlow PIP package, whereas the instructions in the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/\" rel=\"noreferrer\">image retraining tutorial<\/a> assume that you have cloned the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"noreferrer\">Git repository<\/a> (and can use <code>bazel<\/code> to build TensorFlow).<\/p>\n\n<p>However, fortunately the script (<code>retrain.py<\/code>) for image retraining is a simple Python script, which you can download and run without building anything. Simply download the copy of <code>retrain.py<\/code> from the branch of the TensorFlow repository that matches your installed package (e.g. if you've installed TensorFlow 0.12, you can download <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.12\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"noreferrer\">this version<\/a>), and you should be able to run it by typing <code>python retrain.py<\/code> at the Command Prompt.<\/p>\n"},{"AnswerId":"45126137","CreationDate":"2017-07-16T07:29:02.580","ParentId":null,"OwnerUserId":"5305401","Title":null,"Body":"<p>In my case tensorflow version is 1.2 and corresponding retrain.py is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r1.2\/tensorflow\/examples\/image_retraining\" rel=\"noreferrer\">here<\/a>. \n<a href=\"https:\/\/i.stack.imgur.com\/OWyNB.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OWyNB.png\" alt=\"enter image description here\"><\/a>\nDownload and extract flowers images from <a href=\"http:\/\/download.tensorflow.org\/example_images\/flower_photos.tgz\" rel=\"noreferrer\">here<\/a>. \nNow run the the retrain.py file as <\/p>\n\n<p><code>python retrain.py --image_dir=path\\to\\dir\\where\\flowers\\images\\where\\extracted --output_lables=retrained_labels.txt --output_graph=retrained_graph.pb<\/code><\/p>\n\n<p>note: the last two arguments in the above command are optional.<\/p>\n\n<p><strong>Now to test the retrained model:<\/strong><\/p>\n\n<ul>\n<li>go the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/image_retraining\" rel=\"noreferrer\">master branch<\/a> and download the label_image.py code as shown below<\/li>\n<li><a href=\"https:\/\/i.stack.imgur.com\/dJJFv.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dJJFv.png\" alt=\"enter image description here\"><\/a><\/li>\n<li>Then run <code>python label_image.py --image=image\/path\/to\/test\/classfication --graph=retrained_graph.pb --labels=retrained_labels.txt<\/code><\/li>\n<li>The result will be like <a href=\"https:\/\/i.stack.imgur.com\/RqbKv.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RqbKv.png\" alt=\"enter image description here\"><\/a><\/li>\n<\/ul>\n"},{"AnswerId":"48430124","CreationDate":"2018-01-24T19:22:29.277","ParentId":null,"OwnerUserId":"2500344","Title":null,"Body":"<p>I had the same problem on windows. My windows could not find <code>script.retrain<\/code>. I downloaded <code>retrain.py<\/code> file from tensoflow website at <a href=\"https:\/\/raw.githubusercontent.com\/tensorflow\/tensorflow\/r1.1\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">here<\/a>. Then, copied the file in the tensorflow folder and run the retrain script using Python command. <\/p>\n"}]}
{"QuestionId":41434206,"AnswerCount":0,"Tags":"<python><tensorflow><softmax>","CreationDate":"2017-01-02T22:51:47.587","AcceptedAnswerId":null,"OwnerUserId":7238927.0,"Title":"Taking softmax of scores in TensorFlow to output probabilities in addition to predictions","Body":"<p>I would like to be able to access the probability of my binary classifier trained in TensorFlow to allow me to tune the probability threshold without retraining the full model. The following two lines have been working fine for me for months:<\/p>\n\n<p><\/p>\n\n<p>I added the following line to evaluate the softmax of the scores:<\/p>\n\n<p><\/p>\n\n<p>Which threw the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I tried removing the optional name argument, and ran the call as follows:<\/p>\n\n<p><\/p>\n\n<p>In this case, the following error was thrown:<\/p>\n\n<pre><\/pre>\n\n<p>It appears that the softmax function is inheriting the name from the input tensor. What is the best way to address this confusion in the namespace? Many thanks for your time and thoughts.<\/p>\n","answers":[]}
{"QuestionId":41435286,"AnswerCount":1,"Tags":"<c++><opencv><machine-learning><tensorflow>","CreationDate":"2017-01-03T01:35:12.100","AcceptedAnswerId":41435557.0,"OwnerUserId":5695896.0,"Title":"Using OpenCV as a machine learning libarary instead of TensorFlow?","Body":"<p>I was wondering in what cases using C++ should i use OpenCV's ML libraries CvSVM or CvANN_MLP instead of building them with tensorflow ? which one is better and in what cases ?<\/p>\n","answers":[{"AnswerId":"41435557","CreationDate":"2017-01-03T02:20:27.340","ParentId":null,"OwnerUserId":"4480763","Title":null,"Body":"<p>If you specifically want to target 'handwritten recognition,' or similar pattern recognition outside of image analysis, neural networks using Tensor Flow is the way to go. OpenCV, on the other hand, is designed specifically for computer vision (essentially high-level image analysis to interpret certain characteristics of digital images).<\/p>\n\n<p>If you are a beginner in ML, I would highly recommend starting out with TensorFlow as a more broad introduction to ML and specifically to neural networks. There are many great and easy-to-follow <a href=\"https:\/\/www.tensorflow.org\/get_started\/\" rel=\"nofollow noreferrer\">resources<\/a> to get started.<\/p>\n"}]}
{"QuestionId":41435673,"AnswerCount":1,"Tags":"<tensorflow><bazel><syntaxnet>","CreationDate":"2017-01-03T02:39:06.847","AcceptedAnswerId":null,"OwnerUserId":4160135.0,"Title":"syntaxnet .\/configure error","Body":"<p>I was trying to using syntaxnet and I have finished most of processes. Upgrade bazel version to 0.43 in case of errors (Ubuntu 16.04 Ver, Anaconda python 2.7).\nHowever, I am having a troubles with .\/configure part. I am reading the official instruction via tensorflow github.<\/p>\n\n<pre><\/pre>\n\n<p>Following logs will help you to understand what\u2019s going on my machine.  Thanks for the advice <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41626274","CreationDate":"2017-01-13T02:02:12.947","ParentId":null,"OwnerUserId":"7329144","Title":null,"Body":"<p>I think the version of your bazel is too high for Syntaxnet. you can try bazel-0.3.1 please.<\/p>\n"}]}
{"QuestionId":41436068,"AnswerCount":2,"Tags":"<python><theano>","CreationDate":"2017-01-03T03:36:28.420","AcceptedAnswerId":null,"OwnerUserId":6310830.0,"Title":"THEANO_FLAGS: Command not found","Body":"<p>My settings of  file is , but I want to know if I can run one program with , and run another with , I tried  but it raised <\/p>\n","answers":[{"AnswerId":"41482177","CreationDate":"2017-01-05T10:08:47.247","ParentId":null,"OwnerUserId":"4303084","Title":null,"Body":"<p><code>THEANO_FLAGS='device=gpu1' python myscript.py<\/code> is working for me. \nAre you sure you have no space character after <code>THEANO_FLAGS<\/code>, like this: <code>THEANO_FLAGS ='device=gpu1' python myscript.py<\/code>.\nThis would raise the <code>THEANO_FLAGS: Command not found<\/code><\/p>\n"},{"AnswerId":"41440329","CreationDate":"2017-01-03T09:33:54.950","ParentId":null,"OwnerUserId":"6310830","Title":null,"Body":"<p>just add these 2 lines in the script:<\/p>\n\n<pre><code>import os\nos.environ[\"THEANO_FLAGS\"] = \"device=gpu1\"\n<\/code><\/pre>\n"}]}
{"QuestionId":41437364,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2017-01-03T06:03:52.940","AcceptedAnswerId":null,"OwnerUserId":5367279.0,"Title":"Tensorflow Network Save and Restore","Body":"<p>I am trying to prepare a network with this architecture,<\/p>\n\n<blockquote>\n  <p>Convolution Layer<\/p>\n  \n  <p>Max Pooling Layer<\/p>\n  \n  <p>Convolution Layer<\/p>\n  \n  <p>Max Pooling Layer<\/p>\n  \n  <p>Convolution Layer<\/p>\n  \n  <p>Max Pooling Layer<\/p>\n  \n  <p>BLSTM Layer 1<\/p>\n  \n  <p>BLSTM Layer 2<\/p>\n  \n  <p>Fully Connected Layer<\/p>\n  \n  <p>CTC Layer<\/p>\n<\/blockquote>\n\n<p>My code structure is like <\/p>\n\n<pre><\/pre>\n\n<p>Now, this model is tested with a online handwriting data. The network is working fine. But the problem is that the saving is restoring is not working properly. My CNN layer weight and bias variables have distinct names. I run the training for 100 epochs (pass 1). Then I am loading it again for pass 2 training. I have checked the weights and biases are loaded in pass 2 exactly as they are saved at the end of pass 1. But at the start of pass 2 the training loss is increasing w.r.t the loss at the end of pass 1. What am I doing wrong? Is it due to the optimizer configuration? Any help will be highly appreciated.<\/p>\n","answers":[]}
{"QuestionId":41437483,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-03T06:14:27.733","AcceptedAnswerId":null,"OwnerUserId":3529361.0,"Title":"How does tf.nn.conv2d calculate its values?","Body":"<p>I've looked at the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn\/convolution\" rel=\"nofollow noreferrer\">documentation<\/a> for  but it didn't really help much. So I tried to multiply the first 4 values of my input array that form a square  with the first column of the  array . But regardless of how I multiply these values I'm not getting , I don't know what I'm doing wrong. <\/p>\n\n<p>Below I have the code that I used.<\/p>\n\n<p>Given an array:<\/p>\n\n<pre><\/pre>\n\n<p>And weights:<\/p>\n\n<pre><\/pre>\n\n<p>Which prints:<\/p>\n\n<pre><\/pre>\n\n<p>How does:<\/p>\n\n<pre><\/pre>\n\n<p>Give:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41438472,"AnswerCount":0,"Tags":"<python><python-2.7><numpy><caffe><pycaffe>","CreationDate":"2017-01-03T07:29:16.323","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"I am getting error with caffe.io.array_to_datum() for single channel images, how to solve?","Body":"<p>I am trying to convert a set of grayscale images and ground truth images to LMDB for segmentation task. I have created 4 text files (, , , ) indicating path to each image in each folder. \nFor many days, it is taking my time and still I am not successful. I am a beginner in  and  I will be thankful if somebody helps me here.\nI tried the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I was getting an error:<\/p>\n\n<pre><\/pre>\n\n<p>And:<\/p>\n\n<pre><\/pre>\n\n<p>Since my images and labels are in grayscale (i.e., only one channel), I added , and now I am getting the following error: <\/p>\n\n<pre><\/pre>\n\n<ol>\n<li>How can I fix this error?<\/li>\n<li>How can I convert these four image sets into LMDB?<\/li>\n<li>Is it possible I use  (Caffe built-in converter)\nfor this task? If yes, please let me know how can I do that?<\/li>\n<\/ol>\n\n<p>Your help is really appreciated. <\/p>\n","answers":[]}
{"QuestionId":41438885,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-03T08:01:15.800","AcceptedAnswerId":41439738.0,"OwnerUserId":126164.0,"Title":"TensorFlow: how to find the samples that are learned\/recalled most poorly?","Body":"<p>With TensorFlow one can collect error information, such as:<\/p>\n\n<pre><\/pre>\n\n<p>But it only provides a vector telling which samples are recalled correctly.<\/p>\n\n<p>However, to really understand what's going on, or what needs to improve, I'd like to find out the most difficult samples to recall. It's like to compute the Euclidean distance between the logits value produced by the neural network and the target one_hot vector for each sample, and select those of the top distances.<\/p>\n\n<p>Could you share your insight how to do it with TensorFlow?<\/p>\n","answers":[{"AnswerId":"41439738","CreationDate":"2017-01-03T08:58:13.473","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/\" rel=\"nofollow noreferrer\"><code>math_ops<\/code><\/a> documentation!<\/p>\n\n<pre><code>sq_difference = tf.square(tf.add(logits, tf.neg(one_hot_y)))\nsq_distances = tf.reduce_sum(sq_difference, axis=1)\n<\/code><\/pre>\n\n<p>Note that, <code>sq_difference<\/code> has a dimension <code>batch_size * output_len<\/code> and <code>sq_distances<\/code> has a dimension <code>output_len<\/code>.<\/p>\n"}]}
{"QuestionId":41439136,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-01-03T08:19:58.790","AcceptedAnswerId":null,"OwnerUserId":7361818.0,"Title":"Wide & Deep learning for large data error: GraphDef cannot be larger than 2GB","Body":"<p>Inserting 1MM+ rows into the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/wide_and_deep\/\" rel=\"noreferrer\">wide and deep learning model<\/a> throws :<\/p>\n\n<pre><\/pre>\n\n<p>I defined the same  as in the example:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a replacement to  and  that allows for inserting in batches and avoiding the memory error?<\/p>\n","answers":[{"AnswerId":"43025768","CreationDate":"2017-03-26T06:25:10.207","ParentId":null,"OwnerUserId":"7361818","Title":null,"Body":"<p>I solve this problem, at last, using the following code which is inspired by others. Hope this code will be helpful to others. Thank you all for comments above.<\/p>\n\n<pre><code>def input_fn(batch_size,filename):\n    examples_op = tf.contrib.learn.read_batch_examples(\n        filename,\n        batch_size=batch_size,\n        reader=tf.TextLineReader,\n        num_epochs=1,\n        parse_fn=lambda x: tf.decode_csv(x, [tf.constant([''], dtype=tf.string)] * len(HEADERS)))\n\n    examples_dict = {}\n    for i, header in enumerate(HEADERS):\n        examples_dict[header] = examples_op[:, i]\n\n    feature_cols = {k: tf.string_to_number(examples_dict[k], out_type=tf.float32)\n                    for k in CONTINUOUS_COLUMNS}\n\n    feature_cols.update({k: dense_to_sparse(examples_dict[k])\n                         for k in CATEGORICAL_COLUMNS})\n\n    label = tf.string_to_number(examples_dict[LABEL_COLUMN], out_type=tf.int32)\n\n    return feature_cols, label\n\ndef input_fn_pre(batch_size,filename):\n    examples_op = tf.contrib.learn.read_batch_examples(\n        filename,\n        batch_size=batch_size,\n        reader=tf.TextLineReader,\n        num_epochs=1,\n        parse_fn=lambda x: tf.decode_csv(x, [tf.constant([''], dtype=tf.string)] * len(HEADERS)))\n\n    examples_dict = {}\n    for i, header in enumerate(HEADERS):\n        examples_dict[header] = examples_op[:, i]\n\n    feature_cols = {k: tf.string_to_number(examples_dict[k], out_type=tf.float32)\n                    for k in CONTINUOUS_COLUMNS}\n\n    feature_cols.update({k: dense_to_sparse(examples_dict[k])\n                         for k in CATEGORICAL_COLUMNS})\n    return feature_cols\ndef dense_to_sparse(dense_tensor):\n    indices = tf.to_int64(tf.transpose([tf.range(tf.shape(dense_tensor)[0]), tf.zeros_like(dense_tensor, dtype=tf.int32)]))\n    values = dense_tensor\n    shape = tf.to_int64([tf.shape(dense_tensor)[0], tf.constant(1)])\n\n    return tf.SparseTensor(\n        indices=indices,\n        values=values,\n        shape=shape\n    )\n\ndef train_and_eval():\n  \"\"\"Train and evaluate the model.\"\"\"\n  data = pd.read_csv('spark_traindata_forrun_no_nanindex.csv',skipinitialspace=True,\n      engine=\"python\")\n  value_range={}\n  for column in CATEGORICAL_COLUMNS:\n    data[column]=data[column].astype(str)\n    value_range[column]=list(set(data[column]))\n  model_dir = '.\/model6'\n  print(\"model directory = %s\" % model_dir)\n  test = pd.read_csv('test.csv',names= HEADERS)\n  m = build_estimator(model_dir,value_range)\n  m.fit(input_fn=lambda: input_fn(128,'train.csv'), steps=FLAGS.train_steps)\n  results = m.evaluate(input_fn=lambda: input_fn(5000,'test.csv'), steps=1)\n  for key in sorted(results):\n    print(\"%s: %s\" % (key, results[key]))\n<\/code><\/pre>\n"},{"AnswerId":"42240477","CreationDate":"2017-02-15T03:59:25.457","ParentId":null,"OwnerUserId":"4922660","Title":null,"Body":"<p>The example of wide and deep need load the data set all into memory. If you have large data set, you may need <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard3\/tf.decode_csv.md\" rel=\"nofollow noreferrer\">tf.decode_csv<\/a> for csv formatted input. If your input format is self defined, you should create a <a href=\"https:\/\/www.tensorflow.org\/how_tos\/new_data_formats\/\" rel=\"nofollow noreferrer\">custom data reader<\/a> .<\/p>\n"},{"AnswerId":"42653239","CreationDate":"2017-03-07T16:21:48.267","ParentId":null,"OwnerUserId":"690430","Title":null,"Body":"<p>One <a href=\"https:\/\/medium.com\/@ilblackdragon\/tensorflow-tutorial-part-4-958c29c717a0#.ubvuss70i\" rel=\"nofollow noreferrer\">solution<\/a> created by @ilblackdragon is to use <code>dataframe.queues<\/code>. Unfortunately, for categorical variables, you have to integer encode them first otherwise you get the error:<\/p>\n\n<pre><code>ValueError: Data types for extracting pandas data must be int, float, or bool. Found: 'sex' type='object', 'embarked' type='object'\n<\/code><\/pre>\n\n<p>Also using a DNNClassifier doesn't work and gives a key error on the encoded categories (e.g. <code>KeyError: 'Embarked_ids'<\/code>). So in the post, the author creates his own classifier model.<\/p>\n\n<p>Example code:<\/p>\n\n<pre><code># -*- coding: utf-8 -*-\n# flake8: noqa ignore=E501\nimport tempfile\n\nimport pandas as pd\nimport tensorflow as tf\nimport tensorflow.contrib.learn as tf_learn\nimport tensorflow.contrib.layers as tf_layers\nfrom sklearn.metrics import accuracy_score\nfrom sklearn.preprocessing import LabelEncoder\nfrom sklearn.model_selection import train_test_split\n\n# Define the column names for the data sets.\nLABEL_COLUMN = 'Survived'\nCONTINUOUS_COLUMNS = ['Age', 'SibSp', 'Parch', 'Fare']\nCATEGORICAL_COLUMNS = ['Pclass', 'Sex', 'Embarked']\nCATEGORICAL_ID_COLUMNS = [col + '_ids' for col in CATEGORICAL_COLUMNS]\nFEATURE_COLUMNS = CONTINUOUS_COLUMNS + CATEGORICAL_ID_COLUMNS\n\nHIDDEN_UNITS = [10, 20, 10]\nCATEGORICAL_EMBED_SIZE = 10\n\nLABEL_ENCODERS = {}\n\ndef get_feature_cols():\n    # used in DNNClassifier which doesn't work\n    continuous_features = [tf_layers.real_valued_column(col) for col in\n                           CONTINUOUS_COLUMNS]\n    categorical_features = [\n        tf_layers.embedding_column(\n            tf_layers.sparse_column_with_integerized_feature(col + '_ids', len(\n                LABEL_ENCODERS[col].classes_)),\n            CATEGORICAL_EMBED_SIZE)\n        for col in CATEGORICAL_COLUMNS\n        ]\n    return continuous_features + categorical_features\n\n\ndef pandas_input_fn(X, y=None, batch_size=128, num_epochs=None):\n    def input_fn():\n        if y is not None:\n            X['target'] = y\n        queue = tf_learn.dataframe.queues.feeding_functions.enqueue_data(\n            X, 1000, shuffle=num_epochs is None, num_epochs=num_epochs)\n        if num_epochs is None:\n            features = queue.dequeue_many(batch_size)\n        else:\n            features = queue.dequeue_up_to(batch_size)\n\n        features = dict(zip(['index'] + list(X.columns), features))\n\n        if y is not None:\n            target = features.pop('target')\n            return features, target\n        return features\n\n    return input_fn\n\n\ndef encode_categorical(df):\n    global LABEL_ENCODERS\n    for col in CATEGORICAL_COLUMNS:\n        if df[col].dtype == 'object':\n            df[col] = df[col].astype(str)\n        encoder = LabelEncoder().fit(df[col])\n        df[col + '_ids'] = encoder.transform(df[col])\n        df.pop(col)\n        LABEL_ENCODERS[col] = encoder\n    return df, LABEL_ENCODERS\n\n\ndef dnn_tanh(features, target, hidden_units=HIDDEN_UNITS):\n    global LABEL_ENCODERS\n    target = tf.one_hot(target, 2, 1.0, 0.0)\n\n    # Organize continuous features.\n    final_features = [tf.expand_dims(tf.cast(features[col], tf.float32), 1) for\n                      col in CONTINUOUS_COLUMNS]\n\n    # Embed categorical variables into distributed representation.\n    for col in CATEGORICAL_COLUMNS:\n        feature = tf_learn.ops.categorical_variable(\n            features[col + '_ids'],\n            len(LABEL_ENCODERS[col].classes_),\n            embedding_size=CATEGORICAL_EMBED_SIZE,\n            name=col)\n        final_features.append(feature)\n\n    # Concatenate all features into one vector.\n    features = tf.concat(1, final_features)\n\n    # Deep Neural Network\n    logits = tf_layers.stack(features,\n                             tf_layers.fully_connected,\n                             stack_args=hidden_units,\n                             activation_fn=tf.tanh)\n    prediction, loss = tf_learn.models.logistic_regression(logits, target)\n    train_op = tf_layers.optimize_loss(loss,\n                                       tf.contrib.framework.get_global_step(),\n                                       optimizer='SGD',\n                                       learning_rate=0.05)\n    return tf.argmax(prediction, dimension=1), loss, train_op\n\n\ndef process_input_df(df):\n    df, label_encoders = encode_categorical(df)\n    y = df.pop(LABEL_COLUMN)\n    X = df[CATEGORICAL_ID_COLUMNS + CONTINUOUS_COLUMNS].fillna(0)\n    return X, y\n\n\ndef train(X, y, steps=100):\n    model_dir = tempfile.mkdtemp()\n    print(\"model dir: \", model_dir)\n    classifier = tf_learn.Estimator(model_fn=dnn_tanh, model_dir=model_dir)\n    classifier.fit(input_fn=pandas_input_fn(X, y), steps=steps)\n\n    '''\n    # Using DNNClassifier gives KeyError (e.g on EmbedIds)\n    classifier = learn.DNNClassifier(hidden_units=[10, 20, 10],\n        n_classes=2,\n        feature_columns=get_feature_cols(),\n        optimizer=tf.train.GradientDescentOptimizer(learning_rate=0.05))\n    classifier.fit(X, y, batch_size=128, steps=500)\n    '''\n\n    return classifier\n\n\ndef predict(classifier, X):\n    return list(classifier.predict(input_fn=pandas_input_fn(X, num_epochs=1),\n                                   as_iterable=True))\n\n\ndef evaluate(classifier, X, y, steps=1):\n    results = classifier.evaluate(input_fn=pandas_input_fn(X[FEATURE_COLUMNS], y),\n                                  steps=steps)\n    for key in sorted(results):\n        print(\"%s: %s\" % (key, results[key]))\n\n\nif __name__ == '__main__':\n    # DOWNLOAD TITANIC TRAIN DATA\n    data = pd.read_csv('~\/titanic_train.csv')  # LOAD DATA\n    X, y = process_input_df(data)\n    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)\n    classifier = train(X_train, y_train, steps=100)\n    print(\"accuracy_score\", accuracy_score(y_test, predict(classifier, X_test)))\n    evaluate(classifier, X_test, y_test, steps=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41439254,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-03T08:26:43.960","AcceptedAnswerId":41439546.0,"OwnerUserId":6733064.0,"Title":"What are the differences between tf.initialize_all_variables() and tf.global_variables_initializer()","Body":"<p>On Tensorflow official website, it gives explantions of the  and  functions as follow  <\/p>\n\n<blockquote>\n  <h3>tf.initialize_all_variables():<\/h3>\n  \n  <p>Returns an op that initializes all variables.<\/p>\n  \n  <h3>tf.global_variables_initializer():<\/h3>\n  \n  <p>Adds an op to initialize all variables in the model<\/p>\n<\/blockquote>\n\n<p>It seems like both can be used to initialize all variables in graphs. Can we use these two functions exchangbly? If not, what would be the differences? <\/p>\n","answers":[{"AnswerId":"43444670","CreationDate":"2017-04-17T03:20:28.053","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md#release-0120\" rel=\"nofollow noreferrer\">changelog for 0.12 version<\/a> tells you that both both functions do the same thing because:<\/p>\n\n<blockquote>\n  <p>tf.initialize_all_variables renamed to tf.global_variables_initializer<\/p>\n<\/blockquote>\n\n<p>and as Martianwars mentioned the documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/initialize_all_variables\" rel=\"nofollow noreferrer\">initialize_all_variables<\/a> tells that:<\/p>\n\n<blockquote>\n  <p>THIS FUNCTION IS DEPRECATED. It will be removed after 2017-03-02.\n  Instructions for updating: Use tf.global_variables_initializer\n  instead.<\/p>\n<\/blockquote>\n\n<p>If you will call it, you will get a warning as well. So you should always use <code>tf.global_variables_initializer()<\/code><\/p>\n"},{"AnswerId":"41439546","CreationDate":"2017-01-03T08:45:52.913","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Unfortunately, you forgot to read an important line in the documentation of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/initialize_all_variables\" rel=\"noreferrer\"><code>tf.initialize_all_variables<\/code><\/a>.<\/p>\n\n<blockquote>\n  <p>THIS FUNCTION IS DEPRECATED. It will be removed after 2017-03-02. Instructions for updating: Use <code>tf.global_variables_initializer<\/code> instead.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41439411,"AnswerCount":1,"Tags":"<python><image><machine-learning><tensorflow><computer-vision>","CreationDate":"2017-01-03T08:37:25.597","AcceptedAnswerId":41439596.0,"OwnerUserId":7089821.0,"Title":"Cannot read image successfully in tensorflow","Body":"<p>I want to read the jpeg images into batch for image recognation. The images are in the \/Image_p\/ file and the image name are listed in the label.csv file, which are presented like 14634_right. <\/p>\n\n<p>My question is <strong>how to fix my code to read the images into batch successfully?<\/strong>\n    To be more spcific, I don't know whether should I write a  loop and where to implement it. <\/p>\n\n<p>For the original code, I got the error message on  function: <\/p>\n\n<pre><\/pre>\n\n<p>My origin code: <\/p>\n\n<pre><\/pre>\n\n<p>Plus, I have also writen a new file which can read a image successfully (thanks for the help from martianwars). \nMy test code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41439596","CreationDate":"2017-01-03T08:48:33.920","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p><code>image<\/code> will have a <code>(?, ?, 3)<\/code> shape since it's not yet been read, but you have specified the channel in the <code>decode_jpeg()<\/code> function. Try to print this instead,<\/p>\n\n<pre><code>with tf.Session() as sess:   \n    img = sess.run(image)\n    print(img)\n<\/code><\/pre>\n"}]}
{"QuestionId":41439846,"AnswerCount":0,"Tags":"<python><tensorflow><avx><python-3.6>","CreationDate":"2017-01-03T09:03:44.230","AcceptedAnswerId":null,"OwnerUserId":878520.0,"Title":"Tensorflow - wasn't compiled to use SSE4.2, AVX, AVX2, FMA","Body":"<p>I installed tensorflow from source code. <\/p>\n\n<p>After installation when importing tensorflow on Python, it notify me of this. <\/p>\n\n<pre><\/pre>\n\n<p>it says SSE4.2, AVX, AVX2, and FMA are not compiled to use. <\/p>\n\n<p>how can I solve this problem?<\/p>\n","answers":[]}
{"QuestionId":41441638,"AnswerCount":1,"Tags":"<python><tensorflow><theano><keras><keras-layer>","CreationDate":"2017-01-03T10:41:20.530","AcceptedAnswerId":null,"OwnerUserId":3670532.0,"Title":"What is extra with Keras functional API?","Body":"<p>What extra can be done using Keras functional API, which could not be done using keras sequential models?\nApart from the fact that a simple model can be reused for a time bases data using \u201cTimeDistributed\u201d layer wrapper ?<\/p>\n","answers":[{"AnswerId":"41454359","CreationDate":"2017-01-04T00:17:47.873","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>It is much more than model reuse, the functional API allows you to <em>easily<\/em> define models where layers connect to more than just the previous and next layers. You can connect layers to any other layers as you wish, so siamese networks, densely connected networks and such become possible. The old Graph API allowed the same level of connectivity but it was a PITA due to its use of layer node names to define connectivity.<\/p>\n\n<p>The sequential model is just a sequential set of layers, and new neural network architectures at this time are moving away from such pattern.<\/p>\n"}]}
{"QuestionId":41442197,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><textsum>","CreationDate":"2017-01-03T11:13:16.830","AcceptedAnswerId":null,"OwnerUserId":7369124.0,"Title":"Textsum(tensorflow): Assertion error while using vocab file generated from dataset","Body":"<p>Im having slight issue running on CNN data.\nThe vocabulary file generated using the code above gives assertion error. Im not able to understand what is causing this issue.<\/p>\n\n<p>This is the error i get :<\/p>\n\n<pre><\/pre>\n\n<p>the function in seq2seq_attention.py:<\/p>\n\n<p>def main(unused_argv): vocab = data.Vocab(FLAGS.vocab_path, 10000000) Check for presence of required special tokens. assert vocab.CheckVocab(data.PAD_TOKEN) > 0 assert vocab.CheckVocab(data.UNKNOWN_TOKEN) >= 0 assert vocab.CheckVocab(data.SENTENCE_START) > 0 assert vocab.CheckVocab(data.SENTENCE_END) > 0 \u2013<\/p>\n","answers":[{"AnswerId":"41796555","CreationDate":"2017-01-22T21:53:58.597","ParentId":null,"OwnerUserId":"5573987","Title":null,"Body":"<p>What about these? You miss some of them in your vocabulary i.e. SENTENSE_START.<\/p>\n\n<pre><code># Special tokens\nPARAGRAPH_START = '&lt;p&gt;'\nPARAGRAPH_END = '&lt;\/p&gt;'\nSENTENCE_START = '&lt;s&gt;'\nSENTENCE_END = '&lt;\/s&gt;'\nUNKNOWN_TOKEN = '&lt;UNK&gt;'\nPAD_TOKEN = '&lt;PAD&gt;'\nDOCUMENT_START = '&lt;d&gt;'\nDOCUMENT_END = '&lt;\/d&gt;'\n<\/code><\/pre>\n\n<p>source: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/textsum\/data.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/textsum\/data.py<\/a><\/p>\n"}]}
{"QuestionId":41442276,"AnswerCount":5,"Tags":"<tensorflow><progress-bar><theano><keras><keras-layer>","CreationDate":"2017-01-03T11:17:47.283","AcceptedAnswerId":null,"OwnerUserId":7295598.0,"Title":"Keras verbose training progress bar writing a new line on each batch issue","Body":"<p>running a Dense feed-forward neural net in Keras.\nthere are class_weights for two outputs, and sample_weights for a third output. fore some reason it prints the progress verbose display for each batch calculated, and not updating the print on the same line as its supposed to...\nDid this ever happens to you?\nHow is it fixed?\nFrom the shell:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44516293","CreationDate":"2017-06-13T08:37:00.287","ParentId":null,"OwnerUserId":"6026478","Title":null,"Body":"<p>This seems to be a consistent problem with Keras. I tried to find the lines<\/p>\n\n<p><code>sys.stdout.write('\\b' * prev_total_width)<\/code><\/p>\n\n<p><code>sys.stdout.write('\\r')<\/code><\/p>\n\n<p>at the Keras\/utils\/generic_utils.py file and they are (as of the current version) at 258 and 259 accordingly. I commented like 258 but this seems not to resolve the issue. I did manage to make the progress bar work by commenting the line:<\/p>\n\n<p>line 303: <code>sys.stdout.write(info)<\/code><\/p>\n\n<p>It seems as if the info makes the bar too long for the terminal and so it breaks to a new line.<\/p>\n\n<p>So I finally resolved the issue. It seems like it was rather simple at the end....<\/p>\n\n<blockquote>\n  <p>Just make the terminal wider... <\/p>\n<\/blockquote>\n\n<p>Note: Tested on Linux Ubuntu 16.04 | Keras Version 2.0.5<\/p>\n"},{"AnswerId":"41462115","CreationDate":"2017-01-04T10:55:46.487","ParentId":null,"OwnerUserId":"7373680","Title":null,"Body":"<p>I had a similar issue, but have not had the time to investigate it further. The problem seems to be related to the class Progbar in generic_utils.py of keras, see <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/utils\/generic_utils.py\" rel=\"nofollow noreferrer\">link<\/a>, and perhaps Python >= 3.3.  <\/p>\n\n<p>The following lines are found in the update function of the class:<\/p>\n\n<p>Line 107: <code>sys.stdout.write('\\b' * prev_total_width)<\/code><br>\nLine 108: <code>sys.stdout.write('\\r')<\/code><\/p>\n\n<p>I simply removed line 107 as a quick fix, so instead of backspacing the previous line then performing a shift to the beginning of the line, I only perform the shift. I guess there is some better ways than altering the source code though. <\/p>\n"},{"AnswerId":"47328189","CreationDate":"2017-11-16T11:12:18.803","ParentId":null,"OwnerUserId":"2583998","Title":null,"Body":"<p>It was mentioned before, but I will rewrite it to be more visible for future users.<\/p>\n\n<p><strong>You have too narrow terminal to print all these value<\/strong>s - just set <code>width<\/code> argument of <code>Progbar<\/code> constructor to smaller number or remove\/rename some of the provided values.<\/p>\n"},{"AnswerId":"55659250","CreationDate":"2019-04-12T20:50:36.553","ParentId":null,"OwnerUserId":"11353683","Title":null,"Body":"<p>The workaround of androst (4 Jan 2017) did not work for me. However, I found out that the lines of code which androst cited from generic_utils.py were never executed when I ran my code, because the condition in the embracing \"if-clause\" was always False. When I out-commented the if-clause, and set the corresponding variable (manually) to \"True\", it worked.<\/p>\n\n<p>This is what I changed (for me: lines 311-314 of generic_utils.py):<\/p>\n\n<pre><code>    #self._dynamic_display = ((hasattr(sys.stdout, 'isatty') and\n    #                          sys.stdout.isatty()) or\n    #                         'ipykernel' in sys.modules)\n    self._dynamic_display = True   # inserted to overwrite the above (workaround by KS)\n<\/code><\/pre>\n\n<p>Then the progress bar worked nicely :-)<\/p>\n"},{"AnswerId":"57475559","CreationDate":"2019-08-13T10:08:32.480","ParentId":null,"OwnerUserId":"9531617","Title":null,"Body":"<p>The solutions of @user11353683 and @Panos can be completed in order to not modify the sources (which can create future problems): just install <code>ipykernel<\/code> and import it in your code:<\/p>\n\n<p><code>pip install ipykernel<\/code>\nThen \n<code>import ipykernel<\/code><\/p>\n\n<p>In fact, in the Keras generic_utils.py file, the probematic line was:<\/p>\n\n<pre><code>            if self._dynamic_display:\n                sys.stdout.write('\\b' * prev_total_width)\n                sys.stdout.write('\\r')\n            else:\n                sys.stdout.write('\\n')\n<\/code><\/pre>\n\n<p>And the value self._dynamic_display was initiated such as:<\/p>\n\n<pre><code>        self._dynamic_display = ((hasattr(sys.stdout, 'isatty') and\n                                  sys.stdout.isatty()) or\n                                 'ipykernel' in sys.\n<\/code><\/pre>\n\n<p>So, loading <code>ipykernel<\/code>added it to <code>sys.modules<\/code> and fixed the problem for me.<\/p>\n"}]}
{"QuestionId":41442366,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-03T11:22:08.390","AcceptedAnswerId":null,"OwnerUserId":7345996.0,"Title":"ValueError: setting an array element with a sequence tensorflow","Body":"<p>Here is the code segment which is giving me the error -<\/p>\n\n<pre><\/pre>\n\n<p>here is the feed_dic:<\/p>\n\n<pre><\/pre>\n\n<p>Error details:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41442687,"AnswerCount":2,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-01-03T11:39:40.383","AcceptedAnswerId":41442953.0,"OwnerUserId":859453.0,"Title":"Does make sense use dynamic learning rate in AdamOptimizer?","Body":"<p>I'm dvelopping a convolutional neural network for images recognition based on three own classes.\nI built an AlexNet-based model to train.\nI'd like to know two things:<\/p>\n\n<ol>\n<li><em>AdamOptimizer<\/em> performs a learning rate decay internally (from a fixed given value) or not ?<\/li>\n<li>In case of not, can I use  to perform decay ?<\/li>\n<\/ol>\n\n<p>Small examples are apprecciated.\nThanks<\/p>\n","answers":[{"AnswerId":"41450882","CreationDate":"2017-01-03T19:24:58.097","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<blockquote>\n  <p>AdamOptimizer performs a learning rate decay internally (from a fixed given value) or not ?<\/p>\n<\/blockquote>\n\n<p>Yes, Adam does perform a learning rate decay.<\/p>\n\n<p>You should have a look at how Adam works: <\/p>\n\n<blockquote>\n  <p>D. Kingma and J. Ba, \u201cAdam: A method for stochastic optimization,\u201d\n  arXiv preprint arXiv:1412.6980, Dec. 2014. [Online]. Available:\n  <a href=\"https:\/\/arxiv.org\/abs\/1412.6980\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1412.6980<\/a><\/p>\n<\/blockquote>\n\n<p>To sum it up: Adam is RMSProp with momentum and bias correction. A very nice explanation is here: <a href=\"http:\/\/sebastianruder.com\/optimizing-gradient-descent\/index.html#adam\" rel=\"nofollow noreferrer\">http:\/\/sebastianruder.com\/optimizing-gradient-descent\/index.html#adam<\/a><\/p>\n"},{"AnswerId":"41442953","CreationDate":"2017-01-03T11:53:54.003","ParentId":null,"OwnerUserId":"1246987","Title":null,"Body":"<p>As you can see in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/754048a0453a04a761e112ae5d99c149eb9910dd\/tensorflow\/python\/training\/adam.py#L30\" rel=\"noreferrer\">adam.py<\/a> <code>AdamOptimizer<\/code> will adjust its learning rate.<\/p>\n\n<p>The learning rate you pass to the constructor just gives the initial value to start with.<\/p>\n\n<p>So yes, it does not make much sense to use exponential decay on <code>AdamOptimizer<\/code> but on gradient descent or momentum optimizer. See <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/6c875697dc9ceb0b83e283577c3961c652a92a07\/resnet\/resnet_main.py\" rel=\"noreferrer\">here<\/a> for an example.<\/p>\n"}]}
{"QuestionId":41443306,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><neural-network>","CreationDate":"2017-01-03T12:12:38.273","AcceptedAnswerId":null,"OwnerUserId":1524289.0,"Title":"Failed to implement wide and deep network tensorflow on iris dataset","Body":"<p>I've made some modification on the iris data set,i added two columns, country and forest, so i could add some categorical features on the data and tried to implement tensorflow' s <a href=\"https:\/\/www.tensorflow.org\/tutorials\/wide_and_deep\/\" rel=\"nofollow noreferrer\">wide and deep neural net<\/a> as the code below shows:<\/p>\n\n<pre><\/pre>\n\n<p>However i am getting a tensorflow that is not obvious :<\/p>\n\n<pre><\/pre>\n\n<p>I would appreciate alot for some guidance here since i tried to replicate the dataset type as much as i could in the given example<\/p>\n","answers":[{"AnswerId":"41461548","CreationDate":"2017-01-04T10:27:05.910","ParentId":null,"OwnerUserId":"1524289","Title":null,"Body":"<p>from this <a href=\"https:\/\/stackoverflow.com\/questions\/38924283\/tesnorflow-typeerror-expected-string-got-1-of-type-int64-instead\">answer<\/a> , i found out using numbers as categorical features(country) value were being inferred as numbers on this line<\/p>\n\n<pre><code>categorical_cols = {k: tf.SparseTensor( indices=[[i, 0] for i in range(df[k].size)], values=df[k].values, shape=[df[k].size, 1]) for k in CATEGORICAL_COLUMNS}\n<\/code><\/pre>\n\n<p>as tensorflow store its tensors in numpy array,before that it tries to inferr data type before storing them.<\/p>\n\n<p>As i did not specify their type after reading from csv file,they were sent as objects to the SparseTensor object which in turn, was checking the datatype.<\/p>\n\n<p>I hope it helps<\/p>\n"}]}
{"QuestionId":41443307,"AnswerCount":1,"Tags":"<java><android><tensorflow><apk><bazel>","CreationDate":"2017-01-03T12:12:39.007","AcceptedAnswerId":null,"OwnerUserId":7206319.0,"Title":"can not get APK through bazel build in tensorflow","Body":"<p>1: Get the recommended Bazel version listed at: www.tensorflow.org\/versions\/master\/get_started\/os_setup.html#source\nI use pip install: $ pip install tensorflow<\/p>\n\n<p>2: Install NDK and SDK,I install both SDK and NDK by Android Studio\n3: excute following:\n$ curl -L <a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip<\/a> -o \/tmp\/inception5h.zip\n$ curl -L <a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/mobile_multibox_v1.zip\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/mobile_multibox_v1.zip<\/a> -o \/tmp\/mobile_multibox_v1.zip<\/p>\n\n<p>$ unzip \/tmp\/inception5h.zip -d tensorflow\/examples\/android\/assets\/\n$ unzip \/tmp\/mobile_multibox_v1.zip -d tensorflow\/examples\/android\/assets\/<\/p>\n\n<p>4: $ bazel build \/\/tensorflow\/examples\/android:tensorflow_demo\n after this step i got this error.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41495294","CreationDate":"2017-01-05T21:47:15.733","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>What API level do you have set for your SDK in WORKSPACE file? It needs to be >= 23 or you'll get the missing symbol errors you're experiencing. Target api level needs to be >= 23 in AndroidManifest.xml as well.<\/p>\n\n<p>Note that the app will still run on devices with an api level >= 21, you just need to use a level 23 SDK to build.<\/p>\n"}]}
{"QuestionId":41444526,"AnswerCount":1,"Tags":"<python><deep-learning><caffe><face-detection><keypoint>","CreationDate":"2017-01-03T13:23:13.777","AcceptedAnswerId":41459317.0,"OwnerUserId":7039804.0,"Title":"training data augmentation in caffe along with its multilabel","Body":"<h1>How to randomly crop training images along with its position label.<\/h1>\n\n<p>I'm preprocessing images data and its labels for caffe input. I want to crop my training images randomly online (when caffe running) and I know caffe provide a random crop param in layer <a href=\"https:\/\/software.intel.com\/en-us\/articles\/training-and-deploying-deep-learning-networks-with-caffe-optimized-for-intel-architecture\" rel=\"nofollow noreferrer\">random crop<\/a>. The issues is that my label of each corresponding image is variant when image is transformed. So in this way I have to change my label together with transformed image. I try to search my issue long time, but no use. Then I realized maybe there were two ways to solve the issues:<\/p>\n\n<ol>\n<li>get the offset of random crop image on original image. But I googled it with no answer about it.<\/li>\n<li>I'll specify the random offset online, then crop image by the specific offset <a href=\"https:\/\/stackoverflow.com\/questions\/38588943\/caffe-fully-convolutional-cnn-how-to-use-the-crop-parameters\">crop specific offset<\/a>. But how can i generate a random number when caffe training time. <\/li>\n<li>I kown preprocessing the random crop image and its label offline may be a workable method. However, it will occupied considerable space of disk, for my data-set for training is too large. <\/li>\n<\/ol>\n\n<p><strong>my position corresponding label:<\/strong>\nI split each of training image as  grids. the label is the index of grid for specific keypoint in image. for instance, the label will be 11, when my head key-point is at position of  grid of image.<\/p>\n","answers":[{"AnswerId":"41459317","CreationDate":"2017-01-04T08:30:55.127","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can use a <a href=\"https:\/\/stackoverflow.com\/q\/41344168\/1714410\"><code>\"Python\"<\/code> layer<\/a> for this task, as you already <a href=\"https:\/\/stackoverflow.com\/questions\/41444526\/training-data-augmentation-in-caffe-along-with-its-multilabel#comment70095764_41444526\">pointed out<\/a> (and I suppose this is the easiest way around).<br>\nIf you are concerned with running time of this layer, you can make it multi-thread: have the layer invoke several threads to crop images and re-compute their labels. These threads will run at the background and should be able to provide the net with enough data in time for the next mini batch.<\/p>\n\n<p>See python <a href=\"https:\/\/docs.python.org\/2\/library\/threading.html\" rel=\"nofollow noreferrer\"><code>threading<\/code><\/a> module for the <code>Thread<\/code> class. You might also find it convenient to use <a href=\"https:\/\/docs.python.org\/2\/library\/queue.html\" rel=\"nofollow noreferrer\"><code>Queue<\/code><\/a> class to sync the threads and the main input layer.<\/p>\n"}]}
{"QuestionId":41444787,"AnswerCount":1,"Tags":"<caffe><pycaffe><lmdb><matcaffe>","CreationDate":"2017-01-03T13:37:53.427","AcceptedAnswerId":41452356.0,"OwnerUserId":6494707.0,"Title":"Do I have to transpose my gray scale image while creating LMDB database?","Body":"<p>I am converting a set of images and a set of ground truth images to LMDB database. The purpose is pixel-wise prediction. \nMost of the codes for creating LMDB from RGB images has a line indicating transpose in  order, as follows:<\/p>\n\n<pre><\/pre>\n\n<p>My image data (both train data and corresponding ground truth images) are grayscale (i.e., single channel). My question is, what is the purpose of this line and should I do this matrix transpose for my  images and ground truth or it is not necessary?<\/p>\n\n<p>Your help is appreciated. <\/p>\n","answers":[{"AnswerId":"41452356","CreationDate":"2017-01-03T21:05:22.850","ParentId":null,"OwnerUserId":"628538","Title":null,"Body":"<p>Caffe processes images in BGR format instead of RGB. Since most pre-trained models on ImageNet are trained in BGR format, this operation (transpose) is applied by default. If you are training on grey scale images this is not needed. Moreover, even if you are training from scratch on RGB images, it is not necessary.<\/p>\n"}]}
{"QuestionId":41445740,"AnswerCount":3,"Tags":"<c++><xcode><macos><pthreads><caffe>","CreationDate":"2017-01-03T14:27:32.137","AcceptedAnswerId":null,"OwnerUserId":7140808.0,"Title":"When I install Caffe on Mac, there is a fault \"clang: warning: argument unused during compilation: '-pthread' \"","Body":"<p>I feel that all dependences are installed, and I want to install Caffe without GPU, but when I do , there is an error:<\/p>\n\n<pre><\/pre>\n\n<p>I do not understand why this happens and it seems like my Xcode may have some problems? Could someone help me?<\/p>\n","answers":[{"AnswerId":"42965223","CreationDate":"2017-03-23T00:42:37.127","ParentId":null,"OwnerUserId":"6448577","Title":null,"Body":"<p>I had a similar problem as well. It could be the Makefile.config. Please check the paths you provided for <strong>PYTHON_INCLUDE<\/strong> and <strong>PYTHON_LIB<\/strong>.<\/p>\n\n<p>I had given the brew python path to <strong>PYTHON_INCLUDE<\/strong> and system path to <strong>PYTHON_LIB<\/strong>. I also made sure that the caffe python was added to the python path before compiling.Hope this information is of help.<\/p>\n"},{"AnswerId":"49275476","CreationDate":"2018-03-14T10:37:04.910","ParentId":null,"OwnerUserId":"202229","Title":null,"Body":"<p>There are two separate things here:<\/p>\n\n<ol>\n<li><em>\"clang: warning: argument unused during compilation: '-pthread'\"<\/em> is an ignoreable (although annoying) compiler <strong>warning<\/strong> (not error) that you often get with MacOS clang, these days it has pthreads support builtin. Ignore it, or suppress it with <code>-Wno-error=unused-command-line-argument<\/code><\/li>\n<li>whereas <em>\"Undefined symbols for architecture x86_64: ... clang: error: linker command failed with exit code \"<\/em> is an actual <strong>error<\/strong>. Fix that. (Looks like a missing header file to me, so check your <code>-I...<\/code> include paths. I don't know Caffe so I can't comment more on that.)<\/li>\n<\/ol>\n"},{"AnswerId":"41482181","CreationDate":"2017-01-05T10:08:52.590","ParentId":null,"OwnerUserId":"6573448","Title":null,"Body":"<p>make sure you install openCV before building caffe.<\/p>\n\n<pre><code>brew install -vd snappy leveldb gflags glog szip lmdb\n# need the homebrew science source for OpenCV and hdf5\nbrew tap homebrew\/science\nbrew install hdf5 opencv\n<\/code><\/pre>\n\n<p>please go threw <a href=\"http:\/\/caffe.berkeleyvision.org\/install_osx.html\" rel=\"nofollow noreferrer\">caffe OS X Installation<\/a>.<\/p>\n"}]}
{"QuestionId":41446135,"AnswerCount":1,"Tags":"<java><image-processing><caffe><conv-neural-network>","CreationDate":"2017-01-03T14:48:48.163","AcceptedAnswerId":null,"OwnerUserId":7241519.0,"Title":"age genderclassification using caffe in windows platform","Body":"<p>To achieve age and gender classification, i tried with the framework caffe. <\/p>\n\n<p>1) i downloaded the caffemodel from the link <a href=\"http:\/\/www.openu.ac.il\/home\/hassner\/projects\/cnn_agegender\" rel=\"nofollow noreferrer\">http:\/\/www.openu.ac.il\/home\/hassner\/projects\/cnn_agegender<\/a> and later on downloaded the windows version (<a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/windows\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/tree\/windows<\/a>) i maintained the dependencies Visual Studio 14 2015 Win64 and cmake 3.6.3\nthe error i got is as follows<\/p>\n\n<pre><\/pre>\n\n<p>Please suggest the way forward. And please let me know the steps how to achieve age and gender classification. <\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41457645","CreationDate":"2017-01-04T06:36:13.357","ParentId":null,"OwnerUserId":"7241519","Title":null,"Body":"<p>After i installed visual studio 2015 and then created C++ project which implicitly downloaded C++ libraries. Able to run the <strong>build_win<\/strong> successfully.<\/p>\n"}]}
{"QuestionId":41448559,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-03T16:57:22.157","AcceptedAnswerId":41449558.0,"OwnerUserId":7287271.0,"Title":"What's the point of having more than one tf.Graph?","Body":"<p>What's the point of having more than one ?<\/p>\n\n<p>I'm thinking specifically about hyperparameter tuning of machine learning models, where a model is either a graph on its own, or several models are defined as disconnected components within the same graph.<\/p>\n\n<p>I understand that having more than one  is bad because task scheduling cannot be done right, so I assume it's possible to have multiple  objects in one session (though  begs to differ) but what would be the point of doing that instead of having several components with something like  instead? Is it mostly a matter of what gets saved with , visualized in TensorBoard, and so on? Which method is preferable? Should models share a graph or each have their own for hyperparameter tuning?<\/p>\n\n<p>It seems simpler to just use  and to forget about both  and  throughout the rest of the code base. What am I missing?<\/p>\n","answers":[{"AnswerId":"41449558","CreationDate":"2017-01-03T17:56:32.493","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>If you have a single session, then there's no point in having multiple graphs. Session is linked to a graph, so if you try to run an element from another graph, you'll get <code>xyz is not an element of this graph<\/code> error.<\/p>\n\n<p>It makes sense to have multiple graphs when you have multiple sessions. For instance, suppose you are using distributed TensorFlow, but also want to do some computations locally. You could do something like this<\/p>\n\n<pre><code>local_session = tf.Session(\"\", graph=local_graph)\nremote_session = tf.Session(\"grpc:\/\/...\", graph=remote_graph)\n<\/code><\/pre>\n\n<p>You could potentially use two sessions with the same <code>tf.Graph<\/code> object however, however any additions to this object <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/64edd34ce69b4a8033af5d217cb8894105297d8a\/tensorflow\/python\/client\/session.py#L1037\" rel=\"nofollow noreferrer\">will result<\/a> in a <code>TF_ExtendGraph<\/code> call on the next <code>session.run<\/code> even if it's not necessary for that session. In other words, sharing the graph means sending &lt;=2GB graph description to all sessions when the graph is modified.<\/p>\n"}]}
{"QuestionId":41449338,"AnswerCount":2,"Tags":"<python><tensorflow><multi-gpu>","CreationDate":"2017-01-03T17:44:27.433","AcceptedAnswerId":null,"OwnerUserId":1882931.0,"Title":"TensorFlow for MultiGPU","Body":"<p>If someone can help me understand the situation it would be great. Thanks in advance.\nMy setup:\nOS: Ubuntu 16.04, 2 Titan X GPUs. TensorFlow (version 0.12.1) installed in a conda environment using pip as on TF docs. Python 3.5.<\/p>\n\n<p>Code:\nI ran the following code to test my 2 GPU setup. Once each with  and . The outputs are shown below. <\/p>\n\n<p>Questions:\n1) When I run with . The timings on CPU and GPU are identical. But with  I see that the GPU is faster (as I had expected). Why is  slower on GPU? What am I missing?\n2) I have fixed the global seed and the local seed. Why are the outputs within the GPUs different for the  case?<\/p>\n\n<p>Thanks a lot for any insights in advance.<\/p>\n\n<pre><\/pre>\n\n<p>Timings with tf.random_uniform():<\/p>\n\n<pre><\/pre>\n\n<p>Timings with tf.zeros():<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41449981","CreationDate":"2017-01-03T18:24:28.230","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>I suspect this this related to GPU kernel optimization. If you \"pre-warm\" your GPU by running the same computation shape, the next execution is much faster. There's PTX compilation that <a href=\"https:\/\/stackoverflow.com\/questions\/40410210\/tensorflow-2-gpu-slower-then-single-gpu\/40430717#40430717\">adds a couple of seconds<\/a> to the first usage of kernel on a GPU in a process, but it's peculiar that your runtime increases with size of the matrix, perhaps there's some profiling going on as well.<\/p>\n\n<p>Note that without <code>tf.OptimizerOptions.L0<\/code> it becomes implausibly fast, so there's some caching happening as well.<\/p>\n\n<pre><code>shape = (6000, 6000)\nwith tf.device(\"\/gpu:0\"):\n    random_matrix_gpu = tf.zeros(shape)\n    result_op_gpu = tf.reduce_sum(tf.matmul(random_matrix_gpu,tf.transpose(random_matrix_gpu)))\nwith tf.device(\"\/cpu:0\"):\n    random_matrix_cpu = tf.zeros(shape)\n    result_op_cpu = \n\ntf.reduce_sum(tf.matmul(random_matrix_cpu,tf.transpose(random_matrix_cpu)))\nconfig = tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))\nsess = tf.Session(config=config)\n\ndef profile(op, msg):\n    start_time = time.time()\n    sess.run(op)\n    print(msg, time.time()-start_time)\n\nprofile(result_op_cpu, \"cpu1\")\nprofile(result_op_cpu, \"cpu2\")\nprofile(result_op_gpu, \"gpu1\")\nprofile(result_op_gpu, \"gpu2\")\n<\/code><\/pre>\n\n<p>I see this:<\/p>\n\n<pre><code>cpu1 1.716048240661621\ncpu2 1.509080171585083\ngpu1 4.192790746688843\ngpu2 0.13361549377441406\n<\/code><\/pre>\n"},{"AnswerId":"41454591","CreationDate":"2017-01-04T00:49:48.520","ParentId":null,"OwnerUserId":"1882931","Title":null,"Body":"<p>Thanks Yaroslav! I provide the code and results from my run, just in case somebody else is interested. If you try the code please be patient for a few minutes.<\/p>\n\n<p>Code:<\/p>\n\n<pre><code>import sys\nimport numpy as np\nimport tensorflow as tf\nfrom datetime import datetime\n\n\ndevice_names = [\"\/cpu:0\", \"\/gpu:0\", \"\/gpu:1\"]\nshapes = [(3000, 3000), (6000, 6000), (9000, 9000), (12000, 12000)]\nmessages = [\"RESULTS\\n\"]\n\ndef timing_run(matrix_type, config_name, warmup):\n    configs = {\"simple\": tf.ConfigProto(log_device_placement=False),\n               \"optim\": tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))}\n\n    messages.append(\"matrix={}+config={}+warmup={}\".format(matrix_type, config_name, warmup))\n    tf.set_random_seed(1234)\n    for device_name in device_names:\n        for shape in shapes:\n            with tf.device(device_name):\n                if matrix_type == \"random_uniform\":\n                    random_matrix = tf.random_uniform(shape=shape,\n                                                      minval=0,\n                                                      maxval=1,\n                                                      seed=1234)\n                else:\n                    random_matrix = tf.zeros(shape)\n                result_op = tf.reduce_sum(tf.matmul(random_matrix,tf.transpose(random_matrix)))\n\n            session = tf.Session(config=configs[config_name])\n            result1, result2 = -1.0, -1.0\n            # warm up\n            start_time1 = datetime.now()\n            result1 = session.run(result_op)\n            time_diff1 = datetime.now() - start_time1\n            messages.append((device_name,\n                             \"shape = {}\".format(shape),\n                             \"times = {} seconds\".format(time_diff1.total_seconds()),\n                             \"result = {}\".format(result1)))\n            if warmup:\n                # warmed up - runs if warmup=True.\n                start_time2 = datetime.now()\n                result2 = session.run(result_op)\n                time_diff2 = datetime.now() - start_time2\n                messages.append((device_name,\n                                 \"shape = {}\".format(shape),\n                                 \"times = {} seconds\".format(time_diff2.total_seconds()),\n                                 \"result = {}\".format(result1),\n                                 \"*****WARMED UP*****\"))\n            session.close()\n        messages.append(\"++++++++++++++++++++++++++++++++++++++++++++++++++++\")\n    messages.append(\"\\n\\n\")\n\n\n\nif __name__ == \"__main__\":\n    timing_run(matrix_type=\"random_uniform\", config_name=\"simple\", warmup=False)\n    timing_run(matrix_type=\"random_uniform\", config_name=\"simple\", warmup=True)\n    timing_run(matrix_type=\"random_uniform\", config_name=\"optim\", warmup=False)\n    timing_run(matrix_type=\"zeros\", config_name=\"simple\", warmup=False)\n    timing_run(matrix_type=\"zeros\", config_name=\"simple\", warmup=True)\n    timing_run(matrix_type=\"zeros\", config_name=\"optim\", warmup=False)\n\n    # print timings\n    for e in messages:\n        print(e)\n<\/code><\/pre>\n\n<p>Summary:<\/p>\n\n<pre><code>matrix=random_uniform+config=simple+warmup=False\n('\/cpu:0', 'shape = (3000, 3000)', 'times = 0.428429 seconds', 'result = 6754431488.0')\n('\/cpu:0', 'shape = (6000, 6000)', 'times = 2.806464 seconds', 'result = 54023852032.0')\n('\/cpu:0', 'shape = (9000, 9000)', 'times = 9.36232 seconds', 'result = 184425938944.0')\n('\/cpu:0', 'shape = (12000, 12000)', 'times = 22.376751 seconds', 'result = 439655661568.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:0', 'shape = (3000, 3000)', 'times = 0.392957 seconds', 'result = 6754390016.0')\n('\/gpu:0', 'shape = (6000, 6000)', 'times = 0.082889 seconds', 'result = 54006833152.0')\n('\/gpu:0', 'shape = (9000, 9000)', 'times = 0.221844 seconds', 'result = 182251814912.0')\n('\/gpu:0', 'shape = (12000, 12000)', 'times = 0.438476 seconds', 'result = 431995879424.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:1', 'shape = (3000, 3000)', 'times = 0.483864 seconds', 'result = 6754393088.0')\n('\/gpu:1', 'shape = (6000, 6000)', 'times = 0.097571 seconds', 'result = 54006833152.0')\n('\/gpu:1', 'shape = (9000, 9000)', 'times = 0.250176 seconds', 'result = 182252044288.0')\n('\/gpu:1', 'shape = (12000, 12000)', 'times = 0.473314 seconds', 'result = 431996567552.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n\n\n\nmatrix=random_uniform+config=simple+warmup=True\n('\/cpu:0', 'shape = (3000, 3000)', 'times = 0.475717 seconds', 'result = 6754431488.0')\n('\/cpu:0', 'shape = (3000, 3000)', 'times = 0.379542 seconds', 'result = 6754431488.0', '*****WARMED UP*****')\n('\/cpu:0', 'shape = (6000, 6000)', 'times = 2.856803 seconds', 'result = 54023852032.0')\n('\/cpu:0', 'shape = (6000, 6000)', 'times = 2.798967 seconds', 'result = 54023852032.0', '*****WARMED UP*****')\n('\/cpu:0', 'shape = (9000, 9000)', 'times = 9.447787 seconds', 'result = 184425938944.0')\n('\/cpu:0', 'shape = (9000, 9000)', 'times = 9.385646 seconds', 'result = 184425938944.0', '*****WARMED UP*****')\n('\/cpu:0', 'shape = (12000, 12000)', 'times = 21.752967 seconds', 'result = 439655661568.0')\n('\/cpu:0', 'shape = (12000, 12000)', 'times = 21.832136 seconds', 'result = 439655661568.0', '*****WARMED UP*****')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:0', 'shape = (3000, 3000)', 'times = 0.067066 seconds', 'result = 6754394624.0')\n('\/gpu:0', 'shape = (3000, 3000)', 'times = 0.008072 seconds', 'result = 6754394624.0', '*****WARMED UP*****')\n('\/gpu:0', 'shape = (6000, 6000)', 'times = 0.123611 seconds', 'result = 54006833152.0')\n('\/gpu:0', 'shape = (6000, 6000)', 'times = 0.057391 seconds', 'result = 54006833152.0', '*****WARMED UP*****')\n('\/gpu:0', 'shape = (9000, 9000)', 'times = 0.248432 seconds', 'result = 182251913216.0')\n('\/gpu:0', 'shape = (9000, 9000)', 'times = 0.18535 seconds', 'result = 182251913216.0', '*****WARMED UP*****')\n('\/gpu:0', 'shape = (12000, 12000)', 'times = 0.48081 seconds', 'result = 431996043264.0')\n('\/gpu:0', 'shape = (12000, 12000)', 'times = 0.412447 seconds', 'result = 431996043264.0', '*****WARMED UP*****')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:1', 'shape = (3000, 3000)', 'times = 0.105071 seconds', 'result = 6754395648.0')\n('\/gpu:1', 'shape = (3000, 3000)', 'times = 0.008107 seconds', 'result = 6754395648.0', '*****WARMED UP*****')\n('\/gpu:1', 'shape = (6000, 6000)', 'times = 0.137264 seconds', 'result = 54006849536.0')\n('\/gpu:1', 'shape = (6000, 6000)', 'times = 0.064462 seconds', 'result = 54006849536.0', '*****WARMED UP*****')\n('\/gpu:1', 'shape = (9000, 9000)', 'times = 0.280302 seconds', 'result = 182251831296.0')\n('\/gpu:1', 'shape = (9000, 9000)', 'times = 0.191399 seconds', 'result = 182251831296.0', '*****WARMED UP*****')\n('\/gpu:1', 'shape = (12000, 12000)', 'times = 0.509208 seconds', 'result = 431996534784.0')\n('\/gpu:1', 'shape = (12000, 12000)', 'times = 0.4263 seconds', 'result = 431996534784.0', '*****WARMED UP*****')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n\n\n\nmatrix=random_uniform+config=optim+warmup=False\n('\/cpu:0', 'shape = (3000, 3000)', 'times = 0.552631 seconds', 'result = 6754431488.0')\n('\/cpu:0', 'shape = (6000, 6000)', 'times = 2.894024 seconds', 'result = 54023852032.0')\n('\/cpu:0', 'shape = (9000, 9000)', 'times = 9.394226 seconds', 'result = 184425938944.0')\n('\/cpu:0', 'shape = (12000, 12000)', 'times = 21.870817 seconds', 'result = 439655661568.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:0', 'shape = (3000, 3000)', 'times = 0.107416 seconds', 'result = 6754392576.0')\n('\/gpu:0', 'shape = (6000, 6000)', 'times = 0.163633 seconds', 'result = 54006804480.0')\n('\/gpu:0', 'shape = (9000, 9000)', 'times = 0.304741 seconds', 'result = 182251667456.0')\n('\/gpu:0', 'shape = (12000, 12000)', 'times = 0.526494 seconds', 'result = 431995944960.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:1', 'shape = (3000, 3000)', 'times = 0.119625 seconds', 'result = 6754394624.0')\n('\/gpu:1', 'shape = (6000, 6000)', 'times = 0.203158 seconds', 'result = 54006800384.0')\n('\/gpu:1', 'shape = (9000, 9000)', 'times = 0.317646 seconds', 'result = 182251978752.0')\n('\/gpu:1', 'shape = (12000, 12000)', 'times = 0.544184 seconds', 'result = 431996076032.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n\n\n\nmatrix=zeros+config=simple+warmup=False\n('\/cpu:0', 'shape = (3000, 3000)', 'times = 0.632157 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (6000, 6000)', 'times = 2.901679 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (9000, 9000)', 'times = 9.345713 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (12000, 12000)', 'times = 21.707619 seconds', 'result = 0.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:0', 'shape = (3000, 3000)', 'times = 0.498451 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (6000, 6000)', 'times = 2.900121 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (9000, 9000)', 'times = 9.4296 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (12000, 12000)', 'times = 21.750406 seconds', 'result = 0.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:1', 'shape = (3000, 3000)', 'times = 0.523286 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (6000, 6000)', 'times = 2.887522 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (9000, 9000)', 'times = 9.377383 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (12000, 12000)', 'times = 21.639043 seconds', 'result = 0.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n\n\n\nmatrix=zeros+config=simple+warmup=True\n('\/cpu:0', 'shape = (3000, 3000)', 'times = 0.520212 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (3000, 3000)', 'times = 0.000172 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/cpu:0', 'shape = (6000, 6000)', 'times = 2.914485 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (6000, 6000)', 'times = 0.000166 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/cpu:0', 'shape = (9000, 9000)', 'times = 9.346122 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (9000, 9000)', 'times = 0.000207 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/cpu:0', 'shape = (12000, 12000)', 'times = 21.715376 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (12000, 12000)', 'times = 0.0002 seconds', 'result = 0.0', '*****WARMED UP*****')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:0', 'shape = (3000, 3000)', 'times = 0.556841 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (3000, 3000)', 'times = 0.000234 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/gpu:0', 'shape = (6000, 6000)', 'times = 2.936608 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (6000, 6000)', 'times = 0.000244 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/gpu:0', 'shape = (9000, 9000)', 'times = 9.34956 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (9000, 9000)', 'times = 0.000246 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/gpu:0', 'shape = (12000, 12000)', 'times = 21.634354 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (12000, 12000)', 'times = 0.000221 seconds', 'result = 0.0', '*****WARMED UP*****')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:1', 'shape = (3000, 3000)', 'times = 0.562244 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (3000, 3000)', 'times = 0.000255 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/gpu:1', 'shape = (6000, 6000)', 'times = 2.961658 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (6000, 6000)', 'times = 0.000237 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/gpu:1', 'shape = (9000, 9000)', 'times = 9.308582 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (9000, 9000)', 'times = 0.000239 seconds', 'result = 0.0', '*****WARMED UP*****')\n('\/gpu:1', 'shape = (12000, 12000)', 'times = 21.707127 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (12000, 12000)', 'times = 0.000261 seconds', 'result = 0.0', '*****WARMED UP*****')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n\n\n\nmatrix=zeros+config=optim+warmup=False\n('\/cpu:0', 'shape = (3000, 3000)', 'times = 0.560451 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (6000, 6000)', 'times = 2.978946 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (9000, 9000)', 'times = 9.3279 seconds', 'result = 0.0')\n('\/cpu:0', 'shape = (12000, 12000)', 'times = 21.694664 seconds', 'result = 0.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:0', 'shape = (3000, 3000)', 'times = 0.249778 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (6000, 6000)', 'times = 0.365332 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (9000, 9000)', 'times = 0.663667 seconds', 'result = 0.0')\n('\/gpu:0', 'shape = (12000, 12000)', 'times = 1.032716 seconds', 'result = 0.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n('\/gpu:1', 'shape = (3000, 3000)', 'times = 0.299856 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (6000, 6000)', 'times = 0.294592 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (9000, 9000)', 'times = 0.55067 seconds', 'result = 0.0')\n('\/gpu:1', 'shape = (12000, 12000)', 'times = 0.806868 seconds', 'result = 0.0')\n++++++++++++++++++++++++++++++++++++++++++++++++++++\n<\/code><\/pre>\n"}]}
{"QuestionId":41451273,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-03T19:51:23.880","AcceptedAnswerId":41452102.0,"OwnerUserId":4975126.0,"Title":"TensorFlow: Specifying storage of layer activations","Body":"<p><strong>Motivation:<\/strong> It's easy to imagine a scenario where one could amass a large number of layers so as to overwhelm a GPU's memory capacity.<\/p>\n\n<p>A potential solution to this issue would be to take advantage of any backwards computability between layers. I.e. layer 's output is a function of the next layers output:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AI0p4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AI0p4.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here you could simply store the final such layer activation in memory and compute activations in a 'trickle back' fashion during backprop.<\/p>\n\n<hr>\n\n<p><strong>Problem:<\/strong> I'm unfamiliar with manual memory management in TensorFlow and haven't been able to find information on how one might go about specifying what variables TensorFlow should keep in memory and which can be discarded.<\/p>\n\n<p>The other side of implementing this is specifying a custom gradient computation; this I feel is well-enough documented and should be achievable. However, if anyone has any caveats related to this, I'd appreciate hearing them.<\/p>\n","answers":[{"AnswerId":"41452102","CreationDate":"2017-01-03T20:48:37.987","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>As I understand, you want to save memory by discarding intermediate results and recomputing them later. I can see two potential approaches. One is to rewire the graph to do this recomputation, and another is to use persistent tensors for intermediate results and control their deletion manually.<\/p>\n\n<p>For the first approach, consider the following computation and its gradient graph.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/aM1G8.png\" width=\"342\" height=\"521\"><\/p>\n\n<p>You use <code>contrib.graph_editor<\/code> to modify the graph as follows.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/PEeMh.png\" width=\"347\" height=\"521\"><\/p>\n\n<p>Note that now it's possible to pick an execution order which only needs enough memory for 2 activations. However, TensorFlow typically does not pick this execution order, and instead computes both <code>a2<\/code> tensors in the beginning, hence requiring enough memory to store 4 activations at peak. (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5816#issuecomment-268710077\" rel=\"nofollow noreferrer\">caterpillar graph<\/a> for an extreme example).<\/p>\n\n<p>The solution is to add control dependencies to force a particular execution order. <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/zVMlG.png\" width=\"347\" height=\"521\"><\/p>\n\n<p>This would force second <code>a2<\/code> node to be computed after <code>b3<\/code>. Since TensorFlow releases memory as soon as the tensor is no longer needed, all execution orders in this graph need enough peak memory to store 2 activations instead of 3.<\/p>\n\n<p>Here's a <a href=\"https:\/\/github.com\/yaroslavvb\/notebooks\/blob\/master\/simple_rewiring.ipynb\" rel=\"nofollow noreferrer\">notebook<\/a> that implements the example above.<\/p>\n\n<p>If the computation that goes a2->a3 is invertible, you would rewire your graph as follows<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/XkfVG.png\" width=\"347\" height=\"521\"><\/p>\n\n<p>The second approach would be to use Persistent Tensors. You can tell TensorFlow to keep certain tensors after the run call completes. Unlike Variables, these objects can be deleted to free the memory. You have a downside of more <code>.run<\/code> calls, each incurring additional 200 usec delay per run call, but it may be easier than all the graph rewiring. I haven't explored this route, but here's an <a href=\"https:\/\/stackoverflow.com\/questions\/41001298\/decouple-dequeue-operation-from-gradient-loss-computation\/41002892#41002892\">example<\/a> of using persistent tensors  (deletion is done using <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard0\/tf.delete_session_tensor.md\" rel=\"nofollow noreferrer\">delete_session_tensor<\/a> command)<\/p>\n"}]}
{"QuestionId":41451534,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-01-03T20:09:53.123","AcceptedAnswerId":null,"OwnerUserId":5015382.0,"Title":"Python, Tensorflow: Random Shuffle Queue Error (insufficient elements) while experimenting with \"Tensorflow for Machine Learning\"","Body":"<p>I am running the code found on this <a href=\"https:\/\/github.com\/backstopmedia\/tensorflowbook\/blob\/master\/chapters\/04_machine_learning_basics\/softmax.py\" rel=\"nofollow noreferrer\">GitHub repository<\/a>.\nUnfortunately, it always stops with an error, namely:<\/p>\n\n<pre><\/pre>\n\n<p>Anybody, who knows how to solve this issue?<\/p>\n\n<p>Sidenote1: I have the iris.data dataset from <a href=\"https:\/\/archive.ics.uci.edu\/ml\/machine-learning-databases\/iris\/\" rel=\"nofollow noreferrer\">here<\/a> and it has 150 rows.<\/p>\n\n<p>Sidenote2: I run Python 3.5 and just downloaded Tensorflow today<\/p>\n\n<p>Sidenote3: Don't forget to change the print statement to a function statement.<\/p>\n","answers":[]}
{"QuestionId":41451757,"AnswerCount":2,"Tags":"<audio><machine-learning><tensorflow><neural-network>","CreationDate":"2017-01-03T20:26:19.623","AcceptedAnswerId":null,"OwnerUserId":3633116.0,"Title":"How utterance length affect neural network in speaker recognition?","Body":"<p>I'm learning neural networks and trying to create speaker recognition system with tensorflow. I wanted to know how utterance length affect neural network. For example I have 1000 different sound recordings with the same lengths and 1000 different sound recordings with different lenghts. So how theoretically will work neural network with these kind of datas? Will neural network with database of same length recordings will do better or worse? Why?<\/p>\n","answers":[{"AnswerId":"41469651","CreationDate":"2017-01-04T17:16:31.573","ParentId":null,"OwnerUserId":"1397946","Title":null,"Body":"<p>It depends on the type of neural network. When design such, you usually specify the number of input neurons, sou can't feed it with data of arbitrary length. In case of longer sequences you have to either crop your data or use a sliding window.<\/p>\n\n<p>However, some neural networks allow you to process arbitrary sequence of inputs, like e.g. <a href=\"https:\/\/en.wikipedia.org\/wiki\/Recurrent_neural_network\" rel=\"nofollow noreferrer\">Recurrent Neural Network<\/a>. The latter seem like a very good candidate for your problem. <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\/\" rel=\"nofollow noreferrer\">Here<\/a> is a good article that describes implementation of particular type of RNN, called <a href=\"https:\/\/en.wikipedia.org\/wiki\/Long_short-term_memory\" rel=\"nofollow noreferrer\">Long Short-Term Memory<\/a> that work nicely with speech recognition. <\/p>\n"},{"AnswerId":"41681058","CreationDate":"2017-01-16T16:41:13.657","ParentId":null,"OwnerUserId":"2641587","Title":null,"Body":"<p>I assume your question can be reformulated as <em>How a neural network can process audio of different length?<\/em> <\/p>\n\n<p>The trick is that the signal of an arbitrary size is converted into a sequence of fixed-size feature vectors. See my answers <a href=\"https:\/\/stackoverflow.com\/questions\/41047916\/how-to-train-a-neural-network-with-different-size-of-input\/41056840\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/41047151\/how-to-classify-continuous-audio\/41057147#41057147\">here<\/a>.<\/p>\n"}]}
{"QuestionId":41453083,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-03T22:01:16.100","AcceptedAnswerId":41454763.0,"OwnerUserId":3701294.0,"Title":"why does get_tensor_by_name require appending a port to the tensor name","Body":"<p>I know that when getting a tensor by name that I must append an output index<\/p>\n\n<p>ex)<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Where :0 is the output index. But why is this necessary?<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Graph.get_tensor_by_name\" rel=\"noreferrer\">here<\/a> is a  link to get_tensor_by_name in tensorflow's docs. However, it does not mention specifying an output index.<\/p>\n","answers":[{"AnswerId":"48892990","CreationDate":"2018-02-20T19:39:59.343","ParentId":null,"OwnerUserId":"9352849","Title":null,"Body":"<p>The crux to understand the <code>get_tensor_by_name<\/code> function is to realize that a <code>TensorFlow<\/code> model is first specified as a graph during the <strong>construction<\/strong> phase.  In the question, <code>example<\/code> is the name of an operation node in this graph.   When this operation is run during the <strong>execution<\/strong> phase, output is produced by this <code>example<\/code> node.  This output is indexed by a number indicating its production order.  In this case, it is <code>1<\/code>.<\/p>\n\n<p>The graph and its outputs described above were saved during an earlier run.  Later in a different project execution, this run time state was restored possibly by using <code>tf.train.import_meta_graph<\/code> and <code>graph<\/code> becomes a reference to it.  The function <code>get_tensor_by_name<\/code> simply recovers a reference to the saved second output of the <code>example<\/code> node from the restored run time meta graph.  You can now make use this recovered tensor in your own task during another session run.<\/p>\n"},{"AnswerId":"41454763","CreationDate":"2017-01-04T01:12:57.807","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In TensorFlow, names are given to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Operation\" rel=\"noreferrer\"><code>tf.Operation<\/code><\/a> objects (which correspond to nodes in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Graph\" rel=\"noreferrer\"><code>tf.Graph<\/code><\/a>), and a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Tensor\" rel=\"noreferrer\"><code>tf.Tensor<\/code><\/a> object is named for the <code>tf.Operation<\/code> that produces it as an output.<\/p>\n\n<p>Since a <code>tf.Operation<\/code> can have more than one output, to name a <code>tf.Tensor<\/code> uniquely, we include its index in the the outputs as part of its name.<\/p>\n\n<p>Therefore, we chose the following format for the name of a <code>tf.Tensor<\/code> object, which is also the value returned by the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Tensor.name\" rel=\"noreferrer\"><code>tf.Tensor.name<\/code><\/a> property:<\/p>\n\n<pre><code>&lt;name of operation&gt;:&lt;index of output&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":41454362,"AnswerCount":2,"Tags":"<tensorflow><computer-vision><deep-learning><tensorboard>","CreationDate":"2017-01-04T00:18:01.270","AcceptedAnswerId":null,"OwnerUserId":7367084.0,"Title":"Tensorboard embeddings does not display tensors?","Body":"<p>Tensorboard provides embedding visualization of tensorflow variables by using . The following is a working example (from <a href=\"https:\/\/stackoverflow.com\/a\/41370610\">this answer<\/a>)<\/p>\n\n<pre><\/pre>\n\n<p>How can I visualize embeddings from a tensorflow tensor, like  in the code above?<\/p>\n\n<p>Simply replacing<\/p>\n\n<pre><\/pre>\n\n<p>with<\/p>\n\n<pre><\/pre>\n\n<p>doesn't work, because of the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know of an alternative to generate tensorboard embedding visualizations of a tf.tensor?<\/p>\n","answers":[{"AnswerId":"42683137","CreationDate":"2017-03-08T22:29:38.867","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/tensor_summary\" rel=\"nofollow noreferrer\">tf.summary.tensor_summary<\/a> will save a summary for an arbitrary Tensor.<\/p>\n"},{"AnswerId":"41539759","CreationDate":"2017-01-09T01:26:47.823","ParentId":null,"OwnerUserId":"2082009","Title":null,"Body":"<p>You could create a new variable you can assign the value of <code>y<\/code> to.<\/p>\n\n<pre><code>y_var = tf.Variable(tf.shape(y))\nsaver = tf.train.Saver([y_var])\nassign_op = y_var.assign(y)\n...\n# Training\nsess.run((loss, assign_op))\nsaver.save(...)\n<\/code><\/pre>\n"}]}
{"QuestionId":41454511,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><neural-network><autoencoder>","CreationDate":"2017-01-04T00:40:13.637","AcceptedAnswerId":41454722.0,"OwnerUserId":3993270.0,"Title":"TensorFlow: how is dataset.train.next_batch defined?","Body":"<p>I am trying to learn TensorFlow and studying the example at: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3_NeuralNetworks\/autoencoder.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3_NeuralNetworks\/autoencoder.ipynb<\/a><\/p>\n\n<p>I then have some questions in the code below:<\/p>\n\n<pre><\/pre>\n\n<p>Since mnist is just a dataset, what exactly does  mean? How was the  defined?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41454722","CreationDate":"2017-01-04T01:07:10.747","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>mnist<\/code> object is returned from the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7c36309c37b04843030664cdc64aca2bb7d6ecaa\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L189\" rel=\"noreferrer\"><code>read_data_sets()<\/code> function<\/a> defined in the <code>tf.contrib.learn<\/code> module. The <code>mnist.train.next_batch(batch_size)<\/code> method is implemented <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7c36309c37b04843030664cdc64aca2bb7d6ecaa\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160\" rel=\"noreferrer\">here<\/a>, and it returns a tuple of two arrays, where the first represents a batch of <code>batch_size<\/code> MNIST images, and the second represents a batch of <code>batch-size<\/code> labels corresponding to those images.<\/p>\n\n<p>The images are returned as a 2-D NumPy array of size <code>[batch_size, 784]<\/code> (since there are 784 pixels in an MNIST image), and the labels are returned as either a 1-D NumPy array of size <code>[batch_size]<\/code> (if <code>read_data_sets()<\/code> was called with <code>one_hot=False<\/code>) or a 2-D NumPy array of size <code>[batch_size, 10]<\/code> (if <code>read_data_sets()<\/code> was called with <code>one_hot=True<\/code>).<\/p>\n"}]}
{"QuestionId":41455101,"AnswerCount":10,"Tags":"<tensorflow><machine-learning><neural-network><deep-learning><cross-entropy>","CreationDate":"2017-01-04T02:02:31.570","AcceptedAnswerId":null,"OwnerUserId":6773030.0,"Title":"What is the meaning of the word logits in TensorFlow?","Body":"<p>In the following TensorFlow function, we must feed the activation of artificial neurons in the final layer. That I understand. But I don't understand why it is called logits? Isn't that a mathematical function? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47278567","CreationDate":"2017-11-14T05:51:41.580","ParentId":null,"OwnerUserId":"4422118","Title":null,"Body":"<p>They are basically the fullest learned model you can get from the network, before it's been squashed down to apply to only the number of classes we are interested in.  Check out how some researchers use them to train a shallow neural net based on what a deep network has learned:  <a href=\"https:\/\/arxiv.org\/pdf\/1312.6184.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1312.6184.pdf<\/a><\/p>\n\n<p>It's kind of like how when learning a subject in detail, you will learn a great many minor points, but then when teaching a student, you will try to compress it to the simplest case.  If the student now tried to teach, it'd be quite difficult, but would be able to describe it just well enough to use the language.<\/p>\n"},{"AnswerId":"52111173","CreationDate":"2018-08-31T08:11:44.640","ParentId":null,"OwnerUserId":"207661","Title":null,"Body":"<p><strong>Summary<\/strong><\/p>\n\n<p>In context of deep learning the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/estimators\/cnn#logits_layer\" rel=\"noreferrer\">logits layer<\/a> means the layer that feeds in to softmax (or other such normalization). The output of the softmax are the probabilities for the classification task and its input is logits layer. The logits layer typically produces values from -infinity to +infinity and the softmax layer transforms it to values from 0 to 1.<\/p>\n\n<p><strong>Historical Context<\/strong><\/p>\n\n<p>Where does this term comes from? In 1930s and 40s, several people were trying to adapt linear regression to the problem of predicting probabilities. However linear regression produces output from -infinity to +infinity while for probabilities our desired output is 0 to 1. One way to do this is by somehow mapping the probabilities 0 to 1 to -infinity to +infinity and then use linear regression as usual. One such mapping is cumulative normal distribution that was used by Chester Ittner Bliss in 1934 and he called this \"probit\" model, short for \"probability unit\". However this function is computationally expensive while lacking some of the desirable properties for multi-class classification. In 1944 Joseph Berkson used the function <code>log(p\/(1-p))<\/code> to do this mapping and called it logit, short for \"logistic unit\". The term logistic regression derived from this as well.<\/p>\n\n<p><strong>The Confusion<\/strong><\/p>\n\n<p>Unfortunately the term logits is abused in deep learning. From pure mathematical perspective logit is a <em>function<\/em> that performs above mapping. In deep learning people started calling the layer \"logits layer\" that feeds in to logit function. Then people started calling the output <em>values<\/em> of this layer \"logit\" creating the confusion with logit <em>the function<\/em>.<\/p>\n\n<p><strong>TensorFlow Code<\/strong><\/p>\n\n<p>Unfortunately TensorFlow code further adds in to confusion by names like <code>tf.nn.softmax_cross_entropy_with_logits<\/code>. What does logits mean here? It just means the input of the function is supposed to be the output of last neuron layer as described above. The <code>_with_logits<\/code> suffix is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6531\" rel=\"noreferrer\">redundant, confusing and pointless<\/a>. Functions should be named without regards to such very specific contexts because they are simply mathematical operations that can be performed on values derived from many other domains. In fact TensorFlow has another similar function <code>sparse_softmax_cross_entropy<\/code> where they fortunately forgot to add <code>_with_logits<\/code> suffix creating inconsistency and adding in to confusion. PyTorch on the other hand simply names its function without these kind of suffixes.<\/p>\n\n<p><strong>Reference<\/strong><\/p>\n\n<p>The <a href=\"http:\/\/www.columbia.edu\/~so33\/SusDev\/Lecture_9.pdf\" rel=\"noreferrer\">Logit\/Probit lecture slides<\/a> is one of the best resource to understand logit. I have also updated <a href=\"https:\/\/en.wikipedia.org\/wiki\/Logit\" rel=\"noreferrer\">Wikipedia article<\/a> with some of above information.<\/p>\n"},{"AnswerId":"47852892","CreationDate":"2017-12-17T06:54:04.960","ParentId":null,"OwnerUserId":"1499598","Title":null,"Body":"<p><strong>Logit<\/strong> is a function that maps probabilities <code>[0, 1]<\/code> to <code>[-inf, +inf]<\/code>. <\/p>\n\n<p><strong>Softmax<\/strong> is a function that maps <code>[-inf, +inf]<\/code> to <code>[0, 1]<\/code> similar as Sigmoid. But Softmax also normalizes the sum of the values(output vector) to be 1.<\/p>\n\n<p><strong>Tensorflow \"with logit\"<\/strong>: It means that you are applying a softmax function to logit numbers to normalize it. The input_vector\/logit is not normalized and can scale from [-inf, inf]. <\/p>\n\n<p>This normalization is used for multiclass classification problems. And for multilabel classification problems sigmoid normalization is used i.e. <code>tf.nn.sigmoid_cross_entropy_with_logits<\/code><\/p>\n"},{"AnswerId":"50511692","CreationDate":"2018-05-24T14:19:06.397","ParentId":null,"OwnerUserId":"4334743","Title":null,"Body":"<p>Just adding this clarification so that anyone who scrolls down this much can at least gets it right, since there are so many wrong answers upvoted.<\/p>\n\n<p>Diansheng's <a href=\"https:\/\/stackoverflow.com\/a\/47010867\/4334743\">answer<\/a> and JakeJ's <a href=\"https:\/\/stackoverflow.com\/a\/47278567\/4334743\">answer<\/a> get it right.<br>\n<a href=\"https:\/\/stackoverflow.com\/a\/52111173\/4334743\">A new answer<\/a> posted by Shital Shah is an even better and more complete answer.<\/p>\n\n<hr>\n\n<p>Yes, <code>logit<\/code>  as a mathematical <a href=\"https:\/\/en.wikipedia.org\/wiki\/Logit\" rel=\"noreferrer\">function<\/a> in statistics, <strong>but the <code>logit<\/code> used in context of neural networks is different.<\/strong> Statistical <code>logit<\/code> doesn't even make any sense here.<\/p>\n\n<hr>\n\n<p>I couldn't find a formal definition anywhere, but <code>logit<\/code> basically means:<\/p>\n\n<blockquote>\n  <p>The raw predictions which come out of the last layer of the neural network.<br>\n  1. This is the very tensor on which you apply the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Arg_max\" rel=\"noreferrer\"><code>argmax<\/code><\/a> function to get the predicted class.<br>\n  2. This is the very tensor which you feed into the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Softmax_function\" rel=\"noreferrer\"><code>softmax<\/code><\/a> function to get the probabilities for the predicted classes.<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Also, from a <a href=\"https:\/\/www.tensorflow.org\/tutorials\/layers\" rel=\"noreferrer\">tutorial<\/a> on official tensorflow website:<\/p>\n\n<blockquote>\n  <h3>Logits Layer<\/h3>\n  \n  <p>The final layer in our neural network is the logits layer, which will return the raw values for our predictions. We create a dense layer with 10 neurons (one for each target class 0\u20139), with linear activation (the default):<\/p>\n\n<pre><code>logits = tf.layers.dense(inputs=dropout, units=10)\n<\/code><\/pre>\n<\/blockquote>\n\n<hr>\n\n<p>If you are still confused, the situation is like this:<\/p>\n\n<pre><code>raw_predictions = neural_net(input_layer)\npredicted_class_index_by_raw = argmax(raw_predictions)\nprobabilities = softmax(raw_predictions)\npredicted_class_index_by_prob = argmax(probabilities)\n<\/code><\/pre>\n\n<p>where, <code>predicted_class_index_by_raw<\/code> and <code>predicted_class_index_by_prob<\/code> will be equal.<\/p>\n\n<p><strong>Another name for <code>raw_predictions<\/code> in the above code is <code>logit<\/code><\/strong>.<\/p>\n\n<hr>\n\n<p><del>As for the <strong>why<\/strong> <code>logit<\/code>... I have no idea. Sorry.<\/del><br>\n[Edit: See <a href=\"https:\/\/stackoverflow.com\/a\/52111173\/4334743\">this answer<\/a> for  the historical motivations behind the term.]<\/p>\n\n<hr>\n\n<h2>Trivia<\/h2>\n\n<p>Although, if you want to, you can apply statistical <code>logit<\/code> to <code>probabilities<\/code> that come out of the <code>softmax<\/code> function. <\/p>\n\n<p>If the probability of a certain class is <code>p<\/code>,<br>\nThen the <strong>log-odds<\/strong> of that class is <code>L = logit(p)<\/code>.<\/p>\n\n<p>Also, the probability of that class can be recovered as <code>p = sigmoid(L)<\/code>, using the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Sigmoid_function\" rel=\"noreferrer\"><code>sigmoid<\/code><\/a> function.<\/p>\n\n<p>Not very useful to calculate log-odds though.<\/p>\n"},{"AnswerId":"47010867","CreationDate":"2017-10-30T08:34:22.970","ParentId":null,"OwnerUserId":"4849734","Title":null,"Body":"<p>Personal understanding, in TensorFlow domain, logits are the values to be used as input to softmax. I came to this understanding based on this tensorflow tutorial.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/layers\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/layers<\/a><\/p>\n\n<hr>\n\n<p>Although it is true that logit is a function in maths(especially in statistics), I don't think that's the same 'logit' you are looking at. In the book <em>Deep Learning<\/em> by Ian Goodfellow, he mentioned,<\/p>\n\n<blockquote>\n  <p>The function \u03c3<sup>\u22121<\/sup>(x) is called the logit in statistics, but this term\n  is more rarely used in machine learning. \u03c3<sup>\u22121<\/sup>(x) stands for the\n  inverse function of logistic sigmoid function.<\/p>\n<\/blockquote>\n\n<p>In TensorFlow, it is frequently seen as the name of last layer. In Chapter 10 of the book <em>Hands-on Machine Learning with Scikit-learn and TensorFLow<\/em> by Aur\u00e9lien G\u00e9ron, I came across this paragraph, which stated <code>logits<\/code> layer clearly.<\/p>\n\n<blockquote>\n  <p>note that <code>logits<\/code> is the output of the neural network <em>before<\/em> going\n  through the softmax activation function: for optimization reasons, we\n  will handle the softmax computation later.<\/p>\n<\/blockquote>\n\n<p>That is to say, although we use softmax as the activation function in the last layer in our design, for ease of computation, we take out <code>logits<\/code> separately. This is because it is more efficient to calculate <code>softmax<\/code> and <code>cross-entropy<\/code> loss together. Remember that <code>cross-entropy<\/code> is a cost function, not used in forward propagation. <\/p>\n"},{"AnswerId":"46967826","CreationDate":"2017-10-27T05:02:27.247","ParentId":null,"OwnerUserId":"8841322","Title":null,"Body":"<blockquote>\n  <p>The logit (\/\u02c8lo\u028ad\u0292\u026at\/ LOH-jit) function is the inverse of the sigmoidal \"logistic\" function or logistic transform used in mathematics, especially in statistics. When the function's variable represents a probability p, the logit function gives the log-odds, or the logarithm of the odds p\/(1 \u2212 p).<\/p>\n<\/blockquote>\n\n<p>See here: <a href=\"https:\/\/en.wikipedia.org\/wiki\/Logit\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Logit<\/a><\/p>\n"},{"AnswerId":"56789581","CreationDate":"2019-06-27T11:01:25.710","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>(<a href=\"https:\/\/en.wikipedia.org\/wiki\/Fear_of_missing_out\" rel=\"nofollow noreferrer\">FOMO<\/a>sapiens).<\/p>\n\n<p>If you check math Logit function, it converts real space from <code>[0,1]<\/code> interval to infinity <code>[-inf, inf]<\/code>. <\/p>\n\n<p>Sigmoid and softmax will do exactly the opposite thing. They will convert the <code>[-inf, inf]<\/code> real space to <code>[0, 1]<\/code> real space.<\/p>\n\n<p>This is why it has sense in machine learning we may use logit before sigmoid and softmax function (since they match).<\/p>\n\n<p>And this is why \"we may call\" <em>anything<\/em> in machine learning that goes in front sigmoid or softmax function the logit.<\/p>\n\n<p>Here is J. Hinton <a href=\"https:\/\/youtu.be\/mlaLLQofmR8?t=142\" rel=\"nofollow noreferrer\">video<\/a> using this term. <\/p>\n\n<p><sub>PS. I am not recommending you to watch the video just to check on the term.<\/sub><\/p>\n"},{"AnswerId":"43577384","CreationDate":"2017-04-23T22:51:47.463","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Logits is an overloaded term which can mean many different things:<\/p>\n\n<hr>\n\n<p><strong>In Math<\/strong>, <a href=\"https:\/\/en.wikipedia.org\/wiki\/Logit\" rel=\"noreferrer\">Logit<\/a> is a function that maps probabilities (<code>[0, 1]<\/code>) to R (<code>(-inf, inf)<\/code>)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zto5q.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zto5q.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Probability of 0.5 corresponds to a logit of 0. Negative logit correspond to probabilities less than 0.5, positive to > 0.5.<\/p>\n\n<p><strong>In ML<\/strong>, it <a href=\"https:\/\/developers.google.com\/machine-learning\/glossary\/#logits\" rel=\"noreferrer\">can be<\/a> <\/p>\n\n<blockquote>\n  <p>the vector of raw (non-normalized) predictions that a classification\n  model generates, which is ordinarily then passed to a normalization\n  function. If the model is solving a multi-class classification\n  problem, logits typically become an input to the softmax function. The\n  softmax function then generates a vector of (normalized) probabilities\n  with one value for each possible class.<\/p>\n<\/blockquote>\n\n<p><strong>Logits also<\/strong> <a href=\"https:\/\/developers.google.com\/machine-learning\/glossary\/#logits\" rel=\"noreferrer\">sometimes<\/a> refer to the element-wise inverse of the sigmoid function.<\/p>\n"},{"AnswerId":"53352592","CreationDate":"2018-11-17T15:22:07.213","ParentId":null,"OwnerUserId":"6349441","Title":null,"Body":"<p>Logits often are the values of Z function of the output layer in Tensorflow. <\/p>\n"},{"AnswerId":"52223970","CreationDate":"2018-09-07T13:50:52.530","ParentId":null,"OwnerUserId":"6141807","Title":null,"Body":"<p>Here is a concise answer for future readers. <code>Tensorflow<\/code>'s <code>logit<\/code> is defined as the output of a neuron without applying activation function:<\/p>\n\n<pre><code>logit = w*x + b,\n<\/code><\/pre>\n\n<p>x: input, w: weight, b: bias. That's it.<\/p>\n\n<hr>\n\n<p><strong>The following is irrelevant to this question.<\/strong><\/p>\n\n<p>For historical lectures, read other answers. Hats off to <code>Tensorflow<\/code>'s \"creatively\" confusing naming convention. In <code>PyTorch<\/code>, there is only one <code>CrossEntropyLoss<\/code> and it accepts un-activated outputs. Convolutions, matrix multiplications and activations are same level operations. The design is much more modular and less confusing. This is one of the reasons why I switched from <code>Tensorflow<\/code> to <code>PyTorch<\/code>. <\/p>\n"}]}
{"QuestionId":41455571,"AnswerCount":0,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-04T03:01:05.310","AcceptedAnswerId":null,"OwnerUserId":1653534.0,"Title":"different result for tendorflow relu op","Body":"<p>Sample Code as follow:<\/p>\n\n<pre><\/pre>\n\n<p>one result output by running once as follow:<\/p>\n\n<pre><\/pre>\n\n<p>Question: I know this nn.relu equals max(features, 0), which is proved out by variable a; but it's false for variable a1: <\/p>\n\n<pre><\/pre>\n\n<p>Is there something wrong? Any help is appreciated.<\/p>\n","answers":[]}
{"QuestionId":41455575,"AnswerCount":3,"Tags":"<python><tensorflow><restore>","CreationDate":"2017-01-04T03:01:33.860","AcceptedAnswerId":41458470.0,"OwnerUserId":4763311.0,"Title":"Load \/ restore models into tensorflow at specific iteration or checkpoint","Body":"<p>I have a model , which I am saving at every 10 iterations . So , i am having following files in my saved directory . <\/p>\n\n<pre><\/pre>\n\n<p>and so on up to 100 . I have to load only the model-50. Because I have got \nNaN values after 70 iterations. By deafault, when i am restoring the saver will look for the final checkpoint. So, how could I specifically load the model-50. please help, otherwise, i have to run the model gain from scratch, which is time consuming. <\/p>\n","answers":[{"AnswerId":"52521087","CreationDate":"2018-09-26T15:15:24.183","ParentId":null,"OwnerUserId":"2182622","Title":null,"Body":"<p>I'm not sure if things were different in the past, but at least as of now, you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/get_checkpoint_state\" rel=\"nofollow noreferrer\">tf.train.get_checkpoint_state()<\/a> to get <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/checkpoint_state.proto\" rel=\"nofollow noreferrer\">CheckpointState<\/a> proto which contains all_model_checkpoint_paths.<\/p>\n"},{"AnswerId":"53015473","CreationDate":"2018-10-26T19:52:23.443","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>When you execute the command shown in most of the tutorials about saving\/restoring a model <code>saver.restore(sess, tf.train.latest_checkpoint(_dir_models))<\/code> the second parameter which you are passing is just a string to model path. This is defined in a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver#restore\" rel=\"nofollow noreferrer\">saver.restore<\/a> documentation.<\/p>\n\n<blockquote>\n  <p>save_path: Path where parameters were previously saved.<\/p>\n<\/blockquote>\n\n<p>So you can path any string there and latest_checkpoint is just a convenient function to extract this path from a <code>checkpoint<\/code> file. Open this file in a notebook and you will see all the model paths available and what is the latest.<\/p>\n\n<p>You can substitute that path with any path you want. You can get it from that file (either opening it manually or using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/get_checkpoint_state\" rel=\"nofollow noreferrer\">get_checkpoin_state<\/a> which will programmatically do it for you.<\/p>\n"},{"AnswerId":"41458470","CreationDate":"2017-01-04T07:32:58.157","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Since you are using <code>tf.train.Saver<\/code>'s function <code>restore()<\/code>, you can make use of the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.last_checkpoints\" rel=\"noreferrer\"><code>last_checkpoints<\/code><\/a> functions to get a list of all available checkpoints. You will see both <code>model-50<\/code> and <code>model-60<\/code> in this list.<\/p>\n\n<p>Pick the correct model, and pass it directly to <code>restore()<\/code> like this,<\/p>\n\n<pre><code>saver.restore(sess, ckpt_path)\n<\/code><\/pre>\n"}]}
{"QuestionId":41456454,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-04T04:58:37.423","AcceptedAnswerId":null,"OwnerUserId":4943253.0,"Title":"What are \"model variables\" in Tensorflow and slim?","Body":"<p>From <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/slim\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/slim<\/a>,<\/p>\n\n<blockquote>\n  <p>TF-Slim further differentiates variables by defining model variables,\n  which are variables that represent parameters of a model. Model\n  variables are trained or fine-tuned during learning and are loaded\n  from a checkpoint during evaluation or inference. Examples include the\n  variables created by a slim.fully_connected or slim.conv2d layer.\n  Non-model variables are all other variables that are used during\n  learning or evaluation but are not required for actually performing\n  inference. For example, the global_step is a variable using during\n  learning and evaluation but it is not actually part of the model.\n  Similarly, moving average variables might mirror model variables, but\n  the moving averages are not themselves model variables.<\/p>\n<\/blockquote>\n\n<p>From <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops\/variable_helper_functions\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops\/variable_helper_functions<\/a>,<\/p>\n\n<blockquote>\n  <p>tf.model_variables()<\/p>\n  \n  <p>Returns all variables in the MODEL_VARIABLES collection.<\/p>\n<\/blockquote>\n\n<p>Yet slim creates \"moving mean\" variables as part of its batch norm layers that are included in the MODEL_VARIABLES collection.<\/p>\n\n<p>I can see at least possible definitions of \"model variable\":<\/p>\n\n<ol>\n<li>Used in inference.<\/li>\n<li>Fine tuned during training (whether by an optimizer or some other means such as moving averaging).<\/li>\n<li>Stored in checkpoints<\/li>\n<\/ol>\n\n<p>Is it the case that Tensorflow's \"model variables\" are defined by condition 2, while slim's \"model variables\" are defined by condition 1?<\/p>\n","answers":[{"AnswerId":"43548485","CreationDate":"2017-04-21T17:12:08.847","ParentId":null,"OwnerUserId":"390292","Title":null,"Body":"<p>Simply put, slim uses contrib layers. contrib layers use layer_variable_getter, which is actually _model_variable_getter, to generate \"model_variable\" that are variables added into [tf.GraphKeys.GLOBAL_VARIABLES, tf.GraphKeys.MODEL_VARIABLES] collections.<\/p>\n\n<p>Therefore model variables are common variables plus being added to MODEL_VARIABLES collection. contrib layers' variables are model variables plus being renamed from 'bias' to 'biases' and from 'kernel' to 'weights' if needed.<\/p>\n"}]}
{"QuestionId":41458113,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network><machine-translation>","CreationDate":"2017-01-04T07:10:07.503","AcceptedAnswerId":null,"OwnerUserId":6334174.0,"Title":"Tensorflow seq2seq `feed_previous' argument`","Body":"<p>I'm using the built-in  function and I have some problems with the  argument, during training, the groundtruth is fed into the decoder while during testing we feed the output of the last time step to the decoder. The problem is that once I set the  argument I can't change that argument. I want to test my model every epoch, what should I do?<\/p>\n","answers":[{"AnswerId":"41559515","CreationDate":"2017-01-10T00:52:13.513","ParentId":null,"OwnerUserId":"2082009","Title":null,"Body":"<p>From the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4433079e7f317724eaa92ec120c6b1c3c0c52f2f\/tensorflow\/contrib\/legacy_seq2seq\/python\/ops\/seq2seq.py#L822\" rel=\"nofollow noreferrer\">documentation<\/a>, you can provide a boolean tensor for feed_previous.<\/p>\n\n<pre><code>feed_previous = tf.placeholder(tf.bool)\nmodel = tf.nn.seq2seq.embedding_attention_seq2seq(..feed_previous=feed_previous...)\nsess.run(loss, feed_dict={feed_previous=is_training, ...})\n<\/code><\/pre>\n"}]}
{"QuestionId":41458664,"AnswerCount":1,"Tags":"<computer-vision><neural-network><deep-learning><caffe><image-recognition>","CreationDate":"2017-01-04T07:45:54.127","AcceptedAnswerId":null,"OwnerUserId":919174.0,"Title":"Deep learning\/Image recognition- How to compose dataset for one category recognition","Body":"<p>Im using Intel Deep learning SDK (Based on Caffe).<br>\nI want to compose photos dataset for the ability of recognizing coins.<br>\nThe SDK is expecting dataset in the following format:  <\/p>\n\n<pre><\/pre>\n\n<p>And it required to have minimum 2 categories.<br>\nAny idea?<\/p>\n","answers":[{"AnswerId":"41459094","CreationDate":"2017-01-04T08:14:09.727","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>In your case you should have a \"coins\" category, and \"non coins\" category.<br>\nYour net should see images of \"non coins\" in order to distinguish these from \"coin\" images.<\/p>\n\n<p>You might consider training in stages (aka <em>\"hard negative mining\"<\/em>):<\/p>\n\n<pre><code>(1) model_0 &lt;- training with cons\/non-coins images\n(2) for k = 1..K iterations do:\n(2.1)   use model_{k-1} to classify additional training images\n(2.2)   add images wrongly classified by model_{k-1} to training set\n(2.3)   model_k &lt;- training with extended training set\n(3) output model_K\n<\/code><\/pre>\n"}]}
{"QuestionId":41458859,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2017-01-04T07:58:47.490","AcceptedAnswerId":41717938.0,"OwnerUserId":4013781.0,"Title":"Keras Custom Metric for single class accuracy","Body":"<p>I am building a custom metric to measure the accuracy of one class in my multi-class dataset during training. I am having trouble selecting the class. <\/p>\n\n<p>The targets are one hot (e.g: the class 0 label is [1 0 0 0 0] :<\/p>\n\n<pre><\/pre>\n\n<p>The trouble is, we have to use Keras functions to index tensors. How do you create a boolean mask for a tensor? Thank you. <\/p>\n","answers":[{"AnswerId":"41717938","CreationDate":"2017-01-18T11:13:42.143","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Note that when talking about the <em>accuracy<\/em> of one class one may refer to either of the following (not equivalent) two amounts:<\/p>\n\n<ul>\n<li>The <strong>recall<\/strong>, which, for class <em>C<\/em>, is the ratio of examples labelled with class <em>C<\/em> that are predicted to have class <em>C<\/em>.<\/li>\n<li>The <strong>precision<\/strong>, which, for class <em>C<\/em>, is the ratio of examples predicted to be of class <em>C<\/em> that are in fact labelled with class <em>C<\/em>.<\/li>\n<\/ul>\n\n<p>Instead of doing complex indexing, you can just rely on masking for you computation. Assuming we are talking about precision here (changing to recall would be trivial).<\/p>\n\n<pre><code>from keras import backend as K\n\nINTERESTING_CLASS_ID = 0  # Choose the class of interest\n\ndef single_class_accuracy(y_true, y_pred):\n    class_id_true = K.argmax(y_true, axis=-1)\n    class_id_preds = K.argmax(y_pred, axis=-1)\n    # Replace class_id_preds with class_id_true for recall here\n    accuracy_mask = K.cast(K.equal(class_id_preds, INTERESTING_CLASS_ID), 'int32')\n    class_acc_tensor = K.cast(K.equal(class_id_true, class_id_preds), 'int32') * accuracy_mask\n    class_acc = K.sum(class_acc_tensor) \/ K.maximum(K.sum(accuracy_mask), 1)\n    return class_acc\n<\/code><\/pre>\n\n<p>If you want to be more flexible, you can also have the class of interest parametrised:<\/p>\n\n<pre><code>from keras import backend as K\n\ndef single_class_accuracy(interesting_class_id):\n    def fn(y_true, y_pred):\n        class_id_true = K.argmax(y_true, axis=-1)\n        class_id_preds = K.argmax(y_pred, axis=-1)\n        # Replace class_id_preds with class_id_true for recall here\n        accuracy_mask = K.cast(K.equal(class_id_preds, interesting_class_id), 'int32')\n        class_acc_tensor = K.cast(K.equal(class_id_true, class_id_preds), 'int32') * accuracy_mask\n        class_acc = K.sum(class_acc_tensor) \/ K.maximum(K.sum(accuracy_mask), 1)\n        return class_acc\n    return fn\n<\/code><\/pre>\n\n<p>And the use it as:<\/p>\n\n<pre><code>model.compile(..., metrics=[single_class_accuracy(INTERESTING_CLASS_ID)])\n<\/code><\/pre>\n"}]}
{"QuestionId":41459886,"AnswerCount":0,"Tags":"<python><r><shiny><tensorflow>","CreationDate":"2017-01-04T09:04:37.550","AcceptedAnswerId":null,"OwnerUserId":4653285.0,"Title":"Warning: Error in if: missing value where TRUE\/FALSE needed RStudio(shiny)","Body":"<p>This code always gives me the error:<\/p>\n\n<blockquote>\n  <p>Warning: running command 'C:\/Users\/Madhurima\/Anaconda3\/python\n  C:\/Users\/Madhurima\/Anaconda3\/envs\/tensorflow\/Lib\/site-packages\/tensorflow\/models\/tutorials\/image\/imagenet\/classify_image.py\/classify_image.py\n  --image_file \/tmp\/imagenet\/cropped_panda.jpg' had status 2 Warning in renderPlot(...) : NAs introduced by coercion Warning: Error in if:\n  missing value where TRUE\/FALSE needed<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41459900,"AnswerCount":3,"Tags":"<python><image-processing><tensorflow><jpeg><classification>","CreationDate":"2017-01-04T09:05:38.827","AcceptedAnswerId":41470165.0,"OwnerUserId":1826383.0,"Title":"Tensorflow Label_Image for PNG, GIF etc","Body":"<p>I was playing around with Tensorflow for image classification. I used the image_retraining\/retrain.py to retrain the inception library with new categories and used it to classify images using label_image.py from <a href=\"https:\/\/github.com\/llSourcell\/tensorflow_image_classifier\/blob\/master\/src\/label_image.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/llSourcell\/tensorflow_image_classifier\/blob\/master\/src\/label_image.py<\/a> as below:<\/p>\n\n<pre><\/pre>\n\n<p>I noticed two issues. When I retrain  with new categories, it only trains JPG images. I am a noob in machine learning so not sure whether this is a limitation or is it possible to train other extension images like PNG, GIF?<\/p>\n\n<p>Another one is when classifying the images the input is again only for JPG. I tried to change DecodeJpeg to DecodePng in label_image.py above but couldn't work. Another way I tried was to convert other formats into JPG before passing them in for classification like:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any other way to do this? Does Tensorflow have functions to handle other image formats other than JPG?<\/p>\n\n<p>I tried the following by feeding in parsed image as compared to JPEG as suggested by @mrry<\/p>\n\n<pre><\/pre>\n\n<p>It works for JPEG images but when I use PNG or GIF it throws <\/p>\n\n<pre><\/pre>\n\n<p>Thanks and Regards<\/p>\n","answers":[{"AnswerId":"41464412","CreationDate":"2017-01-04T12:57:01.970","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>You should have a look at the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/image\/\" rel=\"nofollow noreferrer\"><code>tf.image<\/code><\/a> package. It's got good functions to decode \/ encode JPEGs, GIFs and PNGs.<\/p>\n"},{"AnswerId":"41498374","CreationDate":"2017-01-06T03:17:58.783","ParentId":null,"OwnerUserId":"1826383","Title":null,"Body":"<p>Following @mrry's suggestion to feed in parsed image, converted the image data into array and convert into RGB as stated below in the code. Now I am able to feed in JPG,PNG and GIF.<\/p>\n\n<pre><code>import tensorflow as tf\nimport sys\nimport numpy as np\nfrom PIL import Image\n\n# change this as you see fit\nimage_path = sys.argv[1]\n\n# Read in the image_data\nimage_data = tf.gfile.FastGFile(image_path, 'rb').read()\n\nimage = Image.open(image_path)\nimage_array = image.convert('RGB')\n\n# Loads label file, strips off carriage return\nlabel_lines = [line.rstrip() for line\n    in tf.gfile.GFile(\"\/root\/tf_files\/output_labels.txt\")]\n\n# Unpersists graph from file\nwith tf.gfile.FastGFile(\"\/root\/tf_files\/output_graph.pb\", 'rb') as f:\n    graph_def = tf.GraphDef()\n    graph_def.ParseFromString(f.read())\n    _ = tf.import_graph_def(graph_def, name='')\n\nwith tf.Session() as sess:\n# Feed the image_data as input to the graph and get first prediction\n    softmax_tensor = sess.graph.get_tensor_by_name('final_result:0')\n\n        predictions = sess.run(softmax_tensor,{'DecodeJpeg:0': image_array})\n\n# Sort to show labels of first prediction in order of confidence\n    top_k = predictions[0].argsort()[-len(predictions[0]):][::-1]\n\n    for node_id in top_k:\n        human_string = label_lines[node_id]\n        score = predictions[0][node_id]\n        print('%s (score = %.5f)' % (human_string, score))\n<\/code><\/pre>\n"},{"AnswerId":"41470165","CreationDate":"2017-01-04T17:47:53.597","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The model can only train on (and evaluate) JPEG images, because the <code>GraphDef<\/code> that you've saved in <code>\/root\/tf_files\/output_graph.pb<\/code> only contains a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/image\/encoding_and_decoding#decode_jpeg\" rel=\"noreferrer\"><code>tf.image.decode_jpeg()<\/code><\/a> op, and uses the output of that op for making predictions. There are at least a couple of options for using other image formats:<\/p>\n\n<ol>\n<li><p><strong>Feed in parsed images rather than JPEG data.<\/strong> In the current program, you feed in a JPEG-encoded image as a string value for the tensor <code>\"DecodeJpeg\/contents:0\"<\/code>. Instead, you can feed in a 3-D array of decoded image data for the tensor <code>\"DecodeJpeg:0\"<\/code> (which represents the <em>output<\/em> of the <code>tf.image.decode_jpeg()<\/code> op), and you can use NumPy, PIL, or some other Python library to create this array.<\/p><\/li>\n<li><p><strong>Remap the image input in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/utility_functions#import_graph_def\" rel=\"noreferrer\"><code>tf.import_graph_def()<\/code><\/a>.<\/strong> The <code>tf.import_graph_def()<\/code> function enables you to connect two different graphs together by remapping individual tensor values. For example, you could do something like the following to add a new image-processing op to the existing graph:<\/p>\n\n<pre><code>image_string_input = tf.placeholder(tf.string)\nimage_decoded = tf.image.decode_png(image_string_input)\n\n# Unpersists graph from file\nwith tf.gfile.FastGFile(\"\/root\/tf_files\/output_graph.pb\", 'rb') as f:\n    graph_def = tf.GraphDef()\n    graph_def.ParseFromString(f.read())\n    softmax_tensor, = tf.import_graph_def(\n        graph_def,\n        input_map={\"DecodeJpeg:0\": image_decoded},\n        return_operations=[\"final_result:0\"])\n\nwith tf.Session() as sess:\n    # Feed the image_data as input to the graph and get first prediction\n    predictions = sess.run(softmax_tensor, {image_string_input: image_data})\n    # ...\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":41460248,"AnswerCount":0,"Tags":"<tensorflow><gpu><cpu>","CreationDate":"2017-01-04T09:25:25.147","AcceptedAnswerId":null,"OwnerUserId":6064301.0,"Title":"In tensorflow, when should I use CPU, and when should I use GPU?","Body":"<p>I only use CPU in this scenario:<\/p>\n\n<p>when I write a emb function, I added <strong>with tf.device('\/cpu:0')<\/strong><\/p>\n\n<p>Can anyone tell me when should I use CPU, and when should I use GPU?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41460783,"AnswerCount":0,"Tags":"<machine-learning><deep-learning><theano><convolution><deconvolution>","CreationDate":"2017-01-04T09:51:04.507","AcceptedAnswerId":null,"OwnerUserId":3109123.0,"Title":"Transposed convolution on feature maps using Theano","Body":"<p>I asked similar question on <strong><a href=\"https:\/\/stats.stackexchange.com\/questions\/253724\/would-you-interpret-this-image-as-a-correct-deconvolution-process\">CrossValidation<\/a><\/strong> for the image interpretation. I'm moving my detailed question here to include some code details. <\/p>\n\n<p>The results I'm having are not fully desirable So maybe you have faced this issue before and you can help me find it out. <\/p>\n\n<p>It is fully convolution neural network \"no fully connected part\". <\/p>\n\n<p><strong>Training part<\/strong><\/p>\n\n<p>first the images are transposed to match the convolution function. (batch_no,img_channels,width,height)<\/p>\n\n<pre><\/pre>\n\n<p>Learning optimized using learning rate:3e-6, Hu_uniform initialization and nestrove for 500 epochs until this convergence. <\/p>\n\n<ul>\n<li>Training cost: 1.602449<\/li>\n<li>Training loss: 4.610442<\/li>\n<li>validation error: 5.126761<\/li>\n<li>Test loss: 5.885714    <\/li>\n<\/ul>\n\n<p><strong>Backward part<\/strong> <\/p>\n\n<blockquote>\n  <p>Loading Image<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>Reshape to one batch<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>Run the model to extract first feature map after activation using Relu<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>Apply backwork model to reconstruct the image using the only activated\n  neurons<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>Here we have two images results, the first is the transposed image without activation and the second with relu since kernels might have some negative weights. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/neHhm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/neHhm.png\" alt=\"without Relu\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/ExgqI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ExgqI.png\" alt=\"With Relu\"><\/a><\/p>\n\n<p>It is clear from the transposed convolution image that this kernel is learn to detect some useful feature related to this image. But the reconstructing part is breaking the image color scheme during the transpose convolution. It might be because the pixels values are small float numbers. Do you see where is the problem here ? <\/p>\n","answers":[]}
{"QuestionId":41460802,"AnswerCount":2,"Tags":"<python-2.7><tensorflow><google-cloud-platform>","CreationDate":"2017-01-04T09:51:56.730","AcceptedAnswerId":null,"OwnerUserId":7090211.0,"Title":"Accessing data in Google Cloud bucket for a python Tensorflow learning program","Body":"<p>I\u2019m working through the Google quick start examples for Cloud Learning \/ Tensorflow as shown here: <a href=\"https:\/\/cloud.google.com\/ml\/docs\/quickstarts\/training\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/ml\/docs\/quickstarts\/training<\/a><\/p>\n\n<p>I want my python program to access data that I have stored in a Google Cloud bucket such as gs:\/\/mybucket. How do I do this inside of my python program instead of calling it from the command line?<\/p>\n\n<p>Specifically, the quickstart example for cloud learning utilizes data they provided but what if I want to provide my own data that I have stored in a bucket such as gs:\/\/mybucket?<\/p>\n\n<p>I noticed a similar post here: <a href=\"https:\/\/stackoverflow.com\/questions\/40198449\/how-can-i-get-the-cloud-ml-service-account-programmatically-in-python\">How can I get the Cloud ML service account programmatically in Python?<\/a> ... but I can\u2019t seem to install the googleapiclient module.<\/p>\n\n<p>Some posts seem to mention Apache Beam though I can\u2019t tell if that\u2019s relevant to me, but besides I can\u2019t figure out how to download or install that whatever it is.<\/p>\n","answers":[{"AnswerId":"51020582","CreationDate":"2018-06-25T09:49:11.093","ParentId":null,"OwnerUserId":"9979257","Title":null,"Body":"<p>Assuming you are using Ubuntu\/Linux as an OS and already having data in GCS bucket\nExecute following commands from a terminal or can be executed on Jupyter Notebook(just use <code>!<\/code> before commands):<\/p>\n\n<p>--------------------- Installation -----------------<\/p>\n\n<p>1st install storage module: \non Terminal type: <\/p>\n\n<pre><code>pip install google-cloud-storage\n<\/code><\/pre>\n\n<p>2nd to verify storage installed or not type the command:  <\/p>\n\n<pre><code>gsutil \n<\/code><\/pre>\n\n<p>(o\/p will show available options)<\/p>\n\n<p>---------------------- Copy data from GCS bucket --------<\/p>\n\n<p>type this command: to check whether you are able to get information about bucket<\/p>\n\n<pre><code>gsutil acl get gs:\/\/BucketName\n<\/code><\/pre>\n\n<p>Now copy the file from GCS Bucket to your machine:<\/p>\n\n<pre><code>gsutil cp gs:\/\/BucketName\/FileName \/PathToDestinationDir\/\n<\/code><\/pre>\n\n<p>In this way, you will be able to copy data from this bucket to your machine for further processing purpose.<\/p>\n\n<p>NOTE: all the above commands can be run from a Jupyter Notebook just use <code>!<\/code> before commands, it will run e.g.<\/p>\n\n<pre><code>!gsutil cp gs:\/\/BucketName\/FileName \/PathToDestinationDir\/\n<\/code><\/pre>\n"},{"AnswerId":"41463646","CreationDate":"2017-01-04T12:17:37.460","ParentId":null,"OwnerUserId":"2877278","Title":null,"Body":"<p>If I understand your question correctly, you want to programmatically talk to GCS in Python. <\/p>\n\n<p>The <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/google-cloud-python\/tree\/master\/storage\" rel=\"nofollow noreferrer\">official docs<\/a> are a good place to start.<\/p>\n\n<p>First, grab the module using pip:<\/p>\n\n<pre><code>pip install --upgrade google-cloud-storage\n<\/code><\/pre>\n\n<p>Then:<\/p>\n\n<pre><code>from google.cloud import storage\nclient = storage.Client()\nbucket = client.get_bucket('bucket-id-here')\n# Then do other things...\nblob = bucket.get_blob('remote\/path\/to\/file.txt')\nprint(blob.download_as_string())\nblob.upload_from_string('New contents!')\nblob2 = bucket.blob('remote\/path\/storage.txt')\nblob2.upload_from_filename(filename='\/local\/path.txt')\n<\/code><\/pre>\n"}]}
{"QuestionId":41464741,"AnswerCount":1,"Tags":"<tensorflow><nlp><word2vec><tensorflow-serving>","CreationDate":"2017-01-04T13:15:53.660","AcceptedAnswerId":null,"OwnerUserId":2622774.0,"Title":"Prediction using NLP\/ML\/word2vec\/tensorflow","Body":"<p>Prediction using NLP\/ML\/word2vec\/tensorflow.\nI have 1000's of records from service request utility. \nI have list of sevice requests &amp; corresponding action taken by executive for each service request which is in natural english langauge(text could be different for same type of request from different user).<\/p>\n\n<p>Sample record-<\/p>\n\n<p>request: addition to xxx mail alias for the user id yyy<\/p>\n\n<p>action: ask user to get approval from manager and then manually add the user id to mail alias group.<\/p>\n\n<p>My understanding is that this data can be used as training data for my model(NLP\/ML\/word2vec), in such a way that if any new request of same type given as input to model, model should be able to suggest the required actions that executive should take. This is what I want to achieve but not sure how to and from where to start. I have executed samples from <\/p>\n\n<p><a href=\"https:\/\/github.com\/tmikolov\/word2vec\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tmikolov\/word2vec<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r0.12\/tensorflow\/models\/embedding\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r0.12\/tensorflow\/models\/embedding<\/a><\/p>\n\n<p>Any pointers.<\/p>\n","answers":[{"AnswerId":"41472844","CreationDate":"2017-01-04T20:39:54.477","ParentId":null,"OwnerUserId":"3318517","Title":null,"Body":"<p>I think you can approach this as a translation problem. Just instead of translating between languages you translate from request to action. The TensorFlow page has a translation tutorial you could try: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/<\/a><\/p>\n\n<p>For an even simpler go at it you could also just train a character-prediction RNN on the whole text file that contains the requests and answers. This is described well with example code at <a href=\"http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/\" rel=\"nofollow noreferrer\">http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/<\/a>.<\/p>\n\n<p>The RNN will then be able to generate or auto-complete requests as well :).<\/p>\n"}]}
{"QuestionId":41465197,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2017-01-04T13:38:04.100","AcceptedAnswerId":null,"OwnerUserId":6882624.0,"Title":"TypeError while using sparse_categorical_crossentropy in Keras","Body":"<p>When I run this code, I get the following error<\/p>\n\n<pre><\/pre>\n\n<p>The Error is:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41465282,"AnswerCount":2,"Tags":"<opencv><deep-learning><caffe><pycaffe>","CreationDate":"2017-01-04T13:42:54.383","AcceptedAnswerId":null,"OwnerUserId":7140808.0,"Title":"Error using Caffe: \"This tool requires OpenCV; compile with USE_OPENCV.\"","Body":"<p>I want to use Caffe's tool,  to compute the mean image, but the command line tells me that:<\/p>\n\n<blockquote>\n  <p>This tool requires OpenCV; compile with USE_OPENCV.<\/p>\n<\/blockquote>\n\n<p>I've searched the Internet, and one guy said:<\/p>\n\n<blockquote>\n  <p>Well, this message suggests that you have compiled caffe without\n  OpenCV, i.e; with USE_OPENCV set to 0 in your Makefile.config.\n  Recompile with that settings commented out (as is the default) and try\n  again. And of course you need to have OpenCV installed for this to\n  work.<\/p>\n<\/blockquote>\n\n<p>But I have already set  when I compiled my Caffe. <\/p>\n\n<p>Could someone help me to solve this problem?<\/p>\n","answers":[{"AnswerId":"41466435","CreationDate":"2017-01-04T14:37:45.833","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>I think you have misread the statement you found, it says you are compiling without OpenCV, which means WITH_OPENCV := 0 , to compile Caffe with OpenCV you have to set WITH_OPENCV := 1 , or just comment it out.<\/p>\n"},{"AnswerId":"42451547","CreationDate":"2017-02-25T03:11:37.027","ParentId":null,"OwnerUserId":"7614679","Title":null,"Body":"<p>I met the same error information when I used caffe to do my work. I solved this problem by modifying something in the \"Makefile.config\" file.<\/p>\n\n<p>The same as the lastest answer\uff0c I commented this line \"USE_OPENCV := 0\" with \"#\", so after compiling I can use OpenCV's function. I can use this way to deal with it, but you said that when you did this action and got another error. I think you didn't install your OpenCV correctly. Maybe it's your version problem or environment variable error. You can try to install OpneCV again.<\/p>\n"}]}
{"QuestionId":41465291,"AnswerCount":0,"Tags":"<asynchronous><parameters><tensorflow><worker>","CreationDate":"2017-01-04T13:43:24.653","AcceptedAnswerId":null,"OwnerUserId":5488129.0,"Title":"How between-graph replication with asynchronous updates in Tensorflow runs?","Body":"<p><strong>We focus on this situation: between-graph replication with asynchronous updates.<\/strong> <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/ecosystem\" rel=\"nofollow noreferrer\">The following figure<\/a> shows how it works. And I known that each worker will send gradients to every ps and receive its updated parameters from every ps.(Is it correct? Is it path 1 and path 2?) <\/p>\n\n<p><strong>But can anyone explain in detail what does the figure 1,2,3,4,5,6 mean respectively?<\/strong><\/p>\n\n<p>Thank you in advance!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9hgIp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9hgIp.png\" alt=\"between-graph replication with asynchronous updates\"><\/a><\/p>\n","answers":[]}
{"QuestionId":41465368,"AnswerCount":1,"Tags":"<lua><command-line-arguments><torch>","CreationDate":"2017-01-04T13:47:21.450","AcceptedAnswerId":41473055.0,"OwnerUserId":1237313.0,"Title":"Torch: CmdLine() : passing array of elements to a variable from command line","Body":"<p>I am using torch.CmdLine() to parse command line arguments in Torch. I need to pass an array of elements for one of the variables. I am not aware of any mechanisms for passing an array of elements for a variable. So I am treating that variable as a string and passing the array of elements separated by space and enclosed within double quotes from the command line.\nThe code for doing this looks as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Here since Lua string indexing is not provided by default, I had to override __index to enable indexing of string as follows,<\/p>\n\n<pre><\/pre>\n\n<p>This works for parsing the string to an array of numbers.<\/p>\n\n<p>However, overriding __index breaks the code somewhere else, throwing the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I can do some workarounds to fix this (instead of overriding __index use string.sub(str,i,i) directly)  but I would like to know your suggestions in passing an array of elements using torch.CmdLine() in an elegant way--if applicable.<\/p>\n\n<p>Thanks in Advance.<\/p>\n","answers":[{"AnswerId":"41473055","CreationDate":"2017-01-04T20:53:49.840","ParentId":null,"OwnerUserId":"7066323","Title":null,"Body":"<p>You can pass the list as a sequence of words separated by spaces as you have done, and after cmd:parse  you can explode the string into an array using:<\/p>\n\n<pre><code>params = cmd:parse()\nlocal tmptab = {}\nfor word in params.cuda_device:gmatch(\"%w+\") do\n  tmptab[#tmptab +1] = word\nend\nparams.cuda_device = tmptab\nfor i,v in pairs(params.cuda_device) do\n   -- some code here\nend\n<\/code><\/pre>\n\n<p>That will explode the string parsed by cmd:parse() into an table with each word on it's on index, and there would be no need to delve into string metamethods...<\/p>\n\n<p>Also, you can avoid the error doing this:<\/p>\n\n<pre><code>getmetatable('').__index = function(str,i)\n  if(type(i) == \"number\") then\n    return string.sub(str,i,i)\n  else\n    return \"\"\n  end \nend\n<\/code><\/pre>\n\n<p>It is odd that you have tried to index a string using another string, however...<\/p>\n"}]}
{"QuestionId":41465681,"AnswerCount":1,"Tags":"<tensorflow><tflearn>","CreationDate":"2017-01-04T14:01:56.523","AcceptedAnswerId":41562382.0,"OwnerUserId":173355.0,"Title":"Use V2 checkpoint for TFLearn (TensorFlow) r0.12.1","Body":"<p>Is there a way to tell TFLearn to save checkpoints in the V2 format?  I am using the current (r0.12.1) release of TensorFlow.  If you follow a simple example, such as:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/tflearn\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/tflearn\/<\/a><\/p>\n\n<p>You will get flooded with:<\/p>\n\n<pre><\/pre>\n\n<p>While the warnings are easy enough to ignore.  I would like to save these files into the correct version so that I do not have to upgrade them later on.<\/p>\n","answers":[{"AnswerId":"41562382","CreationDate":"2017-01-10T06:24:20.717","ParentId":null,"OwnerUserId":"226509","Title":null,"Body":"<p>The issue is that some of the higher level API calls are using the older checkpoint format.  At this point, in r.0.12.1, it is best to suppress or ignore the warnings.<\/p>\n\n<p>This might be fixed in the new version <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r1.0\/<\/a><\/p>\n"}]}
{"QuestionId":41466789,"AnswerCount":2,"Tags":"<python><tensorflow><normalization>","CreationDate":"2017-01-04T14:55:14.787","AcceptedAnswerId":41467218.0,"OwnerUserId":1934212.0,"Title":"Correct dim parameter in tf.nn.l2_normalize(x, dim, epsilon=1e-12, name=None) for RGB picture","Body":"<p>I want to use tensorflows<\/p>\n\n<pre><\/pre>\n\n<p>method to normalize each channel seperately in a (32x32x3) RGB image.<\/p>\n\n<p>How do I have to set the <\/p>\n\n<pre><\/pre>\n\n<p>parameter to achieve that? <\/p>\n","answers":[{"AnswerId":"49021007","CreationDate":"2018-02-28T02:20:45.163","ParentId":null,"OwnerUserId":"7461983","Title":null,"Body":"<p>You should set <code>dim = [0, 1]<\/code>.\nSo each channel <code>[32, 32]<\/code> would be normalized as <code>[32*32]<\/code>, a 1-D tensor.<\/p>\n"},{"AnswerId":"41467218","CreationDate":"2017-01-04T15:18:08.470","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Quoting the official documentation, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/normalization#l2_normalize\" rel=\"nofollow noreferrer\"><code>tf.nn.l2_normalize<\/code><\/a>,<\/p>\n\n<blockquote>\n  <p>For x with more dimensions, independently normalizes each 1-D slice along dimension dim.<\/p>\n<\/blockquote>\n\n<p>Since your channel is the third dimension, you can pass in <code>dim=2<\/code> (since dimensions start from 0).<\/p>\n"}]}
{"QuestionId":41467115,"AnswerCount":1,"Tags":"<python><tensorflow><gensim><word2vec>","CreationDate":"2017-01-04T15:13:46.557","AcceptedAnswerId":null,"OwnerUserId":3009947.0,"Title":"Using word2vec pretrained vectors, how to generate ids of a sentence as input to tf.nn.embedding_lookup function in tensorflow?","Body":"<p>To extract the embedding representations of input data, the tensorflow documentation says we can use the following:<\/p>\n\n<pre><\/pre>\n\n<p>Accdg to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/embeddings#embedding_lookup\" rel=\"nofollow noreferrer\">TF documentation<\/a>, the 2nd parameter of the function tf.nn.embedding_lookup is a tensor of ids:<\/p>\n\n<blockquote>\n  <p>ids: A Tensor with type int32 or int64 containing the ids to be looked up in params.<\/p>\n<\/blockquote>\n\n<p>My question is: Given a sentence, say, <\/p>\n\n<blockquote>\n  <p>\"Welcome to the world\"<\/p>\n<\/blockquote>\n\n<p>how can I represent and transform it into ? In the code below, how can I transform my sentence into . <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41469347","CreationDate":"2017-01-04T16:58:49.210","ParentId":null,"OwnerUserId":"3009947","Title":null,"Body":"<p>I was looking for a code something like this. I wasn't aware of the class attributes of <code>w<\/code> - couldn't find them in the gensim website.<\/p>\n\n<pre><code>s = \"hello, how are you?\"\ntokens = tokenize(s)\/\/function that returns a list of the tokens in a sentence\n\nids = []\nfor key in tokens:\n  try:\n    ids.append(w.vocab[key].index)\n  except:\n    ids.append(w.vocab['UNK'].index)\n<\/code><\/pre>\n"}]}
{"QuestionId":41469128,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><conv-neural-network>","CreationDate":"2017-01-04T16:48:19.443","AcceptedAnswerId":null,"OwnerUserId":3827807.0,"Title":"GAN with not a random input","Body":"<p>I'm very interested in GAN those times.<\/p>\n\n<p>I coded one for MNIST with the following structure :\nGenerator model\nDiscriminator model\nGen + Dis model<\/p>\n\n<p>Generator model generate batches of image from random distribution.\nDiscrimator is trained over it and real images.\nThen Discriminator is freeze in Gen+Dis model and Generator trained. (With the frozen Discriminator who says if the generator is good or not)<\/p>\n\n<p>Now, imagine I don't want to feed my generator with a random distribution but with images. (For upscaling for example, or generate an real image from a draw)<\/p>\n\n<p>Do I need to change something in it ?\n(Except the conv model who will be more complex)\nShould I continue to use the binary_crossentropy as loss function ?<\/p>\n\n<p>Thanks you very much!<\/p>\n","answers":[{"AnswerId":"41470175","CreationDate":"2017-01-04T17:48:22.740","ParentId":null,"OwnerUserId":"3863500","Title":null,"Body":"<p>You can indeed put a variational autoencoder (VAE) in front in order to generate the initial distribution z (see <a href=\"https:\/\/arxiv.org\/abs\/1512.09300\" rel=\"nofollow noreferrer\">paper<\/a>). <\/p>\n\n<p>If you are interested in the topic I can recommend the this <a href=\"https:\/\/www.kadenze.com\/courses\/creative-applications-of-deep-learning-with-tensorflow\/info\" rel=\"nofollow noreferrer\">course<\/a> at Kadenze.<\/p>\n"}]}
{"QuestionId":41471540,"AnswerCount":1,"Tags":"<r><tensorflow>","CreationDate":"2017-01-04T19:11:44.013","AcceptedAnswerId":41471616.0,"OwnerUserId":3882526.0,"Title":"TensorFlow apply a function to each row of a matrix variable","Body":"<p>Hi I'm a newbie to Tensorflow. What I want to do is something like this in R:<\/p>\n\n<pre><\/pre>\n\n<p>Is this do-able in Tensorflow, either in Python API or R tensorflow package? Thanks!<\/p>\n\n<p>EDIT:  is actually what I want.<\/p>\n","answers":[{"AnswerId":"41471616","CreationDate":"2017-01-04T19:16:25.590","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorFlow Python API includes the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/functional_ops\/higher_order_operators#map_fn\" rel=\"noreferrer\"><code>tf.map_fn(fn, elems)<\/code><\/a> higher-order operator, which allows you to specify a (Python) function <code>fn<\/code> that will be applied to each slice of <code>elems<\/code> in the 0th dimension (i.e. to each row if <code>elems<\/code> is a matrix).<\/p>\n\n<p>Note that, while <code>tf.map_fn()<\/code> is very general, it may be more efficient to use specialized ops that either broadcast their arguments on one or more dimensions (e.g. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/arithmetic_operators#multiply\" rel=\"noreferrer\"><code>tf.multiply()<\/code><\/a>), or reduce in parallel across one or more dimensions (e.g. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/reduction#reduce_sum\" rel=\"noreferrer\"><code>tf.reduce_sum()<\/code><\/a>). However, <code>tf.map_fn()<\/code> is useful when there is no built-in operator to do what you want.<\/p>\n"}]}
{"QuestionId":41471614,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-04T19:16:19.283","AcceptedAnswerId":null,"OwnerUserId":7375883.0,"Title":"Why does tf.cast not really work after image resize?","Body":"<p>I read an image and resize it, the result is right, but after cast to uint8, it is all wrong. The output image is wrong too. What happened?<\/p>\n\n<p>The code is below:<\/p>\n\n<pre><\/pre>\n\n<p>BUT, If I did like below, it works well, but WHY!!!??<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41473261,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><neural-network>","CreationDate":"2017-01-04T21:07:22.300","AcceptedAnswerId":null,"OwnerUserId":7376251.0,"Title":"Python \/ Tensorflow \/ Numpy setting an array element with a sequence error","Body":"<p>I'm experimenting with the Generative Adversarial Network found here:\n<a href=\"https:\/\/github.com\/carpedm20\/DCGAN-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/carpedm20\/DCGAN-tensorflow<\/a>\nI've installed all the prereqs and downloaded the code and I've hit an error that I'm not sure where it's coming from. After running main.py with any image set, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The line it mentions, line 153 of model.py, is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>'sample_images' seems to be set to a tensorflow placeholder, based on lines 75  and 76 of model.py:<\/p>\n\n<pre><\/pre>\n\n<p>Some googling has led me towards using something like 'feed_dict', but a few experiments with this haven't worked. Additionally feed_dict is used later on in the code, like this at line 187:<\/p>\n\n<pre><\/pre>\n\n<p>I'm not sure if this is Python, Tensorflow, Numpy or something else causing the error so don't really know where to start. Help appreciated.<\/p>\n\n<p>For reference, I'm running Python 3.5 64-bit on Windows 7<\/p>\n\n<p>Thanks<\/p>\n\n<p>Nathan<\/p>\n\n<p>UPDATE: After far too long spent trying to fix this, I eventually spun up an Ubuntu VM, installed Python and Tensorflow and the whole thing works without a hitch. Not quite an answer to the original question, but definitely a workaround.<\/p>\n","answers":[{"AnswerId":"41473653","CreationDate":"2017-01-04T21:38:42.887","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>I don't have <code>tensorflow<\/code>, so can't help you directly.  But I can create your error with something similar<\/p>\n\n<p>Make an dtype object array with lists of various sizes:<\/p>\n\n<pre><code>In [343]: x=np.array([3, [1,2], [1,2,3]],dtype=object)\nIn [344]: x\nOut[344]: array([3, [1, 2], [1, 2, 3]], dtype=object)\nIn [345]: x.astype(np.float32)\n...\nValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>The same would apply if one or more of the elements were an array, and I imagine a <code>tensorflow<\/code> object, if iterable (as sequence), would produce the same error.<\/p>\n\n<p>So - check the <code>dtype<\/code> of <code>sample_images<\/code>, and identify the type of the elements.  Also check the <code>tensorflow<\/code> has its own methods of converting its objects to arrays of floats.  Don't assume that <code>numpy<\/code> functions and methods work directly on <code>tensorflow<\/code> objects.<\/p>\n"}]}
{"QuestionId":41473397,"AnswerCount":1,"Tags":"<python><architecture><tensorflow>","CreationDate":"2017-01-04T21:18:27.567","AcceptedAnswerId":41473746.0,"OwnerUserId":918967.0,"Title":"How to obtain information on tensorflow architecture","Body":"<p>I have been working on retraining TensorFlow Inception v3 (see <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">TensorFlow Github<\/a>) and was curious as to how I would obtain some general \"metadata\" about the model.<\/p>\n\n<p>For instance:<\/p>\n\n<ul>\n<li>How many hidden layers are there?<\/li>\n<li>Which is the \"last\" layer that gets retrained?<\/li>\n<li>How many total neurons (or neurons \/ layer)?<\/li>\n<li>How many convolutions were used?<\/li>\n<li>How many poolings were used?<\/li>\n<\/ul>\n\n<p>I basically want to be able to write a sentence or two describing the model. There is sentence on the site <\/p>\n\n<blockquote>\n  <p>...single frame evaluation with a computational cost of 5 billion multiply-adds per inference and with using less than 25 million parameters.<\/p>\n<\/blockquote>\n\n<p>but I am not quit sure what that means.  Is that 25 million neurons?  <\/p>\n\n<p>Looking at the document <a href=\"http:\/\/arxiv.org\/abs\/1512.00567\" rel=\"nofollow noreferrer\">Rethinking the Inception Architecture for Computer Vision<\/a> which was referenced on the GitHub, I think much of what I need is in Table 1?<\/p>\n\n<p>Would I write this ?  <\/p>\n\n<p>\"TensorFlow InceptionV3 is a deep convolution neural network that has 13 hidden layers and uses six convolutions, two pooling steps, and three inception modules to perform a softmax classification of images\"<\/p>\n\n<p>Of course I want the correct data in the sentence and would love to know how many neurons are in the model as well<\/p>\n\n<p>Thanks! <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Tsy2A.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Tsy2A.png\" alt=\"Table 1\"><\/a><\/p>\n","answers":[{"AnswerId":"41473746","CreationDate":"2017-01-04T21:44:40.263","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>You're very much on the right track.  Yes, parameters are neurons.  For instance, the output of the first conv layer is 32 filters (kernels) of grid size 149^2.  That's a total of 710,432 neurons\/parameters for that layer alone.<\/p>\n\n<p>The critical part of training is back-propagation which adjusts the weights between one layer and the next.  The result of that SoftMax operation is the 1000 output predictions; the last trained layer would be its connection to the prior layer.<\/p>\n\n<p>You can read the simple convolutions and poolings from the chart.  I'm not sure whether or not you are supposed to include the ones inside the inception layers.<\/p>\n\n<p>Finally, if I read this correctly, the inceptions are multiple units of each type, applied in series.  That would mean that we have 10 inception layers, not 3.<\/p>\n"}]}
{"QuestionId":41474136,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-04T22:15:53.293","AcceptedAnswerId":41477681.0,"OwnerUserId":1044011.0,"Title":"Disable SSE4.1 when compiling TensorFlow","Body":"<p>I followed the instruction on TF's website and install the TensorFlow from the source code. I did not change any configurations, all are the default values.<\/p>\n\n<p>When I run my program (which works fine when using the pre-complied TensorFlow 0.12 wheel), it gives me the following error<\/p>\n\n<p><\/p>\n\n<p>By default TensorFlow enables SSE4 support, I guess there is a way to disable it? Thanks for any inputs.<\/p>\n","answers":[{"AnswerId":"41477681","CreationDate":"2017-01-05T05:10:41.163","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/c4b09b5df79625a70853fd66b5caa7dd92fb4d1f\/tensorflow\/tensorflow.bzl#L130\" rel=\"noreferrer\">This line<\/a> in <code>tensorflow\/tensorflow.bzl<\/code> is responsible for enabling SSE 4.1 instructions in all x86 builds. If you delete that line, the resulting build should work on your machine.<\/p>\n"}]}
{"QuestionId":41474312,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2017-01-04T22:30:15.703","AcceptedAnswerId":null,"OwnerUserId":4359986.0,"Title":"Keras\/Tensorflow model returning zeros for everything","Body":"<p>I have four variables , where are the training set, and are the test set. I have the following Keras neural network<\/p>\n\n<pre><\/pre>\n\n<p>However, the following line is giving me zeros for everything:<\/p>\n\n<pre><\/pre>\n\n<p>I checked obvious things, such as  being zero. My guess is that the problem is some kind of vanishing gradient issue. Any help is appreciated; thanks!<\/p>\n","answers":[]}
{"QuestionId":41474592,"AnswerCount":0,"Tags":"<tensorflow><ocr><image-recognition><pre-trained-model>","CreationDate":"2017-01-04T22:53:18.873","AcceptedAnswerId":null,"OwnerUserId":1122983.0,"Title":"Trained models for tensorflow ocr","Body":"<p>I start the course of tensorflow in udacity, and simultaneously I am looking on the web for the topic.<\/p>\n\n<p>I suppose that the typical use cases are well solved already, in a better way that i can achieve by my own. In other words in some place exists trained models for usual cases ready to use. I found zooModels that if I undestand properly is the thing that i looking for. but I can't realize that there does not exist a ocr model published that can recognise a number in a image:<\/p>\n\n<p><a href=\"http:\/\/static.wixstatic.com\/media\/1caf47_8756d9328dcd470583148d69b76a078e.jpg_srz_1024_688_85_22_0.50_1.20_0.00_jpg_srz\" rel=\"nofollow noreferrer\">image example<\/a><\/p>\n\n<p>Do i need to train my own model? Is there a repository that i don't know?<\/p>\n","answers":[]}
{"QuestionId":41474654,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-04T22:58:49.637","AcceptedAnswerId":null,"OwnerUserId":2446538.0,"Title":"AttributeError: 'module' object has no attribute 'reset_default_graph'","Body":"<p>While using U-Net in tensorflow, using the command:\n  gives attribute error  . There is no tensorflow.py file which is being imported by default as in few similar issues I came across.<\/p>\n\n<p>How can I fix this?<\/p>\n","answers":[]}
{"QuestionId":41475180,"AnswerCount":5,"Tags":"<python><tensorflow>","CreationDate":"2017-01-04T23:54:19.503","AcceptedAnswerId":44099690.0,"OwnerUserId":6733064.0,"Title":"Understanding `tf.nn.nce_loss()` in tensorflow","Body":"<p>I am trying to understand the NCE loss function in Tensorflow. NCE loss is employed for a word2vec task, for instance:<\/p>\n\n<pre><\/pre>\n\n<p>more details, please reference Tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"noreferrer\">word2vec_basic.py<\/a><\/p>\n\n<ol>\n<li>What are the input and output matrices in the NCE function? <\/li>\n<\/ol>\n\n<p>In a word2vec model, we are interested in building representations for words. In the training process, given a slid window, every word will have two embeddings: 1) when the word is a centre word; 2) when the word is a context word. These two embeddings are called input and output vectors, respectively. (<a href=\"http:\/\/www-personal.umich.edu\/~ronxin\/pdf\/w2vexp.pdf\" rel=\"noreferrer\">more explanations of input and output matrices<\/a>)<\/p>\n\n<p>In my opinion, the input matrix is  and the output matrix is . Is it right?<\/p>\n\n<ol start=\"2\">\n<li>What is the final embedding?<\/li>\n<\/ol>\n\n<p>According to a <a href=\"https:\/\/stackoverflow.com\/questions\/37982478\/what-is-the-purpose-of-weights-and-biases-in-tensorflow-word2vec-example\">post<\/a> by s0urcer also relating to , it says the final embedding matrix is just the input matrix. While, <a href=\"https:\/\/stats.stackexchange.com\/questions\/177667\/input-vector-representation-vs-output-vector-representation-in-word2vec\">some others saying<\/a>, the . Which is right\/more common?<\/p>\n","answers":[{"AnswerId":"41957436","CreationDate":"2017-01-31T12:19:04.790","ParentId":null,"OwnerUserId":"4646735","Title":null,"Body":"<p>The <code>embeddings<\/code> Tensor is your final output matrix. It maps words to vectors. Use this in your word prediction graph.<\/p>\n\n<p>The input matrix is a batch of <code>centre-word : context-word<\/code> pairs (<code>train_input<\/code> and <code>train_label<\/code> respectively) generated from the training text. <\/p>\n\n<p>While the exact workings of the <code>nce_loss<\/code> op are not yet know to me, the basic idea is that it uses a single layer network (parameters <code>nce_weights<\/code> and <code>nce_biases<\/code>) to map an input vector (selected from <code>embeddings<\/code> using the <code>embed<\/code> op) to an output word, and then compares the output to the training label (an adjacent word in the training text) and also to a random sub-sample (<code>num_sampled<\/code>) of all other words in the vocab, and then modifies the input vector (stored in <code>embeddings<\/code>) and the network parameters to minimise the error.<\/p>\n"},{"AnswerId":"49295301","CreationDate":"2018-03-15T09:11:00.120","ParentId":null,"OwnerUserId":"5163915","Title":null,"Body":"<p>From the paper <strong><em>Learning word embeddings efficiently with noise-contrastive estimation<\/em><\/strong>:<\/p>\n\n<blockquote>\n  <p>NCE is based on the reduction of density estimationto probabilistic binary classification. The basic idea is to train a logistic regression classifier to\n  discriminate between samples from the data distribution and samples from some \u201cnoise\u201d distribution<\/p>\n<\/blockquote>\n\n<p>We could find out that in word embedding, the NCE is actually negative sampling. (For the difference between this two, see paper: <strong><em>Notes on Noise Contrastive Estimation and Negative Sampling<\/em><\/strong>)<\/p>\n\n<p>Therefore, you do not need to input the noise distribution. And also from the quote, you will find out that it is actually a logistic regression: weight and bias would be the one need for logistic regression. If you are familiar with word2vec, it is just adding a bias. <\/p>\n"},{"AnswerId":"44078789","CreationDate":"2017-05-19T20:50:38.450","ParentId":null,"OwnerUserId":"4670794","Title":null,"Body":"<p>1) In short, it is right in general, but just partly right for the function in question. See <a href=\"https:\/\/www.tensorflow.org\/tutorials\/word2vec#training_the_model\" rel=\"nofollow noreferrer\">tutorial<\/a>:<\/p>\n\n<blockquote>\n  <p>The noise-contrastive estimation loss is defined in terms of a\n  logistic regression model. For this, we need to define the weights and\n  biases for each word in the vocabulary (also called the <code>output weights<\/code>\n  <strong>as opposed<\/strong> to the <code>input embeddings<\/code>).<\/p>\n<\/blockquote>\n\n<p>So inputs to the function <code>nce_loss<\/code> are <code>output weights<\/code> and a small <strong>part<\/strong> of <code>input embeddings<\/code>, among the other stuff.<\/p>\n\n<p>2) 'Final' embedding (aka Word vectors, aka Vector representations of Words) is what you call <code>input matrix<\/code>. Embeddings are strings (vectors) of that matrix, corresponding to each word.<\/p>\n\n<p><strong>Warn<\/strong> In fact, this terminology is confusing because of input and output concepts usage in NN environment. Embeddings matrix is not an input to NN, as input to NN is technically an input layer. You obtain the final state of this matrix during training process. Nonetheless, this matrix should be initialised in programming, because an algorithm has to start from some random state of this matrix to gradually update it during training.<\/p>\n\n<p>The same is true for weights - they are to be initialised also. It happens in the following line:<\/p>\n\n<pre><code>nce_weights = tf.Variable(\n        tf.truncated_normal([50000, 128], stddev=1.0 \/ math.sqrt(128)))\n<\/code><\/pre>\n\n<p>Each embedding vector can be multiplied by a vector from weights matrix (in a string to column manner). So we will get the scalar in the NN <code>output layer<\/code>. The norm of this scalar is interpreted as probability that the target word (from input layer) will be accompanied by label [or context] word corresponding to the place of scalar in <code>output layer<\/code>.<\/p>\n\n<hr>\n\n<p>So, if we are saying about inputs (arguments) to the function, then both matrixes are such: weights and a batch sized extraction from embeddings:<\/p>\n\n<pre><code>tf.nn.nce_loss(weights=nce_weights,            # Tensor of shape(50000, 128)\n               biases=nce_biases,              # vector of zeros; len(128)\n               labels=train_labels,            # labels == context words enums\n               inputs=embed,                   # Tensor of shape(128, 128)\n               num_sampled=num_sampled,        # 64: randomly chosen negative (rare) words\n               num_classes=vocabulary_size))   # 50000: by construction\n<\/code><\/pre>\n\n<p>This <code>nce_loss<\/code> function outputs a vector of batch_size - in the TensorFlow example a <code>shape(128,)<\/code> tensor.\nThen <code>reduce_mean()<\/code> reduces this result to a scalar taken the mean of those 128 values, which is in fact an objective for further minimization.<\/p>\n\n<p>Hope this helps.<\/p>\n"},{"AnswerId":"44099690","CreationDate":"2017-05-21T17:19:09.850","ParentId":null,"OwnerUserId":"8003239","Title":null,"Body":"<p>Let's look at the relative code in <strong>word2vec<\/strong> example (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"noreferrer\">examples\/tutorials\/word2vec<\/a>).<\/p>\n\n<pre><code>embeddings = tf.Variable(\n    tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))\nembed = tf.nn.embedding_lookup(embeddings, train_inputs)\n<\/code><\/pre>\n\n<p>These two lines create embedding representations. <code>embeddings<\/code> is a matrix where each row represents a word vector. <code>embedding_lookup<\/code> is a quick way to get vectors corresponding to <code>train_inputs<\/code>. In word2vec example, <code>train_inputs<\/code> consists of some <code>int32<\/code> number, representing the <code>id<\/code> of target words. Basically, <strong>it can be placed by hidden layer feature<\/strong>.<\/p>\n\n<pre><code># Construct the variables for the NCE loss\nnce_weights = tf.Variable(\n    tf.truncated_normal([vocabulary_size, embedding_size],\n                        stddev=1.0 \/ math.sqrt(embedding_size)))\nnce_biases = tf.Variable(tf.zeros([vocabulary_size]))\n<\/code><\/pre>\n\n<p>These two lines create parameters. They will be updated by optimizer during training. We can use <code>tf.matmul(embed, tf.transpose(nce_weights)) + nce_biases<\/code> to get final output score. In other words, <strong>last inner-product layer in classification can be replaced by it<\/strong>.<\/p>\n\n<pre><code>loss = tf.reduce_mean(\n      tf.nn.nce_loss(weights=nce_weights,     # [vocab_size, embed_size]\n                   biases=nce_biases,         # [embed_size]\n                   labels=train_labels,       # [bs, 1]\n                   inputs=embed,              # [bs, embed_size]\n                   num_sampled=num_sampled, \n                   num_classes=vocabulary_size))\n<\/code><\/pre>\n\n<p>These lines create <code>nce loss<\/code>, <strong>@garej<\/strong> has given a very good explanation. <code>num_sampled<\/code> refers to the number of negative sampling in <code>nce<\/code> algorithm.<\/p>\n\n<hr>\n\n<p>To illustrate the usage of <code>nce<\/code>, we can apply it in <strong>mnist<\/strong> example (<a href=\"https:\/\/github.com\/yl-1993\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_deep.py\" rel=\"noreferrer\">examples\/tutorials\/mnist\/mnist_deep.py<\/a>) with following 2 steps:<\/p>\n\n<p><strong>1. Replace embed with hidden layer output<\/strong>. The dimension of hidden layer is <code>1024<\/code> and num_output is <code>10<\/code>. Minimum value of <code>num_sampled<\/code> is 1. Remember to remove the last inner-product layer in <code>deepnn()<\/code>.<\/p>\n\n<pre><code>y_conv, keep_prob = deepnn(x)                                            \n\nnum_sampled = 1                                                          \nvocabulary_size = 10                                                     \nembedding_size = 1024                                                    \nwith tf.device('\/cpu:0'):                                                \n  embed = y_conv                                                         \n  # Construct the variables for the NCE loss                             \n  nce_weights = tf.Variable(                                             \n      tf.truncated_normal([vocabulary_size, embedding_size],             \n                          stddev=1.0 \/ math.sqrt(embedding_size)))       \n  nce_biases = tf.Variable(tf.zeros([vocabulary_size])) \n<\/code><\/pre>\n\n<p><strong>2. Create loss and compute output<\/strong>. After computing the output, we can use it to calculate accuracy. Note that the label here is not one-hot vector as used in softmax. Labels are the original label of training samples.<\/p>\n\n<pre><code>loss = tf.reduce_mean(                                   \n    tf.nn.nce_loss(weights=nce_weights,                           \n                   biases=nce_biases,                             \n                   labels=y_idx,                                  \n                   inputs=embed,                                  \n                   num_sampled=num_sampled,                       \n                   num_classes=vocabulary_size))                  \n\noutput = tf.matmul(y_conv, tf.transpose(nce_weights)) + nce_biases\ncorrect_prediction = tf.equal(tf.argmax(output, 1), tf.argmax(y_, 1))\n<\/code><\/pre>\n\n<p>When we set <code>num_sampled=1<\/code>, the val accuracy will end at around <code>98.8%<\/code>. And if we set <code>num_sampled=9<\/code>, we can get almost the same val accuracy as trained by softmax. But note that <code>nce<\/code> is different from <code>softmax<\/code>.<\/p>\n\n<p>Full code of training <code>mnist<\/code> by <code>nce<\/code> can be found <a href=\"https:\/\/github.com\/yl-1993\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_deep_nce.py\" rel=\"noreferrer\">here<\/a>. Hope it is helpful.<\/p>\n"},{"AnswerId":"42392407","CreationDate":"2017-02-22T13:09:22.383","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<blockquote>\n  <p>What are the input and output matrices in the NCE function?<\/p>\n<\/blockquote>\n\n<p>Take for example the skip gram model, for this sentence: <\/p>\n\n<p><code>the quick brown fox jumped over the lazy dog<\/code><\/p>\n\n<p>the input and output pairs are: <\/p>\n\n<p><code>(quick, the), (quick, brown), (brown, quick), (brown, fox), ...<\/code><\/p>\n\n<p>For more information please refer to <a href=\"https:\/\/www.tensorflow.org\/tutorials\/word2vec\" rel=\"nofollow noreferrer\">the tutorial<\/a>.<\/p>\n\n<blockquote>\n  <p>What is the final embedding?<\/p>\n<\/blockquote>\n\n<p>The final embedding you should extract is usually <strong>the {w} between the input and hidden layer<\/strong>. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/447Nn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/447Nn.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>To illustrate more intuitively take a look at the following picture: \n<a href=\"https:\/\/i.stack.imgur.com\/Rlo4M.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Rlo4M.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The one hot vector [0, 0, 0, 1, 0] is the input layer in the above graph, the output is the word embedding [10, 12, 19], and W(in the graph above) is the matrix in between. <\/p>\n\n<p>For detailed explanation please read <a href=\"http:\/\/mccormickml.com\/2016\/04\/19\/word2vec-tutorial-the-skip-gram-model\/\" rel=\"nofollow noreferrer\">this tutorial<\/a>. <\/p>\n"}]}
{"QuestionId":41476065,"AnswerCount":0,"Tags":"<python><tensorflow><minimize><gradient-descent>","CreationDate":"2017-01-05T01:43:24.630","AcceptedAnswerId":null,"OwnerUserId":7376934.0,"Title":"Set variable list for gradient descent in TensorFlow","Body":"<p>I want to use gradient descent to minimize the variable \"loss\", but I don't want it to be able to update EVERY variable, just a few of them(<em>There are some variables in the same graph that I would like to keep constant for now.<\/em>). <\/p>\n\n<p>So, I listed the variables that I want to update in  in the last line of this <strong>code snippet<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Error I am getting<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p>I'm a bit confused why the elements in the list would need to have operations because I'm just updating the weights of the neural network. These are just variables, right? Does anybody guess what I'm doing wrong or what other approach do I take?<\/p>\n\n<p>Also, all this code works correctly when I don't specify .<\/p>\n","answers":[]}
{"QuestionId":41476404,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><eigen>","CreationDate":"2017-01-05T02:23:58.263","AcceptedAnswerId":null,"OwnerUserId":878520.0,"Title":"Tensorflow Installation failed. TensorContractionThreadPool.h: No such file or directory","Body":"<p>I try to install Tensorflow from source. <\/p>\n\n<pre><\/pre>\n\n<p>but when I do 'bazel build ...' it raise an Exception like this.. <\/p>\n\n<pre><\/pre>\n\n<p>To put it simple it says <strong>\"src\/Tensor\/TensorContractionThreadPool.h: No such file or directory\"<\/strong><\/p>\n\n<p>What can I do to solve this problem?<\/p>\n","answers":[{"AnswerId":"41491453","CreationDate":"2017-01-05T17:44:19.660","ParentId":null,"OwnerUserId":"316723","Title":null,"Body":"<p>Here's how I did it:<\/p>\n\n<ol>\n<li>Cloned <a href=\"https:\/\/bitbucket.org\/eigen\/eigen\" rel=\"nofollow noreferrer\">https:\/\/bitbucket.org\/eigen\/eigen<\/a><\/li>\n<li>Copied the <code>Tensor<\/code> folder into tensorflow: <code>cp -r eigen\/unsupported\/Eigen\/CXX11\/src\/Tensor tensorflow\/third_party\/eigen3\/unsupported\/Eigen\/CXX11\/src\/<\/code><\/li>\n<li><p>Edit <code>tensorflow\/third_party\/eigen3\/BUILD<\/code> by changing:<\/p>\n\n<pre><code>glob([\"unsupported\/Eigen\/CXX11\/src\/FixedPoint\/*.h\"]) + [\n<\/code><\/pre>\n\n<p>into:<\/p>\n\n<pre><code>hdrs = glob([\"unsupported\/Eigen\/CXX11\/src\/FixedPoint\/*.h\"]) +\n       glob([\"unsupported\/Eigen\/CXX11\/src\/Tensor\/*.h\"]) + [\n<\/code><\/pre>\n\n<p>You could probably also use <code>...\/src\/*\/*.h<\/code><\/p><\/li>\n<li>Run <code>bazel build<\/code> as normal<\/li>\n<\/ol>\n"},{"AnswerId":"41674898","CreationDate":"2017-01-16T11:02:40.267","ParentId":null,"OwnerUserId":"4732532","Title":null,"Body":"<p>@naktinis is right!\nbut hg clone repo from bitbucket is bother\uff0cu should install Mercurial first by:<\/p>\n\n<pre><code># Debian\/Ubuntu\n$ apt-get install mercurial\n\n# Fedora\n$ yum install mercurial\n\n# Gentoo\n$ emerge mercurial\n\n# Mac OS (homebrew)\n$ brew install mercurial\n\n# FreeBSD\n$ cd \/usr\/ports\/devel\/mercurial\n$ make install\n\n# Solaris 11 Express\n$ pkg install SUNWmercurial\n<\/code><\/pre>\n"}]}
{"QuestionId":41476519,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-05T02:38:33.647","AcceptedAnswerId":41491669.0,"OwnerUserId":3084800.0,"Title":"using MultiRNNCell in tensorflow 0.12","Body":"<p>With Tensorflow 0.12, there have been changes to the way that  works, for starters,  is now set to  by default, furthermore, there is this discussion on it:<\/p>\n\n<blockquote>\n  <p>: If , accepted and returned states are -tuples, where . If , the states are all concatenated along the column axis. This latter behavior will soon be deprecated.<\/p>\n<\/blockquote>\n\n<p>I'm wondering how exactly I could use a multi layer RNN with GRU cells, here is my code so far:<\/p>\n\n<pre><\/pre>\n\n<p>My inputs look up word ids and return a corresponding embedding vectors. Now, running with the code above I'm greeted by the following error:<\/p>\n\n<p><\/p>\n\n<p>The places I've got a  in is within my placeholders:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41491669","CreationDate":"2017-01-05T17:57:38.837","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Your main issue is in the setting of the <code>initial_state<\/code>. Since your state is now a tuple, (more specifically an <code>LSTMStateTuple<\/code>, you cannot directly assign it to <code>tf.zeros<\/code>. Instead use,<\/p>\n\n<pre><code>self.initial_state = cell.zero_state(self._batch_size, tf.float32)\n<\/code><\/pre>\n\n<p>Have a look at the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\/#stacking_multiple_lstms\" rel=\"nofollow noreferrer\">documentation<\/a> for more.<\/p>\n\n<hr>\n\n<p>To use this in code, you will need to pass this <em>tensor<\/em> in the <code>feed_dict<\/code>. Do something like this,<\/p>\n\n<pre><code>state = sess.run(model.initial_state)\nfor batch in batches:\n    # Logic to add input placeholder in `feed_dict`\n    feed_dict[model.initial_state] = state\n    # Note I'm re-using `state` below\n    (loss, state) = sess.run([model.loss, model.final_state], feed_dict=feed_dict)\n<\/code><\/pre>\n"}]}
{"QuestionId":41476613,"AnswerCount":0,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-05T02:49:37.240","AcceptedAnswerId":null,"OwnerUserId":4229089.0,"Title":"TensorFlow: getting invalid argument error, even though shape and dtype of feed dict is correct","Body":"<p>I'm having an issue with a ML model I'm building with TensorFlow. In the graph, I have a placeholder of type  and shape :<\/p>\n\n<pre><\/pre>\n\n<p>In the feed_dict, I pass in a NumPy array of type np.float32 and shape (100, 1):<\/p>\n\n<pre><\/pre>\n\n<p>I know the size and type are correct because when I run this,<\/p>\n\n<pre><\/pre>\n\n<p>it outputs the following:<\/p>\n\n<pre><\/pre>\n\n<p>However, I'm still getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>What could the issue be? I've run into it before, but I simply had to update TensorFlow and it worked. This time, however, I'm out of tricks, as TensorFlow is up to date. Any suggestions?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>I also got an insight from someone that it could be that I set h_prime_previous to something else somewhere in the code. That, however, is not happening, so that is definitely not the issue.<\/p>\n","answers":[]}
{"QuestionId":41476668,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-05T02:58:54.690","AcceptedAnswerId":null,"OwnerUserId":7377054.0,"Title":"Bad result of the tensorflow sequence to sequence demo","Body":"<p>I do what the official tutorial tells about the tensorflow sequence to sequence\nmodel . This is my steps and results:\nTraining Commands:<\/p>\n\n<pre><\/pre>\n\n<p>--data_dir [your_data_directory] --train_dir [checkpoints_directory]\n  --size=256 --num_layers=2 --steps_per_checkpoint=50<\/p>\n\n<p>Then I have run it for one week.Now this is the log result:\n<a href=\"https:\/\/i.stack.imgur.com\/5b06Q.png\" rel=\"nofollow noreferrer\">log result<\/a><\/p>\n\n<p>Since the training corpus contains ~22M sentences, one epoch (going through the training data once) takes about 340K steps with batch-size of 64. \nYou can see it will run over.And the learing rate is very very small.\nNow I decode ,this is my command:<\/p>\n\n<p>\nThis is the decode result:\n<a href=\"https:\/\/i.stack.imgur.com\/F479L.png\" rel=\"nofollow noreferrer\">decode result<\/a><\/p>\n\n<p>The first sentence is what the example of tutorial, you can see it is good.\nBut the rest sentences is bad ,so I am confused .it is really so bad or should I have to chang the model parameters ? I want to konw if anyone runs \nthis demo and the result is just like me or not? Am I wrong with something?\nThank you very much!!<\/p>\n","answers":[]}
{"QuestionId":41476955,"AnswerCount":1,"Tags":"<tensorflow><tflearn>","CreationDate":"2017-01-05T03:38:40.960","AcceptedAnswerId":null,"OwnerUserId":7338784.0,"Title":"how can i change the code for tensorflow 0.12 version","Body":"<p>I have a code, which was running before but not now.<\/p>\n\n<pre><\/pre>\n\n<p>But now I'm getting error which says <\/p>\n\n<p>How can I modify this instruction for Tensorflow 0.12?<\/p>\n","answers":[{"AnswerId":"41491567","CreationDate":"2017-01-05T17:51:50.757","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>You should replace <code>TensorFlowEstimator<\/code> with <code>Estimator<\/code>. There has been a significant re-factor in 0.12, please have a look at the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.learn\/estimators\" rel=\"nofollow noreferrer\"><code>Estimator<\/code><\/a> docs.<\/p>\n"}]}
{"QuestionId":41477094,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-05T03:57:05.903","AcceptedAnswerId":null,"OwnerUserId":7377224.0,"Title":"Adjust the number of classifications in the existing model provided by TensorFlow","Body":"<p>I have tried Denny Britz's code on TensorFlow CNN Text Classification. Currently, the code is fixated on the number of text classifications as provided by the existing model of TensorFlow. For example, the model has 8 types of classifications. But I wish to increase\/decrease the number of classifications. Is there a way to make it work?<\/p>\n","answers":[]}
{"QuestionId":41477722,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-05T05:15:27.193","AcceptedAnswerId":null,"OwnerUserId":7345996.0,"Title":"How to use collection in tensorflow to preserve my own object","Body":"<p>I want to use tf.add_to_collection() to preserve my own object in order to get them easily later.\nHere is the code segment:<\/p>\n\n<pre><\/pre>\n\n<p>add object to collection:<\/p>\n\n<pre><\/pre>\n\n<p>get object from the collection:<\/p>\n\n<pre><\/pre>\n\n<p>the code works fine, but i get a warning:<\/p>\n\n<pre><\/pre>\n\n<p>what shoud i do to avoid this warning?<\/p>\n","answers":[{"AnswerId":"41478762","CreationDate":"2017-01-05T06:43:59.813","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The warning is (probably) produced when you call <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.save\" rel=\"nofollow noreferrer\"><code>tf.train.Saver.save()<\/code><\/a>, and it attempts to write out a \"MetaGraph\" that represents the contents of the <code>tf.Graph<\/code>, including the contents of all of the graph's collections.<\/p>\n\n<p>The simplest way to avoid the warning is to pass <code>write_meta_graph=False<\/code> when calling <code>saver.save()<\/code>. However, this leaves you without a MetaGraph to import later.<\/p>\n\n<p>If you want to save the MetaGraph <em>and<\/em>  avoid the warning, you need to implement the necessary hooks (<code>to_proto<\/code> and <code>from_proto<\/code>) for serializing your <code>Model<\/code> object as a protocol buffer in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/d3822f06ca4440bf33d9034022e81ba4620f99c1\/tensorflow\/core\/protobuf\/meta_graph.proto#L31\" rel=\"nofollow noreferrer\"><code>tf.train.MetaGraphDef<\/code><\/a> serialization format. The <a href=\"https:\/\/www.tensorflow.org\/how_tos\/meta_graph\/\" rel=\"nofollow noreferrer\">MetaGraph tutorial<\/a> explains how to do this, but the basic idea is as follows:<\/p>\n\n<ol>\n<li><p>Define a protocol buffer (<code>ModelProto<\/code>) that describes the content of the <code>Model<\/code> object.<\/p><\/li>\n<li><p>Define a <code>model_to_proto()<\/code> function that serializes a <code>Model<\/code> into a <code>ModelProto<\/code>:<\/p>\n\n<pre><code>def model_to_proto(model):\n    ret = ModelProto()\n    # Set fields of `ret` from `model`.\n    return ret\n<\/code><\/pre><\/li>\n<li><p>Define a <code>model_from_proto()<\/code> function that deserializes a <code>ModelProto<\/code> and returns a <code>Model<\/code>:<\/p>\n\n<pre><code>def model_from_proto(model_proto):\n    # Construct a `Model` from the fields of `model_proto`.\n    return Model(...)\n<\/code><\/pre><\/li>\n<li><p>Register your functions for the <code>\"train_model\"<\/code> collection. This currently uses an undocumented function, called <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/d3822f06ca4440bf33d9034022e81ba4620f99c1\/tensorflow\/python\/framework\/ops.py#L4274\" rel=\"nofollow noreferrer\"><code>register_proto_function()<\/code><\/a>:<\/p>\n\n<pre><code>from tensorflow.python.framework import ops\n\nops.register_proto_function(\"train_model\",\n                            proto_type=ModelProto,\n                            to_proto=model_to_proto,\n                            from_proto=model_from_proto)\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":41478027,"AnswerCount":1,"Tags":"<python><session><tensorflow><distributed-computing><fault-tolerance>","CreationDate":"2017-01-05T05:43:26.890","AcceptedAnswerId":41478373.0,"OwnerUserId":6527887.0,"Title":"tf.train.MonitoredTrainingSession arguments","Body":"<p>what arguments does  take in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/distributed_execution#MonitoredTrainingSession\" rel=\"nofollow noreferrer\">tf.train.MonitoredTrainingSession<\/a>?. How can I specify the master node (at for eg ) with the proper syntax?<\/p>\n\n<p>Below is is the error I am encountering when i use  :-<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41478373","CreationDate":"2017-01-05T06:13:44.293","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>config<\/code> argument to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/distributed_execution#MonitoredTrainingSession\" rel=\"nofollow noreferrer\"><code>tf.train.MonitoredTrainingSession<\/code><\/a> takes a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/d3822f06ca4440bf33d9034022e81ba4620f99c1\/tensorflow\/core\/protobuf\/config.proto#L158\" rel=\"nofollow noreferrer\"><code>tf.ConfigProto<\/code><\/a> protocol buffer message.<\/p>\n\n<p>It looks like you should actually pass your argument (<code>\"grpc:\/\/localhost:2222\"<\/code>) as the <code>master<\/code> argument, which takes the same values as the <code>target<\/code> argument to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session.__init__\" rel=\"nofollow noreferrer\"><code>tf.Session<\/code> initializer<\/a>: e.g. <code>\"\"<\/code> means \"in-process runtime\", and <code>\"grpc:\/\/localhost:2222\"<\/code> means \"the gRPC-based <code>tf.train.Server<\/code> listening on <code>localhost:2222<\/code>.<\/p>\n"}]}
{"QuestionId":41478165,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-05T05:56:31.353","AcceptedAnswerId":null,"OwnerUserId":2878004.0,"Title":"Instability in asynchronous distributed tensorflow training","Body":"<p>I've been experimenting with using distributed Tensorflow on a language modeling task with two to three workers and a single parameter server. My workers are doing asynchronous gradient updates. This setup is definitely more efficient in terms of number of gradient updates per second but I'm able to get a lower error faster by just using a single training process.<\/p>\n\n<p>Are there any tips to getting distributed training to work? <\/p>\n\n<p>Is it important that there not be too many gradient updates per second? <\/p>\n\n<p>Do I need to use a much lower learning rate? What should I be trying?<\/p>\n\n<p>Another thing I noticed is that when using distributed training some mini batches can have a log loss that is 50 times bigger than what is expected. I'm curious what could be causing these spikes in the cost function and if it is important.<\/p>\n","answers":[]}
{"QuestionId":41478820,"AnswerCount":1,"Tags":"<android><tensorflow><android-gradle-plugin><bazel>","CreationDate":"2017-01-05T06:48:29.723","AcceptedAnswerId":41479456.0,"OwnerUserId":4917824.0,"Title":"Setting targetSdkVersion and compileSdkVersion for bazel, android tensorflow","Body":"<p>I need help in how to set the target and compile Sdk versions separately for bazel.<\/p>\n\n<p>I want to run tensorflow on the phone following <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\" rel=\"nofollow noreferrer\">this<\/a> tutorial. After editing WORKSPACE for bazel as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I get PARSE ERROR on the phone upon apk installing:<\/p>\n\n<pre><\/pre>\n\n<p>I have usb debugging and 3rd party apk installation enabled. The phone uses Android 4.4.4 so I changed WORKSPACE configuration to match it (tools and API installed by SDK manager properly) as:<\/p>\n\n<pre><\/pre>\n\n<p>Upon building as: <\/p>\n\n<pre><\/pre>\n\n<p>I end up with few errors of that type:<\/p>\n\n<pre><\/pre>\n\n<p>Following <a href=\"https:\/\/stackoverflow.com\/questions\/27812302\/error-retrieving-parent-for-item-no-resource-found-that-matches-the-given-name\">this<\/a> answer, I should build for compile API version > 21 but at the same time target API version 19. It can be easily set for gradle. However I have not found any clue how to do that for <a href=\"https:\/\/bazel.build\/versions\/master\/docs\/bazel-user-manual.html\" rel=\"nofollow noreferrer\">bazel<\/a>. Any suggestions are appreciated!<\/p>\n","answers":[{"AnswerId":"41479456","CreationDate":"2017-01-05T07:28:52.163","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>If you want to target a lower Android API level, the file you want to edit is actually tensorflow\/examples\/android\/AndroidManifest.xml. Leave the WORKSPACE file settings at the highest API level you have installed via your sdk (APIs can be installed with $ANDROID_HOME\/tools\/android).<\/p>\n\n<p>You can lower the min API level to 19 in AndroidManifest.xml, but be aware you will encounter runtime errors on your phone running the demo unless you replace the demo code that interfaces with the camera2 api (which was added with API level 21) with analogous android.hardware.camera code.<\/p>\n"}]}
{"QuestionId":41478897,"AnswerCount":1,"Tags":"<python><anaconda><deep-learning><caffe>","CreationDate":"2017-01-05T06:53:07.387","AcceptedAnswerId":null,"OwnerUserId":7377690.0,"Title":"make all fatal error,when i make caffe","Body":"<p>I compiled caffe on manjaro.When I run make all in the caffe root folder,I get:<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix this?\nThis is some of my Makefile.config:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41481619","CreationDate":"2017-01-05T09:41:50.047","ParentId":null,"OwnerUserId":"6573448","Title":null,"Body":"<p>go threw <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow noreferrer\">caffe installation guide<\/a>. <\/p>\n\n<p>before building caffe, make sure you install:<\/p>\n\n<pre><code>sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libboost-all-dev libhdf5-serial-dev libgflags-dev libgoogle-glog-dev liblmdb-dev protobuf-compiler libatlas-base-dev\nsudo apt-get install python-dev python-pip gfortran\n<\/code><\/pre>\n\n<p>then, try to build build caffe that way:<\/p>\n\n<pre><code>cd caffe\ncp Makefile.config.example Makefile.config\nand change setting in Makefile.config\nmkdir build\ncd build\ncmake ..\n.\/configure\nmake all\n<\/code><\/pre>\n"}]}
{"QuestionId":41479097,"AnswerCount":0,"Tags":"<c><tensorflow>","CreationDate":"2017-01-05T07:06:59.227","AcceptedAnswerId":null,"OwnerUserId":6234868.0,"Title":"tensorflow how to get the data type of a node using c api?","Body":"<p>Suppose my graph have a node below:<\/p>\n\n<pre><\/pre>\n\n<p>and I want to get the data type of it (which is DT_STRING in this case).<\/p>\n\n<p>I tried this:<\/p>\n\n<pre><\/pre>\n\n<p>but the result of type is TF_ATTR_PLACEHOLDER.<\/p>\n\n<p>and this:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41479303,"AnswerCount":1,"Tags":"<deep-learning><caffe><conv-neural-network><pycaffe><matcaffe>","CreationDate":"2017-01-05T07:20:04.317","AcceptedAnswerId":41481379.0,"OwnerUserId":6494707.0,"Title":"I am getting Incorrect data field size error while generating the mean image of training data, even after generating by using create_imagenet.sh,why?","Body":"<p>I am trying to generate the mean image from training data in . My data is 256x256 grayscale images. I created lmdb by using  by replecing  with . <\/p>\n\n<p>I edited create_imagenet.sh as follows:<\/p>\n\n<pre><\/pre>\n\n<p>But I am still getting error while creating .<\/p>\n\n<pre><\/pre>\n\n<p>Here is the error:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone have any suggestion to tackle this error? <\/p>\n\n<p>Your help is really appreciated. <\/p>\n","answers":[{"AnswerId":"41481379","CreationDate":"2017-01-05T09:31:02.367","ParentId":null,"OwnerUserId":"6573448","Title":null,"Body":"<p>Make sure your images are <code>256X256<\/code>, you must convert all the images into a <strong>same size<\/strong> before runing the net.<\/p>\n\n<p>adding <code>--gray<\/code> to your <code>LMDB<\/code> will create your images as single-channel.<\/p>\n\n<pre><code>layer {\n  name: \"data\"\n  type: \"Data\"\n  [...]\n  transform_param {\n    scale: 0.1\n    mean_file_size: mean.binaryproto\n    # for images in particular horizontal mirroring and random cropping\n    # can be done as simple data augmentations.\n    mirror: 1  # 1 = on, 0 = off\n    # crop a `crop_size` x `crop_size` patch:\n    # - at random during training\n    # - from the center during testing\n    crop_size: 256 # cropping your images\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":41480149,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-01-05T08:17:15.590","AcceptedAnswerId":null,"OwnerUserId":7372608.0,"Title":"How does ps work in distribute Tensorflow?","Body":"<p>After learning multi-gpu sample(<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.11\/tensorflow\/models\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"nofollow noreferrer\">multi_gpu_sample<\/a>), I think the distribute tensorflow is similar with multi-gpu, cpu to ps and gpu to worker.<\/p>\n\n<p>So, ps's work is collecting gradients from worker , then update the parameter and send and share them to the workers.<\/p>\n\n<p>But after reading the distribute Tensorflow sample below, I felt confused.\nIt seems that ps do nothing but join() ops.<\/p>\n\n<p>How to understand this? Thanks! <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"41482827","CreationDate":"2017-01-05T10:39:35.583","ParentId":null,"OwnerUserId":"7287271","Title":null,"Body":"<p>It's a pretty confusing example (<a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">which came from the manual, I believe<\/a>). It will probably be changed as Distributed TensorFlow matures.<\/p>\n\n<p>Anyway, both \"worker\" and \"ps\" are <em>tasks<\/em> (or <em>jobs<\/em> rather, which are just groups of <em>tasks<\/em>), so they're really no different. The difference is what they're supposed to be used for. The idea is that state (e.g. <code>tf.Variable<\/code>) should be on the parameter servers, while operations for calculating state should be on the workers. Instead of achieving this by calling <code>tf.device<\/code> manually everywhere, a helper function named <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/distributed_execution#replica_device_setter\" rel=\"nofollow noreferrer\"><code>tf.train.replica_device_setter<\/code><\/a> is used that sets a <code>tf.Variable<\/code>'s device to a parameter server, and the other operations to a worker.<\/p>\n\n<p><code>server.join()<\/code> just means that parameter servers will wait on the workers, instead of terminating their client processes immediately.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.device(tf.replica_device_setter(\n  worker_device=\"\/job:worker\/task:%d\" % FLAGS.task_index, \n  cluster=cluster_spec)):\n\n  v1 = tf.Variable(...)  # Automatically assigned to a parameter server.\n  train_op = ... # Automatically assigned to the worker.\n<\/code><\/pre>\n"}]}
{"QuestionId":41482329,"AnswerCount":1,"Tags":"<ubuntu><caffe>","CreationDate":"2017-01-05T10:16:29.590","AcceptedAnswerId":null,"OwnerUserId":7173164.0,"Title":"Makefile:619: recipe for target '.build_release\/tools\/convert_imageset.bin' failed error at caffe installation on Ubuntu","Body":"<p>I'm trying install BVLC Caffe framework on ubuntu 16.04. I go through all installation steps at their own github <a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow noreferrer\">site<\/a>. When I type \"make all\" I get this errors :<\/p>\n\n<blockquote>\n  <p>CXX\/LD -o .build_release\/tools\/convert_imageset.bin \/usr\/bin\/ld:\n  cannot find -lLIBRARIES \/usr\/bin\/ld: cannot find -l+= \/usr\/bin\/ld:\n  cannot find -lopencv_imgcodecs collect2: error: ld returned 1 exit\n  status Makefile:619: recipe for target\n  '.build_release\/tools\/convert_imageset.bin' failed make: ***\n  [.build_release\/tools\/convert_imageset.bin] Error 1<\/p>\n<\/blockquote>\n\n<p>I search about it and tried<\/p>\n\n<pre><\/pre>\n\n<p>It works fine for me and I add caffe to PYTHONPATH then I can use \"import caffe\" line. But when I tried one of pre-trained caffe models from this <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html\" rel=\"nofollow noreferrer\">site<\/a> and write that to console :<\/p>\n\n<pre><\/pre>\n\n<p>I get this error:<\/p>\n\n<blockquote>\n  <p>No command 'caffe' found, did you mean:  Command 'caff' from package\n  'signing-party' (universe) caffe: command not found<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"54557141","CreationDate":"2019-02-06T15:30:12.583","ParentId":null,"OwnerUserId":"6818619","Title":null,"Body":"<p>I myself solve the error<\/p>\n\n<blockquote>\n  <p>No command 'caffe' found, did you mean: Command 'caff' from package 'signing-party' (universe) caffe: command not found<\/p>\n<\/blockquote>\n\n<p>this way:\nInstead of writing in terminal just <code>caffe &lt;...&gt;<\/code> I write <code>\/my\/path\/to\/caffe\/build\/tools\/caffe &lt;...&gt;<\/code>. The same way I do, for example, <code>convert_imageset<\/code> to get lmdb: <code>\/my\/path\/to\/caffe\/build\/tools\/convert_imageset &lt;...&gt;<\/code>. In other words, I always specify Caffe directory.<\/p>\n"}]}
{"QuestionId":41482823,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-05T10:39:32.397","AcceptedAnswerId":41492093.0,"OwnerUserId":7378465.0,"Title":"What do the functions tf.squeeze and tf.nn.rnn do?","Body":"<p>What do the functions tf.squeeze and tf.nn.rnn do?<\/p>\n\n<p>I searched these API, but I can't find argument, examples etc.\nAlso, what is the shape of  formed by the following code using , and what is the meaning and case of using ?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"58843005","CreationDate":"2019-11-13T18:00:35.853","ParentId":null,"OwnerUserId":"10123804","Title":null,"Body":"<p>tf.squeeze removes deimesion whose size is \"1\".Below example will show use of tf.squeeze.<\/p>\n\n<pre><code>import tensorflow as tf\ntf.enable_eager_execution() ##if using TF1.4 for TF2.0 eager mode is the default mode.\n####example 1\na = tf.constant(value=[1,3,4,5],shape=(1,4))\nprint(a)\nOutput : tf.Tensor([[1 3 4 5]], shape=(1, 4), dtype=int32)\n\n#after applying tf.squeeze shape has been changed from  (4,1) to (4, )\nb = tf.squeeze(input=a)\nprint(b)\noutput: tf.Tensor([1 3 4 5], shape=(4,), dtype=int32)\n####example2\na = tf.constant(value=[1,3,4,5,4,6], shape=(3,1,2))\nprint(a)\nOutput:tf.Tensor(\n[[[1 3]]\n [[4 5]]\n [[4 6]]], shape=(3, 1, 2), dtype=int32)\n\n#after applying tf.squeeze shape has been chnaged from (3, 1, 2) to (3, 2)\nb = tf.squeeze(input=a)\nprint(b)\nOutput:tf.Tensor(\n[[1 3]\n [4 5]\n [4 6]], shape=(3, 2), dtype=int32)\n<\/code><\/pre>\n"},{"AnswerId":"41492093","CreationDate":"2017-01-05T18:25:32.787","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The best source of answers to questions like these is the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/\" rel=\"nofollow noreferrer\">TensorFlow API documentation<\/a>. The two functions you mentioned create operations and symbolic tensors in a dataflow graph. In particular:<\/p>\n\n<ul>\n<li><p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/squeeze\" rel=\"nofollow noreferrer\"><code>tf.squeeze()<\/code><\/a> function returns a tensor with the same value as its first argument, but a different shape. It removes dimensions whose size is one. For example, if <code>t<\/code> is a tensor with shape <code>[batch_num, 1, elem_num]<\/code> (as in your question), <code>tf.squeeze(t, [1])<\/code> will return a tensor with the same contents but size <code>[batch_num, elem_num]<\/code>.<\/p><\/li>\n<li><p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/static_rnn\" rel=\"nofollow noreferrer\"><code>tf.nn.rnn()<\/code><\/a> function returns a pair of results, where the first element represents the outputs of a recurrent neural network for some given input, and the second element represents the final state of that network for that input. The TensorFlow website has a <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\/\" rel=\"nofollow noreferrer\">tutorial on recurrent neural networks<\/a> with more details.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":41482913,"AnswerCount":3,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2017-01-05T10:43:17.747","AcceptedAnswerId":41483033.0,"OwnerUserId":68571.0,"Title":"'module' object has no attribute 'SummaryWriter'","Body":"<p>I'm using Tensorflow version 0.12.head with Python 2.7 on a linux CentOS 7 and when I run this:<\/p>\n\n<pre><\/pre>\n\n<p>I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>I have run these two commands because there is bug <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1645\" rel=\"noreferrer\">issue<\/a> on Github for the same problem:<\/p>\n\n<pre><\/pre>\n\n<p>I'm new in Python and in Tensorflow. Do you know how can I fix this error?<\/p>\n\n<p>I have changed  with :<\/p>\n\n<pre><\/pre>\n\n<p>And I get the same error but with  function:<\/p>\n\n<pre><\/pre>\n\n<p>I have also run it in a terminal and I get the same result:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41483033","CreationDate":"2017-01-05T10:49:01.360","ParentId":null,"OwnerUserId":"499581","Title":null,"Body":"<p><code>tf.train.SummaryWriter<\/code> is deprecated, instead use <code>tf.summary.FileWriter<\/code>.<\/p>\n\n<p>\u21b3 <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/FileWriter\" rel=\"noreferrer\">Adding Summaries to Event Files<\/a><\/p>\n\n<blockquote>\n  <p>It will be removed after 2016-11-30.\n  Instructions for updating: Please switch to <code>tf.summary.FileWriter<\/code>. \n  The interface and behavior is the same; this is just a rename.<\/p>\n<\/blockquote>\n\n<p>&lt;<strong><a href=\"https:\/\/web.archive.org\/web\/20180911010215\/https:\/\/www.tensorflow.org\/install\/migration\" rel=\"noreferrer\"><code>TF Official Migration Page<\/code><\/a><\/strong>> \u2733\ufe0e includes all current deprecated\/renamed functions \u2733\ufe0e <\/p>\n"},{"AnswerId":"42513770","CreationDate":"2017-02-28T16:16:44.153","ParentId":null,"OwnerUserId":"7588645","Title":null,"Body":"<p>i had the same problem...i am using pything 3.5.2...see solution below...hope this works for you..it did for me (it will create a log in your tmp folder):<\/p>\n\n<pre><code>import tensorflow as tf\na = tf.constant(5, name=\"input_a\")\nb = tf.constant(3, name=\"input_a\")\nc = tf.multiply(a,b, name=\"mul_c\")\nd = tf.add(a,b, name=\"add_d\")\ne = tf.add(c,d, name=\"add_e\")\n\nsess = tf.Session()\nsess.run(e)\noutput = sess.run(e)\n\nwriter = tf.summary.FileWriter('\/tmp\/tensorflow_logs', graph=sess.graph)\n\nprint(sess.run(e))\n<\/code><\/pre>\n"},{"AnswerId":"43752678","CreationDate":"2017-05-03T06:31:29.437","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>In a new version of TF, all <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"noreferrer\">summary functions were renamed<\/a>.<\/p>\n\n<blockquote>\n  <p>Summary functions have been consolidated under the <strong>tf.summary<\/strong>\n  namespace.<\/p>\n<\/blockquote>\n\n<pre><code> Deprecated                                               Replacement\n----------------------------------------------------------------------------------\n tf.audio_summary                                         tf.summary.audio\n tf.contrib.deprecated.histogram_summary                  tf.summary.histogram\n tf.contrib.deprecated.scalar_summary                     tf.summary.scalar\n tf.histogram_summary                                     tf.summary.histogram\n tf.image_summary                                         tf.summary.image\n tf.merge_all_summaries                                   tf.summary.merge_all\n tf.merge_summary                                         tf.summary.merge\n tf.scalar_summary                                        tf.summary.scalar\n tf.train.SummaryWriter                                   tf.summary.FileWriter\n----------------------------------------------------------------------------------\n<\/code><\/pre>\n"}]}
{"QuestionId":41483772,"AnswerCount":1,"Tags":"<python><python-3.x><queue><tensorflow><neural-network>","CreationDate":"2017-01-05T11:22:24.500","AcceptedAnswerId":41546353.0,"OwnerUserId":1116408.0,"Title":"Using queues in TensorFlow to load images and labels from text file","Body":"<p>I am trying to run a very simple neural network in TensorFlow which will learn to classify images. So far it is extremely simple, because I am still learning the framework.<\/p>\n\n<p>So far I am struggling to load the data - my data is in TXT file. Every line contains ID of the photo and a binary number which is used as a label.<\/p>\n\n<p>This is my code so far (I have stripped unrelated parts):<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, the network is a very simple one, with only one neuron. I was inspired by the code listed here: <a href=\"https:\/\/stackoverflow.com\/questions\/34340489\/tensorflow-read-images-with-labels\">Tensorflow read images with labels<\/a> After running the code I am getting the following error on the very first iteration:<\/p>\n\n<pre><\/pre>\n\n<p>I have tried to fix the problem for a couple of hours now. So far I checked that:<\/p>\n\n<ul>\n<li>filenames_list and labels_list are properly loaded,<\/li>\n<li>the shapes of tensors (x, y, y_, W and b) are correct<\/li>\n<li>TensorFlow graph is properly constructed and visible in TensorBoard.<\/li>\n<\/ul>\n\n<p>No idea what else should I check. It seems that I don't understand something about Queues in TensorFlow, but I do not know what exactly. Thanks in advance for any help!<\/p>\n","answers":[{"AnswerId":"41546353","CreationDate":"2017-01-09T10:52:49.977","ParentId":null,"OwnerUserId":"4834515","Title":null,"Body":"<p>It might be caused by <code>num_epochs=1<\/code> here <code>tf.train.slice_input_producer([filenames, labels], num_epochs=1, shuffle=True)<\/code>. You can check api of <code>slice_input_producer<\/code>, where it explains: num_epochs: An integer (optional). If specified, slice_input_producer produces each slice num_epochs times before generating an OutOfRange error.<\/p>\n"}]}
{"QuestionId":41484413,"AnswerCount":1,"Tags":"<deep-learning><keras><keras-layer>","CreationDate":"2017-01-05T11:52:21.920","AcceptedAnswerId":47185032.0,"OwnerUserId":6354331.0,"Title":"How to split a tensor column-wise in Keras to implement STFCN","Body":"<p>I'd like to implement the Spatiotemporal Fully Convolutional Network (STFCN) in Keras. I need to feed each depth column of a 3D convolutional output, e.g. a tensor with shape , as input to a separate LSTM.<\/p>\n\n<p>To make this clear, I have a  tensor of dimensions . I need to split (explicitly or implicitly) the tensor into 16 * 16 = 256 tensors of shape .<\/p>\n\n<p>Here is a diagram from the STFCN paper to illustrate the Spatio-Temporal module. What I described above is the arrow between 'Spatial Features' and 'Spatio-Temporal Module'.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/V253r.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/V253r.png\" alt=\"The connection between FCn and Spatio-Temporal Module is the relevant part of the diagram.\"><\/a><\/p>\n\n<p>How would this idea be best implemented in Keras?<\/p>\n","answers":[{"AnswerId":"47185032","CreationDate":"2017-11-08T16:36:49.927","ParentId":null,"OwnerUserId":"5165781","Title":null,"Body":"<p>You can use <code>tf.split<\/code> from Tensorflow using Keras <code>Lambda<\/code> layer<\/p>\n\n<p>Use Lambda to split a tensor of shape <code>(64,16,16)<\/code> into <code>(64,1,1,256)<\/code> and then  subset any indexes you need.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nimport keras.backend as K\nfrom keras.models import  Model\nfrom keras.layers import Input, Lambda\n\n# input data\ndata = np.ones((3,64,16,16))\n\n# define lambda function to split\ndef lambda_fun(x) : \n    x = K.expand_dims(x, 4)\n    split1 = tf.split(x, 16, 2)\n    x = K.concatenate(split1, 4)\n    split2 = tf.split(x, 16, 3)\n    x = K.concatenate(split2, 4)\n    return x\n\n## check thet splitting works fine\ninput = Input(shape= (64,16,16))\nll = Lambda(lambda_fun)(input)\nmodel = Model(inputs=input, outputs=ll)\nres = model.predict(data)\nprint(np.shape(res))    #(3, 64, 1, 1, 256)\n<\/code><\/pre>\n"}]}
{"QuestionId":41484580,"AnswerCount":3,"Tags":"<python><machine-learning><keras>","CreationDate":"2017-01-05T12:01:07.547","AcceptedAnswerId":41501173.0,"OwnerUserId":5080218.0,"Title":"GridSearch with Keras Neural Networks","Body":"<p>I'm trying to perform parameters tuning for a neural network built with keras. This is my code with a comment on the line that causes the error:<\/p>\n\n<pre><\/pre>\n\n<p>And this is the errors report:<\/p>\n\n<pre><\/pre>\n\n<p>Am I missing something? The grid search goes well with random forests, svm and logistic regression. I only have problems with Neural Networks.<\/p>\n","answers":[{"AnswerId":"45604630","CreationDate":"2017-08-10T04:41:35.067","ParentId":null,"OwnerUserId":"1917882","Title":null,"Body":"<p>I think you maybe use scikit-learn 0.16 or earlier version.<br>\nI just had ran into same issue yesterday and after some workarounds I came to know that upgrading scikit-learn from 0.16 to 0.18 solves the issue.<\/p>\n\n<pre><code>clf.fit(X_train, Y_train) #SHOULD WORK with scikit-learn 0.18   \n<\/code><\/pre>\n\n<p>One more thing that 0.18 differs from 0.16 is the GridSearchCV doesn't come up with <code>sklearn.grid_search<\/code> but with <code>sklearn.model_selection<\/code><\/p>\n"},{"AnswerId":"53478944","CreationDate":"2018-11-26T10:16:36.860","ParentId":null,"OwnerUserId":"7435524","Title":null,"Body":"<p>I hope that you've solved the problem by now.<\/p>\n\n<p>a) I guess the problem is that you're not returning the model at the end of the wrapper function <code>tuning()<\/code>. Use <code>return model<\/code><\/p>\n\n<p>b) <code>k_model = KerasClassifier(build_fn=model, verbose=0)<\/code> I think should be <code>build_fn=tuning<\/code> according to how you named your function.<\/p>\n\n<p>c) The method's signature <code>def tuning(X_train,Y_train,X_test,Y_test)\n<\/code> isn't correct. Instead the parameters to be passed to the function, are those that need to be replaced after every iteration (i.e. the hyperparameters you specified in <code>param_grid<\/code>). Use <code>def tuning(batch_size, nb_epoch)<\/code> instead.<\/p>\n\n<p>I hope that was helpful!<\/p>\n"},{"AnswerId":"41501173","CreationDate":"2017-01-06T07:43:34.880","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Here the error indicates that the <code>build_fn<\/code> needs to have 2 arguments as indicated from the # of parameters from <code>param_grid<\/code>. <\/p>\n\n<p>So you need to explicitly define an new function and use that as <code>build_fn=make_model<\/code><\/p>\n\n<pre><code>def make_model(batch_size, nb_epoch):\n    model = Sequential()\n    model.add(Dense(in_size, input_dim=in_size, init='uniform', activation='relu'))\n    model.add(Dense(8, init='uniform', activation='relu'))\n    model.add(Dense(1, init='uniform', activation='sigmoid'))\n    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n    return model\n<\/code><\/pre>\n\n<p>Also check <code>keras\/examples\/mnist_sklearn_wrapper.py<\/code> where <code>GridSearchCV<\/code> is used for hyper-parameter search. <\/p>\n"}]}
{"QuestionId":41485640,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-05T12:52:32.007","AcceptedAnswerId":null,"OwnerUserId":7378936.0,"Title":"Tensorflow profiling timelines inconsistent with wall time","Body":"<p>I am currently developing and profiling a neural network in tensorflow on a GTX 1070. My first implementation turned out to be rather inefficient due to some expensive FFT ops.\nThe upper timeline in the attatched pic reveals that one training step takes about 115ms. A second (mathematically equivalent) implementation seems to consume much less time, one step takes under 10ms.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2djsH.png\" rel=\"nofollow noreferrer\">timelines<\/a>\n(Sorry, I am not allowed to post pics yet)<\/p>\n\n<p>Surprisingly both methods take approx. the same time when measured in python via:<\/p>\n\n<pre><\/pre>\n\n<p>In both cases 100 runs take ~17s > 100*115ms > 100*10ms which is more than I would expect from the python overhead. \nThe network is fed using queues but the results are the same directly feeding numpy arrays into the graph so I concluded the empyt queues cannot be the bottleneck.<\/p>\n\n<p>The timelines are in principle recorded with the difference that I measure two runs:<\/p>\n\n<pre><\/pre>\n\n<p>The first runs always take much longer than the corresponding following runs, is tensorflow doing some optimization here or allocating memory for the first time?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mtqdN.png\" rel=\"nofollow noreferrer\">timelines_first_runs<\/a><\/p>\n\n<p>Can someone explain these inconistent results?\nWhy does my second implementation take the same time measured in python when the profiling timeline shows that it should run much faster?<\/p>\n","answers":[{"AnswerId":"41490776","CreationDate":"2017-01-05T17:08:07.173","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>First run can take longer because of global memory pool growing, PTX compilation and optimizations.<\/p>\n\n<p>You could turn off some of the optimizations as follows<\/p>\n\n<pre><code>sess = tf.Session(config=tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))\n<\/code><\/pre>\n\n<p>However, PTX compilation can still add 1-2 seconds to the first run. So for profiling purposes people usually include a \"pre-warming\" run to run the same computation before profiling it.<\/p>\n\n<p>Also, if your <code>train_op<\/code> is something like <code>assign<\/code> or <code>assign_add<\/code>, running <code>train_op<\/code> will fetch the result back into Python. Python&lt;->TensorFlow memory transfers are not reflected in the timeline (grpc transfers are also missing, tracking <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4809\" rel=\"nofollow noreferrer\">issue<\/a>). For profiling purposes you could run the op and omit the Python transfer as follows<\/p>\n\n<pre><code>sess.run(train_op.op)\n<\/code><\/pre>\n"}]}
{"QuestionId":41486353,"AnswerCount":0,"Tags":"<tensorflow><neural-network><keras><recurrent-neural-network><lstm>","CreationDate":"2017-01-05T13:28:37.870","AcceptedAnswerId":null,"OwnerUserId":6469175.0,"Title":"Wrong Number of LSTM Dimensions in Keras","Body":"<p>I am trying to classify texts using Recurrent Neural Networks and am facing issue while implementing the following Python Code (backend is on TensorFlow):<\/p>\n\n<pre><\/pre>\n\n<p>My input ie X is an Nd Array consisting of 411 rows &amp; 407 columns while Y consists of 411 rows &amp; 7 columns and following is the error I am getting:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41487069,"AnswerCount":2,"Tags":"<android><tensorflow>","CreationDate":"2017-01-05T14:04:20.127","AcceptedAnswerId":41495062.0,"OwnerUserId":4099895.0,"Title":"Build error with TensorFlow Android demo","Body":"<p>Getting protobuf error while compiling TensorFlow Android demo\nUsing Android Studio on OSX.<\/p>\n\n<p>Steps taken:<\/p>\n\n<ol>\n<li><p>installed Bazel upgraded view brew<\/p><\/li>\n<li><p>Cloned tensorflow repo with --recursive flags.<\/p><\/li>\n<li><p>run .configure .<\/p><\/li>\n<li><p>run the project<\/p><\/li>\n<\/ol>\n\n<blockquote>\n  <p>Error:\/private\/var\/tmp\/_bazel_xxxxx\/1b55123784db7232f203473c18395341\/external\/protobuf\/BUILD:73:1:\n  C++ compilation of rule '@protobuf\/\/:protobuf_lite' failed: false\n  failed: error executing command \/bin\/false -MD -MF\n  bazel-out\/stub_armeabi-v7a-opt\/bin\/external\/protobuf\/_objs\/protobuf_lite\/external\/protobuf\/src\/google\/protobuf\/wire_format_lite.pic.d\n  ... (remaining 26 argument(s) skipped):\n  com.google.devtools.build.lib.shell.BadExitStatusException: Process\n  exited with status 1.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42614810","CreationDate":"2017-03-05T22:24:26.100","ParentId":null,"OwnerUserId":"112705","Title":null,"Body":"<p>If you're looking for a quick way to build the TensorFlow Android demo, you can follow the instructions I wrote in this blog post:<br>\n<a href=\"https:\/\/medium.com\/@daj\/creating-an-image-classifier-on-android-using-tensorflow-part-2-9ec0698d801e#.a7kxjliws\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@daj\/creating-an-image-classifier-on-android-using-tensorflow-part-2-9ec0698d801e#.a7kxjliws<\/a><\/p>\n\n<p>To help make this easier, I've created a Docker image that has all the dependencies and configuration you need, so you should just be able to install Docker, download my Docker image, then build.  Good luck!<\/p>\n\n<pre><code>docker pull danjarvis\/tensorflow-android:1.0.0\n<\/code><\/pre>\n"},{"AnswerId":"41495062","CreationDate":"2017-01-05T21:33:05.343","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>You'll need to also edit your WORKSPACE file with your NDK and SDK settings according to the directions at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android<\/a>. \/bin\/false errors are typically seen when Bazel can't find the appropriate executable to run, in this case the ndk gcc compiler.<\/p>\n"}]}
{"QuestionId":41487396,"AnswerCount":1,"Tags":"<theano>","CreationDate":"2017-01-05T14:21:28.597","AcceptedAnswerId":null,"OwnerUserId":2997232.0,"Title":"Nan error of gradients in Theano","Body":"<p>When I construct such a computation<\/p>\n\n<pre><\/pre>\n\n<p>w1, w2, b1, b2 are parameters. And I construct a cross-entropy as loss function.<\/p>\n\n<p>But when I take gradients<\/p>\n\n<pre><\/pre>\n\n<p>All the resulted gradients are nan.<\/p>\n\n<p>Is there any possible reasons for this problem? And any solution to it?<\/p>\n\n<p>Thank you in advance!<\/p>\n","answers":[{"AnswerId":"41503747","CreationDate":"2017-01-06T10:32:57.903","ParentId":null,"OwnerUserId":"2997232","Title":null,"Body":"<p>I got the solution. Basically numerical instability. By changing log sigmoid to softplus solved this problem.<\/p>\n"}]}
{"QuestionId":41487791,"AnswerCount":0,"Tags":"<python><vector><tensorflow><embedding><lstm>","CreationDate":"2017-01-05T14:41:48.180","AcceptedAnswerId":null,"OwnerUserId":7378648.0,"Title":"How to use pretrained GloVe vectors in a tensorflow LSTM generative model","Body":"<p>My goal is to ask if it is possible to use pre-trained GloVe vectors in a Tensorflow word-rnn LSTM generative model, and if so any guidance on how to achieve this?<\/p>\n\n<p>I am referencing this from <a href=\"https:\/\/github.com\/hunkim\/word-rnn-tensorflow\" rel=\"nofollow noreferrer\">here<\/a> \nand I understand(I think) that I am supposed to put the vectors in the embeddings in  of the model.py. From the code, I see that he is not using any pre-trained vectors but the words from the input text.<\/p>\n\n<p>I have seen other answers like <a href=\"https:\/\/stackoverflow.com\/questions\/35687678\/using-a-pre-trained-word-embedding-word2vec-or-glove-in-tensorflow\">this<\/a> but as I am new to Tensorflow and Python I do not fully understand how to apply this into the code.<\/p>\n\n<p>GloVe generates two files, namely: <\/p>\n\n<ol>\n<li>vocabulary file, with the count of all word occurrences<\/li>\n<li>vector file. e.g the word []<\/li>\n<\/ol>\n\n<p>Also, do I have to generate the GloVe vectors and train the LSTM model on the same corpus or can they be separate? eg. GloVe(),  text_to_train()<\/p>\n\n<p>Thank you for the assistance!<\/p>\n","answers":[]}
{"QuestionId":41488031,"AnswerCount":1,"Tags":"<neural-network><deep-learning><keras><conv-neural-network><keras-layer>","CreationDate":"2017-01-05T14:54:43.867","AcceptedAnswerId":41488168.0,"OwnerUserId":3827807.0,"Title":"Get desired output from convolution (GAN)","Body":"<p>I'm trying to code a GAN model for cifar10.\nBut I have a problem.<\/p>\n\n<p>How to get the desired ouput (3x32x32) from a convolutional network ?\nI actually inspire my model from one I found for the mnist :<\/p>\n\n<pre><\/pre>\n\n<p>So, from there, I have an output of 3x28x28\nDo you know how I can get 3x32x32 ?\nThanks!<\/p>\n","answers":[{"AnswerId":"41488168","CreationDate":"2017-01-05T15:00:31.270","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>You can either do PaddingLayers (<a href=\"https:\/\/keras.io\/layers\/convolutional\/#zeropadding2d\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/convolutional\/#zeropadding2d<\/a>) and then apply convolutions to acquire a sensible output or do another Upsampling and then apply continous Convolutions with border_mode='valid' to acquire the corect output size. You can do the convolutions earlier so you do not need as many.<\/p>\n"}]}
{"QuestionId":41488276,"AnswerCount":2,"Tags":"<python><tensorflow><protocol-buffers>","CreationDate":"2017-01-05T15:06:52.070","AcceptedAnswerId":41489484.0,"OwnerUserId":1195250.0,"Title":"Cannot get a simple TFRecord reader to work","Body":"<p>I am trying to get a very simple TFRecord reader to work, but to no avail. (I can get the writer to work just fine). <\/p>\n\n<p>From <a href=\"https:\/\/github.com\/jkschin\/tf_code\/tree\/master\/TFRecords%20Example\" rel=\"nofollow noreferrer\">this github repo<\/a>, there is a  file, and it looks like so:<\/p>\n\n<pre><\/pre>\n\n<p>The problem is that when I run it, I get the following error:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/BLxrl.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BLxrl.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So, I have been banging my head on this for the last day. I am not sure what to do, or even why it does not work. It seems like a simple enough example that should have no issue. I am using TF010.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41489133","CreationDate":"2017-01-05T15:48:25.150","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p>The problem may as well be not with TFRecords, but with string_input_producer using LOCAL_VARIABLES collection when num_epochs is set. See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1045\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>The op <code>tf.initialize_all_variables()<\/code> is not initializing <strong>all<\/strong> variables, as it name suggests. As a quick fix, use the following:<\/p>\n\n<p><code>init_op = tf.group(tf.initialize_all_variables(), tf.initialize_local_variables())<\/code><\/p>\n\n<p>, but consider going to Tensorflow r0.12, where this op is deprecated in favor of <code>tf.global_variables_initializer<\/code> and <code>tf.local_variables_initializer<\/code><\/p>\n"},{"AnswerId":"41489484","CreationDate":"2017-01-05T16:03:05.210","ParentId":null,"OwnerUserId":"5703903","Title":null,"Body":"<p>In the newer version of TensorFlow:<\/p>\n\n<blockquote>\n  <p><code>tf.initialize_all_variables()<\/code> is deprecated.<\/p>\n<\/blockquote>\n\n<p>They mention that you have to use:<\/p>\n\n<blockquote>\n  <p><code>tf.global_variables_initializer()<\/code><\/p>\n<\/blockquote>\n\n<p>This doesn't solve the problem. If we look at the newer API of <code>tf.train.string_input_producer()<\/code>, it mentions that num_epochs will be created as a local variable. What is happening here is that there is nothing in the queue for it to read, hence it says requested 1 current 0. Just add this:<\/p>\n\n<blockquote>\n  <p><code>tf.local_variables_initializer()<\/code><\/p>\n<\/blockquote>\n\n<p>I have pushed the update.<\/p>\n"}]}
{"QuestionId":41488580,"AnswerCount":0,"Tags":"<deep-learning><caffe><pycaffe><matconvnet>","CreationDate":"2017-01-05T15:21:39.373","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"\"Check failed: size_in_datum == data_size\" error while creating mean image, how to solve?","Body":"<p>I am trying to generate the mean image from training data in caffe. My data is 256x256 grayscale images. I created lmdb by using  by replacing  with .<\/p>\n\n<p>I edited create_imagenet.sh as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I successfully generated my LMDB databases. But I am still getting the following error while creating .<\/p>\n\n<pre><\/pre>\n\n<p>I am sure all images have the same size and format. Does anyone have any suggestion to tackle this error?<\/p>\n\n<p>Your help is really appreciated.<\/p>\n","answers":[]}
{"QuestionId":41489238,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><conv-neural-network>","CreationDate":"2017-01-05T15:52:34.877","AcceptedAnswerId":null,"OwnerUserId":7353845.0,"Title":"Why use TensorFlow for Convolutional Neural Networks","Body":"<p>I recently took a courser by Andrew Ng on Coursera. After that I shifted to Python and used Pandas, Numpy, Sklearn to implement ML algorithms. Now while surfing I came across tensorFLow and found it pretty amazing, and implemented this <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">example<\/a> which takes MNIST data as input.\nBut I am unsure why use such as library(TensorFlow)?\nWe are not doing any parallel calculations, since the weights updated in the previous epoch are used in the next one???\nI am finding it difficult to find a reason to use such a Library?<\/p>\n","answers":[{"AnswerId":"41491866","CreationDate":"2017-01-05T18:10:05.290","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are several forms of parallelism that TensorFlow provides when training a convolutional neural network (and many other machine learning models), including:<\/p>\n\n<ol>\n<li><p><strong>Parallelism within individual operations<\/strong> (such as <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/convolution#conv2d\" rel=\"nofollow noreferrer\"><code>tf.nn.conv2d()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/matrix_math_functions#matmul\" rel=\"nofollow noreferrer\"><code>tf.matmul()<\/code><\/a>). These operations have efficient parallel implementations for multi-core CPUs and GPUs, and TensorFlow uses these implementations wherever available.<\/p><\/li>\n<li><p><strong>Parallelism between operations<\/strong>. TensorFlow uses a dataflow graph representation for your model, and where there are two nodes that aren't connected by a directed path in the dataflow graph, these may execute in parallel. For example, the Inception image recognition model has many parallel branches in its dataflow graph (see figure 3 in <a href=\"https:\/\/arxiv.org\/pdf\/1409.4842.pdf\" rel=\"nofollow noreferrer\">this paper<\/a>), and TensorFlow can exploit this to run many operations at the same time. The <a href=\"https:\/\/arxiv.org\/pdf\/1404.5997v2.pdf\" rel=\"nofollow noreferrer\">AlexNet paper<\/a> also describes how to use \"model parallelism\" to run operations in parallel on different parts of the model, and TensorFlow supports that using the same mechanism.<\/p><\/li>\n<li><p><strong>Parallelism between model replicas<\/strong>. TensorFlow is also designed for <a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">distributed execution<\/a>. One common scheme for parallel training (\"data parallelism\") involves sharding your dataset across a set of identical workers, performing the same training computation on each of those workers for different data, and sharing the model parameters between the workers.<\/p><\/li>\n<\/ol>\n\n<p>In addition, libraries like TensorFlow and Theano can perform various optimizations when they can work with the whole dataflow graph of your model. For example, they can eliminate common subexpressions, avoid recomputing constant values, and generate more efficient fused code.<\/p>\n"},{"AnswerId":"41493215","CreationDate":"2017-01-05T19:33:13.660","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>You might be able to find pre-baked models in sklearn or other libraries, but TensorFlow allows for really fast iteration of custom machine learning models. It also comes with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/\" rel=\"nofollow noreferrer\">a ton of useful functions<\/a> that you would have to (and probably shouldn't) write yourself.<\/p>\n\n<p>To me, it's less about performance (though they certainly care about performance), and more about whipping out neural networks really quickly.<\/p>\n"}]}
{"QuestionId":41489907,"AnswerCount":2,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-05T16:25:37.940","AcceptedAnswerId":null,"OwnerUserId":3827807.0,"Title":"Generative adversarial networks tanh?","Body":"<p>I was wondering, why in most of the models of GAN (in MNIST at least) I saw, the activation function (for the discriminator and the generator) was tanh ?\nIsn't ReLu more efficient ? (I always read that for predictive networks)<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41493133","CreationDate":"2017-01-05T19:28:09.873","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>Sometimes it depends on the range that you want the activations to fall into. Whenever you hear \"gates\" in ML literature, you'll probably see a sigmoid, which is between 0 and 1. In this case, maybe they want activations to fall between -1 and 1, so they use tanh. <a href=\"https:\/\/github.com\/soumith\/ganhacks\" rel=\"nofollow noreferrer\">This page<\/a> says to use tanh, but they don't give an explanation. <a href=\"https:\/\/github.com\/Newmu\/dcgan_code\" rel=\"nofollow noreferrer\">DCGAN<\/a> uses ReLUs or leaky ReLUs except for the output of the generator. Makes sense - what if half of your embedding becomes zeros? Might be better to have a smoothly varying embedding between -1 and 1.<\/p>\n\n<p>I'd love to hear someone else's input, as I'm not sure.<\/p>\n"},{"AnswerId":"42909035","CreationDate":"2017-03-20T16:32:52.840","ParentId":null,"OwnerUserId":"5337078","Title":null,"Body":"<p>From the DCGAN paper [Radford et al. <a href=\"https:\/\/arxiv.org\/pdf\/1511.06434.pdf]\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1511.06434.pdf]<\/a>...<\/p>\n\n<p>\"The ReLU activation (Nair &amp; Hinton, 2010) is used in the generator with the exception of the output\nlayer which uses the Tanh function. We observed that using a bounded activation allowed the model\nto learn more quickly to saturate and cover the color space of the training distribution.  Within the\ndiscriminator we found the leaky rectified activation (Maas et al., 2013) (Xu et al., 2015) to work\nwell, especially for higher resolution modeling. This is in contrast to the original GAN paper, which\nused the maxout activation (Goodfellow et al., 2013).\"<\/p>\n\n<p>It could be that the symmetry of tanh is an advantage here, since the network should be treating darker colours and lighter colours in a symmetric way. <\/p>\n"}]}
{"QuestionId":41490250,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-05T16:42:16.193","AcceptedAnswerId":41490871.0,"OwnerUserId":5407700.0,"Title":"Passing bool to feed dict","Body":"<p>So here is an example of using batch normalization over a 1-D input vector. Batch normalization is performed over 100 training examples . I then want to test on say just 1 example later on .<\/p>\n\n<pre><\/pre>\n\n<p>But I can't pass a  to a function expecting a normal python bool. What is the best way of going about this so I can change a bool during a session.<\/p>\n","answers":[{"AnswerId":"41490871","CreationDate":"2017-01-05T17:12:54.820","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The current implementation of the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.layers\/higher_level_ops_for_building_neural_network_layers_#batch_norm\" rel=\"nofollow noreferrer\"><code>tf.contrib.layers.batch_norm()<\/code><\/a> function is designed to accept a <code>tf.Tensor<\/code> as the <code>is_training<\/code> argument (although this fact doesn't appear to be documented), and looking at the revision history, it was added in the TensorFlow 0.10 release. If you are using an older version, please try upgrading to the latest release (currently 0.12), and your existing code should work. Among other improvements, it contains a fused implementation of batch normalization that should make a significant performance improvement.<\/p>\n"}]}
{"QuestionId":41490526,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-01-05T16:56:37.933","AcceptedAnswerId":41667083.0,"OwnerUserId":1714410.0,"Title":"What are solvers callback functions in pycaffe, and how can I use them?","Body":"<p>Looking at <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3020\" rel=\"nofollow noreferrer\">this PR<\/a>, I see that one can define  and  callbacks for  object.<\/p>\n\n<pre><\/pre>\n\n<p>What type of objects are  and ?<br>\nWhat are these callbacks for?<br>\nHow can one use them (an example would be nice...)?<\/p>\n","answers":[{"AnswerId":"41667083","CreationDate":"2017-01-15T22:34:59.023","ParentId":null,"OwnerUserId":"4221706","Title":null,"Body":"<p><strong>1. Where and how are the callbacks defined?<\/strong><\/p>\n\n<p>The callbacks are part of the Solver, and are thus defined in the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/solver.hpp\" rel=\"nofollow noreferrer\"><code>solver.hpp<\/code><\/a> file. To be exact, there is a <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/solver.hpp#L78\" rel=\"nofollow noreferrer\"><code>Callback<\/code><\/a> class, which looks like this:<\/p>\n\n<pre><code>  \/\/ Invoked at specific points during an iteration\n  class Callback {\n   protected:\n    virtual void on_start() = 0;\n    virtual void on_gradients_ready() = 0;\n\n    template &lt;typename T&gt;\n    friend class Solver;\n  };\n  const vector&lt;Callback*&gt;&amp; callbacks() const { return callbacks_; }\n  void add_callback(Callback* value) {\n    callbacks_.push_back(value);\n  }\n<\/code><\/pre>\n\n<p>and a <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/solver.hpp#L117\" rel=\"nofollow noreferrer\">protected vector<\/a> of such callbacks, which is a member of the <code>Solver<\/code> class.<\/p>\n\n<pre><code>  vector&lt;Callback*&gt; callbacks_;\n<\/code><\/pre>\n\n<p>So, this basically provides an <code>add_callback<\/code> function to the <code>Solver<\/code> class, which allows one to add an object of type <code>Callback<\/code> to a vector. This is to make sure, that each callback has two methods: <code>on_start()<\/code> and <code>on_gradients_ready()<\/code>.<\/p>\n\n<p><strong>2. Where are the callbacks called?<\/strong><\/p>\n\n<p>This happens in the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/solver.cpp\" rel=\"nofollow noreferrer\"><code>solver.cpp<\/code><\/a> file, in the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/solver.cpp#L194\" rel=\"nofollow noreferrer\"><code>step()<\/code><\/a> function, which contains the main worker loop. Here's that main loop part (with lots of things stripped out for simplicity):<\/p>\n\n<pre><code>while (iter_ &lt; stop_iter) {\n\n    for (int i = 0; i &lt; callbacks_.size(); ++i) {\n        callbacks_[i]-&gt;on_start();\n    }\n\n    \/\/ accumulate the loss and gradient\n    Dtype loss = 0;\n    for (int i = 0; i &lt; param_.iter_size(); ++i) {\n        loss += net_-&gt;ForwardBackward();\n    }\n    loss \/= param_.iter_size();\n\n    for (int i = 0; i &lt; callbacks_.size(); ++i) {\n      callbacks_[i]-&gt;on_gradients_ready();\n    }\n\n    ApplyUpdate();\n\n    ++iter_;\n}\n<\/code><\/pre>\n\n<p><strong>3. Where is this used?<\/strong><\/p>\n\n<p>This callback feature was implemented when multi-GPU support was added. The only place (that I know of), where callbacks are used, is to synchronize the solver between multiple GPUs:<\/p>\n\n<p>The <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/parallel.hpp#L85\" rel=\"nofollow noreferrer\"><code>P2PSync<\/code><\/a> class in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/parallel.hpp\" rel=\"nofollow noreferrer\"><code>parallel.hpp<\/code><\/a> inherits from the <code>Solver::Callback<\/code> class, and implements an <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/parallel.cpp#L287\" rel=\"nofollow noreferrer\"><code>on_start()<\/code><\/a> and <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/parallel.cpp#L325\" rel=\"nofollow noreferrer\"><code>on_gradients_ready()<\/code><\/a> method, which synchronize the GPUs and finally accumulate the all gradient updates.<\/p>\n\n<p><strong>4. How to use callbacks from Python?<\/strong><\/p>\n\n<p>As the pull request <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3020\" rel=\"nofollow noreferrer\">#3020<\/a> explains,<\/p>\n\n<blockquote>\n  <p><code>on_start<\/code> and <code>on_gradient<\/code> are python functions.<\/p>\n<\/blockquote>\n\n<p>so it should be straight-forward to use. A full, runnable example is shown in <a href=\"https:\/\/gist.github.com\/HBadertscher\/9c38dfebc91186887f11a4c2c80ead52#file-test_callbacks-py\" rel=\"nofollow noreferrer\">this Github Gist<\/a> I created. <\/p>\n\n<p><strong>5. How is this useful?<\/strong><\/p>\n\n<p>As the two callback functions do not take <em>any<\/em> arguments, you can't simply use them to keep track of the loss or similar things. To do that, you have to create a wrapper function around the <code>Solver<\/code> class, and call <code>add_callback<\/code> with two methods as callback functions. This allows you to access the net from within the callback, by using <code>self.solver.net<\/code>. In the following example, I use the <code>on_start<\/code> callback to load data into the net, and the <code>on_gradients_ready<\/code> callback to print the loss function.<\/p>\n\n<pre><code>class SolverWithCallback:\n    def __init__(self, solver_file):\n        self.solver = caffe.SGDSolver(solver_file)\n        self.solver.add_callback(self.load, self.loss)\n\n    def solve(self):\n        self.solver.solve()\n\n    def load(self):\n        inp = np.random.randint(0, 255)\n        self.solver.net.blobs['data'].data[...] = inp\n        self.solver.net.blobs['labels'].data[...] = 2 * inp\n\n    def loss(self):\n        print \"loss: \" + str(self.solver.net.blobs['loss'].data)\n\nif __name__=='__main__':\n    solver = SolverWithCallback('solver.prototxt')\n    solver.solve()\n<\/code><\/pre>\n"}]}
{"QuestionId":41492141,"AnswerCount":0,"Tags":"<memory><machine-learning><tensorflow><neural-network>","CreationDate":"2017-01-05T18:28:26.620","AcceptedAnswerId":null,"OwnerUserId":96603.0,"Title":"How do I learn parameters per input example in TensorFlow without using too much memory?","Body":"<p>I want to train my TensorFlow model to learn some parameters <em>separately for each input element<\/em> - in particular, I want it to learn the initial states of an LSTM network for every input in my training set. The goal is to use these initial states as an embedding.<\/p>\n\n<p>In order to loop through my input data, I use a  that goes over my input stored in  files as . If I just allocate a huge tensor for the learned initial states, my memory gets exhausted. As my input data are loaded in batches, I was wondering if there is some way to split the learned initial states up in batches, too. But when and how can I save those data to disk? And how do I make sure the same initial states come up when the same batch is read again? Also, I know that Adam keeps track of changes to all trainable variables \u2013 can I somehow tell Adam to swap out its memory, too?<\/p>\n","answers":[]}
{"QuestionId":41493858,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2017-01-05T20:13:18.247","AcceptedAnswerId":null,"OwnerUserId":6174837.0,"Title":"New to theano. Trying to add a term to a loss function to penalize negative weights","Body":"<p>To be clear, by weights I mean the entries in the matrices (Ws) of the affine transformation in a node of a neural net. <\/p>\n\n<p>I start with <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/nnet\/nnet.html#theano.tensor.nnet.nnet.categorical_crossentropy\" rel=\"nofollow noreferrer\"><\/a> as my loss function. And I want to add an additional term to penalize negative weights. \nTo this end I want to introduce a term of the form <\/p>\n\n<p><\/p>\n\n<p>Where \"\" are the weights. <\/p>\n\n<p>If I follow the source code of :<\/p>\n\n<pre><\/pre>\n\n<p>Seems like I should update  the third line (from the bottom) to read<\/p>\n\n<pre><\/pre>\n\n<p>And change the declaration of the function to be\n Where in calling for  I write <\/p>\n\n<p> <\/p>\n\n<p>with, for a start,  to be the weights coming from the first layer of my neural net. <\/p>\n\n<p>With those updates, I go on writing: <\/p>\n\n<pre><\/pre>\n\n<p>This passes the compiler and everything runs smoothly, the training of the network completes. However, for some reason the additional term \"  is ignored, it seems not to effect the loss value. <\/p>\n\n<p>I was trying to look into Theano documentation, but so far I could not figure out what might be wrong? The weighs  are shared variables, so I could not pass those as    <\/p>\n\n<p><\/p>\n\n<p>Any comments are welcome. Thanks!<\/p>\n\n<h2><strong>Solution<\/strong><\/h2>\n\n<p>Though, I didn't find why what I did, didn't work, adding the penalty term outside the 'categorical_crossentropy' as suggested in the comments did solve the problem: <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41494625,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2017-01-05T21:01:25.860","AcceptedAnswerId":null,"OwnerUserId":7381295.0,"Title":"Issues using Keras np_utils.to_categorical","Body":"<p>I'm trying to make an array of one-hot vector of integers into an array of one-hot vector that keras will be able to use to fit my model. Here's the relevant part of the code:<\/p>\n\n<pre><\/pre>\n\n<p>Below is an image showing what  and  actually are.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/j8qFx.png\" alt=\"\"><\/p>\n\n<p>I couldn't find any documentation for  that could help me.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"43314437","CreationDate":"2017-04-10T03:09:44.247","ParentId":null,"OwnerUserId":"4572274","Title":null,"Body":"<p><code>np.utils.to_categorical<\/code> is used to convert array of labeled data(from 0 to nb_classes-1) to one-hot vector. <\/p>\n\n<p>The official doc with an example.<\/p>\n\n<pre><code>In [1]: from keras.utils import np_utils # from keras import utils as np_utils\nUsing Theano backend.\n\nIn [2]: np_utils.to_categorical?\nSignature: np_utils.to_categorical(y, num_classes=None)\nDocstring:\nConvert class vector (integers from 0 to nb_classes) to binary class matrix, for use with categorical_crossentropy.\n\n# Arguments\n    y: class vector to be converted into a matrix\n    nb_classes: total number of classes\n\n# Returns\n    A binary matrix representation of the input.\nFile:      \/usr\/local\/lib\/python3.5\/dist-packages\/keras\/utils\/np_utils.py\nType:      function\n\nIn [3]: y_train = [1, 0, 3, 4, 5, 0, 2, 1]\n\nIn [4]: \"\"\" Assuming the labeled dataset has total six classes (0 to 5), y_train is the true label array \"\"\"\n\nIn [5]: np_utils.to_categorical(y_train, num_classes=6)\nOut[5]:\narray([[ 0.,  1.,  0.,  0.,  0.,  0.],\n       [ 1.,  0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  0.,  1.,  0.,  0.],\n       [ 0.,  0.,  0.,  0.,  1.,  0.],\n       [ 0.,  0.,  0.,  0.,  0.,  1.],\n       [ 1.,  0.,  0.,  0.,  0.,  0.],\n       [ 0.,  0.,  1.,  0.,  0.,  0.],\n       [ 0.,  1.,  0.,  0.,  0.,  0.]])\n<\/code><\/pre>\n"},{"AnswerId":"53430549","CreationDate":"2018-11-22T11:59:24.967","ParentId":null,"OwnerUserId":"7592832","Title":null,"Body":"<pre><code>from keras.utils.np_utils import to_categorical\n\nto_categorical(0, max_value_of_array)\n<\/code><\/pre>\n\n<p><strong>It assumes the class values were in string and you will be label encoding them, hence starting everytime from 0 to n-classes.<\/strong><\/p>\n\n<pre><code>for the same example:- consider an array of {1,2,3,4,2}\n\nThe output will be [zero value, one value, two value, three value, four value]\n\narray([[ 0.,  1.,  0., 0., 0.],\n       [ 0.,  0.,  1., 0., 0.],\n       [ 0.,  0.,  0., 1., 0.],\n       [ 0.,  0.,  0., 0., 1.],\n       [ 0.,  0.,  1., 0., 0.]],\n<\/code><\/pre>\n\n<p><strong>Let's look at another example:-<\/strong><\/p>\n\n<pre><code>Again, for an array having 3 classes, Y = {4, 8, 9, 4, 9}\n\nto_categorical(Y) will output\n\narray([[0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  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       [0.,  0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0. ],\n       [0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1. ]]\n<\/code><\/pre>\n"}]}
{"QuestionId":41495757,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2017-01-05T22:19:54.060","AcceptedAnswerId":41499257.0,"OwnerUserId":4787803.0,"Title":"Error performing lstm using BasicLSTMCell in v0.12.0","Body":"<p>I have v0.12.0 installed on my system and I'm facing some issue while running a simple LSTM. While trying to get it to work I have even reduced my code to the examples at <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/recurrent\/<\/a>, but still facing the same issue. I'm attaching snippets from my code and the error log. The \"lstm\" function takes input from a convolving function that produces latent representations (size: 1024) of a sequence of 40 frames.<\/p>\n\n<pre><\/pre>\n\n<p>The shapes of tensors as per output are:<\/p>\n\n<pre><\/pre>\n\n<p>And the snippet of the error log is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41499257","CreationDate":"2017-01-06T05:04:37.787","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>Since <code>state_is_tuple<\/code> is <code>True<\/code> by default, you need to pass in <code>lstm.zero_state(batch_size, tf.float32)<\/code> to the <code>state<\/code> variable.<\/p>\n\n<p>Replace,<\/p>\n\n<pre><code>state = initial_state if initial_state else tf.zeros([batch_size, state_size], name=\"lstm_state\")\n<\/code><\/pre>\n\n<p>with,<\/p>\n\n<pre><code>state = initial_state if initial_state else lstm.zero_state(batch_size, tf.float32)\n<\/code><\/pre>\n\n<p>Also, make sure you pass <code>LSTMStateTuple<\/code> objects to the <code>initial_state<\/code> argument.<\/p>\n"}]}
{"QuestionId":41496251,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-05T23:01:15.133","AcceptedAnswerId":null,"OwnerUserId":4660976.0,"Title":"Fitting Large Matrix Calculations into Memory when using Tensorflow","Body":"<p>I am attempting to build a model which has two phases.\nThe first takes an input image and passes it through a conv-deconv network. The resulting Tensor has entries corresponding to pixels in a desired output image (same size as the input image). <\/p>\n\n<p>To calculate the final output image I want to take the value generated at each pixel location from the first phase and use it as an additional input to a reduction function that is applied over the entire input image. This second step has no trainable variables, but it does have computation\/memory costs that grow exponentially with the size of the input (each output pixel is a function of all input pixels).<\/p>\n\n<p>I'm currently using the tf.map_fn to calculate the output image. I'm mapping the output pixel calculation function onto the results from the first phase. My desire is that tensorflow would allocate the memory to store the intermediate tensors needed for each pixel calculation and then free that memory before moving on to the next pixel calculation. But instead it seems to never free the intermediate calculations causing OOM errors.<\/p>\n\n<p>Is there someway to tell tensorflow (either explicitly or implicitly) that  it should free the memory allocated to hold the data of a Tensor that is no longer needed in the calculation?<\/p>\n","answers":[{"AnswerId":"41496415","CreationDate":"2017-01-05T23:16:45.703","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>TensorFlow deallocates memory for the tensor as soon as the tensor is no longer needed for any future calculations. You can verify this by looking at memory deallocation messages as shown in this <a href=\"https:\/\/github.com\/yaroslavvb\/notebooks\/blob\/master\/mnist-memory.ipynb\" rel=\"nofollow noreferrer\">notebook<\/a>.<\/p>\n\n<p>It's possible you are running out of memory because TensorFlow executes nodes in a memory inefficient order.<\/p>\n\n<p>As an example, consider following computation:<\/p>\n\n<pre><code>k = 2000\na = tf.random_uniform(shape=(k,k))\nfor i in range(n):\n    a = tf.matmul(a, tf.random_uniform(shape=(k,k)))\n<\/code><\/pre>\n\n<p>The order in which it is evaluated can be shown below<\/p>\n\n<p><img src=\"https:\/\/cloud.githubusercontent.com\/assets\/23068\/21414406\/fe679a46-c7b3-11e6-9e7b-3d38403f5ea3.png\"><\/p>\n\n<p>All the circles (<code>tf.random_uniform<\/code>) nodes are evaluated first, followed by squares (<code>tf.matmul<\/code>). This has <code>O(n)<\/code> memory requirement compared to <code>O(1)<\/code> for the optimal order.<\/p>\n\n<p>You can use control dependencies to force a specific execution order, ie, using helper function as below:<\/p>\n\n<pre><code>import tensorflow.contrib.graph_editor as ge\ndef run_after(a_tensor, b_tensor):\n    \"\"\"Force a to run after b\"\"\"\n    ge.reroute.add_control_inputs(a_tensor.op, [b_tensor.op])\n<\/code><\/pre>\n"}]}
{"QuestionId":41496842,"AnswerCount":1,"Tags":"<theano><conv-neural-network><lasagne><nolearn>","CreationDate":"2017-01-06T00:02:06.510","AcceptedAnswerId":41606347.0,"OwnerUserId":2280645.0,"Title":"Convolutional Neural Network - Visualizing weights","Body":"<p><strong>Main Problem<\/strong><\/p>\n\n<p>I cannot understand the Plot of the weights of a specific layer.\nI used a method from no-learn : <\/p>\n\n<p>Im using <strong><em>lasagne<\/em><\/strong> as my neural-network library.<\/p>\n\n<p>The plot comes out fine, but I dont know how i should interpret it. <\/p>\n\n<p><strong>Neural Network Structure<\/strong><\/p>\n\n<p>The structure im using :<\/p>\n\n<pre><\/pre>\n\n<p>Here are the weights of the first 3 Layers :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/aBsiCm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aBsiCm.png\" alt=\"&quot;Image one &quot;\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Oxjnvm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Oxjnvm.png\" alt=\"&quot;Image two &quot;\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qIfoAm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qIfoAm.png\" alt=\"&quot;Image Three &quot;\"><\/a><\/p>\n\n<p>** About the Images **<\/p>\n\n<p>So for me, they look random and i cannot interpret them!<\/p>\n\n<p>However, on Cs231, it says the following :<\/p>\n\n<blockquote>\n  <p>Conv\/FC Filters. The second common strategy is to visualize the\n  weights. These are usually most interpretable on the first CONV layer\n  which is looking directly at the raw pixel data, but it is possible to\n  also show the filter weights deeper in the network. The weights are\n  useful to visualize because well-trained networks usually display nice\n  and smooth filters without any noisy patterns. Noisy patterns can be\n  an indicator of a network that hasn\u2019t been trained for long enough, or\n  possibly a very low regularization strength that may have led to\n  overfitting\n  <a href=\"http:\/\/cs231n.github.io\/understanding-cnn\/\" rel=\"nofollow noreferrer\">http:\/\/cs231n.github.io\/understanding-cnn\/<\/a><\/p>\n<\/blockquote>\n\n<p>Then why mine are random?<\/p>\n\n<p>The structure is trained and performs well for its task.<\/p>\n\n<p><strong>References<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41606347","CreationDate":"2017-01-12T06:21:22.207","ParentId":null,"OwnerUserId":"3737792","Title":null,"Body":"<p>Normally when you visualize the weights you want to check 2 things:<\/p>\n\n<ul>\n<li>That they are smooth and cover a wide range of values, i.e. it's not a bunch of 1's and 0's. That would mean the non-linearity is being saturated.<\/li>\n<li>That they have some kind of structure. Normally you tend to see oriented edges although this is more difficult to see when you have small filters like 3x3.<\/li>\n<\/ul>\n\n<p>That being said, your weights do not appear to be saturated, but they indeed seem to be too random. \nDuring training, did the network converge correctly?\nI am also surprised at how big your filters are (30x30). Not sure what you are trying to accomplish with that.<\/p>\n"}]}
{"QuestionId":41497710,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-06T01:52:44.150","AcceptedAnswerId":null,"OwnerUserId":1191087.0,"Title":"TensorFlow - Setting some variables as `Trainable=False` vs passing variables to the optimizer?","Body":"<p>To only train specific variables in TensorFlow, you can either pass a specific list to the optimizer to train, or you can set some variables as untrainable. Are there any benefits to choosing one method over the other (aside from ease of including\/excluding variables)?<\/p>\n\n<p>I could see it being possible that if a variable was set to , then the graph might be built differently (i.e. without gradient paths). I could see this potentially having an effect on efficiency. Is this the case? And, if it is, does it make any (noticeable) difference?<\/p>\n\n<p>Also, are there any other \"gotchas\" to look out for when using one method over the other? For example, when passing variables to the optimizer, are there common global variables or such that might easily accidentally be missed? Any other things to note?<\/p>\n\n<p>Thank you for your time!<\/p>\n","answers":[]}
{"QuestionId":41497820,"AnswerCount":1,"Tags":"<ubuntu><tensorflow><bazel>","CreationDate":"2017-01-06T02:06:29.060","AcceptedAnswerId":null,"OwnerUserId":7381990.0,"Title":"everytime my Ubuntu is rebooted, Tensorflow with GPU support needs to be rebuild","Body":"<p>When I build tensorflow with gpu support(in configure I set cuda support to yes), \nbazel build -c opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package\neverything works well.<\/p>\n\n<p>but after I reboot the system, and try to build tensorflow using above command, it gives me following error:<\/p>\n\n<h1>ERROR<\/h1>\n\n<pre>\/home\/wolfson\/.cache\/bazel\/_bazel_wolfson\/41eaf6c788f09c81cffb135517d04fa2\/external\/local_config_cuda\/crosstool\/BUILD:4:1: Traceback (most recent call last):\n    File \"\/home\/wolfson\/.cache\/bazel\/_bazel_wolfson\/41eaf6c788f09c81cffb135517d04fa2\/external\/local_config_cuda\/crosstool\/BUILD\", line 4\n        error_gpu_disabled()\n    File \"\/home\/wolfson\/.cache\/bazel\/_bazel_wolfson\/41eaf6c788f09c81cffb135517d04fa2\/external\/local_config_cuda\/crosstool\/error_gpu_disabled.bzl\", line 3, in error_gpu_disabled\n        fail(\"ERROR: Building with --config=c...\")\nERROR: Building with --config=cuda but TensorFlow is not configured to build with GPU support. Please re-run .\/configure and enter 'Y' at the prompt to build with GPU support.\nERROR: no such target '@local_config_cuda\/\/crosstool:toolchain': target 'toolchain' not declared in package 'crosstool' defined by \/home\/wolfson\/.cache\/bazel\/_bazel_wolfson\/41eaf6c788f09c81cffb135517d04fa2\/external\/local_config_cuda\/crosstool\/BUILD.<\/pre>\n","answers":[{"AnswerId":"41502895","CreationDate":"2017-01-06T09:44:44.223","ParentId":null,"OwnerUserId":"4800090","Title":null,"Body":"<p>The bazel build checks if there is an environment variable TF_NEED_CUDA set to 1, otherwise it will complain.\nThis variable is exported in the configure script but is then missing after rebooting.<\/p>\n\n<p>Maybe executing <code>export TF_NEED_CUDA=1<\/code> before building will solve your problem, but there might be other environment variables or files that are missing after a reboot.<\/p>\n"}]}
{"QuestionId":41497982,"AnswerCount":1,"Tags":"<python><c++><tensorflow>","CreationDate":"2017-01-06T02:26:02.567","AcceptedAnswerId":42122446.0,"OwnerUserId":1524166.0,"Title":"Tensorflow model works in Python but not in C++","Body":"<p>For a little background, my main goal is to use Tensorflow's C++ API to classify an image and time it on different systems.<\/p>\n\n<p>I have used <a href=\"https:\/\/github.com\/ry\/tensorflow-vgg16\" rel=\"nofollow noreferrer\" title=\"Ry&#39;s model converter\">Ry's model converter<\/a> to convert his Caffe model to Tensorflow, and it produces the vgg16.tfmodel file, which appears to be a .pb file, once you open it up. <\/p>\n\n<p>Using Ry's tf_forward.py to run this resulting file seems to work perfectly, classifying cats, dogs, etc. However, when I modify the label_image example (tensorflow\/examples\/label_image\/) to use my new vgg16.pb file, something appears to go wrong. <\/p>\n\n<p>Here's the output of classifying the picture of the cat from the tensorflow-vgg16 example:<\/p>\n\n<pre><\/pre>\n\n<p>and here are the settings that I use in my modification of label_image\/main.cc:<\/p>\n\n<pre><\/pre>\n\n<p>I've also tried freeze_graph.py to get the model and weights, but I could never get it to work properly. After tracing everything back a bit,I ran into protobuf code and got a bit stuck. Any help would be really appreciated :)<\/p>\n","answers":[{"AnswerId":"42122446","CreationDate":"2017-02-08T19:56:36.600","ParentId":null,"OwnerUserId":"1524166","Title":null,"Body":"<p>For anyone looking at this in the future, this problem was caused by using the wrong input layer name.<\/p>\n"}]}
{"QuestionId":41498883,"AnswerCount":1,"Tags":"<python><session><tensorflow><distributed-computing>","CreationDate":"2017-01-06T04:21:12.057","AcceptedAnswerId":41501486.0,"OwnerUserId":6527887.0,"Title":"TypeError: Variable to save is not a Variable","Body":"<p>I have the following code to do some simple arithmetic calculations in distributed tensorflow. A minimal reproducible example would be:-<\/p>\n\n<pre><\/pre>\n\n<p>The following is the error:-<\/p>\n\n<pre><\/pre>\n\n<p>Please help me figure out the correct way to use this function.<\/p>\n","answers":[{"AnswerId":"41501486","CreationDate":"2017-01-06T08:06:37.923","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>When you simply write <code>x + 300<\/code>, you are not creating a <code>tf.Variable<\/code>. You need to explicitly use <code>tf.get_variable()<\/code> or <code>tf.Variable()<\/code> to create a variable which can be saved.<\/p>\n\n<pre><code>y = tf.Variable(x + 300)\n<\/code><\/pre>\n"}]}
{"QuestionId":41499321,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><tensorboard>","CreationDate":"2017-01-06T05:13:08.430","AcceptedAnswerId":null,"OwnerUserId":5082661.0,"Title":"TensorFlow Combining Dense Layer with LSTM Cell","Body":"<p>How do i combine a Tensorflow Dense Layer, which is then followed by a LSTM! <\/p>\n\n<p>Given a sequence of variable length, i want to backprop through both the layers, since i will be using this for RL.<\/p>\n\n<p>How do i format my input sequence\/ define my layers to be consistent with size requirements?<\/p>\n","answers":[]}
{"QuestionId":41499509,"AnswerCount":1,"Tags":"<deep-learning><caffe><convolution><pycaffe><matcaffe>","CreationDate":"2017-01-06T05:31:06.177","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"Getting error on [base_conv_layer.cpp:122] Check failed: channels_ % group_ == 0 (1 vs. 0) , how to solve it?","Body":"<p>When I am trying to train  for <a href=\"https:\/\/github.com\/yawaterss\/caffe-pascalcontext-fcn32s\" rel=\"nofollow noreferrer\">semantic segmentation<\/a> on my own data, I am getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>I have not included the creation of previous layers. But it seems  creates previous layers successfully, and it reaches to the , the error comes. I changed  as follows:<\/p>\n\n<pre><\/pre>\n\n<p>and I changed the number of outputs from 60 to 5 (based on the number of classes in my data):  <\/p>\n\n<p>Can someone suggest any solution or idea about this? What I have set\/changed wrongly? What\/where is my mistake?<\/p>\n\n<p>Your help is appreciated.<\/p>\n","answers":[{"AnswerId":"43433506","CreationDate":"2017-04-16T03:49:56.193","ParentId":null,"OwnerUserId":"7873305","Title":null,"Body":"<p>Check failed: channels_ % group_ == 0 (1 vs. 0) \nThis line really matters! You should check your num_output and group and find out if channels_ % group_ == 0.<\/p>\n"}]}
{"QuestionId":41500065,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><theano><softmax>","CreationDate":"2017-01-06T06:19:10.087","AcceptedAnswerId":41520734.0,"OwnerUserId":6333771.0,"Title":"RBM: Deriving the Replicated Softmax Model (RSM)","Body":"<p>I am trying to derive the conditional distribution of the visible variables, <a href=\"https:\/\/i.stack.imgur.com\/dj9T6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dj9T6.png\" alt=\"enter image description here\"><\/a>, for the Replicated Softmax Model (RSM) or equivalently, the Restricted Boltzmann Machine (RBM) for word counts, according to the paper: \"Replicated Softmax: an Undirected Topic Model\" by Salakhutdinov and Hinton. <br><\/p>\n\n<p>Paper can be found at: <a href=\"http:\/\/citeseerx.ist.psu.edu\/viewdoc\/download;jsessionid=B04C8D67D381B8106FF6FA4203A86264?doi=10.1.1.164.71&amp;rep=rep1&amp;type=pdf\" rel=\"nofollow noreferrer\">http:\/\/citeseerx.ist.psu.edu\/viewdoc\/download;jsessionid=B04C8D67D381B8106FF6FA4203A86264?doi=10.1.1.164.71&amp;rep=rep1&amp;type=pdf<\/a> <br><\/p>\n\n<p>However, despite all efforts, I've been unable to get how the conditional can turn out to be a softmax distribtution: <br>\n<a href=\"https:\/\/i.stack.imgur.com\/ZS32L.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZS32L.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Also, I'm confused if <a href=\"https:\/\/i.stack.imgur.com\/rP4AE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rP4AE.png\" alt=\"enter image description here\"><\/a> is a 3D matrix and <a href=\"https:\/\/i.stack.imgur.com\/6YpCT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6YpCT.png\" alt=\"enter image description here\"><\/a> a 2D matrix or is it instead a 2D matrix and vector respectively. I believe it is the latter. Hoping someone can demonstrate the derivations.<\/p>\n\n<p>I am looking to implement the RSM to do topic modelling in python's theano. I am aware that there are codes out there but I prefer to understand the derivation myself so that I can extend or optimize the codes without the risk of breaking the model.<\/p>\n\n<hr>\n\n<p>p.s. apologies, this is a repost of <a href=\"https:\/\/math.stackexchange.com\/questions\/2085616\/rbm-deriving-the-replicated-softmax-model-rsm\">https:\/\/math.stackexchange.com\/questions\/2085616\/rbm-deriving-the-replicated-softmax-model-rsm<\/a> but i did so as aren't as many mathstackexchange users.<\/p>\n","answers":[{"AnswerId":"41520734","CreationDate":"2017-01-07T11:05:02.127","ParentId":null,"OwnerUserId":"6333771","Title":null,"Body":"<p>After sometime I found out where I misunderstood things and managed to derive the equations. Please refer to math.stackexchange:<br>\n<a href=\"https:\/\/math.stackexchange.com\/questions\/2085616\/rbm-deriving-the-replicated-softmax-model-rsm\/2087272#2087272\">https:\/\/math.stackexchange.com\/questions\/2085616\/rbm-deriving-the-replicated-softmax-model-rsm\/2087272#2087272<\/a><\/p>\n"}]}
{"QuestionId":41500466,"AnswerCount":1,"Tags":"<timeout><tensorflow><grpc><serving>","CreationDate":"2017-01-06T06:50:03.030","AcceptedAnswerId":41513403.0,"OwnerUserId":2502090.0,"Title":"TensorFlow Server close the connection within client timeout","Body":"<p>We have used TensorFlow Serving to load the model and implement the Java gRPC client.<\/p>\n\n<p>Normal it works for small data. But if we request with larger batch size and data is almost 1~2M, the server closes the connection and throw the internal error quickly.<\/p>\n\n<p>We have also open an issue to track this in <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/284\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/issues\/284<\/a>.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41513403","CreationDate":"2017-01-06T19:58:40.387","ParentId":null,"OwnerUserId":"4690866","Title":null,"Body":"<p>As can be seen in the <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/284\" rel=\"nofollow noreferrer\">above issue<\/a>, this was caused by the message exceeding the default maximum message size of 4 MiB. The receiver of larger messages needs to explicitly permit larger sizes, or sender send smaller messages.<\/p>\n\n<p>gRPC is fine with larger messages (even 100s MBs), but applications frequently aren't. The max message size is in place to permit \"large\" messages only in applications that are prepared to accept them.<\/p>\n"}]}
{"QuestionId":41500768,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-06T07:13:31.293","AcceptedAnswerId":41503526.0,"OwnerUserId":482601.0,"Title":"Which operations support automatic gradients?","Body":"<p>I have a fairly complex quantisation layer in my network which contains, among other operations,  and  ops. I noticed that my network did not train well. Looking at some debug output, I saw that the fully connected layer before this quantisation layer got zero gradients for its weights (I used  to determine this). Does this mean that what ever is before the quantisation layer does not update in training?<\/p>\n\n<p>In general: How do I figure out which operations let gradients pass through and which do not? For instance, do the above mentionied  and  let gradients pass through`<\/p>\n","answers":[{"AnswerId":"41503526","CreationDate":"2017-01-06T10:19:57.193","ParentId":null,"OwnerUserId":"4800090","Title":null,"Body":"<p>If there is an operation without gradients in your model you will get an error:<\/p>\n\n<pre><code>LookupError: No gradient defined for operation [...]\n<\/code><\/pre>\n\n<p>So your problem seems to be somewhere else, maybe you have a multiplication by zero somewhere which kills the gradients. There is not enough information in your question to find the real reason of your problem.<\/p>\n\n<p>Edit:\nI didn't directly answer the question which operations support automatic gradients.<\/p>\n\n<p>It is not listed in the documentation and I think you can only see it by checking the source code or using the operation and see if you get the mentioned error when you try to optimize the model.<\/p>\n\n<p>For <code>tf.tile<\/code> and <code>tf.expand_dims<\/code> there are gradients defined.<\/p>\n"}]}
{"QuestionId":41501959,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2017-01-06T08:46:53.173","AcceptedAnswerId":null,"OwnerUserId":5019377.0,"Title":"Why is the global Step not incrementing in custom Tensorflow model function","Body":"<p>I've been trying to update <a href=\"https:\/\/github.com\/mouradmourafiq\/tensorflow-lstm-regression\" rel=\"nofollow noreferrer\">this tutorial<\/a> to the latest version of tensorflow ie 0.12 and I've hit a snag in the custom model function definition to train the lstm model.  <\/p>\n\n<pre><\/pre>\n\n<p>I've tried both specifying the global step and not specifying it - and I've ended up with the same result - the step remains at 0 with loss being optimized and this continues until I stop the entire script. The code below is what I use to create an Estimator and try to fit the model.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41568676","CreationDate":"2017-01-10T12:17:20.003","ParentId":null,"OwnerUserId":"5019377","Title":null,"Body":"<p>I realized that the adamOptimizer was not to be used on it's own - but with the neural network that I had created. The following edit on my code - helped me achieve what I wanted to :<\/p>\n\n<pre><code># create model and features\nstacked_lstm = tf.nn.rnn_cell.MultiRNNCell(\n    lstm_cells(params['rnn_layers']),\n    state_is_tuple=True)\nfeatures = tf.cast(features, tf.float32)\ntargets = tf.cast(targets, tf.float32)\nfeatures = tf.unpack(features, axis=1, num=params['time_steps'])\noutput, layers = tf.nn.rnn(stacked_lstm, features, dtype=dtypes.float32)\noutput = dnn_layers(output[-1], params['dense_layers'])\n\n# Define Loss\n(predictions, loss) = learn.models.linear_regression(output, targets)\n\n# train_op\ntrain_op = tf.contrib.layers.optimize_loss(\n    loss=loss,\n    global_step=contrib.framework.get_global_step(),\n    learning_rate=params['learning_rate'],\n    optimizer='Adagrad'\n)\n<\/code><\/pre>\n\n<p>Here I created the neural Network layers first and then tried to find the optimization on that layer - this is what seemed to have caused me problems.\nThe tensorflow documentation helped me out quite a bit - hence I was able to run the code on tensorflow - v0.12.<\/p>\n"}]}
{"QuestionId":41503404,"AnswerCount":2,"Tags":"<python><c++><caffe>","CreationDate":"2017-01-06T10:13:32.223","AcceptedAnswerId":41513933.0,"OwnerUserId":5258060.0,"Title":"How to convert .npy file into .binaryproto?","Body":"<p>I have created a mean image file using python and saved it into numpy file. I would like to know how I could convert this .npy file into .binaryproto file. I am using this file to train using the GoogLeNet.<\/p>\n","answers":[{"AnswerId":"46088716","CreationDate":"2017-09-07T05:49:12.670","ParentId":null,"OwnerUserId":"2324271","Title":null,"Body":"<p>Here's an improved version of @Kev1n91's code.<\/p>\n\n<pre><code>import caffe\nimport numpy as np\n\nmean_npy = np.load('mean.npy') # Input numpy array\nblob = caffe.io.array_to_blobproto(mean_npy)\nmean_binproto = 'mean.binaryproto' # Output binaryproto file\nwith open(mean_binproto, 'wb') as f :\n    f.write( blob.SerializeToString())\n<\/code><\/pre>\n"},{"AnswerId":"41513933","CreationDate":"2017-01-06T20:36:09.543","ParentId":null,"OwnerUserId":"6786718","Title":null,"Body":"<p>You can simply use numpy to creat the .binaryproto and the given caffe io functions<\/p>\n\n<pre><code>import caffe\n#avg_img is your numpy array with the average data \nblob = caffe.io.array_to_blobproto( avg_img)\nwith open( mean.binaryproto, 'wb' ) as f :\n    f.write( blob.SerializeToString())\n<\/code><\/pre>\n"}]}
{"QuestionId":41504125,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2017-01-06T10:54:18.530","AcceptedAnswerId":null,"OwnerUserId":7378465.0,"Title":"How do I get the value of weight and bias from auto-encoder program","Body":"<p>I ran <a href=\"https:\/\/gist.github.com\/tomokishii\/7ddde510edb1c4273438ba0663b26fc6#file-mnist_ae1-py\" rel=\"nofollow noreferrer\">mnist_ae1.py<\/a>(very simple model of auto-encoder), and want to get value of  and .\nSo, I added some process as below.<\/p>\n\n<pre><\/pre>\n\n<p>However, the following error was thrown.<\/p>\n\n<pre><\/pre>\n\n<p>How do I get and save the value of  and ?<\/p>\n","answers":[{"AnswerId":"42683192","CreationDate":"2017-03-08T22:33:24.907","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The second argument to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation#run\" rel=\"nofollow noreferrer\">Operation.run<\/a> which you call in the <code>train_step.run<\/code> line is the session. By passing a dictionary there you confuse the runtime into treating that dictionary as a session. Try doing <code>feed_dict={x...}<\/code> instead.<\/p>\n"}]}
{"QuestionId":41504815,"AnswerCount":2,"Tags":"<python><binary><deep-learning><caffe><pycaffe>","CreationDate":"2017-01-06T11:33:00.557","AcceptedAnswerId":41512898.0,"OwnerUserId":6786718.0,"Title":"How to create mean.binaryproto for caffe out of a text file for data?","Body":"<p>I am using two text files where each has   paths to my validation \/ training images. <\/p>\n\n<p>I now want to create a mean.binaryproto out of these images to give into my input layer. However, I only found examples where this is done using a leveldb input layer. \nI can create my own mean image easily using a python script, but I have no idea how to continue after this, so to say how to write the image as a binaryproto at the end of my script. Are there any pointers?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41513314","CreationDate":"2017-01-06T19:52:36.517","ParentId":null,"OwnerUserId":"5635517","Title":null,"Body":"<p>You can first convert your images to generate an lmdb database using the <em>convert_imageset<\/em> tool. Then, using that and the <em>compute_image_mean<\/em> tool you can generate a mean.binaryproto file.<\/p>\n\n<p>See my answer to: <a href=\"https:\/\/stackoverflow.com\/questions\/41503404\/how-to-convert-npy-file-into-binaryproto\/41513102#41513102\">How to convert .npy file into .binaryproto?<\/a> <\/p>\n"},{"AnswerId":"41512898","CreationDate":"2017-01-06T19:27:35.720","ParentId":null,"OwnerUserId":"6786718","Title":null,"Body":"<p>Since the Average Picture is given in a numpy array the caffe function can be used to write as .binaryproto<\/p>\n\n<pre><code>import caffe\n\nblob = caffe.io.array_to_blobproto( avg_img)\nwith open( mean.binaryproto, 'wb' ) as f :\n    f.write( blob.SerializeToString())\n<\/code><\/pre>\n"}]}
{"QuestionId":41505746,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-06T12:27:50.553","AcceptedAnswerId":41507765.0,"OwnerUserId":853462.0,"Title":"What is the use of tf.select","Body":"<p>(<strong>edited<\/strong> w.r.t. @quirk's answer)<\/p>\n\n<p>I was reading some tensorflow-code online and saw this statements:<\/p>\n\n<pre><\/pre>\n\n<p>source: <a href=\"https:\/\/github.com\/Raverss\/tensorflow-RLSA-NMS\/blob\/master\/source.py#L31\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Raverss\/tensorflow-RLSA-NMS\/blob\/master\/source.py#L31<\/a><\/p>\n\n<p> Is a tensor with just 's,  also of the same size with 's and input is some heatmap(\/tensor) of the same size (all of type ).<\/p>\n\n<p>The code snippet seems reasonably advanced for me to assume that the authors would have just used  if there was no specific reason for the  expression. Especially since this would have eliminated the need for the variables  and , and would save memory, as they are just expensively redundant ways of storing  and .<\/p>\n\n<p>Is the aforementioned  expression equivalent to ? If not, why not?<\/p>\n\n<p><strong>Note:<\/strong> I usually use Keras, and am sorry if I'm touching something very trivial here.<\/p>\n","answers":[{"AnswerId":"41508451","CreationDate":"2017-01-06T15:06:24.473","ParentId":null,"OwnerUserId":"3572366","Title":null,"Body":"<p>Best way you can understand it is by trying it out yourself:<\/p>\n\n<pre><code>In [86]: s = tf.InteractiveSession()\n\nIn [87]: inputs = tf.random_uniform([10], 0., 1.)\n\nIn [88]: positives = tf.ones([10])\n\nIn [89]: negatives = tf.zeros([10])    \n\nIn [90]: s.run([inputs, tf.select(inputs &gt; .5, positives, negatives)])\nOut[90]: \n[array([ 0.13187623,  0.77344072,  0.29853749,  0.29245567,  0.53489852,\n         0.34861541,  0.15090156,  0.40595055,  0.34910154,  0.24349082], dtype=float32),\n array([ 0.,  1.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  0.], dtype=float32)]\n<\/code><\/pre>\n\n<p>For every value > <code>0.5<\/code> in tensor <code>inputs<\/code> you'll get a <code>1.<\/code> at the same index, otherwise the value is <code>0.<\/code>.<\/p>\n\n<p>The result of <code>inputs &gt; .5<\/code> is a tensor of booleans (<code>True<\/code> for values that meet the condition, <code>False<\/code> otherwise).<\/p>\n\n<pre><code>In [92]: s.run(inputs &gt; .5)\nOut[92]: array([ True, False,  True,  True,  True,  True,  True,  True, False,  True], dtype=bool)\n<\/code><\/pre>\n"},{"AnswerId":"41507765","CreationDate":"2017-01-06T14:25:47.623","ParentId":null,"OwnerUserId":"2844164","Title":null,"Body":"<p>Umm, RTD(Read the docs)! <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/comparison_operators#select\" rel=\"nofollow noreferrer\">tf.select<\/a> selects elements from <code>positive<\/code> or <code>negative<\/code> tensors based on the <strong>boolness<\/strong> of the elements in the <code>condition<\/code> tensor. <\/p>\n\n<blockquote>\n  <p><code>tf.select(condition, t, e, name=None)<\/code><br>\n  <strong>Selects elements from t or e, depending on condition.<\/strong><br>\n  The t, and e tensors must all have the same shape, and the output will also have that shape.<\/p>\n<\/blockquote>\n\n<p>(from the official docs.)<\/p>\n\n<p>So in your case:<\/p>\n\n<p><code>threshold = tf.select(input &gt; RLSA_THRESHOLD, positive, negative)<\/code><\/p>\n\n<p><code>input &gt; RLSA_THRESHOLD<\/code> will be a tensor of <code>bool<\/code> or logical values (<code>0<\/code> or <code>1<\/code> symbolically), which will help choose a value from either the <code>positive<\/code> vector or the <code>negative<\/code> vector.<\/p>\n\n<p>For example, say you have a <code>RLSA_THRESHOLD<\/code> of 0.5 and your <code>input<\/code> vector is a 4-dimensional vector of real continuous values ranging from 0 to 1. Your <code>positive<\/code> and <code>negative<\/code> vectors are essentially <code>[1, 1, 1, 1]<\/code> and <code>[0, 0, 0, 0]<\/code>, respectively. <code>input<\/code> is <code>[0.8, 0.2, 0.5, 0.6]<\/code>.<\/p>\n\n<p><code>threshold<\/code> will be <code>[1, 0, 0, 1]<\/code>.<\/p>\n\n<p><strong>NOTE:<\/strong> <code>positive<\/code> and <code>negative<\/code> could be any kind of tensor as long as the dimensions agree with the <code>condition<\/code> tensor. Had <code>positive<\/code> and <code>negative<\/code> been, say, <code>[2, 4, 6, 8]<\/code> and <code>[1, 3, 5, 7]<\/code> respectively, your <code>threshold<\/code> would have been <code>[2, 3, 5, 8]<\/code>.<\/p>\n\n<hr>\n\n<blockquote>\n  <p>The code snippet seems reasonably advanced for me to assume that the authors would have just used <code>input &gt; RLSA_THRESHOLD<\/code> if there was no specific reason for the <code>tf.select<\/code>.<\/p>\n<\/blockquote>\n\n<p>There is a very good reason for that. <code>input &gt; RLSA_THRESHOLD<\/code> would simply return a tensor of logical (boolean) values. Logical values <strong>do not<\/strong> mix well with numerical values. You cannot use them for any realistic numerical computation. Had the <code>positive<\/code> and\/or <code>negative<\/code> tensors been real valued, you might have required your <code>threshold<\/code> tensor to also have real values, in case you planned to use them further along.<\/p>\n\n<hr>\n\n<blockquote>\n  <p>Is the <code>tf.select<\/code> equivalent to <code>input &gt; RLSA_THRESHOLD<\/code>? If not, why not?<\/p>\n<\/blockquote>\n\n<p>No they are not. One is a function, the other is a tensor. <\/p>\n\n<p>I am going to give you the benefit of doubt and assume you meant to ask:<\/p>\n\n<blockquote>\n  <p>Is the <code>threshold<\/code> equivalent to <code>input &gt; RLSA_THRESHOLD<\/code>? If not, why not?<\/p>\n<\/blockquote>\n\n<p>No they are not. As explained above, <code>input &gt; RLSA_THRESHOLD<\/code> is a logical  tensor with a data type of <code>bool<\/code>. <code>threshold<\/code>, on the other hand, is a tensor with the same data type as <code>positive<\/code> and <code>negative<\/code>.<\/p>\n\n<p><strong>NOTE:<\/strong> You can always cast your logical tensors to numerical (or any other supported data type) tensors using any of the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/casting\" rel=\"nofollow noreferrer\">casting<\/a> methods available in <a href=\"\/questions\/tagged\/tensorflow\" class=\"post-tag\" title=\"show questions tagged &#39;tensorflow&#39;\" rel=\"tag\">tensorflow<\/a>.<\/p>\n"}]}
{"QuestionId":41505926,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-06T12:37:15.627","AcceptedAnswerId":41523137.0,"OwnerUserId":4875916.0,"Title":"Batched 4D tensor Tensorflow indexing","Body":"<p><strong>Given<\/strong><\/p>\n\n<ul>\n<li>: 4D tensor of shape <\/li>\n<li>: 3D tensor of shape <\/li>\n<li>: 3D tensor of shape <\/li>\n<\/ul>\n\n<p><strong>Goal<\/strong><\/p>\n\n<p>How can I index into  using the  and  coordinates to obtain a 4D tensor of shape . That is, I want to obtain for each batch, and for each pair  a tensor of shape . <\/p>\n\n<p>In numpy, this would be achieved using  for example but I can't seem to make it work in tensorflow.<\/p>\n\n<p><strong>My attempt so far<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>which is returning an incorrect tensor of shape .<\/p>\n","answers":[{"AnswerId":"41523137","CreationDate":"2017-01-07T15:38:59.373","ParentId":null,"OwnerUserId":"681714","Title":null,"Body":"<p>It should be possible to do it by flattening the tensor as you've done, but the batch dimension has to be taken into account in the index calculation.\nIn order to do this, you'll have to make an additional dummy batch index tensor with the same shape as <code>x<\/code> and <code>y<\/code> that always contains the index of the current batch.\nThis is basically the <code>np.arange(B)<\/code> from your numpy example, which is missing from your TensorFlow code.<\/p>\n\n<p>You can also simplify things a bit by using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#gather_nd\" rel=\"nofollow noreferrer\"><code>tf.gather_nd<\/code><\/a>, which does the index calculations for you.<\/p>\n\n<p>Here's an example:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\n# Example tensors\nM = np.random.uniform(size=(3, 4, 5, 6))\nx = np.random.randint(0, 5, size=(3, 4, 5))\ny = np.random.randint(0, 4, size=(3, 4, 5))\n\ndef get_pixel_value(img, x, y):\n    \"\"\"\n    Utility function that composes a new image, with pixels taken\n    from the coordinates given in x and y.\n    The shapes of x and y have to match.\n    The batch order is preserved.\n    \"\"\"\n\n    # We assume that x and y have the same shape.\n    shape = tf.shape(x)\n    batch_size = shape[0]\n    height = shape[1]\n    width = shape[2]\n\n    # Create a tensor that indexes into the same batch.\n    # This is needed for gather_nd to work.\n    batch_idx = tf.range(0, batch_size)\n    batch_idx = tf.reshape(batch_idx, (batch_size, 1, 1))\n    b = tf.tile(batch_idx, (1, height, width))\n\n    indices = tf.pack([b, y, x], 3)\n    return tf.gather_nd(img, indices)\n\ns = tf.Session()\nprint(s.run(get_pixel_value(M, x, y)).shape)\n# Should print (3, 4, 5, 6).\n# We've composed a new image of the same size from randomly picked x and y\n# coordinates of each original image.\n<\/code><\/pre>\n"}]}
{"QuestionId":41506137,"AnswerCount":1,"Tags":"<optimization><lua><neural-network><torch>","CreationDate":"2017-01-06T12:49:34.920","AcceptedAnswerId":41515774.0,"OwnerUserId":5776252.0,"Title":"Torch: How to apply the torch built-in optimizers to tandem models?","Body":"<p>Recently, I start learning torch. But, the following question really stumps me. A seq2seq demo is available <a href=\"https:\/\/github.com\/Element-Research\/rnn\/blob\/master\/examples\/encoder-decoder-coupling.lua\" rel=\"nofollow noreferrer\">here<\/a>, and the definition of model is showed as follow:<\/p>\n\n<pre><\/pre>\n\n<p>In the original version, the parameters of this model is updated in the following way:<\/p>\n\n<pre><\/pre>\n\n<p>However, I really wonder whether I can use the built-in optimizers in <em>optim<\/em> to train the model above. So, I tried the following method:<\/p>\n\n<pre><\/pre>\n\n<p>But, it didn't work and I got an error:<\/p>\n\n<pre><\/pre>\n\n<p>All in all, how to apply the built-in optimizers in <em>optim<\/em> to update the parameters in more than one models? Do I have to overwrite the <em>adadelta.lua<\/em>?<\/p>\n","answers":[{"AnswerId":"41515774","CreationDate":"2017-01-06T23:03:21.347","ParentId":null,"OwnerUserId":"49985","Title":null,"Body":"<p>It is hard to say exactly what the error is without having access to your complete script. In any case, I have suggestions regarding the above:<\/p>\n\n<ol>\n<li>avoid creating tensors in feval (or any other inner loop). Use buffers that you can resize and reuse to avoid memory allocations.<\/li>\n<li>don't use torch.cat to concatenate parameters. It allocates memory every time it is called. Instead use <code>nn.Container():add(enc):add(dec):getParameters()<\/code> instead of <code>torch.cat(enc:getParameters(), dec:getParameters())<\/code>.<\/li>\n<\/ol>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":41507694,"AnswerCount":3,"Tags":"<tensorflow><anaconda><spyder>","CreationDate":"2017-01-06T14:21:22.827","AcceptedAnswerId":41511246.0,"OwnerUserId":3552975.0,"Title":"Why cannot I import tfprof_log_pb2 in tensorflow?","Body":"<p>From <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/tfprof\/python\/tools\/tfprof\/tfprof_logger.py#L30\" rel=\"nofollow noreferrer\">this line<\/a> I get to know that the new added profile tool is imported, but it seems strange: 1) it cannot be imported from the interactive interface; 2) and when running the mnist example in spyder I also get an error sayinng . <\/p>\n\n<p>I thought since the tensorflow was instlled in anaconda and then it might be an old version and hence has no tfprof in it. But this would only explain why I cannot run  directly in the interactive interface(not in the anaconda environment) and cannot explain why it raises an error when I run it in anaconda. It raises this error in spyder, then the version which supports tfprof should be there. <\/p>\n","answers":[{"AnswerId":"41507724","CreationDate":"2017-01-06T14:23:01.313","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>Both of the two answers are right. <\/p>\n\n<p>In <strong>Ubuntu<\/strong>(python2.7 and GPU) I take a short cut and updated the tensorflow from 10 to 12 by: \n<code>\npip install \\\n  -i https:\/\/pypi.tuna.tsinghua.edu.cn\/simple\/ \\\n  https:\/\/mirrors.tuna.tsinghua.edu.cn\/tensorflow\/linux\/gpu\/tensorflow_gpu-0.12.1-cp27-none-linux_x86_64.whl\n<\/code>\nIn China I use the mirror in Tsinghua University for much faster installation.<br>\nIt goes well directly.   <\/p>\n\n<p>But this seems don't apply to <strong>Mac<\/strong>(at least for python3.5 and cpu only) even if your version of the tensorflow is 12 or above that file has not been generated(even other two files are already listed there(<code>\/\/anaconda\/lib\/python3.5\/site-packages\/tensorflow\/tools\/tfprof\/<\/code>). So I just copied that file(it must be the same as that in my Ubuntu, right?) privided by Yaroslav Bulatov to the folder. And it works. <\/p>\n"},{"AnswerId":"41511246","CreationDate":"2017-01-06T17:41:50.917","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Is <code>tfprof_log_pb2.py<\/code> present on your disk? It is an automatically generated file, and it's not present in the source tree. Instead it's placed in <code>bazel-genfiles<\/code> directory and should've been copied into distribution during build_pip_package.sh<\/p>\n\n<p>There's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5014\" rel=\"nofollow noreferrer\">this issue<\/a> that fixed some of the missing files for other tools, I wonder if tfprof files escaped the fix (ie, if it's not part of <code>tools:all_files<\/code> target)<\/p>\n\n<p>BTW, a copy of that file from my local\u00a0build -- <a href=\"http:\/\/pastebin.com\/dJnnWKuy\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/dJnnWKuy<\/a><\/p>\n"},{"AnswerId":"41511495","CreationDate":"2017-01-06T17:56:15.027","ParentId":null,"OwnerUserId":"1325423","Title":null,"Body":"<p>The problem happens if protobuf for python is older then 3.1.0 and TF is older too. I had exact problem as below:<\/p>\n\n<p>$ python<\/p>\n\n<pre><code>Python 2.7.10 (default, Jul 30 2016, 19:40:32)\n&gt;&gt;&gt; from tensorflow.tools.tfprof import tfprof_log_pb2\nTraceback (most recent call last):\nFile \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nImportError: No module named tfprof\n<\/code><\/pre>\n\n<p>This is how i solved it:<\/p>\n\n<p>Updated setuptools to latest 32.x or above version (I did this because my TF install was failing to update setuptools to version 32.x):<\/p>\n\n<pre><code>pip install update --user setuptools\n<\/code><\/pre>\n\n<p>After that I installed TF as below:<\/p>\n\n<pre><code>export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.1-py2-none-any.whl\nsudo pip install --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n\n<p>You will note that TF 12.1 installs the following:<\/p>\n\n<pre><code>Collecting tensorflow==0.12.1 from https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.1-py2-none-any.whl\nCollecting numpy&gt;=1.11.0 (from tensorflow==0.12.1)\nCollecting protobuf&gt;=3.1.0 (from tensorflow==0.12.1)\nCollecting setuptools (from protobuf&gt;=3.1.0-&gt;tensorflow==0.12.1)\n<\/code><\/pre>\n\n<p>After successful TF Install:<\/p>\n\n<pre><code>Successfully installed protobuf-3.1.0.post1 tensorflow-0.12.1\n<\/code><\/pre>\n\n<p>I tried the very first command as below:<\/p>\n\n<p>$ python       <\/p>\n\n<pre><code>Python 2.7.10 (default, Jul 30 2016, 19:40:32)\n&gt;&gt;&gt; from tensorflow.tools.tfprof import tfprof_log_pb2\n&gt;&gt;&gt; tfprof_log_pb2\n&lt;module 'tensorflow.tools.tfprof.tfprof_log_pb2' from '\/Library\/Python\/2.7\/site-packages\/tensorflow\/tools\/tfprof\/tfprof_log_pb2.pyc'&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":41508973,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-06T15:37:05.223","AcceptedAnswerId":null,"OwnerUserId":7384452.0,"Title":"AttributeError in module 'tensorflow.python.ops.variable_scope'","Body":"<p>setup:\nwindows 10,\ngpu Nvidia 760,\npython 3.5.2,\nAnaconda 4.2.0,\nTensorflow 0.12,<\/p>\n\n<p>when running either the Iris example or Deep &amp; wide, I'm getting the same error related to fit_model in case of Iris:<\/p>\n\n<pre><\/pre>\n\n<p>produces the error<\/p>\n\n<p>and for deep &amp; wide<\/p>\n\n<pre><\/pre>\n\n<p>produces the same error<\/p>\n\n<pre><\/pre>\n\n<p>I appropriate any help on this<\/p>\n","answers":[]}
{"QuestionId":41509145,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><gradient-descent><reinforcement-learning>","CreationDate":"2017-01-06T15:46:08.753","AcceptedAnswerId":null,"OwnerUserId":3949261.0,"Title":"Deep neural network diverges after convergence","Body":"<p>I implemented the A3C network in <a href=\"https:\/\/arxiv.org\/abs\/1602.01783\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1602.01783<\/a> in TensorFlow.<\/p>\n\n<p>At this point I'm 90% sure the algorithm is implemented correctly. However, the network diverges after convergence. See the attached image that I got from a toy example where the maximum episode reward is 7. <\/p>\n\n<p>When it diverges, policy network starts giving a single action very high probability (>0.9) for most states. <\/p>\n\n<p>What should I check for this kind of problem? Is there any reference for it? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/aYiNG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aYiNG.png\" alt=\"This is from a toy example, the maximum episode reward is 7.\"><\/a><\/p>\n","answers":[{"AnswerId":"51084679","CreationDate":"2018-06-28T13:52:14.187","ParentId":null,"OwnerUserId":"597746","Title":null,"Body":"<p>Note that in Figure 1 of the original paper the authors say:<\/p>\n\n<blockquote>\n  <p>For asynchronous methods we average over the best 5\n  models from 50 experiments.<\/p>\n<\/blockquote>\n\n<p>That can mean that in lot of cases the algorithm does not work that well. From my experience, A3C often diverges, even after convergence. Carefull learning-rate scheduling can help. Or do what the authors did - learn several agents with different seed and pick the one performing the best on your validation data. You could also employ early stopping when validation error becomes to increase.<\/p>\n"}]}
{"QuestionId":41509456,"AnswerCount":1,"Tags":"<python><tensorflow><windows-10>","CreationDate":"2017-01-06T16:03:07.367","AcceptedAnswerId":null,"OwnerUserId":5908470.0,"Title":"Python module installed on Windows 10 with Ubuntu bash is not recognized","Body":"<p>I have to use TensorFlow on my Windows 10 laptop, on which I have installed python 2.7, 3.2 and 3.4.2. I followed the instructions on the <a href=\"https:\/\/stackoverflow.com\/questions\/34785414\/how-to-install-tensorflow-on-windows\">first answer here<\/a>. I managed to install  and  with the Ubuntu bash environment and I successfully tested TensorFlow on the command prompt. However, when I try to import it in my IEP environment (using Python 3.4), the module is not recognized. I am pretty new in managing python modules. What am I missing?<\/p>\n\n<p>UPDATE: I found out that, when using Python via the ubuntu bash, the version is 3.4.3, but I never installed this version. Is there a way I can make it work with the versions I installed?<\/p>\n","answers":[{"AnswerId":"41510222","CreationDate":"2017-01-06T16:43:30.243","ParentId":null,"OwnerUserId":"1325423","Title":null,"Body":"<p>Looks like you are trying to make things complex to start with and having 3 (actually 4) python environment is making it even more harder if you are new to python. If you want to start using TF quickly and avoid platform specific problem, the quick route will be to remove most of platform 3.x installs and keep one if you really need 3.x otherwise you can do a lot with python 2.7. Once you have only 1 python environment, it will be very easy for you to manage and run faster. If you don't want to change anything and get you going with python do the following:<\/p>\n\n<ul>\n<li>Install pip 8.1 or above<\/li>\n<li>Install virtualenv for python 2.7 using pip - (This will help you to keep all the python specific modules within a specific folder)<\/li>\n<li>Install TF for python 2.7 from  <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/os_setup<\/a><\/li>\n<li>Use python 2.7 shell <\/li>\n<\/ul>\n\n<p>If you run this command it will show which TF module you are using and where it is install (you must have TF installed):<\/p>\n\n<blockquote>\n  <p>python -c 'import os; import inspect; import tensorflow; print(os.path.dirname(inspect.getfile(tensorflow)))'<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41509508,"AnswerCount":1,"Tags":"<python><deep-learning><keras><keras-layer>","CreationDate":"2017-01-06T16:05:49.670","AcceptedAnswerId":null,"OwnerUserId":4730816.0,"Title":"Fetch weights of the previous layer","Body":"<p>Is it possible to fetch the weights of the previous layer, modify them and set again to the next layer. I want to introduce a custom layer in the network which will modify the weights( as per the desired logic ) and then set the modified weight values to the next layer. Similar to what is depicted in the figure below:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4meFe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4meFe.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am not sure if this is possible or not. I know that we can dump the snapshot and then use it to set the new weights. I can also converted the weights using the snapshots. But, I dont know how to do this within the network itself ( without taking or using any snapshot).<\/p>\n\n<p>Thanks<\/p>\n\n<p>KK<\/p>\n","answers":[{"AnswerId":"41594384","CreationDate":"2017-01-11T15:12:33.247","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p><strong>tl;dr<\/strong>: Load one model (without compiling) and use the weights you want to initialize a model. Create new weights for the layers you want to change.<\/p>\n\n<p><strong>Full version<\/strong>:<\/p>\n\n<p>As per <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/1796\" rel=\"nofollow noreferrer\" title=\"this thread\">this thread<\/a> and as explained by fchollet himself, The canonical way to do this is to load your weights into the previous Keras model (you don't need to compile it, so it's instant) and use that model as a query-able datastructure to access the weights.<\/p>\n\n<p>For a sequential model you can do it like this:<\/p>\n\n<pre><code>weights = model.layers[5].get_weights()\nmodel.layers[5].set_weights(weights)\n<\/code><\/pre>\n\n<p>See also: <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1873\" rel=\"nofollow noreferrer\" title=\"Further discussion\">another discussion<\/a> on this topic with fchollet.<\/p>\n"}]}
{"QuestionId":41511096,"AnswerCount":0,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-01-06T17:34:09.690","AcceptedAnswerId":null,"OwnerUserId":3784885.0,"Title":"my deep learning code's accuray for solving xor logic is just 50%","Body":"<pre><\/pre>\n\n<p>Above is my tensorflow code for solving xor logic. But problem is that accuracy is just 50 percent. And cost Converge to 0.69321.\nActually I have seen lots of code that solves xor implemented in tensorflow, I can't find what is wrong.<\/p>\n\n<p>belows are images indicate how my code works.<\/p>\n\n<p><img src=\"https:\/\/i.imgur.com\/mSokJs1.png\" alt=\"graph node\">\n<img src=\"https:\/\/i.imgur.com\/W6ObkyK.png\" alt=\"cost\"><\/p>\n","answers":[]}
{"QuestionId":41511514,"AnswerCount":2,"Tags":"<ubuntu><caffe><hdf5>","CreationDate":"2017-01-06T17:57:33.573","AcceptedAnswerId":41525431.0,"OwnerUserId":7385135.0,"Title":"hdf5 install on Ubuntu 14.04","Body":"<p>I installed libhdf5 on Ubuntu via apt-get but there are no directories. <\/p>\n\n<blockquote>\n  <p>sudo find -type d -name '<em>hdf5<\/em>'<\/p>\n  \n  <p>.\/share\/doc\/hdf5-tools<\/p>\n  \n  <p>.\/share\/doc\/libhdf5-mpich2-dev<\/p>\n  \n  <p>.\/share\/doc\/libjhdf5-java<\/p>\n  \n  <p>.\/share\/doc\/hdf5-helpers<\/p>\n  \n  <p>.\/share\/doc\/libjhdf5-jni<\/p>\n  \n  <p>.\/share\/doc\/libhdf5-mpich2-7<\/p>\n  \n  <p>.\/local\/MATLAB\/R2016b\/toolbox\/matlab\/imagesci\/@hdf5<\/p>\n  \n  <p>.\/local\/MATLAB\/R2016b\/toolbox\/matlab\/imagesci\/@hdf5\/@hdf5type<\/p>\n<\/blockquote>\n\n<p>I have some issue with hdf5 during compiling CAFFE. It's suggested to add path to hdf5 directory manually to Makefile.config but I can't find directory at all.<\/p>\n\n<p>Hope someone can help me.\nThanks.<\/p>\n","answers":[{"AnswerId":"41512326","CreationDate":"2017-01-06T18:50:26.370","ParentId":null,"OwnerUserId":"7282091","Title":null,"Body":"<p>In my Linux Mint (similar to Ubuntu, at least apt-get...) I can find the headers under <code>\/usr\/include\/hdf5<\/code> and the libraries under <code>\/usr\/lib\/x86_64-linux-gnu<\/code>.<br>\nIf the installation was done correctly, I believe you should find the HDF5 stuff in similar places in your Ubuntu installation.<br>\nTo locate the files you could try to execute in the terminal  <\/p>\n\n<pre><code>locate hdf5\n<\/code><\/pre>\n\n<p>To be noticed that your command  <\/p>\n\n<pre><code>sudo find -type d -name 'hdf5'  \n<\/code><\/pre>\n\n<p>doesn't find the headers and libraries on my machine too.<\/p>\n"},{"AnswerId":"41525431","CreationDate":"2017-01-07T19:14:27.473","ParentId":null,"OwnerUserId":"7385135","Title":null,"Body":"<p>Files were located in <code>\/usr\/include<\/code>. To make caffe compiled I manually created <code>\/hdf5\/serial<\/code> and put header files (<code>hdf5.h  hdf5_hl.h<\/code>) there.<\/p>\n"}]}
{"QuestionId":41512691,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-01-06T19:11:43.660","AcceptedAnswerId":41976995.0,"OwnerUserId":64106.0,"Title":"tensorflow - creating proto files error \"assert d in name_to_node_map,\"","Body":"<p>I'm using keras with tensorflow. \nI'm exporting the model correcty but fail to create proto files.<\/p>\n\n<p>The problem is that on i get an error:<\/p>\n\n<blockquote>\n  <p>File \"keras_indians.py\", line 103, in \n      minimal_graph = convert_variables_to_constants(sess, sess.graph_def, ['Dense3\/output_node'])   File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/graph_util.py\",\n  line 234, in convert_variables_to_constants\n      inference_graph = extract_sub_graph(input_graph_def, output_node_names)   File\n  \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/graph_util.py\",\n  line 158, in extract_sub_graph\n      assert d in name_to_node_map, \"%s is not in graph\" % d AssertionError: Dense3\/output_node is not in graph<\/p>\n<\/blockquote>\n\n<p>I tried any possible variation for the last node name and it doesn't work\nAny help is highly appreciated !<\/p>\n\n<p>Below is the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41976995","CreationDate":"2017-02-01T10:09:43.850","ParentId":null,"OwnerUserId":"6737609","Title":null,"Body":"<p>I struggled with the same problem until now, when I am looping through all nodes created in my graph. I found out that saver rewrite the names of all nodes with the prefix \"restore\/***\" after debugging graph with the code below, found out last node and use it in freezeGraph script.<\/p>\n\n<pre><code> graph_def = sess.graph.as_graph_def()\n for node in graph_def.node:\n     print node\n<\/code><\/pre>\n\n<p>More info :<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3986\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3986<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/38958662\/tensorflow-what-are-the-output-node-names-for-freeze-graph-py-in-the-model-wi\">Tensorflow: What are the &quot;output_node_names&quot; for freeze_graph.py in the model_with_buckets model?<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/how_tos\/tool_developers\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/tool_developers\/<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/37704362\/tensorflow-freeze-graph-script-failing-on-model-defined-with-keras\">Tensorflow freeze_graph script failing on model defined with Keras<\/a><\/p>\n"}]}
{"QuestionId":41512790,"AnswerCount":1,"Tags":"<machine-learning><initialization><neural-network><deep-learning><caffe>","CreationDate":"2017-01-06T19:19:00.907","AcceptedAnswerId":41577316.0,"OwnerUserId":5635517.0,"Title":"Use a Pre-trained Network(1 branch) to initialize weights of layers in a Network with 2 branches in caffe","Body":"<p>I have a Multi Task Network with two similar branches and a pre-trained network with only one branch (which is also same).\nI want to initialize the weights of the layers in the two branches(in my multi task network) with the weights of the layers in my pre-trained network.<\/p>\n\n<p>Now, I can initialize one of the branch correctly by using the same name for the layers as in the pre-trained network.\nBut, I have to keep the names of the layers in the other branch different, and thus those layers won't take the pre-trained weights.<\/p>\n\n<p>Also, I don't want to share the weights in the two branches. So, giving the same name to the weights in the corresponding layers in the two branches won't work.<\/p>\n\n<p>Is there a nice way\/hack to do this ? <\/p>\n\n<p>PS: I would want to avoid Network Surgery, but any comments, explaining a nice way to do it, are also welcome.<\/p>\n\n<p>Clarification : I just want to initialize the two branches with the same weights. They can learn different weights during the training phase, since they are governed by different loss layers.<\/p>\n","answers":[{"AnswerId":"41577316","CreationDate":"2017-01-10T19:57:41.997","ParentId":null,"OwnerUserId":"5635517","Title":null,"Body":"<p>The <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/E3GakB5DENE\" rel=\"nofollow noreferrer\">answer<\/a> by <em>Przemak D<\/em> is a nice hack to do the above.<\/p>\n\n<ol>\n<li>give different names to the layers in the two branches and enable weight sharing<\/li>\n<li>initialize the network and train for 1-2 iterations<\/li>\n<li>then train the original network(without weight sharing) initializing the weights with the caffemodel obtained as a result after step 2.<\/li>\n<\/ol>\n\n<p>The above is a nice hack, but net surgery is a better way to do this.<\/p>\n"}]}
{"QuestionId":41512845,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2017-01-06T19:23:01.457","AcceptedAnswerId":41512903.0,"OwnerUserId":6786718.0,"Title":"What does caffe do with the mean-binary file ?","Body":"<p>In the caffe-input layer one can define a mean image that holds mean values of all the images used. From the image net example: \"The model requires us to subtract the image mean from each image, so we have to compute the mean\".<\/p>\n\n<p>My question is: What is the implementation of this subtraction? Is it simply :<\/p>\n\n<pre><\/pre>\n\n<p>or <\/p>\n\n<pre><\/pre>\n\n<p>or <\/p>\n\n<pre><\/pre>\n\n<p>if it is one of the first two, then how are negative pixels handeld ? Since the pictures are usually stored in uint8 it would mean that it simply starts from the beginning. e.g <\/p>\n\n<blockquote>\n  <p>200 - 255 = 56 <\/p>\n<\/blockquote>\n\n<p>Why I need to know this? I made tests and I know that the second example or the third example would work better. <\/p>\n","answers":[{"AnswerId":"41512903","CreationDate":"2017-01-06T19:27:56.143","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>It's the first one, a trivial normalization step.  Using the second instead wouldn't really matter: the weights would invert.<\/p>\n\n<p>There are no \"negative pixels\", per se: this is simply integer input to the matrix operations.  <em>You<\/em> are welcome to interpret this as a visual alteration of some sort, but the arithmetic doesn't care.<\/p>\n"}]}
{"QuestionId":41513973,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><out-of-memory><deep-learning><underflow>","CreationDate":"2017-01-06T20:38:51.180","AcceptedAnswerId":null,"OwnerUserId":7233512.0,"Title":"Buffer underrun and ResourceExhausted errors with tensorflow","Body":"<p>I'm in high school and I'm trying to do a project involving neural networks.  I am using Ubuntu and trying to do reinforcement learning with tensorflow, but I consistently get lots of underrun warnings when I train a neural network.  They take the form of .  This message is printed to the screen more and more frequently as training progresses.  Eventually, I get a ResourceExhaustedError and the program terminates.  Here is the full error message:<\/p>\n\n<pre><\/pre>\n\n<p>I researched these problems but didn't get a clear idea of how I could fix them.  I am pretty new to programming so I don't know much about how buffers and data reading\/writing works.  I am very perplexed by these errors.  Does anyone know what parts of my code might be causing this and how to fix it?  Thanks for taking the time to consider this question!<\/p>\n\n<p>Here is my code for defining the neural network (based on <a href=\"https:\/\/medium.com\/emergent-future\/simple-reinforcement-learning-with-tensorflow-part-8-asynchronous-actor-critic-agents-a3c-c88f72a5e9f2#.964v242sd\" rel=\"nofollow noreferrer\">this tutorial<\/a>):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Here is my code for training the neural network:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"41537743","CreationDate":"2017-01-08T20:50:04.710","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You are running out of memory. It's possible that your network requires more memory than you have to run, so the first step to tracking down excessive memory usage is to figure out what is using so much memory.<\/p>\n\n<p>Here's one approach that uses timeline and statssummarizer:\n<a href=\"https:\/\/gist.github.com\/yaroslavvb\/08afccbe087171881ceafc0c98abca05\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/yaroslavvb\/08afccbe087171881ceafc0c98abca05<\/a><\/p>\n\n<p>This will print out several tables, one of the tables is the tensors sorted by top memory usage. You should check that you don't have something unusually large in there.<\/p>\n\n<p>You can also see memory timeline using Chrome visualizer, as detailed <a href=\"https:\/\/stackoverflow.com\/questions\/36123740\/is-there-a-way-of-determining-how-much-gpu-memory-is-in-use-by-tensorflow\/37931964#37931964\">here<\/a><\/p>\n\n<p>A more advanced technique is to plot a timeline of memory allocations\/deallocations, as done in this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6019#issuecomment-267881864\" rel=\"nofollow noreferrer\">issue<\/a><\/p>\n\n<p>Theoretically your memory usage shouldn't grow between steps if you aren't creating new stateful ops (Variables), but I found that global memory allocation can grow if sizes of your tensors change between steps. <\/p>\n\n<p>A work-around is to periodically save your parameters to checkpoint and restart your script.<\/p>\n"}]}
{"QuestionId":41514520,"AnswerCount":1,"Tags":"<neural-network><keras><conv-neural-network><dimensions><autoencoder>","CreationDate":"2017-01-06T21:19:32.320","AcceptedAnswerId":null,"OwnerUserId":2145297.0,"Title":"Convolution layers decrease dimension of output in Keras?","Body":"<p>I'm trying to build a convolutional autoencoder in Keras. I was under the impression that convolutional layers have the same output dimension as the input unless if you set the stride>1 or set border_mode='valid'. Here is my network<\/p>\n\n<pre><\/pre>\n\n<p>When I check to see the output dimension of the first layer it seems like the length of the y coordinate has been decreased.<\/p>\n\n<pre><\/pre>\n\n<p>What's causing this? I can't figure out where that 16 is coming from. It doesn't seem to have any simple relationship to 1080.<\/p>\n","answers":[{"AnswerId":"41516589","CreationDate":"2017-01-07T00:41:53.777","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The problem lies in a so called <code>image_ordering<\/code>. Depending on which backend you use - you need to provide data in a specified format:<\/p>\n\n<ul>\n<li><p>for <code>th<\/code> (Theano) image ordering you should provide data in a format:<\/p>\n\n<pre><code>[batches, channels, image_width, image_weight]\n<\/code><\/pre>\n\n<p>This is a format you are assuming as correct.<\/p><\/li>\n<li><p>for <code>tf<\/code> (TensorFlow) image ordering is:<\/p>\n\n<pre><code> [batches, width, height, channels]\n<\/code><\/pre>\n\n<p>This is a format by which your Keras installation is interpretting your input.<\/p><\/li>\n<\/ul>\n\n<p>A default image ordering in keras is <code>tf<\/code> and I'm assuming that this is causing your problem. You may check your image ordering by the following commands:<\/p>\n\n<pre><code>from keras.backend import image_dim_ordering\nprint image_dim_ordering() # Assuming that you're using a Python 2.*\n<\/code><\/pre>\n\n<p>You may change it by changing a <code>keras.json<\/code> what is described <a href=\"https:\/\/keras.io\/backend\/#kerasjson-details\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>You may also check your model details to assure yourself that this is actually the case by printing the result of <code>model.summary()<\/code>. <\/p>\n"}]}
{"QuestionId":41514722,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-01-06T21:34:23.160","AcceptedAnswerId":41523937.0,"OwnerUserId":2213913.0,"Title":"Convert the strictly upper triangular part of a matrix into an array in Tensorflow","Body":"<p>I was trying to convert the strictly upper triangular part of a matrix into an array in Tensorflow. Here is an example:<\/p>\n\n<p>Input:<\/p>\n\n<pre><\/pre>\n\n<p>Output:<\/p>\n\n<pre><\/pre>\n\n<p>I tried the following code but it did not work (an error was reported):<\/p>\n\n<pre><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"46614084","CreationDate":"2017-10-06T21:21:35.760","ParentId":null,"OwnerUserId":"7146757","Title":null,"Body":"<p>The following answer follows closely the answer from @Cech_Cohomology, but it doesn't use Numpy in the process, only TensorFlow.<\/p>\n\n<pre><code>import tensorflow as tf\n\n# The matrix has size n-by-n\nn = 3\n\n# A is the matrix\nA = [[1, 2, 3], [4, 5, 6], [7, 8, 9]]\n\nones = tf.ones_like(A)\nmask_a = tf.matrix_band_part(ones, 0, -1) # Upper triangular matrix of 0s and 1s\nmask_b = tf.matrix_band_part(ones, 0, 0)  # Diagonal matrix of 0s and 1s\nmask = tf.cast(mask_a - mask_b, dtype=tf.bool) # Make a bool mask\n\nupper_triangular_flat = tf.boolean_mask(A, mask)\n\nsess = tf.Session()\nprint(sess.run(upper_triangular_flat))\n<\/code><\/pre>\n\n<p>This outputs:<\/p>\n\n<pre><code>[2 3 6]\n<\/code><\/pre>\n\n<p>The advantage of this method is that when the graph is run there is no need to give a <code>feed_dict<\/code>.<\/p>\n"},{"AnswerId":"41515440","CreationDate":"2017-01-06T22:32:04.683","ParentId":null,"OwnerUserId":"2701363","Title":null,"Body":"<p>If you are using python 2.7, then for an NxN-element array, you can use a list comprehension with a conditional:<\/p>\n\n<pre><code>def upper_triangular_to_array(A):\n    N = A.shape[0]\n    return np.array([p for i, p in enumerate(A.flatten()) if i &gt; (i \/ N) * (1 + N)])\n<\/code><\/pre>\n\n<p>This function requires that A is a 2-dimensional square numpy array to return the correct result. It also relies on floor division of integers, which you will need to correct for if you use python 3.x<\/p>\n"},{"AnswerId":"41523937","CreationDate":"2017-01-07T16:52:36.423","ParentId":null,"OwnerUserId":"2213913","Title":null,"Body":"<p>I finally figured out how to do that using Tensorflow.<\/p>\n\n<p>The idea is to define a placeholder as the boolean mask and then use numpy to pass a boolean matrix to the boolean mask in the runtime. I share my code below:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# The matrix has size n-by-n\nn = 3\n# define a boolean mask as a placeholder\nmask = tf.placeholder(tf.bool, shape=(n, n))\n# A is the matrix\nA = [[1, 2, 3], [4, 5, 6], [7, 8, 9]]\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    npmask = np.triu(np.ones((n, n), dtype=np.bool_), 1)\n    A_upper_triangular = tf.boolean_mask(A, mask)\n    print(sess.run(A_upper_triangular, feed_dict={mask: npmask}))\n<\/code><\/pre>\n\n<p>My Python version is 3.6 and my Tensorflow version is 0.12.0rc1. The output of the above code is <\/p>\n\n<pre><code>[2, 3, 6]\n<\/code><\/pre>\n\n<p>This method can be further generalized. We can use numpy to construct any kind of mask and then pass the mask to the Tensorflow to extract the part of the tensor of interest.<\/p>\n"}]}
{"QuestionId":41514872,"AnswerCount":0,"Tags":"<linux><theano><theano-cuda>","CreationDate":"2017-01-06T21:45:15.800","AcceptedAnswerId":null,"OwnerUserId":7385862.0,"Title":"My CUDA and Theano do not work together","Body":"<p>I am trying to test my GPU sanity by running this code <\/p>\n\n<pre><\/pre>\n\n<p>I got this error:<\/p>\n\n<blockquote>\n  <p>mod.cu(941): warning: pointless comparison of unsigned integer with zero<br>\n  mod.cu(3001): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3004): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3006): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3009): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3011): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3014): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3017): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3020): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3022): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3025): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3027): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3030): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3032): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3035): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3038): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3041): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3043): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3046): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3048): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3051): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(941): warning: pointless comparison of unsigned integer with zero<br>\n  mod.cu(3001): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3004): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3006): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3009): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3011): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3014): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3017): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3020): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3022): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3025): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3027): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3030): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3032): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3035): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3038): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3041): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3043): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3046): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3048): warning: conversion from a string literal to \"char *\" is deprecated<br>\n  mod.cu(3051): warning: conversion from a string literal to \"char <em>\" is deprecated<br>\n  \/usr\/include\/string.h: In function \u2018void<\/em> __mempcpy_inline(void*, const void*, size_t)\u2019:<br>\n  \/usr\/include\/string.h:652:42: error: \u2018memcpy\u2019 was not declared in this scope<br>\n     return (char <em>) memcpy (__dest, __src, __n) + __n;<br>\n                                            ^<br>\n  mod.cu: In function \u2018PyObject<\/em> CudaNdarray_Reshape(CudaNdarray*, PyObject*)\u2019:  <\/p>\n  \n  <p>mod.cu:955:122: warning: format \u2018%lld\u2019 expects argument of type \u2018long long int\u2019, but argument 3 has type \u2018size_t {aka long unsigned int}\u2019 [-Wformat=]<br>\n           PyErr_Format(PyExc_ValueError, \"size must remain unchanged, changed from %lld to %lld\", CudaNdarray_SIZE(self), rval_size);<br>\n                                                                                                                            ^<br>\n  ['nvcc', '-shared', '-O3', '-m64', '-Xcompiler', '-DCUDA_NDARRAY_CUH=c72d035fdf91890f3b36710688069b2e,-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION,-fPIC,-fvisibility=hidden', '-Xlinker', '-rpath,\/home\/rkenaya\/.theano\/compiledir_Linux-4.4--generic-x86_64-with-Ubuntu-16.04-xenial-x86_64-2.7.12-64\/cuda_ndarray', '-I\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/sandbox\/cuda', '-I\/usr\/local\/lib\/python2.7\/dist-packages\/numpy\/core\/include', '-I\/usr\/include\/python2.7', '-I\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/gof', '-o', '\/home\/rkenaya\/.theano\/compiledir_Linux-4.4--generic-x86_64-with-Ubuntu-16.04-xenial-x86_64-2.7.12-64\/cuda_ndarray\/cuda_ndarray.so', 'mod.cu', '-L\/usr\/lib', '-lcublas', '-lpython2.7', '-lcudart']\n  mod.cu:955:122: warning: format \u2018%lld\u2019 expects argument of type \u2018long long int\u2019, but argument 4 has type \u2018size_t {aka long unsigned int}\u2019 [-Wformat=]\n  ERROR (theano.sandbox.cuda): Failed to compile cuda_ndarray.cu: ('nvcc return status', 1, 'for cmd', 'nvcc -shared -O3 -m64 -Xcompiler -DCUDA_NDARRAY_CUH=c72d035fdf91890f3b36710688069b2e,-DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION,-fPIC,-fvisibility=hidden -Xlinker -rpath,\/home\/rkenaya\/.theano\/compiledir_Linux-4.4--generic-x86_64-with-Ubuntu-16.04-xenial-x86_64-2.7.12-64\/cuda_ndarray -I\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/sandbox\/cuda -I\/usr\/local\/lib\/python2.7\/dist-packages\/numpy\/core\/include -I\/usr\/include\/python2.7 -I\/usr\/local\/lib\/python2.7\/dist-packages\/theano\/gof -o \/home\/rkenaya\/.theano\/compiledir_Linux-4.4--generic-x86_64-with-Ubuntu-16.04-xenial-x86_64-2.7.12-64\/cuda_ndarray\/cuda_ndarray.so mod.cu -L\/usr\/lib -lcublas -lpython2.7 -lcudart')<br>\n  WARNING (theano.sandbox.cuda): CUDA is installed, but device gpu is not available  (error: cuda unavailable)<br>\n  [Elemwise{exp,no_inplace}()]<br>\n  Looping 1000 times took 4.233657 seconds<br>\n  Result is [ 1.23178029  1.61879337  1.52278066 ...,  2.20771813  2.29967761\n    1.62323284]<br>\n  Used the cpu<\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":41515716,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-06T22:56:57.890","AcceptedAnswerId":null,"OwnerUserId":7382474.0,"Title":"Tensorflow tf.contrib.learn.DNNClassifer estimation accuracy does not align to DNNClassifier prediction accuracy","Body":"<p>In general, what alignment should we expect between the accuracy achieved during estimation, and the accuracy achieved during prediction? Simply put, if we achieved an 85% accuracy during estimation, should we expect a similar result during prediction? There are a few other posts discussion prediction accuracy issues, but none appear to ask this question directly.<\/p>\n\n<p>To be more specific, I am using a DNNClassifier to learn against 1000 rows of feature data, each with 53 characteristics (i.e. columns). This is then tested against 100 samples. The model is configured to classify against two labels: 0 and 1, and achieves the prediction accuracy below.<\/p>\n\n<p>INFO:tensorflow:Saving evaluation summary for step 45300:\naccuracy: 0.846154\naccuracy\/baseline_label_mean: 0.197802\naccuracy\/threshold_0.500000_mean: 0.846154\nauc: 0.743912\nglobal_step: 45300\nlabels\/actual_label_mean: 0.197802\nlabels\/prediction_mean: 0.283492\nloss: 0.398202\nprecision\/positive_threshold_0.500000_mean: 1.0\nrecall\/positive_threshold_0.500000_mean: 0.222222<\/p>\n\n<p>However, when 15 examples are then used during the prediction step, and where 50% should be classified with a label of 1, they are generally all returned as 0s. Occasionally, a single example will be correctly classified as a 1.<\/p>\n\n<p>For example:\nPredictions: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]<\/p>\n\n<p>My DNNClassifier is configured as follows:<\/p>\n\n<pre><\/pre>\n\n<p>and the predict function is called as below:<\/p>\n\n<pre><\/pre>\n\n<p>I have two questions.\n(1) Am I correct to assume that there should be a direct correlation between estimation and prediction accuracy, and therefore that an 85% estimation accuracy should result in a similar prediction accuracy?<\/p>\n\n<p>(2) Do the other results provided when training completes (i.e. \"labels\/prediction_mean: 0.283492\") provide information that is relevant to the low level of prediction accuracy being achieved? I have not been able to find a definition of these result labels in the TensorFlow documentation, and my apologies if I have missed something.<\/p>\n\n<p>Many thanks for any thoughts on this.<\/p>\n","answers":[]}
{"QuestionId":41516058,"AnswerCount":2,"Tags":"<tensorflow><theano><keras>","CreationDate":"2017-01-06T23:31:50.143","AcceptedAnswerId":null,"OwnerUserId":4015544.0,"Title":"tensorflow theano.tensor.set_subtensor equivalent","Body":"<p>I am implementing an operation in keras, such that it can work on both theano and tensorflow backend. Suppose the input of the operation is:<\/p>\n\n<pre><\/pre>\n\n<p>then its output should be:<\/p>\n\n<pre><\/pre>\n\n<p>My codes are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I have successfully implemented it by theano, but I can not figure out how to assign a tensor by tensorflow. The last two lines of tensor assignment in tensorflow will report error. Is there a T.set_subtensor equivalence in tensorflow? or can you please recommend a better implementation of the operation? Thanks.<\/p>\n","answers":[{"AnswerId":"41516524","CreationDate":"2017-01-07T00:30:31.240","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>TensorFlow tensors are read-only. In order to modify things you need to use variables and <code>.assign<\/code> (= can not be overriden in Python)<\/p>\n\n<pre><code>tensor = tf.Variable(tf.ones((3,3)))\nsess.run(tf.initialize_all_variables())\nsess.run(tensor[1:, 1:].assign(2*tensor[1:,1:]))\nprint(tensor.eval())\n<\/code><\/pre>\n\n<p>Output<\/p>\n\n<pre><code>[[ 1.  1.  1.]\n [ 1.  2.  2.]\n [ 1.  2.  2.]]\n<\/code><\/pre>\n"},{"AnswerId":"42798450","CreationDate":"2017-03-14T23:30:53.003","ParentId":null,"OwnerUserId":"6666226","Title":null,"Body":"<p>It took a lot of searching, but the closest functions to <code>theano.tensor.set_subtensor<\/code> are <code>gather<\/code>, <code>gather_nd<\/code>, <code>scatter<\/code>, and <code>scatter_nd<\/code>. If you are trying to perform sparse updates to a variable, the other answers could work. But if you're trying to dynamically create a tensor from indexing another tensor, those are the functions to use.<\/p>\n\n<p>The point of these functions is to be able to dynamically create a tensor (not a variable) from something else. My use-case is I am generating a flat tensor and I am trying to reshape it into various triangular matrices.<\/p>\n\n<p><code>gather<\/code> is what you use if you're trying to create a smaller matrix from a large sparse matrix. <code>scatter<\/code> is what you use if you're trying to embed a smaller matrix into a large zero matrix.<\/p>\n\n<p>Some combination of <code>gather<\/code>, <code>scatter<\/code> and addition and multiplication can recreate <code>theano.tensor.set_subtensor<\/code>.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scatter_nd\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/scatter_nd<\/a>\n<a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/array_ops#Slicing_and_Joining\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/array_ops#Slicing_and_Joining<\/a><\/p>\n\n<p>You could also use a very complicated set of slicing and concatenation, but gather and scatter would be preferred.<\/p>\n\n<p>Cheers<\/p>\n"}]}
{"QuestionId":41516939,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-07T01:38:49.717","AcceptedAnswerId":41520931.0,"OwnerUserId":3792504.0,"Title":"Tensorflow mask from one-hot encoding","Body":"<p>I have labels that are OHE in the form of  where each example is an  in the range .<\/p>\n\n<p>My output is in the form of a softmax probability distribution with a shape <\/p>\n\n<p>I'm trying to work out how to create a mask that will give me just the probability distribution for each example. e.g.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41520931","CreationDate":"2017-01-07T11:27:43.387","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p>If I understood your example correctly, you need <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#gather_nd\" rel=\"nofollow noreferrer\">tf.gather_nd<\/a><\/p>\n\n<pre><code>range = tf.range(tf.shape(examples)[0])\nindices = tf.pack([range, examples], axis=1)\nresult = tf.gather_nd(probs, indices)\n<\/code><\/pre>\n"}]}
{"QuestionId":41517145,"AnswerCount":0,"Tags":"<memory-leaks><tensorflow>","CreationDate":"2017-01-07T02:18:23.290","AcceptedAnswerId":null,"OwnerUserId":3654968.0,"Title":"Outer product based Conv filters consume disproportionately high memory","Body":"<p>The minimum reproducible example is in this <a href=\"https:\/\/gist.github.com\/sahiliitm\/bed9511d15ba0c45adf3c33210d59232\" rel=\"nofollow noreferrer\">github gist<\/a>. The issue is of a disproportionately large memory usage.<\/p>\n\n<p>I modified the cifar10 example which ships with tensorflow to use outer-product of 3 vectors as the weights of the convolutional layers. This change can be seen in this <a href=\"https:\/\/gist.github.com\/sahiliitm\/bed9511d15ba0c45adf3c33210d59232#file-cifar10-py-L91\" rel=\"nofollow noreferrer\">part of the code<\/a>.<\/p>\n\n<p>For simplicity, i have removed all parameter training operations and even loss computations. The current model only computes logits (forward pass) again and again.\nThe unmodified code (can be executed by setting the  flag to ) uses approximately .4 GB RAM\nwhereas the modified code (with outer product of vectors used as the convolutional weight tensor) uses a disproportionately high 5.6 GB RAM.<\/p>\n\n<p>Any idea why this is the case?\nMy intuition as to why this might happen is that maybe the outer product operations are being executed every single time that the conv filter is needed instead of being executed exactly once in every forward pass. Is this really the case? Is there a way to fix this?<\/p>\n\n<p>Steps to reproduce:<\/p>\n\n<p>To run the default version of the code (low memory footprint):<\/p>\n\n<p><\/p>\n\n<p>To run the modified version of the code (high memory footprint):<\/p>\n\n<p><\/p>\n\n<p>Operating System:<\/p>\n\n<p>Ubuntu 14.04<\/p>\n\n<p>Installed version of CUDA and cuDNN:<\/p>\n\n<p>None, I'm using the CPU version of TF.<\/p>\n\n<p>The output from :<\/p>\n\n<p>0.10.0rc0<\/p>\n\n<p>Using tcmalloc as suggested <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6688\" rel=\"nofollow noreferrer\">here<\/a> didn't help.<\/p>\n","answers":[]}
{"QuestionId":41517202,"AnswerCount":1,"Tags":"<python><machine-learning><scikit-learn><keras><roc>","CreationDate":"2017-01-07T02:29:36.190","AcceptedAnswerId":41618751.0,"OwnerUserId":4592311.0,"Title":"roc_curve from multilabel classification has slope","Body":"<p>I have a multilabel classifier written in Keras from which I want to compute AUC and plot a ROC curve for every element classified from my test set. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/IUWLe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IUWLe.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Everything seems fine, except that some elements  have a roc curve that have a slope as follows: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XCNCA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XCNCA.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I don't know how to interpret the slope in such cases.<\/p>\n\n<p>Basically my workflow goes as follows, I have a pre-trained , instance of Keras, and I have the features  and the binarized labels , every element in  is an array of length 1000, as it is a multilabel classification problem each element in  might contain many 1s, indicating that the element belongs to multiples classes, so I used the built-in loss of  and my outputs of the model prediction are score probailities. Then I plot the roc curve as follows.<\/p>\n\n<pre><\/pre>\n\n<p>The predict method returns probabilities, as I'm using the functional api of keras.<\/p>\n\n<p>Does anyone knows why roc curves looks like this?<\/p>\n","answers":[{"AnswerId":"41618751","CreationDate":"2017-01-12T16:52:06.280","ParentId":null,"OwnerUserId":"4592311","Title":null,"Body":"<p>Asking in the mailing list of scikit-learn, they answered: <\/p>\n\n<blockquote>\n  <p>Slope usually means there are ties in your predictions.<\/p>\n<\/blockquote>\n\n<p>Which is the case in this problem.<\/p>\n"}]}
{"QuestionId":41517497,"AnswerCount":0,"Tags":"<python><tensorflow><distributed>","CreationDate":"2017-01-07T03:27:05.713","AcceptedAnswerId":null,"OwnerUserId":7071746.0,"Title":"Distributed tensorflow: non chief worker stuck at starter session","Body":"<p>I have two workers and a parameter server. The chief worker works well, but the non chief worker stuck at this line of code<\/p>\n\n<pre><\/pre>\n\n<p>and produce output as follows:<\/p>\n\n<pre><\/pre>\n\n<p>It start master session again and again.  <\/p>\n\n<p>train supervisor is set as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestions? Thanks a lot!<\/p>\n\n<h3>update<\/h3>\n\n<p>quote from tensorflow documentation:<\/p>\n\n<blockquote>\n  <p>In the chief task, the Supervisor works exactly as in the first example above. In the other tasks sv.managed_session() waits for the Model to have been initialized before returning a session to the training code. The non-chief tasks depend on the chief task for initializing the model.<\/p>\n  \n  <p>If one of the tasks crashes and restarts, managed_session() checks if the Model is initialized. If yes, it just creates a session and returns it to the training code that proceeds normally. If the model needs to be initialized, the chief task takes care of reinitializing it; the other tasks just wait for the model to have been initialized.<\/p>\n<\/blockquote>\n\n<p>The distributed approach I used here is <em>Between-graph replication<\/em>, The reason is I defined two graphs in a workers, one of them is not complete. and I try to complete that graph in non-chief worker. Therefore it crash at managed_session() and restart. This problem is fixed when I define two graphs in chief worker completely. <\/p>\n","answers":[]}
{"QuestionId":41518379,"AnswerCount":1,"Tags":"<c++><tensorflow><eigen>","CreationDate":"2017-01-07T06:03:44.140","AcceptedAnswerId":41564099.0,"OwnerUserId":5716954.0,"Title":"Why was Eigen chosen for TensorFlow?","Body":"<p>The TensorFlow white paper mentions that Eigen is used. Are there public explanations for how Eigen was chosen, and are they motivation for using Eigen in TensorFlow C++ op kernels?<\/p>\n","answers":[{"AnswerId":"41564099","CreationDate":"2017-01-10T08:15:18.450","ParentId":null,"OwnerUserId":"1641621","Title":null,"Body":"<p>I think that one of the key feature that drove the use of Eigen in the first place is because Eigen features its own highly optimized matrix product kernels whereas all other competitors have to be linked to some BLAS libraries. Moreover, the code of Eigen's product kernel is C++ with easy access to low-level internal kernels, so it was 'easy' for them to tweak and extend it to match their needs. This way Google has been able to develop the Tensor module with high CPU performance in a pure header-only fashion. The support for CUDA and now OpenCL via SyCL came later, those are not intrinsic features of Eigen that drove the initial choice.<\/p>\n"}]}
{"QuestionId":41518602,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-07T06:33:51.673","AcceptedAnswerId":41520773.0,"OwnerUserId":3102085.0,"Title":"Output of loss is None","Body":"<p>I have to finetune VGG.There are five convolutional layers and then three fully connected layers. Output from the last fully connected layer is the input of the loss function. Following is my code:<\/p>\n\n<pre><\/pre>\n\n<p>The problem is that the output of prob is None. Kindly indicate what I am doing wrong.<\/p>\n\n<p>PS: I am following siamese architecture. Input to both branches are different images here.<\/p>\n","answers":[{"AnswerId":"41520773","CreationDate":"2017-01-07T11:09:13.373","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p>The op <code>self.train_step<\/code> does not return anything, it just calculates gradients and updates variables. See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/optimizers#Optimizer.minimize\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>What you need to do is to save reference to <code>loss<\/code> tensor in your <code>vgg16<\/code> class like this:<\/p>\n\n<pre><code>self.loss=tf.nn.l2_loss(self.fc3l)\n<\/code><\/pre>\n\n<p>and then execute both train_step and loss operations in single <code>sess.run<\/code>:<\/p>\n\n<pre><code>_, loss_value = sess.run([vgg.train_step, vgg.loss], feed_dict=...)\nprint('loss is:')\nprint(loss_value)\n<\/code><\/pre>\n"}]}
{"QuestionId":41518869,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-07T07:11:57.643","AcceptedAnswerId":null,"OwnerUserId":933882.0,"Title":"how does TensorFlow handle the differentials for L1 regularization?","Body":"<p>it seems that you can just declare a cost function by tf.abs() and then pass it down to auto-gradient generation (see <a href=\"https:\/\/github.com\/nfmcclure\/tensorflow_cookbook\/blob\/master\/03_Linear_Regression\/04_Loss_Functions_in_Linear_Regressions\/04_lin_reg_l1_vs_l2.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nfmcclure\/tensorflow_cookbook\/blob\/master\/03_Linear_Regression\/04_Loss_Functions_in_Linear_Regressions\/04_lin_reg_l1_vs_l2.py<\/a>)<\/p>\n\n<p>. but we know abs() is not differentiable.<\/p>\n\n<p>how is this done in Tensorflow? does it just randomly throw a number in [-1,1] ?<\/p>\n\n<p>if someone could please point me to the implementation that would be great. Thanks!<\/p>\n\n<p>(I looked for tensorflow.py in the git, but it does not even exist)<\/p>\n","answers":[{"AnswerId":"41520694","CreationDate":"2017-01-07T11:01:01.593","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p><code>f(x) = abs(x)<\/code> is differentiable everywhere, except at <code>x=0<\/code>. It derivative equals:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AMl8U.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/AMl8U.gif\" alt=\"abs derivative\"><\/a><\/p>\n\n<p>So the only question is how tensorflow implements derivative at <code>x=0<\/code>. You can check this manually:<\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.Variable(0.0)\ny = tf.abs(x)\ngrad = tf.gradients(y, [x])[0]\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(sess.run(grad))\n<\/code><\/pre>\n\n<p>It prints <code>0.0<\/code>.<\/p>\n"}]}
{"QuestionId":41519081,"AnswerCount":1,"Tags":"<python><tensorflow><classification>","CreationDate":"2017-01-07T07:38:41.050","AcceptedAnswerId":41520268.0,"OwnerUserId":7089821.0,"Title":"How to avoid the out of range error using shuffle_batch() function?","Body":"<h2>Updated Question<\/h2>\n\n<p>I am trying to match labels with images in tensorflow with the  function, but there is an <strong>out of range error of RandomShuffleQueue<\/strong> when I startd trainning loop with shuffle_batch() function. <\/p>\n\n<h2>1. My updated question<\/h2>\n\n<p>How to avoid the queue out of range error when using shuffle_batch function?<\/p>\n\n<h2>2.My updated code<\/h2>\n\n<p>The following code works well for the first 90 step or so with increasing accuracy, untill it invokes an error.<\/p>\n\n<pre><\/pre>\n\n<h2>3.Updated running result<\/h2>\n\n<p>The updated code invokes an error on step 90 or so:<\/p>\n\n<pre><\/pre>\n\n<p>then<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<h2>Old Question:<\/h2>\n\n<p><em>Why the label of the image is stuck to the first sample when I am matching labels with images using tf.batch()?<\/em>\nMy old question is solved with great help of Standy and My friend Dong. Thx them a lot!<\/p>\n\n<p>1.My questiones<\/p>\n\n<ol>\n<li><p>Where to put the 'shuffle_batch()'function, on position1 or position2? (you can find it in with tf.session() part of the code below)<\/p><\/li>\n<li><p>How to fix the code to make batch sucessfully combine the image and the label, not stucking on the first example?<\/p><\/li>\n<\/ol>\n\n<p>2.csv file<\/p>\n\n<p>The image name and the relevant label are stored in the label.csv in the following format(withour header):<\/p>\n\n<pre><\/pre>\n\n<p>3.Origin code<\/p>\n\n<p>The purpose of the code is to use RNN as a image classifier.<\/p>\n\n<p>The CNN structure is based on the tensorflow example file.<\/p>\n\n<p>You can focus on the Reading data part and lauch the graph part. <\/p>\n\n<pre><\/pre>\n\n<p>` <\/p>\n\n<p>4.Running result\n - In IPython console<br>\n\n - In varible explore<\/p>\n\n<p>The variable labels and images are unchange as the label and image first example, which is stored  in the first line of the label.csv file <\/p>\n\n<p>Therefore, I infer that the reading file queue has stuck on the first line, causing the CNN quickly converge with the accuracy of 1.<\/p>\n","answers":[{"AnswerId":"41520268","CreationDate":"2017-01-07T10:14:21.210","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p><code>shuffle_batch<\/code> <strong>accepts tensors and returns a tensor, so it is a tensorflow op and should be placed in Graph<\/strong>. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#shuffle_batch\" rel=\"nofollow noreferrer\">shuffle_batch docs<\/a><\/p>\n\n<p>I would place it right after you decode single image:<\/p>\n\n<pre><code>image = tf.image.decode_jpeg(im_content, channels=3)\nimages_batch, labels_batch = tf.train.shuffle_batch([image, label], batch_size, batch_size * 50, batch_size * 10)\n# images_batch is now Tensor of shape (batch_size, height, weight, channels)\n...\nh_conv1 = tf.nn.relu(conv2d(images_batch, W_conv1) + b_conv1)\n<\/code><\/pre>\n\n<p>You don't need variables <code>x<\/code> and <code>y_<\/code> now, and you do not need to manually assign your inputs when you are using <code>tf.train.shuffle_batch<\/code>.<\/p>\n\n<p>It may seem counter-intuitive that <code>tf.train.shuffle_batch<\/code> accepts tensors for single example and produces the whole batch, but remember, that the tensors that you feed to this op are from queue so tf.train.shuffle_batch can \"wait\" for multiple elements (under the hood it actually uses another queue to do shuffling and storing intermediate elements, shuffle_batch implementation is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.12\/tensorflow\/python\/training\/input.py#L811\" rel=\"nofollow noreferrer\">here<\/a>)<\/p>\n"}]}
{"QuestionId":41519297,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-07T08:08:53.600","AcceptedAnswerId":null,"OwnerUserId":1712607.0,"Title":"Tensorflow Inception Train and Create GraphDef","Body":"<p>I'm quite new to using Tensorflow. I was playing around the Inception-V3 model for image recognition in this repo: <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception#how-to-train-from-scratch\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception#how-to-train-from-scratch<\/a><\/p>\n\n<p>It provides binaries to help you train your own model. However in the Tensorflow docs on Image Recognition,<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_recognition\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/image_recognition\/<\/a><\/p>\n\n<p>They allow you to download and utilize a pre-trained Inception-V3 model trained on ImageNet data. The model that you download, is in the format of a  GraphDef whereas the model you generate in the  repo is a  \"checkpoint` formatted file.<\/p>\n\n<p>What's the difference between these two types of files and how can I convert my generated inception model into a GraphDef file?<\/p>\n","answers":[{"AnswerId":"42683217","CreationDate":"2017-03-08T22:35:03.593","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/meta_graph\" rel=\"nofollow noreferrer\">documentation on metagraphs<\/a> should be helpful. It shows to save and load these structures which contain enough information to reproduce the model.<\/p>\n"}]}
{"QuestionId":41519569,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-07T08:45:39.460","AcceptedAnswerId":null,"OwnerUserId":4922660.0,"Title":"How to save and restore ResourceBase?","Body":"<p>I'm add a self-defined ops and using ResourceMgr to a create child class of ResourceBase to store some status, and how to write the status into checkpoint file.<\/p>\n\n<p> <\/p>\n\n<pre><\/pre>\n\n<p>as in code above, how can I write feature_index_  to checkpoint files.<\/p>\n","answers":[{"AnswerId":"43329050","CreationDate":"2017-04-10T17:04:01.937","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There is no generic way to save a <code>tensorflow::ResourceBase<\/code> instance, but you can implement your own checkpointing support as follows:<\/p>\n\n<ol>\n<li><p>Define methods on your <code>FeatureTransformMap<\/code> class that serialize the state of the map to, and deserialize it from, one or more <code>tensorflow::Tensor<\/code> objects. For an example, see  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4e3864cb351da6f639b5adfa5a375d826ade26b4\/tensorflow\/core\/kernels\/lookup_table_op.cc#L212\" rel=\"nofollow noreferrer\"><code>MutableHashTableOfScalars::ExportValues()<\/code><\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4e3864cb351da6f639b5adfa5a375d826ade26b4\/tensorflow\/core\/kernels\/lookup_table_op.cc#L207\" rel=\"nofollow noreferrer\"><code>MutableHashTableOfScalars::ImportValues()<\/code><\/a>.<\/p><\/li>\n<li><p>Implement new TensorFlow <code>OpKernel<\/code> classes that invoke your serialization and deserialization methods. For an example, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4e3864cb351da6f639b5adfa5a375d826ade26b4\/tensorflow\/core\/kernels\/lookup_table_op.cc#L818\" rel=\"nofollow noreferrer\"><code>LookupTableExportOp<\/code><\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4e3864cb351da6f639b5adfa5a375d826ade26b4\/tensorflow\/core\/kernels\/lookup_table_op.cc#L835\" rel=\"nofollow noreferrer\"><code>LookupTableImportOp<\/code><\/a>.<\/p><\/li>\n<li><p>In Python, implement a subclass of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4e3864cb351da6f639b5adfa5a375d826ade26b4\/tensorflow\/python\/training\/saver.py#L103\" rel=\"nofollow noreferrer\"><code>BaseSaverBuilder.SaveableObject<\/code><\/a> for your resource, which includes invocations of your new ops. For an example, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4e3864cb351da6f639b5adfa5a375d826ade26b4\/tensorflow\/contrib\/lookup\/lookup_ops.py#L1650\" rel=\"nofollow noreferrer\"><code>MutableDenseHashTable._Saveable<\/code><\/a>.<\/p><\/li>\n<li><p>In Python, when you create an instance of your resource, add it to the collection of <code>tf.GraphKeys.SAVEABLE_OBJECTS<\/code>. For an example, see how this is done for <code>MutableDenseHashTable<\/code> <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/4e3864cb351da6f639b5adfa5a375d826ade26b4\/tensorflow\/contrib\/lookup\/lookup_ops.py#L1357\" rel=\"nofollow noreferrer\">here<\/a>.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":41519890,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-07T09:26:44.540","AcceptedAnswerId":null,"OwnerUserId":6856923.0,"Title":"How to use TensorFlow in OOP style?","Body":"<p>Specifically, when using TensorFlow to build my model in OOP style, where should I build the graph? Where should I start a session to run the graph? What's the best practice for this case?<\/p>\n\n<p>In <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/tf\/\" rel=\"noreferrer\">TensorFlow Mechanics 101<\/a>, the MNIST example just simply define the ,  and  function in the module  and build the graph in . But in my opinion, the graph is actually part of the model and should be built inside the model, maybe in its  method.<\/p>\n\n<p>I have seen many other models using TensorFlow in its <a href=\"https:\/\/github.com\/tensorflow\/models\" rel=\"noreferrer\">model zoo<\/a> and each have their own practice, so I am a little confused here. Is there a best practice or any recommended programming paradigms when using TensorFlow\uff1f<\/p>\n","answers":[{"AnswerId":"42161675","CreationDate":"2017-02-10T14:22:46.960","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>Also check out a nice article about this topic:<\/p>\n\n<p><a href=\"https:\/\/danijar.com\/structuring-your-tensorflow-models\/\" rel=\"nofollow noreferrer\">https:\/\/danijar.com\/structuring-your-tensorflow-models\/<\/a><\/p>\n\n<p>In this article, Danijar Hafner introduces lazy property:<\/p>\n\n<pre><code>class Model:\n\n    def __init__(self, data, target):\n        self.data = data\n        self.target = target\n        self.prediction\n        self.optimize\n        self.error\n\n    @lazy_property\n    def prediction(self):\n        data_size = int(self.data.get_shape()[1])\n        target_size = int(self.target.get_shape()[1])\n        weight = tf.Variable(tf.truncated_normal([data_size, target_size]))\n        bias = tf.Variable(tf.constant(0.1, shape=[target_size]))\n        incoming = tf.matmul(self.data, weight) + bias\n        return tf.nn.softmax(incoming)\n\n    @lazy_property\n    def optimize(self):\n        cross_entropy = -tf.reduce_sum(self.target, tf.log(self.prediction))\n        optimizer = tf.train.RMSPropOptimizer(0.03)\n        return optimizer.minimize(cross_entropy)\n\n    @lazy_property\n    def error(self):\n        mistakes = tf.not_equal(\n            tf.argmax(self.target, 1), tf.argmax(self.prediction, 1))\n        return tf.reduce_mean(tf.cast(mistakes, tf.float32))\n<\/code><\/pre>\n\n<p>See more in the article.<\/p>\n"},{"AnswerId":"41522177","CreationDate":"2017-01-07T13:53:59.727","ParentId":null,"OwnerUserId":"2876799","Title":null,"Body":"<p>I usually build my graphs in the <strong>init<\/strong> but I sometime create a separate compile function.  I have a unique variable scope for the entire class and the class provided save and restore and init functions for its variables.  I also provide functions to train and predict.  I don't think there is really any standard practice but this makes sense to me.  Here is an example of how I build a generative model with image pyramids.<\/p>\n\n<pre><code>class PyramidGenerator:\n    def __init__(self,\n                 session,\n                 log2_input_size,\n                 log2_output_size,\n                 num_features,\n                 convs_per_cell,\n                 filter_size,\n                 conv_activation,\n                 num_attributes,\n                 name = 'pyrgen'):\n\n        self.session = session\n        self.log2_input_size = log2_input_size\n        self.log2_output_size = log2_output_size\n        self.num_attributes = num_attributes\n\n        if not hasattr(num_features, '__iter__'):\n            num_features = [num_features] * (log2_output_size - log2_input_size)\n        if not hasattr(convs_per_cell, '__iter__'):\n            convs_per_cell = [convs_per_cell] * (log2_output_size - log2_input_size)\n        if not hasattr(filter_size, '__iter__'):\n            filter_size = [filter_size] * (log2_output_size - log2_input_size)\n\n        with tf.variable_scope(name) as scope:\n            self.training_images = tf.placeholder(tf.float32, (None, 2 ** log2_output_size, 2 ** log2_output_size, 3), 'training_images')\n            if num_attributes:\n                self.image_attributes = tf.placeholder(tf.float32, (None, num_attributes))\n            self.seed_images = tf.placeholder(tf.float32, (None, 2 ** log2_input_size, 2 ** log2_input_size, 3), 'seed_images')\n            self.learning_rate = tf.placeholder(tf.float32, (), 'learning_rate')\n\n            self.scope_name = scope.name\n            self.cost = 0\n\n            def _augment(img):\n                img = tf.image.random_flip_left_right(img)\n                return img\n\n            augmented = tf.map_fn(_augment, self.training_images)\n            training_scales = {s:tf.image.resize_area(augmented, (2 ** s, 2 ** s)) for s in range(log2_input_size, log2_output_size + 1)}\n            x_gen = self.seed_images\n            x_train = None\n            if num_attributes:\n                h_gen = h_train = tf.tile(tf.reshape(self.image_attributes, (-1, 1, 1, num_attributes)), (1, 2 ** log2_input_size, 2 ** log2_input_size, 1))\n            else:\n                h_gen = h_train = None\n\n            self.generator_outputs = []\n\n            for n_features, conv_size, n_convs, log2_size in zip(num_features, filter_size, convs_per_cell, range(log2_input_size, log2_output_size)):\n                size = 2 ** log2_size\n                with tf.variable_scope('level_%d' % size) as level_scope:\n                    y_train = training_scales[log2_size + 1]\n                    x_train = training_scales[log2_size]\n\n                    x_train, h_train = ops.sharpen_cell(x_train, h_train, 2, n_features, conv_size, n_convs, conv_activation, 'upsampler')\n                    self.cost += tf.reduce_mean((x_train - y_train) ** 2)\n\n                    level_scope.reuse_variables()\n\n                    x_gen, h_gen = ops.sharpen_cell(x_gen, h_gen, 2, n_features, conv_size, n_convs, conv_activation, 'upsampler')\n                    self.generator_outputs.append(tf.clip_by_value(x_gen, -1, 1))\n\n            with tf.variable_scope('training'):\n                opt = tf.train.AdamOptimizer(self.learning_rate)\n                grads = opt.compute_gradients(self.cost)\n                grads = [(tf.clip_by_value(g, -1.0, 1.0), v) for g, v in grads]\n                self.train_step = opt.apply_gradients(grads)\n\n            self.variables = tf.get_collection(tf.GraphKeys.VARIABLES, self.scope_name)\n            self.init_vars = tf.initialize_variables(self.variables)\n            self.saver = tf.train.Saver(self.variables)\n\n    def save(self, fn):\n        self.saver.save(self.session, fn)\n\n    def restore(self, fn):\n        self.saver.restore(self.session, fn)\n\n    def initialize(self):\n        self.session.run(self.init_vars)\n\n    def train(self, training_images, validation_images = [], learning_rate = 1e-3, batch_size = 32):\n        with ThreadPoolExecutor(max(os.cpu_count(), batch_size)) as exc:\n            def _loadImage(fn):\n                img = cv2.imread(fn, cv2.IMREAD_COLOR)\n                img = cv2.resize(img, (2 ** self.log2_output_size, 2 ** self.log2_output_size))\n                return np.float32(img \/ 128.0 - 1.0)\n\n            def _loadBatch(b):\n                if self.num_attributes:\n                    imgs, attrs = zip(*b)\n                else:\n                    imgs = b\n                    attrs = None\n                imgs = list(exc.map(_loadImage, imgs))\n                return imgs, attrs\n\n            total_cost = 0\n            batches = list(_batch(training_images, batch_size, False))\n            loader = exc.submit(_loadBatch, batches[0])\n            for i in range(len(batches)):\n                imgs, attrs = loader.result()\n                if i &lt; len(batches) - 1:\n                    loader = exc.submit(_loadBatch, batches[i + 1])\n                feed_dict = {self.training_images: imgs, self.learning_rate: learning_rate}\n                if self.num_attributes:\n                    feed_dict.update({self.image_attributes: attrs})\n                total_cost += self.session.run((self.cost, self.train_step), feed_dict)[0]\n                print('Training Batch(%d\/%d) Cost(%e)' % (i + 1, len(batches), total_cost \/ (i + 1)), end = '\\r')\n            print()\n            return total_cost \/ (i + 1)\n\n    def generate_random(self):\n        img = np.clip(np.random.randn(1, 2 ** self.log2_input_size, 2 ** self.log2_input_size, 3), -1, 1)\n        if self.num_attributes:\n            attrs = np.random.choice((1.0, -1.0), size = (1, self.num_attributes))\n            feed = {self.seed_images: img, self.image_attributes: attrs}\n        else:\n            feed = {self.seed_images: img}\n        y = self.session.run(self.generator_outputs, feed)\n        return [img] + y\n\n    def generate_from(self, seed_image):\n        if self.num_attributes:\n            img, attrs = seed_image\n        else:\n            img = seed_image\n        img = cv2.imread(img, cv2.IMREAD_COLOR)\n        img = cv2.resize(img, (2 ** self.log2_input_size, 2 ** self.log2_input_size))\n        img = np.expand_dims(np.float32(img \/ 128.0 - 1.0), 0)\n        if self.num_attributes:\n            feed = {self.seed_images: img, self.image_attributes: [attrs]}\n        else:\n            feed = {self.seed_images: img}\n        y = self.session.run(self.generator_outputs, feed)\n        return [img] + y\n<\/code><\/pre>\n"}]}
{"QuestionId":41521250,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-07T12:05:01.550","AcceptedAnswerId":41526081.0,"OwnerUserId":6552386.0,"Title":"How can I assemble a new Operation from a subgraph, and apply gradient to the newly assembled operation","Body":"<p> I want to assemble a new Operation from a sub-graph (that made of several connected operation nodes). and then apply self-designed gradient on the new Operation. (the point is to ignore the gradient flow in the sub-graph, and bridge gradients from output tensor of the NEW OP to the input tensor of the NEW OP). Hope somebody can help!!<\/p>\n","answers":[{"AnswerId":"41526081","CreationDate":"2017-01-07T20:17:33.950","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You can wrap your subgraph into a TensorFlow function, and specify custom gradient for that function, as done in <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/27711108b5fce2e1692f9440631a183b3808fa01\/tensorflow\/python\/framework\/function_test.py#L136\" rel=\"nofollow noreferrer\">python\/framework\/function_test.py<\/a><\/p>\n\n<pre><code>    @function.Defun(dtype, dtype, dtype)\n    def XentLossGrad(logits, labels, dloss):\n      dlogits = array_ops.reshape(dloss, [-1, 1]) * (\n          nn_ops.softmax(logits) - labels)\n      dlabels = array_ops.zeros_like(labels)\n      # Takes exp(dlogits) to differentiate it from the \"correct\" gradient.\n      return math_ops.exp(dlogits), dlabels\n\n    @function.Defun(dtype, dtype, grad_func=XentLossGrad)\n    def XentLoss(logits, labels):\n      return math_ops.reduce_sum(labels * math_ops.log(nn_ops.softmax(logits)),\n                                 1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41521268,"AnswerCount":0,"Tags":"<tensorflow><training-data><on-the-fly>","CreationDate":"2017-01-07T12:06:39.053","AcceptedAnswerId":null,"OwnerUserId":6001053.0,"Title":"How can I generate training data on the fly in TensorFlow?","Body":"<p>I want to create the Moving MNIST training set in TensorFlow. I've been thinking about creating a subclass of  and emitting  training sequences in  or using , but I am not sure whether this is the right way to approach it and whether it is a good idea in the first place (regarding CPU load for creating the sequences during training).<\/p>\n","answers":[]}
{"QuestionId":41521962,"AnswerCount":1,"Tags":"<tensorflow><bazel><magenta>","CreationDate":"2017-01-07T13:29:19.720","AcceptedAnswerId":null,"OwnerUserId":7387730.0,"Title":"Bazel error: \"No test targets were found, yet testing was requested\"","Body":"<p>I have a small assignment where I used TensorFlow to create music:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/magenta\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/magenta<\/a><\/p>\n\n<p>When I run the code \nI get the following error:<\/p>\n\n<pre>\nWARNING: \/home\/admin\/.cache\/bazel\/_bazel_admin\/fb30f33370a5b97d4f9b1dde06f8f344\/external\/protobuf\/protobuf.bzl:90:19: Variables HOST_CFG and DATA_CFG are deprecated in favor of strings \"host\" and \"data\" correspondingly.\nWARNING: \/home\/admin\/.cache\/bazel\/_bazel_admin\/fb30f33370a5b97d4f9b1dde06f8f344\/external\/protobuf\/protobuf.bzl:96:28: Variables HOST_CFG and DATA_CFG are deprecated in favor of strings \"host\" and \"data\" correspondingly.\nINFO: Found 2 targets and 0 test targets...\nINFO: Elapsed time: 4.977s, Critical Path: 0.66s\nERROR: No test targets were found, yet testing was requested.<\/pre>\n","answers":[{"AnswerId":"41527095","CreationDate":"2017-01-07T22:05:31.400","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>When you run<\/p>\n\n<pre><code>bazel test magenta:all\n<\/code><\/pre>\n\n<p>This means \"execute all *_test rules defined in file magenta\/BUILD.\nWhen I look at that file, there are no tests defined there.\n<a href=\"https:\/\/github.com\/tensorflow\/magenta\/blob\/master\/magenta\/BUILD\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/magenta\/blob\/master\/magenta\/BUILD<\/a><\/p>\n\n<p>You should try:<\/p>\n\n<pre><code>bazel test magenta\/...\n<\/code><\/pre>\n\n<p>This translates to all things that are included in magenta folder, including other packages. For more information, please see:\n<a href=\"https:\/\/bazel.build\/versions\/master\/docs\/command-line-reference.html\" rel=\"nofollow noreferrer\">https:\/\/bazel.build\/versions\/master\/docs\/command-line-reference.html<\/a><\/p>\n"}]}
{"QuestionId":41522354,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-07T14:13:42.930","AcceptedAnswerId":null,"OwnerUserId":7387749.0,"Title":"Tensorflow: How to find the correct path to save model parameters - model_dir","Body":"<p>I am a beginner in TensorFlow. I am trying to use <em>tf.contrib.learn.DNNClassifier<\/em>, but I don't know how to find the path to insert in <em>model_dir<\/em>.\nI have tried this:<\/p>\n\n<pre><\/pre>\n\n<p>And I have this error:<\/p>\n\n<pre><\/pre>\n\n<p>Which is the correct path to create the directory?<\/p>\n","answers":[{"AnswerId":"41524274","CreationDate":"2017-01-07T17:20:58.437","ParentId":null,"OwnerUserId":"1325423","Title":null,"Body":"<p>You are providing linux file system naming  \"\/My_folder\" as model_dir so python looks My_folder within the current working directory (cwd) in its run time environment. On Windows file system if you provide a full paths i.e. \"c:\\folder1\\folder2folder3\" then python will be absolute path given as model_dir. <\/p>\n\n<p>What you really need is to first create a separate directory in your windows machine which is write permission for your current user profile. Use python os module to check you can save filed from python runtime to this folder. <\/p>\n\n<p>Once above folder is confirmed to have write permission, just use that path will full absolute name and model will be saved outside python run time directory. <\/p>\n"}]}
{"QuestionId":41522364,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-07T14:15:30.610","AcceptedAnswerId":44654244.0,"OwnerUserId":1616677.0,"Title":"Incremental model training with tensorflow","Body":"<p>I have a simple linear model which inputs (x, y) pairs and deduces b0 and b1 in y = b0 + b1 * x ; the key code is below.\nIt trains on a dataset of known size. Now I want to add the ability of training it constantly: i.e. add every other batch of (x, y), and update cofficients according to the new data. There will be unlimited amount of input. <\/p>\n\n<pre><\/pre>\n\n<p>My idea is to remember so-far-trained coefficients, init a model with them, then just repeat again the training with the new portion of data. Repeat on each input. Is this a right way to go? The model will probably be promoted later to something more complex..<\/p>\n","answers":[{"AnswerId":"44654244","CreationDate":"2017-06-20T13:19:10.300","ParentId":null,"OwnerUserId":"6907100","Title":null,"Body":"<p>It sounds like you're talking about on-line training, i.e. continuously train a model with incoming data while simultaneously using it.  You're right in that you should be able to pick up where you left off and just feed in new data.  What you'll need is a way to save and load the variables between training sessions.   You can use a tf.Saver to do this in \"raw\" tensorflow.  <\/p>\n\n<p>You can also use a tf.contrib.learn.Estimator to do this for you.  You just give it a model_fn that constructs your model, and a model_dir to save the model in, and it will take care of the rest.  Of course, there's already a linear model in tf.contrib.learn.LinearEstimator.  With estimators, you'd just call fit(...) whenever you have new data and it will load your variables and continue running the training steps you've defined.<\/p>\n"}]}
{"QuestionId":41522517,"AnswerCount":4,"Tags":"<tensorflow><neural-network><deep-learning><conv-neural-network>","CreationDate":"2017-01-07T14:31:48.450","AcceptedAnswerId":null,"OwnerUserId":2682877.0,"Title":"Custom median pooling in tensorflow","Body":"<p>I am trying to implement a median pooling layer in tensorflow.<\/p>\n\n<p>However there is neither  and neither .<\/p>\n\n<p>Is there a way to implement such pooling layer with the python api ?<\/p>\n","answers":[{"AnswerId":"52743040","CreationDate":"2018-10-10T14:53:01.470","ParentId":null,"OwnerUserId":"9606634","Title":null,"Body":"<p>For me, Alex's answer is not working for tf 1.4.1. \n<code>tf.top_k<\/code> should be <code>tf.nn.top_k<\/code>\nand should get values of <code>tf.nn.top_k<\/code><\/p>\n\n<p>Also, if the input is [1, H, W, C], either answer could not only work on height and width  and neglect the channel.<\/p>\n"},{"AnswerId":"43554072","CreationDate":"2017-04-22T01:00:04.817","ParentId":null,"OwnerUserId":"4902348","Title":null,"Body":"<p>You could use something like:<\/p>\n\n<pre><code>patches = tf.extract_image_patches(tensor, [1, k, k, 1], ...)\nm_idx = int(k*k\/2+1)\ntop = tf.top_k(patches, m_idx, sorted=True)\nmedian = tf.slice(top, [0, 0, 0, m_idx-1], [-1, -1, -1, 1])\n<\/code><\/pre>\n\n<p>To accommodate even sized median kernels and multiple channels, you will need to extend this, but this should get you most of the way.<\/p>\n"},{"AnswerId":"55633334","CreationDate":"2019-04-11T13:12:09.343","ParentId":null,"OwnerUserId":"688080","Title":null,"Body":"<p>Channel-wise median-pooling can be done by some addition reshapes on top of other answers:<\/p>\n\n<pre><code># assuming NHWC layout\nstrides = rates = [1, 1, 1, 1]\npatches = tf.extract_image_patches(x, [1, k, k, 1], strides, rates, 'VALID')\nbatch_size = tf.shape(x)[0]\nn_channels = tf.shape(x)[-1]\nn_patches_h = (tf.shape(x)[1] - k) \/\/ strides[1] + 1\nn_patches_w = (tf.shape(x)[2] - k) \/\/ strides[2] + 1\nn_patches = tf.shape(patches)[-1] \/\/ n_channels\npatches = tf.reshape(patches, [batch_size, k, k, n_patches_h * n_patches_w, n_channels])\nmedians = tf.contrib.distributions.percentile(patches, 50, axis=[1,2])\nmedians = tf.reshape(medians, (batch_size, n_patches_h, n_patches_w, n_channels))\n<\/code><\/pre>\n\n<p>Not very efficient though.<\/p>\n"},{"AnswerId":"47586632","CreationDate":"2017-12-01T04:26:21.300","ParentId":null,"OwnerUserId":"5775429","Title":null,"Body":"<p>As of March 2017, an easier answer (that under the hood works similarly to how Alex suggested) is to do this:<\/p>\n\n<pre><code>patches = tf.extract_image_patches(x, [1, k, k, 1], [1, k, k, 1], 4*[1], 'VALID')\nmedians = tf.contrib.distributions.percentile(patches, 50, axis=3)\n<\/code><\/pre>\n"}]}
{"QuestionId":41523005,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-01-07T15:22:50.313","AcceptedAnswerId":41523423.0,"OwnerUserId":7216602.0,"Title":"How to use TensorBoard in a Docker container (on Windows)","Body":"<p>I have installed tensorflow on windows through docker toolbox. Everything goes well except I can't use tensorboard. The command line shows 'Starting Tensorboard 29 on port 6006. You can navigate to <a href=\"http:\/\/localhost:6006\/\" rel=\"noreferrer\">http:\/\/localhost:6006\/<\/a>'.However, when I opened this address on my webbrowser, it just can not connect to it. Does anyone know how to solve this problem?<\/p>\n","answers":[{"AnswerId":"41523423","CreationDate":"2017-01-07T16:06:06.330","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>If you're running TensorBoard inside a Docker container, and trying to use a web browser in Windows to view it, you will need to set up port forwarding from the container to your Windows machine. See <a href=\"https:\/\/stackoverflow.com\/a\/33838211\/3574081\">this answer<\/a> for a longer discussion about port forwarding for TensorBoard, but you should be able to make progress by using the following command:<\/p>\n\n<pre><code>docker run -p 0.0.0.0:6006:6006 -it b.gcr.io\/tensorflow\/tensorflow\n<\/code><\/pre>\n\n<p><strong>However,<\/strong> it may be easier to install TensorFlow directly on Windows, and run TensorBoard there. If you install Python 3.5 for Windows, you can install TensorFlow and TensorBoard by running:<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>You can then run TensorBoard directly from the command prompt, and you will not need to worry about port forwarding. See the <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#pip_installation_on_windows\" rel=\"nofollow noreferrer\">Windows installation instructions<\/a> for more details.<\/p>\n"},{"AnswerId":"45212288","CreationDate":"2017-07-20T10:34:57.723","ParentId":null,"OwnerUserId":"5015616","Title":null,"Body":"<p>Maybe you should map your volumes to the folder with the logs and enter with bash well:<\/p>\n\n<pre><code>docker run -v \/\/c\/pathto\/tf_logs:\/tf_logs  \n-p 0.0.0.0:6006:6006 -p 8888:8888 -it b.gcr.io\/tensorflow\/tensorflow bash\ncd ..\ntensorboard --logdir tf_logs\/\n<\/code><\/pre>\n\n<p>hit the mapping in your browser<\/p>\n\n<p><a href=\"http:\/\/192.168.99.100:6006\" rel=\"nofollow noreferrer\">http:\/\/192.168.99.100:6006<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/iOG9t.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iOG9t.jpg\" alt=\"see your graph\"><\/a><\/p>\n"}]}
{"QuestionId":41523886,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-07T16:48:32.300","AcceptedAnswerId":41523939.0,"OwnerUserId":3102085.0,"Title":"How to perform thresholding on a tensor","Body":"<p>Following is my code:<\/p>\n\n<pre><\/pre>\n\n<p>If the value in a tensor is less than 0.5 then it should be replaced with 0 otherwise 1. But it keeps returning the error.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41523939","CreationDate":"2017-01-07T16:52:48.803","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p>In this simple case, you can use<\/p>\n\n<pre><code>result = tf.cast(result + 0.5, tf.int32)\n<\/code><\/pre>\n\n<p>When if-statement becomes more complex, consider using <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops\/control_flow_operations#cond\" rel=\"nofollow noreferrer\"><code>tf.cond<\/code><\/a><\/p>\n"}]}
{"QuestionId":41524393,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><conda>","CreationDate":"2017-01-07T17:33:13.513","AcceptedAnswerId":41736031.0,"OwnerUserId":5173924.0,"Title":"Link Tensorflow from usr\/bin\/python to a conda environment python","Body":"<p>I have Tensorflow with GPU installed on my main usr\/bin\/python, e.g. on an AWS EC2 machine and I'd like to use it within a conda environment. Is this possible without reinstalling tensorflow for each environment ?<\/p>\n\n<p>EDIT: I found out about the virtual-env --system-site-packages option. Maybe an equivalent for Conda ? And especially for packages like Tensorflow-gpu that need to be built for the specific hardware, and are not available directly on pip.<\/p>\n","answers":[{"AnswerId":"41736031","CreationDate":"2017-01-19T07:23:34.737","ParentId":null,"OwnerUserId":"1005215","Title":null,"Body":"<p>If tensorflow is installed in the system, then it will probably be present in one of these four directories:<\/p>\n\n<pre><code>\/usr\/local\/lib\/python2.7\/dist-packages\n\/usr\/local\/lib\/python2.7\/site-packages\n\/usr\/lib\/python2.7\/dist-packages\n\/usr\/lib\/python2.7\/site-packages\n<\/code><\/pre>\n\n<p>To have the system packages also be available in conda, you can do the following:<\/p>\n\n<pre><code>(root) ~\/condaexpts $ cat &lt;&lt;EOF &gt; $CONDA_PREFIX\/lib\/python2.7\/site-packages\/systempkgs.pth\n\/usr\/local\/lib\/python2.7\/dist-packages\n\/usr\/local\/lib\/python2.7\/site-packages\n\/usr\/lib\/python2.7\/dist-packages\n\/usr\/lib\/python2.7\/site-packages\nEOF\n<\/code><\/pre>\n\n<p>Then verify that these paths have been added:<\/p>\n\n<pre><code>(root) ~\/condaexpts $ python\nPython 2.7.12 |Continuum Analytics, Inc.| (default, Jul  2 2016, 17:42:40) \n[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\nAnaconda is brought to you by Continuum Analytics.\nPlease check out: http:\/\/continuum.io\/thanks and https:\/\/anaconda.org\n&gt;&gt;&gt; import sys\n&gt;&gt;&gt; sys.path\n['', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python27.zip', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/plat-linux2', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/lib-tk', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/lib-old', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/lib-dynload', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/site-packages', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/site-packages\/setuptools-27.2.0-py2.7.egg', '\/usr\/local\/lib\/python2.7\/dist-packages', '\/usr\/local\/lib\/python2.7\/site-packages', '\/usr\/lib\/python2.7\/dist-packages']\n<\/code><\/pre>\n\n<p>And now you can import any system package.<\/p>\n\n<p>If you don't want this change to be permanent for your conda environment, you can always add the directory containing system packages to <code>PYTHONPATH<\/code> environment variable:<\/p>\n\n<pre><code>(root) ~\/condaexpts $ PYTHONPATH=$PYTHONPATH:\/usr\/local\/lib\/python2.7\/dist-packages\/ python\nPython 2.7.12 |Continuum Analytics, Inc.| (default, Jul  2 2016, 17:42:40) \n[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\nAnaconda is brought to you by Continuum Analytics.\nPlease check out: http:\/\/continuum.io\/thanks and https:\/\/anaconda.org\n&gt;&gt;&gt; import sys\n&gt;&gt;&gt; sys.path\n['', '\/home\/ubuntu\/condaexpts', '\/usr\/local\/lib\/python2.7\/dist-packages', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python27.zip', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/plat-linux2', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/lib-tk', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/lib-old', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/lib-dynload', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/site-packages', '\/home\/ubuntu\/condaexpts\/m2\/lib\/python2.7\/site-packages\/setuptools-27.2.0-py2.7.egg']\n<\/code><\/pre>\n"}]}
{"QuestionId":41524862,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-07T18:18:52.457","AcceptedAnswerId":null,"OwnerUserId":5974525.0,"Title":"TensorFlow why my cost function doesn't decrease?","Body":"<p>I'm using a very simple NN with a normalized word2vec as input.\nWhen running my train (based on the mini batch) the train cost start around 1020 and decrease around 1000 but never less than this and my accuracy is around 50%.<\/p>\n\n<p>Why doesn't the cost decrease ? How can I verify that the weigth matrice is updated at each run?<\/p>\n\n<pre><\/pre>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41525947","CreationDate":"2017-01-07T20:07:01.693","ParentId":null,"OwnerUserId":"7388732","Title":null,"Body":"<p>try using <a href=\"https:\/\/en.wikipedia.org\/wiki\/Cross_entropy\" rel=\"nofollow noreferrer\">cross entropy<\/a> instead of the L2 loss, also there is no real point in having an activation function on your output layer.<\/p>\n\n<p>The examples that ship with tensorflow actually have a <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/2_BasicModels\/logistic_regression.py\" rel=\"nofollow noreferrer\">basic model<\/a> that is very similar to what you are trying.<\/p>\n\n<p>btw: it might also be that the problem you are trying to learn is simply not solvable by a simple linear model (i.e. what you are trying to do), try using a deeper model. Here is an example of a 2 layer deep <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/multilayer_perceptron.py\" rel=\"nofollow noreferrer\">multilayer perceptron<\/a>.<\/p>\n"}]}
{"QuestionId":41524909,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-07T18:22:25.317","AcceptedAnswerId":41525110.0,"OwnerUserId":6842947.0,"Title":"Trouble with variable initialization","Body":"<p>I am building a graph via a function and am trying to extract the value of a variable to add further operations. A part of the function I have written is shown below :<\/p>\n\n<pre><\/pre>\n\n<p>At the line  i get the following error :<\/p>\n\n<pre><\/pre>\n\n<p>What is the reason for this error when I have run the init_op operation in the session before ? Exactly how does this work and what am I doing wrong over here ?<\/p>\n","answers":[{"AnswerId":"41525110","CreationDate":"2017-01-07T18:41:56.147","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p>You need to define your <code>init_op<\/code> (i.e. call tf.initialize_all_variables()) after you declared all variables.\nCreating a variable via <code>tf.get_variable<\/code> or <code>tf.Variable<\/code> places it in GLOBAL_VARIABLES collection (unless otherwise specified with <code>collections<\/code> kwarg). <code>tf.initialize_all_variables()<\/code> takes a look at this collection and creates an op that initializes variables listed.<\/p>\n\n<p>To see GLOBAL_VARIABLES collection, you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/graph_collections#get_collection\" rel=\"nofollow noreferrer\"><code>tf.get_collection<\/code><\/a> with <code>tf.GraphKeys.GLOBAL_VARIABLES<\/code> as argument.<\/p>\n\n<p><strong>TL;DR<\/strong> Place <code>init_op = tf.initialize_all_variables()<\/code> after the graph was created.<\/p>\n"}]}
{"QuestionId":41525251,"AnswerCount":1,"Tags":"<audio><mobile><tensorflow><signal-processing>","CreationDate":"2017-01-07T18:56:41.980","AcceptedAnswerId":null,"OwnerUserId":257093.0,"Title":"Feeding real-time audio data to tensorflow on a mobile device","Body":"<p>I am building a prototype of a sound detection app that will ultimately run on a phone (iPhone\/Android). It needs to be near real-time to give fast enough response to the user when a particular sound is recognized. I am hoping to use tensorflow to actually build and train the model and then deploy it on mobile device.<\/p>\n\n<p>What I am unsure about is best way to feed data to tensorflow for inference in this case. <\/p>\n\n<p>Option 1: Feed only newly acquired samples to the model. <\/p>\n\n<p>Here the model itself keeps a buffer of previous signal samples, to which new samples are appended and the whole thing get processed.\nSomething like:<\/p>\n\n<pre><\/pre>\n\n<p>This seems to work, but if the samples are pushed to the model 100 times a second, what's happening inside tf.assign (the buffer can grow big enough, and if tf.assign constantly allocates memory this may not work well)? <\/p>\n\n<p>Option 2: Feed the whole recording to the model<\/p>\n\n<p>Here the iPhone app keeps the state\/recording samples, and feeds the whole recording to the model. The input can get quite large, and re-running the detection op on the whole recording will have to keep recomputing the same values each cycle.<\/p>\n\n<p>Option 3: Feed a sliding window of data<\/p>\n\n<p>Here the app keeps the data for the whole recording, but feeds only the latest slice of data to the model. E.g. last 2 sec at 2000 sampling rate == 4000 sample fed fed at the rate of 1\/100 sec (each new 20 samples). The model may also need to keep some running totals for the whole recording.<\/p>\n\n<p>Advise?<\/p>\n","answers":[{"AnswerId":"42989313","CreationDate":"2017-03-24T00:24:22.327","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>I'd need to know a bit more about your application requirements, but for simplicities sake I recommend starting with option #3. The usual way to approach this problem for arbitrary sounds is:<\/p>\n\n<ul>\n<li>Have some trigger to detect the start of a sound or speech utterance. This can just be sustained audio levels, or something more advanced.<\/li>\n<li>Run a spectrogram over a fixed size window, aligned with the start of the noise.<\/li>\n<li>The rest of the network can just be a standard image detection one (usually cut down in size) to classify the sound.<\/li>\n<\/ul>\n\n<p>There are a lot of variations and other possible approaches. For example for speech it's typical to use MFCC as your feature generator, and then run an LSTM to separate out phonemes, but since you mention sound detection I'm guessing you don't need anything this advanced.<\/p>\n"}]}
{"QuestionId":41525409,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-07T19:11:27.467","AcceptedAnswerId":41525764.0,"OwnerUserId":4394807.0,"Title":"how does tensorflow assign Ops to run on GPU?","Body":"<p>I'm confused about the mechanism that tensorflow uses to assign different Ops to CPUs or GPUs. <\/p>\n\n<ol>\n<li><p>Taking below pseudo code as an example. Can we say: as long as the  is\ncreated within the context of , it will\nsurely run on GPU (suppose the GPU implementation of the \nis available), <strong>no matter<\/strong> its input variables ( and ) are\ncreated on CPU or GPU? <\/p>\n\n<pre><\/pre><\/li>\n<li><p>I understand by creating a  with\n    , tensorflow outputs the device\n    placements of all variables\/Ops. However, is there a method allowing\n    me to check only one Op's device assignment?<\/p><\/li>\n<\/ol>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"41525662","CreationDate":"2017-01-07T19:38:39.603","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<ol>\n<li><p>Yes. In fact, it will fail, if no kernel is available for specified device. But two things should be considered:<\/p>\n\n<ul>\n<li>This can be overridden with <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/protobuf\/config.proto#L214\" rel=\"nofollow noreferrer\"><code>allow_soft_placement=True<\/code><\/a> if session's <code>config<\/code>.<\/li>\n<li><code>tf.device<\/code> context managers can be nested, so if <code>SimpleOp<\/code> wasn't so simple, it might have wrapped some of its parts with <code>with tf.device(\"\/cpu:0\"):<\/code><\/li>\n<\/ul><\/li>\n<li><p>Not that I know of (comments are welcome). You can always <code>grep<\/code> your script output, if you are using *nix, <code>python script.py | grep your_op_name<\/code>. The downside is that you'll need to rerun your script twice: first run with <code>log_device_placement=True<\/code> and <code>grep<\/code>, second run without them.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"41525764","CreationDate":"2017-01-07T19:49:09.100","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>TLDR; your op created in <code>with tf.device(\"\/gpu:0\")<\/code> will always run on GPU. If you specify input to be placed on <code>cpu<\/code>, then they will get placed on CPU. If you omit device specifications for inputs, they will get placed on GPU to be closer to your op. You can use <code>run_metadata<\/code> to get a Python object with all device assignments, and lookup your op there. <\/p>\n\n<p>Placement is done by misleadingly named <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/beb10ceb086fe94a6b1247b45397aafddd47e05d\/tensorflow\/core\/common_runtime\/simple_placer.cc\" rel=\"noreferrer\">simple_placer.cc<\/a>, and while the comments specify the mechanics, there are still some bugs getting hashed out (ie, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/6615#issuecomment-271051156\" rel=\"noreferrer\">here<\/a>), so the best way is to check it in practice.<\/p>\n\n<p>When you say that variables are created on GPU, there's actually two kinds of placement -- explicit, when you create the relevant op inside the <code>with tf.device<\/code> block, and implicit, outside of such block. Creating ops outside of <code>with tf.device<\/code> is equivalent to creating ops in <code>with tf.device(None)<\/code>  block.<\/p>\n\n<p>So here's a simple experiment<\/p>\n\n<pre><code>n = 10**6\ndef inputs_cpu():\n    tf.reset_default_graph()\n    with tf.device(\"\/cpu:0\"):\n        a = tf.ones((n,), name=\"A\")\n        b = tf.ones((n,), name=\"B\")\n    with tf.device(\"\/gpu:0\"):\n        c = tf.add(a, b, name=\"C\")\n    return c\n\ndef inputs_none():\n    tf.reset_default_graph()\n    a = tf.ones((n,), name=\"A\")\n    b = tf.ones((n,), name=\"B\")\n    with tf.device(\"\/gpu:0\"):\n        c = tf.add(a, b, name=\"C\")\n    return c\n\ndef run_and_summarize(target):\n    # turn off graph-rewriting optimizations\n    sess = tf.Session(config=tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0))))\n    run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)\n    run_metadata = tf.RunMetadata()\n    sess.run(target, options=run_options, run_metadata=run_metadata)\n\n    for device in run_metadata.step_stats.dev_stats:\n        device_name = device.device\n        if not (device_name.endswith(\"\/cpu:0\") or device_name.endswith(\"\/gpu:0\")):\n            continue\n        print(device.device)\n        for node in device.node_stats:\n            print(\"   \", node.node_name)\n<\/code><\/pre>\n\n<p>Now you can do this<\/p>\n\n<pre><code>run_and_summarize(inputs_cpu())\n<\/code><\/pre>\n\n<p>That runs with inputs pinned to CPU and you'll see this placement is respected<\/p>\n\n<pre><code>\/job:localhost\/replica:0\/task:0\/gpu:0\n    _SOURCE\n    C\n\/job:localhost\/replica:0\/task:0\/cpu:0\n    _SOURCE\n    A\n    B\n<\/code><\/pre>\n\n<p>On other hand when inputs are not specified<\/p>\n\n<pre><code>run_and_summarize(inputs_none())\n<\/code><\/pre>\n\n<p>You can see that now all ops are placed on GPU<\/p>\n\n<pre><code>\/job:localhost\/replica:0\/task:0\/cpu:0\n    _SOURCE\n\/job:localhost\/replica:0\/task:0\/gpu:0\n    _SOURCE\n    A\n    B\n    C\n<\/code><\/pre>\n"}]}
{"QuestionId":41525774,"AnswerCount":1,"Tags":"<image><tensorflow><google-cloud-dataflow><gcloud><google-cloud-ml>","CreationDate":"2017-01-07T19:50:16.513","AcceptedAnswerId":41526600.0,"OwnerUserId":2984051.0,"Title":"Google CloudML Permissions from Dataflow - cannot recreate google tensorflow image example","Body":"<p>I am trying to follow this blog post from Google using their new CLOUDML tools.<\/p>\n\n<p><a href=\"https:\/\/cloud.google.com\/blog\/big-data\/2016\/12\/how-to-train-and-classify-images-using-google-cloud-machine-learning-and-cloud-dataflow\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/blog\/big-data\/2016\/12\/how-to-train-and-classify-images-using-google-cloud-machine-learning-and-cloud-dataflow<\/a><\/p>\n\n<p>Running from within their provided docker instance<\/p>\n\n<pre><\/pre>\n\n<p>starting from:\n<\/p>\n\n<p>To the run the first preprocessing step:<\/p>\n\n<h2>Assign appropriate values.<\/h2>\n\n<pre><\/pre>\n\n<h2>Preprocess the eval set.<\/h2>\n\n<pre><\/pre>\n\n<p>returns<\/p>\n\n<pre><\/pre>\n\n<p>Heading to the console, the logs read:<\/p>\n\n<pre><\/pre>\n\n<p>I tried again authenticating with <\/p>\n\n<pre><\/pre>\n\n<p>and getting the key from the browser. Nothing seems wrong there.<\/p>\n\n<p>I have successfully run the MNIST cloud learning tutorial, so there is no  authentication issues communicating with google compute engine.<\/p>\n\n<p>I can confirm the path to my bucket is correct:<\/p>\n\n<pre><\/pre>\n\n<p>but no folder flowers__20170106_165608 is ever created (due to permissions).<\/p>\n\n<p>Does Dataflow need seperate credentials? I went to the console and made sure my account is open to the dataflow API. Anything beyond<\/p>\n\n<pre><\/pre>\n\n<p>Edit: To show that the service accounts tab on the console. \n<a href=\"https:\/\/i.stack.imgur.com\/xqrMl.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xqrMl.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Edit: Accepted answer below. I'm accepting this answer because Jeremy Lewi is correct. The problem is not that dataflow does have permissions, but because the GCS object was never created. Going into the preprocess logger you can see<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dbR9f.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dbR9f.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The tutorial google has shown is probably not well configured for the free tier, i'm guessing it distributes to too many instances and exceeds the CPU quota. If i cannot solve, I will open a correctly framed question.<\/p>\n","answers":[{"AnswerId":"41526600","CreationDate":"2017-01-07T21:10:43.630","ParentId":null,"OwnerUserId":"4392693","Title":null,"Body":"<p>Please see the information about service accounts at the link provided by the error message.  I suspect the service account is not authorized correctly to view the staged file.<\/p>\n"}]}
{"QuestionId":41526071,"AnswerCount":2,"Tags":"<tensorflow><neural-network><out-of-memory><gpu><keras>","CreationDate":"2017-01-07T20:17:10.783","AcceptedAnswerId":null,"OwnerUserId":2145297.0,"Title":"Why is Keras throwing a ResourceExhaustedError?","Body":"<p>Keras is throwing a ResourceExhaustedError when training a convolutional autoencoder. I'm running the Tensorflow backend. The computer has both an Nvidia Tesla with 11 Gbs of memory and an Nvidia Quadro with 6 Gbs of memory. It seems like Tensorflow is using both GPUs? But I'm not too clear on that. Here is a minimal example of the code I'm using. In my example data is a numpy array of dimension=(100,1080,1920,1).<\/p>\n\n<pre><\/pre>\n\n<p>It seems like the GPUs are running out of memory. The autoencoder has 2625 variables. So that doesn't seem like enough to fill up the video ram. The array data has a size of 1600 MBs. So that also shouldn't fill up the video ram. I'm guessing the problem has something to do with nb_epoch and batch_size, but I'm not entirely clear on what those parameters do. Is there a way alter those parameters to fix my problem?<\/p>\n","answers":[{"AnswerId":"46032606","CreationDate":"2017-09-04T07:55:27.900","ParentId":null,"OwnerUserId":"7148586","Title":null,"Body":"<p>change <code>Convolution2D<\/code> - > <code>Conv2D<\/code><br>\nrefer to : <a href=\"https:\/\/stackoverflow.com\/a\/46032412\/7148586\">https:\/\/stackoverflow.com\/a\/46032412\/7148586<\/a><\/p>\n"},{"AnswerId":"41527523","CreationDate":"2017-01-07T22:59:26.950","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Let's try to approximate how much of memory do you need for this network (despite the memory for parameters):<\/p>\n\n<ol>\n<li>Input: 1600MB<\/li>\n<li>First layer : ~ 1600MB * 16 \/ 4 = 6400MB (you have 16 times more channels and pooling layer - decreases the size of its input by 4).<\/li>\n<li>Second layer : 6400MB (the input and the output of a layer is the same).<\/li>\n<li>Third Layer : 25600MB (your are increasing the amount of data 4 times).<\/li>\n<li>Output layer : 1600MB (the same as input).<\/li>\n<\/ol>\n\n<p>You may see that third layer alone need more than 11GB of memory. Moreover - all forward values are stored - for a backpropagation algorithm - so in fact you need to sum up all of this values in order to obtain the final estimation of a memory need. <\/p>\n"}]}
{"QuestionId":41526260,"AnswerCount":1,"Tags":"<python><serialization><neural-network><theano><keras>","CreationDate":"2017-01-07T20:35:57.547","AcceptedAnswerId":41527441.0,"OwnerUserId":3512709.0,"Title":"model.save_weights and model.load_weights not working as expected","Body":"<p>I'm new to machine learning and am going through the courses on <a href=\"http:\/\/course.fast.ai\/index.html\" rel=\"nofollow noreferrer\">fast.ai<\/a>. We're learning about vgg16, and I'm having trouble saving my model. I wonder what I'm doing wrong. When I start my model from scratch, training to learn the difference between cats and dogs, I get:<\/p>\n\n<pre><\/pre>\n\n<p>But now when I go to load one of the weight files, the model starts from scratch! For example, I would have expected the model below to have a val_acc of 0.9750! Am I misunderstanding something or doing something wrong? Why is the val_acc so low with this loaded model?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41527441","CreationDate":"2017-01-07T22:49:08.903","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The problem lies in a <code>finetune<\/code> function. When you get deeper into its definition:<\/p>\n\n<pre><code>def finetune(self, batches):\n    model = self.model\n    model.pop()\n    for layer in model.layers: layer.trainable=False\n    model.add(Dense(batches.nb_class, activation='softmax'))\n    self.compile()\n<\/code><\/pre>\n\n<p>... one can see that by calling <code>pop<\/code> function - the last layer of your model is deleted. By doing this you are losing information from a trained model. The last layer is added again with random weights and then training starts again. This is the reason behind the accuracy drop.<\/p>\n"}]}
{"QuestionId":41526858,"AnswerCount":2,"Tags":"<python><tensorflow><imagenet>","CreationDate":"2017-01-07T21:39:13.490","AcceptedAnswerId":null,"OwnerUserId":7388993.0,"Title":"Inception retraining issue \"Nan in summary histogram for: HistogramSummary\"","Body":"<p>I'm trying to retrain inceptionV3 on my RPi3. I'm getting this histogram error message.<\/p>\n\n<pre><\/pre>\n\n<p>I tried changing  in  after reading <a href=\"https:\/\/stackoverflow.com\/questions\/36666331\/learning-rate-larger-than-0-001-results-in-error\">this<\/a> \nbut it didnt work.<\/p>\n\n<p>Also, the first time I tried to retrain, I got different results for step 0 before hitting an error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44211784","CreationDate":"2017-05-27T00:36:51.747","ParentId":null,"OwnerUserId":"1004331","Title":null,"Body":"<p>If you're using tf.contrib.learn you'll want to use the following:<\/p>\n\n<pre><code>debug_hook = tf_debug.LocalCLIDebugHook()\ndebug_hook.add_tensor_filter(\"has_inf_or_nan\", tf_debug.has_inf_or_nan)\nhooks = [debug_hook]\n...\nclassifier.fit(..., monitors=hooks)\n<\/code><\/pre>\n"},{"AnswerId":"41529428","CreationDate":"2017-01-08T04:22:11.410","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>Sounds like that it might help to know where the NaN values are coming from. For that, take a look at tensorflow debugger (tfdbg): \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/how_tos\/debugger\/index.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/how_tos\/debugger\/index.md<\/a><\/p>\n\n<p>In your retrain.py, you can make a change like<\/p>\n\n<pre><code>from tensorflow.python import debug as tf_debug\n\n# ... \n# In def main(_)\nif debug:\n  sess = tf_debug.LocalCLIDebugWrapperSession(sess)\n  sess.add_tensor_filter(\"has_inf_or_nan\", tf_debug.has_inf_or_nan)\n\n# ...\n<\/code><\/pre>\n\n<p>Then when the <code>sess.run()<\/code> happens for the training and evaluation, you will  drop into the command-line interface of the debugger. At the <code>tfdbg&gt;<\/code> prompt, you can enter command to let the code run until any NaNs or Infinities appear in the TensorFlow graph:<\/p>\n\n<pre><code>tfdbg&gt; run -f has_inf_or_nan\n<\/code><\/pre>\n\n<p>When the tensor filter <code>has_inf_or_nan<\/code> is hit, the interface will give you a list of Tensors containing Infs or Nans, sorted in time order. The one on the top should be the \"culprit\", i.e., the one that first generated the bad numerical values. Say its name is <code>node_1<\/code>, you can use the following tfdbg commands to look at its inputs and node attributes:<\/p>\n\n<pre><code>tfdbg&gt; li -r node_1\ntfdbg&gt; ni -a node_1\n<\/code><\/pre>\n"}]}
{"QuestionId":41527430,"AnswerCount":0,"Tags":"<python><image-processing><tensorflow><computer-vision>","CreationDate":"2017-01-07T22:47:32.447","AcceptedAnswerId":null,"OwnerUserId":7089821.0,"Title":"Failed to read the image with Tensorflow shuffle_batch() function within user defined function","Body":"<p>I am trying to read the images and relevant labels using  function with my user-defined function, But it seems cannot start the reading file queue. <\/p>\n\n<p><strong>1.My Questiones:<\/strong><\/p>\n\n<ol>\n<li>Why can not the code read images successfully with calling shuffle_batch() in my user-defined function?   <\/li>\n<li>How to fix the problem? <\/li>\n<\/ol>\n\n<p><strong>2.Code with user defined function<\/strong><\/p>\n\n<p><strong>Note print1,2,3 in Session() part.<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Print result of code with user defined function<\/strong><\/p>\n\n<p>print1 works exactly the same as print2: <\/p>\n\n<pre><\/pre>\n\n<p>print3 is always stuck with nothing display on the console(No matter what order they are)<\/p>\n\n<p><strong>3.Code without user defined function<\/strong><\/p>\n\n<p>It is almost the same as the code with user defined function. <\/p>\n\n<pre><\/pre>\n\n<p><strong>4.Print result of code without user defined function<\/strong><\/p>\n\n<p>print1 and print2 works again with the same result:<\/p>\n\n<pre><\/pre>\n\n<p>print3 also works: <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41527484,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-07T22:54:50.010","AcceptedAnswerId":null,"OwnerUserId":1710775.0,"Title":"Shift dimension or rollaxis in tensorflow","Body":"<p>I would like to perform shift dimention(matlab) or rollaxis(numpy) in tensorflow, with the matrix. \nFor eg. a matrix with shape (512,7,7) to (7,7,512). Reshape will not work as I will get a garbage data. \nHow can I do that in tensorflow? <\/p>\n","answers":[{"AnswerId":"41531326","CreationDate":"2017-01-08T09:40:10.600","ParentId":null,"OwnerUserId":"1710775","Title":null,"Body":"<p>Thanks, I will answer my question then.\ntf.transpose is the function to use. The perm parameter allow us to set the new dimension. \nIn my case, the perm will be (1,2,0) since I am going from [512,7,7] to [7,7,512] preserving the order of the 7s in both cases. <\/p>\n\n<p>tf.transpose(x, [1,2,0])<\/p>\n"}]}
{"QuestionId":41528837,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-08T02:27:47.457","AcceptedAnswerId":42408906.0,"OwnerUserId":3994228.0,"Title":"Tensorflow : Android demo accuracy","Body":"<p><strong>What related GitHub issues or Stack Overflow threads have you found by searching the web for your problem?<\/strong><\/p>\n\n<p>I searched #1269 #504<\/p>\n\n<p><strong>Environment info<\/strong><\/p>\n\n<p>Mac OS for build and Android version 5 to run .apk demo.<\/p>\n\n<p><strong>If possible, provide a minimal reproducible example (We usually don't have time to read hundreds of lines of your code)<\/strong><\/p>\n\n<p>I followed the steps mentioned in #1269 and could able to run the example successfully, but the accuracy of the result is very low and often wrong. I have trained my systems on 25 different daily used products like soap, soup, noodles, etc.\nWhere as when i run the same example using following script it give me very high accuracy (approx. 90-95%)<\/p>\n\n<pre><\/pre>\n\n<p>The only difference I see here is that the model file used in the Android demo is stripped because it does not support , whereas in the above code its the actually generated unstripped model. Is there any specific reason or somewhere I am wrong here?<\/p>\n\n<p>I also tried using <\/p>\n\n<p>but unfortunately, it fails with following error: <\/p>\n\n<pre><\/pre>\n\n<p>I suspect that this problem is due to the android not being parse , but please correct me if i am wrong. <\/p>\n\n<p><strong>What other attempted solutions have you tried?<\/strong><\/p>\n\n<p>Yes, I the above script and it gives me quite high accuracy result.<\/p>\n","answers":[{"AnswerId":"42408906","CreationDate":"2017-02-23T06:52:26.320","ParentId":null,"OwnerUserId":"3994228","Title":null,"Body":"<p>Well, the reason for bad accuracy is following: <\/p>\n\n<p>I ran this example code on Lenovo Vibe K5 mobile (This has SanpDragon 415), this wasn't compiled for hexagon DSP, even through the DSP on 415 is very old as compared to 835 (Hexagon DSP 682), in fact i am not quite sure if the Hexagon SDK will work with 415 or not, i haven't tried this though. This means the example was running on CPU to first detect motion and later classify them and hence the poor performance.<\/p>\n\n<ol>\n<li>Slow FPS, will capture the images very slowly and hence moving objects will be really difficult. <\/li>\n<li>So if you have bad image, there are very strong chances that the prediction will also be bad. <\/li>\n<li>Camera capture and classification is taking long time, due to latency its not quite real-time. <\/li>\n<\/ol>\n"}]}
{"QuestionId":41529464,"AnswerCount":1,"Tags":"<caffe><pycaffe><deeplearning4j><matcaffe>","CreationDate":"2017-01-08T04:30:05.747","AcceptedAnswerId":41530143.0,"OwnerUserId":6494707.0,"Title":"What parts an parameters of train_val.prototxt, solver.prototxt, and deploy.prototxt should be changed based on our own data?","Body":"<p>I am confused about changing the parameters for training a  in caffe based on our own data. <\/p>\n\n<ul>\n<li>Which layers of net we should pay attention more to train the net on\nour own data? For example, the number of outputs based on the number\nof classes.<\/li>\n<\/ul>\n\n<p>I tried to train FCN32 for semantic segmentation. I changed the number of outputs in  to the number of classes in my data, but it is giving an error.  <\/p>\n\n<ul>\n<li>We have different outputs in different convolutional layers. How can\nI detect different outputs from each other and which one should I\nchange?<\/li>\n<li>The next question is what is the difference between \nand .<\/li>\n<li>And what is the application of deploy.prototxt? Should I change the\nlayers in  based on the train_val.prototxt?<\/li>\n<\/ul>\n\n<p>I really appreciate if someone knows, please share your knowledge.<\/p>\n\n<p>Thanks <\/p>\n","answers":[{"AnswerId":"41530143","CreationDate":"2017-01-08T06:32:21.197","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<blockquote>\n  <p>Which layer of the net we should pay attention <\/p>\n<\/blockquote>\n\n<p>When finetuning a model to a new task, with probably different number of labels, one <strong>must<\/strong> change <code>num_output<\/code> of the last layer.<br>\nThe last layer serves as a probability prediction layer, outputting a vector of probabilities for each predicted variable: For image classification, the net predicts a prob vector per image; In FCN the net outputs prob vector <em>pre pixel<\/em>. The prob vector is of length = number of labels = <code>num_output<\/code> of last layer. Thus you must change this value to accommodate for the new number of labels in your new task.<br>\nSee <a href=\"https:\/\/stackoverflow.com\/a\/36842553\/1714410\" title=\"Fine Tuning of GoogLeNet Model\">this post for more details<\/a>.<\/p>\n\n<blockquote>\n  <p>What is the difference between <code>'deploy.prototxt'<\/code> and <code>'train_val.prototxt'<\/code><\/p>\n<\/blockquote>\n\n<p>This difference is explained in <a href=\"https:\/\/stackoverflow.com\/a\/34973370\/1714410\">this post<\/a>. <\/p>\n\n<blockquote>\n  <p>Should I change the layers in <code>'deploy.prototxt'<\/code> based on the <code>'train_val.prototxt'<\/code>?<\/p>\n<\/blockquote>\n\n<p>Yes!<\/p>\n"}]}
{"QuestionId":41529526,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2017-01-08T04:41:30.673","AcceptedAnswerId":null,"OwnerUserId":7211739.0,"Title":"tensorflow AttributError: type object 'NewBase' has no attribute 'is_abstract'","Body":"<p>Already upgraded six. pip sucessfully installed tensorflow. Went through several steps\/upgrades just to get to this point. Now, when I try to import into python that this error comes up.<\/p>\n\n<pre><\/pre>\n\n<p>Help!<\/p>\n","answers":[{"AnswerId":"46450701","CreationDate":"2017-09-27T14:37:48.893","ParentId":null,"OwnerUserId":"8167930","Title":null,"Body":"<p>I ran into a similar issue but likely for a different reason.  I happen to have a package called rdkit also installed (recently).  It happens to have a much older version of six.py embedded (v1.6.1 why - idk?).  However, I was able to copy in a newer replacement for six.py and it fixed everything after adding a small fix to rdkit as well.  For work I am doing I need rdkit in the same environment unfortunately.  I guess this is something else to look for if you work with packages that use their own version of six.py.<\/p>\n"},{"AnswerId":"42665044","CreationDate":"2017-03-08T07:08:47.913","ParentId":null,"OwnerUserId":"1384641","Title":null,"Body":"<p>Upgrading six will resolve the issue. \nTensorFlow requires <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/pip_package\/setup.py#L41\" rel=\"nofollow noreferrer\">six => 1.10<\/a> - to check the six version, via interpreter:<\/p>\n\n<pre><code>$ python\n&gt;&gt;&gt;import six\n&gt;&gt;&gt;print six.\\__version\\__\n<\/code><\/pre>\n\n<p>Or, you can simply check it using pip:<\/p>\n\n<pre><code>$ pip show six\n<\/code><\/pre>\n\n<p>If you have six version less than 1.10, upgrade it:<\/p>\n\n<pre><code>$ pip install --upgrade six\n<\/code><\/pre>\n"},{"AnswerId":"43712626","CreationDate":"2017-05-01T00:02:41.303","ParentId":null,"OwnerUserId":"7678312","Title":null,"Body":"<p>It might to be too late, but I put a suggestion.<\/p>\n\n<p>In case of Mac OS X, the Python modules can be placed in two different places and this might to cause the problem.<\/p>\n\n<p>Please check the module, six, this case, in:<\/p>\n\n<ul>\n<li>\/System\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/Extras\/lib\/python<\/li>\n<li>\/Library\/Python\/2.7\/site-packages\/<\/li>\n<\/ul>\n\n<p>If you find an older version in the folder\/file, delete it.\nHope this helps.<\/p>\n"},{"AnswerId":"42664197","CreationDate":"2017-03-08T06:12:22.677","ParentId":null,"OwnerUserId":"1242045","Title":null,"Body":"<p>That's because your \"six\" package is out-of-date. Please do below:<\/p>\n\n<ol>\n<li>check your six version via: python; import six; print(six.<strong>version<\/strong>)<\/li>\n<li>If version &lt; 1.10, then you need to upgrade six to newer version<\/li>\n<li>sudo easy_install --upgrade six<\/li>\n<li>check again six version and make sure the version >= 1.10<\/li>\n<li>redo python; import tensorflow as tf;<\/li>\n<\/ol>\n\n<p>Good Luck!<\/p>\n"}]}
{"QuestionId":41529636,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-01-08T05:00:57.350","AcceptedAnswerId":null,"OwnerUserId":1625695.0,"Title":"Why shuffling data gives significantly higher accuracy?","Body":"<p>In Tensorflow, I've wrote a big model for 2 image classes problem. My question is concerned with the following code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>X is an np.array of shape: (25000,244,244,3). That code results in accuracy=0.5834 (towards random accuracy) and loss=2.7106. But\nwhen I shuffle the data, by adding these 3 lines after the first line:<\/p>\n\n<pre><\/pre>\n\n<p>, the results become convenient: accuracy=0.9933 and loss=0.0208.\nWhy shuffling data can give significantly higher accuracy ? or what can be a reason for that ?<\/p>\n\n<p>The function calc_probs is mainly a run call:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Update:<\/strong>\nAfter hours of debugging, I figured out that evaluating a single image gives different result. For example, if you run the following line of code multiple times, you get a different result each time:<\/p>\n\n<pre><\/pre>\n\n<p>My data is normally sorted, X contains class 1 samples first then class 2. And in calc_probs function, I run using each batch of the data sequentially. So, without shuffling, each run has data of a single class.\nI've also noted that with shuffling, if batch size is very small, I get the random accuracy.<\/p>\n","answers":[{"AnswerId":"41530045","CreationDate":"2017-01-08T06:13:57.443","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>There is some mathematical justification for this in the context of randomized Kaczmarz algorithm. Regular Kaczmarz algorithm is an old algorithm which can be seen as an non-shuffling SGD on a least squares problem, and there are guaranteed faster convergence rates that come out if you use randomization, follow references in <a href=\"http:\/\/www.cs.ubc.ca\/~nickhar\/W15\/Lecture21Notes.pdf\" rel=\"nofollow noreferrer\">http:\/\/www.cs.ubc.ca\/~nickhar\/W15\/Lecture21Notes.pdf<\/a><\/p>\n"}]}
{"QuestionId":41529997,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><computer-vision><neural-network><deep-learning>","CreationDate":"2017-01-08T06:05:58.537","AcceptedAnswerId":null,"OwnerUserId":3863647.0,"Title":"tensorflow conv2d unexpected convolution result","Body":"<p>I try to migrate a Caffe network and model(weights) to tensorflow.\nThe original first layer is defined as shown at last, which is a stride one convolution on  gray image with kernel size 5x5, output channel 96. <\/p>\n\n<p>I converted the weights from  file to  following this procedure:<\/p>\n\n<pre><\/pre>\n\n<p>. After the transpose, new array of , are used to initialize  filter.<\/p>\n\n<p> code is as followed:<\/p>\n\n<pre><\/pre>\n\n<p>I checked the input data of  net and 's tensor , they are the same numbers with the same 2D layout.  and ,  use a batch size of 10.<\/p>\n\n<p>I also checked the kernel through 's  and the  used to initalize  with .<\/p>\n\n<p>the  kernel's first layer screen shot is shown below:\n<a href=\"https:\/\/i.stack.imgur.com\/lUgCb.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lUgCb.png\" alt=\"Caffe&#39;s kernel\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/9kfBT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9kfBT.png\" alt=\"tensorflow&#39;s kernel\"><\/a><\/p>\n\n<p>the bias is  and the upper left corner of the image is:<\/p>\n\n<pre><\/pre>\n\n<p>However, the two's upper left corner of 's first feature map are <strong>different<\/strong>.(please ignore the irrrelevant 256)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/E1hKj.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/E1hKj.png\" alt=\"conv results\"><\/a><\/p>\n\n<p>Only the leftmost column is consistent. I manually calculated and checked the results, the first and the second value of 's () are correct according to the definition of convolution, the second value in 's () is <strong>incorrect<\/strong>.<\/p>\n\n<p>Taking into account the padding, the first value is from 3x3 block of the image, the second should be the 3x4 block.<\/p>\n\n<p>Since  has the correct first value, computed from the 3x3 block of image corner, I assume the kernel layout and image layout and 'SAME' padding are correct. I thought it was a problem with stride that caused the incorrect second value, but the stride must be one, otherwise 's  feature map's size won't be .<\/p>\n\n<p>Caffe's convolution layer def:<\/p>\n\n<pre><\/pre>\n\n<p>UPDATE:<\/p>\n\n<p>Another contrived experiment(see code bolow) shows the  implementation is able to compute the correct second value. However, the error remains in the above situation. What is it that caused the error in the converted version?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41530218,"AnswerCount":1,"Tags":"<python><neural-network><keras>","CreationDate":"2017-01-08T06:44:34.523","AcceptedAnswerId":41532267.0,"OwnerUserId":785494.0,"Title":"Keras: Loading minibatches from HDF5 and CSV","Body":"<p>I have a large dataset, too large to fit into RAM, which is available either as HDF5 or CSV.  How can I feed it into Keras in minibatches?  Also, will this shuffle it for me, or do I need to pre-shuffle the dataset?<\/p>\n\n<p>(I'm also interested in this when the input is a Numpy recarray; since Keras I believe wants the input to be a ndarray.)<\/p>\n\n<p>And, if I want to do some lightweight preprocessing in Keras before learning (e.g. apply a few Python functions to the data to change the representation), hcan that be added?<\/p>\n","answers":[{"AnswerId":"41532267","CreationDate":"2017-01-08T11:38:09.753","ParentId":null,"OwnerUserId":"6769011","Title":null,"Body":"<p>Have a look at the fit_generator method available with Keras here: <a href=\"https:\/\/keras.io\/models\/sequential\/#sequential-model-methods\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/sequential\/#sequential-model-methods<\/a>\nIt fits the model on data generated batch-by-batch by a Python generator  (Where you can write shuffling logic, since generator is under your control).<\/p>\n\n<p>You may apply call pre-processing within the generator itself.<\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":41530801,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-01-08T08:19:16.047","AcceptedAnswerId":41530840.0,"OwnerUserId":5279545.0,"Title":"How to view trained weights and biases (in human readable format) from a trained Tensorflow Model (.ckpt) file","Body":"<p>I have trained a CNN and created a tensorflow model file (.ckpt).<\/p>\n\n<p>I have to implement the forward propagation on a hardware using the trained model and for that I need the updated weights and biases (as I got after the training).<\/p>\n\n<p>However, the .ckpt file has binary encoding. <\/p>\n\n<p>How do I convert it to some human readable format so I can read the updated model parameters?<\/p>\n","answers":[{"AnswerId":"41530840","CreationDate":"2017-01-08T08:25:04.987","ParentId":null,"OwnerUserId":"2005810","Title":null,"Body":"<p>Use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session.run\" rel=\"nofollow noreferrer\">sess.run<\/a> to evaluate whatever weight\/bias in your model. The result is basically a numpy array, then use <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.savetxt.html\" rel=\"nofollow noreferrer\">numpy.savetxt<\/a>.<\/p>\n"}]}
{"QuestionId":41530950,"AnswerCount":1,"Tags":"<matplotlib><tensorflow><ipython-notebook><jupyter>","CreationDate":"2017-01-08T08:42:51.037","AcceptedAnswerId":null,"OwnerUserId":7389955.0,"Title":"ipython cannot search matplotlib while using tensorflow and jupyter also has import error","Body":"<p>I am using python 2.7 in Ubuntu and recently updated tensorflow 0.12.1.\nI installed jupyter today for my sample code of tf and I need to use matplotlib. It does not find module name matplotlib and ipython in tensorflow has same error.<\/p>\n\n<h3>1. How can I set path in virtualenv or ipython or jupyter?<\/h3>\n\n<p>After activate tensorflow, I need to use jupyter notebook.<\/p>\n\n<p>This below in the script for error does not work.<\/p>\n\n<pre><\/pre>\n\n<h3>2. other information: My environments are below.<\/h3>\n\n<blockquote>\n  <p>mickyefromsd@DEKSTOP~$source ~\/tensorflow\/bin\/activate <\/p>\n<\/blockquote>\n\n<p>When I find matplotlib by python script under TF condition and before TF activation, it has below;<\/p>\n\n<blockquote>\n  <p>\/usr\/lib\/python2.7\/dist-packages\/matplotlib\/<\/p>\n<\/blockquote>\n\n<p>When I type 'which ipython', it has below (not by \/usr\/bin\/ipython) ;<\/p>\n\n<blockquote>\n  <p>\/home\/mickeyfromd\/tensorflow\/bin\/ipython<\/p>\n<\/blockquote>\n\n<p>Btw, \/tensorflow\/lib\/python2.7\/site-packages\/ it has ipython and jupyter. <\/p>\n\n<p>(not in the same path of matplotlib)<\/p>\n\n<h3>3. My ipython under TF cannot find my existing matplotlib.<\/h3>\n\n<blockquote>\n  <p>(tensorflow) mickeyfromd@DK-DESKTOP:~$ ipython<\/p>\n  \n  <p>Python 2.7.11+ (default, Apr 17 2016, 14:00:29) <\/p>\n  \n  <p>In [1]: import matplotlib ImportError: No module named matplotlib<\/p>\n<\/blockquote>\n\n<h3>4. I wanted to setup virtualenv, so I just run this<\/h3>\n\n<p>I followed this site. site:<a href=\"http:\/\/help.pythonanywhere.com\/pages\/IPythonNotebookVirtualenvs\" rel=\"nofollow noreferrer\">http:\/\/help.pythonanywhere.com\/pages\/IPythonNotebookVirtualenvs<\/a><\/p>\n\n<blockquote>\n  <p>(tensorflow) mickeyfromd@ipython kernelspec install-self --user\n  .....\n  Installed kernelspec python2 in \/home\/mickeyfromd\/.local\/share\/jupyter\/kernels\/python2<\/p>\n  \n  <p>(tensorflow) mickeyfromd@DK-DESKTOP:~$<\/p>\n<\/blockquote>\n\n<p>I cannot move the the folder (in the second step)<\/p>\n\n<p>How can I make ipython to have path for Matplotlib?<\/p>\n","answers":[{"AnswerId":"46505960","CreationDate":"2017-09-30T18:45:33.650","ParentId":null,"OwnerUserId":"4444871","Title":null,"Body":"<p>That import error is due to change in environment of the jupyter notebook. You might have installed the packages in one environment and you are running the jupyter notebook in another environment. <\/p>\n\n<ol>\n<li>I have got two environments (envs) in my Anaconda folder ( I have Anaconda3 folder to be specific ).<\/li>\n<li>(windows key+cmd ) -> open the windows command prompt run as administrator.<\/li>\n<li>Activate (name of the environment) -> eg.: activate tensorflow-gpu<\/li>\n<li>Start installing packages using conda install <\/li>\n<\/ol>\n\n<p>Note: For each environment you need to install all the packages you want to use, separately using the same process. This solution is for windows users, might work for linux users not sure though.<\/p>\n\n<p>Additionally to make sure your conda environment is up to date run:\n<pre>\nconda update conda<\/p>\n\n<code>conda update anaconda\n<\/code><\/pre>\n\n<p>\ncheck this out : <a href=\"https:\/\/pradyumnamajumder.wordpress.com\/2017\/09\/30\/solution-to-the-python-packages-import-error-in-jupyter\/\" rel=\"noreferrer\">https:\/\/pradyumnamajumder.wordpress.com\/2017\/09\/30\/solution-to-the-python-packages-import-error-in-jupyter\/<\/a><\/p>\n"}]}
{"QuestionId":41531002,"AnswerCount":1,"Tags":"<python><neural-network><theano>","CreationDate":"2017-01-08T08:53:33.493","AcceptedAnswerId":null,"OwnerUserId":3101193.0,"Title":"nested loop in theano","Body":"<p>I have the following definitions inspired by the MLP Theano tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to comput a Gaussian activation like this:<\/p>\n\n<pre><\/pre>\n\n<p>However, this is not possible directly because neither W nor X nor s are iterable. How to code a Theano version in an efficient way?<\/p>\n\n<p>Thank you very much.<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I have configured a solution:<\/p>\n\n<pre><\/pre>\n\n<p>However.. I have observed that this returns a all items zeros in . It seems that even the difference  cant be computed correctly. <\/p>\n\n<p><strong>EDIT_2<\/strong> I have fixed the problem, however I think the mine is not the most efficient solution, can anyone suggest me something better please?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41540527","CreationDate":"2017-01-09T03:18:13.787","ParentId":null,"OwnerUserId":"2845293","Title":null,"Body":"<p><code>T.set_subtensor(x[...], y)<\/code> will return a \"symbolic\" variable with the given subtensor replaced by <code>y<\/code>. It won't actually operate inside the <code>x<\/code>. To change the shared varaible <code>x<\/code>, you'll need to use <code>theano.function<\/code> and \"<code>update<\/code>\" <code>x<\/code> with a new value.<\/p>\n"}]}
{"QuestionId":41531038,"AnswerCount":1,"Tags":"<c++><neural-network><deep-learning><caffe>","CreationDate":"2017-01-08T08:58:09.630","AcceptedAnswerId":null,"OwnerUserId":7091998.0,"Title":"caffe error == cudaSuccess (77 vs. 0) an illegal memory access was encountered","Body":"<p>Recently,I need to change the code of softmax_loss_layer in caffe.\nThe caffe version is: <a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe<\/a><\/p>\n\n<p>The code between ... is my code,and the rest is the same.Here is how I modified the code\nFirst,I added rs_ in loss_layers.hpp<\/p>\n\n<pre><\/pre>\n\n<p>Then,I reshape rs_ in softmax_loss_layer.cpp<\/p>\n\n<pre><\/pre>\n\n<p>Bottom[2] is from dense_image_data_layer.cpp<\/p>\n\n<pre><\/pre>\n\n<p>At last,I added this code in softmax_loss_layer.cu<\/p>\n\n<pre><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><\/pre>\n\n<p>I think that this error is because I didn't use rs_ correctly or bottom[2] is not right.\nAll the code that I changed is posted above,so can anybody tell me what to do?\nIf you need extra information,please tell me?<\/p>\n","answers":[{"AnswerId":"41531250","CreationDate":"2017-01-08T09:28:43.763","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It is difficult to follow your code and changes.<br>\nSome comments;<\/p>\n\n<ol>\n<li><code>reshape<\/code>ing <code>rs_<\/code> should occur in <code>reshape()<\/code> method, rather than in <code>setup()<\/code>.  <\/li>\n<li>It is best using <code>rs_.ReshapeLike(*bottom[2])<\/code> than explicitly enumerating <code>num<\/code>, <code>channels<\/code> etc. What if you are going to have a Blob with different number of dimensions?  <\/li>\n<li><p>Have you <em>tested<\/em> your modified layer? From <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Development\" rel=\"nofollow noreferrer\">caffe wiki<\/a>:<\/p>\n\n<blockquote>\n  <p>Write tests in <code>test\/test_your_layer.cpp<\/code>. Use <code>test\/test_gradient_check_util.hpp<\/code> to check that your Forward and Backward implementations are in numerical agreement.<\/p>\n<\/blockquote><\/li>\n<\/ol>\n"}]}
{"QuestionId":41531657,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-08T10:23:46.973","AcceptedAnswerId":41536532.0,"OwnerUserId":1710775.0,"Title":"Creating a new matrix from a tensor in tensorflow","Body":"<p>Let's say I have a (7,7,3) image with 3 channels. \nHow can I create a matrix A using the image such that the each row consist of just the pairwise connections of the neighbouring pixel? I know this can be done easily in python with for loops but how can we do this within a tensorflow graph? <\/p>\n\n<p>Example of matrix A: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41536532","CreationDate":"2017-01-08T18:55:08.943","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You can do this using some matrix algebra. To illustrate the idea, suppose you wanted to do this for a 1D vector.<\/p>\n\n<p>You can stack the vector with a shifted version of itself to get pairs of neighbors<\/p>\n\n<pre><code>n = 5\na = tf.range(n)\nleft = tf.stack([a[1:], a[:n-1]])\nleft = tf.transpose(left)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iuWjl.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iuWjl.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>By chopping off the tails and repeating for different offset you can get left neighbors and right neighbors<\/p>\n\n<pre><code>right = tf.stack([a[:n-1], a[1:]])\nright = tf.transpose(right)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GWKdA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GWKdA.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>To ignore edge effects you can chop off the ends and stack again into rank-3 matrix<\/p>\n\n<pre><code>stacked_neighbors = tf.stack([left[:-1], right[1:]])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/G7tVg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/G7tVg.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Now to interleave the neighbors we can use a trick with transpose and reshape. <\/p>\n\n<pre><code>stacked_neighbors = tf.transpose(stacked_neighbors, [1, 0, 2])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9Mp4E.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9Mp4E.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Since data storage is in row-major order, reshaping into less dimensions than original, reshape flattens excess dimensions on the left<\/p>\n\n<pre><code>stacked_neighbors = tf.reshape(stacked_neighbors, [6,2])\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5WzDn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5WzDn.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":41531695,"AnswerCount":1,"Tags":"<python><tensorflow><machine-learning><deep-learning><keras>","CreationDate":"2017-01-08T10:28:55.723","AcceptedAnswerId":41534323.0,"OwnerUserId":785494.0,"Title":"How does keras define \"accuracy\" and \"loss\"?","Body":"<p>I can't find how Keras defines \"accuracy\" and \"loss\".  I know I can specify different metrics (e.g. mse, cross entropy) - but keras prints out a standard \"accuracy\".  How is that defined? Likewise for loss: I know I can specify different types of regularization -- are those in the loss?<\/p>\n\n<p>Ideally, I'd like to print out the equation used to define it; if not, I'll settle for an answer here.<\/p>\n","answers":[{"AnswerId":"41534323","CreationDate":"2017-01-08T15:27:59.947","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>Have a look at <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/c2e36f369b411ad1d0a40ac096fe35f73b9dffd3\/keras\/metrics.py\" rel=\"noreferrer\"><code>metrics.py<\/code><\/a>, there you can find definition of all available metrics including different types of accuracy. Accuracy is not printed unless you add it to the list of desired metrics when you compile your model.<\/p>\n\n<p>Regularizers are by definition added to the loss. For example, see <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/f573a86b42e49754386e536358e08e861d40d24c\/keras\/engine\/topology.py#L417\" rel=\"noreferrer\"><code>add_loss<\/code><\/a> method of the <code>Layer<\/code>class.<\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>The type of <code>accuracy<\/code> is determined based on the objective function, see <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/d8b226f26b35348d934edb1213061993e7e5a1fa\/keras\/engine\/training.py#L651\" rel=\"noreferrer\"><code>training.py<\/code><\/a>. The default choice is <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/c2e36f369b411ad1d0a40ac096fe35f73b9dffd3\/keras\/metrics.py#L13\" rel=\"noreferrer\"><code>categorical_accuracy<\/code><\/a>. Other types like <code>binary_accuracy<\/code> and <code>sparse_categorical_accuracy<\/code> are selected when the objective function is either binary or sparse. <\/p>\n"}]}
{"QuestionId":41533489,"AnswerCount":6,"Tags":"<python><tensorflow>","CreationDate":"2017-01-08T13:59:36.457","AcceptedAnswerId":50434439.0,"OwnerUserId":3401555.0,"Title":"How to initialise only optimizer variables in Tensorflow?","Body":"<p>I want to use  in Tensorflow. However, since this optimizer uses some internal variable, attempting to use it without initializing this variable yields an error: <\/p>\n\n<blockquote>\n  <p> (see above for traceback): Attempting to use\n   uninitialized value <\/p>\n<\/blockquote>\n\n<p>This can be easily solved by initializing all variables, using for example<\/p>\n\n<pre><\/pre>\n\n<p>However, I do not want to initialize <strong>all<\/strong> the variables - only those of optimizer. Is there any way to do this?<\/p>\n","answers":[{"AnswerId":"41535957","CreationDate":"2017-01-08T18:03:18.203","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You can filter variables by name and only initialize those. IE<\/p>\n\n<pre><code>momentum_initializers = [var.initializer for var in tf.global_variables() if 'Momentum' in var.name]\nsess.run(momentum_initializers)\n<\/code><\/pre>\n"},{"AnswerId":"50434439","CreationDate":"2018-05-20T11:38:34.773","ParentId":null,"OwnerUserId":"5655977","Title":null,"Body":"<p>There is a more straightforward way:<\/p>\n\n<pre><code>optimizer = tf.train.AdamOptimizer()\nsession.run(tf.variables_initializer(optimizer.variables()))\n<\/code><\/pre>\n"},{"AnswerId":"45624533","CreationDate":"2017-08-10T22:46:08.003","ParentId":null,"OwnerUserId":"2366315","Title":null,"Body":"<p>Both current answers kinda work by filtering the variable name using the 'Momentum' string. But that is very brittle on two sides:<\/p>\n\n<ol>\n<li>It could silently (re-)initialize some other variables you don't actually want to reset! Either simply because of a name-clash, or because you have a more complex graph and optimize different parts separately, for example.<\/li>\n<li>It will only work for one specific optimizer, and how do you know the names to look out for for others?<\/li>\n<li>Bonus: an update to tensorflow might <strong>silently<\/strong> break your code.<\/li>\n<\/ol>\n\n<p>Fortunately, tensorflow's abstract <code>Optimizer<\/code> class has a mechanism for that, these extra optimizer variables are called <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Optimizer#slots\" rel=\"noreferrer\">\"slots\"<\/a>, and you can get all slot names of an optimizer using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Optimizer#get_slot_names\" rel=\"noreferrer\"><code>get_slot_names()<\/code><\/a> method:<\/p>\n\n<pre><code>opt = tf.train.MomentumOptimizer(...)\nprint(opt.get_slot_names())\n# prints ['momentum']\n<\/code><\/pre>\n\n<p>And you can get the variable corresponding to the slot for a specific (trainable) variable <code>v<\/code> using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Optimizer#get_slot\" rel=\"noreferrer\"><code>get_slot(var, slot_name)<\/code><\/a> method:<\/p>\n\n<pre><code>opt.get_slot(some_var, 'momentum')\n<\/code><\/pre>\n\n<p>Putting all this together, you can create an op that initializes the optimizer's state as follows:<\/p>\n\n<pre><code>var_list = # list of vars to optimize, e.g. \n           # tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)\nopt = tf.train.MomentumOptimizer(0.1, 0.95)\nstep_op = opt.minimize(loss, var_list=var_list)\nreset_opt_op = tf.variables_initializer([opt.get_slot(var, name) for name in opt.get_slot_names() for var in var_list])\n<\/code><\/pre>\n\n<p>This will really only reset the correct variables, and be robust across optimizers.<\/p>\n\n<p>Except for one <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8057\" rel=\"noreferrer\">unfortunate caveat<\/a>: <code>AdamOptimizer<\/code>. That one also keeps a counter for how often it's been called. That means you should really think hard about what you're doing here anyways, but for completeness' sake, you can get its extra states as <code>opt._get_beta_accumulators()<\/code>. The returned list should be added to the list in the above <code>reset_opt_op<\/code> line.<\/p>\n"},{"AnswerId":"44743216","CreationDate":"2017-06-25T04:34:58.557","ParentId":null,"OwnerUserId":"4115369","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variables_initializer\" rel=\"nofollow noreferrer\"><code>tf.variables_initializer<\/code><\/a> seems to be the preferred way to initialize a specific set of variables:<\/p>\n\n<pre><code>var_list = [var for var in tf.global_variables() if 'Momentum' in var.name]\nvar_list_init = tf.variables_initializer(var_list)\n...\nsess = tf.Session()\nsess.run(var_list_init)\n<\/code><\/pre>\n"},{"AnswerId":"48946346","CreationDate":"2018-02-23T11:02:47.120","ParentId":null,"OwnerUserId":"9385301","Title":null,"Body":"<p>To fix the None problem just do:<\/p>\n\n<pre><code>  self.opt_vars = [opt.get_slot(var, name) for name in opt.get_slot_names() \n                   for var in self.vars_to_train\n                   if opt.get_slot(var, name) is not None]\n<\/code><\/pre>\n"},{"AnswerId":"46367188","CreationDate":"2017-09-22T14:32:16.580","ParentId":null,"OwnerUserId":"209246","Title":null,"Body":"<p>Building off of LucasB's answer about <code>AdamOptimizer<\/code>, this function takes an <code>AdamOptimizer<\/code> instance <code>adam_opt<\/code> that has its <code>Variables<\/code> created (one of these two called: <code>adam_opt.minimize(loss, var_list=var_list)<\/code> or <code>adam_opt.apply_gradients(zip(grads, var_list))<\/code>. The function creates an <code>Op<\/code> that, when called, re-initializes the optimizer's variables for the passed variable, as well as the global counting state.<\/p>\n\n<pre><code>def adam_variables_initializer(adam_opt, var_list):\n    adam_vars = [adam_opt.get_slot(var, name)\n                 for name in adam_opt.get_slot_names()\n                 for var in var_list if var is not None]\n    adam_vars.extend(list(adam_opt._get_beta_accumulators()))\n    return tf.variables_initializer(adam_vars)\n<\/code><\/pre>\n\n<p>e.g.:<\/p>\n\n<pre><code>opt = tf.train.AdamOptimizer(learning_rate=1e-4)\nfit_op = opt.minimize(loss, var_list=var_list)\nreset_opt_vars = adam_variables_initializer(opt, var_list)\n<\/code><\/pre>\n"}]}
{"QuestionId":41534560,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2017-01-08T15:52:42.713","AcceptedAnswerId":null,"OwnerUserId":4613543.0,"Title":"How to use user_ops in Tensorflow Android APP","Body":"<p>I want to transplantation Faster RCNN to Android device, and I meet some problems.\nI build my Android project by Bazel, just like the tensorflow demo. I add the op  and  to user_ops, but I seems Android does build user_ops, the logcat shows as following:<\/p>\n\n<pre><\/pre>\n\n<p>I try to solve this problem, and I move  and  to jni folder(maybe it is a bad ideal), the previous error is disappeared, but I meet another problem, outputs are all not found, I do not know why, logcat as following:<\/p>\n\n<pre><\/pre>\n\n<p>I do not know how to solve them and I to debug, could you help me.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"41555466","CreationDate":"2017-01-09T19:28:51.877","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>Is there an output index on these nodes, perhaps? E.G. \"rois:0\" or \"rois:1\".<\/p>\n\n<p>If you add a log line for node names\/types to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/util\/stat_summarizer.cc#L41\" rel=\"nofollow noreferrer\">stat_summarizer.cc<\/a> in the constructor you can see what nodes are defined in your graph which might illuminate the problem.<\/p>\n\n<p>Adding the ops to the jni directory as you've done should work fine as a quick and dirty solution, and it seems to have done so if it got you past your initial problem.<\/p>\n"}]}
{"QuestionId":41534593,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-08T15:55:47.323","AcceptedAnswerId":41534724.0,"OwnerUserId":7287271.0,"Title":"Why would I ever use tf.concat instead of tf.stack?","Body":"<p>Is there a good reason to use  instead of ? They seem very similar. Is it just to guarantee that the resulting tensor will have the same number of dimensions as the input list of tensors?<\/p>\n","answers":[{"AnswerId":"41534724","CreationDate":"2017-01-08T16:08:26.497","ParentId":null,"OwnerUserId":"7287271","Title":null,"Body":"<p>Actually, I've misunderstood how <code>tf.stack<\/code> works. If the <code>axis<\/code> parameter is within the range of the existing dimensions, a new axis will be inserted at that index.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nt1 = tf.random_normal([1, 3])\nt2 = tf.random_normal([1, 3])\n\ntf.stack([t1, t2], axis=1).shape.as_list() == [1, 2, 3]\ntf.concat([t1, t2], axis=1).shape.as_list() == [1, 6]\n<\/code><\/pre>\n"}]}
{"QuestionId":41534866,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-01-08T16:23:19.083","AcceptedAnswerId":41536952.0,"OwnerUserId":6112394.0,"Title":"Image distortion returns error: The tensor returned for Reshape_5:0 was not valid","Body":"<p>I am attempting to add image distortion to my ConvNet model and am getting a really odd error. My data is in  format and I am using the  function from the  code. Below is some dummy code I threw together to see that everything was doing what I expected it to do. When I view the images after they are distorted there is no problem. The problem appears to arise after flattening the tensor or after it is run. For some reason, it will execute once, but the second time it throws an error. Below is my code and the error it returns. My current suspicion is that it might be the  but I don't know. <\/p>\n\n<pre><\/pre>\n\n<p>Error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41536952","CreationDate":"2017-01-08T19:34:18.103","ParentId":null,"OwnerUserId":"6112394","Title":null,"Body":"<p>Ok so what I believe was happening was I resized <code>X_test_batch<\/code> back to the shape of an image. I then ran my distortion on it and called it <code>result<\/code>. It was this that I ran and viewed. After viewing, I resized it for the model back to being a flat image. During this, I change the names of some things calling it first <code>image<\/code> and then <code>result<\/code>. If the code just red from top to bottom I think it would have been fine but Tensorflow has the graph that it works from and so it was expecting result to be a 3d tensor when it got to the second iteration, but it was really a flat tensor that was trying to be put through <code>tf.map_fn<\/code>. <code>tf.map_fn<\/code> is something that resides on the graph and does not get executed until <code>sess.run()<\/code> and so this would be the cause of the error that I was seeing. I hope this makes sense and helps someone else down the road. <\/p>\n\n<pre><code>with tf.name_scope('TestingData'):\n    X_test_batch, y_test_batch = inputs(FLAGS.train_dir,\n                                       FLAGS.test_file,\n                                       FLAGS.batch_size,\n                                       FLAGS.n_epochs,\n                                       FLAGS.n_classes,\n                                       one_hot_labels=True,\n                                       imshape=160*160*3)\n\nimage = tf.reshape(X_test_batch, [-1, 160, 160, 3])\n\nimage = tf.map_fn(lambda img: color_distorer(img), image)\nwith tf.Session() as sess:\n\n    init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())\n     sess.run(init_op)\n     coord = tf.train.Coordinator()\n     threads = tf.train.start_queue_runners(coord=coord)\n     # X = sess.run([X_test_batch])\n     for i in range(5):\n         # dX, dy = sess.run([result, y_test_batch])\n         print('******************')\n         image = tf.reshape(image, [-1, 76800])\n         dX, dy = sess.run([image, y_test_batch])\n         distorted = np.reshape(dX, (-1, 160, 160, 3))\n         print(dX.shape)\n         print(distorted.shape)\n         for i in range(50):\n             if i%25 ==0:\n                  print(distorted[i], dy[i])\n                  plt.title(dy[i])\n                  plt.imshow(distorted[i])\n                  plt.show()\n                  print(dX.shape)\n<\/code><\/pre>\n"}]}
{"QuestionId":41535347,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-01-08T17:06:14.807","AcceptedAnswerId":41535880.0,"OwnerUserId":6306884.0,"Title":"How Gradient passed by tf.py_func","Body":"<p>This is Faster R-CNN implement in tensorflow.<br>\nThe <a href=\"https:\/\/github.com\/smallcorgi\/Faster-RCNN_TF\/blob\/master\/lib\/networks\/network.py#L159\" rel=\"nofollow noreferrer\">proposal_layer<\/a> is <a href=\"https:\/\/github.com\/smallcorgi\/Faster-RCNN_TF\/blob\/master\/lib\/rpn_msr\/proposal_layer_tf.py\" rel=\"nofollow noreferrer\">implement<\/a> by python    <\/p>\n\n<p>i am curious about if gradient can pass by <br>\nthe weights and biases are keep changing<br>\nso I think the gradient deliver back successful<\/p>\n\n<p>Then I do a small test  <\/p>\n\n<pre><\/pre>\n\n<p>Variable  keep changing, but  keep the value after initialize and never change    <\/p>\n\n<p> get value<br>\n get <\/p>\n\n<p>I know some questions suggest use  and <br>\nbut I can't find these in the faster rcnn repo(link on top of post)<br>\nam I do something wrong or missing something so that  is freeze<\/p>\n","answers":[{"AnswerId":"41535880","CreationDate":"2017-01-08T17:56:27.577","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Gradient of <code>py_func<\/code> is <code>None<\/code> (just check <code>ops.get_gradient_function(y2.op)<\/code>). There's this <a href=\"https:\/\/gist.github.com\/harpone\/3453185b41d8d985356cbe5e57d67342\" rel=\"noreferrer\">gist<\/a> by @harpone which shows how to use gradient override map for py_func.<\/p>\n\n<p>Here's your example modified to use that recipe<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ndef addone(x):\n    # print(type(x)\n    return x + 1\n\ndef addone_grad(op, grad):\n    x = op.inputs[0]\n    return x\n\nfrom tensorflow.python.framework import ops\nimport numpy as np\n\n# Define custom py_func which takes also a grad op as argument:\ndef py_func(func, inp, Tout, stateful=True, name=None, grad=None):\n\n    # Need to generate a unique name to avoid duplicates:\n    rnd_name = 'PyFuncGrad' + str(np.random.randint(0, 1E+8))\n\n    tf.RegisterGradient(rnd_name)(grad)  # see _MySquareGrad for grad example\n    g = tf.get_default_graph()\n    with g.gradient_override_map({\"PyFunc\": rnd_name}):\n        return tf.py_func(func, inp, Tout, stateful=stateful, name=name)\n\ndef pyfunc_test():\n\n    # create data\n    x_data = tf.placeholder(dtype=tf.float32, shape=[None])\n    y_data = tf.placeholder(dtype=tf.float32, shape=[None])\n\n    w = tf.Variable(tf.constant([0.5]))\n    b = tf.Variable(tf.zeros([1]))\n\n    y1 = tf.mul(w, x_data, name='y1')\n    y2 = py_func(addone, [y1], [tf.float32], grad=addone_grad)[0]\n    y = tf.add(y2, b)\n\n    loss = tf.reduce_mean(tf.square(y - y_data))\n    optimizer = tf.train.GradientDescentOptimizer(0.01)\n    train = optimizer.minimize(loss)\n\n    print(\"Pyfunc grad\", ops.get_gradient_function(y2.op))\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n\n        for step in range(10):\n            #            ran = np.random.rand(115).astype(np.float32)\n            ran = np.ones((115)).astype(np.float32)\n            ans = ran * 1.5 + 3\n            dic = {x_data: ran, y_data: ans}\n            tt, yy, yy1= sess.run([train, y1, y2], feed_dict=dic)\n            if step % 1 == 0:\n                print('step {}'.format(step))\n                print('{}, {}'.format(w.eval(), b.eval()))\n\n        test = sess.run(y, feed_dict={x_data:[1]})\n        print('test = {}'.format(test))\n\n\nif __name__ == '__main__':\n    pyfunc_test()\n<\/code><\/pre>\n"}]}
{"QuestionId":41535845,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><skflow><tflearn>","CreationDate":"2017-01-08T17:52:49.100","AcceptedAnswerId":41545410.0,"OwnerUserId":3169556.0,"Title":"which higher layer abstraction to use for tensorflow","Body":"<p>I am looking for higher layer abstractions for my deep learning project.<\/p>\n\n<p>Few doubts lately.<\/p>\n\n<ol>\n<li><p>I am really confused about which is more actively maintained <a href=\"https:\/\/github.com\/tflearn\/tflearn\" rel=\"nofollow noreferrer\">tflearn<\/a>(<a href=\"http:\/\/tflearn.org\" rel=\"nofollow noreferrer\">docs<\/a>), or <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/learn\/python\/learn\" rel=\"nofollow noreferrer\">tensorflow.contrib.learn<\/a>. But projects are different and actively contributed on Github. I did not find why are people working this way, same goal, same name, but working differently.<\/p><\/li>\n<li><p>That was not enough, we also have <a href=\"https:\/\/github.com\/tensorflow\/skflow\" rel=\"nofollow noreferrer\">skflow<\/a>, why do we have this project separately, this aims to mimic scikit-learn like functionality for deep learning(just like <strong>tflearn<\/strong> do).<\/p><\/li>\n<li><p>There are more and more coming, which one choose, and which one will be maintained in future?  <\/p><\/li>\n<\/ol>\n\n<p>Any ideas?<\/p>\n\n<p><strong>PS<\/strong>: I know this might get closed. but I would definitely want some answers first. Those wanting it closed, please care to drop a reason\/hint\/link in comments<\/p>\n","answers":[{"AnswerId":"41545410","CreationDate":"2017-01-09T10:05:09.767","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>What about keras (<a href=\"https:\/\/keras.io\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/<\/a>)? It is easy to use. However you can do pretty much everything you want with it. It uses either theano or tensorflow as its backend. Kaggle contests are often solved using keras (e.g. <a href=\"https:\/\/github.com\/EdwardTyantov\/ultrasound-nerve-segmentation\" rel=\"nofollow noreferrer\">https:\/\/github.com\/EdwardTyantov\/ultrasound-nerve-segmentation<\/a>).<\/p>\n\n<p>Edit:<\/p>\n\n<p>Because you did not specify python I would also recommend matconvnet if you look for more abstraction. <\/p>\n"}]}
{"QuestionId":41537583,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-08T20:35:23.450","AcceptedAnswerId":41539608.0,"OwnerUserId":1501285.0,"Title":"How to make a simple prediction in Tensorflow on a trained model?","Body":"<p>I have just trained a model like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now I have loaded an image like this: <\/p>\n\n<p>Now the only thing I would like to do is to make a prediction based on .<\/p>\n\n<p><em>How do I do this?<\/em><\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>Added my softmax function:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41539608","CreationDate":"2017-01-09T00:58:33.867","ParentId":null,"OwnerUserId":"6772672","Title":null,"Body":"<p>That depends on how you defined your graph, and depends on how you defined the shape of the 'x' placeholder.<br>\nSupposing 'x' is defined like this:<\/p>\n\n<pre><code>x = tf.placeholder(shape=[None, IMG_WIDTH, IMG_HEIGHT, NUM_COLOR_CHANNELS], dtype=tf.float32)\n<\/code><\/pre>\n\n<p>And supposing 'pred' is the tensor that that gives you the predictions, you just need to evaluate this tensor:<\/p>\n\n<pre><code>predictions = sess.run(pred, feed_dict={x: img1})\n<\/code><\/pre>\n"}]}
{"QuestionId":41538692,"AnswerCount":2,"Tags":"<tensorflow><sparse-matrix><keras>","CreationDate":"2017-01-08T22:42:43.650","AcceptedAnswerId":null,"OwnerUserId":785494.0,"Title":"Using sparse matrices with Keras and Tensorflow","Body":"<p>My data can be viewed as a matrix of 10B entries (100M x 100), which is very sparse (&lt; 1\/100 * 1\/100 of entries are non-zero).  I would like to feed the data into into a Keras Neural Network model which I have made, using a Tensorflow backend.<\/p>\n\n<p>My first thought was to expand the data to be dense, that is, write out all 10B entries into a series of CSVs, with most entries zero.  However, this is quickly overwhelming my resources (even doing the ETL overwhelmed pandas and is causing postgres to struggle).  So I need to use true sparse matrices.<\/p>\n\n<p>How can I do that with Keras (and Tensorflow)? While numpy doesn't support sparse matrices, scipy and tensorflow both do.  There's lots of discussion (e.g.  <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/1886\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/pull\/1886<\/a> <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/3695\/files\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/pull\/3695\/files<\/a> <a href=\"https:\/\/github.com\/pplonski\/keras-sparse-check\" rel=\"noreferrer\">https:\/\/github.com\/pplonski\/keras-sparse-check<\/a> <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/odsQBcNCdZg\" rel=\"noreferrer\">https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/odsQBcNCdZg<\/a> ) about this idea - either using scipy's sparse matrixcs or going directly to Tensorflow's sparse matrices.  But I can't find a clear conclusion, and I haven't been able to get anything to work (or even know clearly which way to go!).<\/p>\n\n<p>How can I do this?<\/p>\n\n<p>I believe there are two possible approaches:<\/p>\n\n<ol>\n<li>Keep it as a scipy sparse matrix, then, when giving Keras a minibatch, make it dense<\/li>\n<li>Keep it sparse all the way through, and use Tensorflow Sparse Tensors<\/li>\n<\/ol>\n\n<p>I also think #2 is preferred, because you'll get much better performance all the way through (I believe), but #1 is probably easier and will be adequate.  I'll be happy with either.<\/p>\n\n<p>How can either be implemented?<\/p>\n","answers":[{"AnswerId":"43188970","CreationDate":"2017-04-03T15:46:58.477","ParentId":null,"OwnerUserId":"4161128","Title":null,"Body":"<p>Sorry, don't have the reputation to comment, but I think you should take a look at the answer here: <a href=\"https:\/\/stackoverflow.com\/questions\/37609892\/keras-sparse-matrix-issue\">Keras, sparse matrix issue<\/a>. I have tried it and it works correctly, just one note though, at least in my case, the shuffling led to really bad results, so I used this slightly modified non-shuffled alternative:<\/p>\n\n<pre><code>def nn_batch_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    index = np.arange(np.shape(y_data)[0])\n    while 1:\n        index_batch = index[batch_size*counter:batch_size*(counter+1)]\n        X_batch = X_data[index_batch,:].todense()\n        y_batch = y_data[index_batch]\n        counter += 1\n        yield np.array(X_batch),y_batch\n        if (counter &gt; number_of_batches):\n            counter=0\n<\/code><\/pre>\n\n<p>It produces comparable accuracies to the ones achieved by keras's shuffled implementation (setting <code>shuffle=True<\/code> in <code>fit<\/code>).<\/p>\n"},{"AnswerId":"56326515","CreationDate":"2019-05-27T12:53:05.373","ParentId":null,"OwnerUserId":"4805432","Title":null,"Body":"<p>This answer addresses the second approach mentioned in the question.\nIt is possible to use sparse matrices as inputs to a Keras model with the Tensorflow backend if you write a custom training loop. In the example below, the model takes a sparse matrix as an input and outputs a dense matrix.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import Dense, Input\nfrom keras.models import Model\nimport scipy\nimport numpy as np\n\ntrainX = scipy.sparse.random(1024, 1024)\ntrainY = np.random.rand(1024, 1024)\n\ninputs = Input(shape=(trainX.shape[1],), sparse=True)\noutputs = Dense(trainY.shape[1], activation='softmax')(inputs)\nmodel = Model(inputs=inputs, outputs=outputs)\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n\nsteps = 10\nfor i in range(steps):\n  # For simplicity, we directly use trainX and trainY in this example\n  # Usually, this is where batches are prepared\n  print(model.train_on_batch(trainX, trainY))\n# [3549.2546, 0.0]\n# ...\n# [3545.6448, 0.0009765625]\n<\/code><\/pre>\n\n<p>However, the usefulness of this approach depends on whether your model needs to densify the sparse matrix.\nIndeed, the above model has one layer which transforms the sparse matrix into a dense one. This can be a problem if your sparse matrix doesn't fit in memory.<\/p>\n"}]}
{"QuestionId":41539156,"AnswerCount":2,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-08T23:42:13.940","AcceptedAnswerId":null,"OwnerUserId":2241766.0,"Title":"Pre-train neural networks in TensorFlow","Body":"<p>I'm training a deep neural network. Training directly the full network is difficult, therefore, I prefer to train layer by layer. Firstly, I train a network with a single hidden layer. After that, I save the model using . Next time, I restore variables by calling: <\/p>\n\n<pre><\/pre>\n\n<p>Of course, this time, the architecture has changed because I insert a new layer between the hidden layer and the output layer, and a new variable is used. For this reason, Tensorflow throws an error: <\/p>\n\n<p>My question is that: How to train neural network layer by layer in Tensorflow? I want to do it before final fine-tuning of the full network.\nMany thanks.<\/p>\n","answers":[{"AnswerId":"41593333","CreationDate":"2017-01-11T14:21:56.110","ParentId":null,"OwnerUserId":"2241766","Title":null,"Body":"<p>Finally, I come to a simple solution. In Python, a numpy array can be saved easily by using <code>numpy.save<\/code>. After that, we can load the array by: <code>numpy.load<\/code>. Therefore, after training, I can evaluate the variable using <code>var.eval()<\/code> to convert it to a numpy array, then save it.<\/p>\n"},{"AnswerId":"41593173","CreationDate":"2017-01-11T14:13:32.757","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>Try to define the entire network with shared weights.\nInstead of training over all the layers define a function that runs just the layers that you want and finally save the entire network.<\/p>\n"}]}
{"QuestionId":41539658,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><convolution>","CreationDate":"2017-01-09T01:08:29.357","AcceptedAnswerId":41540092.0,"OwnerUserId":1145744.0,"Title":"Tensorflow error when I try to use tf.contrib.layers.convolution2d","Body":"<p>When I invoke tf.contrib.layers.convolution2d the tensorflow execution terminates with an error about one of the parameters used<\/p>\n\n<pre><\/pre>\n\n<p>The parameter passed are the follows:<\/p>\n\n<pre><\/pre>\n\n<p>That is exactly as described in the book that I'm reading. I suspect a possible syntax problem with  written directly inside the call, but I don't know how to fix. I'm using Tensorflow 0.12.0 <\/p>\n","answers":[{"AnswerId":"41540092","CreationDate":"2017-01-09T02:18:41.740","ParentId":null,"OwnerUserId":"1310924","Title":null,"Body":"<p>The book that you are reading (You didn't mention which one) might be using an older version of TensorFlow when the initial values for the weight tensor was passed through the <code>weight_init<\/code> argument. In the TensorFlow library version you are using (You didn't mention your TF version), probably that argument is replaced with <code>weight_initializer<\/code>. The latest (TensorFlow v0.12.0) documentation for <code>tf.contrib.layers.convolution2d<\/code> is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.layers\/higher_level_ops_for_building_neural_network_layers_#convolution2d\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>To fix your problem, you can change the following line in your code:<\/p>\n\n<p><code>weight_init=tf.random_normal<\/code><\/p>\n\n<p>to<\/p>\n\n<p><code>weight_initializer=tf.random_normal_initializer()<\/code><\/p>\n\n<p>According to the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops\/sharing_variables?hl=bn#random_normal_initializer\" rel=\"nofollow noreferrer\">documentation<\/a>, by default, <code>tf.random_normal_initialier<\/code> uses a 0.0 mean, a standard deviation of 1.0 and the datatype to be tf.float32. You may change the arguments as per your need using this line instead:\n<code>weight_initializer=tf.random_normal_initializer(mean=0.0, stddev=1.0, seed=None, dtype=tf.float32)<\/code><\/p>\n"}]}
{"QuestionId":41539934,"AnswerCount":0,"Tags":"<python><tensorflow><distributed-computing>","CreationDate":"2017-01-09T01:54:34.950","AcceptedAnswerId":null,"OwnerUserId":6527887.0,"Title":"tensorflow Parameter to MergeFrom() must be instance of same class: expected TensorProto got Variable. for field Value.tensor","Body":"<p>I am trying to save variables through checkpoints to introduce fault tolerance to my program. The following is my configuration:-<\/p>\n\n<pre><\/pre>\n\n<p>In this, the  saves tf.Variable in  and  takes a tensor input in . The error I get is:-<\/p>\n\n<pre><\/pre>\n\n<p>How do I rectify this?<\/p>\n","answers":[]}
{"QuestionId":41540073,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-09T02:14:27.113","AcceptedAnswerId":null,"OwnerUserId":7392469.0,"Title":"why the GPU memory is not the same?","Body":"<p>I use two methods to train my own data, the first method is training a model from scratch, the second is using fine-tuning(according to <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim<\/a>), all parameters are the same for two methods, except checkpoints setting, however, the first method always occurs out of memory (GPU). When decreasing the batch size for the first method, it's ok. what's the reason?<\/p>\n\n<p>the more information can be found in website: <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/848\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/issues\/848<\/a><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41543118","CreationDate":"2017-01-09T07:39:42.827","ParentId":null,"OwnerUserId":"639559","Title":null,"Body":"<p>The batch size indicates how many test-cases you are uploading to the gpu for training at the same time. Reducing the batch-size will reduce the memory-footprint for each training step. To get the maximum performance of your training, you would probably like to find the \"sweet spot\"(the largest batch size without running into OOM) for your training-set and hardware.<\/p>\n"}]}
{"QuestionId":41540177,"AnswerCount":1,"Tags":"<python><csv><tensorflow>","CreationDate":"2017-01-09T02:29:29.823","AcceptedAnswerId":41540755.0,"OwnerUserId":7238189.0,"Title":"TensorFlow Iris load_csv_with_header( )","Body":"<p>In the TensorFlow quick start tutorial which folder am I suppose to put the iris training\/test csv files as to let me use this code. I'm specifically having problems using the load_csv_with_header module and being new to programming I'm at a lost when they say \"Place these files in the same directory as your Python code.\" To my eyes there appears to be a several directories which may satisfy this but at last no results. I've also tried typing getcwd() into the interpreter but it says no module defined. I'm using python 2.7 with the most current version of andaconda to do this. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41540755","CreationDate":"2017-01-09T03:47:30.270","ParentId":null,"OwnerUserId":"1310924","Title":null,"Body":"<p>Let's say you have saved the tutorial code in your computer as <strong>tf_learn_tutorial.py<\/strong>\nin a folder named <strong>tutorial1<\/strong>. Then you should make sure you download and save the <strong>iris_training.csv<\/strong> file and <strong>iris_test.csv<\/strong> file in the same folder named <strong>tutorial1<\/strong> where your <strong>.py<\/strong> file exists.<\/p>\n\n<p>After which the folder <strong>tutorial1<\/strong> on your computer should have (at least) these three files:<\/p>\n\n<ul>\n<li>tf_learn_tutorial.py<\/li>\n<li>iris_training.csv<\/li>\n<li>iris_test.csv<\/li>\n<\/ul>\n\n<p>To make it even simpler for you to quick start the tutorial,<\/p>\n\n<p><strong>Step 1<\/strong>: Download this <a href=\"http:\/\/pastebin.com\/raw\/C7J2Nm6c\" rel=\"nofollow noreferrer\">file (tf_learn_tutorial.py)<\/a> and save it as <strong>tf_learn_tutorial.py<\/strong> on your computer in a folder named <strong>tutorial1<\/strong> anywhere on your hard disk as you like.<\/p>\n\n<p><strong>Step 2<\/strong>: Download this <a href=\"http:\/\/download.tensorflow.org\/data\/iris_training.csv\" rel=\"nofollow noreferrer\">file (iris_training.csv)<\/a> and save it on your computer in the same folder as above named <strong>tutorial1<\/strong>.<\/p>\n\n<p><strong>Step 3<\/strong>: Download this <a href=\"http:\/\/download.tensorflow.org\/data\/iris_test.csv\" rel=\"nofollow noreferrer\">file (iris_test.csv)<\/a> and save it on your computer in the same folder as above named <strong>tutorial1<\/strong>.<\/p>\n\n<p><strong>Step 4<\/strong>: Run the <strong>tf_learn_tutorial.py<\/strong> file from the <strong>tutorial1<\/strong> directory:<\/p>\n\n<p><code>python tf_learn_tutorial.py<\/code>.<\/p>\n"}]}
{"QuestionId":41541626,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><deep-learning><tf-slim>","CreationDate":"2017-01-09T05:38:27.730","AcceptedAnswerId":null,"OwnerUserId":5107084.0,"Title":"TensorFlow: Do we exclude a scope or operation when we fine-tune certain layers of a pre-trained model?","Body":"<p>I'm quite confused about what a scope or operation means when I'm referring to a layer of a network. If I want to fine-tune the last softmax layer of the network, do I go to the scope, the operation or the variables to work on it? What are the differences between all these terms?<\/p>\n\n<p>I have looked at the TF-slim walkthrough ipynb tutorial, and here was how they excluded some scopes for fine-tuning:<\/p>\n\n<pre><\/pre>\n\n<p>It seems that they excluded certain scopes , from which they excluded each variable from the scope they want to exclude, and included whatever variable from scopes they did not list into . Is it safe to say that the scope actually refer to the layers?<\/p>\n\n<p>If so, this is the confusing part: what is the relevance of variables to the op? I have the impression that an op.name is used to find the scope_name like in , for instance, through writing . If that is the case, why would a variable still have an ?<\/p>\n\n<p>How should one find the scope names to select certain layers to fine-tune? I think this will be very important in clearing my confusion.<\/p>\n\n<p>Thank you all for your help.<\/p>\n","answers":[]}
{"QuestionId":41542571,"AnswerCount":0,"Tags":"<python><garbage-collection><tensorflow><pyqtgraph>","CreationDate":"2017-01-09T06:59:55.917","AcceptedAnswerId":null,"OwnerUserId":1616677.0,"Title":"PyQtGraph tries to put TensorFlow stuff onto a QGraphicsScene() on cleanup","Body":"<pre><\/pre>\n\n<p>Looking at the code I see  comes from . I cannot realize how garbage collector can hold a deleted weak reference if its main purpose is to cleanup such an objects?? Also don't know how to make any workaround. This is extract from real code, I need all the libraries imported.<\/p>\n\n<p>This is target source:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41542706,"AnswerCount":1,"Tags":"<image-processing><machine-learning><tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-01-09T07:09:04.270","AcceptedAnswerId":41550026.0,"OwnerUserId":3733913.0,"Title":"Tensorflow: Label inception has no images in the category training","Body":"<p>I am retraining the Inception v3 network based on the code here: <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/#0\" rel=\"nofollow noreferrer\">https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/#0<\/a>.\nI have a set of 200 labels. Each label has images with their count ranging from 50 to 15000. While retraining I get the following error: <\/p>\n\n<pre><\/pre>\n\n<p>Search on google has made me understand that such errors can happen if images are less than 20. This can happen because their may not be sufficient images left for validation. However I have at least 50 images. So why am I sill getting this error?<\/p>\n","answers":[{"AnswerId":"41550026","CreationDate":"2017-01-09T14:20:18.993","ParentId":null,"OwnerUserId":"2314651","Title":null,"Body":"<p>I looked at the <code>tensorflow\/examples\/image_retraining\/retrain.py<\/code> file which is linked from the tutorial in your post.<\/p>\n\n<p>Part of the code that runs is a train\/validation split that takes place in the function <code>create_image_lists(image_dir, testing_percentage, validation_percentage)<\/code><\/p>\n\n<p>The default value for the split is 10% and it's stated in the <code>FLAGS.validation_percentage<\/code><\/p>\n\n<p>Since you have classes that have less than 200 images, the split return classes with less than 20 photos for validation and that's where the error you are getting occur.<\/p>\n\n<p>Try to run the code with only the classes with more than 200 images and see if it helps. if it does you can consider adding more images or manipulate the create_image_lists function to return at least 20 photos for validation.<\/p>\n"}]}
{"QuestionId":41542806,"AnswerCount":1,"Tags":"<python-2.7><deep-learning><pickle><theano><convolution>","CreationDate":"2017-01-09T07:15:19.660","AcceptedAnswerId":null,"OwnerUserId":6738654.0,"Title":"difficulty in saving the trained model for my CNN sentence classifer at test it on a single instance later by loading the model","Body":"<p>I am using the cnn text classification written by <a href=\"https:\/\/github.com\/yoonkim\/CNN_sentence\" rel=\"nofollow noreferrer\">Yoon Kim<\/a> for sentiment analysis. This code applies cross validation to check the quality of learned model. However, I want to save the learned weights and biases, so I can apply the learned model on new instances one by one for the prediction purpose. I appreciate if someone provides an example of how I can do that. I know that I should use pickle load and dumb to do that, but I am not sure exactly which part of the code I should use them I want to have a separate test.py file so I can only test the trained model on a test sample without training the model again. how I should save and then predict based on saved model?<\/p>\n","answers":[{"AnswerId":"41563604","CreationDate":"2017-01-10T07:42:49.843","ParentId":null,"OwnerUserId":"3737792","Title":null,"Body":"<p>If you don't need pickle for some special reason. You can easily use numpy's save and load functions:<\/p>\n\n<pre><code>#saving the parameters\nparams = classifier.params\nnp.save(\"net_parameters\", params)\n\n#loading the parameters\nparams = np.load(\"net_parameters.npy\")\nclassifier.params = params\n<\/code><\/pre>\n"}]}
{"QuestionId":41543687,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><tf-slim>","CreationDate":"2017-01-09T08:19:34.077","AcceptedAnswerId":42683243.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: What is the purpose of endpoints for data parallelism when training across multiple machines?","Body":"<p>In the  source code, there was an endpoint indicated in the creation of its loss function:<\/p>\n\n<pre><\/pre>\n\n<p>Source: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/train_image_classifier.py#L471-L477\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/train_image_classifier.py#L471-L477<\/a><\/p>\n\n<p>My idea is that there are multiple identical networks that are trained in separate machines, and the variables and parameters are averaged out in the end to merge into one network (is this correct?). But I don't quite get what the purpose of endpoints are in this case, as I thought the network_fn ought to produce only logits for predictions. What's the use of end_points?<\/p>\n","answers":[{"AnswerId":"42683243","CreationDate":"2017-03-08T22:36:39.687","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The <code>endpoints<\/code> in this case just track the different outputs of the model. The <code>AuxLogits<\/code> one, for example, has the logits. <\/p>\n"}]}
{"QuestionId":41543774,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-09T08:25:07.680","AcceptedAnswerId":null,"OwnerUserId":1499038.0,"Title":"InvalidArgumentError: for tensor bool tensorflow==0.12.1","Body":"<p>I have experimented with  and I have following problem:<\/p>\n\n<p>I am trying to run:<\/p>\n\n<pre><\/pre>\n\n<p>But I'm getting following stack trace:<\/p>\n\n<blockquote>\n  <p>File \"myscript.py\", line 35, in tf_run\n      sess.run(tf.global_variables_initializer())\n      File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\",\n  line 766, in run\n      run_metadata_ptr)\n      File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\",\n  line 964, in _run\n      feed_dict_string, options, run_metadata)\n      File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\",\n  line 1014, in _do_run\n      target_list, options, run_metadata)\n      File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/client\/session.py\",\n  line 1034, in _do_call\n      raise type(e)(node_def, op, message)\n      InvalidArgumentError: You must feed a value for placeholder tensor 'training' with dtype bool      [[Node: training =\n  Placeholderdtype=DT_BOOL, shape=[],\n  _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n\n<pre><\/pre>\n  \n  <p>line 1587, in placeholder\n      name=name)\n          File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/ops\/gen_array_ops.py\",\n  line 2043, in _placeholder\n      name=name)\n          File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/op_def_library.py\",\n  line 759, in apply_op\n      op_def=op_def)\n         File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 2240, in create_op\n      original_op=self._default_original_op, op_def=op_def)\n         File \"\/usr\/local\/lib\/python2.7\/dist-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 1128, in <strong>init<\/strong>\n      self._traceback = _extract_stack()<\/p>\n<\/blockquote>\n\n<p>This problem appears on:\n installation\n only installation<\/p>\n\n<p>When I try to downgrade to  and replace line <\/p>\n\n<p>\n by \n, then everything works well.<\/p>\n\n<p>Related issue in tensorflow repo is here <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6661\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6661<\/a><\/p>\n","answers":[]}
{"QuestionId":41543954,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-01-09T08:37:50.127","AcceptedAnswerId":41690170.0,"OwnerUserId":482601.0,"Title":"Remove image outputs from TensorBoard","Body":"<p>My TensorBoard log files grow huge because \u2013 it seems \u2013 every image summary ever generated is stored. This even though in TensorBoard, it seems like I can only look at the most recent image. And I only need the most recent image anyway.<\/p>\n\n<p>Is there a way to let TensorBoard know that I only need the latest iamge? I looked at the  API docs but there is no obvious flag.<\/p>\n","answers":[{"AnswerId":"41690170","CreationDate":"2017-01-17T06:14:52.560","ParentId":null,"OwnerUserId":"1653720","Title":null,"Body":"<p>Hi I work on TensorBoard. To the best of my knowledge, the logs are append only. However when TensorBoard loads them into memory, it uses reservoir sampling so they don't consume all your memory. In the future, we may be implementing a system that will reservoir sampling during the writing phase, or possibly, a tool for compressing logs so they only contain what TensorBoard needs.<\/p>\n\n<p>While TensorBoard image dashboard only shows the most recent image at the moment, we'd be hesitant to write tools to remove the previous ones, since we may be extending the dashboard to show more than the most recent sample.<\/p>\n"}]}
{"QuestionId":41544009,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-09T08:42:06.247","AcceptedAnswerId":41551769.0,"OwnerUserId":5443758.0,"Title":"Distributed Tensorflow is getting stuck at sess.run()","Body":"<p>I want to run tensorflow on multiple machines, multiple GPUs. As an initial step, trying out distributed tensorflow on single machine (following tensorflow tutorial <a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/distributed\/<\/a>)<\/p>\n\n<p>Bellow are the lines after which sess.run() stucks<\/p>\n\n<pre><\/pre>\n\n<p>Everything is working fine till here, but when I am running sess.run(), it stucks.<\/p>\n\n<pre><\/pre>\n\n<p>If anybody has already worked on distributed tensorflow, Please let me know the solution or other tutorial which works fine.<\/p>\n","answers":[{"AnswerId":"41551769","CreationDate":"2017-01-09T15:52:12.780","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>By default, Distributed TensorFlow will block until all servers named in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/distributed_execution#ClusterSpec\" rel=\"nofollow noreferrer\"><code>tf.train.ClusterSpec<\/code><\/a> have started. This happens during the first interaction with the server, which will typically be the first <code>sess.run()<\/code> call. Therefore, if you haven't also started a server listening on <code>localhost:2223<\/code>, then TensorFlow will block until you do.<\/p>\n\n<p>There are a few solutions to this problem, depending on your later goals:<\/p>\n\n<ol>\n<li><p>Start a server on <code>localhost:2223<\/code>. In another process, run the following script:<\/p>\n\n<pre><code> import tensorflow as tf\n cluster = tf.train.ClusterSpec({\"local\": [\"localhost:2222\", \"localhost:2223\"]})\n server = tf.train.Server(cluster, job_name=\"local\", task_index=1)\n server.join()  # Wait forever for incoming connections.\n<\/code><\/pre><\/li>\n<li><p>Remove task 1 from the original <code>tf.train.ClusterSpec<\/code>:<\/p>\n\n<pre><code> import tensorflow as tf\n cluster = tf.train.ClusterSpec({\"local\": [\"localhost:2222\"]})\n server = tf.train.Server(cluster, job_name=\"local\", task_index=0)\n # ...\n<\/code><\/pre><\/li>\n<li><p>Specify a \"device filter\" when you create the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session\" rel=\"nofollow noreferrer\"><code>tf.Session<\/code><\/a> so that the session only uses task 0.<\/p>\n\n<pre><code> # ...\n sess = tf.Session(\"grpc:\/\/localhost:2222\",\n                   config=tf.ConfigProto(device_filters=[\"\/job:local\/task:0\"]))\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":41544096,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-09T08:47:33.503","AcceptedAnswerId":null,"OwnerUserId":7393420.0,"Title":"How to concatenate a tf.Variable to achieve Mahalanobis regularization in the loss function","Body":"<p>I'm trying to define a loss in tensorflow, which contains a regularization term. However, this is not the standard l2 or l1 norm. This is a Mahalanobis norm on the weights, with a covariance matrix P. <\/p>\n\n<p>For example, let's say I have a neural net with two layers which contains the following weights (tf.Variables): W1, b1, W2, b2. In order to calculate the Mahalanobis norm I need to concatenate all the weights into one vector, let's call it \"theta\" where theta=[W1(:), b1(:), W2(:), b2(:)]^T. Then I want to perform \"theta^T P theta\". <\/p>\n\n<p>My question is: How do I change the shape of a tf.Variable, while keeping the following properties:<\/p>\n\n<ol>\n<li>The loss contains tf.variables which have 2 representations: One representation of 2-D tensors (W1, b1, W2, b2, inside the neural network) and another representation as a concatenate vector theta (in the regularization term). For example, the weights inside W1 appear both in the 2-D tensor W1 and also inside theta. <\/li>\n<li>The minimizer should \"know\" that when it calculates the gradient with respect to the  weights (W1, b1, W2, b2), it should take into consideration the 2 representations of the weight as explained above.<\/li>\n<\/ol>\n\n<p>I add some code to illustrate my question. How do I define \"theta\" as a concatenated vector of (W1, b1, W2, b2), with the properties mentioned above?<\/p>\n\n<p>I searched for answers to my question, but all the regularizations I saw were l2 or l1 norms on the weights, and not something that requires a concatenation of the weights, as in my case.  <\/p>\n\n<p>Thanks!<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41545692,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><deep-learning><tf-slim>","CreationDate":"2017-01-09T10:19:30.013","AcceptedAnswerId":null,"OwnerUserId":5107084.0,"Title":"TensorFlow: Error reported to Coordinator: <type 'exceptions.TypeError'> despite specifying a checkpoint to restore the model?","Body":"<p>I am trying to test out transfer learning using the TF-slim source code found in here: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/train_image_classifier.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/train_image_classifier.py<\/a><\/p>\n\n<p>The dataset I used was the flowers dataset, which I have already converted into TFRecords as required by the code. I also took the precaution to not save the TFRecords files inside my tmp folder because it will face permission errors, which I had encountered previously. However, despite specifying the checkpoint to restore, I got a  stdout which is puzzling. The error then goes on to give me a TypeError, which can be found in the error trace below.<\/p>\n\n<p>I ran the following command:\n<\/p>\n\n<p>Why was TensorFlow not able to detect the presence of a .ckpt file to restore the model and fine-tune from then onwards?<\/p>\n\n<p>Here is my error stack trace:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41545802,"AnswerCount":1,"Tags":"<python><debugging><tensorflow><anaconda>","CreationDate":"2017-01-09T10:25:01.367","AcceptedAnswerId":null,"OwnerUserId":7393769.0,"Title":"Tensorflow. Kernel died when training. Window Anaconda","Body":"<pre><\/pre>\n\n<h1>Explain.<\/h1>\n\n<p>This is MNIST code using softmax.\nThe problem appears at <\/p>\n\n<p>sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})<\/p>\n\n<p>in the for loop.<\/p>\n\n<p>Just kernel died and restarting with any error message.\nMaybe the code is not the problem because it works well on other guy.<\/p>\n\n<p>I'm using Windows10 Anaconda.<\/p>\n\n<p>What is the problem?<\/p>\n","answers":[{"AnswerId":"45211805","CreationDate":"2017-07-20T10:14:20.353","ParentId":null,"OwnerUserId":"8337439","Title":null,"Body":"<p>I came across similar problem as yours. It is likely that you installed <code>cuda<\/code> and <code>cudnn<\/code> and are running codes on <code>tensorflow-gpu<\/code>. <\/p>\n\n<p>In my case, I at first installed <code>cuda8.0<\/code> and <code>cudnn v6.0 for cuda8.0<\/code>, and got the kernel died problem. <\/p>\n\n<p>Then I changed the <code>cudnn<\/code> version to <code>cudnn v5.1 for cuda8.0<\/code> and solved this problem. Now I am working fine with my environment.<\/p>\n"}]}
{"QuestionId":41547497,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-01-09T11:53:42.047","AcceptedAnswerId":null,"OwnerUserId":3282997.0,"Title":"Keras input issue","Body":"<p>I'm struggling with very strange issue with Keras (version 1.2.0).\nHere is my NN model:<\/p>\n\n<pre><\/pre>\n\n<p>the input is a vector of 302 floats and I have 376 of them in the training sample. \nnow, when I run model.fit(),I get error message: <\/p>\n\n<pre><\/pre>\n\n<p>This is how I create my input:<\/p>\n\n<pre><\/pre>\n\n<p>I've tried converting indata to numpy array using asarray etc, but nothing works! Do you have any suggestions? It is second day I'm looking for a solution...<\/p>\n\n<p>Mariusz<\/p>\n","answers":[{"AnswerId":"42031713","CreationDate":"2017-02-03T19:31:56.250","ParentId":null,"OwnerUserId":"2235576","Title":null,"Body":"<p>try to modify your loop like:<\/p>\n\n<pre><code>for line in datasample:\nelems=line.split(\",\")\nindata= [[float(elems[i])] for i in range(2,len(elems)-1)]\nall_indata.append(indata)\n<\/code><\/pre>\n\n<p>After the loop apply numpy trasformation<\/p>\n\n<pre><code>numpy.array(all_indata)\n<\/code><\/pre>\n"}]}
{"QuestionId":41547764,"AnswerCount":1,"Tags":"<python><optimization><tensorflow><deep-learning>","CreationDate":"2017-01-09T12:08:07.727","AcceptedAnswerId":null,"OwnerUserId":4776477.0,"Title":"Tensorflow: How to use Adam optimizer properly","Body":"<p>Somebody have already asked a <a href=\"https:\/\/stackoverflow.com\/questions\/33788989\/tensorflow-using-adam-optimizer\">similar question<\/a>, but the solution, which is given there, does not work for me.<\/p>\n\n<p>I am trying to use Adam optimizer in tensorflow. Here is a part of my code about it :<\/p>\n\n<pre><\/pre>\n\n<p>The following error is raised after train_op is used for the first time:<\/p>\n\n<blockquote>\n  <p>FailedPreconditionError (see above for traceback): Attempting to use uninitialized value pretrain_1\/beta2_power\n       [[Node: pretrain_1\/beta2_power\/read = IdentityT=DT_FLOAT, _class=[\"loc:@autoencoder_variables\/weights1\"], _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n<\/blockquote>\n\n<p>If I try to add a line<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the following error:<\/p>\n\n<blockquote>\n  <p>NameError: global name 'beta2_power' is not defined<\/p>\n<\/blockquote>\n\n<p>If I follow an advice for the <a href=\"https:\/\/stackoverflow.com\/questions\/33788989\/tensorflow-using-adam-optimizer\">similar question<\/a> and replace sess.run(tf.variables_initializer(vars_to_init)) by sess.run(tf.initialize_all_variables()),\n  I am getting the following error after running this line:<\/p>\n\n<blockquote>\n  <p>FailedPreconditionError: Attempting to use uninitialized value autoencoder_variables\/biases1\n       [[Node: autoencoder_variables\/biases1\/read = IdentityT=DT_FLOAT, _class=[\"loc:@autoencoder_variables\/biases1\"], _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n<\/blockquote>\n\n<p>I had no problems when I was using Gradient Descent optimizer...<\/p>\n\n<p>What am I doing wrong? What is the proper way to use this optimizer?<\/p>\n\n<p>EDIT\nMore details about the class to clarify autoencoder_variables:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41633228","CreationDate":"2017-01-13T11:08:38.653","ParentId":null,"OwnerUserId":"4776477","Title":null,"Body":"<p>The problem was that I was using one variable values in order to initialize other variable (it raised an error of using uninitialized variables during initialization).<\/p>\n\n<p>Instead of using another variable during initialization<\/p>\n\n<pre><code>self[name_b + \"_fixed\"] = tf.Variable(tf.identity(self[name_b]),\n                                            name=name_b + \"_fixed\",\n                                            trainable=False)\n<\/code><\/pre>\n\n<p>I initialize it randomly<\/p>\n\n<pre><code>self[name_b + \"_fixed\"] = tf.Variable(init_b,\n                                            name=name_b + \"_fixed\",\n                                            trainable=False)\n<\/code><\/pre>\n\n<p>And assign it to another variable once it was trained:<\/p>\n\n<pre><code> ae[name_w + \"_fixed\"] = tf.identity(ae[name_w])\n<\/code><\/pre>\n"}]}
{"QuestionId":41548409,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2017-01-09T12:46:15.603","AcceptedAnswerId":41563735.0,"OwnerUserId":5341238.0,"Title":"NameError: global name 'linear' is not defined","Body":"<p>I am trying to run an implementation of attention mechanism by Google DeepMind. However it is based on an older version of TensorFlow and I am getting this error<\/p>\n\n<blockquote>\n  <p>from tensorflow.models.rnn.rnn_cell import RNNCell, linear<br>\n  concat = linear([inputs, h, self.c], 4 * self._num_units, True)<\/p>\n  \n  <p><strong>NameError: global name 'linear' is not defined<\/strong><\/p>\n<\/blockquote>\n\n<p>I couldn't find the linear model\/function in the new tensorflow documentation. Can anyone help me with this? Thanks!<\/p>\n","answers":[{"AnswerId":"41563735","CreationDate":"2017-01-10T07:52:03.997","ParentId":null,"OwnerUserId":"5080995","Title":null,"Body":"<p>You need to use the <code>tf.nn.rnn_cell._linear<\/code> function to make the code work. Have a look at this <a href=\"http:\/\/r2rt.com\/recurrent-neural-networks-in-tensorflow-ii.html\" rel=\"nofollow noreferrer\">tutorial<\/a> for a sample usage.<\/p>\n"}]}
{"QuestionId":41549784,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-09T14:06:42.340","AcceptedAnswerId":44239474.0,"OwnerUserId":7287271.0,"Title":"Can TensorFlow RNNCell have more dimensions than 3D?","Body":"<p>I'd like to use  with 5D tensors of video frames (samples, time, height, width, channels) without having to flatten and reshape on each . Is this possible?<\/p>\n","answers":[{"AnswerId":"44239474","CreationDate":"2017-05-29T09:51:49.553","ParentId":null,"OwnerUserId":"7287271","Title":null,"Body":"<p>As of TensorFlow 1.2 this will be supported. It's also possible already by setting <code>time_major=True<\/code> and reordering the dims accordingly outside of the <code>tf.nn.rnn<\/code> call.<\/p>\n"}]}
{"QuestionId":41549908,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-09T14:13:27.420","AcceptedAnswerId":41550876.0,"OwnerUserId":6667826.0,"Title":"loading text file and accessing data using tensorflow","Body":"<p>I have a text file of size (20480,8). I want the data in 4th column in to one array. I am able to do it using python as <\/p>\n\n<pre><\/pre>\n\n<p>I get the size of the y as (20480,)<\/p>\n\n<p>but I am trying to copy the same as a tensor. How to access the data <\/p>\n\n<pre><\/pre>\n\n<p>I cant understand whether I have loaded the file correctly as I am getting it as empty array\n    <\/p>\n","answers":[{"AnswerId":"41550876","CreationDate":"2017-01-09T15:04:05.843","ParentId":null,"OwnerUserId":"6667826","Title":null,"Body":"<p>found the solution posting if it helps someone<\/p>\n\n<pre><code>x1 = tf.constant(y,name = 'x1')\nmodel = tf.initialize_all_variables()\nwith tf.Session() as session:\n     session.run(model)\nprint(np.shape(session.run(x1)))\n<\/code><\/pre>\n"}]}
{"QuestionId":41550913,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2017-01-09T15:05:36.177","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Deploy error axis 2 out of range for alexnet","Body":"<p>I am using the Alexnet and try to deploy my network. But when I do so I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have no clue why. It has always been working for me and now this error occurs.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>Above is my .prototxt (Should be the same as Alexnet)<\/p>\n\n<pre><\/pre>\n\n<p><strong>Example 2:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41550966,"AnswerCount":3,"Tags":"<tensorflow><neural-network><regression><deep-learning><gradient-descent>","CreationDate":"2017-01-09T15:08:27.957","AcceptedAnswerId":41565985.0,"OwnerUserId":1876627.0,"Title":"Why deep NN can't approximate simple ln(x) function?","Body":"<p>I have created ANN with two RELU hidden layers + linear activation layer and trying to approximate simple ln(x) function. And I am can't do this good. I am confused because lx(x) in x:[0.0-1.0] range should be approximated without problems (I am using learning rate 0.01 and basic grad descent optimization).<\/p>\n\n<pre><\/pre>\n\n<p>For the configuration above NN is just learning to guess y = -1.00. I have tried different learning rates,  couple optimizers and different configurations with no success - learning does not converge in any case. I did something like that with logarithm in past in other deep learning framework without problem.. Can be the TF specific issue? What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"41560916","CreationDate":"2017-01-10T04:00:58.093","ParentId":null,"OwnerUserId":"7397431","Title":null,"Body":"<p>I have not look at the code but this is the theory. If you use an activation function like \"tanh\", then for small weights the activation function is in the linear region and for large weights the activation function is either -1 or +1. If you are in the linear region across all layers then you can not approximate complex functions (i.e. you have a sandwich of linear layers hence the best you can do is linear aproximations)  but if you have bigger weights then the nonlinearly allow you to approximate a wide range of functions. There are no free lunches, the weights need to be at the right values to avoid over-fitting and under-fitting. This process is called regularization. <\/p>\n"},{"AnswerId":"41564050","CreationDate":"2017-01-10T08:12:16.480","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>First of all, your input data is in range [0, 1), which is not a good input to a neural network. Subtract mean from <code>x<\/code> after computing <code>y<\/code> to make it normalized (also ideally divide by standard deviation).<\/p>\n\n<p>However, in your particular case it was not enough to make it work.<\/p>\n\n<p>I played with it and found two ways to make it work (both require data normalization as described above):<\/p>\n\n<ol>\n<li>Either completely remove the second layer<\/li>\n<\/ol>\n\n<p>or<\/p>\n\n<ol start=\"2\">\n<li>Make the number of neurons in the second layer 50.<\/li>\n<\/ol>\n\n<p>My guess would be that 10 neurons do not have sufficient representation power to pass enough information to the last layer (obviously, a perfectly smart NN would learn to ignore the second layer in this case passing the answer in one of the neurons, but the theoretical possibility doesn't mean that gradient descent will learn to do so). <\/p>\n"},{"AnswerId":"41565985","CreationDate":"2017-01-10T09:57:14.753","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<h2>What your network has to predict<\/h2>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zLgos.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/zLgos.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Source: <a href=\"http:\/\/www.wolframalpha.com\/input\/?i=plot+ln(x)+from+0+to+1\" rel=\"noreferrer\">WolframAlpha<\/a><\/p>\n\n<h2>What your architecture is<\/h2>\n\n<p>ReLU(ReLU(x * W_1 + b_1) * W_2 + b_2)*W_out + b_out<\/p>\n\n<h2>Thoughts<\/h2>\n\n<p>My first thought was that ReLU is the problem. However, you don't apply relu to the output, so that should not cause the problem.<\/p>\n\n<p>Changing the initialization (from uniform to normal) and the Optimizer (from SGD to ADAM) seems to fix the problem:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\nimport tensorflow as tf\nimport numpy as np\n\n\ndef get_target_result(x):\n    return np.log(x)\n\n\ndef multilayer_perceptron(x, weights, biases):\n    \"\"\"Create model.\"\"\"\n    # Hidden layer with RELU activation\n    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])\n    layer_1 = tf.nn.relu(layer_1)\n    # # Hidden layer with RELU activation\n    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])\n    layer_2 = tf.nn.relu(layer_2)\n\n    # Output layer with linear activation\n    out_layer = tf.matmul(layer_2, weights['out']) + biases['out']\n    return out_layer\n\n# Parameters\nlearning_rate = 0.01\ntraining_epochs = 10**6\nbatch_size = 500\ndisplay_step = 500\n\n# Network Parameters\nn_hidden_1 = 50  # 1st layer number of features\nn_hidden_2 = 10  # 2nd layer number of features\nn_input = 1\n\n\n# Store layers weight &amp; bias\nweights = {\n    'h1': tf.Variable(tf.truncated_normal([n_input, n_hidden_1], stddev=0.1)),\n    'h2': tf.Variable(tf.truncated_normal([n_hidden_1, n_hidden_2], stddev=0.1)),\n    'out': tf.Variable(tf.truncated_normal([n_hidden_2, 1], stddev=0.1))\n}\n\nbiases = {\n    'b1': tf.Variable(tf.constant(0.1, shape=[n_hidden_1])),\n    'b2': tf.Variable(tf.constant(0.1, shape=[n_hidden_2])),\n    'out': tf.Variable(tf.constant(0.1, shape=[1]))\n}\n\nx_data = tf.placeholder(tf.float32, [None, 1])\ny_data = tf.placeholder(tf.float32, [None, 1])\n\n# Construct model\npred = multilayer_perceptron(x_data, weights, biases)\n\n# Minimize the mean squared errors.\nloss = tf.reduce_mean(tf.square(pred - y_data))\noptimizer = tf.train.GradientDescentOptimizer(learning_rate)\n# train = optimizer.minimize(loss)\ntrain = tf.train.AdamOptimizer(1e-4).minimize(loss)\n\n# Before starting, initialize the variables.  We will 'run' this first.\ninit = tf.initialize_all_variables()\n\n# Launch the graph.\nsess = tf.Session()\nsess.run(init)\n\nfor step in range(training_epochs):\n    x_in = np.random.rand(batch_size, 1).astype(np.float32)\n    y_in = get_target_result(x_in)\n    sess.run(train, feed_dict={x_data: x_in, y_data: y_in})\n    if(step % display_step == 0):\n        curX = np.random.rand(1, 1).astype(np.float32)\n        curY = get_target_result(curX)\n\n        curPrediction = sess.run(pred, feed_dict={x_data: curX})\n        curLoss = sess.run(loss, feed_dict={x_data: curX, y_data: curY})\n        print((\"For x = {0} and target y = {1} prediction was y = {2} and \"\n               \"squared loss was = {3}\").format(curX, curY,\n                                                curPrediction, curLoss))\n<\/code><\/pre>\n\n<p>Training this for 1 minute gave me:<\/p>\n\n<pre><code>For x = [[ 0.19118255]] and target y = [[-1.65452647]] prediction was y = [[-1.65021849]] and squared loss was = 1.85587377928e-05\nFor x = [[ 0.17362741]] and target y = [[-1.75084364]] prediction was y = [[-1.74087048]] and squared loss was = 9.94640868157e-05\nFor x = [[ 0.60853624]] and target y = [[-0.4966988]] prediction was y = [[-0.49964082]] and squared loss was = 8.65551464813e-06\nFor x = [[ 0.33864763]] and target y = [[-1.08279514]] prediction was y = [[-1.08586168]] and squared loss was = 9.4036658993e-06\nFor x = [[ 0.79126364]] and target y = [[-0.23412406]] prediction was y = [[-0.24541236]] and squared loss was = 0.000127425722894\nFor x = [[ 0.09994856]] and target y = [[-2.30309963]] prediction was y = [[-2.29796076]] and squared loss was = 2.6408026315e-05\nFor x = [[ 0.31053194]] and target y = [[-1.16946852]] prediction was y = [[-1.17038012]] and squared loss was = 8.31002580526e-07\nFor x = [[ 0.0512077]] and target y = [[-2.97186542]] prediction was y = [[-2.96796203]] and squared loss was = 1.52364455062e-05\nFor x = [[ 0.120253]] and target y = [[-2.11815739]] prediction was y = [[-2.12729549]] and squared loss was = 8.35050013848e-05\n<\/code><\/pre>\n\n<p>So the answer <em>might<\/em> be that your optimizer is not good \/ the optimization problem starts at a bad point. See<\/p>\n\n<ul>\n<li>Xavier Glorot, Yoshua Bengio: <a href=\"http:\/\/jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf\" rel=\"noreferrer\">Understanding the difficulty of training deep feedforward neural networks<\/a><\/li>\n<li><a href=\"https:\/\/imgur.com\/a\/Hqolp\" rel=\"noreferrer\">Visualizing Optimization Algos<\/a><\/li>\n<\/ul>\n\n<p>The following image is from Alec Radfords nice gifs. It does not contain ADAM, but you get a feeling for how much better one can do than SGD:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fSiiG.gif\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fSiiG.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Two idea how this might be improved<\/p>\n\n<ul>\n<li>try dropout<\/li>\n<li>try not to use x values close to 0. I would rather sample values in [0.01, 1].<\/li>\n<\/ul>\n\n<p>However, my experience with regression problems is quite limited.<\/p>\n"}]}
{"QuestionId":41551044,"AnswerCount":0,"Tags":"<python-2.7><anaconda><deep-learning><keras>","CreationDate":"2017-01-09T15:13:07.093","AcceptedAnswerId":null,"OwnerUserId":6601038.0,"Title":"Prediction on test data using the bottleneck features of a pre-trained network in Keras","Body":"<p>I have followed the blog <a href=\"https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html\" rel=\"nofollow noreferrer\">building-powerful-image-classification-models-using-very-little-data<\/a>.<br>\nIn section : Using the bottleneck features of a pre-trained network, VGG16 weights have been incorporated using bottle neck features.\nI have got the fitted model and accuracy of training and validation set.<\/p>\n\n<p>Now, I want to predict the test data. My test data is in the test folder which contains 100 images of cats and dogs.<br>\nHow can I get Predictions of test data from model ?<\/p>\n\n<p>I have tried the following :<\/p>\n\n<pre><\/pre>\n\n<p>But, getting errors. <\/p>\n","answers":[]}
{"QuestionId":41551910,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-09T15:59:41.237","AcceptedAnswerId":null,"OwnerUserId":7387749.0,"Title":"How to create an input function , input_fn()","Body":"<p>I am following <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/input_fn\/\" rel=\"nofollow noreferrer\">this<\/a> tutorial from _tensorflow.org .\nI am trying to correctly handle the <em>input_fn_<\/em>, to use as argument in <em>.fit()<\/em> .\nI have created the classifier:<\/p>\n\n<pre><\/pre>\n\n<p>Then the input function:<\/p>\n\n<pre><\/pre>\n\n<p>Finally I have put <em>input_fn()<\/em> in <em>fit()<\/em> by:<\/p>\n\n<pre><\/pre>\n\n<p>When I run the code, I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know if it is about the <em>input_fn<\/em> definition, or the <em>fit<\/em> arguments<\/p>\n","answers":[{"AnswerId":"42410697","CreationDate":"2017-02-23T08:38:50.700","ParentId":null,"OwnerUserId":"4581256","Title":null,"Body":"<p>Don't use SKCompat if you want to use input_fn, replace first line with:<\/p>\n\n<pre><code>classifier = tf.contrib.learn.DNNClassifier(\n<\/code><\/pre>\n\n<p>And adjust parenthesis as needed.<\/p>\n"}]}
{"QuestionId":41554050,"AnswerCount":0,"Tags":"<tensorflow><bazel>","CreationDate":"2017-01-09T17:59:20.297","AcceptedAnswerId":null,"OwnerUserId":1576602.0,"Title":"Compiling a custom TensorFlow operation outside TF source tree","Body":"<p>I have a custom TensorFlow operation which currently sits inside  and uses the  bazel function in . I would like to separate it from the TF source tree and ship in an independent repository. How could I compile such operation?<\/p>\n","answers":[]}
{"QuestionId":41554540,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-09T18:28:12.743","AcceptedAnswerId":null,"OwnerUserId":1501285.0,"Title":"How to set image shape in Python for Tensorflow prediction?","Body":"<p>I'm dealing with the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The placeholder is set to: <\/p>\n\n<p>And the image (when running ) has the output: <\/p>\n\n<p>How can I update the image to be aligned when running: <\/p>\n","answers":[{"AnswerId":"41557337","CreationDate":"2017-01-09T21:32:56.410","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The placeholder <code>x<\/code> in your program represents a <strong>batch<\/strong> of 32x32 (presumably) RGB images, for which predictions will be computed in a single step. If you want to compute a prediction on a single image&mdash;i.e. an array of shape <code>(32, 32, 3)<\/code>&mdash;you must reshape it to have an additional leading dimension. There are many ways to do this, but <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#numpy.newaxis\" rel=\"nofollow noreferrer\"><code>np.newaxis<\/code><\/a> is one nice way to do it:<\/p>\n\n<pre><code> img1 = ...                             # Array of shape (32, 32, 3)\n img1_as_batch = img1[np.newaxis, ...]  # Array of shape (1, 32, 32, 3)\n\n print(sess.run(correct_prediction, feed_dict={x: img1_as_batch}))\n<\/code><\/pre>\n"}]}
{"QuestionId":41555039,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><tflearn>","CreationDate":"2017-01-09T18:59:12.527","AcceptedAnswerId":null,"OwnerUserId":1095682.0,"Title":"Predicting values using TFLearn neural networks","Body":"<p>I am new to TFLearn and I am trying out a simple neural network to predict the output array value when an input array is given.<\/p>\n\n<p>Actual input for this code would be either pixel values of a grayscale image or features extracted from a grayscale image. Hence the input is in a 2d array format. The output would be the predicted color for each pixel.<\/p>\n\n<p>In the example code I have used two random arrays of size 9. I need to train the network to predict the 't_y' array when 't_x' array is given as input.\nThe code runs, but the prediction is very poor.<\/p>\n\n<p>The code has been adapted from MNIST example of TFLearn found <a href=\"https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/examples\/images\/convnet_mnist.py\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>This is my code<\/p>\n\n<pre><\/pre>\n\n<p>I am assuming it has something to do with the parameter values specified in the 'conv_2d' and 'fully_connnected' functions.<\/p>\n\n<p>What values would I have to set to get an accurate prediction ?<\/p>\n","answers":[{"AnswerId":"41555251","CreationDate":"2017-01-09T19:14:30.913","ParentId":null,"OwnerUserId":"366553","Title":null,"Body":"<h2>Format of output<\/h2>\n\n<p>The last layer of your code (fully_connected(network, 9, activation='softmax')) results in 9 neurons with a softmax function, i.e. normalised so that their total sum will add up to 1. This is generally usable (and used in MNIST) for selecting\/optimizing a function that selects one of 9 possible output values - the network will output something like [0.01 0.01 0.01 0.9 0.03 0.01 0.01 0.01 0.01], \"predicting\" that the fourth value is the correct one, and this would be matched against a one-hot target vector (e.g. [0 0 0 1 0 0 0 0 0]).<\/p>\n\n<p>Needless to say, a softmax output cannot ever be equal to [9, 5, 1, 4, 7, 9, 7, 3, 6], and not even close to that, since the output of all values softmax will add up to 1. Even the layer before that cannot output such values since tanh can only produce values between -1 and 1, and can't ever result in 9.<\/p>\n\n<p>If you want to predict 9 numbers in the range 1-9, then you might want to use a fully connected layer instead of softmax, and scale your output so that the expected output is in the range of 0 to 1. There's more to that, but this would be a good start.<\/p>\n"}]}
{"QuestionId":41555348,"AnswerCount":0,"Tags":"<deep-learning><caffe>","CreationDate":"2017-01-09T19:20:31.357","AcceptedAnswerId":null,"OwnerUserId":5108943.0,"Title":"Training Caffe Alexnet from Snapshot not Qualitatively the same as Continous Training","Body":"<p>I am seeking help concerning training Caffe AlexNet. A summary of the problem is that reported accuracy of continuous training of Caffe network (AlexNet on a randomly selected 200,000 ILSVRC 2012 images) differs from that of accuracy reported from resuming training from a snapshot of solver state.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/K2qTi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/K2qTi.png\" alt=\"attached picture\"><\/a> Picture above shows that reported accuracies differ based on number of iterations before snapshot and do not match training without snapshot. X-axis on the plot is the number of iterations (or number of images processed if you multiply each point by mini batch size of 100), and Y-axis, the test accuracy. The expectation is that all lines should match up irrespective of the frequency of snapshot.<\/p>\n\n<p>Please note that all experiments are setup the same way with same training\/validation data and parameters (i.e., same seed, same mini batch size, fixed learning rate policy, zero momentum etc).<\/p>\n\n<p>Thanks for your help.<\/p>\n","answers":[]}
{"QuestionId":41555590,"AnswerCount":0,"Tags":"<python><caffe><lmdb>","CreationDate":"2017-01-09T19:36:57.470","AcceptedAnswerId":null,"OwnerUserId":2832033.0,"Title":"Insert images as encoded file to LMDB using Python","Body":"<p>I'm trying to insert a bunch of images into the LMDB format using the following snippet:<\/p>\n\n<pre><\/pre>\n\n<p>However, as the uncompressed bytes are written on the disk, the resultant file is so huge!! Accordingly, I'm wondering how I can set the encoding property to JPG in python. I am already aware that the very option is available in C++ api.<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[]}
{"QuestionId":41555612,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-09T19:38:00.237","AcceptedAnswerId":41557248.0,"OwnerUserId":2454808.0,"Title":"Tensorflow imprecise timeouts","Body":"<p>I've been testing out the the timeout functionality for sess.runs (applied to a convolutional neural network), and it seems like the timeouts aren't very precise. <\/p>\n\n<p>For example, if I set the timeout to be 800 ms, there might be a 1-2 second delay before the timeout exception is triggered. This sort of leads me to believe that cancellation notifications aren't caught between computational nodes. (Which according to the timeline are .2-.5 s each)<\/p>\n\n<p>So<\/p>\n\n<p>1) Is there a way to make the timeouts more precise?<\/p>\n\n<p>2) Are Tensorflow cancellation notifications caught between node computations?<\/p>\n","answers":[{"AnswerId":"41557248","CreationDate":"2017-01-09T21:27:18.223","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The cancellation and timeout mechanism in TensorFlow was only designed to cancel a small number of <strong>blocking<\/strong> operations, in particular: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/queues#QueueBase.dequeue\" rel=\"nofollow noreferrer\">dequeuing<\/a> from an empty queue, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/queues#QueueBase.enqueue\" rel=\"nofollow noreferrer\">enqueuing<\/a> to a full queue, and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/readers#ReaderBase.read\" rel=\"nofollow noreferrer\">reading<\/a> from a file.<\/p>\n\n<p>If you run a graph containing <strong>non-blocking<\/strong> operations, such as <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/matrix_math_functions#matmul\" rel=\"nofollow noreferrer\"><code>tf.matmul()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/convolution#conv2d\" rel=\"nofollow noreferrer\"><code>tf.nn.conv2d()<\/code><\/a>, and the timeout expires, TensorFlow will typically wait until these operations have completed before returning with a \"deadline exceeded\" error.<\/p>\n\n<p>Why is this the case? We added cancellation because users started to build pipelines of blocking operations into their graphs (e.g. for <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow noreferrer\">reading data<\/a>) and some form of cancellation was needed to shut down these pipelines cleanly. Timeouts also help to debug deadlocks that can unfortunately occur in these pipelines. By contrast, TensorFlow is designed to dispatch non-blocking operations as efficiently as possible: for example, when running on a GPU, TensorFlow will asynchronously enqueue multiple operations on the GPU compute stream without blocking on their completion. Although it would technically be possible to check for cancellation between the execution of each operation, this would add latency to operation dispatch, and reduce overall performance in the common case.<\/p>\n\n<p>However, if timeouts\/cancellation for non-blocking operations would be useful for your use case, please feel free to open a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">GitHub issue<\/a> as a feature request!<\/p>\n"}]}
{"QuestionId":41556179,"AnswerCount":1,"Tags":"<tensorflow><grpc><protocol-buffers><tensorflow-serving>","CreationDate":"2017-01-09T20:15:04.950","AcceptedAnswerId":null,"OwnerUserId":64106.0,"Title":"tensorflow serving impossible proto files structure for grpc interface","Body":"<p>I'm trying to compile pb files for GRPC calls to Tensorflow Serving (in php, but the question is not PHP related)<\/p>\n\n<p>The file <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/apis\/predict.proto\" rel=\"nofollow noreferrer\">serving\/tensorflow_serving\/apis\/predict.proto<\/a> has:<\/p>\n\n<pre><\/pre>\n\n<p>However in a normal setup tensorflow and tensorflow serving are not located in a hierarchy that has a common folder from which the two import can work together.<\/p>\n\n<p>Assuming that compiling the proto files to pb files for grpc keeps the hirarchy , it cannot work without locating tensorflow serving under \/tensorflow\/. What am I missing here?<\/p>\n\n<p>What is the best practice in order to compile pb files for grpc client?<\/p>\n\n<p>Another issue: if the pb files are created - they include the imports with same hirarchy so it will force the same folder structure on the client ??? this is against the meaning of GRPC which is isolation and seperation between the entities. <\/p>\n","answers":[{"AnswerId":"41557020","CreationDate":"2017-01-09T21:12:11.030","ParentId":null,"OwnerUserId":"886653","Title":null,"Body":"<p>I don't know anything about tensorflow, but i'm approaching the problem from a <em>just-another-protobuf-creation<\/em> point of view. Here <a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving<\/a> i'm seeing both <code>tensorflow_serving<\/code> and a <em>submodule<\/em> <code>tensorflow<\/code> which is a root of your desired dependency (i.e. it has another <code>tensorflow<\/code> subfolder in it). So i guess that you either missed some configuration step, which would have copied the folder into right relative location, or you are running an incomplete\/incorrect <code>protoc<\/code> command line, i.e. you are missing some <code>-I &lt;path&gt;<\/code><\/p>\n"}]}
{"QuestionId":41556222,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-01-09T20:18:26.247","AcceptedAnswerId":41556815.0,"OwnerUserId":2836240.0,"Title":"Time series prediction with Keras - values close to average","Body":"<p>I am doing a one step ahead prediction using 15 previous samples on a dataset using LSTMs in Keras. <\/p>\n\n<p>The data csv file can be found here: \n(<a href=\"https:\/\/drive.google.com\/file\/d\/0Byiipc0dArG0LVZJelB4NFBucms\/view?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/0Byiipc0dArG0LVZJelB4NFBucms\/view?usp=sharing<\/a>)<\/p>\n\n<p>The second column col[1] values are used. The values in the first column (timestamps) are not used at all.<\/p>\n\n<p>I use the following code:<\/p>\n\n<pre><\/pre>\n\n<p>The problem:<\/p>\n\n<p>I have used this code to predict few more periodical and clean time series and it works well. For this set of data however, I used different parameters of Adam (learning rate, etc). Still, I get the prediction with a large offset from the actual values. It seems like the predicted values are always close to the average of the data values. Please see the following graph. I have 1850 data points in the csv file. These are grouped into sequences of size 15. The input is a sequence of size 15. The output should be the next value predicted. 70% of the data is used for training and the rest is the test data set. Prediction is done on train and test datasets shown in green and red respectively in the following image. <\/p>\n\n<p>(<a href=\"https:\/\/drive.google.com\/file\/d\/0Byiipc0dArG0OEN5el9lc0puNGM\/view?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/0Byiipc0dArG0OEN5el9lc0puNGM\/view?usp=sharing<\/a>)<\/p>\n\n<p>Do you have any idea why this is happening and what can be causing it?<\/p>\n\n<p>Thanks,\nFaezeh<\/p>\n","answers":[{"AnswerId":"41556815","CreationDate":"2017-01-09T20:58:36.373","ParentId":null,"OwnerUserId":"3038791","Title":null,"Body":"<p>Your data looks a lot like binary data + noise. \nBelow is a histogram of the raw data and a histogram of the first difference of the raw data. If the exact value of the prediction is not important to you, I would suggest making the data binary and using a different cost function, e.g. binary crossentropy, though I'm skeptical that it will work giving the explanation below.<\/p>\n\n<p>If there are similar sequences with multiple possible next values, the network will try to predict the average value. For example, consider the sequences (0,0,1) and (0,0,0) and (0,0,-1) and a network trying to learn their last values. With this toy data, any prediction model's best MSE is to predict the average of the last values, 0 in this case...<\/p>\n\n<p>I suggest checking what the network is learning by plotting target values, Y, and predicted target values, Y_hat.\nHope this helps!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xHtJDm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xHtJDm.png\" alt=\"Raw Data Histogram\"><\/a> <a href=\"https:\/\/i.stack.imgur.com\/IwXtbm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IwXtbm.png\" alt=\"1st Difference Histogram\"><\/a><\/p>\n"}]}
{"QuestionId":41556387,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-09T20:29:02.850","AcceptedAnswerId":null,"OwnerUserId":6890895.0,"Title":"how to concat each row's max divided by two part in tensorflow?","Body":"<p>I am a new to tensorflow.<\/p>\n\n<p>Acutually, My question is how to concat each row's max divided by two part in tensorflow?<\/p>\n\n<p>Here's an example of what I want to do in numpy code.<\/p>\n\n<pre><\/pre>\n\n<p>I am tring to use tf.py_func, but I don't know idea of gradient implementation.<\/p>\n","answers":[{"AnswerId":"41557239","CreationDate":"2017-01-09T21:26:44.733","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You can use <code>map_fn<\/code> to apply your max\/min functions to each row<\/p>\n\n<pre><code>def process_row(args):\n    \"\"\"Returns max value before sep_id and max value after sep_id in row\"\"\"\n    row = args[0]\n    sep_id = args[1]\n    return tf.stack([tf.reduce_max(row[:sep_id+1]), tf.reduce_max(row[sep_id:])])\n\narray = tf.reshape(tf.range(20), (4, 5))\nsepids = tf.constant([2,1,0,3], dtype=tf.int32)\nsess = tf.InteractiveSession()\nsess.run(tf.map_fn(process_row, (array, sepids), dtype=(tf.int32)))    \n<\/code><\/pre>\n\n<p>This should print<\/p>\n\n<pre><code>array([[ 2,  4],\n       [ 6,  9],\n       [10, 14],\n       [18, 19]], dtype=int32)\n<\/code><\/pre>\n"}]}
{"QuestionId":41557196,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-09T21:24:08.573","AcceptedAnswerId":null,"OwnerUserId":5296362.0,"Title":"Tensorflow bytenet tensor shape mismatch error","Body":"<p>I am trying to use the French-to-English translation implementation given in this link.<\/p>\n\n<p><a href=\"https:\/\/github.com\/buriburisuri\/ByteNet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/buriburisuri\/ByteNet<\/a><\/p>\n\n<p>When I run the  file therein, I get the following error. <\/p>\n\n<pre><\/pre>\n\n<p>The gist of it seems to be the shape mismatch error. I tried the  option in , but to no avail. Can anybody help me figure this out? if I could fix this error somehow, I would be done. Thanks a lot in advance!<\/p>\n","answers":[]}
{"QuestionId":41557405,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><nlp>","CreationDate":"2017-01-09T21:37:09.247","AcceptedAnswerId":null,"OwnerUserId":6186560.0,"Title":"NLP Recurrent Neural Network always gives constant values","Body":"<p>I've written a simple recurrent network in TensorFlow based on this video that I watched: <a href=\"https:\/\/youtu.be\/vq2nnJ4g6N0?t=8546\" rel=\"nofollow noreferrer\">https:\/\/youtu.be\/vq2nnJ4g6N0?t=8546<\/a><\/p>\n\n<p>In the video the RNN is demonstrated to produce Shakespeare plays by having the network produce words one character at a time. The output of the network is fed back into the input on the next iteration.<\/p>\n\n<p>Here's a diagram of my network:<\/p>\n\n<pre><\/pre>\n\n<p>I expect the network to at least do the copying bit correctly. Unfortunately my network always outputs 32 for all values (ASCII space character). I'm not sure what is causing the issue...<\/p>\n\n<p>Please help me get my network producing poetry!<\/p>\n\n<p>My code is here:\n<a href=\"https:\/\/github.com\/calebh\/namepoet\/blob\/03f112ced94c3319055fbcc74a2acdb4a9b0d41c\/main.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/calebh\/namepoet\/blob\/03f112ced94c3319055fbcc74a2acdb4a9b0d41c\/main.py<\/a><\/p>\n\n<p>The corpus can be replaced by a few paragraphs of Lorem Ipsum to speed up training (the network has the same bad behavior).<\/p>\n","answers":[{"AnswerId":"41784810","CreationDate":"2017-01-21T21:38:24.170","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>I'd try to run the code for longer time? You have batch_size = 10, sequence_size = 30 and 20 iterations, your network essentially has seen 6000 characters in total, maybe with a learning rate of 0.001, it wasn't enough to move away from your initialization.<\/p>\n\n<p>Hence, I'd try to raise the learning rate to a very high value (e.g. 1 or 100) and see if it starts outputting different letters to confirm your implementation is somewhat correct. A network trained with such a high learning rate is usually not going to be accurate at all.<\/p>\n"},{"AnswerId":"41775878","CreationDate":"2017-01-21T04:52:48.153","ParentId":null,"OwnerUserId":"7449151","Title":null,"Body":"<p>Sounds like it might be saturating your filters (ergo the activation function is way at the far end of the spectrum and thus has a very low gradient and gets stuck). You might want to try initializing your neurons parameters with a different method. <\/p>\n\n<p>Also, is there a particular reason you're using GRU? In my experience LSTM units are more reliable, if a bit less efficient.<\/p>\n"}]}
{"QuestionId":41558455,"AnswerCount":1,"Tags":"<python><caffe>","CreationDate":"2017-01-09T23:03:27.677","AcceptedAnswerId":null,"OwnerUserId":7396740.0,"Title":"Python cannot find modules while being called by a caffe .prototxt file as an input layer","Body":"<p>Hi,<\/p>\n\n<p>I have my input data sequence and label constructed by a sequence_layer.py python file. Then input data layer is:<\/p>\n\n<pre><\/pre>\n\n<p>But, once I run the train.sh through terminal, python cannot find modules, error is the following:<\/p>\n\n<pre><\/pre>\n\n<p>But if I type python in the terminal and import the modules, it worked pretty fine:<\/p>\n\n<pre><\/pre>\n\n<p>And I checked the PYTHONPATH while the sequence_layer.py is called by .prototxt using \"print(sys.path)\". The paths are correct. <\/p>\n\n<p>I am using Ubuntu 14.04LTS, and caffe is compiled with \/usr\/include\/python2.7 and I did set WITH_PYTHON_LAYER := 1. <\/p>\n\n<p>Anyone has a solution? Thank you so much for the help. <\/p>\n\n<p>Best,\nXin<\/p>\n","answers":[{"AnswerId":"41661309","CreationDate":"2017-01-15T12:54:02.133","ParentId":null,"OwnerUserId":"7396740","Title":null,"Body":"<p>Problem is solved by using Anaconda2 instead of system default Python2.7. Thanks.<\/p>\n"}]}
{"QuestionId":41558702,"AnswerCount":1,"Tags":"<memory><torch>","CreationDate":"2017-01-09T23:27:22.617","AcceptedAnswerId":null,"OwnerUserId":3207716.0,"Title":"efficiently load and access torch tensor","Body":"<p>I have a big binary file (almost 2GB) containing float32. I load it by<\/p>\n\n<p><\/p>\n\n<p>I will keep accessing this big tensor for 1 to 2 hours when executing my program. I observed that it's very slow for the first 10 to 20 minutes.<\/p>\n\n<p>Can anyone explain why and provide some advice?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"46915906","CreationDate":"2017-10-24T16:45:59.263","ParentId":null,"OwnerUserId":"6869035","Title":null,"Body":"<p>its may be because <\/p>\n\n<blockquote>\n  <p>OS caches things in memory, in the background, without you having to care about it...<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/groups.google.com\/d\/msg\/torch7\/E4SaoAEnit0\/pHMSiCdOlrIJ\" rel=\"nofollow noreferrer\">Preferred Handling of Large Files<\/a><\/p>\n"}]}
{"QuestionId":41559190,"AnswerCount":2,"Tags":"<python><r><tensorflow>","CreationDate":"2017-01-10T00:15:09.920","AcceptedAnswerId":null,"OwnerUserId":3297472.0,"Title":"How can I access weights from a layer created with tf.contrib.layers.fully_connected?","Body":"<p>I'm using  to create a layer in the following code. <\/p>\n\n<pre><\/pre>\n\n<p>How can I access the weights as I would in the following block of code with ?<\/p>\n\n<pre><\/pre>\n\n<p>Note: I'm using TensorFlow for R but this should be the same as TensorFlow for Python by changing  for .<\/p>\n","answers":[{"AnswerId":"42257432","CreationDate":"2017-02-15T18:32:54.137","ParentId":null,"OwnerUserId":"3945436","Title":null,"Body":"<p>You can get the list of all global variables using <code>tf$global_variables()<\/code>. Not an ideal solution (because it retrieves a list of unnamed variables), but it should get you what you need. Reproducible example below<\/p>\n\n<pre><code>library(tensorflow)\n\ndatasets &lt;- tf$contrib$learn$datasets\nmnist &lt;- datasets$mnist$read_data_sets(\"MNIST-data\", one_hot = TRUE)\n\nx &lt;- tf$placeholder(tf$float32, shape(NULL, 784L))\nlogits &lt;- tf$contrib$layers$fully_connected(x, 10L)\n\ny &lt;- tf$nn$softmax(logits)\ny_ &lt;- tf$placeholder(tf$float32, shape(NULL,10L))\n\ncross_entropy &lt;- tf$reduce_mean(-tf$reduce_sum(y_ * tf$log(y), reduction_indices=1L))\ntrain_step &lt;- tf$train$GradientDescentOptimizer(0.5)$minimize(cross_entropy)\n\nsess &lt;- tf$Session()\nsess$run(tf$global_variables_initializer())\n\nfor (i in 1:1000) {\n  batches &lt;- mnist$train$next_batch(100L)\n  batch_xs &lt;- batches[[1]]\n  batch_ys &lt;- batches[[2]]\n  sess$run(train_step,\n           feed_dict = dict(x = batch_xs, y_ = batch_ys))\n}\n\nlst.variables &lt;- sess$run(tf$global_variables())\nstr(lst.variables)\n<\/code><\/pre>\n"},{"AnswerId":"41559474","CreationDate":"2017-01-10T00:47:54.020","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You pass the name of the tensor to the <code>run<\/code> function. You should inspect the graph to see the name of the tensor added to the graph from the function. <\/p>\n"}]}
{"QuestionId":41559348,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-10T00:32:26.927","AcceptedAnswerId":41559455.0,"OwnerUserId":3297472.0,"Title":"What's the difference between those two commands in TensorFlow","Body":"<p>Suppose you have:<\/p>\n\n<pre><\/pre>\n\n<h3>Statement 1:<\/h3>\n\n<pre><\/pre>\n\n<h3>Statement 2:<\/h3>\n\n<pre><\/pre>\n\n<p>Does output in both statements return the same result?<\/p>\n","answers":[{"AnswerId":"41559455","CreationDate":"2017-01-10T00:45:23.897","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>No, they are not equivalent. Statement 1 also adds <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/27711108b5fce2e1692f9440631a183b3808fa01\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L1329\" rel=\"nofollow noreferrer\">an activation function<\/a>.<\/p>\n\n<p>If you set <code>activation=None<\/code>, then they are equal.<\/p>\n"}]}
{"QuestionId":41559425,"AnswerCount":1,"Tags":"<machine-learning><neural-network><keras><autoencoder>","CreationDate":"2017-01-10T00:41:34.593","AcceptedAnswerId":41592486.0,"OwnerUserId":3082832.0,"Title":"Autoencoder not learning identity function","Body":"<p>I'm somewhat new to machine learning in general, and I wanted to make a simple experiment to get more familiar with neural network autoencoders: To make an extremely basic autoencoder that would learn the identity function.<\/p>\n\n<p>I'm using Keras to make life easier, so I did this first to make sure it works:<\/p>\n\n<pre><\/pre>\n\n<p>As expected, the loss is zero, both in train and validation data:<\/p>\n\n<pre><\/pre>\n\n<p>Then I tried to do the same but without initializing the weights to the identity function, expecting that after a while of training it would learn it. It didn't. I've let it run for 200 epochs various times in different configurations, playing with different optimizers, loss functions, and adding L1 and L2 activity regularizers. The results vary, but the best I've got is still really bad, looking nothing like the original data, just being kinda in the same numeric range.\nThe data is simply some numbers oscillating around 1.1. I don't know if an activation layer makes sense for this problem, should I be using one?<\/p>\n\n<p>If this \"neural network\" of one layer can't learn something as simple as the identity function, how can I expect it to learn anything more complex? What am I doing wrong?<\/p>\n\n<h3>EDIT<\/h3>\n\n<p>To have better context, here's a way to generate a dataset very similar to the one I'm using:<\/p>\n\n<pre><\/pre>\n\n<p>I'm suspecting that the variations between the values might be too small. The loss function ends up having decent values (around ), but it's not enough precision for the result to have a similar shape to the original data. Maybe I should scale\/normalize it somehow? Thanks for any advice!<\/p>\n\n<h3>UPDATE<\/h3>\n\n<p>In the end, as it was suggested, the issue was with the dataset having too small variations between the 84 values, so the resulting prediction was actually pretty good in absolute terms (loss function) but comparing it to the original data, the variations were far off. I solved it by normalizing the 84 values in each sample around the sample's mean and dividing by the sample's standard deviation. Then I used the original mean and standard deviation to denormalize the predictions at the other end. I guess this could be done in a few different ways, but I did it by adding this normalization\/denormalization into the model itself by using some Lambda layers that operated on the tensors. That way all the data processing was incorporated into the model, which made it nicer to work with. Let me know if you would like to see the actual code.<\/p>\n","answers":[{"AnswerId":"41592486","CreationDate":"2017-01-11T13:39:41.367","ParentId":null,"OwnerUserId":"2314651","Title":null,"Body":"<p>I believe the problem could be either the number of epoch or the way you inizialize X.\nI ran your code with an X of mine for 100 epochs and printed the argmax() and max values of the weights, it gets really close to the identity function.<\/p>\n\n<p>I'm adding the code snippet that I used<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense\nimport numpy as np\nimport random\nimport pandas as pd\n\nX = np.array([[random.random() for r in xrange(84)] for i in xrange(1,100000)])\nmodel = Sequential([Dense(84, input_dim=84)], name=\"layer1\")\nmodel.compile(optimizer='sgd', loss='mean_squared_error')\nmodel.fit(X, X, nb_epoch=100, batch_size=80, validation_split=0.3)\n\nl_weights = np.round(model.layers[0].get_weights()[0],3)\n\nprint l_weights.argmax(axis=0)\nprint l_weights.max(axis=0)\n<\/code><\/pre>\n\n<p>And I'm getting:<\/p>\n\n<pre><code>Train on 69999 samples, validate on 30000 samples\nEpoch 1\/100\n69999\/69999 [==============================] - 1s - loss: 0.2092 - val_loss: 0.1564\nEpoch 2\/100\n69999\/69999 [==============================] - 1s - loss: 0.1536 - val_loss: 0.1510\nEpoch 3\/100\n69999\/69999 [==============================] - 1s - loss: 0.1484 - val_loss: 0.1459\n.\n.\n.\nEpoch 98\/100\n69999\/69999 [==============================] - 1s - loss: 0.0055 - val_loss: 0.0054\nEpoch 99\/100\n69999\/69999 [==============================] - 1s - loss: 0.0053 - val_loss: 0.0053\nEpoch 100\/100\n69999\/69999 [==============================] - 1s - loss: 0.0051 - val_loss: 0.0051\n[ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83]\n[ 0.85000002  0.85100001  0.79799998  0.80500001  0.82700002  0.81900001\n  0.792       0.829       0.81099999  0.80800003  0.84899998  0.829       0.852\n  0.79500002  0.84100002  0.81099999  0.792       0.80800003  0.85399997\n  0.82999998  0.85100001  0.84500003  0.847       0.79699999  0.81400001\n  0.84100002  0.81        0.85100001  0.80599999  0.84500003  0.824\n  0.81999999  0.82999998  0.79100001  0.81199998  0.829       0.85600001\n  0.84100002  0.792       0.847       0.82499999  0.84500003  0.796\n  0.82099998  0.81900001  0.84200001  0.83999997  0.815       0.79500002\n  0.85100001  0.83700001  0.85000002  0.79900002  0.84100002  0.79699999\n  0.838       0.847       0.84899998  0.83700001  0.80299997  0.85399997\n  0.84500003  0.83399999  0.83200002  0.80900002  0.85500002  0.83899999\n  0.79900002  0.83399999  0.81        0.79100001  0.81800002  0.82200003\n  0.79100001  0.83700001  0.83600003  0.824       0.829       0.82800001\n  0.83700001  0.85799998  0.81999999  0.84299999  0.83999997]\n<\/code><\/pre>\n\n<p>When I used only 5 numbers as an input and printed the actual weights I got this:<\/p>\n\n<pre><code>array([[ 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.]], dtype=float32)\n<\/code><\/pre>\n"}]}
{"QuestionId":41559723,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-10T01:21:07.513","AcceptedAnswerId":null,"OwnerUserId":7345996.0,"Title":"Does the tf.nn.conv2d_transpose transpose the filter?","Body":"<p>I have a trained feedforward CNN. The shape of filter is [height, width, in_channels, out_channels]. And i want to use those filter to do deconv, we know that the deconv process needs the transpose of the filter.Do i need to transpose the filter mannuly, or the  will do it inside the and all we need to do is pass the trained filter to ?<\/p>\n","answers":[{"AnswerId":"52070350","CreationDate":"2018-08-29T05:54:32.700","ParentId":null,"OwnerUserId":"5915270","Title":null,"Body":"<p>According to tensorflow documentation, you have to change the filter shape. Here I will put how this fact has described in latest TF documentation 1.9.<\/p>\n\n<p>For the tf.nn.conv_2d the filter variable should be -  A 4-D tensor of <strong>shape [filter_height, filter_width, <em>in_channels, out_channels<\/em>]<\/strong><\/p>\n\n<p>For the tf.nn.conv_2d_transpose the filter variable should be -  A 4-D tensor of <strong>shape [filter_height, filter_width, <em>out_channels,in_channels<\/em>]<\/strong><\/p>\n\n<ul>\n<li>Here clearly we can see we need to change the change the shape with respect to out_channels<\/li>\n<\/ul>\n"},{"AnswerId":"41568172","CreationDate":"2017-01-10T11:50:01.707","ParentId":null,"OwnerUserId":"7345996","Title":null,"Body":"<p>We needn't to transpose the filter manually.In general, wo organize our code in the following way.<\/p>\n\n<pre><code>stride = [1,1,1,1]\nconv1W = tf.Variable(tf.random.normal[4,4,3,20])\nconv1 = tf.nn.conv2d(input, conv1W, strides=stride, padding='SAME')\nconv1 = tf.nn.relu(conv1)\n<\/code><\/pre>\n\n<p>Then, do the deconv process<\/p>\n\n<pre><code>deconv1 = tf.nn.conv2d_transpose(conv1, conv1W, output_shape=[batch_size,output_height, output_width, output_channels],strides=stride)\nres = tf.nn.relu(deconv1)\n<\/code><\/pre>\n\n<p>The <code>res<\/code> is the result of deconv process.<\/p>\n\n<p>In a word, the <code>filter<\/code> and <code>stride<\/code> using in deconv process is the same as the <code>filter<\/code> and <code>stride<\/code> using in conv process.<\/p>\n"}]}
{"QuestionId":41559814,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-10T01:33:56.923","AcceptedAnswerId":41560826.0,"OwnerUserId":6776821.0,"Title":"Build conditional graph with first axis shape is \"None\" in tensorflow","Body":"<p>When at the graph building phase, suppose the tensor  which is a <strong>neural network's fully connected layer<\/strong>. <\/p>\n\n<p>So assume the shape of  is . I want to set the last column like this in python:<\/p>\n\n<pre><\/pre>\n\n<p>I can only use  when  only has 1 row like this:<\/p>\n\n<pre><\/pre>\n\n<p>How can I check above when first dimension of   is ?<\/p>\n","answers":[{"AnswerId":"41560826","CreationDate":"2017-01-10T03:49:50.293","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>Like this?<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\na = tf.placeholder(tf.int32, shape=[None, 5])\n\nr, c = a.get_shape()\nx_split = tf.split(1, c, a) # split a along axis 1\nlast_col = x_split[-1]\n\n\nmask = tf.greater(last_col, tf.constant(6))\ncond = tf.where(mask,\n                tf.add(last_col, tf.constant(1)), # if true, add one\n                tf.add(last_col, tf.constant(-1)))# if false, minus one\n\nx_split = x_split[0:-1]\nx_split.append(cond)\nans = tf.concat(1, x_split)\n\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    arr = np.array([[1, 2, 3, 4, 5],\n                    [6, 7, 8, 9, 10]])\n    the_ans = sess.run(ans, feed_dict={a: arr})\n'''\nthe_ans is \n[[ 1  2  3  4  4]\n [ 6  7  8  9 11]]\n'''\n<\/code><\/pre>\n"}]}
{"QuestionId":41560212,"AnswerCount":1,"Tags":"<python><lambda><tensorflow><convolution>","CreationDate":"2017-01-10T02:33:09.947","AcceptedAnswerId":41560295.0,"OwnerUserId":1145744.0,"Title":"Tensorflow <lambda>() got an unexpected keyword argument 'partition_info'","Body":"<p>I'm trying to run the code probably written for a previous Tensorflow version. I have fixed many errors related to keywords mismatch<\/p>\n\n<p>Now I get <\/p>\n\n<p>But there isn't <em>partition_info<\/em> keyword in my code. <\/p>\n\n<p>Here the piece of code that causes the problem:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41560295","CreationDate":"2017-01-10T02:43:54.337","ParentId":null,"OwnerUserId":"2132213","Title":null,"Body":"<p>It seems that the function you provide as the <code>weights_initializer<\/code> argument should expect a keyword argument <code>partition_info<\/code>, but your lambda does not.<\/p>\n\n<p>See the <code>_initializer<\/code> function defined in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/init_ops.py#L76\" rel=\"nofollow noreferrer\">ones_initalizer<\/a> for an example of the function signature expected for <code>weights_initializer<\/code>:<\/p>\n\n<pre><code>  def _initializer(shape, dtype=dtype, partition_info=None):\n    return constant_op.constant(1, dtype=dtype, shape=shape)\n<\/code><\/pre>\n\n<p>In that same file, if you look down at <code>uniform_unit_scaling_initializer<\/code> and <code>variance_scaling_initializer<\/code>, you can see examples of weights initializers that actually use the <code>partition_info<\/code> argument; but most of the initializers in the file just ignore it.<\/p>\n"}]}
{"QuestionId":41560621,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-10T03:23:08.860","AcceptedAnswerId":null,"OwnerUserId":6057413.0,"Title":"Tensorflow in Python","Body":"<p>While training tensorflow in python, the program gets killed after sometime. There is nothing getting printed in the logs as error or warning. <\/p>\n\n<p>I am using Fedora24, Python 2.7 and Tensorflow 0.12<\/p>\n\n<p>This is the link of the code I tried. <a href=\"https:\/\/github.com\/inikdom\/neural-sentiment\/blob\/master\/train.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/inikdom\/neural-sentiment\/blob\/master\/train.py<\/a><\/p>\n\n<p>The output is as follows:<\/p>\n\n<p>python train.py\nNo hyper parameter changed detected, using old checkpoint...\nvocab mapping found...\nCreating model with...\nNumber of hidden layers: 2\nNumber of units per layer: 50\nDropout: 0.5\nVocab size is: 20002\n['data2.npy', 'data3.npy', 'data0.npy', 'data1.npy']\nNumber of training examples per batch: 200,<br>\nNumber of batches per epoch: 250\nCreated model with fresh parameters.\nBeggining training...\nMaximum number of epochs to train for: 50\nBatch size: 200\nStarting learning rate: 0.01\nLearning rate decay factor: 0.97\nTest size is: 14800, splitting into 74 batches\nKilled<\/p>\n\n<p>I do not know if this is caused by the session.run method. I cannot avoid that method as well. Please help..<\/p>\n","answers":[{"AnswerId":"41561546","CreationDate":"2017-01-10T05:12:04.613","ParentId":null,"OwnerUserId":"6057413","Title":null,"Body":"<p>I figured it out. Its because of memory usage. Just increase the memory, if you are using a VM.<\/p>\n"}]}
{"QuestionId":41560780,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-10T03:44:31.480","AcceptedAnswerId":null,"OwnerUserId":2388930.0,"Title":"TensorFlow: Read batch features to an array","Body":"<p>I am using  (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/contrib.learn\/input_processing#read_batch_features\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/contrib.learn\/input_processing#read_batch_features<\/a>) to read in  protos as part of my input function, which returns a dict of  objects. After training my model, calling  on my  returns one batch of predictions as an array, which I would like to compare to the known values.<\/p>\n\n<p>I try to obtain the known values by calling , where  is a  of known values, returned from the input function. However, at this point, my program hangs. I suspect it is stuck in an infinite loop reading labels from the disk, rather than just reading one batch as I would like.<\/p>\n\n<p>Is this the correct way to obtain one batch of values in the  ?<\/p>\n\n<p>Edit: I have tried to start the queue runners, is the following correct?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41566133","CreationDate":"2017-01-10T10:04:03.107","ParentId":null,"OwnerUserId":"1246987","Title":null,"Body":"<p>The whole setup you need is:<\/p>\n\n<pre><code>_, labels = eval_input_fn()\n\nwith tf.Session() as sess:\n        sess.run([\n            tf.local_variables_initializer(),\n            tf.global_variables_initializer()\n        ])\n\n        coord = tf.train.Coordinator()\n        threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\n        try:\n            while not coord.should_stop():\n                print(sess.run(label))\n\n        except tf.errors.OutOfRangeError as error:\n            coord.request_stop(error)\n        finally:\n            coord.request_stop()\n            coord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":41560982,"AnswerCount":2,"Tags":"<python><r><tensorflow><anaconda>","CreationDate":"2017-01-10T04:08:28.867","AcceptedAnswerId":42159731.0,"OwnerUserId":4092281.0,"Title":"How to use TensorFlow in R if I have Anaconda Python already installed?","Body":"<p>In the <a href=\"https:\/\/rstudio.github.io\/tensorflow\/\" rel=\"nofollow noreferrer\">guide about using TensorFlow<\/a> in R, they suggest not to use the Anaconda TensorFlow installation. Does this mean Anaconda Python cannot coexist with using TensorFlow in R?<\/p>\n","answers":[{"AnswerId":"41583760","CreationDate":"2017-01-11T06:03:28.680","ParentId":null,"OwnerUserId":"4989496","Title":null,"Body":"<p>There is an issue using anaconda along with R. Please follow <a href=\"https:\/\/github.com\/rstudio\/tensorflow\" rel=\"nofollow noreferrer\">this<\/a> link for more info. <\/p>\n\n<p>Directly from the link above:<\/p>\n\n<p>TensorFlow for R is currently only compatible with OS X and Linux (support for Windows will likely be added over the next few months).<\/p>\n\n<p>You should NOT install TensorFlow with Anaconda as there are issues with the way Anaconda builds the python shared library that prevent dynamic linking from R. There are potential solution to the above problem see <a href=\"https:\/\/github.com\/ContinuumIO\/anaconda-issues\/issues\/498\" rel=\"nofollow noreferrer\">this<\/a> for more details.<\/p>\n"},{"AnswerId":"42159731","CreationDate":"2017-02-10T12:44:19.387","ParentId":null,"OwnerUserId":"468305","Title":null,"Body":"<p>This is no longer an issue, the documentation was updated too.<\/p>\n\n<p>See here: <a href=\"https:\/\/github.com\/rstudio\/tensorflow\/commit\/4e1e11d6ba2fe7efe1a03356f96172dbf8db365e\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rstudio\/tensorflow\/commit\/4e1e11d6ba2fe7efe1a03356f96172dbf8db365e<\/a><\/p>\n"}]}
{"QuestionId":41561902,"AnswerCount":2,"Tags":"<python><linux><tensorflow>","CreationDate":"2017-01-10T05:44:36.633","AcceptedAnswerId":null,"OwnerUserId":7397359.0,"Title":"TensorFlow: binary installation error on CentOS","Body":"<p>I was trying to have TensorFlow installed on my 32-bit CentOS 6.8 linux machine. I have python 3.6 installed with pip 9.0.1. However as I try this command:<\/p>\n\n<pre><\/pre>\n\n<p>I got:<\/p>\n\n<pre><\/pre>\n\n<p>The full log info are pasted below:<\/p>\n\n<pre><\/pre>\n\n<p>Not sure what the problem could be. Any suggestion or advice would be greatly appreciated!<\/p>\n\n<hr>\n\n<p>Update:\n  I have tried the same code on a 64bit centOS 6.5\uff0c getting the same error. So maybe I can conclude this has nothing to do with the x86_64 part?<\/p>\n","answers":[{"AnswerId":"41659071","CreationDate":"2017-01-15T08:04:40.970","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>The issue is, you are using py3.6 and 32 bit.\npip looks at filenames to figure out what the package is built for:<\/p>\n\n<pre><code>tensorflow-0.7.1-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>This package name says it is for python 2.7 for a 64 bit linux system.\n<a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0423\/\" rel=\"nofollow noreferrer\">https:\/\/www.python.org\/dev\/peps\/pep-0423\/<\/a><\/p>\n\n<p>for python 3.6, it is not a guaranteed way, but you may download the pip package for python 3.5 and rename it by replacing all \"35\"s in the name with \"36\"<\/p>\n"},{"AnswerId":"41561987","CreationDate":"2017-01-10T05:51:58.280","ParentId":null,"OwnerUserId":"1646185","Title":null,"Body":"<p>Your machine\/os is 32-bit, however tensorflow-0.7.1-cp27-none-linux_x86_64.whl only fit 64-bit machine.\nAnd I can't find 32-bit version tensorflow package in tensorflow home page:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/os_setup<\/a><\/p>\n\n<p>It seems the only way to solve it is try a 64-bit machine and 64-bit os.<\/p>\n"}]}
{"QuestionId":41562441,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-10T06:28:10.897","AcceptedAnswerId":null,"OwnerUserId":7397753.0,"Title":"Tensorflow confusing error reading images with labels","Body":"<p>I'm trying to read a set of images and labels for batch training, but I keep getting the error: <\/p>\n\n<pre><\/pre>\n\n<p>Here is my very reduced code that reproduces the error:<\/p>\n\n<pre><\/pre>\n\n<p>I only used one image in this example.\nWhat is the cause of this error in this simple example?<\/p>\n\n<p>A related question that didn't help:\n<a href=\"https:\/\/stackoverflow.com\/questions\/34340489\/tensorflow-read-images-with-labels\">Tensorflow read images with labels<\/a><\/p>\n","answers":[{"AnswerId":"41599387","CreationDate":"2017-01-11T19:42:23.853","ParentId":null,"OwnerUserId":"1246987","Title":null,"Body":"<p>If you want evaluate multiple tensors in a <code>session<\/code> you need to pass them in an array <code>fetches<\/code> to <code>session.run(fetches)<\/code>.<\/p>\n\n<p>So change <code>sess.run(ibatch, lbatch)<\/code> to <code>sess.run([ibatch, lbatch])<\/code> see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management\" rel=\"nofollow noreferrer\">docs<\/a> for more.<\/p>\n"}]}
{"QuestionId":41562652,"AnswerCount":0,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-01-10T06:42:42.737","AcceptedAnswerId":null,"OwnerUserId":5082661.0,"Title":"Fully Connected Layer Followed by LSTM Layer","Body":"<p>How do i combine a Tensorflow Fully Connected  Layer, which is then followed by a LSTM Layer. My goal is to feed data of batch_size=batch_size, sequence length seq_length, and dimension 1. Target is 21 dim one_hot vector.<\/p>\n\n<p>Here is the code i tried. It throws an error <strong>Shape must be rank 2 but is rank 3 for W_first line<\/strong>. What am i doing wrong<\/p>\n\n<pre><\/pre>\n\n<p>Thanks in advance!<\/p>\n","answers":[]}
{"QuestionId":41562718,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-10T06:46:53.330","AcceptedAnswerId":null,"OwnerUserId":1279459.0,"Title":"FLOPs for tf.embedding_lookup_sparse","Body":"<p>When executing tf.embedding_lookup_sparse in tensorflow, are there any floating operations performed? And if so, is it possible to find the number of FLOPs carried by the operation?<\/p>\n","answers":[{"AnswerId":"41598817","CreationDate":"2017-01-11T19:07:52.813","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>Depending on the passed parameters, any number of FLOPs are performed. While the actual <code>array_ops.gather(...)<\/code> call does not perform FLOPs, there are a number of other places where computation may or may not occur. <\/p>\n\n<ul>\n<li>Integer ops are performed when sharding the the input tensors. (Can be either <code>mod<\/code> or <code>div<\/code>.) <\/li>\n<li>Incorporating weights (if not <code>None<\/code>)<\/li>\n<li>The combiner to reduce values together (can be <code>None<\/code>, <code>\"sum\"<\/code>, <code>\"mean\"<\/code>, or <code>\"sqrtn\"<\/code>.<\/li>\n<\/ul>\n\n<p>You can see the implementation here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/embedding_ops.py#L203-L366\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/embedding_ops.py#L203-L366<\/a><\/p>\n"}]}
{"QuestionId":41563720,"AnswerCount":5,"Tags":"<deep-learning><keras><keras-layer>","CreationDate":"2017-01-10T07:51:25.727","AcceptedAnswerId":41595178.0,"OwnerUserId":1934212.0,"Title":"Error when checking model input: expected convolution2d_input_1 to have 4 dimensions, but got array with shape (32, 32, 3)","Body":"<p>I want to train a deep network starting with the following layer:<\/p>\n\n<pre><\/pre>\n\n<p>using <\/p>\n\n<pre><\/pre>\n\n<p>with the following generator:<\/p>\n\n<pre><\/pre>\n\n<p>(validation generator looks similar).<\/p>\n\n<p>During training, I get the error: <\/p>\n\n<pre><\/pre>\n\n<p>How can that be, with a first layer<\/p>\n\n<pre><\/pre>\n\n<p>?<\/p>\n","answers":[{"AnswerId":"59102840","CreationDate":"2019-11-29T10:16:04.467","ParentId":null,"OwnerUserId":"2525417","Title":null,"Body":"<p>Probably <strong>very trivial<\/strong>, but I solved it by just converting the input to <em>numpy array<\/em>. <\/p>\n\n<p>For the neural network architecture,<\/p>\n\n<pre><code>    model = Sequential()\n    model.add(Conv2D(32, (5, 5), activation=\"relu\", input_shape=(32, 32, 3)))\n<\/code><\/pre>\n\n<p>When the input was,<\/p>\n\n<pre><code>    n_train = len(train_y_raw)\n    train_X = [train_X_raw[:,:,:,i] for i in range(n_train)]\n    train_y = [train_y_raw[i][0] for i in range(n_train)]\n<\/code><\/pre>\n\n<p>I got the error,<a href=\"https:\/\/i.stack.imgur.com\/USoqV.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/USoqV.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>But when I changed it to,<\/p>\n\n<pre><code>   n_train = len(train_y_raw)\n   train_X = np.asarray([train_X_raw[:,:,:,i] for i in range(n_train)])\n   train_y = np.asarray([train_y_raw[i][0] for i in range(n_train)])\n<\/code><\/pre>\n\n<p>It fixed the issue.<\/p>\n"},{"AnswerId":"57066969","CreationDate":"2019-07-17T00:18:22.520","ParentId":null,"OwnerUserId":"10480983","Title":null,"Body":"<p>It is as simple as to Add one dimension, so I was going through the tutorial taught by Siraj Rawal on CNN Code Deployment tutorial, it was working on his terminal, but the same code was not working on my terminal, so I did some research about it and solved, I don't know if that works for you all. Here I have come up with solution;<\/p>\n\n<p>Unsolved code lines which gives you problem:<\/p>\n\n<pre><code>if K.image_data_format() == 'channels_first':\n    x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)\n    x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)\n    print(x_train.shape)\n    input_shape = (1, img_rows, img_cols)\nelse:\n    x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols)\n    x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols)\n    input_shape = (img_rows, img_cols, 1)\n<\/code><\/pre>\n\n<p>Solved Code: <\/p>\n\n<pre><code>if K.image_data_format() == 'channels_first':\n    x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)\n    x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)\n    print(x_train.shape)\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<\/code><\/pre>\n\n<p>Please share the feedback here if that worked for you.<\/p>\n"},{"AnswerId":"41595178","CreationDate":"2017-01-11T15:50:53.760","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>The input shape you have defined is the shape of a single sample. The model itself expects some array of samples as input (even if its an array of length 1).<\/p>\n\n<p>Your output really should be 4-d, with the 1st dimension to enumerate the samples. i.e. for a single image you should return a shape of (1, 32, 32, 3).<\/p>\n\n<p>You can find more information <a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"noreferrer\" title=\"here\">here<\/a> under \"Convolution2D\"\/\"Input shape\"<\/p>\n"},{"AnswerId":"58884508","CreationDate":"2019-11-15T20:52:24.080","ParentId":null,"OwnerUserId":"12380606","Title":null,"Body":"<p>it depends on how you actually order your data,if its on a channel first basis then you should reshape your data:\nx_train=x_train.reshape(x_train.shape[0],channel,width,height)<\/p>\n\n<p>if its channel last:\nx_train=s_train.reshape(x_train.shape[0],width,height,channel)<\/p>\n"},{"AnswerId":"55950416","CreationDate":"2019-05-02T10:26:54.637","ParentId":null,"OwnerUserId":"11441177","Title":null,"Body":"<pre><code>x_train = x_train.reshape(-1,28, 28, 1)   #Reshape for CNN -  should work!!\nx_test = x_test.reshape(-1,28, 28, 1)\nhistory_cnn = cnn.fit(x_train, y_train, epochs=5, validation_data=(x_test, y_test))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<blockquote>\n  <p>Train on 60000 samples, validate on 10000 samples\n  Epoch 1\/5\n  60000\/60000 [==============================] - 157s 3ms\/step - loss: 0.0981 - acc: 0.9692 - val_loss: 0.0468 - val_acc: 0.9861\n  Epoch 2\/5\n  60000\/60000 [==============================] - 157s 3ms\/step - loss: 0.0352 - acc: 0.9892 - val_loss: 0.0408 - val_acc: 0.9879\n  Epoch 3\/5\n  60000\/60000 [==============================] - 159s 3ms\/step - loss: 0.0242 - acc: 0.9924 - val_loss: 0.0291 - val_acc: 0.9913\n  Epoch 4\/5\n  60000\/60000 [==============================] - 165s 3ms\/step - loss: 0.0181 - acc: 0.9945 - val_loss: 0.0361 - val_acc: 0.9888\n  Epoch 5\/5\n  60000\/60000 [==============================] - 168s 3ms\/step - loss: 0.0142 - acc: 0.9958 - val_loss: 0.0354 - val_acc: 0.9906<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41564212,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-10T08:21:39.180","AcceptedAnswerId":null,"OwnerUserId":6727583.0,"Title":"How does Tensorflow do quantization and dequantization?","Body":"<p>According to the blog post \"<a href=\"https:\/\/petewarden.com\/2016\/05\/03\/how-to-quantize-neural-networks-with-tensorflow\/\" rel=\"noreferrer\">https:\/\/petewarden.com\/2016\/05\/03\/how-to-quantize-neural-networks-with-tensorflow\/<\/a>\", Tensorflow quantizes values before they go into a layer. After being processed by the layer, the values are dequantized. Tensorflow quantizes values by rescaling the values between 0 and 255, so it needs to keep \"min\" and \"max\" to dequantize the values.<\/p>\n\n<p>I would like to ask:\n1. how the \"min\" and \"max\" in the outputs of a \"quantization\" op are determined? I mean, if we simply find the minimum and maximum value and set them to 0 and 255, we will get data overflow or underflow when doing convolution.\n2. how the \"min\" and \"max\" in the outputs of a \"convolution\" op are determined? Both weights and activations are quantized, so there are two sets of \"min\" and \"max\". How does a convolution op combine them to form a single set of \"min\" and \"max\"?<\/p>\n","answers":[{"AnswerId":"45174137","CreationDate":"2017-07-18T18:07:30.190","ParentId":null,"OwnerUserId":"2910092","Title":null,"Body":"<p>TensorFlow uses i.a. <a href=\"https:\/\/github.com\/google\/gemmlowp\" rel=\"nofollow noreferrer\">gemmlowp<\/a> for low-precision matrix multiplications. Although 8-bit values are used as inputs, intermediate results are 32-bit values. These 32-bit values are converted back to 8-bit before returning the results.<\/p>\n\n<p>From <a href=\"https:\/\/github.com\/google\/gemmlowp\/blob\/master\/doc\/low-precision.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/google\/gemmlowp\/blob\/master\/doc\/low-precision.md<\/a> :<\/p>\n\n<blockquote>\n  <p>To avoid overflow, we internally accumulate results on more than 8 bits, and at the end we keep only some significant 8 bits.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41564321,"AnswerCount":1,"Tags":"<image-processing><tensorflow><reshape>","CreationDate":"2017-01-10T08:28:30.903","AcceptedAnswerId":41599044.0,"OwnerUserId":1246987.0,"Title":"Split image tensor into small patches","Body":"<p>I have an  of shape  which I want to split into 7x7 patches.<\/p>\n\n<pre><\/pre>\n\n<p>Using<\/p>\n\n<pre><\/pre>\n\n<p>unfortunately does not work in practice as  mixes up the pixel order (if you view  you will only see noise).<\/p>\n\n<p>So my new approach was to use :<\/p>\n\n<pre><\/pre>\n\n<p>this indeed preserves the image pixel order unfortunately it creates a lot of OPs which is not very good.<\/p>\n\n<p><strong>How do I split an image into smaller patches with less OPs?<\/strong><\/p>\n\n<p><strong>Update1:<\/strong><\/p>\n\n<p>I ported the <a href=\"https:\/\/stackoverflow.com\/questions\/16856788\/slice-2d-array-into-smaller-2d-arrays\">answer of this question for numpy<\/a> to tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>but I think there is still room for improvement.<\/p>\n\n<p><strong>Update2:<\/strong><\/p>\n\n<p>This will convert patches back to the original image.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41599044","CreationDate":"2017-01-11T19:20:04.720","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>I think your issue is somewhere else. I wrote the following code snippet (using a smaller 14x14 image so that I could hand-check all the values), and confirmed that your initial code did the correct operations:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nIMAGE_SIZE = [1, 14, 14, 1]\nPATCH_SIZE = [1, 7, 7, 1]\n\ninput_image = np.reshape(np.array(xrange(14*14)), IMAGE_SIZE)\nimage = tf.placeholder(dtype=tf.int32, shape=IMAGE_SIZE)\nimage_patches = tf.extract_image_patches(\n    image, PATCH_SIZE, PATCH_SIZE, [1, 1, 1, 1], 'VALID')\nimage_patches_reshaped = tf.reshape(image_patches, [-1, 7, 7, 1])\n\nsess = tf.Session()\n\n(output, output_reshaped) = sess.run(\n    (image_patches, image_patches_reshaped),\n    feed_dict={image: input_image})\n\nprint \"Output (shape: %s):\" % (output.shape,)\nprint output\n\nprint \"Reshaped (shape: %s):\" % (output_reshaped.shape,)\nprint output_reshaped\n<\/code><\/pre>\n\n<p>The output was:<\/p>\n\n<pre><code>python resize.py \nOutput (shape: (1, 2, 2, 49)):\n[[[[  0   1   2   3   4   5   6  14  15  16  17  18  19  20  28  29  30  31\n     32  33  34  42  43  44  45  46  47  48  56  57  58  59  60  61  62  70\n     71  72  73  74  75  76  84  85  86  87  88  89  90]\n   [  7   8   9  10  11  12  13  21  22  23  24  25  26  27  35  36  37  38\n     39  40  41  49  50  51  52  53  54  55  63  64  65  66  67  68  69  77\n     78  79  80  81  82  83  91  92  93  94  95  96  97]]\n\n  [[ 98  99 100 101 102 103 104 112 113 114 115 116 117 118 126 127 128 129\n    130 131 132 140 141 142 143 144 145 146 154 155 156 157 158 159 160 168\n    169 170 171 172 173 174 182 183 184 185 186 187 188]\n   [105 106 107 108 109 110 111 119 120 121 122 123 124 125 133 134 135 136\n    137 138 139 147 148 149 150 151 152 153 161 162 163 164 165 166 167 175\n    176 177 178 179 180 181 189 190 191 192 193 194 195]]]]\nReshaped (shape: (4, 7, 7, 1)):\n[[[[  0]\n   [  1]\n   [  2]\n   [  3]\n   [  4]\n   [  5]\n   [  6]]\n\n  [[ 14]\n   [ 15]\n   [ 16]\n   [ 17]\n   [ 18]\n   [ 19]\n   [ 20]]\n\n  [[ 28]\n   [ 29]\n   [ 30]\n   [ 31]\n   [ 32]\n   [ 33]\n   [ 34]]\n\n  [[ 42]\n   [ 43]\n   [ 44]\n   [ 45]\n   [ 46]\n   [ 47]\n   [ 48]]\n\n  [[ 56]\n   [ 57]\n   [ 58]\n   [ 59]\n   [ 60]\n   [ 61]\n   [ 62]]\n\n  [[ 70]\n   [ 71]\n   [ 72]\n   [ 73]\n   [ 74]\n   [ 75]\n   [ 76]]\n\n  [[ 84]\n   [ 85]\n   [ 86]\n   [ 87]\n   [ 88]\n   [ 89]\n   [ 90]]]\n\n\n [[[  7]\n   [  8]\n   [  9]\n   [ 10]\n   [ 11]\n   [ 12]\n   [ 13]]\n\n  [[ 21]\n   [ 22]\n   [ 23]\n   [ 24]\n   [ 25]\n   [ 26]\n   [ 27]]\n\n  [[ 35]\n   [ 36]\n   [ 37]\n   [ 38]\n   [ 39]\n   [ 40]\n   [ 41]]\n\n  [[ 49]\n   [ 50]\n   [ 51]\n   [ 52]\n   [ 53]\n   [ 54]\n   [ 55]]\n\n  [[ 63]\n   [ 64]\n   [ 65]\n   [ 66]\n   [ 67]\n   [ 68]\n   [ 69]]\n\n  [[ 77]\n   [ 78]\n   [ 79]\n   [ 80]\n   [ 81]\n   [ 82]\n   [ 83]]\n\n  [[ 91]\n   [ 92]\n   [ 93]\n   [ 94]\n   [ 95]\n   [ 96]\n   [ 97]]]\n\n\n [[[ 98]\n   [ 99]\n   [100]\n   [101]\n   [102]\n   [103]\n   [104]]\n\n  [[112]\n   [113]\n   [114]\n   [115]\n   [116]\n   [117]\n   [118]]\n\n  [[126]\n   [127]\n   [128]\n   [129]\n   [130]\n   [131]\n   [132]]\n\n  [[140]\n   [141]\n   [142]\n   [143]\n   [144]\n   [145]\n   [146]]\n\n  [[154]\n   [155]\n   [156]\n   [157]\n   [158]\n   [159]\n   [160]]\n\n  [[168]\n   [169]\n   [170]\n   [171]\n   [172]\n   [173]\n   [174]]\n\n  [[182]\n   [183]\n   [184]\n   [185]\n   [186]\n   [187]\n   [188]]]\n\n\n [[[105]\n   [106]\n   [107]\n   [108]\n   [109]\n   [110]\n   [111]]\n\n  [[119]\n   [120]\n   [121]\n   [122]\n   [123]\n   [124]\n   [125]]\n\n  [[133]\n   [134]\n   [135]\n   [136]\n   [137]\n   [138]\n   [139]]\n\n  [[147]\n   [148]\n   [149]\n   [150]\n   [151]\n   [152]\n   [153]]\n\n  [[161]\n   [162]\n   [163]\n   [164]\n   [165]\n   [166]\n   [167]]\n\n  [[175]\n   [176]\n   [177]\n   [178]\n   [179]\n   [180]\n   [181]]\n\n  [[189]\n   [190]\n   [191]\n   [192]\n   [193]\n   [194]\n   [195]]]]\n<\/code><\/pre>\n\n<p>Based on the reshaped output, you can see it is a 4x7x7x1 with values for the first patch as: [0-7),[14-21), [28-35), [42-49), [56-63), [70-77), and [84-91), which corresponds to the upper left 7x7 grid.<\/p>\n\n<p>Perhaps you can explain a bit further what's going on when it doesn't work correctly?<\/p>\n"}]}
{"QuestionId":41567895,"AnswerCount":2,"Tags":"<python><scikit-learn><tensorflow><k-means>","CreationDate":"2017-01-10T11:37:23.563","AcceptedAnswerId":41568439.0,"OwnerUserId":470184.0,"Title":"Will scikit-learn utilize GPU?","Body":"<p>Reading implementation of scikit-learn in tensroflow : <a href=\"http:\/\/learningtensorflow.com\/lesson6\/\" rel=\"noreferrer\">http:\/\/learningtensorflow.com\/lesson6\/<\/a> and scikit-learn : <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.cluster.KMeans.html\" rel=\"noreferrer\">http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.cluster.KMeans.html<\/a> I'm struggling to decide which implementation to use.<\/p>\n\n<p>scikit-learn is installed as part of the tensorflow docker container so can use either implementation. <\/p>\n\n<p>Reason to use scikit-learn : <\/p>\n\n<blockquote>\n  <p>scikit-learn contains less boiler plate than the tensorflow\n  implementation.<\/p>\n<\/blockquote>\n\n<p>Reason to use tensorflow :<\/p>\n\n<blockquote>\n  <p>If running on Nvidia GPU the algorithm wilk be run against in parallel\n  , I'm not sure if scikit-learn will utilise all available GPU's ?<\/p>\n<\/blockquote>\n\n<p>Reading <a href=\"https:\/\/www.quora.com\/What-are-the-main-differences-between-TensorFlow-and-SciKit-Learn\" rel=\"noreferrer\">https:\/\/www.quora.com\/What-are-the-main-differences-between-TensorFlow-and-SciKit-Learn<\/a> <\/p>\n\n<blockquote>\n  <p>TensorFlow is more  low-level; basically, the Lego bricks that help\n  you to implement machine learning algorithms whereas scikit-learn\n  offers you off-the-shelf algorithms, e.g., algorithms for\n  classification such as SVMs, Random Forests, Logistic Regression, and\n  many, many more. TensorFlow really shines if you want to implement\n  deep learning algorithms, since it allows you to take advantage of\n  GPUs for more efficient training.<\/p>\n<\/blockquote>\n\n<p>This statement re-enforces my assertion that \"scikit-learn contains less boiler plate than the tensorflow implementation\" but also suggests scikit-learn will not utilise all available GPU's ?<\/p>\n","answers":[{"AnswerId":"55810574","CreationDate":"2019-04-23T11:49:18.887","ParentId":null,"OwnerUserId":"8044221","Title":null,"Body":"<p>Simply No.<\/p>\n\n<p>Take a look at FAQ \"Will you add GPU support?\" provided by scikit-learn <a href=\"https:\/\/scikit-learn.org\/stable\/faq.html\" rel=\"nofollow noreferrer\">here<\/a>. They have given a clear explanation of why.<\/p>\n"},{"AnswerId":"41568439","CreationDate":"2017-01-10T12:04:14.840","ParentId":null,"OwnerUserId":"5213005","Title":null,"Body":"<p>Tensorflow only uses GPU if it is built against Cuda and CuDNN. By default none of both are going to use GPU, especially if it is running inside Docker, unless you use <a href=\"https:\/\/github.com\/NVIDIA\/nvidia-docker\" rel=\"noreferrer\">nvidia-docker<\/a> and an image capable of doing it.<\/p>\n\n<p>Scikit-learn is not intended to be used as a deep-learning framework, and seems that it doesn't support GPU computations.<\/p>\n\n<blockquote>\n  <p><strong>Why is there no support for deep or reinforcement learning \/ Will there be support for deep or reinforcement learning in scikit-learn?<\/strong><\/p>\n  \n  <p>Deep learning and reinforcement learning both require a rich\n  vocabulary to define an architecture, with deep learning additionally\n  requiring GPUs for efficient computing. However, neither of these fit\n  within the design constraints of scikit-learn; as a result, deep\n  learning and reinforcement learning are currently out of scope for\n  what scikit-learn seeks to achieve.<\/p>\n<\/blockquote>\n\n<p>Extracted from <a href=\"http:\/\/scikit-learn.org\/stable\/faq.html#why-is-there-no-support-for-deep-or-reinforcement-learning-will-there-be-support-for-deep-or-reinforcement-learning-in-scikit-learn\" rel=\"noreferrer\">http:\/\/scikit-learn.org\/stable\/faq.html#why-is-there-no-support-for-deep-or-reinforcement-learning-will-there-be-support-for-deep-or-reinforcement-learning-in-scikit-learn<\/a><\/p>\n\n<blockquote>\n  <p><strong>Will you add GPU support in scikit-learn?<\/strong><\/p>\n  \n  <p>No, or at least not in the near future. The main reason is that GPU\n  support will introduce many software dependencies and introduce\n  platform specific issues. scikit-learn is designed to be easy to\n  install on a wide variety of platforms. Outside of neural networks,\n  GPUs don\u2019t play a large role in machine learning today, and much\n  larger gains in speed can often be achieved by a careful choice of\n  algorithms.<\/p>\n<\/blockquote>\n\n<p>Extracted from <a href=\"http:\/\/scikit-learn.org\/stable\/faq.html#will-you-add-gpu-support\" rel=\"noreferrer\">http:\/\/scikit-learn.org\/stable\/faq.html#will-you-add-gpu-support<\/a><\/p>\n"}]}
{"QuestionId":41568239,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><keras>","CreationDate":"2017-01-10T11:54:26.360","AcceptedAnswerId":null,"OwnerUserId":6469175.0,"Title":"Issue with Python Neural Network Code (Keras)","Body":"<p>I want to build an neural net model where say the data of 100 rows is divided into 5 stacks of 20 rows each, now rather than iterating through it all at once I want to build a neural net based on the first 20 rows of data (ie first stack), then save the model(weights etc) and then pass the next stack (Next 20 rows ie row no 21-40) into the updated model (ie weights updated from previous model) rather and so on. Can someone tell me what such type of neural network is called?\nI just tried my first neural net yesterday where I iterated all data in batches (which I believe takes place in one epoch rather than multiple).<\/p>\n\n<p>Following is the Neural Net I made in Python using Keras(Tensorflow backend) can someone suggest me edits to make the following model as per my requirement?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41568980","CreationDate":"2017-01-10T12:33:10.593","ParentId":null,"OwnerUserId":"3697202","Title":null,"Body":"<p>It sounds like you want to train on your data in minibatches of size 20, and save the model after each minibatch. You don't need to change the shape of your input data for this - a matrix of shape (nb_datapoints, nb_features) works. Make sure to specify <code>batch_size=20<\/code> when you call <code>model.fit()<\/code>.<\/p>\n\n<p>In order to save your model after each minibatch, look into Keras <a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">callbacks<\/a>. You'd need to write your own custom callback, but you can model it after the existing <a href=\"https:\/\/keras.io\/callbacks\/#modelcheckpoint\" rel=\"nofollow noreferrer\"><code>ModelCheckpoint<\/code><\/a> callback - that saves the model after each epoch, so it should be relatively simple to customize it for your needs.<\/p>\n"}]}
{"QuestionId":41568721,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-10T12:18:54.077","AcceptedAnswerId":41573686.0,"OwnerUserId":3009947.0,"Title":"Using Pearson Correlation instead of accuracy in tensorflow to report performance","Body":"<p>In my TF code I defined the following in my model:<\/p>\n\n<pre><\/pre>\n\n<p>But I realised that it's more appropriate to use pearson correlation in my task. So, I changed it to this:<\/p>\n\n<pre><\/pre>\n\n<p>I call my model in the computational graph and in the training loop I have this evaluation code:<\/p>\n\n<pre><\/pre>\n\n<p>which calls this method: <\/p>\n\n<pre><\/pre>\n\n<p>All this I do within a TF session:<\/p>\n\n<pre><\/pre>\n\n<p>But I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>Sorry if the question is very basic - I'm quite a novice yet in TF. Could you please point out what's wrong here and how it can be fixed?<\/p>\n","answers":[{"AnswerId":"41573686","CreationDate":"2017-01-10T16:32:45.687","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.metrics\/metric_ops#streaming_pearson_correlation\" rel=\"nofollow noreferrer\"><code>streaming_pearson_correlation<\/code><\/a> function delegates to <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.metrics\/metric_ops#streaming_covariance\" rel=\"nofollow noreferrer\"><code>streaming_covariance<\/code><\/a>, and it creates <strong>four local variables<\/strong>, according to documentation.<\/p>\n\n<p><code>tf.initialize_all_variables()<\/code> does not initialize local variables.<\/p>\n\n<p>For Tensorflow r0.12+, use<\/p>\n\n<pre><code>tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())\n<\/code><\/pre>\n\n<p>For Tensorflow &lt;= r0.11, use<\/p>\n\n<pre><code>tf.group(tf.initialize_all_variables(), tf.initialize_variables(tf.local_variables()))\n<\/code><\/pre>\n\n<p>More: <a href=\"https:\/\/stackoverflow.com\/questions\/38910198\/what-is-a-local-variable-in-tensorflow\">What is a local variable in tensorflow?<\/a><\/p>\n"}]}
{"QuestionId":41569242,"AnswerCount":2,"Tags":"<tensorflow><multi-gpu><multiple-gpu>","CreationDate":"2017-01-10T12:49:01.803","AcceptedAnswerId":null,"OwnerUserId":4171200.0,"Title":"Tensorflow processing performance with multiple gpu","Body":"<p>friends!\nI have a question about processing with multiple gpu.\nI'm using 4 gpus and tried simple A^n + B^n example in 3 way like below.<\/p>\n\n<ol>\n<li><p>Single GPU<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Multiple GPU<\/p>\n\n<pre><\/pre><\/li>\n<li><p>No specific gpu designated (I think maybe all of gpu used)<\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n\n<p>when tried this, the result was incomprehensible.\nthe result was\n1. single gpu : 6.x seconds\n2. multiple gpu(2 gpus) : 2.x seconds\n3. no specific gpu designated(maybe 4 gpus) : 4.x seconds<\/p>\n\n<p>I cannot understand why #2 is faster than #3.\nAnyone can help me?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"42493208","CreationDate":"2017-02-27T18:29:42.327","ParentId":null,"OwnerUserId":"6626887","Title":null,"Body":"<p>In the third code sample (where no GPU was designated) Tensorflow didn't use all of your GPUs. By default if Tensorflow can find a GPU (\"\/gpu:0\") to use it assigns as many calculations as possible to that GPU. You would need to tell it specifically that you want want it to use all 4 like you did in the second code sample.<\/p>\n\n<p>From the Tensorflow documentation:<\/p>\n\n<blockquote>\n  <p>If you have more than one GPU in your system, the GPU with the lowest ID will be selected by default. If you would like to run on a different GPU, you will need to specify the preference explicitly:<\/p>\n<\/blockquote>\n\n<pre><code>with tf.device('\/gpu:2'):\n    tf code here\n<\/code><\/pre>\n"},{"AnswerId":"41597366","CreationDate":"2017-01-11T17:43:26.160","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>While the Tensorflow scheduler works well for single GPUs, it is not as good at optimizing the placement of computations on multiple GPUs yet. (Although it is being worked on presently.) Without further details, it's hard to know exactly what's going on. To get a get a better picture, you can log where the computations are actually being placed by the scheduler. You can do this by setting the <code>log_device_placement<\/code> flag on when creating the <code>tf.Session<\/code>:<\/p>\n\n<pre><code># Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n<\/code><\/pre>\n"}]}
{"QuestionId":41569858,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><data-science>","CreationDate":"2017-01-10T13:21:07.097","AcceptedAnswerId":41586755.0,"OwnerUserId":5628686.0,"Title":"Keras model.predict for multinomial logistic regression","Body":"<p>I'm training a model whose output is a softmax layer of size 19. When I try , for each input, I get what appears to be a probability distribution across the 19 classes. I tried , and got a numpy array of the size of , with each output equal to 0. How can I get one hot vectors for the output? <\/p>\n","answers":[{"AnswerId":"41586755","CreationDate":"2017-01-11T09:13:39.633","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So a documentation of <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\"><code>predcit_classes<\/code><\/a> is somehow misleading because if you check carefully its <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/8f7574437951a8f1042407d25bffb57292e3b6a6\/keras\/models.py\" rel=\"nofollow noreferrer\">implementation<\/a>, you'll find out that it works only for binary classification. In order to solve your problem you may use the <code>numpy<\/code> library (basically - a function <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.argmax.html\" rel=\"nofollow noreferrer\"><code>argmax<\/code><\/a>) in a following way:<\/p>\n\n<pre><code>import numpy as np\nclasses = np.argmax(model.predict(x), axis = 1)\n<\/code><\/pre>\n\n<p>.. in order to get an array with a class number for each example. In order to get a one-hot vector - you might use a <code>keras<\/code> built-in function <a href=\"https:\/\/keras.io\/utils\/np_utils\/#to_categorical\" rel=\"nofollow noreferrer\"><code>to_categorical<\/code><\/a> in a following manner:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.utils.np_utils import to_categorical\nclasses_one_hot = to_categorical(np.argmax(model.predict(x), axis = 1))\n<\/code><\/pre>\n"}]}
{"QuestionId":41570099,"AnswerCount":1,"Tags":"<deep-learning><theano><keras><keras-layer>","CreationDate":"2017-01-10T13:31:58.273","AcceptedAnswerId":null,"OwnerUserId":6828811.0,"Title":"Keras Convolution2d layer input shape","Body":"<p>i am basically trying to build a deep model which consists of many convolution2d layers followed by maxpooling 2d as follows :<\/p>\n\n<pre><\/pre>\n\n<p>However , i am getting the following error :<\/p>\n\n<blockquote>\n  <p>File\n  \"\/home\/user\/anaconda2\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\",\n  line 100, in standardize_input_data\n      str(array.shape)) Exception: Error when checking model input: expected convolution2d_input_1 to have 4 dimensions, but got array\n  with shape (8000, 180, 54)\n  Blockquote<\/p>\n<\/blockquote>\n\n<p>But i am following the (samples, channels, rows, cols) norm. Why is this happening ?<\/p>\n","answers":[{"AnswerId":"41570921","CreationDate":"2017-01-10T14:14:12.447","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>It seems like your input data has the wrong shape. You should print out the shape of the data that you are feeding into the network. <\/p>\n\n<p>It seems like your array are gray input images and they normally only use 2 dimensions because they only have 1 channel. Therefore the np array is ordered without the third dimension. Normally you have to add this by using np.reshape or allocating your array in another way. When I get an error message like yours i would try:<\/p>\n\n<pre><code>X # training data\nX = np.transpose(X, (0, 2, 1))\nX = np.reshape(X, (X.shape[0], 1, X.shape[1], X.shape[2]))\n<\/code><\/pre>\n"}]}
{"QuestionId":41570901,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><keras><keras-layer>","CreationDate":"2017-01-10T14:13:11.890","AcceptedAnswerId":41586658.0,"OwnerUserId":7017670.0,"Title":"Get Ouput Of A Keras Model\/Layer","Body":"<p>My Keras model is the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/babi_rnn.py\" rel=\"nofollow noreferrer\">babi_rnn example<\/a> in the Keras repository.<\/p>\n\n<p>I want to get the output (in words) of the model on a dataset.<\/p>\n\n<p>I tried:<\/p>\n\n<pre><\/pre>\n\n<p>but I get the error:<\/p>\n\n<pre><\/pre>\n\n<p><strong>How can I simply get the output of the model or layer upon being fed by an input?<\/strong><\/p>\n\n<p>That is, I need <\/p>\n\n<pre><\/pre>\n\n<p>So that I could do:  for seeing <strong>what is the model actually doing<\/strong>.<\/p>\n\n<p>I think this should have been the simplest use case, but it looks troublesome to implement.<\/p>\n\n<p>Any help would be greatly appreciated. Thanks.<\/p>\n","answers":[{"AnswerId":"41586658","CreationDate":"2017-01-11T09:08:02.603","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can simply use <code>model.predict<\/code> to get <code>Y<\/code> the <code>predict<\/code> function internally calls <code>_make_predict_function()<\/code> which does what you are trying to do.<\/p>\n\n<p>But your model is trained to map certain type of input to certain type of output... so you need to take care of those while using the <code>predict<\/code> function and also interpreting the same. In this example this conversion is done in <code>vectorize_stories()<\/code> so try and understand what it is doing.<\/p>\n\n<p>In this case to get the word predicted all you need to do after training the model is:<\/p>\n\n<pre><code>Y_pred = model.predict([tX, tXq])\nfor pred in Y_pred:\n    print (vocab[pred.argmax()-1])\n<\/code><\/pre>\n\n<p>Note again <code>tX<\/code> is the vectorized test story <code>tXq<\/code> is the vectorized test query and <code>Y_pred<\/code> is your vectorized predicted answer by the model. <\/p>\n"}]}
{"QuestionId":41572660,"AnswerCount":1,"Tags":"<android><cmake><tensorflow>","CreationDate":"2017-01-10T15:41:49.560","AcceptedAnswerId":null,"OwnerUserId":3836902.0,"Title":"Pass a compiler flag in Tensorflow android","Body":"<p>I am using Android Java Inference Interface for TensorFlow (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/f94e20f2c704f037a113d350fd7c431e3fd75df8\/tensorflow\/contrib\/android\" rel=\"nofollow noreferrer\">TensorFlow-Android<\/a> ). <\/p>\n\n<p>I have two questions.<\/p>\n\n<ol>\n<li>How can I pass compiler flag to build core tensorflow with selective registration? Should I put it into <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f94e20f2c704f037a113d350fd7c431e3fd75df8\/tensorflow\/contrib\/android\/cmake\/CMakeLists.txt\" rel=\"nofollow noreferrer\">CMakeLists.txt<\/a> ?<\/li>\n<li>How does  variable affect build? I mean initially, we build  static lib (by uncommenting build_all_android.sh in gralde) then we build .so shared lib using , and passing this flag to cmake seems redundant.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"42614875","CreationDate":"2017-03-05T22:32:26.880","ParentId":null,"OwnerUserId":"112705","Title":null,"Body":"<p>You can pass a <code>--copt<\/code> parameter to bazel (<a href=\"https:\/\/bazel.build\/versions\/master\/docs\/command-line-reference.html\" rel=\"nofollow noreferrer\">basic docs<\/a>):<\/p>\n\n<pre><code>--copt=&lt;a string&gt; multiple uses are accumulated\nAdditional options to pass to gcc.\n<\/code><\/pre>\n\n<p>The user manual has more extensive documentation on this option:<br>\n<a href=\"https:\/\/bazel.build\/versions\/master\/docs\/bazel-user-manual.html#flag--copt\" rel=\"nofollow noreferrer\">https:\/\/bazel.build\/versions\/master\/docs\/bazel-user-manual.html#flag--copt<\/a><\/p>\n"}]}
{"QuestionId":41574431,"AnswerCount":1,"Tags":"<scikit-learn><keras>","CreationDate":"2017-01-10T17:10:28.877","AcceptedAnswerId":41585626.0,"OwnerUserId":6828811.0,"Title":"How to use classification report from sklearn for keras models?","Body":"<p>So i have written a network which consists of the followings for multi-class classification :\n-y_labels transformed with to_categorical\n-last layer uses a sigmoid function with 3 neurons as my classes\n-model compile uses  categorical_crossentropy as loss function\nSo i used<\/p>\n\n<pre><\/pre>\n\n<p>and then i used it as  <\/p>\n\n<pre><\/pre>\n\n<p>y_test has the  form to_categorical\nAnd i am getting the following error :<\/p>\n\n<pre><\/pre>\n\n<p>My question is how can i transform it back in order to use it as such?<\/p>\n","answers":[{"AnswerId":"41585626","CreationDate":"2017-01-11T08:06:04.483","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>The error simply indicates that <code>y_test<\/code> and <code>pred<\/code> are of different types. Check function <code>type_of_target<\/code> in <a href=\"https:\/\/github.com\/scikit-learn\/scikit-learn\/blob\/master\/sklearn\/utils\/multiclass.py#L206-L229\" rel=\"noreferrer\">multiclass.py<\/a>. As indicated here one of the <code>y<\/code> is indicator of classes and another of class vector. You can infer which one is what just by printing the shape, <code>y_test.shape , pred.shape<\/code>. <\/p>\n\n<p>More over since you are using <code>model.predict_classes<\/code> instead of <code>model.predict<\/code> your output of <code>model.predict_classes<\/code> will be just classes and not class vector. <\/p>\n\n<p>So either you need to convert one of them by:<\/p>\n\n<pre><code># class --&gt; class vector\nfrom keras.utils import np_utils\nx_vec = np_utils.to_categorical(x, nb_classes)\n\n# class vector --&gt; class \nx = x_vec.argmax(axis=-1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41574802,"AnswerCount":1,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2017-01-10T17:30:07.683","AcceptedAnswerId":41579291.0,"OwnerUserId":2984051.0,"Title":"Re-training inception google cloud stuck at global step 0","Body":"<p>I am following the flowers tutorials for re-training inception on google cloud ml. I can run the tutorial, train, predict, just fine. <\/p>\n\n<p>I then substituted the flowers dataset for a test dataset of my own. Optical character recognition of image digits.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/f49ku.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/f49ku.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My full code is <a href=\"https:\/\/github.com\/bw4sz\/MeerkatReader\/blob\/master\/cloudML\/MeerkatReader.sh\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>Dict File for <a href=\"https:\/\/github.com\/bw4sz\/MeerkatReader\/blob\/master\/cloudML\/dict.txt\" rel=\"nofollow noreferrer\">labels<\/a><\/p>\n\n<p>Eval <a href=\"https:\/\/github.com\/bw4sz\/MeerkatReader\/blob\/master\/cloudML\/testing_data.csv\" rel=\"nofollow noreferrer\">set<\/a><\/p>\n\n<p>Training <a href=\"https:\/\/github.com\/bw4sz\/MeerkatReader\/blob\/master\/cloudML\/training_data.csv\" rel=\"nofollow noreferrer\">Set<\/a><\/p>\n\n<p>Running from recent docker build provided by google.<\/p>\n\n<pre><\/pre>\n\n<p>I can preprocess files, and submit the training job using<\/p>\n\n<pre><\/pre>\n\n<p>but it never makes it past global step 0. The flowers tutorial trained in about ~ 1 hr on the free tier. I have let my training go as long as 11 hrs. No movement.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jyjUM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jyjUM.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Looking over at stackdriver, nothing progresses.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/r4bTl.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/r4bTl.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have also tried a tiny toy dataset of 20 training images, and 10 eval images. Same issue.<\/p>\n\n<p>The GCS Bucket ends up looking like this\n<a href=\"https:\/\/i.stack.imgur.com\/w0SQe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/w0SQe.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Perhaps unsurprisingly, I can't visualize this log in tensorboard, nothing to show.<\/p>\n\n<p>Full training log:<\/p>\n\n<pre><\/pre>\n\n<p>Just repeating the last line until I kill it.<\/p>\n\n<p>Is my mental model for this service incorrect? All suggestions welcome.<\/p>\n","answers":[{"AnswerId":"41579291","CreationDate":"2017-01-10T22:07:00.487","ParentId":null,"OwnerUserId":"4392784","Title":null,"Body":"<p>Everything looks fine. My suspicion is that there is a problem with your data. Specifically I suspect TF is unable to read any data from your GCS files (are they empty?)? As a result when you invoke train, TF ends up blocking trying to read a batch of data which it can't do.<\/p>\n\n<p>I would suggest adding logging statements around the call to session.run in <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/blob\/master\/flowers\/trainer\/task.py#L226\" rel=\"nofollow noreferrer\">Trainer.run_training<\/a>. This will tell you whether that is the line where it is getting stuck.<\/p>\n\n<p>I'd also suggest checking the sizes of your GCS files.<\/p>\n\n<p>TensorFlow also has an experimental <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a304537954a865752ad1b18461e6bd67b36082db\/tensorflow\/core\/protobuf\/config.proto#L251\" rel=\"nofollow noreferrer\">RunOptions<\/a> which allows you to specify a timeout for Session.run. Once this feature is ready, this might be useful for ensuring code doesn't block forever.<\/p>\n"}]}
{"QuestionId":41575442,"AnswerCount":1,"Tags":"<json><xml><tensorflow><export>","CreationDate":"2017-01-10T18:04:05.933","AcceptedAnswerId":41576075.0,"OwnerUserId":3002273.0,"Title":"TensorFlow export compute graph to XML, JSON, etc","Body":"<p>I want to export a TensorFlow compute graph to XML or something similar so I can modify it with an external program and then re-import it. I found <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/meta_graph\/\" rel=\"nofollow noreferrer\">Meta Graph<\/a> but this exports in a binary format which I wouldn't know how to modify.<\/p>\n\n<p>Does such capability exist?<\/p>\n","answers":[{"AnswerId":"41576075","CreationDate":"2017-01-10T18:42:52.833","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The native serialization format for TensorFlow's dataflow graph uses <a href=\"https:\/\/developers.google.com\/protocol-buffers\/\" rel=\"noreferrer\">protocol buffers<\/a>, which have bindings in many different languages. You can generate code that should be able to parse the binary data from the two message schemas: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f94e20f2c704f037a113d350fd7c431e3fd75df8\/tensorflow\/core\/framework\/graph.proto\" rel=\"noreferrer\"><code>tensorflow.GraphDef<\/code><\/a> (a lower-level representation) and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f94e20f2c704f037a113d350fd7c431e3fd75df8\/tensorflow\/core\/protobuf\/meta_graph.proto\" rel=\"noreferrer\"><code>tensorflow.MetaGraphDef<\/code><\/a> (a higher-level representation, which includes a <code>GraphDef<\/code> and other information about how to interpret some of the nodes in the graph).<\/p>\n\n<p>If there is no protocol buffer implementation for your target language, you can generate JSON from the Python protocol buffer object. For example, the following generates a string containing a JSON representation of a <code>GraphDef<\/code>:<\/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"}]}
{"QuestionId":41575654,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-10T18:17:33.540","AcceptedAnswerId":null,"OwnerUserId":7375883.0,"Title":"Why the labels from tensorflow's shuffle do not work?","Body":"<p>every one!\nFrom shuffle I got images and labels, images with conv works well.<\/p>\n\n<p>images and labels from tfreocords<\/p>\n\n<pre><\/pre>\n\n<p>BUT  does not work as below:<\/p>\n\n<pre><\/pre>\n\n<p>The result is always wrong, because it cannot get values from  at all.<\/p>\n\n<p>Some details are below:<\/p>\n\n<pre><\/pre>\n\n<p>......some training model<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41580045","CreationDate":"2017-01-10T23:07:15.717","ParentId":null,"OwnerUserId":"7375883","Title":null,"Body":"<p>The reason is that you have to use tensoflow mothods not numpy, below works well by accaray.<\/p>\n\n<pre><code>def correct_rate(out, labels):\n  arg = tf.argmax(out, 1)\n  arg = tf.cast(arg, tf.int32)\n  correct_prediction = tf.equal(labels, arg)\n  accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n  return accuracy\n\naccr = correct_rate(logits, train_labels)\nprint(sess.run(accr))\n<\/code><\/pre>\n"}]}
{"QuestionId":41576944,"AnswerCount":3,"Tags":"<image><python-3.x><tensorflow>","CreationDate":"2017-01-10T19:36:01.157","AcceptedAnswerId":41703607.0,"OwnerUserId":78187.0,"Title":"Dynamic image cropping in Tensorflow","Body":"<p>I'm trying to figure out how to take a crop of an image determined dynamically in Tensorflow.  Below is an example of what I am trying to accomplish, however I can't seem to make it work.  Essentially, I want to feed images and the crop values for that image within the graph, and then continue on with other computations on those cropped pieces.  My current attempt:<\/p>\n\n<pre><\/pre>\n\n<pre>\n\/Users\/p111\/anaconda\/bin\/python \/Users\/p111\/PycharmProjects\/analysis_code\/testing.py\nTraceback (most recent call last):\n  File \"\/Users\/p111\/PycharmProjects\/analysis_code\/testing.py\", line 31, in \n    cropped_image = tf.map_fn(lambda img, crop: crop_image(img, crop), elems=[image_placeholder, crop_placeholder])\n  File \"\/Users\/p111\/anaconda\/lib\/python3.5\/site-packages\/tensorflow\/python\/ops\/functional_ops.py\", line 390, in map_fn\n    swap_memory=swap_memory)\n  File \"\/Users\/p111\/anaconda\/lib\/python3.5\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2636, in while_loop\n    result = context.BuildLoop(cond, body, loop_vars, shape_invariants)\n  File \"\/Users\/p111\/anaconda\/lib\/python3.5\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2469, in BuildLoop\n    pred, body, original_loop_vars, loop_vars, shape_invariants)\n  File \"\/Users\/p111\/anaconda\/lib\/python3.5\/site-packages\/tensorflow\/python\/ops\/control_flow_ops.py\", line 2419, in _BuildLoop\n    body_result = body(*packed_vars_for_body)\n  File \"\/Users\/p111\/anaconda\/lib\/python3.5\/site-packages\/tensorflow\/python\/ops\/functional_ops.py\", line 380, in compute\n    packed_fn_values = fn(packed_values)\nTypeError: () missing 1 required positional argument: 'crop'\n<\/pre>\n\n<p>Edit:  It appears that elems will only accept a single tensor.  Which means I would need to somehow combine my two tensors into one, and then unpack it in my function to get the values out.  I'm not sure how I would perform that kind of tensor manipulation.  I have found the glimpse method already and that does work, however I am wondering if the same can be done with this specific method.  Mostly, I'm wondering how you would combine and then split a pair of tensors so it can be used in this method.<\/p>\n","answers":[{"AnswerId":"41577956","CreationDate":"2017-01-10T20:39:29.303","ParentId":null,"OwnerUserId":"302268","Title":null,"Body":"<p>tf.map_fn(f, l) runs function f for every tensor in list l. In your case, your function expects 2 arguments, but since you supply a flat list, map_fn() sends them one by one. According to docs, map_fn() supports variable arity, so what you should do is something like this<\/p>\n\n<pre>\ntf.map_fn(lambda img, crop: crop_image(img, crop), \n          elems=([image_placeholder, crop_placeholder], ))\n<\/pre>\n\n<p>so the list you pass to map_fn contains pairs of arguments.<\/p>\n"},{"AnswerId":"41703607","CreationDate":"2017-01-17T17:39:58.840","ParentId":null,"OwnerUserId":"4834515","Title":null,"Body":"<p>I saw this code from <a href=\"https:\/\/stackoverflow.com\/questions\/37086098\/does-tensorflow-map-fn-support-taking-more-than-one-tensor\">here<\/a>.<\/p>\n\n<pre><code>elems = (np.array([1, 2, 3]), np.array([-1, 1, -1]))\nalternate = map_fn(lambda x: x[0] * x[1], elems, dtype=tf.int64)\n# alternate == [-1, 2, -3]\n<\/code><\/pre>\n\n<p>It is possible to use a tuple or a list to pack several elements into one so I tried this.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom matplotlib import pyplot as plt\nimport numpy as np\n\nsess = tf.InteractiveSession()\n\nimg1 = np.random.random([400, 600, 3])\nimg2 = np.random.random([400, 600, 3])\nimg3 = np.random.random([400, 600, 3])\nimages = np.array([img1, img2, img3])\n# images = tf.convert_to_tensor(images)  # it can be uncommented.\n\nimg1_crop = [100, 100, 100, 100]\nimg2_crop = [200, 150, 100, 100]\nimg3_crop = [150, 200, 100, 100]\ncrop_values = np.array([img1_crop, img2_crop, img3_crop])\n# crop_values = tf.convert_to_tensor(crop_values)  # it can be uncommented.\n\ndef crop_image(img, crop):\n    return tf.image.crop_to_bounding_box(img,\n                                         crop[0],\n                                         crop[1],\n                                         crop[2],\n                                         crop[3])\n\nfn = lambda x: crop_image(x[0], x[1])\nelems = (images, crop_values)\n\ncropped_image = tf.map_fn(fn, elems=elems, dtype=tf.float64)\nresult = sess.run(cropped_image)\n\nprint result.shape\n\nplt.imshow(result[0])\nplt.show()\n<\/code><\/pre>\n\n<p>It works on my machine with tf version 0.11 and python2. Hope this can help you.<\/p>\n"},{"AnswerId":"41740138","CreationDate":"2017-01-19T10:59:52.623","ParentId":null,"OwnerUserId":"2911687","Title":null,"Body":"<p>Couple of things : <\/p>\n\n<ul>\n<li>You do not have a <code>return<\/code> statement in the <code>crop_image<\/code> function.<\/li>\n<li><code>map_fn<\/code> accepts a single argument.<\/li>\n<li>I strongly advise you to separate the graph definition and the session usage.<\/li>\n<\/ul>\n\n<p>--<\/p>\n\n<pre><code># Graph def\ndef crop_image(img, crop):\n    return tf.image.crop_to_bounding_box(img,\n                                  crop[0],\n                                  crop[1],\n                                  crop[2],\n                                  crop[3])\n\nimage_placeholder = tf.placeholder(tf.float32, [None, 400, 600, 3])\ncrop_placeholder = tf.placeholder(dtype=tf.int32)\ncropped_image = tf.map_fn(lambda inputs: crop_image(*inputs), elems=[image_placeholder, crop_placeholder], dtype=tf.float32)\n\n\n# Session\nsess = tf.InteractiveSession()\n\nimg1 = np.random.random([400, 600, 3])\nimg2 = np.random.random([400, 600, 3])\nimg3 = np.random.random([400, 600, 3])\n\nimages = [img1, img2, img3]\n\nimg1_crop = [100, 100, 100, 100]\nimg2_crop = [200, 150, 100, 100]\nimg3_crop = [150, 200, 100, 100]\n\ncrop_values = [img1_crop, img2_crop, img3_crop]\n\nsess.run(tf.global_variables_initializer())\n\nresult = sess.run(cropped_image, feed_dict={image_placeholder: images, crop_placeholder:crop_values})\n\nplt.imshow(result[0])\nplt.show()\n<\/code><\/pre>\n"}]}
{"QuestionId":41577384,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-10T20:03:11.673","AcceptedAnswerId":41577625.0,"OwnerUserId":2122597.0,"Title":"variable scope issue in Tensorflow","Body":"<pre><\/pre>\n\n<p>In my code I am calling this function twice to create 2 bidirectional LSTMs. Then I got the problem of reusing variables.<\/p>\n\n<blockquote>\n  <p>ValueError: Variable lstm\/BiRNN_FW\/BasicLSTMCell\/Linear\/Matrix\n  already exists, disallowed. Did you mean to set reuse=True in\n  VarScope?<\/p>\n<\/blockquote>\n\n<p>To resolve this I added the LSTM definition in the function within <\/p>\n\n<p>This led to a new issue <\/p>\n\n<blockquote>\n  <p>ValueError: Variable lstm\/BiRNN_FW\/BasicLSTMCell\/Linear\/Matrix does\n  not exist, disallowed. Did you mean to set reuse=None in VarScope?<\/p>\n<\/blockquote>\n\n<p>Please help with a solution to this.<\/p>\n","answers":[{"AnswerId":"50452620","CreationDate":"2018-05-21T16:04:13.000","ParentId":null,"OwnerUserId":"6835092","Title":null,"Body":"<p>I also has the similar problem. However I was using keras implementation with pretrained Resnet50 model.<\/p>\n\n<p>It worked for me when I updated the tensorflow version using following command:<\/p>\n\n<pre><code>conda update -f -c conda-forge tensorflow\n<\/code><\/pre>\n\n<p>and used <\/p>\n\n<pre><code>from keras import backend as K\nK.clear_session\n<\/code><\/pre>\n"},{"AnswerId":"41577625","CreationDate":"2017-01-10T20:18:41.350","ParentId":null,"OwnerUserId":"302268","Title":null,"Body":"<p>When you create BasicLSTMCell(), it creates all the required weights and biases to implement an LSTM cell under the hood. All of these variables are assigned names automatically. If you call the function more than once within the same scope you get the error you get. Since your question seems to state that you want to create two separate LSTM cells, you do not want to reuse the variables, but you do want to create them in separate scopes. You can do this in two different ways (I haven't actually tried to run this code, but it should work). You can call your function from within a unique scope<\/p>\n\n<pre>\ndef biLSTM(data, n_steps):\n    ... blah ...\n\nwith tf.variable_scope('LSTM1'):\n    outputs, hidden = biLSTM(data, steps)\n\nwith tf.variable_scope('LSTM2'):\n    outputs, hidden = biLSTM(data, steps)\n<\/pre>\n\n<p>or you can pass a unique scope name to the function and use the scope inside<\/p>\n\n<pre>\ndef biLSTM(data, n_steps, layer_name):\n    ... blah...\n    with tf.variable_scope(layer_name) as scope:\n        lstm_fw_cell = tf.nn.rnn_cell.BasicLSTMCell(n_hidden, forget_bias=1.0)\n        lstm_bw_cell = tf.nn.rnn_cell.BasicLSTMCell(n_hidden, forget_bias=1.0)\n        outputs, _, _ = tf.nn.bidirectional_rnn(lstm_fw_cell, lstm_bw_cell, data, dtype=tf.float32)\n    return outputs, n_hidden\n\nl1 = biLSTM(data, steps, 'layer1')\nl2 = biLSTM(data, steps, 'layer2')\n<\/pre>\n\n<p>It is up to your coding sensibilities which approach to choose, they are functionally pretty much the same.<\/p>\n"}]}
{"QuestionId":41577654,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-10T20:20:16.797","AcceptedAnswerId":41578733.0,"OwnerUserId":2542516.0,"Title":"tf.Session() seemingly not closed even after exiting the with block","Body":"<p>I am saving two graphs; one with a 2X2 zero tensor and the other with a ones tensor of the same size. I restore one of them depending on a condition.<\/p>\n\n<pre><\/pre>\n\n<p>Both the calls to  return zeros. Observing the ops in the session suggests that the session in  is being reused across the two calls which AFAIK is not supposed to happen as the session is closed when  returns:<\/p>\n\n<pre><\/pre>\n\n<p>Is this a bug or am I missing something?<\/p>\n","answers":[{"AnswerId":"41578733","CreationDate":"2017-01-10T21:27:11.820","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> The <code>tf.Session<\/code> is closed between the two calls to <code>test()<\/code> in your code, but you are running into a problem because the two sessions are sharing the same <code>tf.Graph<\/code>. Create each session with a fresh <code>tf.Graph<\/code> to avoid the problem.<\/p>\n\n<p>In particular, the nodes created when you call <code>tf.train.import_meta_graph()<\/code> in the call to <code>test(False)<\/code> remain in the graph when subsequently call <code>tf.train.import_meta_graph()<\/code> in the call to <code>test(True)<\/code>. This means that each of the two calls to <code>session.graph.get_tensor_by_name('dummy_graph\/a:0')<\/code> will return the same node (which was created when you first called <code>test()<\/code>).<\/p>\n\n<p>There are a few ways to avoid this problem. The simplest is to create the <code>tf.Session<\/code> with its own graph:<\/p>\n\n<pre><code>def test(boolean):\n    # Session will use its own graph.\n    with tf.Session(graph=tf.Graph()) as session:\n        if boolean:\n            # ...\n<\/code><\/pre>\n"},{"AnswerId":"41579051","CreationDate":"2017-01-10T21:49:40.680","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>To add more details to mrry answer, here's exactly what's happening:<\/p>\n\n<p>During checkpoint saving, you are creating new graph each time, so you are saving in both checkpoints a tensor <code>dummy_graph\/a<\/code>:<\/p>\n\n<pre><code>zero.data-00000-of-00001: dummy_graph\/a - [0, 0, \none.data-00000-of-00001: dummy_graph\/a - [1, 1, \n<\/code><\/pre>\n\n<p>During first load call you first create a graph with variable <code>dummy_graph\/a<\/code>, load <code>[0, 0,<\/code> into it, then call init op which will overwrite this value with <code>[0, 0,<\/code><\/p>\n\n<p>During second load call, your <code>import_meta_graph<\/code> will append to existing default graph. Since there's a name conflict, it will append <code>_1<\/code> to nodes, so now your graph will have nodes <code>dummy_graph\/a<\/code> and <code>dummy_graph\/a_1<\/code> and corresponding init nodes <code>init<\/code> and <code>init_1<\/code><\/p>\n\n<p>During second restore, your saver will restore your <code>[1, 1, ...<\/code> checkpoint into <code>dummy_graph\/a<\/code>. Then you call <code>init<\/code> which will overwrite the value of <code>dummy_graph\/a<\/code> with <code>[0, 0, ...<\/code>. You then return the value of <code>dummy_graph\/a<\/code><\/p>\n\n<p>Note that after second restore, your session has two variables, and the second one is not initialized. Curiously,  <code>tf.report_uninitialized_variables()<\/code> does not show it, even though <code>sess.run('dummy_graph\/a_1:0')<\/code> will throw <code>Attempting to use uninitialized value<\/code> error, that seems like a bug.<\/p>\n"}]}
{"QuestionId":41578532,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-01-10T21:14:11.587","AcceptedAnswerId":41580924.0,"OwnerUserId":1714410.0,"Title":"What is \"Parameter\" layer in caffe?","Body":"<p>Recently I came across <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/layers\/parameter_layer.hpp\" rel=\"nofollow noreferrer\"><\/a> layer in caffe.<br>\nIt seems like this layer exposes its internal parameter blob to \"top\".  <\/p>\n\n<p>What is this layer using for?<br>\nCan you give a usage example?<\/p>\n","answers":[{"AnswerId":"41580924","CreationDate":"2017-01-11T00:38:13.400","ParentId":null,"OwnerUserId":"4221706","Title":null,"Body":"<p>This layer was introduced in the pull request <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2079\" rel=\"nofollow noreferrer\">#2079<\/a>, with the following description:<\/p>\n\n<blockquote>\n  <p>This layer simply holds a parameter blob of user-defined shape, and shares it as its single top. <\/p>\n<\/blockquote>\n\n<p>which is exactly what you expected. This was introduced in context of the issue <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1474\" rel=\"nofollow noreferrer\">#1474<\/a>, which basically proposes to treat parameters like normal bottom blobs. To show why this can be useful, consider the following example (taken from issue <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1474\" rel=\"nofollow noreferrer\">#1474<\/a>, by <a href=\"https:\/\/github.com\/longjon\" rel=\"nofollow noreferrer\">@longjon<\/a>):<\/p>\n\n<p>The inner product layer calculates <code>C = A * B<\/code>, where <code>C<\/code> is the top blob (output), <code>B<\/code> is a bottom blob, and <code>A<\/code> <em>must<\/em> be a parameter blob. This is very restrictive, as this makes it impossible to use the inner product layer for multiplying the inner product between two bottom blobs, e.g. multiplying two input matrices. The issue <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1474\" rel=\"nofollow noreferrer\">#1474<\/a> suggests to make a fundamental change: make parameters independent of the layer. Instead, treat them like a normal bottom blob.<\/p>\n\n<p>As a first step in that direction, the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/layers\/parameter_layer.hpp\" rel=\"nofollow noreferrer\"><code>Parameter<\/code> layer<\/a> was introduced. This allows you to define a new parameter, which you can then feed into the bottom of another layer. \nThe counterpart - a method for feeding parameters into a layer as a bottom blob - is proposed in pull request <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2166\" rel=\"nofollow noreferrer\">#2166<\/a>, which <strong>isn't merged<\/strong> yet, as of Jan. 2017.\nThough this is not merged yet, you can still use <code>Parameter<\/code> layer to define new learnable parameters to feed into other layers as bottom blob.<\/p>\n"}]}
{"QuestionId":41579250,"AnswerCount":0,"Tags":"<docker><path><directory><tensorflow>","CreationDate":"2017-01-10T22:03:46.593","AcceptedAnswerId":null,"OwnerUserId":5007558.0,"Title":"Where can i find the tensorflow folder which i pulled from DockerHub? (on windows7)","Body":"<p>I wanted to pull the latest updations of the tensorflow repo on github, but in order to do that i first have to find the image i pulled from DockerHub on my drive.<\/p>\n\n<p>Can anyone tell where to find the tensorflow directory??<\/p>\n\n<p>I have tried searching everywhere but really can't find it. I really need it for my practical implementation of small machine learning projects i've been thinking. Any help will be appreciated.!<\/p>\n","answers":[]}
{"QuestionId":41581402,"AnswerCount":0,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2017-01-11T01:40:16.660","AcceptedAnswerId":null,"OwnerUserId":7402153.0,"Title":"SELECTIVELY Restore Weights in TensorFlow","Body":"<p>I have two neural network (TF Computation Graph) defined like following:<\/p>\n\n<ol>\n<li>Model 1: some input&amp;hidden layers + a 3-way softmax layer<\/li>\n<li>Model 2: some input&amp;hidden layers(same as Model 1) + a binary Sigmoid layer<\/li>\n<\/ol>\n\n<p>The only difference between Model 1 and 2 is the output layer.<\/p>\n\n<p>I have stored some trained parameters for Model 1 and would like to use them (the input &amp; hidden layers) to initialize new Model 2. Ideally, for Model 2, the input &amp; hidden layers should use the saved weights while the output layer will be randomly initialized. I am wondering if it is possible to implement this partial weights recovery in TensorFlow. What is the best practice to do this?<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[]}
{"QuestionId":41583151,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-11T05:12:41.430","AcceptedAnswerId":null,"OwnerUserId":7021851.0,"Title":"tensorflow slim and half precision","Body":"<p>Is there a way to define  in slim library? In general, lets say I want to train with  i.e. want all tensors to be of half precision. Is there a way to specify this to ?<\/p>\n","answers":[{"AnswerId":"41599279","CreationDate":"2017-01-11T19:34:37.280","ParentId":null,"OwnerUserId":"805030","Title":null,"Body":"<p>For half precision in tensorflow use: <code>tf.float16<\/code><br>\nExample:<br><\/p>\n\n<pre>\nimport tensorflow as tf\nimport tensorflow.contrib.slim as slim\nweights = slim.variable('weights', shape=[10, 1], dtype=tf.float16)\n<\/pre>\n"}]}
{"QuestionId":41583540,"AnswerCount":2,"Tags":"<tensorflow><neural-network><dropout>","CreationDate":"2017-01-11T05:47:21.513","AcceptedAnswerId":41584818.0,"OwnerUserId":697065.0,"Title":"Custom dropout in tensorflow","Body":"<p>I'm training a DNN model on some data, and am hoping to analyze the learned weights to learn something about the true system I am studying (signaling cascades in biology). I guess one could say I am using Artificial NNs to learn about Biological NNs.<\/p>\n\n<p>For each of my training examples, I have removed a single gene, that is responsible for signaling at the top layer. <\/p>\n\n<p>As I am modeling this signaling cascade as a NN, and removing one of the nodes in the first hidden layer, I realized that I'm doing a real life version of dropout. <\/p>\n\n<p>I would therefore like to use dropout to train my model, however the implementations of dropout that I have seen online seem to randomly drop out a node. What I need is a way to specify which node to dropout for each training example.<\/p>\n\n<p>Any advice on how to implement this? I'm open to any package, but right now everything i have already done is in Tensorflow so I'd appreciate a solution that uses that framework. <\/p>\n\n<p>For those that prefer the details explained:<\/p>\n\n<p>I have 10 input variables, that are fully connected to 32 relu nodes in the first layer, which are fully connected to a second layer (relu), which is fully connected to the output (linear because I am doing regression). <\/p>\n\n<p>In addition to the 10 input variables, I also happen to know which of the 28 nodes should be dropped out. <\/p>\n\n<p>Is there a way I can specify this when training?<\/p>\n\n<p>Here is the code I currently use:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41600254","CreationDate":"2017-01-11T20:37:28.360","ParentId":null,"OwnerUserId":"697065","Title":null,"Body":"<p>For completeness, here is my final implementation:<\/p>\n\n<pre><code>import numpy as np\nimport pandas as pd\nimport tflearn\nimport tensorflow as tf\n\nmeta = pd.read_csv('..\/..\/input\/nn\/meta.csv')\nexperiments = meta[\"Unnamed: 0\"]\ndel meta[\"Unnamed: 0\"]\n\nstress_one_hot = pd.get_dummies(meta[\"train\"])\n\nkinase_deletion = pd.get_dummies(meta[\"Strain\"])\nkinase_one_hot = 1 - kinase_deletion\n\nexpression = pd.read_csv('..\/..\/input\/nn\/data.csv')\ngenes = expression[\"Unnamed: 0\"]\ndel expression[\"Unnamed: 0\"] # This holds the gene names just so you know...\n\nexpression = expression.transpose()\n\n# Set up data for tensorflow\n# Gene expression\ntarget = expression\ntarget = np.array(expression, dtype='float32')\ntarget_mean = target.mean(axis=0, keepdims=True) \ntarget_std = target.std(axis=0, keepdims=True)\ntarget = target - target_mean\ntarget = target \/ target_std\n\n# Stress information\ndata1 = stress_one_hot\ndata1 = np.array(data1, dtype='float32')\ndata_mean = data1.mean(axis=0, keepdims=True)\ndata_std = data1.std(axis=0, keepdims=True)\ndata1 = data1 - data_mean\ndata1 = data1 \/ data_std\n\n# Kinase information\ndata2 = kinase_one_hot\ndata2 = np.array(data2, dtype='float32')\n\n# For Reference\n# data1.shape\n# #(301, 10)\n# data2.shape\n# #(301, 29)\n\n\n# Build the Neural Network\n\nnum_stresses = 10\nnum_kinase = 29\nnum_transcription_factors = 200\nnum_genes = 6692\n\n# Build neural network\n# Input variables (10)\n# Which Node to dropout (32)\nstress = tflearn.input_data(shape=[None, num_stresses])\nkinase_deletion = tflearn.input_data(shape=[None, num_kinase])\n\n# This is the layer that I want to perform selective dropout on,\n# I should be able to specify which of the 32 nodes should output zero\n# based on a 1X32 vector of ones and zeros.\nkinase = tflearn.fully_connected(stress, num_kinase, activation='relu')\nkinase_dropout = tf.mul(kinase, kinase_deletion)\n\ntranscription_factor = tflearn.fully_connected(kinase_dropout, num_transcription_factors, activation='relu')\n\ngene = tflearn.fully_connected(transcription_factor, num_genes, activation='linear')\n\nadam = tflearn.Adam(learning_rate=0.00001, beta1=0.99)\n\nregression = tflearn.regression(gene, optimizer=adam, loss='mean_square', metric='R2')\n\n# Define model\nmodel = tflearn.DNN(regression, tensorboard_verbose=1)\n\n# Start training (apply gradient descent algorithm)\nmodel.fit([data1, data2], target, n_epoch=20000, show_metric=True, shuffle=True)#,validation_set=0.05)\n<\/code><\/pre>\n"},{"AnswerId":"41584818","CreationDate":"2017-01-11T07:15:33.127","ParentId":null,"OwnerUserId":"3893224","Title":null,"Body":"<p>I would supply your input variables and an equal sized vector of all 1's except for the one you want to drop, that one is a 0.<\/p>\n\n<p>The very first operation should then be multiplication to zero out the gene you want to drop.  From there on out, it should be the exact same as what you have now.<\/p>\n\n<p>You can either multiply (zero out your gene) before handing it to tensorflow or add another place holder and feed it into the graph  in the feed_dict like you do your variables.  The latter one would probably be better.<\/p>\n\n<p>If you need to drop a hidden node (in layer 2), it's just another vector of 1s and a 0.<\/p>\n\n<p>Let me know if that works or if you need more help.<\/p>\n\n<hr>\n\n<p>Edit: \nOk, so I haven't really worked with tflearn very much (I just did regular tensorflow), but I think you can combine tensorflow and tflearn.  Basically, I added <code>tf.multiply<\/code>.  You might have to add in another <code>tflearn.input_data(shape =[num_stresses])<\/code> and <code>tflearn.input_data(shape =[num_kinase])<\/code> to give you placeholders for the <code>stresses_dropout_vector<\/code> and <code>kinase_dropout_vector<\/code>.  And of course, you can change the number and positions of zeros in those two vectors.<\/p>\n\n<pre><code>import tensorflow as tf ###### New ######\nimport tflearn\n\nnum_stresses = 10\nnum_kinase = 32\nnum_transcription_factors = 200\nnum_genes = 6692\n\nstresses_dropout_vector = [1] * num_stresses ###### NEW ######\nstresses_dropout_vector[desired_node_to_drop] = 0 ###### NEW ######\n\nkinase_dropout_vector = [1] * num_kinase ###### NEW ######\nkinase_dropout_vector[desired_hidden_node_to_drop] = 0 ###### NEW ######\n\n# Build neural network\n# Input variables (10)\n# Which Node to dropout (32)\nstress = tflearn.input_data(shape=[None, num_stresses])\nkinase_deletion = tflearn.input_data(shape=[None, num_kinase])\n\n# This is the layer that I want to perform selective dropout on,\n# I should be able to specify which of the 32 nodes should output zero\n# based on a 1X32 vector of ones and zeros. \n\nstress_dropout = tf.multiply(stress, stresses_dropout_vector) ###### NEW ###### Drops out an input\nkinase = tflearn.fully_connected(stress_dropout, num_kinase, activation='relu') ### changed stress to stress_dropout\nkinase_dropout = tf.multiply(kinase, kinase_dropout_vector) ###### NEW ###### Drops out a hidden node\n\ntranscription_factor = tflearn.fully_connected(kinase_dropout, num_transcription_factors, activation='relu') ### changed kinase to kinase_dropout\n\ngene = tflearn.fully_connected(transcription_factor, num_genes, activation='linear')\n\nadam = tflearn.Adam(learning_rate=0.00001, beta1=0.99)\n\nregression = tflearn.regression(gene, optimizer=adam, loss='mean_square', metric='R2')\n\n# Define model\nmodel = tflearn.DNN(regression, tensorboard_verbose=1)\n<\/code><\/pre>\n\n<p>If adding in tensorflow doesn't work, you just have to find a regular old tflearn.multiply function that does an element wise multiplication of two given tensors\/vectors.<\/p>\n\n<p>Hope that helps.<\/p>\n"}]}
{"QuestionId":41583628,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><theano><theano.scan>","CreationDate":"2017-01-11T05:53:09.190","AcceptedAnswerId":41586573.0,"OwnerUserId":3101193.0,"Title":"efficient kernel implementation in theano","Body":"<p>I have just implemented a Gaussian kernel in Theano. However when I tested it as part of a neural network, it takes too long. It seems that the kernel subtractions are not paralellized. The whole training of the network uses a single processing core. So, how to correctly induce Theano to split the kernel operation? <\/p>\n\n<pre><\/pre>\n\n<p>Thak you very much.<\/p>\n","answers":[{"AnswerId":"41586573","CreationDate":"2017-01-11T09:02:54.743","ParentId":null,"OwnerUserId":"3737792","Title":null,"Body":"<p>Removing the loops will allow Theano to optimize the parallelization.<\/p>\n\n<p>First you can avoid the inner loop by doing:<\/p>\n\n<pre><code>for i in range(batch_s):\n    T.exp(-(W.T - X[i]).norm(2,axis=1) ** 2) \/ 2 * S ** 2)\n<\/code><\/pre>\n\n<p>Then you can use map on the outer loop:<\/p>\n\n<pre><code>import theano.tensor as T\nimport numpy\nimport theano\nimport timeit\n\nstart = timeit.default_timer()\nbatch_s=5\ndims=10\nhidd_s=3\nout_s=2\n\nmissing_param = None #\"ignore\"\n\nrng = numpy.random.RandomState(1234)\ninput = T.matrix(\"input\")\nX = numpy.asarray(rng.uniform(low=-2.1, high=5.0, size=(batch_s, dims)))\n\n\n\nW=theano.shared(\n        value=numpy.asarray(\n            rng.uniform(low=0.001, high=1.0, size=(dims, hidd_s)),\n                dtype=theano.config.floatX),\n        name='W', borrow=True)\n\nS=theano.shared(\n        value=numpy.asarray(\n            rng.uniform(low=10.0, high=100.0, size=(hidd_s, )),\n                dtype=theano.config.floatX),\n        name='S', borrow=True)\n\n\nf_func,f_updates = theano.map(lambda i : T.exp(-(W.T - i).norm(2,axis=1) ** 2) \/ 2 * S ** 2,input,[])\n\n\nlayer_out = theano.function([input],                                                        \n                          f_func,\n                          updates=f_updates,\n              on_unused_input=missing_param,\n                          allow_input_downcast=True)\n\n\nprint layer_out(X.astype('float32'))\n\nstop = timeit.default_timer()\n\nprint \"running time: \" + str(stop - start) \n<\/code><\/pre>\n\n<p>The output for the original code is:<\/p>\n\n<pre><code>[[  1.83701953e-25   1.78982216e-26   9.22911484e-27]\n [  1.60078639e-17   9.21553384e-17   7.62476155e-14]\n [  8.13404350e-17   1.88481821e-17   2.44677516e-15]\n [  3.16093011e-29   1.49698827e-27   2.42876079e-27]\n [  9.57103818e-09   3.46683533e-12   6.66103154e-12]]\nrunning time: 1.30477905273\n<\/code><\/pre>\n\n<p>With the new one:<\/p>\n\n<pre><code>[[  1.83701953e-25   1.78982216e-26   9.22911484e-27]\n [  1.60078639e-17   9.21553384e-17   7.62476155e-14]\n [  8.13404350e-17   1.88481821e-17   2.44677516e-15]\n [  3.16093011e-29   1.49698827e-27   2.42876079e-27]\n [  9.57103818e-09   3.46683533e-12   6.66103154e-12]]\nrunning time: 0.589275121689\n<\/code><\/pre>\n"}]}
{"QuestionId":41584312,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-11T06:41:54.960","AcceptedAnswerId":null,"OwnerUserId":582967.0,"Title":"TensorFlow: Specify number of Ops run in parallel","Body":"<p>As I understand, TF invokes multiple operators in parallel as long as they are independent. (<a href=\"https:\/\/stackoverflow.com\/q\/40008117\/582967\">link<\/a>)<\/p>\n\n<p>And the parallelism can be controlled by  and  if operators ar e running on CPU (<a href=\"https:\/\/stackoverflow.com\/q\/41233635\/582967\">link<\/a>). However these parameters does not affect GPU operators at all. How can I control parallelism of GPU? (for example, run operators serially although there are independent operators)<\/p>\n\n<p>EDIT:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41595343","CreationDate":"2017-01-11T15:59:04.747","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's a way to profile execution that avoids common pitfalls:<\/p>\n\n<pre><code># Turn off graph-rewriting optimizations\nconfig = tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))\n\n# throw error if explicit device placement can't be satisfied\nconfig.allow_soft_placement = False\n\nN = 8192\nwith tf.device(\"\/gpu:0\"):\n    input1 = tf.Variable(tf.random_normal([N,N]))\n    input2 = tf.Variable(tf.random_normal([N,N]))\n    result = tf.matmul(input1, input2)\n    result_no_output = result.op # to avoid transferring data back to Python\nsess = tf.Session(config=config)\n\n# load values into GPU\nsess.run(tf.global_variables_initializer())\n\n# pre-warming\nsess.run(result_no_output)\n\nnum_ops = N**3 + N**2*(N-1)  # N^3 muls, N^2 (N-1) adds\nelapsed = []\nfor i in range(10):\n    start = time.time()\n\n    sess.run(result_no_output)\n    elapsed.append(time.time()-start)\n\nprint(\"%d x %d matmul, %.2f elapsed, %.2f G ops\/sec\"%(N, N, min(elapsed), num_ops\/min(elapsed)\/10**9))\n<\/code><\/pre>\n\n<p>On TitanX pascal this shows 9.5 T ops\/sec which is close to theoretical max os 11 T ops\/sec theoretical maximum<\/p>\n\n<pre><code>8192 x 8192 matmul, 0.12 elapsed, 9527.10 G ops\/sec\n<\/code><\/pre>\n"}]}
{"QuestionId":41585226,"AnswerCount":1,"Tags":"<deep-learning><caffe><lasagne><regularized>","CreationDate":"2017-01-11T07:40:28.567","AcceptedAnswerId":41586157.0,"OwnerUserId":5583416.0,"Title":"L2 regularization in caffe, conversion from lasagne","Body":"<p>I have a lasagane code. I want to create the same network using caffe. I could convert the network. But i need help with the hyperparameters in lasagne. The hyperparameters in lasagne look like:<\/p>\n\n<pre><\/pre>\n\n<p>How do i perform the L2 regularization part in caffe? Do I have to add any layer for regularization after each convolution\/inner-product layer? Relevant parts from my solver.prototxt is as below:<\/p>\n\n<pre><\/pre>\n\n<p>also posted in <a href=\"http:\/\/datascience.stackexchange.com\">http:\/\/datascience.stackexchange.com<\/a>. Waiting for answers.<\/p>\n","answers":[{"AnswerId":"41586157","CreationDate":"2017-01-11T08:41:07.497","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like you already got it right.<br>\nThe <a href=\"https:\/\/stackoverflow.com\/q\/32177764\/1714410\"><code>weight_decay<\/code><\/a> meta-parameter combined with <code>regularization_type: \"L2\"<\/code> in your <code>'solver.prototxt'<\/code> tell caffe to use <code>L2<\/code> regularization with <code>weight_decay = 1e-5<\/code>.<\/p>\n\n<p>One more thing you might want to tweak is how much regularization affect each parameter. You can set this <em>for each parameter blob in the net<\/em> via<\/p>\n\n<pre><code>param { decay_mult: 1 }\n<\/code><\/pre>\n\n<p>For example, an <code>\"InnerProduct\"<\/code> layer with bias has two parameters:<\/p>\n\n<pre><code>layer {\n  type: \"InnerProduct\"\n  name: \"fc1\"\n  # bottom and top here\n  inner_product_param { \n    bias_term: true\n    # ... other params\n  }\n  param { decay_mult: 1 } # for weights use regularization\n  param { decay_mult: 0 } # do not regularize the bias\n}\n<\/code><\/pre>\n\n<p>By default, <code>decay_mult<\/code> is set to 1, that is, all weights of the net are regularized the same. You can change that to regularize more\/less specific parameter blobs.<\/p>\n"}]}
{"QuestionId":41585723,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-11T08:12:40.677","AcceptedAnswerId":41585795.0,"OwnerUserId":482601.0,"Title":"Extract patches from 3D Matrix","Body":"<p>I have a 3D matrix  of dimensions . I want to extract patches of dimensions  from each \"channel\" .  divides  and  divides . In this example,<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/46mev.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/46mev.jpg\" alt=\"Example\"><\/a><\/p>\n\n<p>I know how to do this in tensorflow using  but I was hoping for something more efficient in terms of setting it up, because the dimensions will be big and I'd rather not have the indices array of  in memory. Is there possibly an intelligent reshape? Either numpy or tensorflow solution would be very nice!<\/p>\n","answers":[{"AnswerId":"41585795","CreationDate":"2017-01-11T08:17:44.513","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>You could use some reshaping and swapping of axes -<\/p>\n\n<pre><code>A.reshape(h\/\/ph,ph,w\/\/pw,pw,-1).swapaxes(1,2)\n<\/code><\/pre>\n\n<p>Sample run -<\/p>\n\n<pre><code>In [46]: # Sample inputs\n    ...: h,w,c = 10,12,3\n    ...: ph, pw = 2,2\n    ...: A = np.random.randint(0,9,(h,w,c))\n    ...: \n\nIn [47]: A.reshape(h\/\/ph,ph,w\/\/pw,pw,-1).swapaxes(1,2).shape\nOut[47]: (5, 6, 2, 2, 3)\n<\/code><\/pre>\n\n<p>Each element (as block) along first two axes represent the patches. Thus. for the sample provided, we would have <code>5 x 6 = 30<\/code> patches.<\/p>\n\n<p>If you want those patches along one merged first axis, use one more <code>reshape<\/code> -<\/p>\n\n<pre><code>In [85]: out = A.reshape(h\/\/ph,ph,w\/\/pw,pw,-1).swapaxes(1,2).reshape(-1,ph,pw,c)\n\nIn [86]: out.shape\nOut[86]: (30, 2, 2, 3)\n<\/code><\/pre>\n\n<p>Let's verify by manually inspecting values themselves -<\/p>\n\n<pre><code>In [81]: A[:ph,:pw] # First patch\nOut[81]: \narray([[[6, 5, 2],\n        [4, 0, 1]],\n\n       [[0, 0, 4],\n        [2, 3, 0]]])\n\nIn [82]: A[:ph,pw:2*pw] # Second patch\nOut[82]: \narray([[[8, 3, 3],\n        [0, 0, 2]],\n\n       [[8, 5, 4],\n        [3, 4, 6]]])\n\nIn [83]: out[0]\nOut[83]: \narray([[[6, 5, 2],\n        [4, 0, 1]],\n\n       [[0, 0, 4],\n        [2, 3, 0]]])\n\nIn [84]: out[1]\nOut[84]: \narray([[[8, 3, 3],\n        [0, 0, 2]],\n\n       [[8, 5, 4],\n        [3, 4, 6]]])\n<\/code><\/pre>\n"}]}
{"QuestionId":41585866,"AnswerCount":3,"Tags":"<r><tensorflow><install><anaconda>","CreationDate":"2017-01-11T08:23:04.613","AcceptedAnswerId":null,"OwnerUserId":7026104.0,"Title":"How to install Tensorflow for R","Body":"<p>I had Tensorflow installed with Anaconda. Now I want use it in R and I need to reinstall Tensorflow, because the note <a href=\"https:\/\/rstudio.github.io\/tensorflow\" rel=\"nofollow noreferrer\">here<\/a>  <\/p>\n\n<blockquote>\n  <p>NOTE: You should NOT install TensorFlow with Anaconda as there are\n  issues with the way Anaconda builds the python shared library that\n  prevent dynamic linking from R.<\/p>\n<\/blockquote>\n\n<p>I already tried to uninstall from Anaconda and install with pip but its came to the same place in anaconda directory. Tesorflow is working from terminal but in R shows  <\/p>\n\n<p>Anybody can help me to how I can solve the problem? Should I uninstall anaconda and install Tensorflow using pip? <\/p>\n","answers":[{"AnswerId":"41602621","CreationDate":"2017-01-11T23:23:53.980","ParentId":null,"OwnerUserId":"322912","Title":null,"Body":"<p>You have several options on what to do. Probably the cleanest one is to install a system-wide python (if not installed yet) and then create a <a href=\"http:\/\/docs.python-guide.org\/en\/latest\/dev\/virtualenvs\/\" rel=\"nofollow noreferrer\">virtual environment<\/a>. This basically takes your system python binaries and moves them to its own compartment where everythign is isolated from the rest, incl. anaconda. Once you are inside an activated virtual environment you can install all the necessary Python appendages for TensorFlow. Once that is done, make sure you set up a correct environmental PATH for TensorFlow from where R can reach it:<\/p>\n\n<pre><code>Sys.setenv(TENSORFLOW_PYTHON=\"\/path\/to\/virtualenv\/python\/binary\")\ndevtools::install_github(\"rstudio\/tensorflow\")\n<\/code><\/pre>\n\n<p>Example of the path to where you installed the virtual environment project would be, I think, something like <code>~\/minion\/medvedi\/venv_medvedi\/bin\/python<\/code>.<\/p>\n"},{"AnswerId":"57841737","CreationDate":"2019-09-08T11:58:10.617","ParentId":null,"OwnerUserId":"9162933","Title":null,"Body":"<p>With the help of Keras, we can install the TensorFlow package in R.<\/p>\n\n<pre><code>install_keras()\nlibrary(keras)\ndevtools::install_github(\"rstudio\/keras\")\ninstall_tensorflow(package_url = \"https:\/\/pypi.python.org\/packages\/b8\/d6\/af3d52dd52150ec4a6ceb7788bfeb2f62ecb6aa2d1172211c4db39b349a2\/tensorflow-1.3.0rc0-cp27-cp27mu-manylinux1_x86_64.whl#md5=1cf77a2360ae2e38dd3578618eacc03b\")\nlibrary(tensorflow)\n<\/code><\/pre>\n\n<blockquote>\n  <p>Keras is a high-level neural network API for deep learning from TensorFlow Google.<\/p>\n<\/blockquote>\n"},{"AnswerId":"42159358","CreationDate":"2017-02-10T12:24:04.810","ParentId":null,"OwnerUserId":"468305","Title":null,"Body":"<p>This is no longer an issue, the documentation was updated too.<\/p>\n\n<p>See here:\n<a href=\"https:\/\/github.com\/rstudio\/tensorflow\/commit\/4e1e11d6ba2fe7efe1a03356f96172dbf8db365e\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rstudio\/tensorflow\/commit\/4e1e11d6ba2fe7efe1a03356f96172dbf8db365e<\/a><\/p>\n"}]}
{"QuestionId":41586449,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><training-data>","CreationDate":"2017-01-11T08:55:43.390","AcceptedAnswerId":41586823.0,"OwnerUserId":1871955.0,"Title":"how to make train.txt file in caffe","Body":"<p>I am using caffe net with python. I have train.txt file like this:<\/p>\n\n<pre><\/pre>\n\n<p>my questions:<\/p>\n\n<ol>\n<li><p>Can I use hex instead of int at the end of each line?<\/p><\/li>\n<li><p>About the label, does it need to start from 0<\/p><\/li>\n<\/ol>\n\n<p>Or Should I change the above to:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41586823","CreationDate":"2017-01-11T09:17:26.700","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you are using this <code>train.txt<\/code> list as input for <a href=\"https:\/\/stackoverflow.com\/a\/31431716\/1714410\"><code>convert_imageset<\/code><\/a> or as <code>source<\/code> for <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L772-L801\" rel=\"nofollow noreferrer\"><code>\"ImageData\"<\/code> layer<\/a> then the second part of each line <strong>must<\/strong> be a numeric label.<br>\nThe labels <strong>must<\/strong> start with 0 since they are used as indices into the predicted probability vectors when computing loss and accuracy. See <a href=\"https:\/\/stackoverflow.com\/a\/36174031\/1714410\">this thread<\/a> for more information.<\/p>\n"}]}
{"QuestionId":41587689,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-11T09:57:52.423","AcceptedAnswerId":41588099.0,"OwnerUserId":4922660.0,"Title":"How to save a tensor in checkpoint in Tensorflow?","Body":"<p>I want to use tf.train.Saver() to make checkpoint of a tensor, here is my code snippet:<\/p>\n\n<p><\/p>\n\n<pre><\/pre>\n\n<p>but got an error as follows:<\/p>\n\n<p><\/p>\n\n<pre><\/pre>\n\n<p>One way I think about is storing the Tensor in client by sess.run(temp) and save, but is there a more significant way?<\/p>\n","answers":[{"AnswerId":"41588099","CreationDate":"2017-01-11T10:16:52.263","ParentId":null,"OwnerUserId":"1346276","Title":null,"Body":"<p><code>temp<\/code> is not a <code>tf.Variable<\/code>, but an operation. It \"has\" no value or state, it is just a node in the graph. If you want to save the result of adding to <code>var<\/code> explicitely, you can assign <code>temp<\/code> to another variable by <code>tf.assign<\/code> and save this other variable.  The easier way would probably be to save <code>var<\/code> (or the whole session), and after restoring just evaluate <code>temp<\/code> again.<\/p>\n"}]}
{"QuestionId":41588319,"AnswerCount":0,"Tags":"<php><tensorflow><tensorflow-serving>","CreationDate":"2017-01-11T10:26:48.420","AcceptedAnswerId":null,"OwnerUserId":64106.0,"Title":"Connecting to tensorflow serving from PHP","Body":"<p>The connection between Tensorflow serving to PHP is quite immature. I could not find any relevant information online.<\/p>\n\n<p>In theory it should be possible to compile the proto files to pb.php files, but it fails on proto files dependencies due to the fact that some of the proto files are located in tensorflow_serving\/apis\/ and others in tensorflow\/core\/framework\/ . I tried every possible combination without success.<\/p>\n\n<p>If you have any success\/knowledge on connecting to tensorflow serving from PHP gRPC client, please post the details. Specifically how to compile the proto files and how to use pb.php files with protocol buffers runtime to actually connect to tensorflow server through gRPC<\/p>\n\n<p>Alternatively, if you have successfully compiled the proto files please share them and how to use them.<\/p>\n\n<p>Any information will help.  Thanks!<\/p>\n","answers":[]}
{"QuestionId":41588346,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><file-not-found>","CreationDate":"2017-01-11T10:27:48.727","AcceptedAnswerId":41589107.0,"OwnerUserId":3426322.0,"Title":"Can `No such file or directory` only mean that there is `No such file or directory`?","Body":"<p>I am trying to train a network using caffe, but am getting a file-not-found exception:  <\/p>\n\n<pre><\/pre>\n\n<p>There could be all kinds of things causing errors, but this specific file does seem to be there. I already changed the path to the absolute path, checked spelling (there is indeed a file with this exact name in this exact folder) and double checked the permissions of the file to be read-write.<br>\nThe difficulty is finding the cause of this error when the file does seem to be there. I mean, what else could be the cause?<br>\nSo my question is: does anyone know what could be the cause of this error?<\/p>\n\n<p>edit: on request, the output of :  <\/p>\n\n<pre><\/pre>\n\n<p>I'm assuming the relevant part of the prototxt file is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41589107","CreationDate":"2017-01-11T11:01:22.340","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your mean file is indeed in place, but this is not what caffe is complaining about.<br>\nWhat about your database file, <code>'\/home\/myUser\/Documents\/code\/RFNN\/data\/ILSVRC2012\/10class\/ilsvrc12_train_lmdb_10biggest'<\/code>? is it in the right location?<\/p>\n\n<hr>\n\n<h3>TL;DR<\/h3>\n\n<blockquote>\n  <p>Can <code>No such file or directory<\/code> only mean that there is <code>No such file or directory<\/code>?<\/p>\n<\/blockquote>\n\n<p>Yes.<\/p>\n"}]}
{"QuestionId":41588383,"AnswerCount":2,"Tags":"<python><multithreading><python-3.x><tensorflow><keras>","CreationDate":"2017-01-11T10:29:30.457","AcceptedAnswerId":41597238.0,"OwnerUserId":3591044.0,"Title":"How to run Keras on multiple cores?","Body":"<p>I'm using Keras with Tensorflow backend on a cluster (creating neural networks). How can I run it in a multi-threaded way on the cluster (on several cores) or is this done automatically by Keras? For example in Java one can create several threads, each thread running on a core.<\/p>\n\n<p>If possible, how many cores should be used?<\/p>\n","answers":[{"AnswerId":"53940268","CreationDate":"2018-12-27T05:44:49.220","ParentId":null,"OwnerUserId":"2911977","Title":null,"Body":"<p>you can configure session of tensorflow and use this session for keras backend<\/p>\n\n<pre><code> session_conf = tensorflow.ConfigProto(intra_op_parallelism_threads=8, inter_op_parallelism_threads=8)\n tensorflow.set_random_seed(1)\n sess = tensorflow.Session(graph=tensorflow.get_default_graph(), config=session_conf)\n keras.backend.set_session(sess)\n<\/code><\/pre>\n"},{"AnswerId":"41597238","CreationDate":"2017-01-11T17:36:03.927","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>Tensorflow automatically runs the computations on as many cores as are available on a single machine. <\/p>\n\n<p>If you have a distributed cluster, be sure you follow the instructions at <a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/distributed\/<\/a> to configure the cluster. (e.g. create the tf.ClusterSpec correctly, etc.)<\/p>\n\n<p>To help debug, you can use the <code>log_device_placement<\/code> configuration options on the session to have Tensorflow print out where the computations are actually placed. (Note: this works for both GPUs as well as distributed Tensorflow.)<\/p>\n\n<pre><code># Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n<\/code><\/pre>\n\n<p>Note that while Tensorflow's computation placement algorithm works fine for small computational graphs, you might be able to get better performance on large computational graphs by manually placing the computations in specific devices. (e.g. using <code>with tf.device(...):<\/code> blocks.)<\/p>\n"}]}
{"QuestionId":41589126,"AnswerCount":1,"Tags":"<keras><conv-neural-network><mnist><softmax>","CreationDate":"2017-01-11T11:02:17.600","AcceptedAnswerId":41594001.0,"OwnerUserId":2115332.0,"Title":"What is the replace for softmax layer in case more than one output can be activated?","Body":"<p>For example, I have CNN which tries to predict numbers from MNIST dataset (code written using Keras). It has 10 outputs, which form softmax layer. Only one of outputs can be true (independently for each digit from 0 to 9):<\/p>\n\n<pre><\/pre>\n\n<p>Sum of predicted is equal to 1.0 due to definition of softmax.<\/p>\n\n<p>Let's say I have a task where I need to classify some objects that can fall in several categories:<\/p>\n\n<pre><\/pre>\n\n<p>So I need to normalize in some other way. I need function which gives value on range [0, 1] and which sum can be larger than 1.<\/p>\n\n<p>I need something like that:<\/p>\n\n<pre><\/pre>\n\n<p>Each number is probability that object falls in given category. After that I can use some threshold like 0.5 to distinguish categories in which given object falls.<\/p>\n\n<p>The following questions appear:<\/p>\n\n<ol>\n<li>So which activation function can be used for this?<\/li>\n<li>May be this function already exists in Keras? <\/li>\n<li>May be you can propose some other way to predict in this case?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"41594001","CreationDate":"2017-01-11T14:54:14.767","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>Your problem is one of multi-label classification, and in the context of Keras it is discussed, for example, here: <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/741\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/741<\/a><\/p>\n\n<p>In short the suggested solution for it in keras is to replace the softmax layer with a sigmoid layer and use binary_crossentropy as your cost function.<\/p>\n\n<p>an example from that thread:<\/p>\n\n<pre><code># Build a classifier optimized for maximizing f1_score (uses class_weights)\n\nclf = Sequential()\n\nclf.add(Dropout(0.3))\nclf.add(Dense(xt.shape[1], 1600, activation='relu'))\nclf.add(Dropout(0.6))\nclf.add(Dense(1600, 1200, activation='relu'))\nclf.add(Dropout(0.6))\nclf.add(Dense(1200, 800, activation='relu'))\nclf.add(Dropout(0.6))\nclf.add(Dense(800, yt.shape[1], activation='sigmoid'))\n\nclf.compile(optimizer=Adam(), loss='binary_crossentropy')\n\nclf.fit(xt, yt, batch_size=64, nb_epoch=300, validation_data=(xs, ys), class_weight=W, verbose=0)\n\npreds = clf.predict(xs)\n\npreds[preds&gt;=0.5] = 1\npreds[preds&lt;0.5] = 0\n\nprint f1_score(ys, preds, average='macro')\n<\/code><\/pre>\n"}]}
{"QuestionId":41590281,"AnswerCount":0,"Tags":"<python-2.7><machine-learning><tensorflow><computer-vision><neural-network>","CreationDate":"2017-01-11T11:54:03.630","AcceptedAnswerId":null,"OwnerUserId":2106994.0,"Title":"Error in training model for Faster-R-CNN implemented in Tensorflow","Body":"<p>I am using the following github repository which implements a deep learning model called as Faster R-CNN in tensorflow:<\/p>\n\n<p><a href=\"https:\/\/github.com\/smallcorgi\/Faster-RCNN_TF\" rel=\"nofollow noreferrer\">https:\/\/github.com\/smallcorgi\/Faster-RCNN_TF<\/a><\/p>\n\n<p>This model will be used for object detection.<\/p>\n\n<p>I am trying to replicate the steps that are mentioned there to train a model and here is what I am facing:<\/p>\n\n<ol>\n<li>I follow steps from 1 to 5 successfully.<\/li>\n<li><p>Then, I run the command:<\/p>\n\n<p><strong>experiments\/scripts\/faster_rcnn_end2end.sh cpu 0 VGG16 pascal_voc<\/strong><\/p><\/li>\n<\/ol>\n\n<p>And I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>It looks like there is a problem in the creation of <strong>roi_pooling.so<\/strong> file. The file gets created successfully but it is showing this error.<\/p>\n\n<p>Does it mean that the mentioned repository is meant for GPU based computations and not CPU based ?<\/p>\n\n<p>How will I be able to run it on CPU only if possible ? <\/p>\n","answers":[]}
{"QuestionId":41592537,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-11T13:41:46.940","AcceptedAnswerId":41595438.0,"OwnerUserId":3882526.0,"Title":"Tensorflow multiplication broadcasting within batches","Body":"<p>We know that tf.multiply can broadcast like this:<\/p>\n\n<pre><\/pre>\n\n<p>This will give us <\/p>\n\n<pre><\/pre>\n\n<p>But my question is, what should I do if both  and  are in batches? That is,<\/p>\n\n<pre><\/pre>\n\n<p>Then how can I get the result of multiplying (broadcasting) the vector onto the matrix in each batch? Like the following answer:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41595438","CreationDate":"2017-01-11T16:04:50.737","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Broadcasting first adds singleton dimensions to the left until rank is matched. In first case that adds batch dimension. But in second case you already have batch dimension so you need to insert singleton dimension manually in the second position:<\/p>\n\n<pre><code>a = tf.reshape(tf.range(24), (2, 3, 4))\nb = tf.reshape(tf.range(8), (2, 4))\nsess.run(tf.mul(a, tf.expand_dims(b, 1)))\n<\/code><\/pre>\n"}]}
{"QuestionId":41592849,"AnswerCount":1,"Tags":"<python><arrays><numpy><theano>","CreationDate":"2017-01-11T13:58:08.227","AcceptedAnswerId":41895436.0,"OwnerUserId":4730816.0,"Title":"Convert numpy array into theano tensor variable","Body":"<p>I need to convert a given numpy array to theano tensor variable. I need this converted theano obj for multiplication with another theano variable.<\/p>\n\n<pre><\/pre>\n\n<p>I am using theano 0.9.0. I tried the above code ( after looking <a href=\"https:\/\/stackoverflow.com\/questions\/41198973\/convert-python-ndarray-to-theano-tensor-type-variable\">convert python ndarray to theano tensor type variable<\/a> ) but its resulting in the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am a beginner in theano thus any help will be appreciated :)<\/p>\n","answers":[{"AnswerId":"41895436","CreationDate":"2017-01-27T13:46:48.950","ParentId":null,"OwnerUserId":"4730816","Title":null,"Body":"<p>Wrong arguments were leading to the error.<\/p>\n"}]}
{"QuestionId":41593902,"AnswerCount":0,"Tags":"<python><tensorflow><pip><python-3.5><keras>","CreationDate":"2017-01-11T14:48:48.040","AcceptedAnswerId":null,"OwnerUserId":7017670.0,"Title":"Can't Get PyDot to Work in Python 3.5","Body":"<p>While using Keras library for machine learning, I keep getting the error:<\/p>\n\n<pre><\/pre>\n\n<p>My system is Windows 10, Anaconda 3.5 environment. I have tried all of the solutions mentioned in github issues and stack overflow:<\/p>\n\n<p>like installing pyparsing , pydot-ng, graphviz etc. Nothing is working, I keep getting the error.<\/p>\n\n<p>What to install to remove the error? Please help. Thanks.<\/p>\n","answers":[]}
{"QuestionId":41594334,"AnswerCount":1,"Tags":"<arguments><tensorflow><padding><layer>","CreationDate":"2017-01-11T15:10:39.333","AcceptedAnswerId":41693457.0,"OwnerUserId":5621651.0,"Title":"Does locally specifying an argument in a scope in tensorflow overwrite all following arguments of the same type?","Body":"<p>I am currently trying to understand the inception-v3 architecture and was taking a closer look at the definition of the model's layers:<\/p>\n\n<pre><\/pre>\n\n<p>Checking the dimensions of each layer, I first had to take a look at the different padding styles:  and .  will discard edges, while  will actually pad equally on both sides, so convolution still works on edges.\nThis holds for example for the first layer with 299x299 pixels to 149x149 with a stride of 2, so we only consider all odd pixels [Filter size: [3,3]] and end up with a dimension of 149x149, not 150x150 because padding is  (edges are discarded). Convolving this layer again, with the same filter size but now a stride of 1 we get 147x147 due to the edges \"suffering\" from being discarded. This layer then is again convolved but now with the twist, that padding is set to  which results in the same dimension of 147x147 as the layer before.<\/p>\n\n<p>Now comes the spot that confuses me:\nAssuming,  padding was only valid for the  layer and is globally still set to , the dimension for  is correctly shown as 73x73 due to discarding the edge. When now going to the next convolutional layer  I would expect it to become 71x71, taken the  padding as active. However, the output of  remains at 73x73, which means, that  padding is used. But in , the padding now seems to be , reducing the dimension to 71x71 confusing me totally.<\/p>\n\n<p>In the readme on github of slim's  I found, that setting one of the arguments locally overrides the global argument given:<\/p>\n\n<blockquote>\n<pre><\/pre>\n  \n  <p>As the example illustrates, the use of arg_scope makes the code\n  cleaner, simpler and easier to maintain. Notice that while argument\n  values are specifed in the arg_scope, they can be overwritten locally.\n  In particular, while the padding argument has been set to 'SAME', the\n  second convolution overrides it with the value of 'VALID'.<\/p>\n<\/blockquote>\n\n<p>However, this would mean, that  should also have dimension of 73x73 because the padding would be , so preserving the edges and the final pooling layer  would then even be 37x37.<\/p>\n\n<p>What is the thing that I am missing? Where is my mistake?\nThank you for helping me, I hope I have made the confusing problem clear.<\/p>\n","answers":[{"AnswerId":"41693457","CreationDate":"2017-01-17T09:31:14.103","ParentId":null,"OwnerUserId":"5621651","Title":null,"Body":"<p>I didn't see the filter size for the <code>pool1<\/code> layer is actually <code>[1,1]<\/code> so it is not reducing the dimensions and has nothing to do with the <code>arg_scope<\/code> as it stays exactly how it should.<\/p>\n"}]}
{"QuestionId":41595598,"AnswerCount":1,"Tags":"<python><matrix><theano>","CreationDate":"2017-01-11T16:11:47.097","AcceptedAnswerId":41596469.0,"OwnerUserId":245549.0,"Title":"How to do a \"generalized\" matrix operation in Theano?","Body":"<p>Let's assume that I have a 4-dimensional tensor  and I would like to multiply it with a 3-dimensional tensor  in the following way:<\/p>\n\n<pre><\/pre>\n\n<p>This is a generalization of matrix multiplication:<\/p>\n\n<pre><\/pre>\n\n<p>In other words, I pair some indexes of the first tensor with some indexes of the second tensor and then sum over them.<\/p>\n","answers":[{"AnswerId":"41596469","CreationDate":"2017-01-11T16:55:49.687","ParentId":null,"OwnerUserId":"5303618","Title":null,"Body":"<p>If I get you right this is done by the <code>dot<\/code> function of <code>theano.tensor<\/code> module.<\/p>\n\n<p>All the relevant information is in <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#linear-algebra\" rel=\"nofollow noreferrer\">the doc on linear algebra for tensor with theano<\/a><\/p>\n\n<p>You can even go further and specify the axis of summation using <code>theano.tensor.tensordot<\/code>, so to reproduce your example, it could be:<\/p>\n\n<pre><code>import theano.tensor as T\n\nA = T.ftensor4() # create a 4 dimensional float32 tensor\nB = T.ftensor4()\n\nC = T.tensordot(A,B, axes=[[1,3],[1,3]]) # here you can specify on which indices to sum\n\nprint C.ndim\n# &gt;&gt; 4\n<\/code><\/pre>\n"}]}
{"QuestionId":41595966,"AnswerCount":2,"Tags":"<concurrency><tensorflow>","CreationDate":"2017-01-11T16:30:38.960","AcceptedAnswerId":41597060.0,"OwnerUserId":5199939.0,"Title":"Run parallel op with different inputs and same placeholder","Body":"<p>I have the necessity to calculate more then one accuracy in the same time, concurrently. <\/p>\n\n<pre><\/pre>\n\n<p>The piece of code is the same of the <em>mnist<\/em> example in the tutorial of <strong>TensorFlow<\/strong> but instead of having:<\/p>\n\n<pre><\/pre>\n\n<p>I have two placeolder because I already calculated and stored them. <\/p>\n\n<pre><\/pre>\n\n<p>I want to fill the network with the values I aready have and then calculate the accuracy and this have to happen for each network I loaded.<\/p>\n\n<p>So if I load 20 networks I want to calculate in parallel the accuracy for each one. There is a way with the session run to execute the same operation with different input? <\/p>\n","answers":[{"AnswerId":"41599896","CreationDate":"2017-01-11T20:12:22.940","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>One approach to parallelizing TF computations is to execute <code>run<\/code> calls in parallel using threads (TF is incompatible with <code>multiprocessing<\/code>). It's a bit more complicated than other approaches because you have to handle parallelism yourself on the Python side.<\/p>\n\n<p>Here's an example that runs same matmul op in same session in different Python threads with different fed inputs and runs about 4x faster with 4 threads compared to 1 thread<\/p>\n\n<pre><code>import os, sys, queue, threading, time\nimport tensorflow as tf\nimport numpy as np\n\ndef p(s):\n    # helper function for printing from multiple threads\n    # need to append \\n or results get intermixed in notebook\n    print(s+\"\\n\", flush=True, end=\"\")\n\n\nnum_threads = 4\ndata_size = 32  # number of data points to enqueue\nwork_per_thread = data_size\/num_threads\ntimeout = 10  # grace period for dequeing\n\ninput_queue = queue.Queue(data_size)\noutput_queue = queue.Queue(data_size)\ndtype = np.float32\n\n# use matrix vector matmul since it's compute intensive and uses single core\n# see issue #6752\nn = 16*1024\nwith tf.device(\"\/cpu:0\"):\n    x = tf.placeholder(dtype)\n    matrix = tf.Variable(tf.ones((n, n)))\n    vector = tf.Variable(tf.ones((n, 1)))\n    y = tf.matmul(matrix, vector)[0, 0] + x\n\n# turn off graph-rewriting optimizations\nsess = tf.Session(config=tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0))))\nsess.run(tf.global_variables_initializer())\n\ndone = False\ndef runner(runner_id):\n    p(\"Starting runner %s\" % (runner_id,))\n    count = 0\n    while not done:\n        try:\n            x_val = input_queue.get(timeout=1)\n        except queue.Empty:\n            # retry on empty queue\n            continue\n        p(\"Start computing %d on %d\" %(x_val, runner_id))\n        out = sess.run(y, {x: x_val})\n        count+=1\n        output_queue.put(out)\n        if count&gt;=work_per_thread:\n            break\n    else:\n        p(\"Stopping runner \"+str(runner_id))\n\nthreads = []\nprint(\"Creating threads.\")\nfor i in range(num_threads):\n    t = threading.Thread(target=runner, args=(i,))\n    threads.append(t)\n\nfor i in range(data_size):\n    input_queue.put(i, timeout=timeout)\n\n# start threads\np(\"Launching runners.\")\nstart_time = time.time()\nfor t in threads:\n    t.start()\n\np(\"Reading results.\")\nfor i in range(data_size):\n    try:\n        p(\"Main thread: obtained %.2f\" % (output_queue.get(timeout=timeout),))\n    except queue.Empty:\n        print(\"No results after %d, terminating computation.\"%(timeout,))\n        break\nelse:\n    p(\"Computed successfully.\")\ndone = True\n\np(\"Waiting for threads to finish.\")\nfor t in threads:\n    t.join()\nprint(\"Done in %.2f seconds\" %(time.time() - start_time))\n<\/code><\/pre>\n"},{"AnswerId":"41597060","CreationDate":"2017-01-11T17:26:53.840","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>You have multiple options to make things happen in parallel:<\/p>\n\n<ul>\n<li>Parallelize using multiple python threads \/ subprocesses. (See Python's \"multiprocessing\" library.)<\/li>\n<li>Batch up the operations into single larger operations. (e.g. Similar to the image operations that operate on a batch of images simultaneously <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/image\/resizing#resize_bilinear\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/image\/resizing#resize_bilinear<\/a>.)<\/li>\n<li>Make a single graph that has the 20 network accuracy calculations.<\/li>\n<\/ul>\n\n<p>I think the last one is the easiest, so I've included a bit of sample code below to get you started:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef construct_accuracy_calculation(i):\n  W = tf.placeholder(tf.float32, [784, 10], name=(\"%d_W\" % i))\n  b = tf.placeholder(tf.float32, [10], name=(\"%d_b\" % i))\n  # ...\n  correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))\n  accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n  return (W, b, accuracy)\n\n\ndef main():\n  accuracy_computations = []\n  feed_dict={}\n  for i in xrange(NUM_NETWORKS):\n    (W, b) = load_network(i)\n    (W_op, b_op, accuracy) = construct_accuracy_calculation(i)\n    feed_dict[W_op] = W\n    feed_dict[b_op] = b\n    accuracy_computations.append(accuracy)\n\n  # sess = ...\n  accuracy_values = sess.run(accuracy_computations, feed_dict=feed_dict)\n\nif __name__ == \"__main__\":\n  main()\n<\/code><\/pre>\n"}]}
{"QuestionId":41595991,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2017-01-11T16:32:36.930","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"How to train ResNet101 model from scratch in Caffe?","Body":"<p>I'm using the <a href=\"https:\/\/bitbucket.org\/aquariusjay\/deeplab-public-ver2\" rel=\"nofollow noreferrer\">DeepLab_v2<\/a> version of Caffe in order to do semantic segmentation. I can finetune the ResNet101 using imagenet model,  but I cannot train the model from scratch using custom data. Did anyone have similar experience and managed to solve this issue?<\/p>\n\n<p>This is how a functional block of the ResNet looks like, that I'm currently using for training:<\/p>\n\n<pre><\/pre>\n\n<p>I tried all kinds of variations including . I am able to train one single block of the type above, but when I try to use all 101 layers, the model does not converge anymore.\nAny ideas?<\/p>\n","answers":[]}
{"QuestionId":41596619,"AnswerCount":1,"Tags":"<neural-network><keras>","CreationDate":"2017-01-11T17:03:24.960","AcceptedAnswerId":41602393.0,"OwnerUserId":7384314.0,"Title":"How to build simple neural network on keras (not image recognition)","Body":"<p>I am new to keras and I am trying to built my own neural network.<\/p>\n\n<p>A task:<\/p>\n\n<p>I need to write a system that can make decisions for the character, which may meet one or more enemies. The system can be known:<\/p>\n\n<ul>\n<li>Percentage Health character<\/li>\n<li>Presence of the pistol;<\/li>\n<li>The number of enemies.<\/li>\n<\/ul>\n\n<p>The answer must be in the form of one of the following:<\/p>\n\n<ol>\n<li>Attack<\/li>\n<li>Run<\/li>\n<li>Hide (for a surprise attack)<\/li>\n<li>To do nothing<\/li>\n<\/ol>\n\n<p>To train up I made a table of \"lessons\":<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lD0WX.png\" rel=\"nofollow noreferrer\">https:\/\/i.stack.imgur.com\/lD0WX.png<\/a><\/p>\n\n<p>So here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Here the predictions I get.\n[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]<\/p>\n\n<p>The accuracy on each epoch is 0.2727 and the loss is decrease.\nIt's not right.<\/p>\n\n<p>I was trying to devide learning rate by 10, changing activations and optimizers. Even data I input manually.\nCan anyone tell me how to solve my simple problem. thx.<\/p>\n","answers":[{"AnswerId":"41602393","CreationDate":"2017-01-11T23:07:17.350","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>There are several problems in your code.<\/p>\n\n<ol>\n<li>Number of data entries are very small compared to the NN model. <\/li>\n<li>Y is represented as classes number and not as class vector. A regression model can be learnt on this but its a poor design choice.<\/li>\n<li>output of <code>softmax<\/code> function is always between 0-1 .. as this is used your model only knows to spew out values between 0-1.<\/li>\n<\/ol>\n\n<p>Here below is a bit better modified code:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense\nfrom keras.optimizers import SGD\nimport numpy\n# fix random seed for reproducibility\nseed = 7\nnumpy.random.seed(seed)\n# split into input (X) and output (Y) variables\nX = numpy.array([[0.5,1,1], [0.9,1,2], [0.8,0,1], [0.3,1,1], [0.6,1,2], [0.4,0,1], [0.9,1,7], [0.5,1,4], [0.1,0,1], [0.6,1,0], [1,0,0]])\ny = numpy.array([[1],[1],[1],[2],[2],[2],[3],[3],[3],[0],[0]])\n\nfrom keras.utils import np_utils\nY = np_utils.to_categorical(y, 4)\n# print Y\n\n# create model\nmodel = Sequential()\nmodel.add(Dense(3, input_dim=3, activation='relu'))\nmodel.add(Dense(4, activation='softmax'))\n# Compile model\n# sgd = SGD(lr=0.1)\n# model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])\nmodel.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\n\n# Fit the model\nmodel.fit(X, Y, nb_epoch=700)\n\n# calculate predictions\npredictions = model.predict(X)\n\npredictions_class = predictions.argmax(axis=-1)\nprint(predictions_class)\n<\/code><\/pre>\n\n<p>Note I have used the <code>softmax<\/code> activation as the classes are mutually exclusive  <\/p>\n"}]}
{"QuestionId":41596901,"AnswerCount":1,"Tags":"<tensorflow><slack-api><chatbot>","CreationDate":"2017-01-11T17:18:57.580","AcceptedAnswerId":null,"OwnerUserId":5475451.0,"Title":"Chatbot: generative model using Python framework deployed in Slack","Body":"<p>Could somebody tell me whether it is possible to develop a chatbot using python ML frameworks such as tensorflow and deploy in Slack using Slack's apps? <\/p>\n\n<p>As far as I have read we could develop some retrieval based model using node.js. But I'm looking for a generative model. <\/p>\n\n<p>Anything to help me get started is much appreciated. <\/p>\n\n<p>Thanks! <\/p>\n","answers":[{"AnswerId":"45553651","CreationDate":"2017-08-07T18:38:48.570","ParentId":null,"OwnerUserId":"5915270","Title":null,"Body":"<p>In order to use a generative model you have to use <strong>tensorflow sequence to sequence architecture<\/strong> . Here's the <a href=\"https:\/\/github.com\/tensorflow\/nmt\" rel=\"nofollow noreferrer\">official TF tutorial<\/a> from TF on sequence-to-sequence architecture . <\/p>\n\n<p>So after training the model you can set this as an API which is easy. Then may be you can set it with the slack API for sure. <\/p>\n"}]}
{"QuestionId":41597862,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-11T18:13:15.083","AcceptedAnswerId":null,"OwnerUserId":7406007.0,"Title":"Tensorflow You must feed a value for placeholder tensor 'Placeholder' with dtype bool","Body":"<p>I run the code for second respondents and catch an error.\n<a href=\"https:\/\/stackoverflow.com\/questions\/33949786\/how-could-i-use-batch-normalization-in-tensorflow\">tensorflow batch normalizatioon second respondents<\/a>\nError pic here<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/s7KeE.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/s7KeE.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41600214","CreationDate":"2017-01-11T20:34:58.930","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>When you run the op, you must provide a value for <code>feed_dict<\/code>.<\/p>\n\n<p>Here is an example program:<\/p>\n\n<pre><code>import tensorflow as tf\n\n# Define the inputs you will feed into the tensorflow computation graph\na = tf.placeholder(tf.int32, shape=[1], name=\"a\")\nx = tf.placeholder(tf.int32, shape=[4], name=\"x\")\n# This is the actual computation we want to run.\noutput = a * x\n\nwith tf.Session() as sess:\n  # Actually run the computation, feeding in [10] for a, and [1, 2, 3, 4] for x.\n  # This will print out: [10 20 30 40]\n  print sess.run(output, feed_dict={a: [10], x: [1, 2, 3, 4]})\n<\/code><\/pre>\n"}]}
{"QuestionId":41598505,"AnswerCount":1,"Tags":"<image><python-3.x><tensorflow><slice><crop>","CreationDate":"2017-01-11T18:51:12.637","AcceptedAnswerId":41600007.0,"OwnerUserId":78187.0,"Title":"Tensorflow - Dynamic Slicing of Images","Body":"<p>I previously asked this question, but after some investigation of the problem it appears I may have just gone down the wrong path for what I am trying to achieve.  <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/41576944\/dynamic-image-cropping-in-tensorflow\">Dynamic image cropping in Tensorflow<\/a>   <\/p>\n\n<p>I thought maybe this might be a better path to try.  But the part I can't figure out is what I should put for the size parameter on the slice operation.  Fundamentally, what I am trying to achieve is having the capability to dynamically decide how to crop an image and then crop it and then continue with those cropped images in my computation graph.  Feel free to offer an alternative if this seems to be an inefficient way to go about this.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41600007","CreationDate":"2017-01-11T20:20:04.403","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>Instead of using tf.slice (which doesn't let you operate on a batch), I recommend using <code>tf.image.extract_glimpse<\/code>. Here is a toy sample program that operates in a batch:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nNUM_IMAGES = 2\nNUM_CHANNELS = 1\nCROP_SIZE = [3, 4]\nIMG_HEIGHT=10\nIMG_WIDTH=10\n\n# Fake input data, but ordered so we can look at the printed values and\n# map them back. The values of the first image are:\n# array([[ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9],\n#        [10, 11, 12, 13, 14, 15, 16, 17, 18, 19],      \n#        [20, 21, 22, 23, 24, 25, 26, 27, 28, 29],      \n#        [30, 31, 32, 33, 34, 35, 36, 37, 38, 39],      \n#        [40, 41, 42, 43, 44, 45, 46, 47, 48, 49],      \n#        [50, 51, 52, 53, 54, 55, 56, 57, 58, 59],      \n#        [60, 61, 62, 63, 64, 65, 66, 67, 68, 69],      \n#        [70, 71, 72, 73, 74, 75, 76, 77, 78, 79],      \n#        [80, 81, 82, 83, 84, 85, 86, 87, 88, 89],      \n#        [90, 91, 92, 93, 94, 95, 96, 97, 98, 99]])      \nimage1 = np.reshape(\n    np.array(xrange(NUM_IMAGES * IMG_HEIGHT * IMG_WIDTH * NUM_CHANNELS)),\n    [NUM_IMAGES, IMG_HEIGHT, IMG_WIDTH, NUM_CHANNELS])\n\n# We use normalized=False to use pixel indexing.\n# normalized=True means centers are specified between [0,1).\nimage1_center = [0, 0]  # The center of the crop is ~ the center of the image.\nimage2_center = [3, 5]  # Offset down &amp; right in the image.\n\nimg = tf.placeholder(tf.float32, shape=[NUM_IMAGES, IMG_HEIGHT, IMG_WIDTH, NUM_CHANNELS], name=\"img\")\nsize = tf.placeholder(tf.int32, shape=[2], name=\"crop_size\")\ncenters = tf.placeholder(tf.float32, shape=[NUM_IMAGES, 2], name=\"centers\")\noutput = tf.image.extract_glimpse(img, size, centers, normalized=False)\n\nsess = tf.Session()\nfeed_dict = {\n  img: image1,\n  size: CROP_SIZE,\n  centers: [image1_center, image2_center],\n}\nprint sess.run(output, feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>If you would like to extract multiple sizes (and even multiple glimpses per image), check out <code>tf.image.crop_and_resize<\/code>.<\/p>\n\n<p>Docs: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/image\/cropping#extract_glimpse\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/image\/cropping#extract_glimpse<\/a> <\/p>\n"}]}
{"QuestionId":41599063,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving><openai-gym>","CreationDate":"2017-01-11T19:21:24.180","AcceptedAnswerId":null,"OwnerUserId":6609435.0,"Title":"How can I run trained agents in new Tensorflow session?","Body":"<p>I would like to train multiple agents (with possibly very different graphs, variables, ...) using OpenAI's Universe\/gym.<\/p>\n\n<p>I started with the <a href=\"https:\/\/github.com\/openai\/universe-starter-agent\" rel=\"nofollow noreferrer\">universe-starter-agent<\/a> code and adjusted the saver as to also dump the .meta files.<\/p>\n\n<p>The process of recovering a trained Agent and using it to run inference seems very tricky.. What I'm currently doing:<\/p>\n\n<ul>\n<li>I added some variable names to the LSTMPolicy class in model.py: <\/li>\n<\/ul>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"false\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-css lang-css prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<ul>\n<li>I then wrote this snippet to try and load the graph and use it for inference: <\/li>\n<\/ul>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-css lang-css prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>Everything I tried so far gave me an empty list back for logits and the current snippet (above) is giving me errors:\n\"The name 'global\/LSTM_state_in_0' refers to an Operation, not a Tensor.\"<\/p>\n\n<p>Even if I got this to work, I'd have to adjust the code every time I change something to the architecture of my Agent's graph... So I'm thinking there must be an easier way to do this? Anybody who can help me out here?<\/p>\n\n<p>Ideally, I would like to have a function load_model(path_to_model)\nthat starts up a session, loads everything that is needed, returns an object onto which I can call a .predict method of some sort so that I can feed a single numpy array (42,42,1) and get the logits out of the trained Agent...<\/p>\n","answers":[{"AnswerId":"41601346","CreationDate":"2017-01-11T21:51:24.920","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The following lines seem problematic:<\/p>\n\n<pre><code>state_in_0 = graph.get_tensor_by_name(\"global\/LSTM_state_in_0\")\nstate_in_1 = graph.get_tensor_by_name(\"global\/LSTM_state_in_1\")\n<\/code><\/pre>\n\n<p>The problem is that <code>\"global\/LSTM_state_in_0\"<\/code> is the name of an <strong>operation<\/strong>, rather than the name of a <strong>tensor<\/strong> (i.e. a particular output of an operation). Assuming that you're interested in the 0th output of these operations, you should append <code>\":0\"<\/code> to the names, and the correct code would be:<\/p>\n\n<pre><code>state_in_0 = graph.get_tensor_by_name(\"global\/LSTM_state_in_0:0\")\nstate_in_1 = graph.get_tensor_by_name(\"global\/LSTM_state_in_1:0\")\n<\/code><\/pre>\n"}]}
{"QuestionId":41599176,"AnswerCount":1,"Tags":"<format><caffe><hdf5>","CreationDate":"2017-01-11T19:29:13.877","AcceptedAnswerId":41601562.0,"OwnerUserId":5583416.0,"Title":"format for data for HDF5 input to caffe","Body":"<p>My images are 3channel RGB images. I convert each image to vector form.\nI intend to feed this 'data' to caffe using HDF5 format.<\/p>\n\n<p>My code to form the HDF5 data is(using matlab) <\/p>\n\n<pre><\/pre>\n\n<p>The data blob above is of the format (row, col, channel, number_samples). The Data blob should be of form(number_samples,channel,width, height) as per `<a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1Blob.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1Blob.html<\/a><\/p>\n\n<p>When i run  training with (row, col, channel, number_samples), the code runs. When i run  training with (number_samples, channels, row, col, ), the code gives error: <\/p>\n\n<pre><\/pre>\n\n<p>where 7500 is the number of images in the dataset.\nWhat is the correct way to enter the data for HDF5 format?<\/p>\n","answers":[{"AnswerId":"41601562","CreationDate":"2017-01-11T22:06:36.440","ParentId":null,"OwnerUserId":"3327666","Title":null,"Body":"<p>One lead for this is that Matlab stores data in Fortran order and not in C order, which is what HDF5 uses internally. Have a look at the Matlab documentation for HDF5, it should be pretty explicit about this.<\/p>\n"}]}
{"QuestionId":41599798,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-11T20:05:38.413","AcceptedAnswerId":41601297.0,"OwnerUserId":808044.0,"Title":"Where does tensorflow log error?","Body":"<p>I use TF v0.12.1 with GPU support on Ubuntu-16.04 64 bits, but have an internal error in my pipeline:<\/p>\n\n<pre><\/pre>\n\n<p>Where does TF log its error messages?<\/p>\n","answers":[{"AnswerId":"41601297","CreationDate":"2017-01-11T21:48:20.247","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The log corresponding to that exception should be written to standard error in the process that run the op. The error message is produced by <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f8fcd85b8151d287ad09a9323fabf1c5025775b7\/tensorflow\/python\/lib\/core\/py_func.cc#L192\" rel=\"noreferrer\">this code<\/a>, which calls <a href=\"https:\/\/docs.python.org\/2\/c-api\/exceptions.html\" rel=\"noreferrer\"><code>PyErr_Print()<\/code><\/a>, which renders the current Python exception (in the C API) to standard error.<\/p>\n\n<p>(Note that if you are using distributed TensorFlow and the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/script_ops\/script_language_operators_#py_func\" rel=\"noreferrer\"><code>tf.py_func()<\/code><\/a> op is placed in a different task, you should look in the standard error for that task for log messages. However, please note that <code>tf.py_func()<\/code> does not work if the op is placed in a different process from the process that created the graph, because it relies on the code for the function being registered in a process-wide registry.)<\/p>\n"}]}
{"QuestionId":41600067,"AnswerCount":0,"Tags":"<python><apache-spark><pyspark><keras>","CreationDate":"2017-01-11T20:24:19.020","AcceptedAnswerId":null,"OwnerUserId":6348351.0,"Title":"Broadcasting a keras model with pyspark","Body":"<p>I have a Keras model and I need to broadcast it\/include it in a pyspark UDF. It can't be pickled,  so simply sc.broadcast(model) won't work. Any ideas?<\/p>\n","answers":[]}
{"QuestionId":41600321,"AnswerCount":1,"Tags":"<graph><tensorflow><distributed>","CreationDate":"2017-01-11T20:41:51.947","AcceptedAnswerId":41601168.0,"OwnerUserId":4394807.0,"Title":"Distributed tensorflow: the difference between In-graph replication and Between-graph replication","Body":"<p>I got confused about the two concepts:  and  when reading the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/distributed\/#replicated_training\" rel=\"noreferrer\">Replicated training<\/a> in tensorflow's official How-to. <\/p>\n\n<ol>\n<li><p>It's said in above link that <\/p>\n\n<blockquote>\n  <p><strong>In-graph replication.<\/strong> In this approach, the client builds a single\n  tf.Graph that contains one set of parameters (in tf.Variable nodes\n  pinned to \/job:ps); ...<\/p>\n<\/blockquote>\n\n<p>Does this mean there are <strong>multiple<\/strong> s in  approach? If yes, where are the corresponding codes in\nthe provided examples?<\/p><\/li>\n<li><p>While there is already a  example in above link, could anyone provide a \nimplementation (pseudo code is fine)  and highlight its main\ndifferences from ?<\/p>\n\n<p>Thanks in advance! <\/p>\n\n<hr><\/li>\n<\/ol>\n\n<h2><strong>Edit_1: more questions<\/strong><\/h2>\n\n<p>Thanks a lot for your detailed explanations and gist code @mrry @YaroslavBulatov ! After looking\n    your responses, I have the following two questions:<\/p>\n\n<ol start=\"3\">\n<li><p>There is the following statement in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/distributed\/#replicated_training\" rel=\"noreferrer\">Replicated training<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>Between-graph replication.<\/strong> In this approach, there is a separate\n  client for each \/job:worker task, typically in the same process as the\n  worker task. Each client builds a <strong>similar graph<\/strong> containing the\n  parameters (pinned to \/job:ps as before using\n  tf.train.replica_device_setter() to map them deterministically to the\n  same tasks); and a <strong>single copy<\/strong> of the compute-intensive part of the\n  model, pinned to the local task in \/job:worker.<\/p>\n<\/blockquote>\n\n<p>I have two sub-questions related to above words in bold.<\/p>\n\n<p>(A) Why do we say each client builds <strong>similar graph<\/strong>, but not <strong>same graph<\/strong>?\nI wonder the graph built in each client in the example of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/distributed\/#replicated_training\" rel=\"noreferrer\">Replicated training<\/a>\nshould be the same because below graph construction codes are shared within all s.:<\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>(B) Shouldn't it be <strong>multiple copies<\/strong> of compute-intensive part of\nthe model, since we have multiple ?<\/p><\/li>\n<li><p>Does the example in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/distributed\/#replicated_training\" rel=\"noreferrer\">Replicated training<\/a> support training on multiple machines, each of which has multiple GPUs? If not, can we\nuse simultaneously both the  to support training on multiple\nGPUs on each machine and  for\ncross-machine training? I ask this question because\n@mrry indicated that the  is essentially same to the way\nused in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/91c7b91f834a5a857e8168b96d6db3b93d7b9c2a\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"noreferrer\">CIFAR-10 example model for multiple GPUs<\/a>.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"41601168","CreationDate":"2017-01-11T21:39:51.787","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>First of all, for some historical context, \"in-graph replication\" is the first approach that we tried in TensorFlow, and it did not achieve the performance that many users required, so the more complicated \"between-graph\" approach is the current recommended way to perform distributed training. Higher-level libraries such as <code>tf.learn<\/code> use the \"between-graph\" approach for distributed training.<\/p>\n\n<p>To answer your specific questions:<\/p>\n\n<ol>\n<li><blockquote>\n  <p>Does this mean there are multiple <code>tf.Graph<\/code>s in the between-graph\n  replication approach? If yes, where are the corresponding codes in the provided examples?<\/p>\n<\/blockquote>\n\n<p>Yes. The typical between-graph replication setup will use a separate TensorFlow process for each worker replica, and each of this will build a separate <code>tf.Graph<\/code> for the model. Usually each process uses the global default graph (accessible through <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/utility_functions#get_default_graph\" rel=\"noreferrer\"><code>tf.get_default_graph()<\/code><\/a>) and it is not created explicitly.<\/p>\n\n<p>(In principle, you could use a single TensorFlow process with the same <code>tf.Graph<\/code> and multiple <code>tf.Session<\/code> objects that share the same underlying graph, as long as you configured the <code>tf.ConfigProto.device_filters<\/code> option for each session differently, but this is an uncommon setup.)<\/p><\/li>\n<li><blockquote>\n  <p>While there is already a between-graph replication example in above link, could anyone provide an in-graph replication implementation (pseudocode is fine) and highlight its main differences from between-graph replication?<\/p>\n<\/blockquote>\n\n<p>For historical reasons, there are not many examples of in-graph replication (<a href=\"https:\/\/gist.github.com\/yaroslavvb\/ef407a599f0f549f62d91c3a00dcfb6c\" rel=\"noreferrer\">Yaroslav's gist<\/a> is one exception). A program using in-graph replication will typically include a loop that creates the same graph structure for each worker (e.g. the loop on <a href=\"https:\/\/gist.github.com\/yaroslavvb\/ef407a599f0f549f62d91c3a00dcfb6c#file-simple_barrier-py-L74\" rel=\"noreferrer\">line 74 of the gist<\/a>), and use variable sharing between the workers. <\/p>\n\n<p>The one place where in-graph replication persists is for using multiple devices in a single process (e.g. multiple GPUs). The <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/91c7b91f834a5a857e8168b96d6db3b93d7b9c2a\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"noreferrer\">CIFAR-10 example model for multiple GPUs<\/a> is an example of this pattern (see the loop over GPU devices <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/91c7b91f834a5a857e8168b96d6db3b93d7b9c2a\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py#L165\" rel=\"noreferrer\">here<\/a>).<\/p><\/li>\n<\/ol>\n\n<p>(In my opinion, the inconsistency between how multiple workers and multiple devices in a single worker are treated is unfortunate. In-graph replication is simpler to understand than between-graph replication, because it doesn't rely on implicit sharing between the replicas. Higher-level libraries, such as <code>tf.learn<\/code> and TF-Slim, hide some of these issues, and offer hope that we can offer a better replication scheme in the future.)<\/p>\n\n<ol start=\"3\">\n<li><blockquote>\n  <p>Why do we say each client builds a <strong>similar<\/strong> graph, but not the <strong>same<\/strong> graph?<\/p>\n<\/blockquote>\n\n<p>Because they aren't required to be identical (and there is no integrity check that enforces this). In particular, each worker might create a graph with different explicit device assignments (<code>\"\/job:worker\/task:0\"<\/code>, <code>\"\/job:worker\/task:1\"<\/code>, etc.). The chief worker might create additional operations that are not created on (or used by) the non-chief workers. However, in most cases, the graphs are logically (i.e. modulo device assignments) the same.<\/p>\n\n<blockquote>\n  <p>Shouldn't it be <strong>multiple copies<\/strong> of the compute-intensive part of the model, since we have multiple workers?<\/p>\n<\/blockquote>\n\n<p>Typically, each worker has a separate graph that contains a single copy of the compute-intensive part of the model. The graph for worker <em>i<\/em> does not contain the nodes for worker <em>j<\/em> (assuming i &ne; j). (An exception would be the case where you're using between-graph replication for distributed training, and in-graph replication for using multiple GPUs in each worker. In that case, the graph for a worker would typically contain <em>N<\/em> copies of the compute-intensive part of the graph, where <em>N<\/em> is the number of GPUs in that worker.)<\/p><\/li>\n<li><blockquote>\n  <p>Does the example in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/distributed\/#replicated_training\" rel=\"noreferrer\">Replicated training<\/a> support training on multiple machines, each of which has multiple GPUs?<\/p>\n<\/blockquote>\n\n<p>The example code only covers training on multiple machines, and says nothing about how to train on multiple GPUs in each machine. However, the techniques compose easily. In this part of the example:<\/p>\n\n<pre><code># Build model...\nloss = ...\n<\/code><\/pre>\n\n<p>...you could add a loop over the GPUs in the local machine, to achieve distributed training multiple workers each with multiple GPUs.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":41600519,"AnswerCount":2,"Tags":"<machine-learning><neural-network><keras>","CreationDate":"2017-01-11T20:55:00.110","AcceptedAnswerId":41612192.0,"OwnerUserId":7406649.0,"Title":"Accuracy does not go up on a keras model","Body":"<p>I'm trying to train a model on data from the Higgs Boson challenge on kaggle. The first thing I decided to do was to create a simple keras model. I've tried different amount and width of layers, different cost functions, different optimizers different functions in neurons, but the accuracy on the training set is always between 0.65-0.7 range. I don't really understand why. Here's my an example of a model that worked so weird:<\/p>\n\n<pre><\/pre>\n\n<p>I also tried larger models and got such an accuracy too. Please tell me what I am doing wrong.<\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I have tried training this model with 100 epochs and the batch size 0f 100 and got loss equal to 4.9528 and accuracy to 0.6924 again. And it always outputs zero for every example.<\/p>\n","answers":[{"AnswerId":"41612192","CreationDate":"2017-01-12T11:34:32.350","ParentId":null,"OwnerUserId":"2314651","Title":null,"Body":"<p>The problem arise from the fact that your model always outputs the majority class. It's not a weighted problem (one of the classes appears more than the other) and it seems that your network \"learns\" to always output the same class.<\/p>\n\n<p>Try using a different classifier (Random Forest for example) and you'll see that the accuracy is much better.<\/p>\n\n<pre><code>from sklearn.ensemble import RandomForestClassifier\nrf = RandomForestClassifier()\nrf.fit(X_train, y_train)\n<\/code><\/pre>\n\n<p>When trying to address the issue with the neural network I uses SMOTE to balance the train dataset. You should use \"adam\" as the optimizer for the classification. Also, a much smaller network architecture should be enough for this problem.<\/p>\n\n<pre><code>from keras.layers import Dense, Dropout\nfrom keras.models import Sequential\nimport pandas as pd\nimport numpy as np\nfrom sklearn.model_selection import train_test_split\nfrom imblearn.over_sampling import SMOTE\n\ndf = pd.read_csv(\"training.csv\")\ny = np.array(df['Label'].apply(lambda x: 0 if x=='s' else 1))\nX = np.array(df.drop([\"EventId\",\"Label\"], axis=1))\nsm = SMOTE()\n\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)\nX_res, y_res = sm.fit_sample(X_train, y_train)\n\nmodel = Sequential()\nmodel.add(Dense(25, input_shape=(31,),activation=\"relu\"))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(10, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(1, activation='sigmoid'))\nmodel.compile(optimizer=\"adam\",loss='binary_crossentropy',metrics=['accuracy'])\nmodel.fit(X_res, y_res,validation_data=(X_test, y_test),nb_epoch=100,batch_size=100)\n<\/code><\/pre>\n\n<p>An example results:<\/p>\n\n<pre><code>    Epoch 11\/100\n230546\/230546 [==============================] - 5s - loss: 0.5146 - acc: 0.7547 - val_loss: 0.3365 - val_acc: 0.9138\nEpoch 12\/100\n230546\/230546 [==============================] - 5s - loss: 0.4740 - acc: 0.7857 - val_loss: 0.3033 - val_acc: 0.9270\nEpoch 13\/100\n230546\/230546 [==============================] - 5s - loss: 0.4171 - acc: 0.8295 - val_loss: 0.2821 - val_acc: 0.9195\n<\/code><\/pre>\n"},{"AnswerId":"41601200","CreationDate":"2017-01-11T21:42:08.423","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>You are training way too short<\/p>\n\n<blockquote>\n  <p>model.fit(train,labels,nb_epoch=1,batch_size=1)<\/p>\n<\/blockquote>\n\n<p>this means you are going <strong>once<\/strong> through the data, with extremely small batch, it should be something among the lines of<\/p>\n\n<pre><code>model.fit(train, labels, nb_epoch=100, batch_size=100)\n<\/code><\/pre>\n"}]}
{"QuestionId":41601620,"AnswerCount":2,"Tags":"<python><tensorflow><google-cloud-dataflow><google-cloud-ml>","CreationDate":"2017-01-11T22:10:53.847","AcceptedAnswerId":41602040.0,"OwnerUserId":2984051.0,"Title":"cloudml retraining inception - received a label value outside the valid range","Body":"<p>I am following the flowers <a href=\"https:\/\/cloud.google.com\/blog\/big-data\/2016\/12\/how-to-train-and-classify-images-using-google-cloud-machine-learning-and-cloud-dataflow\" rel=\"nofollow noreferrer\">tutorial<\/a> for re-training inception on google cloud ml. I can run the tutorial, train, predict, just fine. <\/p>\n\n<p>I then substituted the flowers dataset for a test dataset of my own. Optical character recognition of image digits.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1nLwX.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1nLwX.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When training the model I receive the error:<\/p>\n\n<pre><\/pre>\n\n<p>The format of training and eval datasets looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>The dict file looks like this<\/p>\n\n<pre><\/pre>\n\n<p>I had originally had labels like 1,2,3,4 and \/, but I changed them to character strings in case they were special characters (especially \/). I can see a somewhat similar message <a href=\"https:\/\/stackoverflow.com\/questions\/33933846\/tensorflow-indices-are-not-valid-out-of-bounds\">here<\/a>, but that had to do with 0th based indexing. <\/p>\n\n<p>What's strange about the message is that there are indeed 13 label types. Somehow tensorflow is looking for only 7 (0-6). My questions is what kind of formatting error might make tensorflow think there are less labels then there are. I can confirm that both the testing and training data 80-20 split have all the label classes (though in different frequencies).<\/p>\n\n<p>I'm running from recent docker build provided by google.<\/p>\n\n<pre><\/pre>\n\n<p>I am submit the training job using<\/p>\n\n<pre><\/pre>\n\n<p>Full error:<\/p>\n\n<pre><\/pre>\n\n<p>Everything looks okay in my bucket<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rxZZl.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rxZZl.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And saving my log events.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/y6stu.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/y6stu.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41602040","CreationDate":"2017-01-11T22:40:10.350","ParentId":null,"OwnerUserId":"3818158","Title":null,"Body":"<p>I think you need to specify <code>--label_count 13<\/code> when you submit your training job.  That flag should go after the in the second group of flags after the <code>--<\/code>, since it needs to be passed to the code you're executing, and not to gcloud\/Cloud ML.<\/p>\n\n<p>The issue is that the TensorFlow training code needs to know how many output logits to make before it starts stepping through the data; so it cannot inspect the intermediate files from the preprocess step.<\/p>\n\n<p>Let me know if this helps.<\/p>\n"},{"AnswerId":"41618273","CreationDate":"2017-01-12T16:27:45.890","ParentId":null,"OwnerUserId":"6915676","Title":null,"Body":"<p>The flags after the -- model specific and are specified in <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/blob\/master\/flowers\/trainer\/model.py#L47\" rel=\"nofollow noreferrer\">model.py<\/a> file:<\/p>\n\n<pre><code>def create_model():\n  \"\"\"Factory method that creates model to be used by generic task.py.\"\"\"\n  parser = argparse.ArgumentParser()\n  # Label count needs to correspond to nubmer of labels in dictionary used\n  # during preprocessing.\n  parser.add_argument('--label_count', type=int, default=5)\n  parser.add_argument('--dropout', type=float, default=0.5)\n  parser.add_argument(\n      '--inception_checkpoint_file',\n      type=str,\n      default=DEFAULT_INCEPTION_CHECKPOINT)\n  args, task_args = parser.parse_known_args()\n  override_if_not_in_args('--max_steps', '1000', task_args)\n  override_if_not_in_args('--batch_size', '100', task_args)\n  override_if_not_in_args('--eval_set_size', '370', task_args)\n  override_if_not_in_args('--eval_interval_secs', '2', task_args)\n  override_if_not_in_args('--log_interval_secs', '2', task_args)\n  override_if_not_in_args('--min_train_eval_rate', '2', task_args)\n  return Model(args.label_count, args.dropout,\n               args.inception_checkpoint_file), task_args\n<\/code><\/pre>\n\n<p>Note you can change things like label_count, dropout, max_steps, etc., to affect the model training.<\/p>\n\n<p>HTH.<\/p>\n"}]}
{"QuestionId":41601660,"AnswerCount":0,"Tags":"<python><pandas><deep-learning><keras><kaggle>","CreationDate":"2017-01-11T22:13:41.533","AcceptedAnswerId":null,"OwnerUserId":657104.0,"Title":"Log loss score in training & validation very different from score on Kaggle test set","Body":"<p>I'm able to get a log loss score as low as 0.24 in training and 0.38 in validation, but once I submit my predictions to Kaggle to score on the test set, the loss is way off (sometimes as high as 4, but mostly never below 0.69). Any ideas what could be going wrong?<\/p>\n\n<p>FYI, here is my code for writing the predictions to file (using Keras &amp; Tensorflow backend):<\/p>\n\n<pre><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[]}
{"QuestionId":41602258,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-01-11T22:57:05.210","AcceptedAnswerId":41606593.0,"OwnerUserId":4525192.0,"Title":"Tokenizer error in Keras pipeline","Body":"<p>I want to use keras for authorship attribution. I have a list of (text,labels). I am trying to use the keras builtin vectorizer but I get the following error:<\/p>\n\n<blockquote>\n  <p>Vectorizing sequence data... Traceback (most recent call last):   File \"\", line 1, in    File\n  \"\/home\/angelo\/org\/courses\/corpusling\/finalproject\/src\/neuralnet.py\",\n  line 46, in \n      X_train = tokenizer.texts_to_matrix(X_train, mode='binary')   File \"\/home\/angelo\/org\/courses\/corpusling\/finalproject\/venv0\/lib\/python3.5\/site-packages\/keras\/preprocessing\/text.py\",\n  line 166, in texts_to_matrix\n      sequences = self.texts_to_sequences(texts)   File \"\/home\/angelo\/org\/courses\/corpusling\/finalproject\/venv0\/lib\/python3.5\/site-packages\/keras\/preprocessing\/text.py\",\n  line 131, in texts_to_sequences\n      for vect in self.texts_to_sequences_generator(texts):   File \"\/home\/angelo\/org\/courses\/corpusling\/finalproject\/venv0\/lib\/python3.5\/site-packages\/keras\/preprocessing\/text.py\",\n  line 150, in texts_to_sequences_generator\n      i = self.word_index.get(w) AttributeError: 'Tokenizer' object has no attribute 'word_index'<\/p>\n<\/blockquote>\n\n<p>The following is my code so far:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41606593","CreationDate":"2017-01-12T06:36:44.350","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You need to use <code>tokenizer.fit_on_texts(texts)<\/code> before using <code>tokenizer.texts_to_matrix()<\/code> <\/p>\n\n<p>here <code>texts<\/code> is the list of the the text data (both train and test). <\/p>\n\n<p><code>fit_on_texts()<\/code> uses it to build <code>word_index<\/code>. Its nothing but unique word to number mapping. And this mapping is later used to generate the matrix.<\/p>\n"}]}
{"QuestionId":41602339,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-11T23:03:08.620","AcceptedAnswerId":41604856.0,"OwnerUserId":5997806.0,"Title":"Python image classification in tensorflow .12 removed?","Body":"<p>There used to be a classify_image.py in 0.11, now the only image classification I see if for C++ in .<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/search?utf8=%E2%9C%93&amp;q=classify_image.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/search?utf8=%E2%9C%93&amp;q=classify_image.py<\/a><\/p>\n\n<p>Is the file renamed and moved? How can I classify an image in python with TF .12?<\/p>\n","answers":[{"AnswerId":"41604856","CreationDate":"2017-01-12T03:56:26.497","ParentId":null,"OwnerUserId":"3799743","Title":null,"Body":"<p>Please, note that now the models directory is a whole new repository in Tensorflow's organization.<\/p>\n\n<p>The file you want has been moved to: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/imagenet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/imagenet<\/a><\/p>\n\n<blockquote>\n  <p>Simple image classification with Inception. Run image classification\n  with Inception trained on ImageNet 2012 Challenge data set.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41602374,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-11T23:05:35.153","AcceptedAnswerId":41603661.0,"OwnerUserId":3873000.0,"Title":"tf.zeros doesn't return a 1D tensor?","Body":"<p>I'm trying to duplicate a tensor across a new axis, like this:<\/p>\n\n<pre><\/pre>\n\n<p>However I'm getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>In the documentation it says the way I wrote tf.constant is supposed to have it return a 1D tensor but when I checked its shape with get_shape(), it has (5,) as its shape. I tried reshaping it but nothing changed.<\/p>\n\n<p>Why am I getting this error? Thanks.<\/p>\n","answers":[{"AnswerId":"41603661","CreationDate":"2017-01-12T01:27:22.910","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Your \"5\" parameter has shape <code>()<\/code> but needs to have rank 1<\/p>\n\n<pre><code>original_tensor = tf.constant([1,2,3,4,5])\nmade_copies_tensor = tf.tile(original_tensor, [5])\nsess.run(made_copies_tensor)\n<\/code><\/pre>\n"}]}
{"QuestionId":41603357,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2017-01-12T00:47:00.640","AcceptedAnswerId":41605232.0,"OwnerUserId":1363671.0,"Title":"model ensemble with shared layers","Body":"<p>In keras, I want to train an ensemble of models that share some layers. They are of the following form:<\/p>\n\n<p>x ---> f(x)  ---> g_1(f(x))<\/p>\n\n<p>x ---> f(x)  ---> g_2(f(x))<\/p>\n\n<p>...<\/p>\n\n<p>x ---> f(x)  ---> g_n(f(x))<\/p>\n\n<p>here f(x) are some nontrivial shared layers.  g_1 through g_n have their specific parameters.<\/p>\n\n<p>At each training stage, data x, is fed into one of the n networks, say, the i-th. A loss on g_i(f(x)) is then minimized\/decreased via gradient based optimizer. How could I define and train such a model?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"41605232","CreationDate":"2017-01-12T04:38:07.683","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can easily do so by using functional Model. <\/p>\n\n<p>A small example .. you can build on it:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import Model\nfrom keras.layers import Dense, Input\n\nX = np.empty(shape=(1000,100))\nY1 = np.empty(shape=(1000))\nY2 = np.empty(shape=(1000,2))\nY3 = np.empty(shape=(1000,3))\n\ninp = Input(shape=(100,))\ndense_f1 = Dense(50)\ndense_f2 = Dense(20)\n\nf = dense_f2(dense_f1(inp))\n\ndense_g1 = Dense(1)\ng1 = dense_g1(f)\n\ndense_g2 = Dense(2)\ng2 = dense_g2(f)\n\ndense_g3 = Dense(3)\ng3 = dense_g3(f)\n\n\nmodel = Model([inp], [g1, g2, g3])\nmodel.compile(loss=['mse', 'binary_crossentropy', 'categorical_crossentropy'], optimizer='rmsprop')\n\nmodel.summary()\n\nmodel.fit([X], [Y1, Y2, Y3], nb_epoch=10)\n<\/code><\/pre>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>Based on your comments, you can always make different models and write the training loop yourself based on how you need your training. You can see in the <code>model.summary()<\/code> all the models are sharing the initial layers. Here is the extension to the example<\/p>\n\n<pre><code>model1 = Model(inp, g1)\nmodel1.compile(loss=['mse'], optimizer='rmsprop')\nmodel2 = Model(inp, g2)\nmodel2.compile(loss=['binary_crossentropy'], optimizer='rmsprop')\nmodel3 = Model(inp, g3)\nmodel3.compile(loss=['categorical_crossentropy'], optimizer='rmsprop')\nmodel1.summary()\nmodel2.summary()\nmodel3.summary()\n\nbatch_size = 10\nnb_epoch=10\nn_batches = X.shape[0]\/batch_size\n\n\nfor iepoch in range(nb_epoch):\n    for ibatch in range(n_batches):\n        x_batch = X[ibatch*batch_size:(ibatch+1)*batch_size]\n        if ibatch%3==0:\n            y_batch = Y1[ibatch*batch_size:(ibatch+1)*batch_size]\n            model1.train_on_batch(x_batch, y_batch)      \n        elif ibatch%3==1:\n            y_batch = Y2[ibatch*batch_size:(ibatch+1)*batch_size]\n            model2.train_on_batch(x_batch, y_batch)      \n        else:\n            y_batch = Y3[ibatch*batch_size:(ibatch+1)*batch_size]\n            model3.train_on_batch(x_batch, y_batch)      \n<\/code><\/pre>\n"}]}
{"QuestionId":41603462,"AnswerCount":1,"Tags":"<tensorflow><mxnet>","CreationDate":"2017-01-12T00:59:09.103","AcceptedAnswerId":null,"OwnerUserId":1558203.0,"Title":"Can I run a model trained using tensorflow on mxnet?","Body":"<p>I have models trained on tensorflow. Can I use mxnet in forward only mode to run these ?\n<a href=\"https:\/\/github.com\/dmlc\/nnvm\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dmlc\/nnvm<\/a> says this should be possible in future, but is the support available today ?<\/p>\n","answers":[{"AnswerId":"44207438","CreationDate":"2017-05-26T17:46:07.393","ParentId":null,"OwnerUserId":"3451801","Title":null,"Body":"<p>MXNet doesn't have tensorflow model converter yet. It does have a caffe-to-mxnet converter. So you can convert your tf model to caffe, that would work.. <\/p>\n\n<p><a href=\"https:\/\/github.com\/dmlc\/mxnet\/tree\/master\/tools\/caffe_converter\" rel=\"nofollow noreferrer\">https:\/\/github.com\/dmlc\/mxnet\/tree\/master\/tools\/caffe_converter<\/a><\/p>\n"}]}
{"QuestionId":41603965,"AnswerCount":1,"Tags":"<tensorflow><neural-network><recurrent-neural-network>","CreationDate":"2017-01-12T02:04:30.097","AcceptedAnswerId":null,"OwnerUserId":6736078.0,"Title":"TensorFlow -- Understanding Embeddings in translate.py","Body":"<p>My understanding was that translate.py implements an encoder-decoder model, one step of which is translating sentences from the input language to fixed-length vetors.  From the post at <a href=\"https:\/\/stackoverflow.com\/questions\/35508356\/translate-py-doesnt-use-embedding\">translate.py doesnt use embedding?<\/a> and my own investigation of the code, I thought the sentence embeddings would be stored in embedding_attention_seq2seq\/RNN\/EmbeddingWrapper\/embedding:0.  But this is a variable of a shape corresponding to my input vocabulary size X rnn layer size.  I don't understand how I can conceptually interpret it as encodings of sentences.  Where's the error in my understanding?<\/p>\n","answers":[{"AnswerId":"44189091","CreationDate":"2017-05-25T20:14:02.063","ParentId":null,"OwnerUserId":"1726622","Title":null,"Body":"<p>Isn't the embedding layer's named <strong>seq2seq\/embedding_attention_seq2seq\/RNN\/EmbeddingWrapper\/embedding:0<\/strong> <\/p>\n\n<p>you can show all variables with :<\/p>\n\n<pre><code>for var in tf.trainable_variables():\n  print var.name\n<\/code><\/pre>\n\n<p>in the <strong>embedding_attention_seq2seq<\/strong>() function definition<\/p>\n"}]}
{"QuestionId":41604112,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-12T02:22:39.843","AcceptedAnswerId":null,"OwnerUserId":6878563.0,"Title":"Tensorflow Custom Cost Function","Body":"<p>I have 2 concrete classes, say A and C.  I want to use an NN to classify them into classes A, B, C, such that samples that are too close to confidently classify are just classed as B.  The cost function should be as follows:  A misclassification, (an A classified as C, or vice versa) will have a very large cost.  A correct classification will have zero cost.  Classifying an item as B will have very low cost.  The result is that we only distinguish samples that we are VERY SURE fit into their respective classes.  <\/p>\n\n<p>I have only worked through the simple tutorials in TensorFlow, but it didn't cover how to define more specific costs functions such as this.  Can anyone explain how this can be accomplished in TensorFlow<\/p>\n\n<p>Here is my relevant code, where I currently classify using only 2 classes. It is straight from the TensorFlow tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>y is the output of the NN (will look like [[1,0,0],[0,1,0]] for a two sample set with 3 classes), and y_ is the correct classes for the sample which might be [[1,0,0],[0,0,1]].  In this example, we would have classified the second sample as B because we were uncertain, but the true class was C. <\/p>\n","answers":[{"AnswerId":"41624461","CreationDate":"2017-01-12T22:50:23.507","ParentId":null,"OwnerUserId":"302268","Title":null,"Body":"<p>I think you have some fundamental misunderstanding of how NN classifiers work. You should probably read up on that a bit if you are going to go deeper into coding them. I highly recommend online book by Michael Nielsen <a href=\"http:\/\/neuralnetworksanddeeplearning.com\/\" rel=\"nofollow noreferrer\">Neural Networks and Deep Learning<\/a>.<\/p>\n\n<p>That said, the solution you are looking for is not in creating a special cost function, but in how you interpret the results you get from NN. You do not have 3 classes, you have 2. The \"I have no idea what this is\" is not a class by itself, but rather a measure of NN's confidence in its answer. So, your network should have 2 outputs, one for each class, just like in the TendorFlow guides. And you should train it just like in the guides. Once your network is trained, when you feed it a sample to classify, you get 2 numbers, lets call them A' and C'. These numbers indicate NN's confidence in what class the sample belongs to. For example, if you get A' == 0.999 and C' == 0.00001, the network is pretty damn sure that your sample is class A. If you get A' == 0.6 and C' == 0.59, your network has no idea if the sample is A or C, but slightly favors the theory that it is class A. It is now up to you to say what your confidence intervals are. To make this easier, you should probably use softmax for the output layer non-linearity (the way TensorFlow MNIST guides do). One of the useful features of softmax is that the sum of all your classes will always be 1 and you can easily make decisions based on the difference between A' and C'.<\/p>\n"}]}
{"QuestionId":41604253,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-12T02:40:31.093","AcceptedAnswerId":null,"OwnerUserId":4407207.0,"Title":"How to fine tune the hyper-parameters of ftrl optimizer in tensorflow","Body":"<p>I found the initial parameters of FTRL optimizer is:<\/p>\n\n<pre><\/pre>\n\n<p>But how could I fine-tune the parameters to make it better?<\/p>\n","answers":[{"AnswerId":"42110983","CreationDate":"2017-02-08T10:46:33.237","ParentId":null,"OwnerUserId":"4206999","Title":null,"Body":"<p>You should perform hyper parameter tuning - Keep a test dataset on the side, and then split your train dataset to train and cross validation sets.<\/p>\n\n<p>You can then use the validation dataset for tuning these parameters, or directly perform this on the train dataset. You could do grid search over all parameters, or any other method in <a href=\"https:\/\/en.wikipedia.org\/wiki\/Hyperparameter_optimization\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Hyperparameter_optimization<\/a><\/p>\n\n<p>You simply pick the parameters with the best performance<\/p>\n"}]}
{"QuestionId":41604484,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><computer-vision><neural-network><conv-neural-network>","CreationDate":"2017-01-12T03:11:00.537","AcceptedAnswerId":null,"OwnerUserId":7361100.0,"Title":"CNN Object Localization Preprocessing?","Body":"<p>I'm trying to use a pretrained VGG16 as an object localizer in Tensorflow on ImageNet data. In their paper, the group mentions that they basically just strip off the softmax layer and either toss on a 4D\/4000D fc layer for bounding box regression. I'm not trying to do anything fancy here (sliding windows, RCNN), just get some mediocre results.<\/p>\n\n<p>I'm sort of new to this and I'm just confused about the preprocessing done here for localization. In the paper, they say that they scale the image to 256 as its shortest side, then take the central 224x224 crop and train on this. I've looked all over and can't find a simple explanation on how to handle localization data.<\/p>\n\n<p>Questions: How do people usually handle the bounding boxes here?... <\/p>\n\n<ul>\n<li>Do you use something like the tf.sample_distorted_bounding_box command, and then rescale the image based on that?<\/li>\n<li>Do you just rescale\/crop the image itself, and then interpolate the bounding box with the transformed scales? Wouldn't this result in negative box coordinates in some cases?<\/li>\n<li><p>How are multiple objects per image handled?<\/p>\n\n<ul>\n<li><p>Do you just choose a single bounding box from the beginning ,crop to that, then train on this crop?<\/p><\/li>\n<li><p>Or, do you feed it the whole (centrally cropped) image, and then try to predict 1 or more boxes somehow?<\/p><\/li>\n<\/ul><\/li>\n<li>Does any of this generalize to the Detection or segmentation (like MS-CoCo) challenges, or is it completely different?<\/li>\n<\/ul>\n\n<p>Anything helps...\nThanks<\/p>\n","answers":[{"AnswerId":"41611808","CreationDate":"2017-01-12T11:15:38.117","ParentId":null,"OwnerUserId":"2444909","Title":null,"Body":"<p>Localization is usually performed as an intersection of sliding windows where the network identifies the presence of the object you want.<\/p>\n\n<p>Generalizing that to multiple objects works the same.<\/p>\n\n<p>Segmentation is more complex. You can train your model on a pixel mask with your object filled, and you try to output a pixel mask of the same size<\/p>\n"}]}
{"QuestionId":41604616,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-12T03:27:45.700","AcceptedAnswerId":42361617.0,"OwnerUserId":3444294.0,"Title":"Save and load Tensorflow model","Body":"<p>I want to save a Tensorflow (0.12.0) model, including graph and variable values, then later load and execute it. I have the read the docs and other posts on this but cannot get the basics to work. I am using the technique from <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/how_tos\/meta_graph\/\" rel=\"nofollow noreferrer\">this page in the Tensorflow docs<\/a>. Code:<\/p>\n\n<p>Save a simple model:<\/p>\n\n<pre><\/pre>\n\n<p>Later, load and execute the model:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Question 1:<\/strong>  The saving code seems to work but the loading code produces this error:<\/p>\n\n<pre><\/pre>\n\n<p>How to fix this?.<\/p>\n\n<p><strong>Question 2:<\/strong> I included this line to follow the pattern in the TF docs...<\/p>\n\n<pre><\/pre>\n\n<p>... but why is that line necessary? Doesn't export the entire graph by default? If not then does one need to add every variable in the graph to the collection before saving? Or do we just add to the collection those variables that will be accessed after the restore?<\/p>\n\n<p>---------------------- <em>Update January 12 2017<\/em> -----------------------------<\/p>\n\n<p>Partial success based on Kashyap's suggestion below but a mystery still exists. The code below works <em>but<\/em> only if I include the lines containing  and . Without those lines, 'load' mode throws an error in the last line:\n. My understanding was that by default  saves and restores all variables in the graph, so why is it necessary to specify the name of variables that will be used in the collection? I assume this has to do with mapping Tensorflow's variable names to Python names, but what are the rules of the game here? For which variables does this need to be done?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42361617","CreationDate":"2017-02-21T07:56:51.483","ParentId":null,"OwnerUserId":"4993842","Title":null,"Body":"<p>I've been trying to figure out the same thing and was able to successfully do it by using <code>Supervisor<\/code>. It automatically loads all variables and your graph etc. Here is the documentation - <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/supervisor\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/supervisor<\/a>. Below is my code -<\/p>\n\n<pre><code>sv = tf.train.Supervisor(logdir=\"\/checkpoint', save_model_secs=60)\n    with sv.managed_session() as sess:\n        if not sv.should_stop(): \n            #Do run\/eval\/train ops on sess as needed. Above works for both saving and loading\n<\/code><\/pre>\n\n<p>As you see, this is much simpler than using the <code>Saver<\/code> object and dealing with individual variables etc as long as the graph stays the same (my understanding is that <code>Saver<\/code> comes handy when we want to reuse a pre-trained model for a different graph).<\/p>\n"},{"AnswerId":"41617748","CreationDate":"2017-01-12T16:02:29.857","ParentId":null,"OwnerUserId":"2704763","Title":null,"Body":"<p><strong><em>Question1<\/em><\/strong><\/p>\n\n<p>This question has been already answered thoroughly <a href=\"https:\/\/stackoverflow.com\/questions\/33759623\/tensorflow-how-to-restore-a-previously-saved-model-python\">here<\/a>. You don't have to explicitly call <code>export_meta_graph<\/code>. Call the <code>save method<\/code>. This will generate the <code>.meta<\/code> file also (since save method will call the <code>export_meta_graph<\/code> method internally.)<\/p>\n\n<p>For example <\/p>\n\n<p><code>saver0.save(sess, '.\/myModel.ckpt')<\/code><\/p>\n\n<p>will produce <code>myModel.ckpt<\/code> file and also the <code>myModel.ckpt.meta<\/code> file.<\/p>\n\n<p>Then you can restore the model using <\/p>\n\n<pre><code>with tf.Session() as sess:\n    saver1 = tf.train.import_meta_graph('.\/myModel.ckpt.meta')\n    saver1.restore(sess, '.\/myModel')\n    print sess.run(myVar)\n<\/code><\/pre>\n\n<p><strong><em>Question2<\/em><\/strong><\/p>\n\n<p>Collections are used to store custom  information like learning rate,the regularisation factor that you have used and other information and these will be stored when you export the graph. Tensorflow itself defines some collections like \"TRAINABLE_VARIABLES\" which are used to get all the trainable variables of the model you built. You can chose to export all the collections in your graph or you can specify which collections to export in the <code>export_meta_graph<\/code> function. <\/p>\n\n<p>Yes tensorflow will export all the variables that you define. But if you need any other information that needs to be exported to the graph then they can be added to the collection.<\/p>\n"}]}
{"QuestionId":41604666,"AnswerCount":1,"Tags":"<python><csv><tensorflow>","CreationDate":"2017-01-12T03:33:14.113","AcceptedAnswerId":41604790.0,"OwnerUserId":6815505.0,"Title":"where and how to put the filename in this tensorflow code?","Body":"<p>I've copied a tensorflow code from Sirajology's github. It should load a .csv into a single layer neural net. <\/p>\n\n<p>My question is, how and where do I put the .csv file into the code? <\/p>\n\n<p>Also, I don't understand if the code will automatically split the .csv into train and test data, or if I need to do that with some different code before it is fed to the neural net?<\/p>\n\n<p>I've been spending a lot of time with python and tensorflow and understand some basic concepts but am still a total newb. Any help is appreciated! Thanks!!!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41604790","CreationDate":"2017-01-12T03:47:50.993","ParentId":null,"OwnerUserId":"3799743","Title":null,"Body":"<p>It's expecting to receive it as argument in the terminal.<\/p>\n\n<p>The lines below are checking for it:<\/p>\n\n<pre><code>tf.app.flags.DEFINE_string('train', None,\n                       'File containing the training data (labels &amp; features).')\ntf.app.flags.DEFINE_string('test', None,\n                       'File containing the test data (labels &amp; features).')\ntf.app.flags.DEFINE_integer('num_epochs', 1,\n                        'Number of examples to separate from the training '\n                        'data for the validation set.')\n<\/code><\/pre>\n\n<p>So, you just have to run it as:<\/p>\n\n<pre><code>python YourScript.py --train FileName.csv --test TestName.csv --num_epochs 5 --verbose True\n<\/code><\/pre>\n"}]}
{"QuestionId":41604686,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-12T03:35:07.357","AcceptedAnswerId":41619982.0,"OwnerUserId":1104696.0,"Title":"How to use tf.while_loop() for variable-length inputs in tensorflow?","Body":"<p>I am trying to use  to process variable-length inputs. However, I can only use it for fixed length. The code no longer works after I change  to .  seems to handle variable-length inputs. I am not sure how  achieves this with .<\/p>\n\n<pre><\/pre>\n\n<p>output (before change):<\/p>\n\n<pre><\/pre>\n\n<p>output (after change):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41619982","CreationDate":"2017-01-12T18:01:57.173","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>It works if you remove shapes from all the variables:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nconfig = tf.ConfigProto(graph_options=tf.GraphOptions(\n  optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))\ntf.reset_default_graph()\nsess = tf.Session(\"\", config=config)\n#initial_m = tf.Variable(0.0, name='m')\n\n#The code no longer works after I change shape=(4) to shape=(None)\ninputs = tf.placeholder(dtype='float32', shape=(None)) \ntime_steps = tf.shape(inputs)[0]\ninitial_outputs = tf.TensorArray(dtype=tf.float32, size=time_steps)\ninitial_t = tf.placeholder(dtype='int32')\ninitial_m = tf.placeholder(dtype=tf.float32)\n\ndef should_continue(t, *args):\n    return t &lt; time_steps\n\ndef iteration(t, m, outputs_):\n    cur = tf.gather(inputs, t)\n    m  = m * 0.5 + cur * 0.5\n    outputs_ = outputs_.write(t, m)\n    return t + 1, m, outputs_\n\nt, m, outputs = tf.while_loop(should_continue, iteration,\n                              [initial_t, initial_m, initial_outputs])\n\noutputs = outputs.stack()\ninit = tf.global_variables_initializer()\nsess.run([init])\nprint(sess.run([outputs],\n               feed_dict={inputs: np.asarray([1, 1, 1, 1]), initial_t: 0,\n                          initial_m: 0.}))\n<\/code><\/pre>\n"},{"AnswerId":"52642944","CreationDate":"2018-10-04T09:04:21.767","ParentId":null,"OwnerUserId":"4576234","Title":null,"Body":"<p><code>tf.while_loop<\/code> has a keyword parameter called <code>shape_invariants<\/code>. There you can give a list with all tensors' expected shapes, so you can specify which dimensions should remain the same and which are of variable size. For example, you can build your loop like this:<\/p>\n\n<pre><code>t, final_tensor = tf.while_loop(should_continue, iteration,\n                                [t0, initial_tensor],\n                                [t0.get_shape(), tf.TensorShape([None,10])])\n<\/code><\/pre>\n\n<p>This would allow the <code>initial_tensor<\/code> to change in its first dimension, as long as the second dimension remains of size 10.<\/p>\n"}]}
{"QuestionId":41605182,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-12T04:32:45.433","AcceptedAnswerId":null,"OwnerUserId":7407833.0,"Title":"tf.train.SequenceExample with lists at each step","Body":"<p>How is a list of values passed into a feature_list?  The documentation suggests that this is valid but it is not clear how this is accomplished given that passing a list result in an error.<\/p>\n\n<pre><\/pre>\n\n<p>This is an example given in the example proto file.\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto<\/a><\/p>\n\n<pre><\/pre>\n\n<p>Is it necessary to use something other than append when adding a list?<\/p>\n\n<p>An example of a sequence would be...<\/p>\n\n<pre><\/pre>\n\n<p>...where there are four steps in this sequence, and at each step there is a set of values.  The desired result would look something like the example below.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"45355407","CreationDate":"2017-07-27T15:43:05.717","ParentId":null,"OwnerUserId":"4101573","Title":null,"Body":"<p>Use extend instead of append.<\/p>\n\n<pre><code>tf.train.SequenceExample().feature_lists.feature_list[\"multiple\"].feature.add().int64_list.value.extend([1,2,3,4])\n<\/code><\/pre>\n"}]}
{"QuestionId":41605355,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><pip>","CreationDate":"2017-01-12T04:53:41.067","AcceptedAnswerId":null,"OwnerUserId":7407893.0,"Title":"Updating the supported tags for pip","Body":"<p>I'm trying to install Tensorflow, and received the following error.<\/p>\n\n<blockquote>\n  <p>tensorflow-0.12.1-cp35-cp35m-win_amd64.whl is not a supported wheel on this platform.<\/p>\n<\/blockquote>\n\n<p>By reading through other questions, I think I've traced the issue to the cp35 tag not being supported by the version of pip I have installed. What's odd is that I believe I installed python 3.5 and the latest version of pip (9.0.1), but have the following supported tags:<\/p>\n\n<blockquote>\n  <p>[('cp27', 'cp27m', 'win_amd64'), ('cp27', 'none', 'win_amd64'), ('py2', 'none', 'win_amd64'), ('cp27', 'none', 'any'), ('cp2', 'none', 'any'), ('py27', 'none', 'any'), ('py2', 'none', 'any'), ('py26', 'none', 'any'), ('py25', 'none', 'any'), ('py24', 'none', 'any'), ('py23', 'none', 'any'), ('py22', 'none', 'any'), ('py21', 'none', 'any'), ('py20', 'none', 'any')]<\/p>\n<\/blockquote>\n\n<p>How can I go about modifying the supported tags, or is that even the right approach?<\/p>\n","answers":[{"AnswerId":"42087865","CreationDate":"2017-02-07T10:49:23.190","ParentId":null,"OwnerUserId":"7528245","Title":null,"Body":"<p>I have the same error when I run this command. I found error that the installed version of python was x86 and TensorFlow is for x64 versions. I reinstalled the python with x64 version and it works now! I hope this works for you too!<\/p>\n"},{"AnswerId":"42066167","CreationDate":"2017-02-06T11:01:05.290","ParentId":null,"OwnerUserId":"766551","Title":null,"Body":"<p>If you're having this problem on windows then it's most likely that you have python 3.6.<\/p>\n\n<p>Tensorflow for python 3.6 on windows is not available atm unless you want to go through some hassle.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6999\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6999<\/a>\n\"In the meantime, if you need the pip package for windows for a python version other than py 3.5, you will need to build from sources. We are sorry for the inconvenience.\"<\/p>\n\n<p>If you wish to use Anaconda Python then tensorflow page is leading to the python 3.6 download which won't work. Python 3.5 Anaconda you can get from the Anaconda archive <a href=\"https:\/\/repo.continuum.io\/archive\/\" rel=\"nofollow noreferrer\">https:\/\/repo.continuum.io\/archive\/<\/a> or use the direct link\n<a href=\"https:\/\/repo.continuum.io\/archive\/Anaconda3-4.2.0-Windows-x86_64.exe\" rel=\"nofollow noreferrer\">https:\/\/repo.continuum.io\/archive\/Anaconda3-4.2.0-Windows-x86_64.exe<\/a><\/p>\n"}]}
{"QuestionId":41605636,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-12T05:20:41.383","AcceptedAnswerId":42641032.0,"OwnerUserId":1050648.0,"Title":"multiple runs of inference model using saved checkpoint produces stochastic errors - Tensorflow","Body":"<p>I am running Tensorflow 0.12.1 on a GPU. I have a trained Deep CNN model whose weights I've saved using a  checkpoint file. During inference, I reload the saved checkpoint using .  The code seems to run without issues, but everytime I re-run the script, I'm getting screwed up outputs. AFAIK, I do not shuffle my test set inputs. The inputs are being loaded and fed to the network properly. It is just the output of different runs of the CNN on the same test set using the same order is producing very different outputs.  I'm perplexed! Also, how do I execute a graph loaded with saved checkpoint without running an  during inference? It seems my code requires all global and local variables to be initialized before execution. (I initialize first, and then only restore the checkpoint!).Here's a snippet of my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42641032","CreationDate":"2017-03-07T05:43:28.800","ParentId":null,"OwnerUserId":"1277928","Title":null,"Body":"<p>Are you running on same machine or different machine\n            #saver = tf.train.Saver() <\/p>\n\n<p>The following comment is in tensorflow docs \n            #NOTE: Restarting training from saved meta_graph only works if the device assignments have not changed.\n            #saver = tf.train.import_meta_graph(metafile)<\/p>\n"}]}
{"QuestionId":41606031,"AnswerCount":1,"Tags":"<ubuntu><cpu><caffe>","CreationDate":"2017-01-12T05:54:41.690","AcceptedAnswerId":null,"OwnerUserId":6907470.0,"Title":"caffe\/include\/caffe\/util\/device_alternate.hpp:35:23: fatal error: cublas_v2.h: No such file or directory","Body":"<p>I have installed caffe and opencv 2.4 on my ubuntu 14.04 machine without CUDA and GPU support. Also I have uncommented CPU_ONLY := 1 in makefile.config file.\nI want to use caffe on cpu mode (on c++ platform). But I am getting error while building classification example file on codelite platform.<\/p>\n\n<p>The error is:<\/p>\n\n<blockquote>\n  <p>\/caffe\/include\/caffe\/util\/device_alternate.hpp:35:23: fatal error: cublas_v2.h: No such file or directory<\/p>\n<\/blockquote>\n\n<p>I am not getting if I am wrong somewhere in installation or I have to make some more changes for CPU only mode.<\/p>\n","answers":[{"AnswerId":"52514849","CreationDate":"2018-09-26T09:51:22.880","ParentId":null,"OwnerUserId":"5056262","Title":null,"Body":"<p>You should check out <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2704\" rel=\"nofollow noreferrer\">this<\/a> issue posted on the caffe github repo. The solution that worked for me was adding <code>add_definitions( -DCPU_ONLY=1 )<\/code> at the second line of CMakeLists.txt<\/p>\n"}]}
{"QuestionId":41606854,"AnswerCount":1,"Tags":"<python><linux><tensorflow>","CreationDate":"2017-01-12T06:53:21.927","AcceptedAnswerId":41610033.0,"OwnerUserId":7238189.0,"Title":"TensorFlow Permission Denied Error \/Location","Body":"<p>Following <a href=\"https:\/\/github.com\/llSourcell\/tensorflow_demo\" rel=\"noreferrer\">this tutorial<\/a> when I run board.py in my terminal I get this error on Ubuntu 16.04:<\/p>\n\n<pre><\/pre>\n\n<p>I'm a relatively new Linux user\/programmer so I'm not entirely sure what to do here\/look up as google hasn't found me the answer. So my two questions are:<\/p>\n\n<ol>\n<li>How should I go about resolving this<\/li>\n<li>In the future how should I go about troubleshooting errors similar to this as to become more independent<\/li>\n<\/ol>\n","answers":[{"AnswerId":"41610033","CreationDate":"2017-01-12T09:57:13.923","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p>Change <code>\/LOCATION\/ON\/YOUR\/COMPUTER\/<\/code> in <a href=\"https:\/\/github.com\/llSourcell\/tensorflow_demo\/blob\/master\/board.py#L56\" rel=\"nofollow noreferrer\">line 56<\/a> to the correct location where the user has write permissions, for example, <code>.\/summaries<\/code> (<code>summaries<\/code> folder in the current working directory).<\/p>\n\n<p><code>\/<\/code> is owned by root and typical users have no write permissions there.<\/p>\n"}]}
{"QuestionId":41607144,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-01-12T07:12:28.070","AcceptedAnswerId":41646443.0,"OwnerUserId":3792504.0,"Title":"Loading two models from Saver in the same Tensorflow session","Body":"<p>I have two networks: a  which generates output and an  which grades the output.<\/p>\n\n<p>Both have been trained separately but now I need to combine their outputs during a single session.<\/p>\n\n<p>I've attempted to implement the solution proposed in this post: <a href=\"https:\/\/stackoverflow.com\/questions\/39175945\/run-multiple-pre-trained-tensorflow-nets-at-the-same-time\">Run multiple pre-trained Tensorflow nets at the same time<\/a><\/p>\n\n<p><strong>My code<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>The problem<\/strong><\/p>\n\n<p>The call to the function  appears to be doing nothing. In another module I use a saver with  and it restores the checkpoint fine.<\/p>\n\n<p>The model has . To check what was happening I used  to extract the  before and after restore. Each time the initial randomly assigned variables are being used and the variables from the checkpoint are not being assigned.<\/p>\n\n<p>Any thoughts on why  appears to be doing nothing?<\/p>\n","answers":[{"AnswerId":"54385492","CreationDate":"2019-01-27T06:01:37.147","ParentId":null,"OwnerUserId":"10973820","Title":null,"Body":"<p>Please check this:<\/p>\n\n<pre><code>adv_varlist = {v.name.lstrip(\"avd\/\")[:-2]: v \n<\/code><\/pre>\n\n<p>It should be \"adv\" , ain't it<\/p>\n"},{"AnswerId":"58898930","CreationDate":"2019-11-17T08:57:39.867","ParentId":null,"OwnerUserId":"7418311","Title":null,"Body":"<p>The answer marked as correct does not tell us how to load two different models into one session explicitly, here is my answer:<\/p>\n\n<ol>\n<li><p>create two different name scopes for the models you want to load.<\/p><\/li>\n<li><p>initialize two savers which are going to load parameters for variables in the two different networks.<\/p><\/li>\n<li><p>load from the corresponding checkpoint files.<\/p><\/li>\n<\/ol>\n\n<pre><code>with tf.Session() as sess:\n    with tf.name_scope(\"net1\"):\n      net1 = Net1()\n    with tf.name_scope(\"net2\"):\n      net2 = Net2()\n\n    net1_varlist = {v.op.name.lstrip(\"net1\/\"): v\n                    for v in tf.get_collection(tf.GraphKeys.VARIABLES, scope=\"net1\/\")}\n    net1_saver = tf.train.Saver(var_list=net1_varlist)\n\n    net2_varlist = {v.op.name.lstrip(\"net2\/\"): v\n                    for v in tf.get_collection(tf.GraphKeys.VARIABLES, scope=\"net2\/\")}\n    net2_saver = tf.train.Saver(var_list=net2_varlist)\n\n    net1_saver.restore(sess, \"net1.ckpt\")\n    net2_saver.restore(sess, \"net2.ckpt\")\n<\/code><\/pre>\n"},{"AnswerId":"41646443","CreationDate":"2017-01-14T03:15:24.120","ParentId":null,"OwnerUserId":"3792504","Title":null,"Body":"<p>Solving this problem took a long time so I'm posting my likely imperfect solution in case anyone else needs it.<\/p>\n\n<p>To diagnose the problem I manually looped through each of the variables and assigned them one by one. Then I noticed that after assigning the variable the name would change. This is described here: <a href=\"https:\/\/stackoverflow.com\/questions\/34112202\/tensorflow-checkpoint-save-and-read\">TensorFlow checkpoint save and read<\/a><\/p>\n\n<p>Based on the advice in that post I ran each of the models in their own graphs. It also means that I had to run each graph in its own session. This meant handling the session management differently.<\/p>\n\n<p>First I created two graphs<\/p>\n\n<pre><code>model_graph = tf.Graph()\nwith model_graph.as_default():\n    model = Model(args)\n\nadv_graph = tf.Graph()\nwith adv_graph.as_default():\n    adversary = Adversary(adv_args)\n<\/code><\/pre>\n\n<p>Then two sessions<\/p>\n\n<pre><code>adv_sess = tf.Session(graph=adv_graph)\nsess = tf.Session(graph=model_graph)\n<\/code><\/pre>\n\n<p>Then I initialised the variables in each session and restored each graph separately<\/p>\n\n<pre><code>with sess.as_default():\n    with model_graph.as_default():\n        tf.global_variables_initializer().run()\n        model_saver = tf.train.Saver(tf.global_variables())\n        model_ckpt = tf.train.get_checkpoint_state(args.save_dir)\n        model_saver.restore(sess, model_ckpt.model_checkpoint_path)\n\nwith adv_sess.as_default():\n    with adv_graph.as_default():\n        tf.global_variables_initializer().run()\n        adv_saver = tf.train.Saver(tf.global_variables())\n        adv_ckpt = tf.train.get_checkpoint_state(adv_args.save_dir)\n        adv_saver.restore(adv_sess, adv_ckpt.model_checkpoint_path)\n<\/code><\/pre>\n\n<p>From here whenever each session was needed I would wrap any <code>tf<\/code> functions in that session with <code>with sess.as_default():<\/code>. At the end I manually close the sessions<\/p>\n\n<pre><code>sess.close()\nadv_sess.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":41607155,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-12T07:13:06.330","AcceptedAnswerId":41608016.0,"OwnerUserId":6430408.0,"Title":"You must feed a value for placeholder tensor 'Placeholder' with dtype float","Body":"<p>I'm a newer to tensorflow, I really don't know how to solve the problem.<\/p>\n\n<p>The code is like:<\/p>\n\n<ol>\n<li><p>Feed the train with values:<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Use the value in CNN:<\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n\n<p>Then have the error<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I print the input valuetypes using  and the result is  and .<\/p>\n\n<p>I really don't know why this error is happening.<\/p>\n\n<hr>\n\n<p>The code format<\/p>\n\n<p>First: <\/p>\n\n<pre><\/pre>\n\n<p>Second:<\/p>\n\n<pre><\/pre>\n\n<p>Third is the session:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41608016","CreationDate":"2017-01-12T08:08:56.810","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>Some questions<\/p>\n\n<p>first<br>\nwhy you use <code>sess = tf.InteractiveSession()<\/code> and <code>with tf.Session() as sess:<\/code> at same time, just curious<\/p>\n\n<p>second\nwhat is your placeholder name <code>x<\/code> or <code>images<\/code>?<br>\nif name is <code>x<\/code>, <code>{images: x_data...}<\/code> won't feed <code>x_data<\/code> to <code>x<\/code>, it override(?) <code>images<\/code><br>\nI think feed_dict should be <code>{x: x_data...}<\/code>  <\/p>\n\n<p>if name is <code>images<\/code>,do you have two <code>images<\/code> in your program, <code>placeholder<\/code> and <code>shuffle data<\/code>, try to modify name of variable<\/p>\n"},{"AnswerId":"46124495","CreationDate":"2017-09-08T21:01:22.360","ParentId":null,"OwnerUserId":"3380501","Title":null,"Body":"<p>I saw one problem with the code. There are two variables with the same name <code>label<\/code>. One of them refers to a Tensor, and the other one refers to some data. When you set <code>label: label<\/code> in the <code>feed_dict<\/code>, you need to distinguish between the two variables.\nMaybe you can try changing the name for one of the variables?<\/p>\n"}]}
{"QuestionId":41607160,"AnswerCount":1,"Tags":"<c><image><torch>","CreationDate":"2017-01-12T07:13:18.300","AcceptedAnswerId":41676407.0,"OwnerUserId":1069497.0,"Title":"How to add image channel with torch?","Body":"<p>I have loaded a jpg image with 3 channels, how to add one last channel to generate an image tensor with 4 channels?<\/p>\n","answers":[{"AnswerId":"41676407","CreationDate":"2017-01-16T12:24:56.740","ParentId":null,"OwnerUserId":"6076729","Title":null,"Body":"<p>Let's say you have an image with 3 channels, height <code>height<\/code> and width <code>width<\/code>:<\/p>\n\n<pre><code>a = torch.Tensor(3, height, width) -- this is your image\nb = torch.Tensor(1, height, width) -- the channel you want to add\nc = torch.cat(a,b,1)\n<\/code><\/pre>\n\n<p>A working example:<\/p>\n\n<pre><code>th&gt; a = torch.Tensor(3,3,3):fill(1)                                                               \nth&gt; b = torch.Tensor(1,3,3):fill(0)\nth&gt; c = torch.cat(a,b,1)                                                                   \nth&gt; c\n(1,.,.) = \n  1  1  1\n  1  1  1\n  1  1  1\n\n(2,.,.) = \n  1  1  1\n  1  1  1\n  1  1  1\n\n(3,.,.) = \n  1  1  1\n  1  1  1\n  1  1  1\n\n(4,.,.) = \n  0  0  0\n  0  0  0\n  0  0  0\n[torch.DoubleTensor of size 4x3x3]\n<\/code><\/pre>\n"}]}
{"QuestionId":41607778,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-01-12T07:54:07.227","AcceptedAnswerId":null,"OwnerUserId":6306884.0,"Title":"How to use tf.nn.top_k with dimension None","Body":"<p>I want to use  to replace ()<br>\nbut it seem that  doesn't accept  dimension<br>\nhere is my code<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41627383","CreationDate":"2017-01-13T04:28:28.657","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>Find the answer from <a href=\"https:\/\/stackoverflow.com\/a\/39834890\/6306884\">here<\/a><br>\ndimension still None but it works, surprised<\/p>\n\n<pre><code>val = tf.gather(val, keep) # shape is (?,)\nshape_list = tf.unpack(tf.shape(val))\nargsort = tf.nn.top_k(val, shape_list[0]) # it works\n<\/code><\/pre>\n"}]}
{"QuestionId":41607987,"AnswerCount":0,"Tags":"<c++><caffe>","CreationDate":"2017-01-12T08:07:50.090","AcceptedAnswerId":null,"OwnerUserId":5532004.0,"Title":"I add a new layer in caffe. There is a parameter in this layer which cannot be changed in test cpp file","Body":"<pre><\/pre>\n\n<p>The default value of \"apart\" is \"true\", then I changed the value to be \"false\", but the last line still display a \"1\" not \"0\". However the value of extract_one_dim_param->apart() is \"0\".<\/p>\n","answers":[]}
{"QuestionId":41608242,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><batch-normalization>","CreationDate":"2017-01-12T08:25:18.150","AcceptedAnswerId":41616692.0,"OwnerUserId":1714410.0,"Title":"How should \"BatchNorm\" layer be used in caffe?","Body":"<p>I am a little confused about how should I use\/insert <a href=\"https:\/\/stackoverflow.com\/q\/41269570\/1714410\"><\/a> layer in my models.<br>\nI see several different approaches, for instance:<\/p>\n\n<h3><a href=\"https:\/\/github.com\/KaimingHe\/deep-residual-networks\/blob\/master\/prototxt\/ResNet-101-deploy.prototxt#L75-L93\" rel=\"nofollow noreferrer\">ResNets<\/a>: + (no parameter sharing)<\/h3>\n\n<p> layer is followed immediately with  layer: <\/p>\n\n<pre><\/pre>\n\n<h3><a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/cifar10\/cifar10_full_sigmoid_train_test_bn.prototxt#L68-L82\" rel=\"nofollow noreferrer\">cifar10 example<\/a>: only <\/h3>\n\n<p>In the cifar10 example provided with caffe,  is used without any  following it:<\/p>\n\n<pre><\/pre>\n\n<h3><a href=\"https:\/\/gist.github.com\/ducha-aiki\/6457bbd49fea8b7634a7#file-cifar10_full_sigmoid_train_test_bn-prototxt-L73-L114\" rel=\"nofollow noreferrer\">cifar10<\/a> Different  for  and <\/h3>\n\n<p> is changed between  and  phase:<\/p>\n\n<pre><\/pre>\n\n<h3>So what should it be?<\/h3>\n\n<p>How should one use layer in caffe?<\/p>\n","answers":[{"AnswerId":"41616692","CreationDate":"2017-01-12T15:15:54.097","ParentId":null,"OwnerUserId":"249226","Title":null,"Body":"<p>If you follow the original paper, the Batch normalization should be followed by Scale and Bias layers (the bias can be included via the Scale, although this makes the Bias parameters inaccessible). <code>use_global_stats<\/code> should also be changed from training (False) to testing\/deployment (True) - which is the default behavior. Note that the first example you give is a prototxt for deployment, so it is correct for it to be set to True.<\/p>\n\n<p>I'm not sure about the shared parameters.<\/p>\n\n<p>I made a pull request to improve the documents on the batch normalization, but then closed it because I wanted to modify it. And then, I never got back to it.<\/p>\n\n<p>Note that I think <code>lr_mult: 0<\/code> for <code>\"BatchNorm\"<\/code> is no longer required (perhaps not allowed?), although I'm not finding the corresponding PR now.<\/p>\n"}]}
{"QuestionId":41608741,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2017-01-12T08:53:21.533","AcceptedAnswerId":null,"OwnerUserId":6777471.0,"Title":"Torch tensor set the negative numbers to zero","Body":"<p><\/p>\n\n<p>How to convert  such that all the negative values in x are replaced with zero without using a loop such that the tensor must look like<\/p>\n\n<p><\/p>\n","answers":[{"AnswerId":"53952681","CreationDate":"2018-12-28T01:22:52.597","ParentId":null,"OwnerUserId":"3639581","Title":null,"Body":"<p>Pytorch supports indexing by operators<\/p>\n\n<p><code>\na = torch.Tensor([1,0,-1])\na[a &lt; 0] = 0\na\ntensor([1., 0., 0.])\n<\/code><\/p>\n"}]}
{"QuestionId":41609678,"AnswerCount":6,"Tags":"<python><import><machine-learning><tensorflow><jupyter-notebook>","CreationDate":"2017-01-12T09:41:00.747","AcceptedAnswerId":null,"OwnerUserId":7408920.0,"Title":"ImportError: No module named data_utils","Body":"<p>I'm extremely new to python and I'm recently trying to understand more about Machine Learning and Neural Nets<\/p>\n\n<p>I know this is a trivial question but I seem to be having problem importing data_utils on jupyter notebook. Can anyone please help<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/R7sVu.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/R7sVu.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Note: I am not using Keras, and I am following the tutorial in <a href=\"https:\/\/youtu.be\/SJDEOWLHYVo?list=PL2-dafEMk2A7EEME489DsI468AB0wQsMV&amp;t=213\" rel=\"nofollow noreferrer\">this video<\/a>.<\/p>\n","answers":[{"AnswerId":"41610416","CreationDate":"2017-01-12T10:13:12.690","ParentId":null,"OwnerUserId":"5359243","Title":null,"Body":"<p>Since you mentioned about Machine Learning and Neural Nets, I'll assume you're referring to <a href=\"https:\/\/github.com\/fchollet\/keras\" rel=\"nofollow noreferrer\">Keras<\/a>.<\/p>\n\n<p>I'll also assume that you installed it via<\/p>\n\n<pre><code>pip install keras\n<\/code><\/pre>\n\n<p>In that case, you need to uninstall it first by running<\/p>\n\n<pre><code>pip uninstall keras\n<\/code><\/pre>\n\n<p>Then, clone it from <a href=\"https:\/\/github.com\/fchollet\/keras\" rel=\"nofollow noreferrer\">Github<\/a>, cd into its directory and run<\/p>\n\n<pre><code>sudo python setup.py install\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/35905264\/keras-importerror-no-module-named-data-utils\/35911254#35911254\">More information here<\/a><\/p>\n"},{"AnswerId":"44251134","CreationDate":"2017-05-29T22:50:40.143","ParentId":null,"OwnerUserId":"1042409","Title":null,"Body":"<p>Based on the link that you provided for the video you are using, go to\n<a href=\"https:\/\/github.com\/llSourcell\/tensorflow_chatbot\" rel=\"nofollow noreferrer\">this link<\/a> and download both files into your working directory.<\/p>\n\n<p>The files you need to download are <a href=\"https:\/\/raw.githubusercontent.com\/llSourcell\/tensorflow_chatbot\/master\/data_utils.py\" rel=\"nofollow noreferrer\">data_utils<\/a> and <a href=\"https:\/\/raw.githubusercontent.com\/llSourcell\/tensorflow_chatbot\/master\/seq2seq_model.py\" rel=\"nofollow noreferrer\">seq2seq_model<\/a>.<\/p>\n\n<p>But before doing this tutorial try the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/\" rel=\"nofollow noreferrer\">tensorflow tutorials<\/a> found on the tensorflow website, to get you started with this library. <\/p>\n\n<p>Also, if you are very new to python, I recommend <a href=\"https:\/\/www.tutorialspoint.com\/python3\/\" rel=\"nofollow noreferrer\">this tutorial<\/a> first.<\/p>\n"},{"AnswerId":"53484433","CreationDate":"2018-11-26T15:32:59.190","ParentId":null,"OwnerUserId":"10296210","Title":null,"Body":"<p>I too was using same files.\nJust open <code>seq2seq_model.py<\/code> and in line 35 remove from TensorFlow and keep it just <code>import dat_<\/code><\/p>\n"},{"AnswerId":"45404361","CreationDate":"2017-07-30T20:26:25.187","ParentId":null,"OwnerUserId":"7633796","Title":null,"Body":"<p>I was also following the video and went through the same issue, after searching for a while; Here is the link from <strong>Github<\/strong> <a href=\"https:\/\/github.com\/llSourcell\/tensorflow_chatbot\" rel=\"nofollow noreferrer\">tensorflow_chatbot_required_files<\/a>. You can download from here and copy it to your working directory (python file directory). <\/p>\n\n<p>Now you will be able to import both.<\/p>\n"},{"AnswerId":"45804994","CreationDate":"2017-08-21T20:23:08.710","ParentId":null,"OwnerUserId":"5899492","Title":null,"Body":"<p>what you are trying to do is importing a self-defined module, in order to do that do as follow:<\/p>\n\n<p>in your case the module is called 'data_util' which contains functions that will be called later as data_util.'function name'.<\/p>\n\n<p>let say that the data_util which is a python file (.py) is in this directory (C:\/Users\/xxx\/modules), so all what you have to do is to run this line of code in order for python to find your modul when you call import data_util:<\/p>\n\n<pre><code>import sys  \nsys.path.append('C:\/Users\/xxx\/modules')  \nimport data_util\n<\/code><\/pre>\n"},{"AnswerId":"44251018","CreationDate":"2017-05-29T22:36:22.700","ParentId":null,"OwnerUserId":"476749","Title":null,"Body":"<p>I'm assuming you are using code from tensorflow's github, in that case you need to download <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/translate\/data_utils.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/translate\/data_utils.py<\/a> into your folder also. <\/p>\n"}]}
{"QuestionId":41610894,"AnswerCount":1,"Tags":"<python><machine-learning><keras><gradient-descent><perceptron>","CreationDate":"2017-01-12T10:35:09.757","AcceptedAnswerId":null,"OwnerUserId":4569000.0,"Title":"Using the delta rule in keras","Body":"<p>I'm trying to build a linear single-layer perceptron (i.e. no hidden layers, all inputs connected to all outputs, linear activation function) and train it, one data point at a time, with the delta rule, but I don't get the results that I'm expecting. I'm using the mean squared error as my loss function, the derivative of which <em>should<\/em> result in weight updates that are simply learning_rate * error (* 2), but somehow the results look very different from my manual calculation. What am I missing?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41631534","CreationDate":"2017-01-13T09:42:19.043","ParentId":null,"OwnerUserId":"4569000","Title":null,"Body":"<p>I found the problem. Per default, dense layers include biases - once that is changed, the network shows the desired behaviour!<\/p>\n"}]}
{"QuestionId":41611130,"AnswerCount":1,"Tags":"<deep-learning><caffe><pycaffe><deeplearning4j><matconvnet>","CreationDate":"2017-01-12T10:46:15.530","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"Why loss is remaining constant during training FCN-8s?","Body":"<p>I am trying to run FCN-8s. I did the following steps:\n1. downloaded <a href=\"https:\/\/github.com\/developmentseed\/caffe-fcn\" rel=\"nofollow noreferrer\">this repository<\/a> \n2. converting my data to LMDB and changing the paths in train_val.prototxt\n3. downloading the fcn8s-heavy-pascal \n4. changing the  in  and  from <strong>60<\/strong> to <strong>5<\/strong> (the number of classes in my data) in the last following layers:<\/p>\n\n<pre><\/pre>\n\n<p>I started the training with the weights of pre-trained model for pascal dataset. But the the loss remains constant (loss = 105476) over the time.<\/p>\n\n<pre><\/pre>\n\n<p>I do not know which part I am doing wrong. I really appreciate your help to solve this issue.<\/p>\n","answers":[{"AnswerId":"42293199","CreationDate":"2017-02-17T08:43:03.957","ParentId":null,"OwnerUserId":"7579660","Title":null,"Body":"<ul>\n<li><p>Did you use the function called <code>surgery.transplant()<\/code> in <code>solve.py<\/code> to transplant the caffemodel of original net to your current net?<\/p><\/li>\n<li><p>Did you add <code>weight-filler<\/code> and <code>bias-filler<\/code> on Deconvolutional layer with initial value in <code>net.py<\/code>?<\/p><\/li>\n<li><p>After doing these two steps, did you execute the <code>net.py<\/code> to generate a updated layer?<\/p><\/li>\n<\/ul>\n\n<p>Check these steps, and see what will happen.<\/p>\n"}]}
{"QuestionId":41611255,"AnswerCount":1,"Tags":"<tensorflow><classification><floating-accuracy><softmax>","CreationDate":"2017-01-12T10:51:18.023","AcceptedAnswerId":null,"OwnerUserId":5974525.0,"Title":"TensorFlow: using F1_score in context of a binary classification","Body":"<p>I'm using a softmax function for a binary classification task.\nMy test label is a one hot list and looks like:\n    test_y = [[1. 0.] [1. 0.]\u2026]<\/p>\n\n<p>And the predicted label is a probability list: <\/p>\n\n<pre><\/pre>\n\n<p>When I try to use the f1_score, I get an error :<\/p>\n\n<p>ValueError: Can't handle mix of binary and continuous<\/p>\n\n<p>How can I handle this issue?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41617021","CreationDate":"2017-01-12T15:30:55.380","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>f1_score will not classify the results for you.<\/p>\n\n<p>Change your predictions to vectors of classes, for example:<\/p>\n\n<pre><code>import numpy as np\ntest_y = [np.argmax(prediction) for prediction in test_y]\ntest_y_pred= [np.argmax(prediction) for prediction in test_y_pred]\n<\/code><\/pre>\n"}]}
{"QuestionId":41611518,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-12T11:03:56.480","AcceptedAnswerId":41615550.0,"OwnerUserId":3104485.0,"Title":"Printing a generator in python tensor flow","Body":"<p>I am trying to follow the  tutorial as described in this <a href=\"https:\/\/github.com\/random-forests\/tutorials\/blob\/master\/ep7.ipynb\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n\n<p>I am trying to print the predicted result as described :<\/p>\n\n<pre><\/pre>\n\n<p>But I am not able to print the result. I am getting the following error.<\/p>\n\n<pre><\/pre>\n\n<p>How do I print the  in .<\/p>\n\n<p>I tried to write a loop and iterate over the elements it didn't work and I tried to use  to print the generator. That also didn't work. How do I print it ?<\/p>\n","answers":[{"AnswerId":"41612438","CreationDate":"2017-01-12T11:46:53.247","ParentId":null,"OwnerUserId":"2749397","Title":null,"Body":"<p>No <code>tensorflow<\/code> here, so let's mock up a generator and test it against your <code>print<\/code> expression<\/p>\n\n<pre><code>In [11]: def predict(a, b):\n    ...:     for i in range(10):\n    ...:         yield i, i*i\n    ...:         \n\nIn [12]: print('a:%d, b:%d'%(predict(0, 0)))\n---------------------------------------------------------------------------\nTypeError                                 Traceback (most recent call last)\n&lt;ipython-input-12-29ec761936ef&gt; in &lt;module&gt;()\n----&gt; 1 print('a:%d, b:%d'%(predict(0, 0)))\n\nTypeError: %d format: a number is required, not generator\n<\/code><\/pre>\n\n<p>So far, so good: I have the same problem that you've experienced.<\/p>\n\n<p>The problem is, of course, that what you get when you call a generator function are not values but a generator object...<\/p>\n\n<p>You have to iterate on the generator objects, using whatever is returned from each iteration, e.g.,<\/p>\n\n<pre><code>In [13]: print('\\n'.join('a:%d, b:%d'%(i,j) for i, j in predict(0,0)))\na:0, b:0\na:1, b:1\na:2, b:4\na:3, b:9\na:4, b:16\na:5, b:25\na:6, b:36\na:7, b:49\na:8, b:64\na:9, b:81\n<\/code><\/pre>\n\n<p>or, if you don't like one-liners,<\/p>\n\n<pre><code>In [14]: for i, j in predict(0, 0):\n    ...:     print('a:%d, b:%d'%(i,j))\n    ...:     \na:0, b:0\na:1, b:1\na:2, b:4\na:3, b:9\na:4, b:16\na:5, b:25\na:6, b:36\na:7, b:49\na:8, b:64\na:9, b:81\n<\/code><\/pre>\n\n<p>In other words, you have <em>to explicitly consume<\/em> what the generator is producing.<\/p>\n"},{"AnswerId":"41615550","CreationDate":"2017-01-12T14:22:32.120","ParentId":null,"OwnerUserId":"3104485","Title":null,"Body":"<p>This is how I solved it<\/p>\n\n<pre><code>new_samples = np.array([test_data[8]], dtype=float)\n\ny = list(classifier.predict(new_samples, as_iterable=True))\nprint('Predictions: {}'.format(str(y)))\n\nprint (\"Predicted %s, Label: %d\" % (str(y), test_labels[8]))\n<\/code><\/pre>\n"},{"AnswerId":"41611923","CreationDate":"2017-01-12T11:20:44.980","ParentId":null,"OwnerUserId":"2776376","Title":null,"Body":"<p>From the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.learn\/estimators#LinearClassifier.predict\" rel=\"nofollow noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>Runs inference to determine the class probability predictions.\n  (deprecated arguments)<\/p>\n  \n  <p>SOME ARGUMENTS ARE DEPRECATED. They will be removed after 2016-09-15.\n  Instructions for updating: The default behavior of predict() is\n  changing. The default value for as_iterable will change to True, and\n  then the flag will be removed altogether. The behavior of this flag is\n  described below.<\/p>\n<\/blockquote>\n\n<p>Try:<\/p>\n\n<pre><code>classifier.predict(x=test_data[0])\n<\/code><\/pre>\n"}]}
{"QuestionId":41612057,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-12T11:27:16.567","AcceptedAnswerId":41617388.0,"OwnerUserId":null,"Title":"how to add label to image data set for classification?","Body":"<p>I am using python 3.6 installed on mac os. I have text file that store name of image and the class number of every single image on.  <\/p>\n\n<pre><\/pre>\n\n<p>I want to give them to my neural network in <a href=\"\/questions\/tagged\/tensorflow\" class=\"post-tag\" title=\"show questions tagged &#39;tensorflow&#39;\" rel=\"tag\">tensorflow<\/a> as label of input data and give the image to network in same time like this <\/p>\n\n<pre><\/pre>\n\n<p>I cannot find any related documentation. could some one tell me what should i do for this pleased ?   <\/p>\n","answers":[{"AnswerId":"41617388","CreationDate":"2017-01-12T15:46:33.817","ParentId":null,"OwnerUserId":"5443758","Title":null,"Body":"<p>Assuming that you wanted to know, how to feed image and its respective label into neural network.<\/p>\n\n<p>There are two things:<\/p>\n\n<ol>\n<li>Reading the images and converting those in numpy array.<\/li>\n<li>Feeding the same and its corresponding label into network.<\/li>\n<\/ol>\n\n<p>As said by <a href=\"https:\/\/stackoverflow.com\/users\/3994824\/thomas-pinetz\">Thomas Pinetz<\/a> , once you calculated names and labels. Create one hot encoding of labels.<\/p>\n\n<pre><code>from PIL import Image\nnumber_of_batches = len(names)\/ batch_size\nfor i in range(number_of_batches):\n     batch_x = names[i*batch_size:i*batch_size+batch_size]\n     batch_y = labels[i*batch_size:i*batch_size+batch_size]\n     batch_image_data = np.empty([batch_size, image_height, image_width, image_depth], dtype=np.int)\n     for ix in range(len(batch_x)):\n        f = batch_x[ix]\n        batch_image_data[ix] = np.array(Image.open(data_dir+f))\n     sess.run(train_op, feed_dict={xs:batch_image_data, ys:batch_y})\n<\/code><\/pre>\n"},{"AnswerId":"41612182","CreationDate":"2017-01-12T11:34:00.357","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>You can use streight forward python i\/o utilities (<a href=\"https:\/\/docs.python.org\/3\/tutorial\/inputoutput.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3\/tutorial\/inputoutput.html<\/a>) like:<\/p>\n\n<pre><code>names = []\nlabels = []\nwith open('label.txt', 'r') as f:\n    for line in f.readlines():\n       tokens = line.split(' ')\n       names.append(tokens[0])\n       labels.append(int(tokens[1]))\n<\/code><\/pre>\n\n<p>Then you can use the names array to load in the images and the labels as your y array.<\/p>\n"}]}
{"QuestionId":41613767,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-01-12T12:54:16.813","AcceptedAnswerId":41675604.0,"OwnerUserId":5685754.0,"Title":"tf.parse_example used in mnist export example","Body":"<p>I am new to tensorflow and are reading mnist_export.py in tensorflow serving example.<\/p>\n\n<p>There is something here I cannot understand:<\/p>\n\n<pre><\/pre>\n\n<p>Above, serialized_tf_example is a Tensor.<\/p>\n\n<p>I have read the api document <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/converting#parse_example\" rel=\"nofollow noreferrer\">tf.parse_example<\/a> but it seems that  is serialized  protos like:<\/p>\n\n<pre><\/pre>\n\n<p>So how to understand  here as  is a Tensor, not  proto?<\/p>\n","answers":[{"AnswerId":"41675604","CreationDate":"2017-01-16T11:41:04.353","ParentId":null,"OwnerUserId":"5685754","Title":null,"Body":"<p>Here <code>serialized_tf_example<\/code> is serialized string of a <code>tf.train.Example<\/code>. See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/converting#parse_example\" rel=\"nofollow noreferrer\">tf.parse_example<\/a> for the usage. <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/#file_formats\" rel=\"nofollow noreferrer\">Reading data<\/a> chapter gives some example link. <\/p>\n\n<p>tf_example.SerializeToString() converts <code>tf.train.Example<\/code> to string and tf.parse_example parses the serialized string to a dict.<\/p>\n"},{"AnswerId":"49644165","CreationDate":"2018-04-04T06:25:23.603","ParentId":null,"OwnerUserId":"9594960","Title":null,"Body":"<p>The below mentioned code provides the simple example of using parse_example<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\nserialized_tf_example = tf.placeholder(tf.string, shape=[1], name='serialized_tf_example')\nfeature_configs = {'x': tf.FixedLenFeature(shape=[1], dtype=tf.float32)}\ntf_example = tf.parse_example(serialized_tf_example, feature_configs)\n\nfeature_dict = {'x': tf.train.Feature(float_list=tf.train.FloatList(value=[25]))}\nexample = tf.train.Example(features=tf.train.Features(feature=feature_dict))\nf = example.SerializeToString()\n\n\nsess.run(tf_example,feed_dict={serialized_tf_example:[f]})\n<\/code><\/pre>\n"}]}
{"QuestionId":41614400,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-12T13:25:40.600","AcceptedAnswerId":41616277.0,"OwnerUserId":1225744.0,"Title":"tensorflow run takes forever on small data","Body":"<p>I'm trying a simple example of a book where I have a train data sample with 892 lines, the usual titanic survival textbook example. I define:<\/p>\n\n<pre><\/pre>\n\n<p>And now I try to do:<\/p>\n\n<pre><\/pre>\n\n<p>and I see <\/p>\n\n<pre><\/pre>\n\n<p>but the  part hangs forever (or at least a very long tiem). I'm running this on Jupyter with a 2.7 kernel and  version <\/p>\n\n<p>What am I missing?<\/p>\n","answers":[{"AnswerId":"41616277","CreationDate":"2017-01-12T14:58:05.630","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>At the line<\/p>\n\n<pre><code>return tf.train.shuffle_batch(decoded,\n                              batch_size=batch_size,\n                              capacity=batch_size * 50,\n                              min_after_dequeue=batch_size)\n<\/code><\/pre>\n\n<p>You're <strong>defining<\/strong> the operation of extraction of values from the queue and the creation of batches.<\/p>\n\n<p>If you look at the complete signature of the method you'll notice that there's a parameter that refers to a number of threads.<\/p>\n\n<pre><code>tf.train.shuffle_batch(\ntensors,\nbatch_size,\ncapacity,\nmin_after_dequeue,\nnum_threads=1,\nseed=None, enqueue_many=False, shapes=None,\nallow_smaller_final_batch=False, shared_name=None, name=None)\n<\/code><\/pre>\n\n<p>I'm pointing this out because the operations you defined are executed from some threads.\nThe thread <strong>must<\/strong> be started and stopped, this function does not do this for you. The only thing that this function does, about the thread handling, is to add <code>num_thread<\/code> to a queue.<\/p>\n\n<p>In practice, to start and stop the threads you need to define within the session an operation that wake up the threads in the queue:<\/p>\n\n<pre><code>graph = tf.Graph()\n\nwith tf.Session(graph=graph) as sess:\n\n    W = tf.Variable(tf.zeros([5, 1]), name=\"weights\")\n    b = tf.Variable(0., name=\"bias\")\n    tf.global_variables_initializer().run()\n    print 'tf was run'\n    X,Y = inputs()\n    # define a coordinator to start and stop the threads\n    coord = tf.train.Coordinator()\n    # wake up the threads\n    threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n    print 'inputs!'\n    sess.run(Y) #execute operation\n    # When done, ask the threads to stop.\n    coord.request_stop()\n    # Wait for threads to finish.\n    coord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":41614718,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-12T13:40:30.763","AcceptedAnswerId":null,"OwnerUserId":4929293.0,"Title":"Why tensor flow give nan instead of number","Body":"<pre><\/pre>\n\n<p>My source code is same as above\nI can't understand why result is NAN.<br>\nI'm beginner in this tensorflow and machine learning area.<br>\nI think my source code is right. Please tell me if there are wrong.<\/p>\n","answers":[{"AnswerId":"41617282","CreationDate":"2017-01-12T15:42:08.110","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>With out seeing your <code>x_<\/code> values, it is difficult to know for certain. But an educated guess is the <code>tf.log<\/code> op you are using received zero or negative inputs. So instead of<\/p>\n\n<pre><code>cost = tf.reduce_mean(-tf.reduce_sum(hypothesis* tf.log(y), reduction_indices=[1]))\n<\/code><\/pre>\n\n<p>Try something like<\/p>\n\n<pre><code>cost = tf.reduce_mean(-tf.reduce_sum(hypothesis* tf.log(tf.clip_by_value(y, 1e-8, 1.0)), reduction_indices=[1]))\n<\/code><\/pre>\n\n<p>Also worth pointing out is that as of version 1.0.0-alpha TensorFlow has a specialized debugger that helps diagnosing this kind of issues, among others. See: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/debugger\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/debugger\/<\/a><\/p>\n\n<p>To try it out, after your line <code>sess.run(init)<\/code> add.<\/p>\n\n<pre><code>from tensorflow.python import debug as tf_debug\nsess = tf_debug.LocalCLIDebugWrapperSession(sess)\nsess.add_tensor_filter(\"has_inf_or_nan\", tf_debug.has_inf_or_nan)\n<\/code><\/pre>\n\n<p>In the command-line interface that shows up on the next <code>sess.run<\/code> call.<\/p>\n\n<pre><code>tfdbg&gt; run -f has_inf_or_nan\n<\/code><\/pre>\n"},{"AnswerId":"41636429","CreationDate":"2017-01-13T14:07:39.353","ParentId":null,"OwnerUserId":"5232403","Title":null,"Body":"<p>Looks like you are trying a simple logistic regression but the variables are all mixed up.<\/p>\n\n<p>First, the cross entropy must be:<\/p>\n\n<pre><code>y*tf.log(hypothesis)\n<\/code><\/pre>\n\n<p>not the other way around. Second, <code>y<\/code> must sampled from your data labels - something that you are not doing. <code>x_<\/code> and <code>y_<\/code> are missing. It is hard to debug without correcting all these two issues.<\/p>\n"}]}
{"QuestionId":41615654,"AnswerCount":1,"Tags":"<python><tensorflow><word2vec><recurrent-neural-network>","CreationDate":"2017-01-12T14:27:34.390","AcceptedAnswerId":41621101.0,"OwnerUserId":2267781.0,"Title":"How to pass 3d Tensor to tensorflow RNN embedding_rnn_seq2seq","Body":"<p>I'm trying to feed sentences in which each world has word2vec representation.\nHow can I do it in tensorflow seq2seq models?<\/p>\n\n<p>Suppose the variable <\/p>\n\n<pre><\/pre>\n\n<p>Which has dimensions [num_of_observations or batch_size x word_vec_representation x  sentense_lenght].<\/p>\n\n<p>when I pass it to embedding_rnn_seq2seq<\/p>\n\n<pre><\/pre>\n\n<p>error occurs<\/p>\n\n<pre><\/pre>\n\n<p>Also there is a more complex problem\nHow can i pas as input a vector, not a scalar to first cell of my RNN?<\/p>\n\n<p>By now it looks like (when we are about any sequence)<\/p>\n\n<ol>\n<li>get first value of sequence (scalar)<\/li>\n<li>compute First layer RNN First layer embedding cell output<\/li>\n<li>compute First layer RNN Second layer embedding cell output<\/li>\n<li>etc<\/li>\n<\/ol>\n\n<p>But this is needed:<\/p>\n\n<ol>\n<li>Get first value of sequence (vector) <\/li>\n<li>compute First layer RNN First layer cell output (as ordinary computing simple perceptron when Input is a vector) <\/li>\n<li>compute First layer RNN Second layer embedding cell output (as ordinary computing simple perceptron when Input is a vector)<\/li>\n<\/ol>\n","answers":[{"AnswerId":"41621101","CreationDate":"2017-01-12T19:12:37.243","ParentId":null,"OwnerUserId":"2267781","Title":null,"Body":"<p>The main point is that: \nseq2seq make inside themself word embedding. \n<a href=\"https:\/\/www.reddit.com\/r\/MachineLearning\/comments\/3t1887\/use_word_embeddings_in_tensor_flows_seq_2_seq\/\" rel=\"nofollow noreferrer\">Here is reddit question and answer<\/a><\/p>\n\n<p>Also, if smbd wants to use pretrained Word2Vec there are ways to do it,\nsee: <\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/35687678\/using-a-pre-trained-word-embedding-word2vec-or-glove-in-tensorflow\">stackoverflow 1<\/a> <\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/33851950\/initialising-seq2seq-embedding-with-pretrained-word2vec\">stackoverflow 2<\/a><\/li>\n<\/ul>\n\n<p>So this can be used no only for word embedding<\/p>\n"}]}
{"QuestionId":41615720,"AnswerCount":1,"Tags":"<macos><python-3.x><cuda><tensorflow><gpu>","CreationDate":"2017-01-12T14:30:44.307","AcceptedAnswerId":null,"OwnerUserId":2938765.0,"Title":"CUDA can't find GPU on mac","Body":"<p>A couple days ago I have managed to get CUDA working with tensorflow on my mac with a GeForce GTX 780M. However today I've noticed it is no longer working. I'm not sure what has changed but I've verified that the libraries (especially cudann) are still properly installed.<\/p>\n\n<p>Neither restarting nor reinstalling tensorflow helped (I installed tensorflow from <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/gpu\/tensorflow_gpu-0.12.1-py3-none-any.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/mac\/gpu\/tensorflow_gpu-0.12.1-py3-none-any.whl<\/a>) Here's the output from running the mnist example from the tensorflow website:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the output of :<\/p>\n\n<pre><\/pre>\n\n<p>Output of: <\/p>\n\n<pre><\/pre>\n\n<p>I have tried reinstalling the drivers, installing old ones and none of it helped<\/p>\n\n<p>According to <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/issues\/38\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/issues\/38<\/a> I did  to prevent memory issues when running tensorflow. If I then run  it fails to find the gpu:<\/p>\n\n<pre><\/pre>\n\n<p>However if I run  then running  gives:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41620786","CreationDate":"2017-01-12T18:50:48.080","ParentId":null,"OwnerUserId":"6088060","Title":null,"Body":"<p>\"CUDA_VISIBLE_DEVICES=1\" means that Cuda only sees gpu_1 on your machine. Do you have two GPUs? What does \"nvidia-smi\" show?<\/p>\n\n<p>In general, if you want to use \"CUDA_VISIBLE_DEVICES\", make sure you point to a GPU that you want to use. <\/p>\n"}]}
{"QuestionId":41616065,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-12T14:48:41.560","AcceptedAnswerId":null,"OwnerUserId":3882526.0,"Title":"Tensorflow print in function","Body":"<p>I have a function in a file  that defines a loss function:<\/p>\n\n<pre><\/pre>\n\n<p>To explain, somewhere in this function I want to print a tensor. However, I don't have any session declared in this function; my sessions are declared in another file called . Therefore, when I try to put  in , I can't because I don't know which session to  with. Is there a way to solve this problem, either by using  or other debug methods? Thanks!<\/p>\n","answers":[{"AnswerId":"41616274","CreationDate":"2017-01-12T14:57:57.990","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/debugging_operations#Print\" rel=\"nofollow noreferrer\"><code>tf.Print<\/code><\/a> works as an identity function which returns the same tensor that you passed as the first parameter, having a side effect of printing the list of tensors specified as the second parameter.<\/p>\n\n<p>So you should use as following:<\/p>\n\n<pre><code>def loss(a, b):\n    ...\n    a = tf.Print(a, [a], message = 'debug: ')\n    return tf.add(a, b)\n<\/code><\/pre>\n\n<p><code>a<\/code> will be printed each time tensor <code>tf.add(a, b)<\/code> is evaluated.<\/p>\n"}]}
{"QuestionId":41616292,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tflearn>","CreationDate":"2017-01-12T14:58:39.073","AcceptedAnswerId":41857516.0,"OwnerUserId":4750553.0,"Title":"How to load and retrain tflean model","Body":"<p>I am trying DNN- RNN on a text data set. It is a simple dummy data and I think the code can be used with most of text data. However I am getting error when I am trying to load the trained model and then retrain it. Please tell me If I am doing it wrong. <\/p>\n\n<pre><\/pre>\n\n<p>return data,vocab_processor, n_words, MAX_DOCUMENT_LENGTH<\/p>\n\n<p>This is for converting a list of text docs into the format required<\/p>\n\n<pre><\/pre>\n\n<p>Here we initialize the RNN model<\/p>\n\n<pre><\/pre>\n\n<p>This is for training the model.. I know this might not be the best way to do it but it was an experiment.. <\/p>\n\n<pre><\/pre>\n\n<p>This function predicts the new strings<\/p>\n\n<pre><\/pre>\n\n<p>This is for saving the trained model <\/p>\n\n<pre><\/pre>\n\n<p>And here I am loading the saved model\nand i am getting this error.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41857516","CreationDate":"2017-01-25T17:10:00.733","ParentId":null,"OwnerUserId":"4750553","Title":null,"Body":"<p>This is to create a graph and save it<\/p>\n\n<pre><code>graph1 = tf.Graph()\nwith graph1.as_default():\n    network = input_data(shape=[None, MAX_DOCUMENT_LENGTH])\n    network = tflearn.embedding(network, input_dim=n_words, output_dim=128)\n    branch1 = conv_1d(network, 128, 3, padding='valid', activation='relu', regularizer=\"L2\")\n    branch2 = conv_1d(network, 128, 4, padding='valid', activation='relu', regularizer=\"L2\")\n    branch3 = conv_1d(network, 128, 5, padding='valid', activation='relu', regularizer=\"L2\")\n    network = merge([branch1, branch2, branch3], mode='concat', axis=1)\n    network = tf.expand_dims(network, 2)\n    network = global_max_pool(network)\n    network = dropout(network, 0.5)\n    network = fully_connected(network, 2, activation='softmax')\n    network = regression(network, optimizer='adam', learning_rate=0.001,loss='categorical_crossentropy', name='target')\n    model = tflearn.DNN(network, tensorboard_verbose=0)\n    clf, acc, roc_auc,fpr,tpr =classify_DNN(data,clas,model)\n    clf.save(model_path)\n<\/code><\/pre>\n\n<p>To reload and retrain or use it for prediction<\/p>\n\n<pre><code>MODEL = None\nwith tf.Graph().as_default():\n## Building deep neural network\n    network = input_data(shape=[None, MAX_DOCUMENT_LENGTH])\n    network = tflearn.embedding(network, input_dim=n_words, output_dim=128)\n    branch1 = conv_1d(network, 128, 3, padding='valid', activation='relu', regularizer=\"L2\")\n    branch2 = conv_1d(network, 128, 4, padding='valid', activation='relu', regularizer=\"L2\")\n    branch3 = conv_1d(network, 128, 5, padding='valid', activation='relu', regularizer=\"L2\")\n    network = merge([branch1, branch2, branch3], mode='concat', axis=1)\n    network = tf.expand_dims(network, 2)\n    network = global_max_pool(network)\n    network = dropout(network, 0.5)\n    network = fully_connected(network, 2, activation='softmax')\n    network = regression(network, optimizer='adam', learning_rate=0.001,loss='categorical_crossentropy', name='target')\n    new_model = tflearn.DNN(network, tensorboard_verbose=3)\n    new_model.load(model_path)\n    MODEL = new_model\n<\/code><\/pre>\n\n<p>Use the MODEL for prediction or retraining.\nThe 1st line and the with loop was important. For anyone who might need help<\/p>\n"}]}
{"QuestionId":41617463,"AnswerCount":4,"Tags":"<python><tensorflow><deep-learning><tensorboard>","CreationDate":"2017-01-12T15:50:20.123","AcceptedAnswerId":null,"OwnerUserId":2616232.0,"Title":"Tensorflow Confusion Matrix in TensorBoard","Body":"<p>I want to have a visual of confusion matrix in tensorboard. To do this, I am modifying Evaluation example of Tensorflow Slim: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/eval_image_classifier.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/eval_image_classifier.py<\/a><\/p>\n\n<p>In this example code, Accuracy already provided but it is not possible to add \"confusion matrix\" metric directly because it is not streaming. <\/p>\n\n<p><strong>What is difference between streaming metrics and non-streaming ones?<\/strong><\/p>\n\n<p>Therefore, I tried to add it like this:<\/p>\n\n<pre><\/pre>\n\n<p>This creates an image in tensorboard but probably there is a formatting problem. Matrix should be normalized between 0-1 so that It produces meaningful image.<\/p>\n\n<p><strong>How can I produce a meaningful confusion matrix? How can I deal with multi batch evaluation process?<\/strong><\/p>\n","answers":[{"AnswerId":"41621582","CreationDate":"2017-01-12T19:39:46.907","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>Re: your image not being meaningful - according to the docs for the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/summary\/generation_of_summaries_#image\" rel=\"nofollow noreferrer\">tf.summary.image<\/a>, for uint8 values are unchanged (won't be normalized), and are interpreted in range [0, 255]. Have you tried re-normalizing your image to [0,255] instead of [0,1]?<\/p>\n"},{"AnswerId":"48030258","CreationDate":"2017-12-30T01:58:57.360","ParentId":null,"OwnerUserId":"7320664","Title":null,"Body":"<p>Here is something I have put together That works reasonably well. Still need to adjust a few things like the tick placements etc. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PEMTG.gif\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PEMTG.gif\" alt=\"Confusion Matrix as Image in Tensorflow\"><\/a><\/p>\n\n<p>Here is the function that will pretty much do everything for you.<\/p>\n\n<pre><code>from textwrap import wrap\nimport re\nimport itertools\nimport tfplot\nimport matplotlib\nimport numpy as np\nfrom sklearn.metrics import confusion_matrix\n\n\n\ndef plot_confusion_matrix(correct_labels, predict_labels, labels, title='Confusion matrix', tensor_name = 'MyFigure\/image', normalize=False):\n''' \nParameters:\n    correct_labels                  : These are your true classification categories.\n    predict_labels                  : These are you predicted classification categories\n    labels                          : This is a lit of labels which will be used to display the axix labels\n    title='Confusion matrix'        : Title for your matrix\n    tensor_name = 'MyFigure\/image'  : Name for the output summay tensor\n\nReturns:\n    summary: TensorFlow summary \n\nOther itema to note:\n    - Depending on the number of category and the data , you may have to modify the figzie, font sizes etc. \n    - Currently, some of the ticks dont line up due to rotations.\n'''\ncm = confusion_matrix(correct_labels, predict_labels, labels=labels)\nif normalize:\n    cm = cm.astype('float')*10 \/ cm.sum(axis=1)[:, np.newaxis]\n    cm = np.nan_to_num(cm, copy=True)\n    cm = cm.astype('int')\n\nnp.set_printoptions(precision=2)\n###fig, ax = matplotlib.figure.Figure()\n\nfig = matplotlib.figure.Figure(figsize=(7, 7), dpi=320, facecolor='w', edgecolor='k')\nax = fig.add_subplot(1, 1, 1)\nim = ax.imshow(cm, cmap='Oranges')\n\nclasses = [re.sub(r'([a-z](?=[A-Z])|[A-Z](?=[A-Z][a-z]))', r'\\1 ', x) for x in labels]\nclasses = ['\\n'.join(wrap(l, 40)) for l in classes]\n\ntick_marks = np.arange(len(classes))\n\nax.set_xlabel('Predicted', fontsize=7)\nax.set_xticks(tick_marks)\nc = ax.set_xticklabels(classes, fontsize=4, rotation=-90,  ha='center')\nax.xaxis.set_label_position('bottom')\nax.xaxis.tick_bottom()\n\nax.set_ylabel('True Label', fontsize=7)\nax.set_yticks(tick_marks)\nax.set_yticklabels(classes, fontsize=4, va ='center')\nax.yaxis.set_label_position('left')\nax.yaxis.tick_left()\n\nfor i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):\n    ax.text(j, i, format(cm[i, j], 'd') if cm[i,j]!=0 else '.', horizontalalignment=\"center\", fontsize=6, verticalalignment='center', color= \"black\")\nfig.set_tight_layout(True)\nsummary = tfplot.figure.to_summary(fig, tag=tensor_name)\nreturn summary\n<\/code><\/pre>\n\n#\n\n<p>And here is the rest of the code that you will need to call this functions. <\/p>\n\n<pre><code>''' confusion matrix summaries '''\nimg_d_summary_dir = os.path.join(checkpoint_dir, \"summaries\", \"img\")\nimg_d_summary_writer = tf.summary.FileWriter(img_d_summary_dir, sess.graph)\nimg_d_summary = plot_confusion_matrix(correct_labels, predict_labels, labels, tensor_name='dev\/cm')\nimg_d_summary_writer.add_summary(img_d_summary, current_step)\n<\/code><\/pre>\n\n<p>Confuse away!!!<\/p>\n"},{"AnswerId":"42857070","CreationDate":"2017-03-17T12:02:16.573","ParentId":null,"OwnerUserId":"7652692","Title":null,"Body":"<p>Here's how I produced and displayed a \"streaming\" confusion matrix for test code (returned <code>test_op<\/code> is evaluated for each batch to test on).<\/p>\n\n<pre><code>def _get_streaming_metrics(prediction,label,num_classes):\n\n    with tf.name_scope(\"test\"):\n        # the streaming accuracy (lookup and update tensors)\n        accuracy,accuracy_update = tf.metrics.accuracy(label, prediction, \n                                               name='accuracy')\n        # Compute a per-batch confusion\n        batch_confusion = tf.confusion_matrix(label, prediction,\n                                             num_classes=num_classes,\n                                             name='batch_confusion')\n        # Create an accumulator variable to hold the counts\n        confusion = tf.Variable( tf.zeros([num_classes,num_classes], \n                                          dtype=tf.int32 ),\n                                 name='confusion' )\n        # Create the update op for doing a \"+=\" accumulation on the batch\n        confusion_update = confusion.assign( confusion + batch_confusion )\n        # Cast counts to float so tf.summary.image renormalizes to [0,255]\n        confusion_image = tf.reshape( tf.cast( confusion, tf.float32),\n                                  [1, num_classes, num_classes, 1])\n        # Combine streaming accuracy and confusion matrix updates in one op\n        test_op = tf.group(accuracy_update, confusion_update)\n\n        tf.summary.image('confusion',confusion_image)\n        tf.summary.scalar('accuracy',accuracy)\n\n    return test_op,accuracy,confusion\n<\/code><\/pre>\n\n<p>After you process all the data batches by running <code>test_op<\/code> , you can simply look up the final confusion matrix (within your session) by <code>confusion.eval()<\/code> or <code>sess.eval(confusion)<\/code> if you prefer.<\/p>\n"},{"AnswerId":"42983389","CreationDate":"2017-03-23T17:47:43.767","ParentId":null,"OwnerUserId":"6659511","Title":null,"Body":"<p>Here is something that works with tf.contrib.metrics.MetricSpec (when you use Estimator). It is inspired from Jerod's answer and the metric_op.py source file. You get a streamed confusion matrix with percentages :<\/p>\n\n<pre><code>from tensorflow.python.framework import ops,dtypes\nfrom tensorflow.python.ops import array_ops,variables\n\ndef _createLocalVariable(name, shape, collections=None, \nvalidate_shape=True,\n              dtype=dtypes.float32):\n  \"\"\"Creates a new local variable.\n  \"\"\"\n  # Make sure local variables are added to \n  # tf.GraphKeys.LOCAL_VARIABLES\n  collections = list(collections or [])\n  collections += [ops.GraphKeys.LOCAL_VARIABLES]\n  return variables.Variable(\n  initial_value=array_ops.zeros(shape, dtype=dtype),\n  name=name,\n  trainable=False,\n  collections=collections,\n  validate_shape=validate_shape)\n\ndef streamingConfusionMatrix(label, prediction, \nweights=None,num_classes=None):\n  \"\"\"\n  Compute a streaming confusion matrix\n  :param label: True labels\n  :param prediction: Predicted labels\n  :param weights: (Optional) weights (unused)\n  :param num_classes: Number of labels for the confusion matrix\n  :return: (percentConfusionMatrix,updateOp)\n  \"\"\"\n  # Compute a per-batch confusion\n\n  batch_confusion = tf.confusion_matrix(label, prediction,\n                                    num_classes=num_classes,\n                                    name='batch_confusion')\n\n  count = _createLocalVariable(None,(),dtype=tf.int32)\n  confusion = _createLocalVariable('streamConfusion',[num_classes, \n  num_classes],dtype=tf.int32)\n\n  # Create the update op for doing a \"+=\" accumulation on the batch\n  countUpdate = count.assign(count + tf.reduce_sum(batch_confusion))\n  confusionUpdate = confusion.assign(confusion + batch_confusion)\n\n  updateOp = tf.group(confusionUpdate,countUpdate)\n\n  percentConfusion = 100 * tf.truediv(confusion,count)\n\n  return percentConfusion,updateOp\n<\/code><\/pre>\n\n<p>You can then use it as evaluation metric in the following way:<\/p>\n\n<pre><code>from tensorflow.contrib import learn,metrics\n#[...]\n\nevalMetrics = {'accuracy': \nlearn.MetricSpec(metric_fn=metrics.streaming_accuracy),\n               'confusionMatrix':learn.MetricSpec(metric_fn=\n                                                  lambda \nlabel,prediction,weights=None:                         \nstreamingConfusionMatrix(                                                    \nlabel,prediction,weights,num_classes=nLabels))}\n<\/code><\/pre>\n\n<p>I suggest you use numpy.set_printoptions(precision=2,suppress=True) to print it out.<\/p>\n"}]}
{"QuestionId":41617597,"AnswerCount":1,"Tags":"<python><theano><lasagne>","CreationDate":"2017-01-12T15:55:18.360","AcceptedAnswerId":41628094.0,"OwnerUserId":2280645.0,"Title":"Theano MiniBatch Iterator not working","Body":"<p><strong>Theano MiniBatch Iterator not working<\/strong><\/p>\n\n<p>I coded a minibatch iterator to get predicted results from my neural network.\nHowever, i made some tests and noticed a few errors.<\/p>\n\n<p><strong>Basically :<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I made a script to show this bug in my code. Its shown below:<\/p>\n\n<pre><\/pre>\n\n<h2>Here are the Tests<\/h2>\n\n<p>For batch_size = 10 :<\/p>\n\n<pre><\/pre>\n\n<p>For batch_size = 11 :<\/p>\n\n<pre><\/pre>\n\n<p>For Batch_size  12<\/p>\n\n<pre><\/pre>\n\n<p><strong>How can i fix this?<\/strong><\/p>\n","answers":[{"AnswerId":"41628094","CreationDate":"2017-01-13T05:44:52.990","ParentId":null,"OwnerUserId":"3737792","Title":null,"Body":"<p>Please try to be more specific in the question. What exactly do you want to fix?<\/p>\n\n<p>There isn't any error.\nWhen the batch size is bigger than the inputs the function <code>minibatch_iterator_predictor<\/code> yields an empty iterator and the loop <code>for batch in minibatch_iterator_predictor(inputs=x, batch_size=batch_size)<\/code> is not executed.<\/p>\n\n<p>What you get when the batch_size is bigger than the number of inputs is just the zeroes from the initialization: <code>prediction = np.empty((x.shape[0], 2), dtype=np.float32)<\/code><\/p>\n\n<p>What you can do is to limit the max batch_size to the number of inputs:<\/p>\n\n<pre><code>def minibatch_iterator_predictor(inputs, batch_size):\n    assert len(inputs) &gt; 0\n    if batch_size &gt; len(inputs):\n        batch_size = len(inputs)\n\n    for start_idx in range(0, len(inputs) - batch_size + 1, batch_size):\n        excerpt = slice(start_idx, start_idx + batch_size)\n        yield inputs[excerpt]\n<\/code><\/pre>\n"}]}
{"QuestionId":41618096,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><computer-vision><deep-learning><convolution>","CreationDate":"2017-01-12T16:19:04.210","AcceptedAnswerId":null,"OwnerUserId":1145744.0,"Title":"How to use a trained CNN model for object identification in Tensorflow","Body":"<p>I have a CNN model that is trained using a set of 120 pictures. \nThe images are converted in TFR record and labeled with this method<\/p>\n\n<pre><\/pre>\n\n<p>The whole <em>model+training<\/em> script ends with \nand I get 2 .tfr files as output (<em>training<\/em> and <em>test<\/em>).<\/p>\n\n<p>Now suppose you have a picture and want to know what is the more similar picture of the 120 pic sample to identify it. How I have to proceed? <\/p>\n\n<p>train_prediction tensor has this format 120 is the total numbers of categories <\/p>\n\n<p>In the book that I'm reading unfortunately there isn't any indication and the chapter end with this trained model that I don't know how to use in a real application, and searching in internet there are many similar sample that end at same point.<\/p>\n","answers":[{"AnswerId":"41734546","CreationDate":"2017-01-19T05:41:37.397","ParentId":null,"OwnerUserId":"2839624","Title":null,"Body":"<p>I didn't understand your question. By 120 samples, do you mean there are 120 classes with few examples images per class? In that case, this is classification problem where you train a model for classification with input as image and output as probability that the input will belong to one of the 120 classes (+ extra class that it doesnt belong to any of those). In this case, fully connected layer is fed into softmax function that outputs this probability for 120 classes which is essentially vector of length 120. But for this, you need multiple training examples per class.<\/p>\n\n<p>In case you just have 120 images and you need to know if another test image similarity with one of these images, you can just take output vector of layer before fully connected layer for those 120 images for a neural network that is already trained on a larger sample of images (like inception model) and then just similarity of the test image vector with those 120 vectors.<\/p>\n\n<p>I have hosted a simple to use demo using above technique <a href=\"http:\/\/demo.nanonets.ai\/ImageCategorization\/\" rel=\"nofollow noreferrer\">here<\/a>. Check if works out for you otherwise refine the problem statement more.<\/p>\n"}]}
{"QuestionId":41618564,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-12T16:42:49.123","AcceptedAnswerId":41639889.0,"OwnerUserId":4782057.0,"Title":"Tensorflow error while making train_step.run()","Body":"<p>I've been around this problem for the whole day and I can't find a solution online for what it could be. <\/p>\n\n<p>I'm coding a convolutional neural network for classifying some BW images. I'm first reading the data, preparing the architecture of the network and then running the training part, but all the time I get this error while trying to train:<\/p>\n\n<pre><\/pre>\n\n<p>I can't find anything on the web that helps, me. And I can't figure out what is wrong. Thank you very much guys, I link you the whole code below:<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT ERROR WITH OPS NAMES<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41639889","CreationDate":"2017-01-13T17:07:25.857","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>From reading the code, the problem stems from the shape of the arguments to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.softmax_cross_entropy_with_logits()<\/code><\/a>. According to the documentation:<\/p>\n\n<blockquote>\n  <p><code>logits<\/code> and <code>labels<\/code> must have the same shape <code>[batch_size, num_classes]<\/code> and the same dtype (either <code>float16<\/code>, <code>float32<\/code>, or <code>float64<\/code>).<\/p>\n<\/blockquote>\n\n<p>Your code calls <code>tf.nn.softmax_cross_entropy_with_logits(prediction, y)<\/code>, so let's look at the shapes of the arguments:<\/p>\n\n<ul>\n<li><p><code>prediction<\/code> is the value returned from <code>convolutional_neural_network(x)<\/code>, and it has shape <code>[batch_size, N_CLASSES]<\/code>. (The placeholder for <code>x<\/code> represents <code>batch_size<\/code> as <code>None<\/code>, so that it can be dynamic.)<\/p><\/li>\n<li><p><code>y<\/code> is defined as <code>y = tf.placeholder(tf.int32)<\/code>. The placeholder has no shape information, so it is statically unknown (which partly explains the poor error message... see below for more about that). To find out the actual shape of <code>y<\/code>, we can look at how you feed the placeholder, and it appears that you feed it with a list of integers, parsed from the input file, where the integer represents the true label for the corresponding example.<\/p><\/li>\n<\/ul>\n\n<p>To fix this problem, you should replace <code>tf.nn.softmax_cross_entropy_with_logits()<\/code> with its <strong>sparse<\/strong> counterpart, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#sparse_softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.sparse_softmax_cross_entropy_with_logits()<\/code><\/a>, which can handle input data in the format that you are using:<\/p>\n\n<pre><code>cross_entropy = tf.reduce_mean(\n    tf.nn.sparse_softmax_cross_entropy_with_logits(prediction, y))\n<\/code><\/pre>\n\n<p>(An alternative approach would be to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#one_hot\" rel=\"nofollow noreferrer\"><code>tf.one_hot(y, N_CLASSES)<\/code><\/a> to convert <code>y<\/code> into the appropriate one-hot encoding for <code>tf.nn.softmax_cross_entropy_with_logits()<\/code>, but that can be less efficient because it has to materialize a potentially large matrix for the target values.)<\/p>\n\n<p>Note that the reason you are getting a runtime error for this problem is because of how the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/placeholders#placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a> for <code>y<\/code> is defined, without a static shape. If you defined it as being a vector, you would get an error at graph construction time:<\/p>\n\n<pre><code># `y` is a (variable-length) vector.\ny = tf.placeholder(tf.int32, shape=[None])\n<\/code><\/pre>\n"}]}
{"QuestionId":41619128,"AnswerCount":1,"Tags":"<theano><keras><recurrent-neural-network><theano-cuda>","CreationDate":"2017-01-12T17:12:13.993","AcceptedAnswerId":null,"OwnerUserId":1462738.0,"Title":"model selection with Keras and Theano takes a very long time","Body":"<p>I am performing nested cross-validation for model selection and performance estimation for a set of recurrent neural networks with different architectures and parameters using Keras and Theano, which are set up to run on a AWS P2 instance which has a Tesla K80 GPU with CUDA and cuDNN installed\/enabled.<\/p>\n\n<p>To perform model selection, I compare 30 models sampled from the parameter space using<\/p>\n\n<pre><\/pre>\n\n<p>I then construct a RNN model using the function  defined below<\/p>\n\n<pre><\/pre>\n\n<p>which construct a RNN whose number of hidden layers is given by the parameter  in  and the number of hidden units in each layer is randomly sampled from the list . At the end, this function s the model and returns it.<\/p>\n\n<p>During the nested cross-validation (CV), the inner loop (which runs  times) compares the performance of the 30 randomly selected model. After this step, I pick the best-performing model in the outer loop and estimate its performance on a hold-out dataset; this scheme is repeated  times. Therefore, I am ing a RNN model xx30 times, and this takes an extremely long time; for example, when  and , my method takes between 6 to 7 hours to finish.<\/p>\n\n<p>I see that the GPU is being used sporadically (but the GPU usage never goes above 40%); however, most of the time, it is the CPU that is being used. My (uneducated) guess is that  is being done on the CPU many many times and takes the bulk of the computing time, whereas model fitting and predicting are done on the GPU and takes a short time.<\/p>\n\n<p>My questions:<\/p>\n\n<ol>\n<li>Is there a way to remedy this situation?<\/li>\n<li>Is  actually done on the CPU?<\/li>\n<li>How do people do nested CV to select the best RNN architecture?<\/li>\n<li>Is it reasonable for me to perform this scheme on the production server? Do you suggest I do one big nested CV, that might take 24 hours, to select the best performing model and just use that one model afterwards on the production server?<\/li>\n<\/ol>\n\n<p>Thank you all.<\/p>\n","answers":[{"AnswerId":"41627818","CreationDate":"2017-01-13T05:19:12.120","ParentId":null,"OwnerUserId":"1950580","Title":null,"Body":"<p>I can't answer all your questions, still hope it helps.<\/p>\n\n<p>Compilation is done in CPU because it's mainly composed of symbolic graph operations and code generation. To make things worse, theano graph optimization uses pure python code, which can be an overhead compared to a C\/C++ implementation.<\/p>\n\n<p>To improve theano compilation time (at the cost of runtime performance):<\/p>\n\n<h3>Use less aggressive optimization<\/h3>\n\n<p>In <code>\/home\/ec2-user\/.theanorc<\/code> add line: <\/p>\n\n<p><code>optimizer = fast_compile<\/code><\/p>\n\n<p>Or totally disable optimization with:<\/p>\n\n<p><code>optimizer = None<\/code><\/p>\n\n<h3>Precompile some blocks<\/h3>\n\n<p>If there are common blocks shared amoung your models, you can precompile them with <code>theano.OpFromGraph<\/code><\/p>\n\n<p>You can't do this in Keras alone, though.<\/p>\n\n<h3>Switch framework<\/h3>\n\n<p>Keras does support tensorflow backend. Compared to theano, tensorflow work more like a VM than a compiler. Typically TF runs slower than theano but compiles much faster.<\/p>\n"}]}
{"QuestionId":41619281,"AnswerCount":1,"Tags":"<variables><testing><caffe><sharing><phase>","CreationDate":"2017-01-12T17:19:58.997","AcceptedAnswerId":null,"OwnerUserId":2774587.0,"Title":"caffe forgets set variable during test phase","Body":"<p>In caffe, if I make my own variable associated with a layer and I modify this variable during training iterations, the test phase does not know about the modifications to that variable, it seems to initialize it for itself.  the training testing phases share weights but nothing else. How do I share other variables between the training and testing phases. Thanks.<\/p>\n\n<p>eg:\nsay I declare in inner_product_layer.hpp,<\/p>\n\n<pre><\/pre>\n\n<p>and in Backward_gpu() I modify,<\/p>\n\n<pre><\/pre>\n\n<p>the testing phase thinks dog=5. I want the testing phase to know that dog is now 6.<\/p>\n","answers":[{"AnswerId":"41619771","CreationDate":"2017-01-12T17:48:38.127","ParentId":null,"OwnerUserId":"628538","Title":null,"Body":"<p>The executable for training and test phase is different, hence, if you modify a variable inside the code during training, it won't be reflected at test time. You could dump the variable to an output file when training ends and read the file as an argument at test time while initializing the layer in the LayerSetUp function. The filename can be added as a layer parameter , where you can check if you are in the test phase and set your parameters from the file you saved.<\/p>\n"}]}
{"QuestionId":41620975,"AnswerCount":1,"Tags":"<python><tensorflow><prediction>","CreationDate":"2017-01-12T19:04:11.963","AcceptedAnswerId":null,"OwnerUserId":6083088.0,"Title":"Predict product returns in e-commerce","Body":"<p>I was trying to build a NN model using TF by following this article on the internet <a href=\"http:\/\/aisel.aisnet.org\/cgi\/viewcontent.cgi?article=1369&amp;context=icis2015\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n\n<p>As of now, I have been able to use some code online <a href=\"https:\/\/github.com\/JRMeyer\/tensorflow-tutorial\/blob\/master\/logistic_regression_train.py\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n\n<p>This uses a Logistic regression in TF, and the tries to predict a binary classification in [1 0] pattern, where [1 0] &amp; [0 1] is is not-returned &amp; returned respectively. <\/p>\n\n<p>I have 13 feature set and 1 label, but post training the model predicts at an accuracy of around 40% <\/p>\n\n<p>I have also used Sklearn's scaler to scale the feature set in -1,1. <\/p>\n\n<p>Is there a definite way to tackle this problem for a beginner?<\/p>\n","answers":[{"AnswerId":"41622077","CreationDate":"2017-01-12T20:10:03.080","ParentId":null,"OwnerUserId":"3893224","Title":null,"Body":"<p>Have you tried other learning models such as a Decision Tree (and by extension Random Forest), or perhaps XGBoost?  In small experiments, I found those to be better at classifying over a neural network.<\/p>\n\n<p>You could also spend a few minutes reading up on feature engineering to some other ideas for getting better results from your classifier.<\/p>\n"}]}
{"QuestionId":41621053,"AnswerCount":2,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2017-01-12T19:10:13.533","AcceptedAnswerId":41621150.0,"OwnerUserId":1537681.0,"Title":"How to inject values into the middle of TensorFlow graph?","Body":"<p>Consider the following code:<\/p>\n\n<pre><\/pre>\n\n<p>The resulting graph is x -> z -> y. Sometimes I'm interested in computing y all the way from from x but sometimes I have z to start and would like inject this value into the graph. So the z needs to behave like a partial placeholder. How can I do that?<\/p>\n\n<p>(For anyone interested why I need this. I am working with an autoencoder network which observes an image x, generates an intermediate compressed representation z, then computes reconstruction of image y. I'd like to see what the network reconstructs when I inject different values for z.)<\/p>\n","answers":[{"AnswerId":"47206674","CreationDate":"2017-11-09T16:15:21.920","ParentId":null,"OwnerUserId":"4572034","Title":null,"Body":"<p>I'm not allowed to comment on your post, @iramusa, so I'll give an answer. You don't need to use a placeholder_with_default. You can just feed in the values to whatever node you want:<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.placeholder(tf.float32,(), name='x')\nz = x + tf.constant(5.0)\ny = z*tf.constant(0.5)\n\nwith tf.Session() as sess:\n    print(sess.run(y, feed_dict={x: 2}))  # get 3.5\n    print(sess.run(y, feed_dict={z: 5}))  # get 2.5\n    print(sess.run(y, feed_dict={y: 5}))  # get 5\n<\/code><\/pre>\n"},{"AnswerId":"41621150","CreationDate":"2017-01-12T19:15:15.380","ParentId":null,"OwnerUserId":"1537681","Title":null,"Body":"<p>Use placeholder with default in the following way:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, (), name='x')\n# z is a placeholder with default value\nz = tf.placeholder_with_default(x+tf.constant(5.0), (), name='z')\ny = tf.mul(z, tf.constant(0.5))\n\nwith tf.Session() as sess:\n    # and feed the z in\n    print(sess.run(y, feed_dict={z: 5}))\n<\/code><\/pre>\n\n<p>Silly me.<\/p>\n"}]}
{"QuestionId":41621071,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2017-01-12T19:11:01.147","AcceptedAnswerId":41642426.0,"OwnerUserId":2847699.0,"Title":"Restore subset of variables in Tensorflow","Body":"<p>I am training a Generative Adversarial Network (GAN) in tensorflow, where basically we have two different networks each one with its own optimizer.<\/p>\n\n<pre><\/pre>\n\n<p>The problem is that I train one of the networks (g) first, and then, I want to train g and d together. However, when I call the load function:<\/p>\n\n<pre><\/pre>\n\n<p>I have an error like this (with a lot more traceback):<\/p>\n\n<pre><\/pre>\n\n<p>I can restore the g network and keep training with that function, but when I want to star d from scratch, and g from the the stored model I have that error.<\/p>\n","answers":[{"AnswerId":"50216949","CreationDate":"2018-05-07T14:47:05.170","ParentId":null,"OwnerUserId":"8671242","Title":null,"Body":"<p>I had a similar problem when restoring only part of my variables from a checkpoint and some of the saved variables did not exist in the new model.\nInspired by @Lidong answer I modified a little the reading function:<\/p>\n\n<pre><code>def get_tensors_in_checkpoint_file(file_name,all_tensors=True,tensor_name=None):\nvarlist=[]\nvar_value =[]\nreader = pywrap_tensorflow.NewCheckpointReader(file_name)\nif all_tensors:\n  var_to_shape_map = reader.get_variable_to_shape_map()\n  for key in sorted(var_to_shape_map):\n    varlist.append(key)\n    var_value.append(reader.get_tensor(key))\nelse:\n    varlist.append(tensor_name)\n    var_value.append(reader.get_tensor(tensor_name))\nreturn (varlist, var_value)\n<\/code><\/pre>\n\n<p>and added a loading function:<\/p>\n\n<pre><code>def build_tensors_in_checkpoint_file(loaded_tensors):\nfull_var_list = list()\n# Loop all loaded tensors\nfor i, tensor_name in enumerate(loaded_tensors[0]):\n    # Extract tensor\n    try:\n        tensor_aux = tf.get_default_graph().get_tensor_by_name(tensor_name+\":0\")\n    except:\n        print('Not found: '+tensor_name)\n    full_var_list.append(tensor_aux)\nreturn full_var_list\n<\/code><\/pre>\n\n<p>Then you can simply load all common variables using:<\/p>\n\n<pre><code>CHECKPOINT_NAME = path to save file\nrestored_vars  = get_tensors_in_checkpoint_file(file_name=CHECKPOINT_NAME)\ntensors_to_load = build_tensors_in_checkpoint_file(restored_vars)\nloader = tf.train.Saver(tensors_to_load)\nloader.restore(sess, CHECKPOINT_NAME)\n<\/code><\/pre>\n\n<p>Edit: I am using tensorflow 1.2<\/p>\n"},{"AnswerId":"47979898","CreationDate":"2017-12-26T14:24:14.190","ParentId":null,"OwnerUserId":"5752878","Title":null,"Body":"<p>Inspired by @mrry, I propose a solution for this problem.\nTo make it clear, I formulate the problem as restoring a subset of the variable from the checkpoint, when the model is built on a pre-trained model.\nFirst, we should use print_tensors_in_checkpoint_file function from the library <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.4\/tensorflow\/python\/tools\/inspect_checkpoint.py\" rel=\"nofollow noreferrer\">inspect_checkpoint<\/a> or just simply extract this function by: <\/p>\n\n<pre><code>from tensorflow.python import pywrap_tensorflow\ndef print_tensors_in_checkpoint_file(file_name, tensor_name, all_tensors):\n    varlist=[]\n    reader = pywrap_tensorflow.NewCheckpointReader(file_name)\n    if all_tensors:\n      var_to_shape_map = reader.get_variable_to_shape_map()\n      for key in sorted(var_to_shape_map):\n        varlist.append(key)\n    return varlist\nvarlist=print_tensors_in_checkpoint_file(file_name=the path of the ckpt file,all_tensors=True,tensor_name=None)\n<\/code><\/pre>\n\n<p>Then we use tf.get_collection() just like @mrry saied:<\/p>\n\n<pre><code>variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)\n<\/code><\/pre>\n\n<p>Finally, we can initialize the saver by:<\/p>\n\n<pre><code>saver = tf.train.Saver(variable[:len(varlist)])\n<\/code><\/pre>\n\n<p>The complete version can be found at my github: <a href=\"https:\/\/github.com\/pobingwanghai\/tensorflow_trick\/blob\/master\/restore_from_checkpoint.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pobingwanghai\/tensorflow_trick\/blob\/master\/restore_from_checkpoint.py<\/a><\/p>\n\n<p>In my situation, the new variables are added at the end of the model, so I can simply use [:length()] to identify the needed variables, for a more complex situation, you might have to do some hand-alignment work or write a simple string matching function to determine the required variables.<\/p>\n"},{"AnswerId":"41642426","CreationDate":"2017-01-13T19:50:19.207","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To restore a subset of variables, you must create a new <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.__init__\" rel=\"noreferrer\"><code>tf.train.Saver<\/code><\/a> and pass it a specific list of variables to restore in the optional <code>var_list<\/code> argument.<\/p>\n\n<p>By default, a <code>tf.train.Saver<\/code> will create ops that (i) save every variable in your graph when you call <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.save\" rel=\"noreferrer\"><code>saver.save()<\/code><\/a> and (ii) lookup (by name) every variable in the given checkpoint when you call <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.restore\" rel=\"noreferrer\"><code>saver.restore()<\/code><\/a>. While this works for most common scenarios, you have to provide more information to work with specific subsets of the variables:<\/p>\n\n<ol>\n<li><p>If you only want to restore a subset of the variables, you can get a list of these variables by calling <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/graph_collections#get_collection\" rel=\"noreferrer\"><code>tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=G_NETWORK_PREFIX)<\/code><\/a>, assuming that you put the \"g\" network in a common <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/utility_functions#name_scope\" rel=\"noreferrer\"><code>with tf.name_scope(G_NETWORK_PREFIX):<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sharing_variables#variable_scope\" rel=\"noreferrer\"><code>tf.variable_scope(G_NETWORK_PREFIX):<\/code><\/a> block. You can then pass this list to the <code>tf.train.Saver<\/code> constructor.<\/p><\/li>\n<li><p>If you want to restore a subset of the variable and\/or they variables in the checkpoint have <strong>different names<\/strong>, you can pass a dictionary as the <code>var_list<\/code> argument. By default, each variable in a checkpoint is associated with a <em>key<\/em>, which is the value of its <code>tf.Variable.name<\/code> property. If the name is different in the target graph (e.g. because you added a scope prefix), you can specify a dictionary that maps string keys (in the checkpoint file) to <code>tf.Variable<\/code> objects (in the target graph).<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"44288910","CreationDate":"2017-05-31T15:31:04.127","ParentId":null,"OwnerUserId":"5281437","Title":null,"Body":"<p>You can create a separate instance of <code>tf.train.Saver()<\/code> with the <code>var_list<\/code> argument set to the variables you want to restore.\nAnd create a separate instance to save the variables<\/p>\n"}]}
{"QuestionId":41621333,"AnswerCount":1,"Tags":"<tensorflow><sequences><lstm>","CreationDate":"2017-01-12T19:25:51.830","AcceptedAnswerId":41625993.0,"OwnerUserId":4646735.0,"Title":"Tensorflow LSTM character by character sequence prediction","Body":"<p>I am attempting to replicate the character level language modeling demonstrated in the excellent article <a href=\"http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/\" rel=\"noreferrer\">http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/<\/a> using Tensorflow.<\/p>\n\n<p>So far my attempts have failed. My network typically outputs a single character after processing 800 or so characters.\nI believe I have fundamentally misunderstood the way tensor flow has implemented LSTMs, and perhaps rnns in general. I am finding the documentation to be difficult to follow.<\/p>\n\n<p>Here is the essence of my code:<\/p>\n\n<p><strong>Graph definition<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Loss Calculation<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Training<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Output, using 1MB Shakespere file to train<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>This is clearly incorrect.<\/p>\n\n<p>I think I am getting confused by the difference between 'batches' and 'sequences', and as to whether or not the state of the LSTM is preserved between what I call 'batches' (ie, sub-sequences)<\/p>\n\n<p>I'm getting the impression that I've trained it using 'batches' of sequences of length 1, and that between each batch, state data is discarded. Consequently it is simply finding the most commonly occurring symbol.<\/p>\n\n<p>Can anyone confirm this, or otherwise correct my mistake, and give some indication of how I should go about the task of character by character prediction using very long training sequences?<\/p>\n\n<p>Many Thanks.<\/p>\n","answers":[{"AnswerId":"41625993","CreationDate":"2017-01-13T01:25:35.960","ParentId":null,"OwnerUserId":"6491684","Title":null,"Body":"<p>So your idata should have a shape of: [batch_size, maximum_sequence_length, 257].\n(If not all sequences in a batch have the same length you need to pad as necessary, and be careful when computing losses that this is done only over non-padded values.)<\/p>\n\n<p>The dynamic_rnn steps through your input by time for you. So, you only need to loop over batches. <\/p>\n\n<p>Since, your second dimension of idata is 1 you are right that your effective sequence length is 1.<\/p>\n\n<p>For a language model not character-based but using word embeddings take a look at <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\/\" rel=\"nofollow noreferrer\">this<\/a> tutorial.<\/p>\n\n<p>Other notes:<\/p>\n\n<ul>\n<li><p>If you want to experiment with different number of units in the LSTM - consider adding a linear layer on top of the output to project each output (for batch entry i at time t) down to 257 which is the number of classes of your target.<\/p><\/li>\n<li><p>No need to do a one-hot encoding of the target. Take a look at sparse_softmax_cross_entropy. <\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":41621340,"AnswerCount":2,"Tags":"<python><arrays><numpy><tensorflow><data-science>","CreationDate":"2017-01-12T19:26:20.630","AcceptedAnswerId":null,"OwnerUserId":7411486.0,"Title":"What is the solution python gives me \"ValueError: setting an array element with a sequence.\"","Body":"<p>I am running the code below but it's giving me an error about arrays. I have tried to find a solution and somehow understand the problem but I couldn't solve the problem. Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>And i get this error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"NeuralNet.py\", line 92, in\n  \n      train_neural_network(x)   File \"NeuralNet.py\", line 83, in train_neural_network\n      _, c = sess.run([optimizer, cross_entropy], feed_dict={x: x_train, y:[t for t in y_train.as_matrix()]})   File\n  \"\/home\/jusuf\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\",\n  line 717, in runrun_metadata_ptr) File\n  \"\/home\/jusuf\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\",\n  line 888, in _run np_val = np.asarray(subfeed_val,\n  dtype=subfeed_dtype) File\n  \"\/home\/jusuf\/anaconda3\/lib\/python3.5\/site-packages\/numpy\/core\/numeric.py\",\n  line 482, in asarray\n      return array(a, dtype, copy=False, order=order) ValueError: setting an array element with a sequence.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"41621618","CreationDate":"2017-01-12T19:42:19.637","ParentId":null,"OwnerUserId":"7382624","Title":null,"Body":"<p>Error message is;<\/p>\n\n<pre><code>ValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>Explains that: You're trying to set an array element with a sequence. Well, where is the mistake, see below;<\/p>\n\n<pre><code>c = sess.run([optimizer, cross_entropy]\n<\/code><\/pre>\n\n<p>So tell me what is the \"c\" . It's probably a float, integer or whatever. But I am pretty sure it is not an array. That's why you are receiving an exception above.<\/p>\n\n<p>But if you want to print out that array, you can do directly;<\/p>\n\n<p><code>print(sess.run([optimizer, cross_entropy])<\/code> instead of run <code>print(c)<\/code> <\/p>\n\n<p>As far as I saw from your code, you are not using \"c\" anywhere.<\/p>\n"},{"AnswerId":"41621693","CreationDate":"2017-01-12T19:46:40.873","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>One action that produces this error is assigning a list to an array element:<\/p>\n\n<pre><code>In [498]: x=np.zeros(3)\nIn [499]: x\nOut[499]: array([ 0.,  0.,  0.])\nIn [500]: x[0] = [1,2,3]\n....\nValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>Since the error is in a <code>np.asarray(subfeed_val, dtype=subfeed_dtype)<\/code> statement it is more likely that it is doing something like:<\/p>\n\n<pre><code>In [502]: np.array([[1,2,3],[1,2]], dtype=int)\nValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>It's still the problem of trying to put a sequence of numbers into one slot.<\/p>\n\n<p>Looking further up the error stack, the error is in:<\/p>\n\n<pre><code>sess.run([optimizer, cross_entropy], feed_dict={x: x_train, y:[t for t in y_train.as_matrix()]})\n<\/code><\/pre>\n\n<p>I don't think it has to do with the assignment to <code>c<\/code>.<\/p>\n\n<p><code>sess.run<\/code> is a <code>tensorflow<\/code> function that I know nothing about.<\/p>\n\n<p>================<\/p>\n\n<p>the error stack, properly formatted is<\/p>\n\n<pre><code>Traceback (most recent call last): \nFile \"NeuralNet.py\", line 92, in train_neural_network(x) \nFile \"NeuralNet.py\", line 83, in train_neural_network\n   _, c = sess.run([optimizer, cross_entropy], feed_dict={x: x_train, y:[t for t in y_train.as_matrix()]}) \nFile \"\/home\/jusuf\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 717, in    \n    runrun_metadata_ptr) \nFile \"\/home\/jusuf\/anaconda3\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\", line 888, in \n   _run np_val = np.asarray(subfeed_val, dtype=subfeed_dtype) \nFile \"\/home\/jusuf\/anaconda3\/lib\/python3.5\/site-packages\/numpy\/core\/numeric.py\", line 482, in \n   asarray return array(a, dtype, copy=False, order=order) \nValueError: setting an array element with a sequence.\n<\/code><\/pre>\n\n<p>I'd suggest reviewing the <code>tensorflow<\/code> documentation, and make sure that inputs to this function are correct.  Focus on allowed types, and if arrays, pay attention to dimensions, shape, and dtype.<\/p>\n"}]}
{"QuestionId":41622883,"AnswerCount":1,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2017-01-12T21:00:37.137","AcceptedAnswerId":null,"OwnerUserId":1537681.0,"Title":"How to get a gradient of a placeholder with default in TensorFlow?","Body":"<p>I am working with an autoencoder. My tensorflow graph looks something like this: (input image) --conv2d--> (compressed representation) --deconv2d--> (reconstructed image).<\/p>\n\n<p>I changed compressed representation tensor to be placeholder_with_default, so that I can plug in any value I want and see what is reconstructed. The default value that should be used when <em>compressed representation<\/em> is not fed is the result of conv2d. <\/p>\n\n<p>However, the optimiser complains that:<\/p>\n\n<pre><\/pre>\n\n<p>How can I let the graph know to just use the usual gradient at the optimisation stage? When <em>compressed representation<\/em> is not a placeholder everything works fine.<\/p>\n","answers":[{"AnswerId":"46817356","CreationDate":"2017-10-18T19:00:20.350","ParentId":null,"OwnerUserId":"7681811","Title":null,"Body":"<p>You don't actually need to use a placeholder to feed in data - any tensor can be the target of the feed_dict argument in session.run(). Hence you should keep the compressed representation as a tensor and just assign it to what you want when you want to provide your own data in the feed_dict argument.<\/p>\n"}]}
{"QuestionId":41623402,"AnswerCount":1,"Tags":"<python><tensorflow><skflow>","CreationDate":"2017-01-12T21:35:34.693","AcceptedAnswerId":null,"OwnerUserId":4529104.0,"Title":"Import skflow on a Windows machine","Body":"<p>I am trying to import skflow on my Windows PC. I have already install and used Anaconda on Python (3.5). I have no trouble to use tensorflow but when I want to use skflow I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know how to fix this?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"41623618","CreationDate":"2017-01-12T21:52:12.110","ParentId":null,"OwnerUserId":"6926393","Title":null,"Body":"<p>Please try this:<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; from tensorflow.contrib import learn as skflow\n<\/code><\/pre>\n\n<p>FYI, skflow is not any special package\/module\/... It's an alias for tensorflow.contrib.learn which is in 'sklearn' style.<\/p>\n"}]}
{"QuestionId":41625325,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2017-01-13T00:11:18.893","AcceptedAnswerId":41625836.0,"OwnerUserId":5551505.0,"Title":"Input images in tensorflow graph as a couple","Body":"<p>I want to input the images in my siamese model as a couple. But if I read them from a list of image names, using <\/p>\n\n<p><\/p>\n\n<p>the images are read read sequentially and their integrity as a couple is destroyed. Here,  is the list of all the images names, such that index 0 and 1 is a pair, 2 and 3 is a pair and so on.<\/p>\n\n<p>Any insights?<\/p>\n","answers":[{"AnswerId":"41625836","CreationDate":"2017-01-13T01:07:51.267","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>Consider using <code>tf.train.batch_join<\/code>, as that maintains grouping between two tensors. See the documentation at: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline<\/a> <\/p>\n"}]}
{"QuestionId":41625833,"AnswerCount":0,"Tags":"<audio><tensorflow><neural-network><keras><conv-neural-network>","CreationDate":"2017-01-13T01:07:41.473","AcceptedAnswerId":null,"OwnerUserId":7412459.0,"Title":"Keras CNN converging on only one classification","Body":"<p>I have been successfully building and  training a standard 4 layer MLP neural network on Keras for an audio classification task. However when I tried to add a convolutional layer to the network, the CNN only successfully trains ~25% of the time. On the other instances, when tested, the network will always predict a single class, giving a 10% accuracy for the 10 class classification problem<\/p>\n\n<p>I am not sure whether the network is not training at all, whether it is converging on some local minima, or whether it is something different altogether.<\/p>\n\n<p>My input data are audio signals ~5 seconds in length, which have been converted into a spectrogram and flattened into a vector. There are 10 classes and the MLP model was able to identify the correct class with a ~55% accuracy (not great, but far better than a random guess).<\/p>\n\n<p>The failed output I am receiving so far looks like this...<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, the network only ever predicted class 5. In addition to this, the loss never actually reduced.<\/p>\n\n<p>In contrast, the successful runs looked like this...<\/p>\n\n<pre><\/pre>\n\n<p>Here is the network I have been running<\/p>\n\n<pre><\/pre>\n\n<p>Thanks in advance for your help. Any assistance is greatly appreciated<\/p>\n","answers":[]}
{"QuestionId":41626542,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-13T02:38:57.280","AcceptedAnswerId":41687434.0,"OwnerUserId":5355831.0,"Title":"Tensorflow: InvalidArgumentError with MNIST, [55000] vs. [10000]","Body":"<p>I'm working through the code from this presentation <a href=\"http:\/\/www.youtube.com\/watch?v=vq2nnJ4g6N0&amp;t=20m28s\" rel=\"nofollow noreferrer\">http:\/\/www.youtube.com\/watch?v=vq2nnJ4g6N0&amp;t=20m28s<\/a> and am getting this error: <\/p>\n\n<p>I've already worked through a few errors regarding tensor shape\/dimensions, but have no idea how to specifically understand this let alone correct it.<\/p>\n\n<p>I'm new to tf and any advice is much appreciated, here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>The total error output is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41687434","CreationDate":"2017-01-17T00:54:27.310","ParentId":null,"OwnerUserId":"5098368","Title":null,"Body":"<p>You are feeding training data (number of rows: 55,000) instead of testing data (number of rows: 10,000) in this line:<\/p>\n\n<pre><code>test_data = {X: mnist.train.images, Y_: mnist.test.labels}\n<\/code><\/pre>\n\n<p>Just fix it with: <\/p>\n\n<pre><code>test_data = {X: mnist.test.images, Y_: mnist.test.labels}\n<\/code><\/pre>\n"}]}
{"QuestionId":41626651,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-01-13T02:50:46.263","AcceptedAnswerId":41657448.0,"OwnerUserId":1050648.0,"Title":"Error loading saved checkpoint in TFSlim","Body":"<p>Using v0.12.1 version of tensorflow, I'm trying to finetune a pre-trained vgg16 model using checkpoint available at  <a href=\"http:\/\/download.tensorflow.org\/models\/vgg_16_2016_08_28.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/vgg_16_2016_08_28.tar.gz<\/a>.<\/p>\n\n<p>I'm getting the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>Need help. I posted a similar question a few months ago - I'm past that but couldn't find a good solution. I use the model definition given in <\/p>\n\n<pre><\/pre>\n\n<p>I understand that there are two versions of ckpt files..v1 and v2. could this be an issue? How to solve this please?<\/p>\n","answers":[{"AnswerId":"41657448","CreationDate":"2017-01-15T03:15:52.583","ParentId":null,"OwnerUserId":"1050648","Title":null,"Body":"<p>There was a mismatch between the variables generated by the model definition and the variables stored in the ckpt file. Specifically, this was the first variable. Solved it by doing this:<\/p>\n\n<pre><code>variables_to_restore = slim.get_variables_to_restore(exclude=['vgg_16\/fc6','vgg_16\/fc7','vgg_16\/fc8'])\nprint [v.name for v in variables_to_restore]\nrestorer = tf.train.Saver(variables_to_restore[1:]) # remove first entry  !\n<\/code><\/pre>\n"}]}
{"QuestionId":41626830,"AnswerCount":4,"Tags":"<tensorflow><pip>","CreationDate":"2017-01-13T03:13:27.420","AcceptedAnswerId":null,"OwnerUserId":5998842.0,"Title":"Pip only install cpu tensorflow of tensorflow 0.11","Body":"<p>I previously installed  which worked fine, but for a code of a colleague I need . So I uninstalled tensorflow and tensorflow-gpu 0.12 and I tried to install v 0.11 with:<\/p>\n\n<pre><\/pre>\n\n<p>It successfully installs tensorflow, but not the gpu version, even though in the link above I took the gpu and not cpu version. Apparently, it doesn't matter if I choose cpu or gpu, it always only installs the cpu version.<\/p>\n\n<p>Any idea how to fix this problem?<\/p>\n","answers":[{"AnswerId":"45034852","CreationDate":"2017-07-11T12:37:05.407","ParentId":null,"OwnerUserId":"8289596","Title":null,"Body":"<p>On running <\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.11.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>I can successfully install <code>tensorflow v 0.11<\/code>.<\/p>\n"},{"AnswerId":"47667567","CreationDate":"2017-12-06T05:43:58.163","ParentId":null,"OwnerUserId":"4179779","Title":null,"Body":"<p>Thaks for @user8289596 answer the following command was useful for my case and i have successfully installed tensorflow 0.11<\/p>\n\n<p><code>\npip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.11.0rc0-cp27-none-linux_x86_64.whl<\/code><\/p>\n\n<p>Note: I am using anaconda with python 2.7 and linux<\/p>\n"},{"AnswerId":"44438335","CreationDate":"2017-06-08T14:21:27.047","ParentId":null,"OwnerUserId":"4138044","Title":null,"Body":"<p>For tensorflow versions <code>0.11.x<\/code> and below there isn't any <code>tensorflow-gpu<\/code> package in pip. The wheel corresponding to <code>https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow-0.11.0-cp27-none-linux_x86_64.whl<\/code> is the correct wheel with <code>gpu<\/code> support. If you install this wheel via pip, it'll say <code>Installed tensorflow==0.11.0<\/code> but has support of CUDA-8. By downloading the individual wheels, for the <code>cpu<\/code> and <code>gpu<\/code> folder have a different md5 signature, hence they are different. <\/p>\n\n<p>See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/10485\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"54637839","CreationDate":"2019-02-11T19:35:54.830","ParentId":null,"OwnerUserId":"6069097","Title":null,"Body":"<p>For MacOS, Python 3, this worked for me:<\/p>\n\n<pre><code>export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.11.0-py3-none-any.whl\n<\/code><\/pre>\n\n<p>then install with:<\/p>\n\n<pre><code>sudo pip3 install --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/chromium.googlesource.com\/external\/github.com\/tensorflow\/tensorflow\/+\/r0.11\/tensorflow\/g3doc\/get_started\/os_setup.md\" rel=\"nofollow noreferrer\">Source<\/a> also has instructions for Python 2 and 3 on a variety of Ubuntu\/Linux and MacOS configurations.<\/p>\n"}]}
{"QuestionId":41627738,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-01-13T05:10:04.853","AcceptedAnswerId":41628828.0,"OwnerUserId":1761806.0,"Title":"Why is there no mention of contrib.layers.linear in the Tensorflow documentation?","Body":"<p>I'm trying to understand someone else's simple tensorflow model and they make use of contrib.layers.linear.<\/p>\n\n<p>However I cannot find any information on this anywhere and it's not mentioned in the tensorflow documentation.<\/p>\n","answers":[{"AnswerId":"42384536","CreationDate":"2017-02-22T06:49:42.183","ParentId":null,"OwnerUserId":"4218883","Title":null,"Body":"<p>According to what I can see in the Master branch, the function linear still exists in <code>contrib.layers<\/code>. It actually is a \"simple alias which removes the <code>activation_fn<\/code> parameter\":<\/p>\n\n<p><code>linear = functools.partial(fully_connected, activation_fn=None)<\/code><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L2140\" rel=\"nofollow noreferrer\">Here<\/a> is a link from the 1.0 branch (to increase link persistence).<\/p>\n\n<p>Though, if the <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.layers\" rel=\"nofollow noreferrer\">doc<\/a> still shows it, the link to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/linear\" rel=\"nofollow noreferrer\">contrib.layers.linear<\/a> seems indeed broken.<\/p>\n"},{"AnswerId":"41627739","CreationDate":"2017-01-13T05:10:04.853","ParentId":null,"OwnerUserId":"1761806","Title":null,"Body":"<p>I managed to find the answer and felt it was still worth posting this to save others wasting their time.<\/p>\n\n<p>\"In general, tf.contrib contains contributed code. It is meant to contain features and contributions that eventually should get merged into core TensorFlow, but whose interfaces may still change, or which require some testing to see whether they can find broader acceptance.<\/p>\n\n<p>Code in tf.contrib isn't supported by the Tensorflow team. It is included in the hope that it is helpful, but it might change or be removed at any time; there are no guarantees.\" <a href=\"https:\/\/stackoverflow.com\/questions\/38060825\/tensorflow-what-is-tf-contrib-and-where-can-i-find-the-source-code-for-tf-con\">source<\/a><\/p>\n"},{"AnswerId":"41628828","CreationDate":"2017-01-13T06:43:56.173","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>tf.contrib.layers<\/code> module has API documentation <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.layers\/\" rel=\"noreferrer\">here<\/a>. As you observed in <a href=\"https:\/\/stackoverflow.com\/a\/41627739\/3574081\">your answer<\/a>, the <code>contrib<\/code> APIs in TensorFlow are (especially) subject to change. The <code>tf.contrib.layers.linear()<\/code> function appears to have been removed, but you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/fully_connected\" rel=\"noreferrer\"><code>tf.contrib.layers.fully_connected(\u2026, activation_fn=None)<\/code><\/a> to achieve the same effect.<\/p>\n"}]}
{"QuestionId":41627902,"AnswerCount":1,"Tags":"<c++><makefile><caffe>","CreationDate":"2017-01-13T05:29:02.500","AcceptedAnswerId":null,"OwnerUserId":7410281.0,"Title":"Error message during installing caffe command 'make all' about \"caffe::cudnn::dataType<double>::zero\"","Body":"<p>I ran <\/p>\n\n<pre><\/pre>\n\n<p>as suggested on the website to complete the installation. I use Ubuntu 15.04 with CUDA and atlas.<\/p>\n\n<p>The error messages showed as follows<\/p>\n\n<blockquote>\n  <p>CXX\/LD -o .build_release\/tools\/compute_image_mean.bin\n        CXX\/LD -o .build_release\/tools\/convert_imageset.bin\n        CXX\/LD -o .build_release\/tools\/extract_features.bin\n        CXX\/LD -o .build_release\/tools\/upgrade_solver_proto_text.bin\n        CXX\/LD -o .build_release\/tools\/upgrade_net_proto_text.bin\n        CXX\/LD -o .build_release\/tools\/upgrade_net_proto_binary.bin\n        CXX\/LD -o .build_release\/tools\/caffe.bin\n        CXX\/LD -o .build_release\/examples\/cpp_classification\/classification.bin\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      collect2: error: ld returned 1 exit status\n      Makefile:619: recipe for target '.build_release\/tools\/upgrade_solver_proto_text.bin' failed\n      make: <strong>* [.build_release\/tools\/upgrade_solver_proto_text.bin] Error 1\n      make: <em><\/strong> \u6b63\u5728\u7b49\u5f85\u672a\u5b8c\u6210\u7684\u4efb\u52a1....\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      collect2: error: ld returned 1 exit status\n      Makefile:619: recipe for target '.build_release\/tools\/compute_image_mean.bin' failed\n      make: <strong><\/em> [.build_release\/tools\/compute_image_mean.bin] Error 1\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      collect2: error: ld returned 1 exit status\n      Makefile:619: recipe for target '.build_release\/tools\/upgrade_net_proto_binary.bin' failed\n      make: <em><\/strong> [.build_release\/tools\/upgrade_net_proto_binary.bin] Error 1\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      collect2: error: ld returned 1 exit status\n      Makefile:619: recipe for target '.build_release\/tools\/upgrade_net_proto_text.bin' failed\n      make: <strong><\/em> [.build_release\/tools\/upgrade_net_proto_text.bin] Error 1\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      collect2: error: ld returned 1 exit status\n      Makefile:619: recipe for target '.build_release\/tools\/convert_imageset.bin' failed\n      make: <em><\/strong> [.build_release\/tools\/convert_imageset.bin] Error 1\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      collect2: error: ld returned 1 exit status\n      Makefile:624: recipe for target '.build_release\/examples\/cpp_classification\/classification.bin' failed\n      make: <strong><\/em> [.build_release\/examples\/cpp_classification\/classification.bin] Error 1\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      collect2: error: ld returned 1 exit status\n      Makefile:619: recipe for target '.build_release\/tools\/caffe.bin' failed\n      make: *<\/strong> [.build_release\/tools\/caffe.bin] Error 1\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::zero\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      .build_release\/lib\/libcaffe.so\uff1a\u5bf9\u2018caffe::cudnn::dataType::one\u2019\u672a\u5b9a\u4e49\u7684\u5f15\u7528\n      collect2: error: ld returned 1 exit status\n      Makefile:619: recipe for target '.build_release\/tools\/extract_features.bin' failed\n      make: *** [.build_release\/tools\/extract_features.bin] Error 1<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42061464","CreationDate":"2017-02-06T06:22:23.377","ParentId":null,"OwnerUserId":"4192309","Title":null,"Body":"<p>I am not sure about your case, but my solution was to uncomment the line with <\/p>\n\n<pre><code># USE_OPENCV := 0\nOPENCV_VERSION := 3\n<\/code><\/pre>\n\n<p>in 'Makefile.config' since I was with OpenCV version > 3.0.<\/p>\n\n<p>make clean and make all -j $(($(nproc) + 1)) with all your processor. It should be fine afterwards.<\/p>\n"}]}
{"QuestionId":41628733,"AnswerCount":1,"Tags":"<python><csv><tensorflow>","CreationDate":"2017-01-13T06:37:26.857","AcceptedAnswerId":41628940.0,"OwnerUserId":6815505.0,"Title":"Using my own .csv in tensorflow","Body":"<p>I asked a previous question about the same code here <a href=\"https:\/\/stackoverflow.com\/questions\/41604666\/where-and-how-to-put-the-filename-in-this-tensorflow-code\/41604790?noredirect=1#comment70413582_41604790\">where and how to put the filename in this tensorflow code?<\/a> <\/p>\n\n<p>Not sure if I should merge that into this question or leave it as is. <\/p>\n\n<p>The following code is from Sirajology's git hub. I have not found a super straight forward tutorial on how to get one's own .csv file into a simple tensorflow neural network so my hope is this thread might provide that instruction for future searchers.<\/p>\n\n<p>The code is as follows<\/p>\n\n<pre><\/pre>\n\n<p>When I run the code from terminal with the following command (windows10 cmd line)  I get these errors. Any help is greatly appreciated!<\/p>\n\n<p><strong>Error #1<\/strong>\n  File \"softmax.py\", line 133, in \n    tf.app.run()<\/p>\n\n<pre><\/pre>\n\n<p><strong>Error #2<\/strong>\n File \"C:\\app.py\", line 43, in run\n    sys.exit(main(sys.argv[:1] + flags_passthrough))<\/p>\n\n<pre><\/pre>\n\n<p><strong>Error #3<\/strong>\n File \"softmax.py\", line 57, in main\n    train_data,train_labels = extract_data(train_data_filename) <\/p>\n\n<pre><\/pre>\n\n<p><strong>Error #4<\/strong> \n File \"softmax.py\", line 31, in extract_data\n    for line in file(filename):\nNameError: name 'file' is not defined<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41628940","CreationDate":"2017-01-13T06:52:15.913","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It looks like the problem stems from this line, which uses a built-in function (<code>file()<\/code>) that is not available in Python 3.5:<\/p>\n\n<pre><code>for line in file(filename):\n<\/code><\/pre>\n\n<p>Replacing it with the following line should fix the error:<\/p>\n\n<pre><code>for line in open(filename):\n<\/code><\/pre>\n"}]}
{"QuestionId":41628883,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2017-01-13T06:48:56.527","AcceptedAnswerId":41640238.0,"OwnerUserId":1992451.0,"Title":"Does Tensorflow TFDetect demo track objects end-to-end?","Body":"<p>This question is regarding TFDetect demo which is available as a part of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\/\" rel=\"nofollow noreferrer\">Tensorflow Android Camera Demo<\/a>. The description says,<\/p>\n\n<blockquote>\n  <p>Demonstrates a model based on Scalable Object Detection using Deep\n  Neural Networks to localize and <strong>track<\/strong> people in the camera preview in\n  real-time.<\/p>\n<\/blockquote>\n\n<p>When I ran the demo, the app was creating boxes around detected objects with a fractional number assigned to each object (I guess the confidence score). My question is, how is tracking being performed here. Is it multiple object tracking (described <a href=\"https:\/\/motchallenge.net\/\" rel=\"nofollow noreferrer\">here<\/a>) where there is an id assigned to each track and the tracks are stored in the memory, or is it just detection of objects across multiple frames to see how the object is moving?<\/p>\n\n<p>Please correct me if I missed out on anything.<\/p>\n","answers":[{"AnswerId":"41640238","CreationDate":"2017-01-13T17:29:49.897","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>Two main things are going on here:<\/p>\n\n<p>1: detection is being done in a background thread. This takes on the order of 100-1000ms depending on device, so is not sufficient to maintain smooth tracking.<\/p>\n\n<p>2: tracking is being done in the UI thread. This generally takes &lt; 5ms per frame, and can be done on every frame once the position of objects is known. The tracker implements pyrimidal lucas-kanade optical flow on the median movement of FAST features -- press the volume key and you'll see the individual features being tracked.<\/p>\n\n<p>The tracker runs on every frame, storing optical flow keypoints at every timestamp. Thus, when a detection comes in, the tracker is able to figure out where it currently is by walking the position forward along the collected keypoints deltas. There is also some non-max suppression being done as well by multiboxtracker.<\/p>\n\n<p>Once an object is tracked by the tracker, no further input from the detector is required. The tracker will automatically drop the track when the normalized cross-correlation with the original detection drops below a certain threshold, or update the position\/appearance when the detector finds a better match with significant overlap.<\/p>\n"}]}
{"QuestionId":41629745,"AnswerCount":1,"Tags":"<deep-learning><caffe><nvidia-digits>","CreationDate":"2017-01-13T07:50:40.680","AcceptedAnswerId":null,"OwnerUserId":6675900.0,"Title":"How to use latest version of Caffe in DIGITS?","Body":"<p>The Caffe from NVidia is older than BVLC\/Caffe. How can I use the latest version of Caffe in NVidia DIGITS?<\/p>\n","answers":[{"AnswerId":"41640142","CreationDate":"2017-01-13T17:23:57.920","ParentId":null,"OwnerUserId":"2404152","Title":null,"Body":"<p>Either set the <code>CAFFE_ROOT<\/code> environment variable, or install your build of Caffe to PATH and PYTHONPATH.<\/p>\n\n<p><a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\/blob\/digits-5.0\/docs\/BuildCaffe.md#download-source\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NVIDIA\/DIGITS\/blob\/digits-5.0\/docs\/BuildCaffe.md#download-source<\/a>\n<a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\/blob\/digits-5.0\/docs\/Configuration.md#environment-variables\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NVIDIA\/DIGITS\/blob\/digits-5.0\/docs\/Configuration.md#environment-variables<\/a><\/p>\n"}]}
{"QuestionId":41630022,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-13T08:11:10.650","AcceptedAnswerId":41632564.0,"OwnerUserId":6334174.0,"Title":"Using placeholder as shape in tensorflow","Body":"<p>I'm try to define a two dimensional placeholder in tensorflow, However, I don't know the size of that in advance. Hence I define another placeholder, but it seems it doesn't work at all. Here is the minimum example:<\/p>\n\n<pre><\/pre>\n\n<p>Error message:<\/p>\n\n<pre><\/pre>\n\n<p>Then I tried to pack the shape, so I have this:<\/p>\n\n<pre><\/pre>\n\n<p>doesn't work either:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41632564","CreationDate":"2017-01-13T10:35:25.770","ParentId":null,"OwnerUserId":"1939748","Title":null,"Body":"<p>Use <code>None<\/code> if you don't know the length in some dimension in advance, e.g.<\/p>\n\n<pre><code>input = tf.placeholder(tf.int32, [None, 5])\n<\/code><\/pre>\n\n<p>When you feed this placeholder a proper array of shape (batch_size, 5), it's dynamic shape will be set correctly, i.e.<\/p>\n\n<pre><code>sess.run(tf.shape(input), feed_dict={input: np.zeros(dtype=np.int32, shape=(10, 5))})\n<\/code><\/pre>\n\n<p>will return<\/p>\n\n<pre><code>array([10,  5], dtype=int32)\n<\/code><\/pre>\n\n<p>as expected<\/p>\n"}]}
{"QuestionId":41630086,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><deep-learning><recurrent-neural-network>","CreationDate":"2017-01-13T08:16:01.977","AcceptedAnswerId":null,"OwnerUserId":6119768.0,"Title":"Tensorflow RNN accuracy and cost: NaN","Body":"<p>I trained a RNN for some signals and their one hot encoded labels. After working through some size and shape mismatches, the network running without error. However I am getting NaN for loss and accuracy.<\/p>\n\n<pre><\/pre>\n\n<p>Here is how I set my input data and labels-<\/p>\n\n<pre><\/pre>\n\n<p>I am feeding the data in batches like this-<\/p>\n\n<pre><\/pre>\n\n<p>After running the code I either get very small cost or NaN. For accuracy I get NaN every time.\nI have checked some of the related stack questions but cannot understand the problem. Please help.<\/p>\n\n<p>My full code gist is <a href=\"https:\/\/gist.github.com\/tu-lee\/9eb8bbfc81317bd3f288e207de7feb5a\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n","answers":[]}
{"QuestionId":41630475,"AnswerCount":0,"Tags":"<c++><tensorflow>","CreationDate":"2017-01-13T08:42:11.177","AcceptedAnswerId":null,"OwnerUserId":187676.0,"Title":"convert_graphdef_memmapped_format produces corrupt graphs","Body":"<p>I tried memory mapping an exported\/frozen graph. Let's say the graph is named . After converting the graph like this<\/p>\n\n<pre><\/pre>\n\n<p>I get<\/p>\n\n<ul>\n<li> for an unquantized graph<\/li>\n<li> for the 8bit quantized version of the graph<\/li>\n<\/ul>\n\n<p>If I try to load any of these output files using the C++ API  and then  the graph the following errors occur<\/p>\n\n<ul>\n<li>Unquantized: <\/li>\n<li>Quantized: <\/li>\n<\/ul>\n\n<p>so I guess the output file is a corrupt protobuf file (or a different version).<\/p>\n\n<p>Anyone else experienced this and know how to resolve it?<\/p>\n","answers":[]}
{"QuestionId":41632785,"AnswerCount":0,"Tags":"<python><image-processing><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2017-01-13T10:46:25.170","AcceptedAnswerId":null,"OwnerUserId":3598726.0,"Title":"How to train tensorflow network with 0D (scalar) labels?","Body":"<p>I have a task, where I'd like to count things on an image. I have labels for the images, which are integer values, but I have never seen such kind of tutorial yet. All the examples are cases, where the final output of the tf network, and the labels are 1D vector objects (one-hot vectors usually). Can you help me with a python code\/tutorial\/paper, where the labels are scalars?<\/p>\n","answers":[]}
{"QuestionId":41633087,"AnswerCount":1,"Tags":"<neural-network><keras><keras-layer>","CreationDate":"2017-01-13T11:01:36.700","AcceptedAnswerId":41645732.0,"OwnerUserId":1211174.0,"Title":"How to use keras ImageDataGenerator with a Siamese or Tripple networks","Body":"<p>I'm trying to build up both a Siamese neural network and <a href=\"https:\/\/arxiv.org\/pdf\/1412.6622v3.pdf\" rel=\"noreferrer\">triple neural network<\/a>  on a custom large dataset<\/p>\n\n<p> has  which makes the generation of input data to a  neural network very easy. <\/p>\n\n<p>I'm interesting to use  or similar ways in order to train a networks with 2() and 3()  inputs.<\/p>\n\n<p>In <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/mnist_siamese_graph.py\" rel=\"noreferrer\">mniset keras siamese example<\/a>, The input generated by a pre-process stage which is done by  method. I don't think this kind of way fit for a large dataset.<\/p>\n\n<p>Is it possible to use  in this case? What are my other options assuming the data-set is very big?<\/p>\n","answers":[{"AnswerId":"41645732","CreationDate":"2017-01-14T01:07:59.183","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>The idea of DataGenerators is to give <code>fit_generator<\/code> a stream of data in batches.. hence giving control to you how you want to produce the data, ie whether you load from files or you do some data augmentation like what is done in <code>ImageDataGenerator<\/code>. <\/p>\n\n<p>Here I posting the modified version of mniset siamese example with custom DataGenerator, you can work it out from here.<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code>import numpy as np\nnp.random.seed(1337)  # for reproducibility\n\nimport random\nfrom keras.datasets import mnist\nfrom keras.models import Sequential, Model\nfrom keras.layers import Dense, Dropout, Input, Lambda\nfrom keras.optimizers import SGD, RMSprop\nfrom keras import backend as K\n\nclass DataGenerator(object):\n    \"\"\"docstring for DataGenerator\"\"\"\n    def __init__(self, batch_sz):\n        # the data, shuffled and split between train and test sets\n        (X_train, y_train), (X_test, y_test) = mnist.load_data()\n        X_train = X_train.reshape(60000, 784)\n        X_test = X_test.reshape(10000, 784)\n        X_train = X_train.astype('float32')\n        X_test = X_test.astype('float32')\n        X_train \/= 255\n        X_test \/= 255\n\n        # create training+test positive and negative pairs\n        digit_indices = [np.where(y_train == i)[0] for i in range(10)]\n        self.tr_pairs, self.tr_y = self.create_pairs(X_train, digit_indices)\n\n        digit_indices = [np.where(y_test == i)[0] for i in range(10)]\n        self.te_pairs, self.te_y = self.create_pairs(X_test, digit_indices)\n\n        self.tr_pairs_0 = self.tr_pairs[:, 0]\n        self.tr_pairs_1 = self.tr_pairs[:, 1]\n        self.te_pairs_0 = self.te_pairs[:, 0]\n        self.te_pairs_1 = self.te_pairs[:, 1]\n\n        self.batch_sz = batch_sz\n        self.samples_per_train  = (self.tr_pairs.shape[0]\/self.batch_sz)*self.batch_sz\n        self.samples_per_val    = (self.te_pairs.shape[0]\/self.batch_sz)*self.batch_sz\n\n\n        self.cur_train_index=0\n        self.cur_val_index=0\n\n    def create_pairs(self, x, digit_indices):\n        '''Positive and negative pair creation.\n        Alternates between positive and negative pairs.\n        '''\n        pairs = []\n        labels = []\n        n = min([len(digit_indices[d]) for d in range(10)]) - 1\n        for d in range(10):\n            for i in range(n):\n                z1, z2 = digit_indices[d][i], digit_indices[d][i+1]\n                pairs += [[x[z1], x[z2]]]\n                inc = random.randrange(1, 10)\n                dn = (d + inc) % 10\n                z1, z2 = digit_indices[d][i], digit_indices[dn][i]\n                pairs += [[x[z1], x[z2]]]\n                labels += [1, 0]\n        return np.array(pairs), np.array(labels)\n\n    def next_train(self):\n        while 1:\n            self.cur_train_index += self.batch_sz\n            if self.cur_train_index &gt;= self.samples_per_train:\n                self.cur_train_index=0\n            yield ([    self.tr_pairs_0[self.cur_train_index:self.cur_train_index+self.batch_sz], \n                        self.tr_pairs_1[self.cur_train_index:self.cur_train_index+self.batch_sz]\n                    ],\n                    self.tr_y[self.cur_train_index:self.cur_train_index+self.batch_sz]\n                )\n\n    def next_val(self):\n        while 1:\n            self.cur_val_index += self.batch_sz\n            if self.cur_val_index &gt;= self.samples_per_val:\n                self.cur_val_index=0\n            yield ([    self.te_pairs_0[self.cur_val_index:self.cur_val_index+self.batch_sz], \n                        self.te_pairs_1[self.cur_val_index:self.cur_val_index+self.batch_sz]\n                    ],\n                    self.te_y[self.cur_val_index:self.cur_val_index+self.batch_sz]\n                )\n\ndef euclidean_distance(vects):\n    x, y = vects\n    return K.sqrt(K.sum(K.square(x - y), axis=1, keepdims=True))\n\n\ndef eucl_dist_output_shape(shapes):\n    shape1, shape2 = shapes\n    return (shape1[0], 1)\n\n\ndef contrastive_loss(y_true, y_pred):\n    '''Contrastive loss from Hadsell-et-al.'06\n    http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/hadsell-chopra-lecun-06.pdf\n    '''\n    margin = 1\n    return K.mean(y_true * K.square(y_pred) + (1 - y_true) * K.square(K.maximum(margin - y_pred, 0)))\n\n\ndef create_base_network(input_dim):\n    '''Base network to be shared (eq. to feature extraction).\n    '''\n    seq = Sequential()\n    seq.add(Dense(128, input_shape=(input_dim,), activation='relu'))\n    seq.add(Dropout(0.1))\n    seq.add(Dense(128, activation='relu'))\n    seq.add(Dropout(0.1))\n    seq.add(Dense(128, activation='relu'))\n    return seq\n\n\ndef compute_accuracy(predictions, labels):\n    '''Compute classification accuracy with a fixed threshold on distances.\n    '''\n    return labels[predictions.ravel() &lt; 0.5].mean()\n\n\ninput_dim = 784\nnb_epoch = 20\nbatch_size=128\n\ndatagen = DataGenerator(batch_size)\n\n# network definition\nbase_network = create_base_network(input_dim)\n\ninput_a = Input(shape=(input_dim,))\ninput_b = Input(shape=(input_dim,))\n\n# because we re-use the same instance `base_network`,\n# the weights of the network\n# will be shared across the two branches\nprocessed_a = base_network(input_a)\nprocessed_b = base_network(input_b)\n\ndistance = Lambda(euclidean_distance, output_shape=eucl_dist_output_shape)([processed_a, processed_b])\n\nmodel = Model(input=[input_a, input_b], output=distance)\n\n# train\nrms = RMSprop()\nmodel.compile(loss=contrastive_loss, optimizer=rms)\nmodel.fit_generator(generator=datagen.next_train(), samples_per_epoch=datagen.samples_per_train, nb_epoch=nb_epoch, validation_data=datagen.next_val(), nb_val_samples=datagen.samples_per_val)\n<\/code><\/pre>\n"}]}
{"QuestionId":41633295,"AnswerCount":1,"Tags":"<keras><lstm><recurrent-neural-network><stateful>","CreationDate":"2017-01-13T11:11:58.827","AcceptedAnswerId":null,"OwnerUserId":6893857.0,"Title":"LSTM state within a batch","Body":"<p>I'm currently working on a <a href=\"http:\/\/machinelearningmastery.com\/understanding-stateful-lstm-recurrent-neural-networks-python-keras\/\" rel=\"nofollow noreferrer\">Keras tutorial for recurrent network training<\/a> and I'm having trouble understanding the Stateful LSTM concept. To keep things as simple as possible, the sequences have the same length . As far as I get it, the input data is of shape  and we then train our LSTM on  batches of size  as follows:<\/p>\n\n<p>For each batch:<\/p>\n\n<ol>\n<li>Feed in the 2D-tensors  and for each input 2D-tensor compute the gradient<\/li>\n<li>Sum these gradients to get the total gradient on the batch<\/li>\n<li>Backpropagate the gradient and update weights<\/li>\n<\/ol>\n\n<p>In the tutorial's example, feeding in the 2D-tensors is feeding in a sequence of size  of letters encoded as vectors of length . However, the tutorial says that in the Keras implementation of LSTMs, the hidden state is not reset after a whole sequence (2D-tensor) is fed in, but after a batch of sequences is fed in to use more context. <\/p>\n\n<p>Why does keeping the hidden state of the previous sequence and using it as initial hidden state for our current sequence improve the learning and the predictions on our test set, since that \"previously learned\" initial hidden state won't be available when making predictions ? Moreover, Keras' default behaviour is to shuffle input samples at the beginning of each epoch so the batch context is changed at each epoch. This behaviour seems contradictory to keeping the hidden state through a batch since batch context is random.  <\/p>\n","answers":[{"AnswerId":"41655240","CreationDate":"2017-01-14T21:34:17.143","ParentId":null,"OwnerUserId":"3697202","Title":null,"Body":"<p>LSTMs in Keras aren't stateful by default - each sequence starts with newly-reset states. By setting <code>stateful=True<\/code> in your recurrent layer, successive inputs in a batch don't reset the network state. This assumes that the sequences are actually successive, and it means that in a (very informal) sense, you're training on sequences of length <code>batch_size * seq_length<\/code>.<\/p>\n\n<blockquote>\n  <p>Why does keeping the hidden state of the previous sequence and using\n  it as initial hidden state for our current sequence improve the\n  learning and the predictions on our test set, since that \"previously\n  learned\" initial hidden state won't be available when making\n  predictions ?<\/p>\n<\/blockquote>\n\n<p>In theory, it improves learning because a longer context can teach the network things about the distribution that are still relevant when testing on the individually shorter sequences. If the network is learning some probability distribution, that distribution should hold over different sequence lengths.<\/p>\n\n<blockquote>\n  <p>Moreover, Keras's default behaviour is to shuffle input samples at the\n  beginning of each epoch so the batch context is changed at each epoch.\n  This behaviour seems contradictory to keeping the hidden state through\n  a batch since batch context is random.<\/p>\n<\/blockquote>\n\n<p>I haven't checked, but I assume that when <code>stateful=True<\/code>, only batches are shuffled - not the sequences within them.<\/p>\n\n<p>In general, when we give the network some initial state, we don't mean for that to be a universally better starting point. It just means that the network can take the information from previous sequences into account when training.<\/p>\n"}]}
{"QuestionId":41633748,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><google-cloud-ml>","CreationDate":"2017-01-13T11:34:18.750","AcceptedAnswerId":41637685.0,"OwnerUserId":4543908.0,"Title":"Load numpy array in google-cloud-ml job","Body":"<p>In the model I want to launch, I have some variables which have to be initialized with specific values.<\/p>\n\n<p>I currently store these variables into numpy arrays but I don't know how to adapt my code to make it work on a google-cloud-ml job.<\/p>\n\n<p>Currently I initialize my variable like this:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone help me ?<\/p>\n","answers":[{"AnswerId":"56798661","CreationDate":"2019-06-27T21:27:03.317","ParentId":null,"OwnerUserId":"10532410","Title":null,"Body":"<p>I tried the accepted answer but ran into some problems.  Eventually this worked for me (Python 3): <\/p>\n\n<pre><code>from io import BytesIO\nimport numpy as np\nfrom tensorflow.python.lib.io import file_io\n<\/code><\/pre>\n\n<p>To save: <\/p>\n\n<pre><code>dest = 'gs:\/\/[BUCKET-NAME]\/' # Destination to save in GCS\nnp.save(file_io.FileIO(dest, 'w'), np.ones((100, )))\n<\/code><\/pre>\n\n<p>To load:<\/p>\n\n<pre><code>f = BytesIO(file_io.read_file_to_string(src, binary_mode=True))\narr = np.load(f)\n<\/code><\/pre>\n"},{"AnswerId":"41637685","CreationDate":"2017-01-13T15:10:32.310","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>First, you'll need to copy\/store the data on GCS (using, e.g., <code>gsutil<\/code>) and ensure your training script has access to that bucket. The easiest way to do so is to copy the array to the same bucket as your data, since you'll likely already have configured that bucket for read access. If the bucket is in the same project as your training job and you followed <a href=\"https:\/\/cloud.google.com\/ml\/docs\/how-tos\/getting-set-up\" rel=\"nofollow noreferrer\">these<\/a> instructions (particularly, <code>gcloud beta ml init-project<\/code>), you should be set. If the data will be in another bucket, see <a href=\"https:\/\/cloud.google.com\/ml\/docs\/how-tos\/using-external-buckets\" rel=\"nofollow noreferrer\">these<\/a> instructions.<\/p>\n\n<p>Then you'll need to use a library capable of loading data from GCS. Tensorflow includes a module that can do this, although you're free to use any client library that can read from GCS. Here's an example of using TensorFlow's <code>file_io<\/code> module:<\/p>\n\n<pre><code>from StringIO import StringIO\nimport tensorflow as tf\nimport numpy as np\nfrom tensorflow.python.lib.io import file_io\n\n# Create a variable initialized to the value of a serialized numpy array\nf = StringIO(file_io.read_file_to_string('gs:\/\/my-bucket\/123.npy'))\nmy_variable = tf.Variable(initial_value=np.load(f), name='my_variable')\n<\/code><\/pre>\n\n<p>Note that we have to read the file into a string and use <code>StringIO<\/code>, since <code>file_io.FileIO<\/code> does not fully implement the seek function required by <code>numpy.load<\/code>.<\/p>\n\n<p>Bonus: in case it's useful, you can directly store a numpy array to GCS using the <code>file_io<\/code> module, e.g.:<\/p>\n\n<pre><code>np.save(file_io.FileIO('gs:\/\/my-bucket\/123', 'w'), np.array([[1,2,3], [4,5,6]]))\n<\/code><\/pre>\n\n<p>For Python 3, use <code>from io import StringIO<\/code> instead of <code>from StringIO import StringIO<\/code>.<\/p>\n"}]}
{"QuestionId":41634674,"AnswerCount":2,"Tags":"<tensorflow><gpu><distributed-system>","CreationDate":"2017-01-13T12:27:54.180","AcceptedAnswerId":41638727.0,"OwnerUserId":4844184.0,"Title":"Tensorflow on shared GPUs: how to automatically select the one that is unused","Body":"<p>I have access through ssh to a cluster of n GPUs. Tensorflow automatically gave them names gpu:0,...,gpu:(n-1).<\/p>\n\n<p>Others have access too and sometimes they take random gpus.\nI did not place any  explicitely because that is cumbersome and even if I selected gpu number j and that someone is already on gpu number j that would be problematic.<\/p>\n\n<p>I would like to go throuh the gpus usage and find the first that is unused and use only this one.\nI guess someone could parse the output of  with bash and get a variable i and feed that variable i to the tensorflow script as the number of the gpu to use.<\/p>\n\n<p>I have never seen any example of this. I imagine it is a pretty common problem. What would be the simplest way to do that ? Is a pure tensorflow one available ?<\/p>\n","answers":[{"AnswerId":"41638727","CreationDate":"2017-01-13T16:02:43.197","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>I'm not aware of pure-TensorFlow solution. The problem is that existing place for TensorFlow configurations is a Session config. However, for GPU memory, a GPU memory pool is\u00a0shared for all TensorFlow sessions within a process, so Session config would be the wrong place to add it, and there's no mechanism for process-global config (but there should be, to also be able to configure process-global Eigen threadpool). So you need to do on on a process level by using <code>CUDA_VISIBLE_DEVICES<\/code> environment variable.<\/p>\n\n<p>Something like this:<\/p>\n\n<pre><code>import subprocess, re\n\n# Nvidia-smi GPU memory parsing.\n# Tested on nvidia-smi 370.23\n\ndef run_command(cmd):\n    \"\"\"Run command, return output as string.\"\"\"\n    output = subprocess.Popen(cmd, stdout=subprocess.PIPE, shell=True).communicate()[0]\n    return output.decode(\"ascii\")\n\ndef list_available_gpus():\n    \"\"\"Returns list of available GPU ids.\"\"\"\n    output = run_command(\"nvidia-smi -L\")\n    # lines of the form GPU 0: TITAN X\n    gpu_regex = re.compile(r\"GPU (?P&lt;gpu_id&gt;\\d+):\")\n    result = []\n    for line in output.strip().split(\"\\n\"):\n        m = gpu_regex.match(line)\n        assert m, \"Couldnt parse \"+line\n        result.append(int(m.group(\"gpu_id\")))\n    return result\n\ndef gpu_memory_map():\n    \"\"\"Returns map of GPU id to memory allocated on that GPU.\"\"\"\n\n    output = run_command(\"nvidia-smi\")\n    gpu_output = output[output.find(\"GPU Memory\"):]\n    # lines of the form\n    # |    0      8734    C   python                                       11705MiB |\n    memory_regex = re.compile(r\"[|]\\s+?(?P&lt;gpu_id&gt;\\d+)\\D+?(?P&lt;pid&gt;\\d+).+[ ](?P&lt;gpu_memory&gt;\\d+)MiB\")\n    rows = gpu_output.split(\"\\n\")\n    result = {gpu_id: 0 for gpu_id in list_available_gpus()}\n    for row in gpu_output.split(\"\\n\"):\n        m = memory_regex.search(row)\n        if not m:\n            continue\n        gpu_id = int(m.group(\"gpu_id\"))\n        gpu_memory = int(m.group(\"gpu_memory\"))\n        result[gpu_id] += gpu_memory\n    return result\n\ndef pick_gpu_lowest_memory():\n    \"\"\"Returns GPU with the least allocated memory\"\"\"\n\n    memory_gpu_map = [(memory, gpu_id) for (gpu_id, memory) in gpu_memory_map().items()]\n    best_memory, best_gpu = sorted(memory_gpu_map)[0]\n    return best_gpu\n<\/code><\/pre>\n\n<p>You can then put it in <code>utils.py<\/code> and set GPU in your TensorFlow script before first <code>tensorflow<\/code> import. IE<\/p>\n\n<pre><code>import utils\nimport os\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = str(utils.pick_gpu_lowest_memory())\nimport tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"50065069","CreationDate":"2018-04-27T14:55:09.587","ParentId":null,"OwnerUserId":"1500585","Title":null,"Body":"<p>An implementation along the lines of Yaroslav Bulatov's solution is available on <a href=\"https:\/\/github.com\/bamos\/setGPU\" rel=\"noreferrer\">https:\/\/github.com\/bamos\/setGPU<\/a>.<\/p>\n"}]}
{"QuestionId":41634915,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><conv-neural-network>","CreationDate":"2017-01-13T12:42:01.370","AcceptedAnswerId":null,"OwnerUserId":6430408.0,"Title":"Tensorflow CNN model get error \" NaN loss during training.\"","Body":"<p>I am a newer to tensorflow. \nWhen I training the data with cnn ,sometimes \"NaN loss during training\" is happened at the first batch.<\/p>\n\n<pre><\/pre>\n\n<p>\"bach_size\" ,\"learning_rate\" are adjust, even learning_rate = 0 is used.<\/p>\n\n<p>What I really confused is when all parameters fixed, the result is sometimes can run normally ,sometimes get the error \"Nan\" at first batch. I want to know how it happens? what factors lead to the result?<\/p>\n","answers":[{"AnswerId":"41648819","CreationDate":"2017-01-14T10:11:04.683","ParentId":null,"OwnerUserId":"2300116","Title":null,"Body":"<ul>\n<li>Check for NANs in your data<\/li>\n<li>RELU activation gives sometimes NANs<\/li>\n<\/ul>\n"}]}
{"QuestionId":41635738,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2017-01-13T13:28:40.973","AcceptedAnswerId":null,"OwnerUserId":7414822.0,"Title":"Get Pretrained Inception v3 model from Open Images dataset working on Android","Body":"<p>I tried a while to get the pretrained model working on android. The problem is, I only got the ckpt and meta file for the pretrained net. In my opinion I need the .pb for the android app. So I tried to convert the given files to an .pb file. <\/p>\n\n<p>Therefore I tried the freeze_graph.py but without succes. So I used the example code from <a href=\"https:\/\/github.com\/openimages\/dataset\/blob\/master\/tools\/classify.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/openimages\/dataset\/blob\/master\/tools\/classify.py<\/a> and modified it to store a pb. file after loading <\/p>\n\n<pre><\/pre>\n\n<p>Now my problem is that I have the .pb file but I don't have any opinion what is the input node name and I am not sure if  is the right output name. In the example android app I have to specify both. And the android app crashed with:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know if there are more problem by trying to fix the .pb problem. Or if anyone knows a better way to port the ckpt and meta files to a .pd file in my case or knows a source for the final file with input and ouput names please give me a hint to complete this task.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41640346","CreationDate":"2017-01-13T17:36:20.720","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>You'll need to use the optimize_for_inference.py script to strip out the unused nodes in your graph. \"decodeJpeg\" is not supported on Android -- pixel values should be fed in directly. ClassifierActivity.java has more detail about the specific nodes to use for inception v3.<\/p>\n"}]}
{"QuestionId":41636901,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-01-13T14:32:09.297","AcceptedAnswerId":42220126.0,"OwnerUserId":802362.0,"Title":"How to generate word suffix using Tensorflow?","Body":"<p>We are working with Tensorflow. We have a very large dataset of words, brands and so on. We are trying to generate suffixes for words. For instance, give 'real' may give 'realtor', 'really', 'realestate', ... In our dataset, each word is separated by \\n. <\/p>\n\n<p>We have tried several things so far, but it does not seem to produce adequate results. It seems to me that this is a fairly simple task that should be solved <\/p>\n\n<p>The basic idea we are working on is adapting <a href=\"https:\/\/github.com\/sherjilozair\/char-rnn-tensorflow\" rel=\"nofollow noreferrer\">char-rnn-tensorflow<\/a>. We have simply changed the sampling to sample (with weighted random probabilities) until a carriage return is sampled. We kept the original implementation, with rnn_decoder and sequence_loss_by_example and used rnn as model. <\/p>\n\n<p>So far, we haven't obtained good samples, almost random ones and sometimes suffixes but without any relation to the beginning of our word (real -> realstation). It is possible that it comes from the fact that this implementation do not separate words, so we would need something more simple ?<\/p>\n\n<p>What technologies would you use for such a task ?<\/p>\n","answers":[{"AnswerId":"42220126","CreationDate":"2017-02-14T07:23:58.287","ParentId":null,"OwnerUserId":"802362","Title":null,"Body":"<p>Finally, we have been able to obtain good results by carefully tuning the parameters of the network and applying temperature to the results to get some good suggestions. The sequence length was very important during training. <\/p>\n"}]}
{"QuestionId":41637162,"AnswerCount":1,"Tags":"<neural-network><caffe><nvidia><tensorrt>","CreationDate":"2017-01-13T14:45:34.867","AcceptedAnswerId":null,"OwnerUserId":7414948.0,"Title":"How cv::MAT convert to NHCW format?","Body":"<p>In User Guide.html, Input\/output of tensorRT is need to use NCHW format.<br>\nWhat's NCHW fomat ?<br>\nHow can I convert cv::MAT to NCHW format?<\/p>\n\n<p>I run inference using TensorRT like below code.<br>\nNothing error. But, It's not right result of output.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"45858916","CreationDate":"2017-08-24T10:13:05.900","ParentId":null,"OwnerUserId":"1630","Title":null,"Body":"<p>NCHW: For a 3 channel image, say BGR, pixels of the B channel are stored first, then the G channel and finally the R channel.<\/p>\n\n<p>NHWC: For each pixel, its 3 colors are stored together in BGR order.<\/p>\n\n<p>TensorRT requires your image data to be in NCHW order. But OpenCV reads it in NHWC order. You can write a simple function to read the data from NHWC to a buffer where you store them in NCHW order. Copy this buffer to device memory and pass to TensorRT.<\/p>\n\n<p>You can find an example of this operation in the <code>samples\/sampleFasterRCNN\/sampleFasterRCNN.cpp<\/code> file in your TensorRT installation. It reads a PPM file, which is also in NHWC order and then converts it to NCHW order and subtracts the mean values, both in a single step. You can modify that to suit your purpose.<\/p>\n"}]}
{"QuestionId":41637800,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-13T15:17:11.700","AcceptedAnswerId":41642230.0,"OwnerUserId":7012118.0,"Title":"How to get a list contains each element in a tensor?","Body":"<p>I am learning  and working on a demo with constraint that each weight be nonnegative.<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/contrib.opt\/other_functions_and_classes#ScipyOptimizerInterface\" rel=\"nofollow noreferrer\">ScipyOptimizerInterface<\/a><\/p>\n\n<blockquote>\n  <p>: Optional list of equality constraint scalar Tensors to\n  be held equal to zero. : Optional list of inequality\n  constraint scalar Tensors to be kept nonnegative.<\/p>\n<\/blockquote>\n\n<p>How can I change  to something like ?<\/p>\n","answers":[{"AnswerId":"41642230","CreationDate":"2017-01-13T19:37:58.000","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>You can set it with:<\/p>\n\n<pre><code>inequalities = [wt[i] for i in range(wt.get_shape()[0])]\n<\/code><\/pre>\n"}]}
{"QuestionId":41637974,"AnswerCount":1,"Tags":"<random><tensorflow><convolution>","CreationDate":"2017-01-13T15:26:03.403","AcceptedAnswerId":41638332.0,"OwnerUserId":7415539.0,"Title":"Running session multiple times with tf.random returns different values for conv2d","Body":"<pre><\/pre>\n\n<p>If I run the code above with option1, I get the same values for  and \nHowever, when I run the above with option2, I get different values for  and .<\/p>\n\n<p>I am guessing the tf.random_normal gets initialized every time the session is ran.\nIs this meant to happen?\nThis happens even if I do the \nCan someone explain how TensorFlow initializes its random variables when the session is ran?\n&nbsp;<\/p>\n","answers":[{"AnswerId":"41638332","CreationDate":"2017-01-13T15:45:08.340","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>All of the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/constant_op\/random_tensors\" rel=\"nofollow noreferrer\">random number ops<\/a> in TensorFlow (including <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/constant_op\/random_tensors#random_normal\" rel=\"nofollow noreferrer\"><code>tf.random_normal()<\/code><\/a>) sample a new random tensor each time they runs:<\/p>\n\n<blockquote>\n  <p>TensorFlow has several ops that create random tensors with different distributions. The random ops are stateful, and create new random values each time they are evaluated.<\/p>\n<\/blockquote>\n\n<p>If you want to sample the distribution once and then re-use the result, you should use a <code>tf.Variable<\/code> and initialize it by running <code>tf.random_normal()<\/code> once. For example, the following code will print the same random value twice:<\/p>\n\n<pre><code>f_tf = tf.Variable(tf.random_normal([1, 3, 1, 1]))\n# ...\ninit_op = tf.global_variables_initializer()\n# ...\nwith tf.Session() as sess:\n  sess.run(init_op)\n  print(sess.run(f_tf))\n  print(sess.run(f_tf))\n<\/code><\/pre>\n"}]}
{"QuestionId":41638234,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-01-13T15:40:13.923","AcceptedAnswerId":41638404.0,"OwnerUserId":4313064.0,"Title":"why does keras.applications inception not recognise my elephant?","Body":"<p>I have just started trying the keras applications. The example for resnet50 works fine. However the inception_v3 model thinks my elephant is a sports car. Where am I going wrong?<\/p>\n\n<pre><\/pre>\n\n<p>Predicted: [('n04285008', 'sports_car', 0.58716565), ('n04041544', 'radio', 0.19915217), ('n03535780', 'horizontal_bar', 0.11363289), ('n03691459', 'loudspeaker', 0.059153806), ('n01669191', 'box_turtle', 0.012588425), ('n04286575', 'spotlight', 0.0068869498), ('n03594945', 'jeep', 0.005987139), ('n01728920', 'ringneck_snake', 0.0050685192), ('n03637318', 'lampshade', 0.0020551293), ('n03000134', 'chainlink_fence', 0.00168875)]<\/p>\n","answers":[{"AnswerId":"41638404","CreationDate":"2017-01-13T15:48:32.797","ParentId":null,"OwnerUserId":"4313064","Title":null,"Body":"<p>I am an idiot. Cut and pasted the imports including the preprocessing and decode predictions from the resnet50 example. Need to import these from inception_v3.<\/p>\n"}]}
{"QuestionId":41638498,"AnswerCount":1,"Tags":"<tensorflow><windows-10><gpu>","CreationDate":"2017-01-13T15:52:44.530","AcceptedAnswerId":41638676.0,"OwnerUserId":7384452.0,"Title":"Tensorflow on Windows-10","Body":"<p>Is GPU version of Tensorflow 0.12 running on Windows 10 stable? If not which is the most stable implementation for version 0.12? <\/p>\n\n<p>Also when is Tensorflow 1.0 is due to be released.<\/p>\n","answers":[{"AnswerId":"41638676","CreationDate":"2017-01-13T16:00:07.667","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow 0.12.1 is currently (Jan. 2017) the latest \"stable\" release of TensorFlow for Windows. As this is the first version of TensorFlow released for Windows, there have been some bugs, and the <code>tf.contrib<\/code> libraries are not fully supported, but the core parts of the TensorFlow library are tested regularly and stable.<\/p>\n\n<p>TensorFlow 1.0.0-alpha was <a href=\"https:\/\/www.google.com\/url?hl=en&amp;q=https:\/\/groups.google.com\/a\/tensorflow.org\/d\/msgid\/discuss\/CAKrkxF_fHRH1etTf%252BoNhHtETe%252BXKnD3kUrMS_Tqf22h4r%252BsK_g%2540mail.gmail.com?utm_medium%3Demail%26utm_source%3Dfooter&amp;source=gmail&amp;ust=1484409413162000&amp;usg=AFQjCNEulIJie-_0UcB-W_47gqvN0SYuMw\" rel=\"nofollow noreferrer\">announced<\/a> earlier this week, but it is not yet available from PyPI, and official binary packages are not yet available. An announcement will be made soon about the release date for TensorFlow 1.0.0, and it will include a Windows package.<\/p>\n"}]}
{"QuestionId":41638887,"AnswerCount":0,"Tags":"<numpy><tensorflow>","CreationDate":"2017-01-13T16:10:50.933","AcceptedAnswerId":null,"OwnerUserId":482601.0,"Title":"Fastest way to calculate distance","Body":"<p>I have<\/p>\n\n<ul>\n<li>a  dimensional matrix <\/li>\n<li>a  dimensional matrix <\/li>\n<\/ul>\n\n<p>Let's denote the b-th row of  with  and the i-th column of  with . I want to calculate the  dimensional matrix  which contains at :<\/p>\n\n<pre><\/pre>\n\n<p>where the bars are the L2-norm.<\/p>\n\n<p>My current solution uses the trick that<\/p>\n\n<pre><\/pre>\n\n<p>which I calculate like so:<\/p>\n\n<pre><\/pre>\n\n<p>It works and is reasonably fast. But I was wondering, is there a trick to make this calculation even faster?<\/p>\n","answers":[]}
{"QuestionId":41639000,"AnswerCount":0,"Tags":"<android><tensorflow>","CreationDate":"2017-01-13T16:17:17.787","AcceptedAnswerId":null,"OwnerUserId":2826310.0,"Title":"tensorflow export model and run it on Android failed","Body":"<p>I train an model using tensorflow on gpu server(K40), and then I export the model using session bundle, and I could run the exported model on PC computer. But when I use it on android, an error occurs:<\/p>\n\n<p>tensorflow_jni.cc:184 Check failed: ::tensorflow::Status::OK() == (tensorflow::serving::GetDefaultSignature(bundle.meta_graph_def, &amp;signature)) (OK vs. Failed precondition: Expected signature Any type_url for: tensorflow.serving.Signatures. Got: type.googleapis.com\/tensorflow.contrib.Signatures.)<\/p>\n\n<p>We are frustrated about this error, I have no idea what is wrong during the process?<\/p>\n","answers":[]}
{"QuestionId":41640037,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><keras>","CreationDate":"2017-01-13T17:17:29.893","AcceptedAnswerId":43446675.0,"OwnerUserId":3827807.0,"Title":"Deconvolution under Keras issues","Body":"<p>I'm trying to use the Deconvolution2D of keras with Tensorflow backend.<\/p>\n\n<p>But I got some issues.\nFirst, in the output_shape, if I pass None for the batch_size, I get this error :<\/p>\n\n<pre><\/pre>\n\n<p>And if I change None by the batch size I use, here is the error.. :<\/p>\n\n<pre><\/pre>\n\n<p>Here is the model I use :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43446675","CreationDate":"2017-04-17T06:54:00.827","ParentId":null,"OwnerUserId":"6271239","Title":null,"Body":"<p>This was an annoyance with deconvolution in previous versions of Keras, always having to give a fixed batch size and manually compute output_shape. This also meant that your data set size had to be divisible by 'batch_size' or an error would be raised on the last (smaller) batch.<\/p>\n\n<p>Fortunately this was fixed in Keras 2.0. Deconvolution2D has been replaced by Conv2DTranspose and you don't even have to give output_shape as an argument any more:<\/p>\n\n<pre><code>    model.add(Conv2DTranspose(filters=256, kernel_size=(h,h), strides=(2,2), padding='same'))\n<\/code><\/pre>\n"}]}
{"QuestionId":41640112,"AnswerCount":1,"Tags":"<python><queue><tensorflow>","CreationDate":"2017-01-13T17:22:04.480","AcceptedAnswerId":41640835.0,"OwnerUserId":7399443.0,"Title":"TensorFlow: Enqueue tuple of tensors throws TypeError","Body":"<p>I have a time based dataset which consists of categorical features, real values features, a mask that states whether the given feature was present at the time, and a \"Deltas\" array that contains the length of time since a value has been present.<\/p>\n\n<p>I want to build a queue of tuples of these tensors so that the categorical features can be converted to one hot and so that the data, mask, and deltas can be used in different parts of the model. Below is some code that I wrote to do this:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>When I run this I get a TypeError:<\/p>\n\n<p>TypeError: Expected uint16, got array(...) of type 'ndarray' instead.<\/p>\n\n<p>I am not sure what I am doing wrong, is it the dtype definition when I create my queue?<\/p>\n","answers":[{"AnswerId":"41640835","CreationDate":"2017-01-13T18:05:49.560","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are a few problems here:<\/p>\n\n<ol>\n<li><p>Your thread is calling <code>q.enqueue_many()<\/code> repeatedly. Despite its (slightly confusing) name, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/queues#QueueBase.enqueue_many\" rel=\"nofollow noreferrer\"><code>q.enqueue_many()<\/code><\/a> method does not immediately enqueue data in the queue, but rather it returns a <code>tf.Operation<\/code> that must be passed to <code>sess.run()<\/code> to add the tensors in the queue. The code that runs in the separate thread is creating 10 enqueue-many operations and discarding them, which is probably not what you intended.<\/p><\/li>\n<li><p>The return value of <code>nextBatch(100)<\/code> is a list of 100 tuples of 4 arrays. The <code>q.enqueue_many()<\/code> method expects a tuple of 4 arrays. If you want to enqueue a list of 100 tuples, you'll need to run a <code>q.enqueue()<\/code> op 100 times, <strong>or<\/strong> stack together the 100 arrays for each tuple component so that you have a single tuple of four arrays.<\/p><\/li>\n<li><p>The arrays produce in <code>nextBatch()<\/code> don't match the shapes of the queue components. Assuming <code>n_steps<\/code> is the dimension that could be variable (for the purposes of padding), the function should produce arrays of <code>(n_steps, 5)<\/code>, <code>(n_steps, 48)<\/code>, <code>(n_steps, 53)<\/code>, and <code>(n_steps, 53)<\/code> to match the queue definition.<\/p><\/li>\n<\/ol>\n\n<p>Here's a version of your code that works as I assume you intended:  <\/p>\n\n<pre><code>import tensorflow as tf\nimport threading\nimport numpy as np\n\n# Function to generate batches of data                                                                                                                         \ndef nextBatch(batch_size):\n  n_steps = 14\n  batch = []\n  for _ in range(batch_size):\n    # Create tuple of tensors                                                                                                                                  \n    ex = (np.random.randint(0,5, (n_steps, 5)),\n          np.random.randn(n_steps, 48),\n          np.random.randint(0,2, (n_steps, 53)),\n          np.random.randint(0,2000, (n_steps, 53)))\n    batch.append(ex)\n  return batch\n\nq = tf.PaddingFIFOQueue(1000,\n                        [tf.uint16, tf.float32, tf.uint16, tf.uint16],\n                        [(None, 5), (None, 48), (None, 53), (None, 53)])\n\n# Define a single op for enqueuing a tuple of placeholder tensors.\nplaceholders = [tf.placeholder(tf.uint16, shape=(None, 5)),\n                tf.placeholder(tf.float32, shape=(None, 48)),\n                tf.placeholder(tf.uint16, shape=(None, 53)),\n                tf.placeholder(tf.uint16, shape=(None, 53))]\nenqueue_op = q.enqueue(placeholders)\n\n# Create a session in order to run the enqueue_op.\nsess = tf.Session()\n\ndef enqueue_thread_fn():\n  for i in range(10):\n    batch = nextBatch(100)\n    for batch_elem in batch:\n      # Each call to `sess.run(enqueue_op, ...)` enqueues a single element in\n      # the queue.\n      sess.run(enqueue_op, feed_dict={placeholders[0]: batch_elem[0],\n                                      placeholders[1]: batch_elem[1],\n                                      placeholders[2]: batch_elem[2],\n                                      placeholders[3]: batch_elem[3]})\n\n# Start enqueuing                                                                                                                                              \nt = threading.Thread(target=enqueue_thread_fn)\nt.start()\nt.join()\nsess.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":41640860,"AnswerCount":1,"Tags":"<image-processing><binary><tensorflow>","CreationDate":"2017-01-13T18:06:59.710","AcceptedAnswerId":41641624.0,"OwnerUserId":4162005.0,"Title":"read and write tfrecords binary file (type missmatch)","Body":"<p>Hi I'm trying to build an image input pipe. My preprocessed training data is stored in a tfrecords file which I've create with the following lines of codes:<\/p>\n\n<pre><\/pre>\n\n<p>..<\/p>\n\n<pre><\/pre>\n\n<p>Now I try to read the binary data to reconstruct a tensor out of it:<\/p>\n\n<pre><\/pre>\n\n<p>Sadly this is not working. I get this error messag:<\/p>\n\n<blockquote>\n  <p>ValueError: Tensor conversion requested dtype string for Tensor with\n  dtype int64: 'Tensor(\"ParseSingleExample\/Squeeze_label:0\", shape=(),\n  dtype=int64)' \n  ...<\/p>\n  \n  <p>TypeError: Input 'bytes' of 'DecodeRaw' Op has type int64 that does not match expected type of string.<\/p>\n<\/blockquote>\n\n<p>Can some give me a hint on how to fix this?<\/p>\n\n<p>Thanks in advance!<\/p>\n\n<p><strong>UPDATE<\/strong>: complete code listing of \"read_and_decode\"<\/p>\n\n<p>@mmry thank you very much. now my code breaks at shuffeling the batch. With:<\/p>\n\n<blockquote>\n  <p>ValueError: All shapes must be fully defined:\n  [TensorShape([Dimension(None), Dimension(None), Dimension(None)]),\n  TensorShape([])]<\/p>\n<\/blockquote>\n\n<p>any suggestions?<\/p>\n","answers":[{"AnswerId":"41641624","CreationDate":"2017-01-13T18:57:35.440","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There is no need to use the <code>tf.decode_raw()<\/code> op in this line:<\/p>\n\n<pre><code>label = tf.decode_raw(features['label'], tf.int64)\n<\/code><\/pre>\n\n<p>Instead, you should be able to write:<\/p>\n\n<pre><code>label = features['label']\n<\/code><\/pre>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/converting#decode_raw\" rel=\"nofollow noreferrer\"><code>tf.decode_raw()<\/code><\/a> op only accepts <code>tf.string<\/code> tensors, and converts the binary representation of some tensor data (as a variable-length string) into a typed representation (as a vector of a particular type of elements). However, you have defined the feature <code>'label'<\/code> as having type <code>tf.int64<\/code>, so there is no need to convert that feature if you want to use it as a <code>tf.int64<\/code>.<\/p>\n"}]}
{"QuestionId":41641598,"AnswerCount":1,"Tags":"<tensorflow><dimensionality-reduction>","CreationDate":"2017-01-13T18:55:51.180","AcceptedAnswerId":43993913.0,"OwnerUserId":7415539.0,"Title":"Tensorflow reduce dimensions of rank 3 tensor","Body":"<p>I am trying to build a CLDNN that is researched in the paper <a href=\"https:\/\/36e9b848-a-62cb3a1a-s-sites.googlegroups.com\/site\/tsainath\/tsainath_cldnn_icassp2015.pdf?attachauth=ANoY7coIUra14fov8me7FImAp0XczhJmdMNt0Sr_E_WxwJYNpuONb2Mm0ODdEdwBq7-gU083lNqY2cNAZOFWD4wl2nNckR81Vt0VUrU_zzkkiy5CR97Px81U-UWjztvDstpu95JtqTO4Y4F9NAAklDpKiSpx1CC8TeTtnJLOEVr7VkcBSzTWBmWvtwPtI52LhES0Rb9L2I8mYTA8_dSHtg3Dyfg2skkEdKKewsSLTUKKMas5Lm03gjk%3D&amp;attredirects=0\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>After the convolutional layers, the features go through a dim-reduction layer. At the point when the features leave the conv layers, the dimensions are . N represents the number of windows and I think the network requires the reduction in the dimension M, so the dimensions of the features after the dim-red layer is  , where . <\/p>\n\n<p>I have two questions.<\/p>\n\n<ol>\n<li><p>How do I do this in TensorFlow? I tried using a weight with <\/p>\n\n<pre><\/pre>\n\n<p>I thought the multiplication of  would yield  but  and  are not valid dimensions for multiplication. I would like to keep N constant and reduce the dimension of M. <\/p><\/li>\n<li><p>What kind of non-linearity should I apply to the outcome of ? I was thinking about using a ReLU but I couldn't even get #1 done.<\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"43993913","CreationDate":"2017-05-16T06:26:18.763","ParentId":null,"OwnerUserId":"4221706","Title":null,"Body":"<p>According to the linked paper (T. N. Sainath <em>et al.<\/em>: \"Convolutional, Long Short-Term Memory, Fully Connected Deep Neural Networks\"), <\/p>\n\n<blockquote>\n  <p>[...] reducing the dimensionality, such that we have 256 outputs from the linear layer, was appropriate.<\/p>\n<\/blockquote>\n\n<p>That means, whatever the input size is, i.e. <code>[?, N, M]<\/code> or any other dimensionality (always assuming that the first dimension is the number of samples in a mini-batch, denoted by <code>?<\/code>), the output will be <code>[?, Q]<\/code>, where typically <code>Q=256<\/code>.<\/p>\n\n<p>As we are doing dimensionality reduction by multiplying the input with a weight matrix, no spatial information will be preserved. This means, that it doesn't matter whether each input is a matrix or a vector, so we can reshape the input to the linear layer <code>x<\/code> to have the dimensions <code>[?, N*M]<\/code>. Then, we can create a simple matrix multiplication <code>tf.matmul(x, W)<\/code> where <code>W<\/code> is a matrix with dimensions <code>[N*M, Q].<\/code><\/p>\n\n<pre><code>W = tf.Variable(tf.truncated_normal([N*M, Q], stddev=0.1))\nx_vec = tf.reshape(x, shape=(-1, N*M))\ny = tf.matmul(x_vec, W)\n<\/code><\/pre>\n\n<p>Finally, regarding question 2: in the paper, the dimensionality reduction layer is a linear layer, i.e. you do not apply a non-linearity to the output. <\/p>\n"}]}
{"QuestionId":41641684,"AnswerCount":2,"Tags":"<python><sequence><keras><recurrent-neural-network>","CreationDate":"2017-01-13T19:01:39.803","AcceptedAnswerId":41648579.0,"OwnerUserId":996366.0,"Title":"char-rnn (multi-layer Recurrent Neural Network) implementation getting IndexError","Body":"<p>I am trying to implement this <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/197#issuecomment-111805144\" rel=\"nofollow noreferrer\">char-rnn.py<\/a> with slight changes in my system. <\/p>\n\n<p>This is my full code:<\/p>\n\n<pre><\/pre>\n\n<p>But I am getting this error:<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"41648579","CreationDate":"2017-01-14T09:40:34.740","ParentId":null,"OwnerUserId":"5114796","Title":null,"Body":"<p>It looks like it is trying to run over a sequence that is longer than the data.<\/p>\n\n<p>Looking at your code, this is the area throwing the error:<\/p>\n\n<pre><code>for i in numpy.arange(n_timestamps):\nprobs = model.predict_proba(sentences)[:,i,:]\n# Go over each sequence and sample the i-th character.\nfor j in numpy.arange(len(sentences)):\n    sentences[j, i+1, mnrnd(probs[j, :])] = 1\n<\/code><\/pre>\n\n<p>The problem is probably that your data is n_timestamps long, but you are trying to predict the n_timestamps + 1 character (when you are predicting i +1).<\/p>\n\n<p>try reducing the length of your loop by one as follows:<\/p>\n\n<pre><code>for i in numpy.arange(n_timestamps - 1):\nprobs = model.predict_proba(sentences)[:,i,:]\n# Go over each sequence and sample the i-th character.\nfor j in numpy.arange(len(sentences)):\n    sentences[j, i+1, mnrnd(probs[j, :])] = 1\n<\/code><\/pre>\n"},{"AnswerId":"41648233","CreationDate":"2017-01-14T08:46:16.127","ParentId":null,"OwnerUserId":"2314651","Title":null,"Body":"<p>There might be a problem with the way you manipulate X, I tried to run your code but encountered different errors than the one you mentioned.<\/p>\n\n<p>I'm adding a different example on an LSTM which does what you wanted and is based on the example <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/lstm_text_generation.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><code>from __future__ import print_function\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Activation\nfrom keras.layers import LSTM\nfrom keras.optimizers import RMSprop\nimport numpy as np\nimport random\nimport sys\n\ntext = \"123456789\"*1000\nprint('corpus length:', len(text))\n\nchars = sorted(list(set(text)))\nprint('total chars:', len(chars))\nchar_indices = dict((c, i) for i, c in enumerate(chars))\nindices_char = dict((i, c) for i, c in enumerate(chars))\n\n# cut the text in semi-redundant sequences of maxlen characters\nmaxlen = 200\nsentences = []\nnext_chars = []\nfor i in range(0, len(text) - maxlen):\n    sentences.append(text[i: i + maxlen])\n    next_chars.append(text[i + maxlen])\nprint('nb sequences:', len(sentences))\n\nprint('Vectorization...')\nX = np.zeros((len(sentences), maxlen, len(chars)), dtype=np.bool)\ny = np.zeros((len(sentences), len(chars)), dtype=np.bool)\nfor i, sentence in enumerate(sentences):\n    for t, char in enumerate(sentence):\n        X[i, t, char_indices[char]] = 1\n    y[i, char_indices[next_chars[i]]] = 1\n\n\n# build the model: a single LSTM\nprint('Build model...')\nmodel = Sequential()\nmodel.add(LSTM(128, input_shape=(maxlen, len(chars))))\nmodel.add(Dense(len(chars)))\nmodel.add(Activation('softmax'))\n\noptimizer = RMSprop(lr=0.01)\nmodel.compile(loss='categorical_crossentropy', optimizer=optimizer)\n\n\ndef sample(preds, temperature=1.0):\n    # helper function to sample an index from a probability array\n    preds = np.asarray(preds).astype('float64')\n    preds = np.log(preds) \/ temperature\n    exp_preds = np.exp(preds)\n    preds = exp_preds \/ np.sum(exp_preds)\n    probas = np.random.multinomial(1, preds, 1)\n    return np.argmax(probas)\n\n# train the model, output generated text after each iteration\nfor iteration in range(1, 60):\n    print()\n    print('-' * 50)\n    print('Iteration', iteration)\n    model.fit(X, y, batch_size=128, nb_epoch=1)\n\n    start_index = random.randint(0, len(text) - maxlen - 1)\n\n    for diversity in [0.2, 0.5, 1.0, 1.2]:\n        print()\n        print('----- diversity:', diversity)\n\n        generated = ''\n        sentence = text[start_index: start_index + maxlen]\n        generated += sentence\n        print('----- Generating with seed: \"' + sentence + '\"')\n        sys.stdout.write(generated)\n\n        for i in range(400):\n            x = np.zeros((1, maxlen, len(chars)))\n            for t, char in enumerate(sentence):\n                x[0, t, char_indices[char]] = 1.\n\n            preds = model.predict(x, verbose=0)[0]\n            next_index = sample(preds, diversity)\n            next_char = indices_char[next_index]\n\n            generated += next_char\n            sentence = sentence[1:] + next_char\n\n            sys.stdout.write(next_char)\n            sys.stdout.flush()\n        print()\n<\/code><\/pre>\n"}]}
{"QuestionId":41643044,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-13T20:36:50.600","AcceptedAnswerId":41656087.0,"OwnerUserId":1259390.0,"Title":"What is the difference between tf.train.MonitoredTrainingSession and tf.train.Supervisor","Body":"<p>I'm wondering what is the difference between these 2 tensorflow object when used to train a neural networks ?<\/p>\n","answers":[{"AnswerId":"41656087","CreationDate":"2017-01-14T23:23:02.220","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Supervisor is on the way to getting deprecated and new users are encouraged to use to tf.train.FooSession classes\n(from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6604#issuecomment-270950324\" rel=\"noreferrer\">comment<\/a>)<\/p>\n"}]}
{"QuestionId":41643295,"AnswerCount":0,"Tags":"<time-series><theano>","CreationDate":"2017-01-13T16:42:42.233","AcceptedAnswerId":null,"OwnerUserId":3716368.0,"Title":"Masking Layer using Keras Functional API","Body":"<p>I'm trying to use Keras' Functional API as I need to add additional input into my recurrent neural net between two layers.  This would be fine however I also have input I want to forecast and so I'm using a masking layer to keep the out of sample data from affecting the model.  The masking layer throws an error when I try to pass it an Input(...) as it's expecting a list of integers not tensor variables.  Is there a specific masking layer for the Functional API?  Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the <a href=\"https:\/\/drive.google.com\/open?id=0B2Kz5NTvWjJuSzcxc0dmX3lVMEE\" rel=\"nofollow noreferrer\">data<\/a>.  Use columns 0 as the test data and columns 1 and 3 as training data in the code.  <\/p>\n","answers":[]}
{"QuestionId":41643365,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-01-13T21:01:48.883","AcceptedAnswerId":null,"OwnerUserId":5703487.0,"Title":"Standalone Tensorflow Projector","Body":"<p>If you visit the <a href=\"http:\/\/projector.tensorflow.org\/\" rel=\"nofollow noreferrer\">http:\/\/projector.tensorflow.org\/<\/a> you can use it with your own dataset (ie a TSV file). I am playing with N-D data and found useful to look at these visualisations after PCA reduction. <\/p>\n\n<p>I am wondering how I can run my own Projector version on my machine.\nLooking at the doc, it seems to be released only as a tensorboard plugin for seeing the embedding results...<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"50397491","CreationDate":"2018-05-17T17:40:40.593","ParentId":null,"OwnerUserId":"2482434","Title":null,"Body":"<p>According to the replies made on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7562\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7562<\/a> more recently than the other answer here, you can get the standalone version at <a href=\"https:\/\/github.com\/tensorflow\/embedding-projector-standalone\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/embedding-projector-standalone\/<\/a> and edit the oss_demo_projector_config.json to point to your datasets. <\/p>\n\n<p>The demo files are binary files ending in .bytes, which can be generated from a numpy array with .tofile:<\/p>\n\n<pre><code>vectors = numpy.zeros(vector_shape, dtype=numpy.float32)\nvectors.tofile('my_tensors.bytes')\n<\/code><\/pre>\n"},{"AnswerId":"42377448","CreationDate":"2017-02-21T20:36:02.417","ParentId":null,"OwnerUserId":"1170391","Title":null,"Body":"<p>It has only been released as a TensorBoard plugin.<\/p>\n"}]}
{"QuestionId":41643637,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-13T21:21:46.543","AcceptedAnswerId":null,"OwnerUserId":2008194.0,"Title":"How can I keep state from train step to train step?","Body":"<p>I have a tensor in my computation graph that I'd like to add a row to after every train step. How can I accomplish this?<\/p>\n\n<p>More detail: I'm grabbing gradients from , and I'd like to modify those gradients based on the gradient history. Here is the code that I'm trying to use:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>I've also tried keeping a variable around that I use to keep track of rows by concatenating new rows onto, but that didn't work.<\/p>\n","answers":[{"AnswerId":"41645904","CreationDate":"2017-01-14T01:35:30.617","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's an example of using persistent Tensors to store gradient history. In the loop below <code>gradient_history<\/code> refers to concatenation of all the gradients so far:<\/p>\n\n<pre><code>n = 2\nx = tf.Variable(tf.zeros((n,)))\nx_target = 10*tf.ones((n,))\nloss = tf.reduce_sum(tf.square(x - x_target))\noptimizer = tf.train.GradientDescentOptimizer(0.1)\ngradient = tf.gradients(loss, [x])[0]\ntrain_op = optimizer.apply_gradients([[gradient, x]])\n\n# initialize history with first gradient\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\ngradient_history0 = sess.run(tf.get_session_handle(tf.stack([gradient])))\nprevious_gradients_in, previous_gradients = tf.get_session_tensor(gradient_history0, dtype=dtype)\ngradient_history = tf.concat(0, [previous_gradients, [gradient]])\ngradient_history_out = tf.get_session_handle(gradient_history)\n\nsess.run(tf.global_variables_initializer())\n\nfor i in range(10):\n    [gradient_history0, _, loss0, gradient0] = sess.run([gradient_history_out, train_op, loss, gradient],\n                                                   feed_dict={previous_gradients_in: gradient_history0.handle})\n    print(loss0, gradient0)\n<\/code><\/pre>\n\n<p>When you run it, you should see something like this:<\/p>\n\n<pre><code>200.0 [-20. -20.]\n128.0 [-16. -16.]\n81.92 [-12.80000019 -12.80000019]\n52.4288 [-10.23999977 -10.23999977]\n33.5544 [-8.19199944 -8.19199944]\n21.4748 [-6.55359936 -6.55359936]\n13.7439 [-5.24287987 -5.24287987]\n8.79609 [-4.19430351 -4.19430351]\n5.6295 [-3.35544205 -3.35544205]\n3.60288 [-2.68435287 -2.68435287]\n<\/code><\/pre>\n\n<p>Note that at each fed step of your computation, <code>gradient_history<\/code> is a Tensor object which refers to gradient history. Meanwhile <code>gradient_history0<\/code> is a TensorHandle object which refers to the saved history which persists between <code>session.run<\/code> calls. You can feed that value back into the graph using <code>feed_dict={...: gradient_history0.handle}<\/code>, but unlike feeding numpy arrays, you are feeding in a \"pointer\" to the data, and the data itself never leaves TensorFlow runtime. Since the handle persists between <code>session.run<\/code> calls, you can also evaluate it directly:<\/p>\n\n<pre><code>In [10]: gradient_history0.eval()\nOut[10]:\narray([[-20.        , -20.        ],\n       [-20.        , -20.        ],\n       [-16.        , -16.        ],\n       [-12.80000019, -12.80000019],\n       [-10.23999977, -10.23999977],\n       [ -8.19199944,  -8.19199944],\n       [ -6.55359936,  -6.55359936],\n       [ -5.24287987,  -5.24287987],\n       [ -4.19430351,  -4.19430351],\n       [ -3.35544205,  -3.35544205],\n       [ -2.68435287,  -2.68435287]], dtype=float32)\n<\/code><\/pre>\n"},{"AnswerId":"41685142","CreationDate":"2017-01-16T21:18:59.460","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>I ended up using <code>tf.py_func<\/code> to accomplish this. I keep track of state in a global list that is accessed in the Python function. Here the gradients are applied:<\/p>\n\n<pre><code># process each individual gradient before applying it\ntrain_step = optimizer.apply_gradients([(process_gradient(grad, str(i)), var) for i, (grad, var) in enumerate(grads_and_vars)])\n<\/code><\/pre>\n\n<p>Here is where I keep track of state over time, and would use the built up state:<\/p>\n\n<pre><code>def construct_processor(name):\n  global_gradients_over_time = {}\n\n  def python_process_gradient(gradient):\n    reshaped_gradient = gradient.flatten()\n\n    if name in global_gradients_over_time:\n      global_gradients_over_time[name].append(reshaped_gradient)\n    else:\n      global_gradients_over_time[name] = [reshaped_gradient]\n\n    # process gradients somehow\n\n    return gradient\n\n  return python_process_gradient\n\ndef process_gradient(gradient, name):\n  return tf.py_func(construct_processor(name), [gradient], tf.float32)\n<\/code><\/pre>\n\n<p><code>construct_processor<\/code> is just there to allow you to process gradients one at a time, giving each set of gradients a name so I can find them in the global dictionary. This approach also keeps memory off the GPU, I think.<\/p>\n"}]}
{"QuestionId":41644164,"AnswerCount":0,"Tags":"<docker><tensorflow>","CreationDate":"2017-01-13T22:05:37.570","AcceptedAnswerId":null,"OwnerUserId":4177177.0,"Title":"distributed tensorflow using docker","Body":"<p>I am experimenting with distributed tensorflow and an <a href=\"https:\/\/ischlag.github.io\/2016\/06\/12\/async-distributed-tensorflow\/\" rel=\"nofollow noreferrer\">example<\/a> project.\nRunning the project on the same docker container seems to work well. As soon as you run the application on different conatiners, they cannot connect to eachother. <\/p>\n\n<p>I don't really know the problem, but I think this is because docker and tensorflow open ports which have to be concatenated to connect to the application like localhost:[docker-port]:[tf-port] <\/p>\n\n<p>Do you think my guess is correct? And how can I solve this problem?   <\/p>\n","answers":[]}
{"QuestionId":41645571,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-14T00:43:15.523","AcceptedAnswerId":null,"OwnerUserId":5827626.0,"Title":"changing number of outputs\/classes in tensorflow","Body":"<p>Here is my simple logistic regression with two features and two classes:<\/p>\n\n<pre><\/pre>\n\n<p>Everything works as expected at this point.<\/p>\n\n<p>Now, my goal is to modify number of outputs\/classes in-between training steps.\nIt seems that there is a way to modify the shape of weights:<\/p>\n\n<pre><\/pre>\n\n<p>This also works. I can run  to add two more classes and then continue running .<\/p>\n\n<p>The problem starts if I want to add bias term:<\/p>\n\n<pre><\/pre>\n\n<p>If I run  before adding more classes - it works.\nIf I run change ops -  they also work fine and checking on w and b individually shows that they increased in size. Even more, running  also works and returns predictions for more classes.<\/p>\n\n<p><strong>The only thing that does not work is !<\/strong><\/p>\n\n<p>It gives this error (3 in the shape correspond to the number of examples in passed x and y values):<\/p>\n\n<pre><\/pre>\n\n<p>From the traceback it seems to be coming from  addition op, but somehow related to gradient calculation. As if the b term is still of shape (1, 2) instead of (1, 4). Although evaluating  separately does return shape (1, 4) array.<\/p>\n\n<p>What am I doing wrong here? Why changing  does not cause problems but changing  does. Is this a bug?<\/p>\n","answers":[{"AnswerId":"41731185","CreationDate":"2017-01-18T23:17:29.253","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You should not change the number of classes while training. If that happens, you have to create a new model for the given classes and re-train from it.<\/p>\n"}]}
{"QuestionId":41645990,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2017-01-14T01:50:20.103","AcceptedAnswerId":41710515.0,"OwnerUserId":5885421.0,"Title":"Persistent Variable in keras Custom Layer","Body":"<p>I want write a custom layer, where I can keep a variable in memory between runs. \nFor example,<\/p>\n\n<pre><\/pre>\n\n<p>When I run , I expect the  to get updated, and print the incremented value.\nBut it looks like it always prints <\/p>\n\n<pre><\/pre>\n\n<p>My question is, how do I make the  increment and save after every run of <\/p>\n\n<p>Thanks,\nNaveen<\/p>\n","answers":[{"AnswerId":"41710515","CreationDate":"2017-01-18T02:40:43.360","ParentId":null,"OwnerUserId":"2924421","Title":null,"Body":"<p>The trick is that you have to call <code>self.add_update(...)<\/code> in your call function to register a function that will be called every time your model is evaluated (I found this by digging into the source code of the stateful rnns). If you do <code>self.stateful = True<\/code> it will call your custom update function for every training and prediction call, otherwise it will only call it during training. For example:<\/p>\n\n<pre><code>import keras.backend as K\nimport numpy as np\nfrom keras.engine.topology import Layer\n\nclass CounterLayer(Layer):\n  def __init__(self, stateful=False,**kwargs):\n    self.stateful = stateful # True means it will increment counter on predict and train, false means it will only increment counter on train \n    super(CounterLayer, self).__init__(**kwargs)\n\n\n  def build(self, input_shape):\n    # Define variables in build\n    self.count = K.variable(0, name=\"count\")\n    super(CounterLayer, self).build(input_shape)\n\n  def call(self, x, mask=None):\n    updates = []\n    # The format is (variable, value setting to)\n    # So this says \n    # self.pos = self.pos + 1\n    updates.append((self.count, self.count+1))\n\n    # You can append more updates to this list or call add_update more\n    # times if you want\n\n    # Add our custom update\n\n    # We stick x here so it calls our update function every time our layer \n    # is given a new x\n    self.add_update(updates, x)\n\n    # This will be an identity layer but keras gets mad for some reason\n    # if you just output x so we'll multiply it by 1 so it thinks it is a\n    # \"new variable\"\n    return self.count\n  # in newer keras versions you might need to name this compute_output_shape instead\n  def get_output_shape_for(self, input_shape):\n    # We will just return our count as an array ([[count]])\n    return (1,1)\n\n  def reset_states(self):\n    self.count.set_value(0)\n<\/code><\/pre>\n\n<p>Example usage:<\/p>\n\n<pre><code>from keras.layers import Input\nfrom keras.models import Model\nfrom keras.optimizers import RMSprop\ninputLayer = Input(shape=(10,))\ncounter = CounterLayer() # Don't update on predict\n# counter = CounterLayer(stateful=True) # This will update each time you call predict\ncounterLayer = counter(inputLayer)\nmodel = Model(input=inputLayer, output=counterLayer)\noptimizer = RMSprop(lr=0.001)\nmodel.compile(loss=\"mse\", optimizer=optimizer)\n\n\n# See the value of our counter\nprint counter.count.get_value()\n\n# This won't actually train anything but each epoch will update our counter\n\n# Note that if you say have a batch size of 5, update will be called 5 times per epoch\nmodel.fit(np.zeros([1, 10]), np.array([0]), batch_size=1, nb_epoch=5)\n\n# The value of our counter has now changed\nprint counter.count.get_value()\n\nmodel.predict(np.zeros([1, 10]))\n\n# If we did stateful=False, this didn't change, otherwise it did\nprint counter.count.get_value()\n<\/code><\/pre>\n"}]}
{"QuestionId":41646447,"AnswerCount":1,"Tags":"<lambda><theano><keras>","CreationDate":"2017-01-14T03:15:59.583","AcceptedAnswerId":53417838.0,"OwnerUserId":5114796.0,"Title":"Keras multiply layer output with scalar","Body":"<p>I have a layer output I want to multiply by a scalar. I can do this with a lambda layer ie<\/p>\n\n<pre><\/pre>\n\n<p>which works fine. But if I want to use a different scalar for each example, I try to supply these as a second input, with shape (Examples, 1)<\/p>\n\n<pre><\/pre>\n\n<p>therefore my lambda layer becomes<\/p>\n\n<pre><\/pre>\n\n<p>But this now throws an error at train time. Note, 32 is the batch-size, and 128 is a dimension of the layer input (and output) - the layer input being multiplied by the scalar is (batch_size x 32(filters in previous layer) x 128(spatial dim) x 128(spatial dim)).<\/p>\n\n<pre><\/pre>\n\n<p>I assume I am not feeding the right shape in via the input layer, but can't work out why.<\/p>\n","answers":[{"AnswerId":"53417838","CreationDate":"2018-11-21T17:45:30.780","ParentId":null,"OwnerUserId":"7168287","Title":null,"Body":"<p>Not sure if it is useful to answer an old question, but maybe someone else has run into the same problem.<\/p>\n\n<p>The issue is indeed the shape of your scalar versus the shape of your input (or x). You should reshape your scalar to have as many dimensions as the matrix you're multiplying with, using <code>np.reshape<\/code>, e.g.:<\/p>\n\n<pre><code>from keras import *\nfrom keras.layers import *\nimport numpy as np\n\n# inputs\nX = np.ones((32,32,128,128))\ns = np.arange(32).reshape(-1,1,1,1) # 1 different scalar per batch example, reshaped\nprint(X.shape, s.shape)\n\n# model\ninput_X = Input(shape=(32,128,128))\ninput_scalar = Input(shape = (1,1,1))\nsc_mult = Lambda(lambda x: x * input_scalar)(input_X)\nmodel = Model(inputs=[input_X, input_scalar], outputs=sc_mult)\n\nout = model.predict([X,s])\nout\n<\/code><\/pre>\n\n<p>Now <code>out[0,:,:,:]<\/code> is all zeros, <code>out[1,:,:,:]<\/code> is all ones, <code>out[31,:,:,:]<\/code> is all <code>31<\/code>s, et cetera.<\/p>\n"}]}
{"QuestionId":41646532,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-14T03:31:53.567","AcceptedAnswerId":null,"OwnerUserId":2658138.0,"Title":"Usage of tf.name_scope","Body":"<p>My question is about the usage of  in the tutorials code  <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/reader.py#L100\" rel=\"nofollow noreferrer\">ptb\/reader.py<\/a>, <\/p>\n\n<pre><\/pre>\n\n<p>what's the purpose of calling tf.name_scope with the list [raw_data, batch_size, num_steps]?   is a python list,  and  are python int. According to the doc for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/utility_functions#name_scope\" rel=\"nofollow noreferrer\">name_scope<\/a>: <\/p>\n\n<p><\/p>\n\n<p>But , ,  are not any node in the graph. What does it mean by validates them are from the same graph? In fact, the code still runs when the list is removed from the tf.name_scope call: <\/p>\n\n<pre><\/pre>\n\n<p>What's the difference of calling  with and without the ? <\/p>\n","answers":[{"AnswerId":"42511756","CreationDate":"2017-02-28T14:45:37.063","ParentId":null,"OwnerUserId":"7590388","Title":null,"Body":"<p>ptb\/reader.py is using the tf.name_scope() <strong>values<\/strong> argument incorrectly. Graph elements should be passed.<\/p>\n\n<p>name_scope() will make the default graph context match the graph of the passed in elements. This is useful if you are working on multiple graphs. <\/p>\n\n<p>tf.name_scope will raise an error if the list of <strong>values<\/strong> (graph elements) are not all from the same graph.<\/p>\n"}]}
{"QuestionId":41647784,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-01-14T07:33:20.007","AcceptedAnswerId":null,"OwnerUserId":882543.0,"Title":"TFRecordReader seems extremely slow , and multi-threads reading not working","Body":"<p>My training process use tfrecord format for train&amp;eval dataset.<\/p>\n\n<p>I test the benchmark of reader , only 8000records\/second. and io speed(see from iotop command) just 400KB-500KB\/s.<\/p>\n\n<p>I'm using the cpp version of protobuf here <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#protobuf-library-related-issues\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#protobuf-library-related-issues<\/a><\/p>\n\n<p>If possible, provide a minimal reproducible example (We usually don't have time to read hundreds of lines of your code)<\/p>\n\n<pre><\/pre>\n\n<p>What other attempted solutions have you tried?<\/p>\n\n<p>I try to set num_threads in tf.train.shuffle_batch but not working. <\/p>\n\n<p>It seems that when set to 2 threads, it work at 8000records\/s, when enlarge the thread number, it get slower. (I remove all ops that cost cpus. Just read data.)<\/p>\n\n<p>My sever are 24 core cpus.<\/p>\n","answers":[{"AnswerId":"41768163","CreationDate":"2017-01-20T16:25:25.223","ParentId":null,"OwnerUserId":"599139","Title":null,"Body":"<p>Here's a simple speedup building on Yaroslav's answer:<\/p>\n\n<p>Tensorflow has a built in function, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/readers#TFRecordReader.read_up_to\" rel=\"nofollow noreferrer\">tf.TFRecordReader.read_up_to<\/a>,  that reads multiple records in each <code>session.run()<\/code> call, thereby removing the excess overhead caused by multiple calls.<\/p>\n\n<pre><code>enqueue_many_size = SOME_ENQUEUE_MANY_SIZE\nreader = tf.TFRecordReader(options = tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB))\n_, queue_batch = reader.read_up_to(filename_queue, enqueue_many_size)\nbatch_serialized_example = tf.train.shuffle_batch(\n    [queue_batch],\n    batch_size=batch_size,\n    num_threads=thread_number,\n    capacity=capacity,\n    min_after_dequeue=min_after_dequeue,\n    enqueue_many=True)\n<\/code><\/pre>\n\n<p>As with Yaroslav's answer, you need to set <code>enqueue_many=True<\/code> so that the batch function knows it is accepting multiple records.<\/p>\n\n<p>This was very fast in my use case.<\/p>\n"},{"AnswerId":"42847417","CreationDate":"2017-03-17T01:02:11.533","ParentId":null,"OwnerUserId":"7723792","Title":null,"Body":"<p>An addendum to Yaroslav's answer:\nYou can use <code>tf.python_io.tf_record_iterator<\/code> to iterate through the examples in order append them to a list which you can pass to <code>tf.train.shuffle_batch<\/code> with <code>enqueue_many=true<\/code>:<\/p>\n\n<pre><code>queue_batch = []\nfor serialized_example in tf.python_io.tf_record_iterator(filename,options = tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB)):\n    queue_batch.append(serialized_example)\nbatch_serialized_example = tf.train.shuffle_batch(\n    [queue_batch],\n    batch_size=batch_size,\n    num_threads=thread_number,\n    capacity=capacity,\n    min_after_dequeue=min_after_dequeue,\n    enqueue_many=True)\n<\/code><\/pre>\n\n<p>It seems that trying to iterate through examples by using <code>reader.read()<\/code> will result in one read per batch. i.e. the nth batch will be <code>batch_num<\/code> copies of the nth record rather than <code>batch_num<\/code> many unique records.<\/p>\n"},{"AnswerId":"41688401","CreationDate":"2017-01-17T03:00:36.097","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>The issue here is that there's a fixed cost overhead to each <code>session.run<\/code>, and filling the queue with many tiny examples to the queue will be slow.<\/p>\n\n<p>In particular, each session.run is about 100-200 usec, so you can only do about 5k-10k <code>session.run<\/code> calls per second.<\/p>\n\n<p>This problem is obvious if doing Python profiling (python -m cProfile), but hard to see if starting from timeline profile, or CPU profile.<\/p>\n\n<p>The work-around is to use <code>enqueue_many<\/code> to add things to your queue in batches. I took your benchmark from <a href=\"https:\/\/gist.github.com\/ericyue\/7705407a88e643f7ab380c6658f641e8\" rel=\"noreferrer\">https:\/\/gist.github.com\/ericyue\/7705407a88e643f7ab380c6658f641e8<\/a> and modified it to enqueue many items per <code>.run<\/code> call, and that gives 10x speed-up.<\/p>\n\n<p>The modification is to modify <code>tf.batch<\/code> call as follows:<\/p>\n\n<pre><code>if enqueue_many:\n    reader = tf.TFRecordReader(options = tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB))\n    queue_batch = []\n    for i in range(enqueue_many_size):\n        _, serialized_example = reader.read(filename_queue)\n        queue_batch.append(serialized_example)\n    batch_serialized_example = tf.train.shuffle_batch(\n        [queue_batch],\n        batch_size=batch_size,\n        num_threads=thread_number,\n        capacity=capacity,\n        min_after_dequeue=min_after_dequeue,\n        enqueue_many=True)\n<\/code><\/pre>\n\n<p>For complete source, check here:\n<a href=\"https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/ericyue-slowreader\/benchmark.py\" rel=\"noreferrer\">https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/ericyue-slowreader\/benchmark.py<\/a><\/p>\n\n<p>It's hard to optimize it to go much faster since now most of the time is spent in queue operations. Looking at <a href=\"https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/ericyue-slowreader\/benchmark-synthetic-batch.py\" rel=\"noreferrer\">stripped down<\/a> version which just adds integers to a queue, you also get similar speed, and looking at timeline, time is spent in dequeue ops.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OcWpt.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OcWpt.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Each dequeue op takes about 60 usec, but there's on average 5 runnning in parallel, so you get 12 usec per dequeue. So that means you'll get &lt;200k examples per second in the best case.<\/p>\n"}]}
{"QuestionId":41648129,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-01-14T08:29:03.283","AcceptedAnswerId":41648242.0,"OwnerUserId":1934212.0,"Title":"balancing an imbalanced dataset with keras image generator","Body":"<p>The keras <\/p>\n\n<pre><\/pre>\n\n<p>can be used to \"Generate batches of tensor image data with real-time data augmentation\" (<a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"noreferrer\">https:\/\/keras.io\/preprocessing\/image\/<\/a>)<\/p>\n\n<p>The tutorial at <\/p>\n\n<p><a href=\"https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html\" rel=\"noreferrer\">https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html<\/a><\/p>\n\n<p>demonstrates how a small but balanced dataset can be augmented using the ImageDataGenerator. Is there an easy way to use this generator to augment a heavily unbalanced dataset, such that the resulting, generated dataset is balanced?<\/p>\n","answers":[{"AnswerId":"41648242","CreationDate":"2017-01-14T08:47:57.463","ParentId":null,"OwnerUserId":"5114796","Title":null,"Body":"<p>This would not be a standard approach to deal with unbalanced data. Nor do I think it would be really justified - you would be significantly changing the distributions of your classes, where the smaller class is now much less variable. The larger class would have rich variation, the smaller would be many similar images with small affine transforms. They would live on a much smaller region in image space than the majority class.<\/p>\n\n<p>The more standard approaches would be:<\/p>\n\n<ul>\n<li>the class_weights argument in model.fit, which you can use to make the model learn more from the minority class. <\/li>\n<li>reducing the size of the majority class.<\/li>\n<li>accepting the imbalance. Deep learning can cope with this, it just needs lots more data (the solution to everything, really).<\/li>\n<\/ul>\n\n<p>The first two options are really kind of hacks, which may harm your ability to cope with real world (imbalanced) data. Neither really solves the problem of low variability, which is inherent in having too little data. If application to a real world dataset after model training isn't a concern and you just want good results on the data you have, then these options are fine (and much easier than making generators for a single class).<\/p>\n\n<p>The third option is the right way to go if you have enough data (as an example, the <a href=\"http:\/\/jamanetwork.com\/journals\/jama\/fullarticle\/2588763\" rel=\"noreferrer\">recent paper from Google about detecting diabetic retinopathy<\/a> achieved high accuracy in a dataset where positive cases were between 10% and 30%).<\/p>\n\n<p>If you truly want to generate a variety of augmented images for one class over another, it would probably be easiest to do it in pre-processing. Take the images of the minority class and generate some augmented versions, and just call it all part of your data. Like I say, this is all pretty hacky.<\/p>\n"}]}
{"QuestionId":41649038,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-14T10:40:54.920","AcceptedAnswerId":null,"OwnerUserId":1616677.0,"Title":"Linear regression with a categorical variable","Body":"<p>I have a linear model with one dependent variable y = b0 + b1 * x and need to add a categorical feature to it. The problem is that training for each category should result in its own curve (b0, b1 pair). Each category is a different process. I'm tempted to use the simplest way: create a dict and train a separate model for each process. <\/p>\n\n<p>Does this problem fall into regular linear regression? Can I use tf.contrib.learn.LinearRegressor?<\/p>\n\n<p>I've read about <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/tutorials\/wide\/\" rel=\"nofollow noreferrer\">TF's linear model<\/a> and <a href=\"http:\/\/www.psychstat.missouristate.edu\/multibook\/mlt08m.html\" rel=\"nofollow noreferrer\">categorical variables in statistics<\/a> and others, but seems its don't address the issue (or I cannot apply it correctly).<\/p>\n","answers":[]}
{"QuestionId":41649708,"AnswerCount":1,"Tags":"<tensorflow><distributed-computing>","CreationDate":"2017-01-14T11:56:25.397","AcceptedAnswerId":41657185.0,"OwnerUserId":4394807.0,"Title":"Tensorflow: is there a rule to set the port of worker\/ps when creating ClusterSpec?","Body":"<p>When creating  in distributed setting, we need to assign  to workers\/ps. Is there a rule for choosing ? Or is it free to set it to be any number?<\/p>\n","answers":[{"AnswerId":"41657185","CreationDate":"2017-01-15T02:22:50.113","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can set it to any available TCP port on the host where you start the <code>tf.train.Server<\/code>.<\/p>\n\n<p>I usually use port <code>2222<\/code> because it's easy to type (and that number has ended up in a lot of the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">documentation<\/a> that I originally wrote).<\/p>\n"}]}
{"QuestionId":41650708,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-14T13:49:05.927","AcceptedAnswerId":41654054.0,"OwnerUserId":783758.0,"Title":"How do I debug why a node is evaluated?","Body":"<p>In Tensorflow, Placeholders must only be fed if the goal depends on it:<\/p>\n\n<pre><\/pre>\n\n<p>Now, in my more complex graph, I have the problem that I get an error that some placeholders are not fed, but I think that they shouldn't be needed, by the same mechanism as illustrated above.<\/p>\n\n<p>How can I debug this? The error message only states which placeholder is needed, but not why. It would be helpful to get the path from the placeholder to the goal.<\/p>\n\n<p>How can I get this information?<\/p>\n","answers":[{"AnswerId":"41654054","CreationDate":"2017-01-14T19:25:05.117","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>If the graph is not huge, you could just do backward graph search from target node<\/p>\n\n<p>ie, <\/p>\n\n<pre><code>def find(start, target):\n    \"\"\"Returns path to parent from given start node\"\"\"\n    if start == target:\n        return [target]\n    for parent in start.op.inputs:\n        found_path = find(parent, target)\n        if found_path:\n            return [start]+found_path\n    return []\n<\/code><\/pre>\n\n<p>To use it<\/p>\n\n<pre><code>tf.reset_default_graph()\na1 = tf.ones(())\nb1 = tf.ones(())\na2 = 2*a1\nb2 = 2*b1\na3 = 2*a2\nb3 = 2*b2\nd4 = b3+a3\nfind(d4, a1)\n<\/code><\/pre>\n\n<p>Should return<\/p>\n\n<pre><code>[&lt;tf.Tensor 'add:0' shape=() dtype=float32&gt;,\n &lt;tf.Tensor 'mul_2:0' shape=() dtype=float32&gt;,\n &lt;tf.Tensor 'mul:0' shape=() dtype=float32&gt;,\n &lt;tf.Tensor 'ones:0' shape=() dtype=float32&gt;]\n<\/code><\/pre>\n\n<p>If the graph is large, you can speed it up by restricting the search to ops between them<\/p>\n\n<pre><code>import tensorflow.contrib.graph_editor as ge\nops_between = ge.get_walks_intersection_ops(source, target)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.graph_editor\/module_select#get_walks_intersection_ops\" rel=\"nofollow noreferrer\">ge.get_walks_intersection_ops doc<\/a><\/p>\n"}]}
{"QuestionId":41650828,"AnswerCount":1,"Tags":"<machine-learning><theano><keras>","CreationDate":"2017-01-14T14:02:44.987","AcceptedAnswerId":null,"OwnerUserId":5977562.0,"Title":"Keras model predict train\/test shape","Body":"<p>I am training a CNN with Keras but with 30x30 patches from an image. I want to test the network with a full image but I get the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: GpuElemwise. Input dimension mis-match. Input 2 (indices\n  start at 0) has shape[1] == 30, but the output's size on that axis is\n  100. Apply node that caused the error: GpuElemwise{Composite{((i0 + i1) - i2)}}[(0, 0)](GpuDimShuffle{0,2,3,1}.0, GpuReshape{4}.0,\n  GpuFromHost.0) Toposort index: 79 Inputs types:\n  [CudaNdarrayType(float32, 4D), CudaNdarrayType(float32, (True, True,\n  True, False)), CudaNdarrayType(float32, 4D)] Inputs shapes: [(10, 100,\n  100, 3), (1, 1, 1, 3), (10, 30, 30, 3)] Inputs strides: [(30000, 100,\n  1, 10000), (0, 0, 0, 1), (2700, 90, 3, 1)] Inputs values: ['not\n  shown', CudaNdarray([[[[ 0.01060364  0.00988821  0.00741314]]]]), 'not\n  shown'] Outputs clients:\n  [[GpuCAReduce{pre=sqr,red=add}{0,1,1,1}(GpuElemwise{Composite{((i0 +\n  i1) - i2)}}[(0, 0)].0)]]<\/p>\n<\/blockquote>\n\n<p>This is my model.predict:<\/p>\n\n<pre><\/pre>\n\n<p>It's seems like the issue is that the input size cannot be anything other than 30x30 but the first input shape for the first layer of my network is none, none, 3. <\/p>\n\n<pre><\/pre>\n\n<p>Is it simply not possible to test an image with different dimensions to the ones I trained with?<\/p>\n","answers":[{"AnswerId":"41661342","CreationDate":"2017-01-15T12:58:26.530","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>As fchollet himself described <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1920\" rel=\"nofollow noreferrer\" title=\"here\">here<\/a>, you should be able to define the input as so:<\/p>\n\n<pre><code>input_shape=(1, None, None)\n<\/code><\/pre>\n\n<p>However this will fail if you have layers that use the Flatten operation.<\/p>\n\n<p>This suggests that you should be able to accomplish your goal with a fully convolutional NN.<\/p>\n"}]}
{"QuestionId":41651011,"AnswerCount":2,"Tags":"<keras><floating-accuracy><threshold>","CreationDate":"2017-01-14T14:22:33.000","AcceptedAnswerId":null,"OwnerUserId":7418689.0,"Title":"binary_accuracy in keras Metrices , what's the threshold value to predicted one sample as positive and negative cases","Body":"<p>What's the threshold value of  in keras Metrices is used to predicted one sample as positive and negative cases? is that threshold value 0.5? how to adjust it? I want to set the threshold value 0.80, if the predicted value is 0.79, then it is considered a negative sample\uff0cotherwise\uff0cif the predicted value is 0.81, then it is considered a positive sample.<\/p>\n","answers":[{"AnswerId":"41657585","CreationDate":"2017-01-15T03:38:08.000","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p><code>binary_accuracy<\/code> don't have threshold param but you can easily define one yourself. <\/p>\n\n<pre><code>import keras\nfrom keras import backend as K\n\ndef threshold_binary_accuracy(y_true, y_pred):\n    threshold = 0.80\n    if K.backend() == 'tensorflow':\n        return K.mean(K.equal(y_true, K.tf.cast(K.lesser(y_pred,threshold), y_true.dtype)))\n    else:\n        return K.mean(K.equal(y_true, K.lesser(y_pred,threshold)))\n\na_pred = K.variable([.1, .2, .6, .79, .8, .9])\na_true = K.variable([0., 0., 0.,  0., 1., 1.])\n\nprint K.eval(keras.metrics.binary_accuracy(a_true, a_pred))\nprint K.eval(threshold_binary_accuracy(a_true, a_pred))\n<\/code><\/pre>\n\n<p>Now you can use it as <code>metrics=[threshold_binary_accuracy]<\/code><\/p>\n"},{"AnswerId":"42564712","CreationDate":"2017-03-02T19:53:49.373","ParentId":null,"OwnerUserId":"7197896","Title":null,"Body":"<p>To answer the initial question, keras uses the round function to assign classes, so the threshold is 0.5.<\/p>\n\n<p><a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/metrics.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/metrics.py<\/a><\/p>\n\n<p>def binary_accuracy(y_true, y_pred):\n    return K.mean(K.equal(y_true, K.round(y_pred)))<\/p>\n"}]}
{"QuestionId":41651034,"AnswerCount":1,"Tags":"<python><input><tensorflow><jpeg>","CreationDate":"2017-01-14T14:25:42.480","AcceptedAnswerId":null,"OwnerUserId":3205742.0,"Title":"input pipeline in tensorflow","Body":"<p>While practicing on the official tensorflow mnist dataset tutorial for beginners, I'm trying to change the mnist data to my own images collected from search engines. <\/p>\n\n<pre><\/pre>\n\n<p>That strFilePaths is a standard python list containing all my image paths , iLabels is a list of lists representing labels. And I have only 2 classes in this case.<\/p>\n\n<p>This program runs without error output but tensorflow just keeps on running and not giving me any output. I've read the \"reading files\" session in tensorflow website for like a thousand times but I still don't have a clue on whether I did things right or not.<\/p>\n\n<p>Q1: What's wrong with this code?\nQ2: Is there any complete example on how to read jpeg files into tensorflow and perform some training tasks on them?<\/p>\n","answers":[{"AnswerId":"43854214","CreationDate":"2017-05-08T17:52:23.090","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>Unfortunately, without access to your code and files, I am unable to help you debug further. However, you can see a full example of how to retrain Inception to recognize new categories of information (e.g. flowers) in the <code>image_retraining<\/code> example at: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/examples\/image_retraining\/retrain.py<\/a> <\/p>\n"}]}
{"QuestionId":41651229,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-14T14:48:14.940","AcceptedAnswerId":null,"OwnerUserId":3267509.0,"Title":"Using retrained data with classify_image.py","Body":"<p>I've been using tensorflow image recognition. I've build many scripts which interact with classify_image.py.<\/p>\n\n<p>I also retrained the model using retrain.py, with my own dataset.<\/p>\n\n<p>How can use the two files generated: output_graph.pb, output_labels.txt with classify_image.py ?<\/p>\n","answers":[{"AnswerId":"43859880","CreationDate":"2017-05-09T02:05:00.477","ParentId":null,"OwnerUserId":"162273","Title":null,"Body":"<p>Ah, <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">the docs say<\/a><\/p>\n\n<blockquote>\n  <p>If you'd like to use the retrained model in a Python program <a href=\"https:\/\/github.com\/eldor4do\/TensorFlow-Examples\/blob\/master\/retraining-example.py\" rel=\"nofollow noreferrer\">this example from @eldor4do shows what you'll need to do.<\/a><\/p>\n<\/blockquote>\n\n<p>Just copied\/edited that one file locally, and <code>python .\\edited-retraining-example.py<\/code>.  And that was easy.<\/p>\n\n<p>Note that if you're on Windows, change all examples of <code>\/tmp\/...<\/code> to <code>c:\/tmp\/...<\/code>.<\/p>\n"}]}
{"QuestionId":41651628,"AnswerCount":5,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-01-14T15:27:53.967","AcceptedAnswerId":41742406.0,"OwnerUserId":2326911.0,"Title":"Negative dimension size caused by subtracting 3 from 1 for 'Conv2D'","Body":"<p>I'm using <a href=\"https:\/\/keras.io\/\" rel=\"noreferrer\">Keras<\/a> with <a href=\"https:\/\/www.tensorflow.org\/\" rel=\"noreferrer\">Tensorflow<\/a> as backend , here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>and Trackback error:<\/p>\n\n<pre><\/pre>\n\n<p>First I saw some answers that problem is with  version so I upgrade  to , but still exist , is that problem with network or I missing something, what should  looks like?<\/p>\n\n<p><strong>Update<\/strong>\nHere is :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"45433384","CreationDate":"2017-08-01T09:00:04.533","ParentId":null,"OwnerUserId":"5187177","Title":null,"Body":"<p>FWIW, I got this error repeatedly with some values of strides or kernel_size but not all, with the backend and image_ordering already set as tensorflow's, and they all disappeared when I added <code>padding=\"same\"<\/code><\/p>\n"},{"AnswerId":"41742406","CreationDate":"2017-01-19T13:00:08.753","ParentId":null,"OwnerUserId":"2911687","Title":null,"Body":"<p>Your issue comes from the <code>image_ordering_dim<\/code> in <code>keras.json<\/code>.<\/p>\n\n<p>From <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"noreferrer\">Keras Image Processing doc<\/a>:     <\/p>\n\n<blockquote>\n  <p>dim_ordering: One of {\"th\", \"tf\"}. \"tf\" mode means that the images should have shape (samples, height, width, channels), \"th\" mode means that the images should have shape (samples, channels, height, width). It defaults to the image_dim_ordering value found in your Keras config file at ~\/.keras\/keras.json. If you never set it, then it will be \"tf\".<\/p>\n<\/blockquote>\n\n<p>Keras maps the convolution operation to the chosen backend (theano or tensorflow). However, both backends have made different choices for the ordering of the dimensions. If your image batch is of N images of HxW size with C channels, theano uses the NCHW ordering while tensorflow uses the NHWC ordering.<\/p>\n\n<p>Keras allows you to choose which ordering you prefer and will do the conversion to map to the backends behind. But if you choose <code>image_ordering_dim=\"th\"<\/code> it expects Theano-style ordering (NCHW, the one you have in your code) and if <code>image_ordering_dim=\"tf\"<\/code> it expects tensorflow-style ordering (NHWC).<\/p>\n\n<p>Since your <code>image_ordering_dim<\/code> is set to <code>\"tf\"<\/code>, if you reshape your data to the tensorflow style it should work:<\/p>\n\n<pre><code>X_train = X_train.reshape(X_train.shape[0], img_cols, img_rows, 1)\nX_test = X_test.reshape(X_test.shape[0], img_cols, img_rows, 1)\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>input_shape=(img_cols, img_rows, 1)\n<\/code><\/pre>\n"},{"AnswerId":"44964655","CreationDate":"2017-07-07T07:10:44.553","ParentId":null,"OwnerUserId":"7715045","Title":null,"Body":"<p>Just add this:<\/p>\n\n<pre><code>from keras import backend as K\nK.set_image_dim_ordering('th')\n<\/code><\/pre>\n"},{"AnswerId":"53689732","CreationDate":"2018-12-09T05:44:08.420","ParentId":null,"OwnerUserId":"9209481","Title":null,"Body":"<p>I faced the same problem, but it was solved by changing the conv2d function:<\/p>\n\n<pre><code>`\nif K.image_data_format=='channels_first':\n    x_train = x_train.reshape(x_train.shape[0], 1,img_cols,img_rows)\n    x_test = x_test.reshape(x_test.shape[0], 1,img_cols,img_rows)\n    input_shape = (1,img_cols,img_rows)\nelse:\n    x_train = x_train.reshape(x_train.shape[0],img_cols,img_rows,1)\n    x_test = x_test.reshape(x_test.shape[0],img_cols,img_rows,1)\n    input_shape = (img_cols,img_rows,1)\nmodel.add(Convolution2D(32,(3, 3), input_shape = input_shape, activation=\"relu\"))\n`\n<\/code><\/pre>\n"},{"AnswerId":"57067482","CreationDate":"2019-07-17T01:49:59.503","ParentId":null,"OwnerUserId":"4374376","Title":null,"Body":"<p>I am also having the same problem. However, each Conv3D layer, I am using, is reducing size of the input. So, including one parameter padding='same' during declaring the Conv2D\/3D layer solved the problem. Here is the demo code <\/p>\n\n<pre><code>model.add(Conv3D(32,kernel_size=(3,3,3),activation='relu',padding='same'))\n<\/code><\/pre>\n\n<p>Also reducing the size of the filter can also solve the problem. <\/p>\n"}]}
{"QuestionId":41651701,"AnswerCount":2,"Tags":"<python-2.7><deep-learning><caffe><pycaffe><deeplearning4j>","CreationDate":"2017-01-14T15:34:50.803","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"Why the FCN32 output is a black image (all pixels have zero values)?","Body":"<p>I am trying to get the output from FCN 32. I trained FCN32 with pascalcontext-fcn32-heavy.caffemodel pre-trained model. I could run for grayscale images with 5 classes. However, during inference, the output is all zero (a black image). This is inference code:<\/p>\n\n<pre><\/pre>\n\n<p>This the log information of inference: <\/p>\n\n<pre><\/pre>\n\n<p>I have two major questions:<\/p>\n\n<ol>\n<li>I am wondering why the output is black? and<\/li>\n<li>How can I know when to stop running the algorithm (i.e., iteration\nnumber)? I really do not know what is the optimum iteration number and\nloss value that I can stop fine tuning in that stage. I stopped\ntraining in , I have no idea about this.<\/li>\n<li>Is it necessary that the result of segmentation be a grayscale image\nas well (like input), or creating RGB result image does not make any\ndifference in the output?<\/li>\n<\/ol>\n\n<p>I really do not know how much I am doing the right way. Quite CONFUSED :(\nDoes anyone have any suggestion? I really appreciate your help.<\/p>\n","answers":[{"AnswerId":"41777703","CreationDate":"2017-01-21T09:27:15.740","ParentId":null,"OwnerUserId":"7082570","Title":null,"Body":"<p>Yes, it usually depends on your image size! Have you checked your data type? Both your images and groundtruths should be uint8! <\/p>\n\n<p>Also have you add the \"group\" line to your Deconv layer?<\/p>\n\n<p>Best<\/p>\n"},{"AnswerId":"41774532","CreationDate":"2017-01-21T00:36:33.457","ParentId":null,"OwnerUserId":"7082570","Title":null,"Body":"<p>Make sure that your labels' data type is uint8! I had the same problem!<\/p>\n\n<p>Also make sure that you have weight filler like below in your prototxt before training!<\/p>\n\n<pre><code>layer {\nname: \"myupscore2\"\ntype: \"Deconvolution\"\nbottom: \"myscore_fr\"\ntop: \"myupscore2\"\nparam {\nlr_mult: 5\n}\nconvolution_param {\ngroup :2\nnum_output: 2\nweight_filler: { type: \"bilinear\" }\nbias_term: false\nkernel_size: 4\nstride: 16\n }\n }\n<\/code><\/pre>\n\n<p>good luck!<\/p>\n"}]}
{"QuestionId":41651842,"AnswerCount":1,"Tags":"<debugging><tensorflow><neural-network><keras>","CreationDate":"2017-01-14T15:47:22.717","AcceptedAnswerId":null,"OwnerUserId":2991813.0,"Title":"Why is my CNN not learning","Body":"<p>I am sorry for such a cliche question, but I really don't know why my CNN is not improving. <\/p>\n\n<p>I am training a CNN for SVHN dataset (single digit) with images of 32x32.<\/p>\n\n<p>For preprocessing, I transform RGB to grayscale and normalize all pixel data by standardization. So the data range becomes (-1,1). To verify that my  and  correspond to each other correctly, I randomly pick an image from  and a label from  with the same index, and it shows that they do.<\/p>\n\n<p><img src=\"https:\/\/www.dropbox.com\/s\/6tmotiuur17c3kv\/random_sample.png?raw=1\" alt=\"\"><\/p>\n\n<p>Here's my code (Keras, tensorflow backend):<\/p>\n\n<pre><\/pre>\n\n<p>After running 10 epochs, the accuracy is still the same as in the first epoch, and that's why I stopped it.<\/p>\n\n<pre><\/pre>\n\n<p>Should I keep trying with more patience or is there something wrong with my CNN?<\/p>\n","answers":[{"AnswerId":"51482292","CreationDate":"2018-07-23T15:22:24.527","ParentId":null,"OwnerUserId":"7042560","Title":null,"Body":"<p>Try switching your optimizer to Adam, as it is more capable than SGD. You can include Nesterov momentum with nAdam. So i would try the following.<\/p>\n\n<pre><code>model.compile(loss='categorical_crossentropy',\n              optimizer='nadam',\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>This will adjust learning rates automatically and you don't need to worry about it as much.<\/p>\n"}]}
{"QuestionId":41651874,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-14T15:50:35.803","AcceptedAnswerId":null,"OwnerUserId":7147097.0,"Title":"Tensorflow: LSTM when each part of a sequence belongs to a class?","Body":"<p>I'm trying to implement LSTM in Tensorflow, but my data is a bit \"untraditional\" in the following way:\nEach hour of the day, one of five events can happen. I want to feed sequences of 24 hours into my LSTM.<\/p>\n\n<p>The thing is, I need it to output a target class for each element of each sequence. <\/p>\n\n<p>This has to be possible with LSTM's, right? The only examples of TF implementations of LSTM's have been just feeding a sequence and getting a target class for that sequence.<\/p>\n\n<p>Any help would be appreciated!  <\/p>\n","answers":[{"AnswerId":"41668629","CreationDate":"2017-01-16T02:27:37.580","ParentId":null,"OwnerUserId":"2491303","Title":null,"Body":"<ol>\n<li><p>Element level classification is definitely possible in TF.\nFor example, take a look at this StackOverflow <a href=\"https:\/\/stackoverflow.com\/questions\/33737611\/using-tensorflow-for-sequence-tagging-synced-sequence-input-and-output\">question<\/a> about POS-tagging and the references therein. (if the reference seems too complicated, I would search for other TF references about POS tagging).<\/p><\/li>\n<li><p>Your question does not describe how sequences differ one from another. What are the actual elements in an input sequence? The hours 1..24?<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":41653515,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-14T18:31:30.040","AcceptedAnswerId":null,"OwnerUserId":3128177.0,"Title":"How to set default variable collection in tensorflow?","Body":"<p>I want to have something like this:<\/p>\n\n<pre><\/pre>\n\n<p>Is there anything I can do about it? <\/p>\n","answers":[{"AnswerId":"41655985","CreationDate":"2017-01-14T23:04:21.363","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Use TF-Slim's <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/framework\/python\/ops\/arg_scope.py\" rel=\"nofollow noreferrer\">arg_scope<\/a> combined with arg_scope aware <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/framework\/python\/ops\/variables.py\" rel=\"nofollow noreferrer\">variable<\/a><\/p>\n"},{"AnswerId":"44277938","CreationDate":"2017-05-31T07:07:28.050","ParentId":null,"OwnerUserId":"298233","Title":null,"Body":"<p>This is not very convenient, but you could define a <code>custom_getter<\/code> using <code>tf.variable_scope<\/code>:<\/p>\n\n<pre><code>def make_local_getter(getter, *args, **kwargs):\n    if kwargs['collections'] is None:\n        kwargs['collections'] = []\n    kwargs['collections'] += [tf.GraphKeys.LOCAL_VARIABLES]\n    return getter(*args, **kwargs)\n\nwith tf.variable_scope(scope, custom_getter=make_local_getter):\n    tf.get_variable(...)\n<\/code><\/pre>\n"}]}
{"QuestionId":41653744,"AnswerCount":1,"Tags":"<python><linear-algebra><theano><matrix-multiplication>","CreationDate":"2017-01-14T18:55:15.037","AcceptedAnswerId":41654447.0,"OwnerUserId":6332289.0,"Title":"Matrix multiplication with bitwise operators","Body":"<p>I want to create a bitmask vector that masks which values are greater than a given value. Something like [1, 2, 3, 4, 5] * [>3, >3, >3, >3, >3] = [0, 0, 0, 1, 1]. I want to be able to run this on theano to get faster computation time for matrix operations. Is there a linear algebra procedure that can be written using bitwise operators or bits to create this bitmask? I am currently looping through this matrix and I would like to move the computation to a GPU using theano which requires more matrix multiplication. Thanks for any help.<\/p>\n","answers":[{"AnswerId":"41654447","CreationDate":"2017-01-14T20:05:51.543","ParentId":null,"OwnerUserId":"888731","Title":null,"Body":"<p>You can get exactly what you want with logical operations between matrices.\nFor example<\/p>\n\n<pre><code>print((np.r_[1, 2, 3, 4, 5] &gt; 3))\n<\/code><\/pre>\n\n<p>will give<\/p>\n\n<pre><code>[False False False  True  True]\n<\/code><\/pre>\n\n<p>And if you want integers you can do<\/p>\n\n<pre><code>print((np.r_[1, 2, 3, 4, 5] &gt; 3).astype(int) )\n<\/code><\/pre>\n\n<p>and get<\/p>\n\n<pre><code>[0 0 0 1 1]\n<\/code><\/pre>\n"}]}
{"QuestionId":41656162,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><neural-network><tflearn>","CreationDate":"2017-01-14T23:33:36.663","AcceptedAnswerId":42578691.0,"OwnerUserId":4099925.0,"Title":"Multi Column Deep Neural Network with TFLearn and Tensorflow","Body":"<p>I am trying to build a multi column deep neural network (MDNN) with tflearn and tensorflow. The MDNN is explained in <a href=\"http:\/\/people.idsia.ch\/~juergen\/nn2012traffic.pdf\" rel=\"nofollow noreferrer\">this paper<\/a>. The part I am struggling with is how I can add two or more inputs together to be fed to tensorflow.<\/p>\n\n<p>For a single column I have:<\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre>\n\n<p>where  is of shape . I am pretty new to numpy, tensorflow and tflearn. The difficulty for now really lays in how to specify multiple inputs to tflearn.<\/p>\n\n<p>Any help is greatly appreciated.<\/p>\n","answers":[{"AnswerId":"42578691","CreationDate":"2017-03-03T12:17:08.700","ParentId":null,"OwnerUserId":"2695508","Title":null,"Body":"<p>The MDNN explained in the paper individually trains several models using random (but bounded) distortions on the data. Once all models are trained, they produce predictions using an ensemble classifier by averaging the output of all the models on different versions of the data.<\/p>\n\n<p>As far as I understand, the columns are not jointly but independently trained.  So you must create different models and call fit on each on them. I recommend you start training a single model and once you have a training setting getting good results, replicate it. To generate predictions, you must compute the average of the predicted probabilities from the <em>predict<\/em> function and take the most probable class.<\/p>\n\n<p>One way to a generate data from your inputs is to use <a href=\"http:\/\/tflearn.org\/data_augmentation\/#data-augmentation\" rel=\"nofollow noreferrer\">data augmentation<\/a>. However, instead of generating new data you must replace it by the modified versions.<\/p>\n"}]}
{"QuestionId":41656564,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-15T00:35:37.993","AcceptedAnswerId":null,"OwnerUserId":1128571.0,"Title":"Cannot imoprt tensorflow on mac with python 2.7","Body":"<p>I installed TensorFlow using pip, but when I try to import it I get error:<\/p>\n\n<pre><\/pre>\n\n<p>I am running numpy version 1.11.3 and python version 2.7.9<\/p>\n","answers":[{"AnswerId":"42665119","CreationDate":"2017-03-08T07:13:05.040","ParentId":null,"OwnerUserId":"1384641","Title":null,"Body":"<p>For <strong>AttributeError: type object 'NewBase' has no attribute 'is_abstract'<\/strong> (mentioned in comments)<\/p>\n\n<p>Upgrading six will resolve the issue. TensorFlow requires <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/pip_package\/setup.py#L41\" rel=\"nofollow noreferrer\">six => 1.10<\/a> - to check the six version, via interpreter:<\/p>\n\n<pre><code>$ python\n&gt;&gt;&gt;import six\n&gt;&gt;&gt;print six.\\__version\\__\n<\/code><\/pre>\n\n<p>Or, you can simply check it using pip:<\/p>\n\n<pre><code>$ pip show six\n<\/code><\/pre>\n\n<p>If you have six version less than 1.10, upgrade it:<\/p>\n\n<pre><code>$ pip install --upgrade six\n<\/code><\/pre>\n\n<hr>\n\n<p>Source: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1965\" rel=\"nofollow noreferrer\">Github thread<\/a>. <\/p>\n"}]}
{"QuestionId":41656682,"AnswerCount":1,"Tags":"<deep-learning><caffe><matrix-multiplication>","CreationDate":"2017-01-15T00:53:05.900","AcceptedAnswerId":null,"OwnerUserId":628538.0,"Title":"Matrix multiplication between two feature maps in Caffe","Body":"<p>Is there a layer (or a simple combination of layers) which can be used to perform matrix multiplication between two feature maps? For example, say there is a blob is of size MxN and another one of size NxP, is there a layer which can multiply these two blobs like a matrix?<\/p>\n","answers":[{"AnswerId":"41658699","CreationDate":"2017-01-15T07:05:56.163","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>There is a planned change to caffe to allow for manipulations as you ask, that is, treating parameter blobs as regular blobs. See <a href=\"https:\/\/stackoverflow.com\/a\/41580924\/1714410\">this answer<\/a> for more information.<\/p>\n\n<p>Until this change is complete, you might consider<\/p>\n\n<ol>\n<li><code>\"Tile\"<\/code>ing the blobs into <code>M<\/code>x<code>N<\/code>x<code>P<\/code> blobs.  <\/li>\n<li>Using <code>\"Eltwise\"<\/code> layer for element-wise multiplication.  <\/li>\n<li>Using <code>\"Pooling\"<\/code> with global pool on the <code>N<\/code> axis.<\/li>\n<\/ol>\n"}]}
{"QuestionId":41656799,"AnswerCount":0,"Tags":"<gpu><caffe><layer><timing>","CreationDate":"2017-01-15T01:10:35.677","AcceptedAnswerId":null,"OwnerUserId":2774587.0,"Title":"timing the Backward_gpu function for a particular layer in caffe","Body":"<p>I have added altered the backward_gpu function and I want to measure the change in the time so I have,<\/p>\n\n<pre><\/pre>\n\n<p>But this returns the same value ( it really shouldn't based on the amount of work I've reduced) as it did before.  Is it possible that caffe is not waiting until everything is done inside to compute the time? Meaning the work inside was never accurately reflected in the time?  When I do similar things to the Forward_gpu function the reduction in time is apparent.  <\/p>\n\n<p>How do I accurately measure the timing of the Backward_gpu function for a particular layer in caffe?<\/p>\n","answers":[]}
{"QuestionId":41657172,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-01-15T02:20:34.577","AcceptedAnswerId":41657318.0,"OwnerUserId":737455.0,"Title":"What does * (asterisk) do when applied to a TensorFlow layer?","Body":"<p>Currently reading a Python implementation of Inception-ResNet to assist in building a model in a different language (Deeplearning4j). This implementation is Inception-ResNet-v1 and I was trying to figure out how it implements the residual shortcuts in ResNet style.<\/p>\n\n<p>In the following code block is .<\/p>\n\n<pre><\/pre>\n\n<p>Scale is a  between 0 and 1.  is a stack of layers, the last being a conv2d layer.<\/p>\n\n<p>What specifically is happening with ?<\/p>\n","answers":[{"AnswerId":"41657318","CreationDate":"2017-01-15T02:48:10.257","ParentId":null,"OwnerUserId":"6862133","Title":null,"Body":"<p>Each layer in <code>up<\/code> is being multiplied by the scalar value in <code>scale<\/code>. And then <code>net<\/code> is being redefined as <code>net + scale * up<\/code>. So <code>net<\/code> should have the same dimensions as <code>up<\/code>.<\/p>\n"}]}
{"QuestionId":41657212,"AnswerCount":0,"Tags":"<python><deep-learning><theano><nvidia>","CreationDate":"2017-01-15T02:28:23.687","AcceptedAnswerId":null,"OwnerUserId":7162700.0,"Title":"Theano: ImportError: cannot import name raise_from","Body":"<p>I am trying to test out Theano on the new TX1.\nI've installed by following the installation instruction.<\/p>\n\n<pre><\/pre>\n\n<p>When I try to import theano in python, I get the following error message.<\/p>\n\n<blockquote>\n  <p>ImportError: cannot import name raise_from<\/p>\n<\/blockquote>\n\n<p>Is there anything wrong with my installation procedure?\nThanks.<\/p>\n","answers":[]}
{"QuestionId":41657305,"AnswerCount":1,"Tags":"<neural-network><normalization><keras>","CreationDate":"2017-01-15T02:46:35.193","AcceptedAnswerId":41696464.0,"OwnerUserId":7400630.0,"Title":"Unormalize an normalized output","Body":"<p>I am currently working with a NN problem, in which i am mapping an input to an output. <\/p>\n\n<p>I've normalized the input and the output such that they are in the same range => -1 to 1 and such that i can use tanh as activation function. <\/p>\n\n<p>I seem to get pretty decent result from the training, val_loss = 0.0156. <\/p>\n\n<p>But to actually use the dataset I have to unormalize the output, which makes it very different from the actual output. <\/p>\n\n<p>I am pretty new to NN , but is unormalizing usually done?? if not? how are these issues then resolved?<\/p>\n\n<p>I am currently normalizing using min-max nomalization, and since i store the max and min value, I use those to return the values. <\/p>\n\n<p><strong>More about the data:<\/strong> <\/p>\n\n<p>The dataset consist of <a href=\"https:\/\/en.wikipedia.org\/wiki\/Short-time_Fourier_transform\" rel=\"nofollow noreferrer\">stft<\/a> audio files, and the output is a feature vector (<a href=\"https:\/\/en.wikipedia.org\/wiki\/Mel-frequency_cepstrum\" rel=\"nofollow noreferrer\">mfcc<\/a>). \nSince stft is giving me complex values, but since keras isn't able to handle complex numbers, I've splitted the real and imaginary part, and concantenated them. <\/p>\n","answers":[{"AnswerId":"41696464","CreationDate":"2017-01-17T11:55:26.877","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>You can use a <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\" title=\"lambda layer\">lambda layer<\/a> to do the unormalization step as part of the network itself.<\/p>\n\n<p>Lets say you can define a normalization function:<\/p>\n\n<pre><code>f = K.function(\/* some function *\/)\n<\/code><\/pre>\n\n<p>And your current model is:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(1000, input_dim=1000))\nmodel.add(Dense(1000))\n<\/code><\/pre>\n\n<p>You can normalize the outputs by doing:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(1000, input_dim=1000))\nmodel.add(Dense(1000))\nmodel.add(Lambda(lambda x: f(x))\n<\/code><\/pre>\n"}]}
{"QuestionId":41657644,"AnswerCount":3,"Tags":"<deep-learning><caffe><pycaffe><deeplearning4j>","CreationDate":"2017-01-15T03:52:19.123","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"How we know when to stop training a model on a pre-trained model?","Body":"<p>My apologies since my question may sound stupid question. But I am quite new in deep learning and caffe.\nHow can we detect how many iterations are required to fine-tune a pre-trained on our own dataset? For example, I am running fcn32 for my own data with 5 classes. When can I stop the fine-tuning process by looking at the loss and accuracy of training phase?<\/p>\n\n<p>Many thanks<\/p>\n","answers":[{"AnswerId":"41659123","CreationDate":"2017-01-15T08:12:32.667","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Normally you converge to a specific validation accuracy for your model. In practice you normally stop training, if the validation loss did not increase in x epochs. Depending on your epoch duration x may vary most commonly between 5 and 20. <\/p>\n\n<p>Edit:<\/p>\n\n<p>An epoch is one iteration over your dataset for trainig in ML terms. You do not seem to have a validation set. Normally the data is split into training and validation data so you can see how well your model performs on unseen data and made decisions about which model to take by looking at this data. You might want to take a look at <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/mnist.html<\/a> to see the usage of a validation set, even though they call it test set. <\/p>\n"},{"AnswerId":"41661473","CreationDate":"2017-01-15T13:15:05.483","ParentId":null,"OwnerUserId":"7396740","Title":null,"Body":"<p>You shouldn't do it by looking at the loss or accuracy of training phase. Theoretically, the training accuracy should always be increasing (also means the training loss should always be decreasing) because you train the network to decrease the training loss. But a high training accuracy doesn't necessary mean a high test accuracy, that's what we referred as over-fitting problem. So what you need to find is a point where the accuracy of test set (or validation set if you have it) stops increasing. And you can simply do it by specifying a relatively larger number of iteration at first, then monitor the test accuracy or test loss, if the test accuracy stops increasing (or the loss stops decreasing) in consistently N iterations (or epochs), where N could be 10 or other number specified by you, then stop the training process. <\/p>\n"},{"AnswerId":"41661630","CreationDate":"2017-01-15T13:30:55.243","ParentId":null,"OwnerUserId":"5687374","Title":null,"Body":"<p>The best thing to do is to track training and validation accuracy and store snapshots of the weights every k iterations. To compute validation accuracy you need to have a sparate set of held out data which you do not use for training.\nThen, you can stop once the validation accuracy stops increasing or starts decreasing. This is called early stopping in the literature. Keras, for example, provides functionality for this: <a href=\"https:\/\/keras.io\/callbacks\/#earlystopping\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/#earlystopping<\/a><\/p>\n\n<p>Also, it's good practice to plot the above quantities, because it gives you important insights into the training process. See <a href=\"http:\/\/cs231n.github.io\/neural-networks-3\/#accuracy\" rel=\"nofollow noreferrer\">http:\/\/cs231n.github.io\/neural-networks-3\/#accuracy<\/a> for a great illustration (not specific to early stopping).<\/p>\n\n<p>Hope this helps<\/p>\n"}]}
{"QuestionId":41658720,"AnswerCount":0,"Tags":"<tensorflow><sublimetext3>","CreationDate":"2017-01-15T07:08:31.500","AcceptedAnswerId":null,"OwnerUserId":7420730.0,"Title":"sublimeJEDI syntax completion not working in sublime text 3 for tensorflow","Body":"<p>Sublime text 3 Build 3126<\/p>\n\n<p>Tensorflow auto-completions not working using sublimeJEDI.<\/p>\n\n<p>Though they work for numpy.<\/p>\n","answers":[]}
{"QuestionId":41658793,"AnswerCount":4,"Tags":"<installation><tensorflow>","CreationDate":"2017-01-15T07:22:41.253","AcceptedAnswerId":null,"OwnerUserId":6380396.0,"Title":"tensorflow installation issues:ImportError: No module named tensorflow","Body":"<p>environments: Ubuntu 14.04(64bit)  Python2.7.11<\/p>\n\n<p>Firstly, I installed tensorflow in the way of Virtualenz installation.<\/p>\n\n<pre><\/pre>\n\n<p>and then, I test my installation and some issue appear. I know I didn't install tensorflow successfully.\nimport tensorflow\nTraceback (most recent call last):\n  File \"\", line 1, in \nImportError: No module named tensorflow<\/p>\n\n<blockquote>\n  <blockquote>\n    <blockquote>\n      <p>import tensorflow as tf\n      Traceback (most recent call last):\n        File \"\", line 1, in \n      ImportError: No module named tensorflow<\/p>\n    <\/blockquote>\n  <\/blockquote>\n<\/blockquote>\n\n<p>I don't know how to solve the problem. Please help me, it cost me one day. I tried to uninstall tensorflow and then I installed in the way of pip installation. But I get the same error.\nThe protocbuf is 3.1.0. <\/p>\n","answers":[{"AnswerId":"51964183","CreationDate":"2018-08-22T09:51:47.313","ParentId":null,"OwnerUserId":"10258870","Title":null,"Body":"<p>INSTALLATION OF TENSORFLOW ON UBUNTU 18.04<\/p>\n\n<p>download anaconda python  package <\/p>\n\n<p>install it via shell using bash<\/p>\n\n<pre><code>$bash anaconda*.sh\n<\/code><\/pre>\n\n<p>editing the .bashrc script \/\/location home<\/p>\n\n<pre><code>$sudo apt-get install python3-pip\n$sudo apt-get update\n$cd\n$nano .bashrc\n<\/code><\/pre>\n\n<p>nano is the text editor\ninsert the given line at the end of the file<\/p>\n\n<pre><code>export PATH=-\/anaconda3\/bin:$PATH\n<\/code><\/pre>\n\n<p>create a virtual environment\nusing conda<\/p>\n\n<pre><code>$conda create -n myenv python=3.5\n<\/code><\/pre>\n\n<p>\/\/SPECIFY THE VERSION REQUIRED DO NOT USE 3.7 AS THERE IS A COMPATIBLITY ISSUE WITH TENSORFLOW 10<\/p>\n\n<pre><code>$source activate myenv\n$pip install -U tensorflow\n$python\n&gt;&gt;import tensorflow as tf\n&gt;&gt;  \/\/get this prompt without an error it means the installation is successful\n&gt;&gt;exit()\nsource deactivate\n<\/code><\/pre>\n\n<p>fully tested if an issue arises do let me know\nwhenever you install python packages i would suggest to do it in a virtual environment<\/p>\n"},{"AnswerId":"49837050","CreationDate":"2018-04-14T23:15:38.223","ParentId":null,"OwnerUserId":"9647167","Title":null,"Body":"<p>In case, your prerequisite python packages are not installed properly,\ncheck several things.<\/p>\n\n<pre><code>$ source $HOME\/tensorflow\/bin\/activate\n$ which python\n$ which pip\n<\/code><\/pre>\n\n<p>please check these binaries are in the path <code>$HOME\/tensorflow\/bin\/activate<\/code>. If so, try<\/p>\n\n<pre><code>$ pip install -I --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n\n<p>where <code>-I<\/code> option forces to install packages.<\/p>\n"},{"AnswerId":"41659039","CreationDate":"2017-01-15T08:01:06.113","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>Are you running python in the same virtual environment you installed tensorflow in?<\/p>\n\n<p>To access your tensorflow installation, you have to first \"activate\" the virtualenv in any new terminals, as follows:<\/p>\n\n<pre><code>source ~\/tensorflow\/bin\/activate\npython\nimport tensorflow as tf\n<\/code><\/pre>\n\n<p>If you run the above in a new terminal, does it solve your problem?<\/p>\n"},{"AnswerId":"44646649","CreationDate":"2017-06-20T07:33:23.323","ParentId":null,"OwnerUserId":"6038077","Title":null,"Body":"<p>When you did\n<code>$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow_gpu-0.12.1-cp27-none-linux_x86_64.whl<\/code><\/p>\n\n<p>this step you are specifying that you are going to use Nvidia card.\nTo run tensorflow with GPU(Nvidia graphics card) you need to satisfy all Nvidia requirements<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#NVIDIARequirements\" rel=\"nofollow noreferrer\">Nvidia requires some special privileges to its CUDA cores<\/a><\/p>\n\n<p>You also need to check for Cuda pathnames to the LD_LIBRARY_PATH environment variable, check in <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow noreferrer\">Nvidia Documentation<\/a>.Also, you need to install an profiling support, this can be done by libcupti-dev library, which is the NVIDIA CUDA Profile Tools Interface. This library provides advanced profiling support. To install this library, issue the following command:\n<code>sudo apt-get install libcupti-dev<\/code><\/p>\n\n<p>But if you want to run tensorflow in CPU mode only, do not specify <code>$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow_gpu-0.12.1-cp27-none-linux_x86_64.whl<\/code>.With this you are overriding TF_BINARY_URL variable to use Nvidia CUDA core\nSo, to use CPU from all your steps remove <code>$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/gpu\/tensorflow_gpu-0.12.1-cp27-none-linux_x86_64.whl<\/code> and include only <code>$export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.12.1-cp27-none-linux_x86_64.whl<\/code> and reinstall<\/p>\n\n<p>I hope this should clear the problem<\/p>\n"}]}
{"QuestionId":41658885,"AnswerCount":3,"Tags":"<caffe><solver><gradient-descent>","CreationDate":"2017-01-15T07:38:42.383","AcceptedAnswerId":null,"OwnerUserId":5583416.0,"Title":"solver.prototxt for adam solver in caffe","Body":"<p>My solver.prototxt using adam is as follows. Do I need to add or remove any terms? The loss doesnt seem to reduce<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43544865","CreationDate":"2017-04-21T14:09:18.053","ParentId":null,"OwnerUserId":"6279632","Title":null,"Body":"<p>How about your work? If you used Adam. I suggest you look at the setting in caffe. I do not know why you have L2 and delta value. <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver_adam.prototxt\" rel=\"nofollow noreferrer\">This<\/a> is standard setting <\/p>\n\n<pre><code># The train\/test net protocol buffer definition\n# this follows \"ADAM: A METHOD FOR STOCHASTIC OPTIMIZATION\"\nnet: \"examples\/mnist\/lenet_train_test.prototxt\"\n# test_iter specifies how many forward passes the test should carry out.\n# In the case of MNIST, we have test batch size 100 and 100 test iterations,\n# covering the full 10,000 testing images.\ntest_iter: 100\n# Carry out testing every 500 training iterations.\ntest_interval: 500\n# All parameters are from the cited paper above\nbase_lr: 0.001\nmomentum: 0.9\nmomentum2: 0.999\n# since Adam dynamically changes the learning rate, we set the base learning\n# rate to a fixed value\nlr_policy: \"fixed\"\n# Display every 100 iterations\ndisplay: 100\n# The maximum number of iterations\nmax_iter: 10000\n# snapshot intermediate results\nsnapshot: 5000\nsnapshot_prefix: \"examples\/mnist\/lenet\"\n# solver mode: CPU or GPU\ntype: \"Adam\"\nsolver_mode: GPU\n<\/code><\/pre>\n"},{"AnswerId":"42958260","CreationDate":"2017-03-22T17:05:50.160","ParentId":null,"OwnerUserId":"7405440","Title":null,"Body":"<p>Try with a learning rate of 0.1 and a slower step size like 300 and watch the behavior, also check if the lmdb\/hdf5 file is well formed and have a correct scale to ease the learning, you can do this by generating the mean file over your dataset.<\/p>\n"},{"AnswerId":"42956162","CreationDate":"2017-03-22T15:31:10.780","ParentId":null,"OwnerUserId":"7709837","Title":null,"Body":"<p>By comparing with <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_solver_adam.prototxt\" rel=\"nofollow noreferrer\">caffe example on mnist<\/a>, 'stepsize' can be deleted since 'lr_policy' is 'fixed'.<\/p>\n"}]}
{"QuestionId":41659710,"AnswerCount":0,"Tags":"<machine-learning><lua><deep-learning><conv-neural-network><torch>","CreationDate":"2017-01-15T09:35:48.143","AcceptedAnswerId":null,"OwnerUserId":6102900.0,"Title":"error using CAddTable and ConcatTable in torch","Body":"<p>I tried to build a few residual blocks  in my  networks using torch, but why torch just gives me errors when using CAddTable and ConcatTable. Could anyone please give some advice? Thanks in advance. The code I use is here:<\/p>\n\n<pre><\/pre>\n\n<p>and the error reads like this:\n...torch\/install\/share\/lua\/5.1\/cudnn\/SpatialConvolution.lua:102: input has to contain: 3 feature maps, but received input of size: 1 x 16 x 60 x 60\nstack traceback:<\/p>\n","answers":[]}
{"QuestionId":41662727,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-15T15:17:18.427","AcceptedAnswerId":null,"OwnerUserId":3863945.0,"Title":"TypeError: DataType float32 for attr 'T' not in list of allowed values: int32, int64","Body":"<p>hallo i am trying to run the code using Anaconda but i got this error even i have already run the program using cpu in another pc , its worked fine. may be because using python v3.5 and in the other pc python v2.5 ?<\/p>\n\n<p>i created screenshot for the error <a href=\"https:\/\/i.stack.imgur.com\/8txul.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8txul.png\" alt=\"enter image description here\"><\/a>  <\/p>\n\n<p>the error line is that <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41663188,"AnswerCount":1,"Tags":"<c++><tensorflow><fft><eigen>","CreationDate":"2017-01-15T16:06:25.560","AcceptedAnswerId":41693967.0,"OwnerUserId":1150961.0,"Title":"Using eigen tensor fft","Body":"<p>I had a look at implementing a CPU fft for tensorflow using the eigen <a href=\"https:\/\/bitbucket.org\/eigen\/eigen\/src\/60f8b0f18340337e036b566bc583c77deb52f965\/unsupported\/Eigen\/CXX11\/src\/Tensor\/TensorFFT.h?at=default&amp;fileviewer=file-view-default\" rel=\"nofollow noreferrer\">tensor fft module<\/a>. Benoit Steiner offered some advice on how to use it <a href=\"https:\/\/listengine.tuxfamily.org\/lists.tuxfamily.org\/eigen\/2016\/02\/msg00071.html\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>The eigen fft uses the templated type of the corresponding tensor to perform its computation. Unfortunately, the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/op_kernel.h#L581\" rel=\"nofollow noreferrer\">inputs<\/a> extracted from the  are declared  because inputs are immutable. I thus end up with a bunch of compiler errors that const variables cannot be assigned to (because the  reuses the templated typename internally).<\/p>\n\n<p>What's the best way to get around this problem? Some thoughts:<\/p>\n\n<ul>\n<li>use a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/op_kernel.h#L614\" rel=\"nofollow noreferrer\"><\/a> but that seems like bending the API just to get around a problem<\/li>\n<li>modify the  with  such that non- variables are used internally<\/li>\n<li>explicitly instantiate the  rather than using the template member functions defined in <a href=\"https:\/\/bitbucket.org\/eigen\/eigen\/src\/60f8b0f18340337e036b566bc583c77deb52f965\/unsupported\/Eigen\/CXX11\/src\/Tensor\/TensorBase.h?at=default&amp;fileviewer=file-view-default#TensorBase.h-501\" rel=\"nofollow noreferrer\"><\/a><\/li>\n<\/ul>\n\n<p>The last option seems to be the most straightforward but I am struggling with getting all the template arguments right. Any suggestions?<\/p>\n","answers":[{"AnswerId":"41693967","CreationDate":"2017-01-17T09:55:54.343","ParentId":null,"OwnerUserId":"1150961","Title":null,"Body":"<p>With the help of a co worker, we managed to get the desired behaviour by simply casting the <code>const<\/code> input array to a non-<code>const<\/code> one.<\/p>\n\n<pre><code>typename TTypes&lt;T, 2&gt;::Tensor&amp; casted_input = *reinterpret_cast&lt;typename TTypes&lt;T, 2&gt;::Tensor*&gt;(&amp;input);\nauto result = casted_input.template fft&lt;Eigen::BothParts, Eigen::FFT_FORWARD&gt;(dims);\n<\/code><\/pre>\n"}]}
{"QuestionId":41664156,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><image-segmentation><pycaffe>","CreationDate":"2017-01-15T17:34:27.353","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"How can I save the probability of the prediction in caffe?","Body":"<p>Does anyone know how can I save the output of predicted class labels of each pixel in FCN semantic segmentation? I would like to see the probability map of the image during inference. The data in which layer should be saved?<\/p>\n\n<p>Many thanks <\/p>\n","answers":[{"AnswerId":"41665820","CreationDate":"2017-01-15T20:17:33.653","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>As  you can see from the  code in <a href=\"https:\/\/github.com\/shelhamer\/fcn.berkeleyvision.org\/blob\/master\/infer.py#L20\" rel=\"nofollow noreferrer\"><code>infer.py<\/code><\/a> the predicted label is <code>argmax<\/code> of <code>'score'<\/code> layer.<\/p>\n\n<pre><code>out = net.blobs['score'].data[0].argmax(axis=0)\n<\/code><\/pre>\n\n<p>The <code>'score'<\/code> is the <a href=\"https:\/\/github.com\/shelhamer\/fcn.berkeleyvision.org\/blob\/master\/voc-fcn32s\/train.prototxt#L520-L530\" rel=\"nofollow noreferrer\">input to <code>\"SoftmaxWithLoss\"<\/code> layer<\/a> during training.  Hence, to get class probabilities from <code>'score'<\/code> you need to add <code>\"Softmax\"<\/code> on top of <code>'score'<\/code>:<\/p>\n\n<pre><code>e_s = np.exp(net.blobs['score'].data[0])\nprob = e_s \/ e_s.sum(axis=0) \n<\/code><\/pre>\n"}]}
{"QuestionId":41664734,"AnswerCount":3,"Tags":"<tensorflow><mnist>","CreationDate":"2017-01-15T18:30:37.743","AcceptedAnswerId":41666715.0,"OwnerUserId":6691357.0,"Title":"Splitting MNIST data tensorflow","Body":"<p>I've been following the tensorflow tutorials. I've imported the MNIST dataset and ran the code for a 2 layer convolutional neural net. It took nearly 45 minutes to train. I want to cut down the training data by discarding some of the data. How do I do that?\nHere's the code:<\/p>\n\n<pre><\/pre>\n\n<p>I cut down the size of testing data since it's a numpy array. How do I do the same for training data?<\/p>\n","answers":[{"AnswerId":"41666715","CreationDate":"2017-01-15T21:50:48.397","ParentId":null,"OwnerUserId":"2616232","Title":null,"Body":"<p>You are using dataset provider defined in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py<\/a><\/p>\n\n<p>To reduce number of training sample, you can change this file (Line 237), or create a modified version and use it, instead of <\/p>\n\n<pre><code>from tensorflow.examples.tutorials.mnist import input_data\n<\/code><\/pre>\n\n<p>Which points to the link that I mentioned above.<\/p>\n"},{"AnswerId":"41666837","CreationDate":"2017-01-15T22:03:58.953","ParentId":null,"OwnerUserId":"7091875","Title":null,"Body":"<p>Cutting your training sample will not help you in any good way - as far as you use minibatches, it doesn't affect performance directly. As a better alternative, you can reduce the number of epochs and\/or increase the learning rate. \nReducing data sample is a very bad idea in this case<\/p>\n"},{"AnswerId":"42178914","CreationDate":"2017-02-11T17:26:38.947","ParentId":null,"OwnerUserId":"7550927","Title":null,"Body":"<p>just one question - we are talking about this code `<a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/<\/a> ???<\/p>\n\n<p>So if this needs 45 minutes, i guess you are running on cpu - you should think of using gpu. I tested the code in an Azure VM N Series with a Tesla K 80 GPU and it finishes in 4 Minutes<\/p>\n"}]}
{"QuestionId":41665109,"AnswerCount":1,"Tags":"<tensorflow><nlp><conv-neural-network>","CreationDate":"2017-01-15T19:02:37.153","AcceptedAnswerId":41987897.0,"OwnerUserId":826983.0,"Title":"Trying to understand CNNs for NLP tutorial using Tensorflow","Body":"<p>I am following <a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"nofollow noreferrer\">this tutorial<\/a> in order to understand CNNs in NLP. There are a few things which I don't understand despite having the code in front of me. I hope somebody can clear a few things up here.<\/p>\n\n<hr>\n\n<p>The first rather minor thing is the parameter of the  object. In the example on github this is just  which I think is the max-length of all sentences in the training data. This means that   is  a 56-dimensional vector which will contain just the indices from the dictionary of a sentence for each word. <\/p>\n\n<p>This list goes into  which will return a matrix consisting of the word embeddings of those words given by . According to <a href=\"https:\/\/stackoverflow.com\/a\/34877590\/826983\">this answer<\/a> this operation is similar to using indexing with numpy:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>But the problem here is that  most of the time looks like . So am I correct if I assume that  ignores the value 0?<\/p>\n\n<hr>\n\n<p>Another thing I don't get is how  is working here:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I assume, taht  is the matrix which is the <em>actual<\/em> input to the CNN where each row represents the word embedding of <em>one<\/em> word. But how can  know about those indices given by ? <\/p>\n\n<hr>\n\n<p>The last thing which I don't understand here is <\/p>\n\n<blockquote>\n  <p> is our embedding matrix that we learn during training. We initialize it using a random uniform distribution.  creates the actual embedding operation. The result of the embedding operation is a 3-dimensional tensor of shape .<\/p>\n<\/blockquote>\n\n<p>Does this mean that we are <em>actually<\/em> learning the word embeddings <em>here<\/em>? The tutorial states at the beginning:<\/p>\n\n<blockquote>\n  <p>We will not used pre-trained word2vec vectors for our word embeddings. Instead, we learn embeddings from scratch.<\/p>\n<\/blockquote>\n\n<p>But I don't see a line of code where this is actually happening. The <a href=\"https:\/\/github.com\/dennybritz\/cnn-text-classification-tf\/blob\/master\/text_cnn.py#L22\" rel=\"nofollow noreferrer\">code of the embedding layer<\/a> does not look like as if there is anything being trained or learned - so where is it happening?<\/p>\n","answers":[{"AnswerId":"41987897","CreationDate":"2017-02-01T19:19:49.373","ParentId":null,"OwnerUserId":"3898714","Title":null,"Body":"<p>Answer to ques 1 (So am I correct if I assume that tf.nn.embedding_lookup ignores the value 0?) : <\/p>\n\n<p>The 0's in the input vector is the index to 0th symbol in the vocabulary, which is the PAD symbol. I don't think it gets ignored when the lookup is performed. 0th row of the embedding matrix will be returned.<\/p>\n\n<p>Answer to ques 2 (But how can tf.nn.embedding_lookup know about those indices given by self.input_x?) : <\/p>\n\n<p>Size of the embedding matrix is [V * E] where is the size of vocabulary and E is dimension of embedding vector. 0th row of matrix is embedding vector for 0th element of vocabulary, 1st row of matrix is embedding vector for 1st element of vocabulary. \nFrom the input vector x, we get the indices of words in vocabulary, which are used for indexing the embedding matrix.<\/p>\n\n<p>Answer to ques 3 (Does this mean that we are actually learning the word embeddings here?). <\/p>\n\n<p>Yes, we are actually learning the embedding matrix. \nIn the embedding layer, in line <code>W = tf.Variable(       tf.random_uniform([vocab_size, embedding_size], -1.0, 1.0),name=\"W\")<\/code> W is the embedding matrix and by default, in tensorflow <code>trainable=TRUE<\/code> for variable. So, W will also be a learned parameter. To use pre- trained model, set <code>trainable = False<\/code>.<\/p>\n\n<p>For detailed explanation of the code you can follow blog: <a href=\"https:\/\/agarnitin86.github.io\/blog\/2016\/12\/23\/text-classification-cnn\" rel=\"nofollow noreferrer\">https:\/\/agarnitin86.github.io\/blog\/2016\/12\/23\/text-classification-cnn<\/a> <\/p>\n"}]}
{"QuestionId":41665321,"AnswerCount":0,"Tags":"<python-3.x><deep-learning><keras><autoencoder>","CreationDate":"2017-01-15T19:21:56.413","AcceptedAnswerId":null,"OwnerUserId":5083043.0,"Title":"high loss rate when training my model","Body":"<p>I'm working on image denoising using autoencoders (working with keras tensorflow backend).<\/p>\n\n<p>when i train my model the loss rate is pretty high and stable(somewhere around 2.x).<\/p>\n\n<p>i can't understand what i'm doing wrong.<\/p>\n\n<p>here's my code:<\/p>\n\n<pre><\/pre>\n\n<p>any suggestions?<\/p>\n","answers":[]}
{"QuestionId":41665390,"AnswerCount":3,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2017-01-15T19:31:01.660","AcceptedAnswerId":41707319.0,"OwnerUserId":590335.0,"Title":"Is it possible to visualize keras embeddings in tensorboard?","Body":"<p>keras has the ability to export some of it's training data in a tensorboard comaptible format using <a href=\"https:\/\/keras.io\/callbacks\/#tensorboard\" rel=\"noreferrer\">keras.callbacks.TensorBoard<\/a><\/p>\n\n<p>However, it doesn't support the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/embedding_viz\/\" rel=\"noreferrer\">embedding visualisation<\/a> in tensorboard.<\/p>\n\n<p>Is there a way around this?<\/p>\n","answers":[{"AnswerId":"57461479","CreationDate":"2019-08-12T12:44:32.713","ParentId":null,"OwnerUserId":"857731","Title":null,"Body":"<p>This is now possible directly with the <a href=\"https:\/\/keras.io\/callbacks\/#tensorboard\" rel=\"nofollow noreferrer\"><code>keras.callbacks.TensorBoard<\/code><\/a> callback:<\/p>\n\n<pre><code>from keras import callbacks\n\nmodel.fit(x_train, y_train,\n        batch_size=batch_size,\n        epochs=10,\n        callbacks=[\n                   callbacks.TensorBoard(batch_size=batch_size,\n                                         embeddings_freq=3,  # Store embeddings every 3 epochs (this can be time consuming)\n                                         embeddings_layer_names=['fc1', 'fc2'],  # Embeddings are taken from layers with names fc1 and fc2\n                                         embeddings_metadata='metadata.tsv',  # This file will describe the embeddings data (see below)\n                                         embeddings_data=x_test),  # Data used for the embeddings\n                   ],\n        )\n\n\n# Use this metadata.tsv file before you have a trained model:\nwith open(\"metadata.tsv\", 'w') as f:\n    f.write(\"label\\tidx\\n\")\n    f.write('\\n'.join([\"{}\\t{}\".format(class_names[int(y.argmax())], i)\n                       for i, y in enumerate(y_test)]))\n\n\n# After the model is trained, you can update the metadata file to include more information, such as the predicted labels and the mistakes:\ny_pred = model.predict(x_test)\nwith open(\"metadata.tsv\", 'w') as f:\n    f.write(\"label\\tidx\\tpredicted\\tcorrect\\n\")\n    f.write('\\n'.join([\"{}\\t{}\\t{}\\t{}\".format(class_names[int(y.argmax())],\n                                               i,\n                                               class_names[int(y_pred[i].argmax())],\n                                               class_names[int(y.argmax())]==class_names[int(y_pred[i].argmax())])\n                       for i, y in enumerate(y_test)]))\n<\/code><\/pre>\n\n<p><strong>Note:<\/strong> Tensorboard will usually look for your <code>metadata.tsv<\/code> in the <code>logs<\/code> directory. If it doesn't find it, it will tell you at which path it was looking and you can copy it there and refresh tensorboard.<\/p>\n"},{"AnswerId":"42370427","CreationDate":"2017-02-21T14:43:36.553","ParentId":null,"OwnerUserId":"892001","Title":null,"Body":"<p>There is pull request with this functionality - <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/5247\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/pull\/5247<\/a> callback is extended to create visualization for specific embedding layers.<\/p>\n"},{"AnswerId":"41707319","CreationDate":"2017-01-17T21:28:06.487","ParentId":null,"OwnerUserId":"590335","Title":null,"Body":"<p>Found a solution:<\/p>\n\n<pre><code>import os\n\nimport keras\nimport tensorflow\n\nROOT_DIR = '\/tmp\/tfboard'\n\nos.makedirs(ROOT_DIR, exist_ok=True)\n\n\nOUTPUT_MODEL_FILE_NAME = os.path.join(ROOT_DIR,'tf.ckpt')\n\n# get the keras model\nmodel = get_model()\n# get the tensor name from the embedding layer\ntensor_name = next(filter(lambda x: x.name == 'embedding', model.layers)).W.name\n\n# the vocabulary\nmetadata_file_name = os.path.join(ROOT_DIR,tensor_name)\n\nembedding_df = get_embedding()\nembedding_df.to_csv(metadata_file_name, header=False, columns=[])\n\nsaver = tensorflow.train.Saver()\nsaver.save(keras.backend.get_session(), OUTPUT_MODEL_FILE_NAME)\n\nsummary_writer = tensorflow.train.SummaryWriter(ROOT_DIR)\n\nconfig = tensorflow.contrib.tensorboard.plugins.projector.ProjectorConfig()\nembedding = config.embeddings.add()\nembedding.tensor_name = tensor_name\nembedding.metadata_path = metadata_file_name\ntensorflow.contrib.tensorboard.plugins.projector.visualize_embeddings(summary_writer, config)\n<\/code><\/pre>\n"}]}
{"QuestionId":41665799,"AnswerCount":7,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-01-15T20:14:32.083","AcceptedAnswerId":45546663.0,"OwnerUserId":3743194.0,"Title":"Keras model.summary() object to string","Body":"<p>I want to write a *.txt file with the neural network hyperparameters and the model architecture. Is it possible to write the object model.summary() to my output file?<\/p>\n\n<pre><\/pre>\n\n<p>It happens that I'm getting \"None\" as you can see below.<\/p>\n\n<pre><\/pre>\n\n<p>Any idea how to deal with that?<\/p>\n","answers":[{"AnswerId":"43033472","CreationDate":"2017-03-26T19:27:33.053","ParentId":null,"OwnerUserId":"2915983","Title":null,"Body":"<p>It's not the best way to do it but one thing you can do is to redirect stdout:<\/p>\n\n<pre><code>orig_stdout = sys.stdout\nf = open('out.txt', 'w')\nsys.stdout = f\nprint(model.summary())\nsys.stdout = orig_stdout\nf.close()\n<\/code><\/pre>\n\n<p>see <a href=\"https:\/\/stackoverflow.com\/a\/7152903\/2915983\">\"How to redirect 'print' output to a file using python?\"<\/a><\/p>\n"},{"AnswerId":"41670607","CreationDate":"2017-01-16T06:27:51.293","ParentId":null,"OwnerUserId":"6769011","Title":null,"Body":"<p>I stumbled upon the same problem as well!\nThere are two possible workarounds:<\/p>\n\n<p>Use <code>to_json()<\/code> method of the model  <\/p>\n\n<pre><code>summary = str(model.to_json()) \n<\/code><\/pre>\n\n<p>This is your case above.<\/p>\n\n<p>Otherwise use the ascii method from keras_diagram<\/p>\n\n<pre><code>from keras_diagram import ascii\nsummary = ascii(model)\n<\/code><\/pre>\n"},{"AnswerId":"53937848","CreationDate":"2018-12-26T22:33:10.720","ParentId":null,"OwnerUserId":"8378726","Title":null,"Body":"<p>I understand the OP has already accepted winni2k's answer, but since the question title actually implies saving the outputs of <code>model.summary()<\/code> to a <em>string<\/em>, not a file, the following code might help others who come to this page looking for that (like I did).<\/p>\n\n<p>The code below was run using TensorFlow <code>1.12.0<\/code>, which comes with Keras <code>2.1.6-tf<\/code> on Python <code>3.6.2<\/code>.<\/p>\n\n<pre><code>from tensorflow.keras.models import Sequential\nfrom tensorflow.keras.layers import Dense, Activation\nimport io\n\n# Example model\nmodel = Sequential([\n    Dense(32, input_shape=(784,)),\n    Activation('relu'),\n    Dense(10),\n    Activation('softmax'),\n])\n\ndef get_model_summary(model):\n    stream = io.StringIO()\n    model.summary(print_fn=lambda x: stream.write(x + '\\n'))\n    summary_string = stream.getvalue()\n    stream.close()\n    return summary_string\n\nmodel_summary_string = get_model_summary(model)\n\nprint(model_summary_string)\n<\/code><\/pre>\n\n<p>Which yields (as a string):<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\ndense (Dense)                (None, 32)                25120     \n_________________________________________________________________\nactivation (Activation)      (None, 32)                0         \n_________________________________________________________________\ndense_1 (Dense)              (None, 10)                330       \n_________________________________________________________________\nactivation_1 (Activation)    (None, 10)                0         \n=================================================================\nTotal params: 25,450\nTrainable params: 25,450\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n"},{"AnswerId":"53668338","CreationDate":"2018-12-07T11:10:28.117","ParentId":null,"OwnerUserId":"10759429","Title":null,"Body":"<p>For me, this worked to just get the model summary as a string:<\/p>\n\n<pre><code>stringlist = []\nmodel.summary(print_fn=lambda x: stringlist.append(x))\nshort_model_summary = \"\\n\".join(stringlist)\nprint(short_model_summary)\n<\/code><\/pre>\n"},{"AnswerId":"44893061","CreationDate":"2017-07-03T19:55:11.553","ParentId":null,"OwnerUserId":"3345375","Title":null,"Body":"<p>One option, although not an exact replacement for model.summary, is to export a model's configuration using <code>model.get_config()<\/code>. From <a href=\"https:\/\/keras.io\/models\/about-keras-models\/\" rel=\"noreferrer\">the docs<\/a>:<\/p>\n\n<blockquote>\n  <p><code>model.get_config()<\/code>: returns a dictionary containing the configuration of the model. The model can be reinstantiated from its config via:<\/p>\n\n<pre><code>config = model.get_config()\nmodel = Model.from_config(config)\n# or, for Sequential:\nmodel = Sequential.from_config(config)\n<\/code><\/pre>\n<\/blockquote>\n"},{"AnswerId":"50077092","CreationDate":"2018-04-28T13:32:22.030","ParentId":null,"OwnerUserId":"3317059","Title":null,"Body":"<p>And if you want to write to a log:<\/p>\n\n<pre><code>import logging\nlogger = logging.getLogger(__name__)\n\nmodel.summary(print_fn=logger.info)\n<\/code><\/pre>\n"},{"AnswerId":"45546663","CreationDate":"2017-08-07T12:14:27.197","ParentId":null,"OwnerUserId":"528691","Title":null,"Body":"<p>With my version of Keras (<code>2.0.6<\/code>) and Python (<code>3.5.0<\/code>), this works for me:<\/p>\n\n<pre><code># Create an empty model\nfrom keras.models import Sequential\nmodel = Sequential()\n\n# Open the file\nwith open(filename + 'report.txt','w') as fh:\n    # Pass the file handle in as a lambda function to make it callable\n    model.summary(print_fn=lambda x: fh.write(x + '\\n'))\n<\/code><\/pre>\n\n<p>This outputs the following lines to the file:<\/p>\n\n<pre><code>_________________________________________________________________\nLayer (type)                 Output Shape              Param #   \n=================================================================\nTotal params: 0\nTrainable params: 0\nNon-trainable params: 0\n_________________________________________________________________\n<\/code><\/pre>\n"}]}
{"QuestionId":41666964,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-15T22:21:36.620","AcceptedAnswerId":null,"OwnerUserId":3873000.0,"Title":"model variables in Tensorflow's batch_norm","Body":"<p>The documentation online says moving_average and moving_variance are both model_variables, and tf.model_variables() returns tensors of the type local_variables. Does that mean model_variables are not saved when I save my state?<\/p>\n\n<p>I'm trying to apply batch normalization to a couple of 3D convolution and fully connected layers. I trained my network with batch_norm and saved a checkpoint file, but when I went to restore my saved state, it said moving_mean could not be located. The exact error was, when TF went to assign the restored value to moving_mean, the shape of the lhs tensor, [], could not be reconciled with the that of the rhs, [20].<\/p>\n\n<p>The graph restores fine when I don't add batch_norm around my layers. \nI'm planning to add a global variable at the end of training that saves my moving_mean and moving_variance values. Is this the way TF intended for me to use batch_norm?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41841207","CreationDate":"2017-01-25T00:08:19.217","ParentId":null,"OwnerUserId":"3873000","Title":null,"Body":"<p>The variables moving_mean and moving_variance were not in my saved stated because I had set updates_collections to its default. Since I never included a control dependency when I ran the layers, those variables never were updated.<\/p>\n\n<p>The code to include is:<\/p>\n\n<pre><code>from tensorflow.python import control_flow_ops\n\nupdate_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\nif update_ops:\n    updates = tf.tuple(update_ops)\n    total_loss = control_flow_ops.with_dependencies(updates, total_loss)\n<\/code><\/pre>\n\n<p>Or set <\/p>\n\n<pre><code>updates_collection=None \n<\/code><\/pre>\n\n<p>to for in-place updates.<\/p>\n\n<p>See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard4\/tf.contrib.layers.batch_norm.md\" rel=\"nofollow noreferrer\">the API description<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4361\" rel=\"nofollow noreferrer\">current github discussion<\/a> for more information. <\/p>\n"}]}
{"QuestionId":41667183,"AnswerCount":2,"Tags":"<python><neural-network><theano><feature-extraction><lasagne>","CreationDate":"2017-01-15T22:47:00.240","AcceptedAnswerId":null,"OwnerUserId":5125320.0,"Title":"Feature Extraction on Neural Networks Using Theano","Body":"<p>I have a trained network, which consists of the following layers:\n {conv1, pool1,  conv2, pool2, conv3, pool3, conv4, pool4, fc5, fc6, output}\nwhich fc means fully connected layers and conv means convolutional layers.<\/p>\n\n<p>I need to do feature extraction for some images. I am using Lasagne and Theano.\nI need to save features from each layer for later analysis. I am a newbie in this language so I tried to find sample codes or some tutorials on this (with theano\/lasagne). However, I failed to understand what should I do by myself. <\/p>\n\n<p>I would appreciate if someone can guide me on this in order to how to implement feature extraction.<\/p>\n\n<p>Thank you in advance<\/p>\n\n<p>Edit: I followed comments by Mr\/Ms gntoni, here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>For my case, I need to save features from each layer. I want to write a function like the one that we have in Caffe:<\/p>\n\n<pre><\/pre>\n\n<p>However, my trained model is written with lasagne and theano, So I have to implement this in lasagne format.<\/p>\n\n<p>After writing the code above (in lasagne), I am getting an empty output.\nI wonder why and how can I fix it.<\/p>\n\n<p>Thank you in advance<\/p>\n","answers":[{"AnswerId":"42078074","CreationDate":"2017-02-06T22:02:40.020","ParentId":null,"OwnerUserId":"1104911","Title":null,"Body":"<p>Keep in mind that in Lasagne the <code>get_output<\/code> method returns Theano tensors, and you cannot directly use them to compute the features from a <code>numpy<\/code> array. However, you can define a Theano <em>function<\/em> and use it to compute the values. In your case:<\/p>\n\n<pre><code>layers = [self.acnn.cnn[layer] for layer in layer_list]\nfeat_fn = theano.function([input_var], lasagne.layers.get_output(layers),\n                          deterministic=True)\n<\/code><\/pre>\n\n<p>where <code>input_var<\/code> is the input tensor to your network. The <code>get_output<\/code> method can accept multiple layers and Theano functions can have multiple outputs, so you can define a single function to extract all the features. Getting the numerical values is then as simple as:<\/p>\n\n<pre><code>feat_all = feat_fn(img)\n<\/code><\/pre>\n"},{"AnswerId":"41668496","CreationDate":"2017-01-16T02:06:30.053","ParentId":null,"OwnerUserId":"3737792","Title":null,"Body":"<p>A Convolutional Neural Network, like yours, consists in two parts:<\/p>\n\n<p>The first one is the feature extraction part and in your case consists in the conv-pool layers {conv1, pool1, conv2, pool2, conv3, pool3, conv4, pool4}.<\/p>\n\n<p>The second is the classification part. In your network: {fc5, fc6, output}.<\/p>\n\n<p>When training, the first part is trying to obtain the best representation of the input data to be classified by the second part.<\/p>\n\n<p>So, if after trained, you disconnect this two parts, <strong>the output of conv4<\/strong> layer will be giving you the features you want. <\/p>\n\n<p>This features can be used with a different classifier. In fact, many people use an already trained network (e.g. AlexNet), remove the last classification layers, and use the features with their own classification system.<\/p>\n"}]}
{"QuestionId":41668158,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow><deep-learning><data-science>","CreationDate":"2017-01-16T01:09:10.693","AcceptedAnswerId":null,"OwnerUserId":3993270.0,"Title":"TensorFlow: how to determine if we want to break the training dataset into batches","Body":"<p>I am learning TensorFlow (as well as general deep learning). I am wondering when do we need to break the input training data into batches? And how do we determine the batch size? Is there a rule of thumb? Thanks!<\/p>\n","answers":[{"AnswerId":"41763164","CreationDate":"2017-01-20T11:58:02.110","ParentId":null,"OwnerUserId":"2606070","Title":null,"Body":"<p>Generally Deep Learning algorithms are ran on GPUs which has limited memory and thus a limited number of input data samples (in the algorithm commonly defined as batch size) could be loaded at a time.<\/p>\n\n<p>In general larger batch size reduces the overall computation time (as the internal matrix multiplications are done in a parallel manner in GPU, thus with large batch sizes the time gets saved in reading\/writing gradients and possibly some other operations output).<\/p>\n\n<p>Another probable benefit of large batch size is: \nIn multi-class classification problems, if the number of classes are large, a \nlarger batch size makes algorithm generalize better(technically avoids over-fitting) over the different classes (while doing this a standard technique is to keep uniform distribution of classes in a batch).<\/p>\n\n<p>While deciding batch size there are some other factors which comes into play are: learning rate and the type of Optimization method.<\/p>\n\n<p>I hope this answers your question to certain extent!<\/p>\n"}]}
{"QuestionId":41668502,"AnswerCount":2,"Tags":"<php><python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-01-16T02:07:23.747","AcceptedAnswerId":null,"OwnerUserId":1552048.0,"Title":"Using tensorflow models in web applications","Body":"<p>So I've recently became very interested in machine learning, and have been using  (python) in some of my projects at work.<\/p>\n\n<p>However, I've now found a use for digit classification in one of my web projects which is all written in PHP for server-side code. Ideally, I'd like to be able to upload a picture via a Javascript interface, and on the PHP receiving end process the picture, slide a window to find the set of digits I'm looking for, and then process this through my trained neural network. I guess I'll summarise in one question: <strong><em>is this possible?<\/em><\/strong> Thank you!<\/p>\n","answers":[{"AnswerId":"41670255","CreationDate":"2017-01-16T05:56:30.137","ParentId":null,"OwnerUserId":"847877","Title":null,"Body":"<p>Tensorflow provides <a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"noreferrer\">serving project<\/a> for communication, so you need to implement gRPC client for PHP, then use it to talk to Tensorflow serving environment.<\/p>\n\n<p>That's how I work with Ruby and Tensorflow, it should be working for you as well!<\/p>\n"},{"AnswerId":"46737248","CreationDate":"2017-10-13T20:01:29.440","ParentId":null,"OwnerUserId":"1795763","Title":null,"Body":"<p>You can try one of these PHP extensions:<\/p>\n\n<p><a href=\"https:\/\/github.com\/absalomedia\/tensile\" rel=\"nofollow noreferrer\">https:\/\/github.com\/absalomedia\/tensile<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/wan2land\/php-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/wan2land\/php-tensorflow<\/a><\/p>\n"}]}
{"QuestionId":41668620,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-16T02:27:00.057","AcceptedAnswerId":41677676.0,"OwnerUserId":7423371.0,"Title":"sparse matrix diagonal multiplication using tensorflow","Body":"<p>I have a symmetric sparse matrix M and a diagonal matrix D. I want to compute DMD<\/p>\n\n<p>M is created as a SparseTensor, and D can be represented either as tf.diag or a SparseTensor - in theory it should be entirely possible to evaluate DMD without going through a dense matrix, but I can't figure out how - does anyone have any tips?<\/p>\n","answers":[{"AnswerId":"41677676","CreationDate":"2017-01-16T13:38:01.367","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>Here is some tips. <\/p>\n\n<p>Note that the effect of multiplying D before and after M is to multiply each non-zero element M(i, j) with D(i, i) and D(j, j). See <a href=\"http:\/\/www.solitaryroad.com\/c108.html\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Say P = DMD. To construct P as a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/sparse_ops\/sparse_tensor_representation#SparseTensor\" rel=\"nofollow noreferrer\">sparse tensor<\/a>, you will need indices, value, and shape. You can set indices and shape as M's indices and shape, and compute value with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/functional_ops\/higher_order_operators#map_fn\" rel=\"nofollow noreferrer\">tf.map_fn<\/a> over M's value and D (D could be a dense 1D Tensor).<\/p>\n"}]}
{"QuestionId":41668786,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-01-16T02:53:15.610","AcceptedAnswerId":41861465.0,"OwnerUserId":6433136.0,"Title":"How do you create a dynamic_rnn with dynamic \"zero_state\" (Fails with Inference)","Body":"<p>I have been working with the \"dynamic_rnn\" to create a model.<\/p>\n\n<p>The model is based upon a 80 time period signal, and I want to zero the \"initial_state\" before each run so I have setup the following code fragment to accomplish this:<\/p>\n\n<pre><\/pre>\n\n<p>This works great for the training process.  The problem is once I go to the inference, where my BatchSize = 1, I get an error back as the rnn \"state\" doesn't match the new Xinputs shape.  So what I figured is I need to make \"self.BatchSize\" based upon the input batch size rather than hard code it.  I tried many different approaches, and none of them have worked.  I would rather not pass a bunch of zeros through the feed_dict as it is a constant based upon the batch size.<\/p>\n\n<p>Here are some of my attempts.  They all generally fail since the input size is unknown upon building the graph:<\/p>\n\n<pre><\/pre>\n\n<p>.....<\/p>\n\n<pre><\/pre>\n\n<p>Another approach, thinking the initializer might not get called until run-time, but still failed at graph build:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to get this constant initial state to be created dynamically or do I need to reset it through the feed_dict with tensor-serving code?  Is there a clever way to do this only once within the graph maybe with an tf.Variable.assign?<\/p>\n","answers":[{"AnswerId":"41861465","CreationDate":"2017-01-25T20:53:39.867","ParentId":null,"OwnerUserId":"6433136","Title":null,"Body":"<p>The solution to the problem was how to obtain the \"batch_size\" such that the variable is not hard coded.<\/p>\n\n<p>This was the correct approach from the given example:<\/p>\n\n<pre><code>Xinputs = tf.placeholder(tf.int32, (None, self.sequence_size, self.num_params), name=\"input\")\nstate = cell_L1.zero_state(Xinputs.get_shape()[0],Xinputs.dtype)\n<\/code><\/pre>\n\n<p>The problem is the use of \"get_shape()[0]\", this returns the \"shape\" of the tensor and takes the batch_size value at [0].  The documentation doesn't seem to be that clear, but this appears to be a constant value so when you load the graph into an inference, this value is still hard coded (maybe only evaluated at graph creation?).<\/p>\n\n<p>Using the \"tf.shape()\" function, seems to do the trick.  This doesn't return the shape, but a tensor.  So this seems to be updated more at run-time.  Using this  code fragment solved the problem of a training batch of 128 and then loading the graph into TensorFlow-Service inference handling a batch of just 1.<\/p>\n\n<pre><code>Xinputs = tf.placeholder(tf.int32, (None, self.sequence_size, self.num_params), name=\"input\")\nbatch_size = tf.shape(Xinputs)[0]\nstate = self.cell_L1.zero_state(batch_size,Xinputs.dtype)\n<\/code><\/pre>\n\n<p>Here is a good link to TensorFlow FAQ which describes this approach '<strong>How do I build a graph that works with variable batch sizes?<\/strong>':\n<a href=\"https:\/\/www.tensorflow.org\/resources\/faq\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/resources\/faq<\/a><\/p>\n"}]}
{"QuestionId":41668813,"AnswerCount":2,"Tags":"<python><theano><keras><keras-layer>","CreationDate":"2017-01-16T02:56:46.877","AcceptedAnswerId":41670915.0,"OwnerUserId":996366.0,"Title":"How to add and remove new layers in keras after loading weights?","Body":"<p>I am trying to do a transfer learning; for that purpose I want to remove the last two layers of the neural network and add another two layers. This is an example code which also output the same error.<\/p>\n\n<pre><\/pre>\n\n<p>I removed the layer using  but when I tried to add its outputting this error <\/p>\n\n<blockquote>\n  <p>AttributeError: 'Model' object has no attribute 'add'<\/p>\n<\/blockquote>\n\n<p>I know the most probable reason for the error is improper use of . what other syntax should I use?<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I tried to remove\/add layers in keras but its not  allowing it to be added after loading external weights.<\/p>\n\n<pre><\/pre>\n\n<p>its showing this error<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"49689175","CreationDate":"2018-04-06T09:03:57.380","ParentId":null,"OwnerUserId":"3127650","Title":null,"Body":"<p>Another way to do it<\/p>\n\n<pre><code>from keras.models import Model\n\nlayer_name = 'relu_conv2'\nmodel2= Model(inputs=model1.input, outputs=model1.get_layer(layer_name).output)\n<\/code><\/pre>\n"},{"AnswerId":"41670915","CreationDate":"2017-01-16T06:53:28.697","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can take the <code>output<\/code> of the last model and create a new model. The lower layers remains the same.<\/p>\n\n<pre><code>model.summary()\nmodel.layers.pop()\nmodel.layers.pop()\nmodel.summary()\n\nx = MaxPooling2D()(model.layers[-1].output)\no = Activation('sigmoid', name='loss')(x)\n\nmodel2 = Model(input=in_img, output=[o])\nmodel2.summary()\n<\/code><\/pre>\n\n<p>Check <a href=\"https:\/\/stackoverflow.com\/questions\/41378461\/how-to-use-models-from-keras-applications-for-transfer-learnig\/41386444#41386444\">How to use models from keras.applications for transfer learnig?<\/a> <\/p>\n\n<p><strong>Update on Edit:<\/strong><\/p>\n\n<p>The new error is because you are trying to create the new model on global <code>in_img<\/code> which is actually not used in the previous model creation.. there you are actually defining a local <code>in_img<\/code>. So the global <code>in_img<\/code> is obviously not connected to the upper layers in the symbolic graph. And it has nothing to do with loading weights.<\/p>\n\n<p>To better resolve this problem you should instead use <code>model.input<\/code> to reference to the input.<\/p>\n\n<p><code>model3 = Model(input=model2.input, output=[o])<\/code><\/p>\n"}]}
{"QuestionId":41669004,"AnswerCount":1,"Tags":"<image><resize><tensorflow>","CreationDate":"2017-01-16T03:24:20.527","AcceptedAnswerId":null,"OwnerUserId":3623196.0,"Title":"pixel value larger than 255 after using bicubic resizing (tf.image.resize_images)","Body":"<p>I first downsample an image by factor of 2 using tf.image.resize_images and then upsample it by factor 2 again using tf.image.resize_images, for these two operations, I used bicubic as the kernel, but the result image has some pixel values larger than 255 or smaller than 0, when I switched to using bilinear kernel, the result images is good. Any one knows the reason?  Thank you very much.<\/p>\n","answers":[{"AnswerId":"41709175","CreationDate":"2017-01-17T23:56:26.180","ParentId":null,"OwnerUserId":"1201422","Title":null,"Body":"<p>There's an explanation for this issue here: <a href=\"https:\/\/stackoverflow.com\/questions\/37032251\/tensorflow-image-resize-mess-up-image-on-unknown-image-size#39339482\">tensorflow image_resize mess up image on unknown image size<\/a><\/p>\n\n<p>You can use tf.saturate_cast(image, dtype=tf.uint8) on the float32 output of resize_images to get pixels between 0 and 255.<\/p>\n"}]}
{"QuestionId":41669170,"AnswerCount":1,"Tags":"<tensorflow><skflow>","CreationDate":"2017-01-16T03:49:51.873","AcceptedAnswerId":null,"OwnerUserId":173355.0,"Title":"How to make contrib.learn (SKFLOW) regression predict multiple values (TensorFlow 0.12.1)","Body":"<p>This is very similar to the question <a href=\"https:\/\/stackoverflow.com\/questions\/34224826\/skflow-regression-predict-multiple-values\">skflow regression predict multiple values<\/a>.  However, later versions of TensorFlow seem to rendered the answer from this question obsolete. <\/p>\n\n<p>I would like to be able to have multiple output neurons in a TensorFlow Learn regression neural network (<strong>DNNRegressor<\/strong>).  I upgraded the code from the referenced question to account for breaking changes in TensorFlow, but still get an error. <\/p>\n\n<pre><\/pre>\n\n<p>But this results in:<\/p>\n\n<pre><\/pre>\n\n<p>I don't seen any release notes about breaking changes that indicate that the method for multiple outputs have changed. Is there another way to do this?<\/p>\n","answers":[{"AnswerId":"47530796","CreationDate":"2017-11-28T11:45:39.147","ParentId":null,"OwnerUserId":"3706049","Title":null,"Body":"<p>As mentioned in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/learn\/DNNRegressor\" rel=\"nofollow noreferrer\">tf.contrib.learn.DNNRegressor<\/a> docs, you may use the  <code>label_dimension<\/code> parameter, which is exactly what you are looking for.<\/p>\n\n<p>Your code line with this param will do what you want:<\/p>\n\n<pre><code>regressor = skflow.DNNRegressor(hidden_units=[5, 5],\n                                feature_columns=feature_columns,\n                                label_dimension=2)\n<\/code><\/pre>\n\n<p>The standard <code>predict()<\/code> returns an generator object. To get an array, you have to add <code>as_iterable=False<\/code>:<\/p>\n\n<pre><code>score = metrics.mean_squared_error(regressor.predict(X, as_iterable=False), y)\n<\/code><\/pre>\n"}]}
{"QuestionId":41669233,"AnswerCount":2,"Tags":"<deep-learning><keras>","CreationDate":"2017-01-16T03:59:06.340","AcceptedAnswerId":null,"OwnerUserId":5046896.0,"Title":"Keras, How to define padding in Convolution2D?","Body":"<p>In the <a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"noreferrer\">Convolution2D docs of keras<\/a>, I found there are no details on definition of <\/p>\n\n<pre><\/pre>\n\n<p>the  argument is <\/p>\n\n<blockquote>\n  <p>tuple of length 2. Factor by which to subsample output. Also called strides elsewhere.<\/p>\n<\/blockquote>\n\n<p>and I think it is stride.<\/p>\n\n<p>And the  argument  is <\/p>\n\n<blockquote>\n  <p>'valid', 'same' or 'full'. ('full' requires the Theano backend.)<\/p>\n<\/blockquote>\n\n<p> and  are also arguments in TensorFlow's conv2d function.<\/p>\n\n<p>How to define the padding, how to set the value of it?<\/p>\n","answers":[{"AnswerId":"41672108","CreationDate":"2017-01-16T08:24:54.133","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>What you want is the <a href=\"https:\/\/keras.io\/layers\/convolutional\/#zeropadding2d\" rel=\"nofollow noreferrer\">ZeroPadding2D<\/a> layer, just put it before a convolutional layer. This is more flexible than just putting this functionality inside Convolution2D.<\/p>\n"},{"AnswerId":"56762922","CreationDate":"2019-06-25T22:43:32.020","ParentId":null,"OwnerUserId":"1651560","Title":null,"Body":"<p>The accepted answer above mentioned the ZeroPadding2D layer in Keras. However, if you want to use some other kinds of padding which are not provided by Keras, such as reflection padding, you should implement them by yourself.<\/p>\n\n<p>As far as I know, there are two ways to custom padding layer in Keras:\n1. Use <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\">Lambda<\/a> layer in Keras, and call the <a href=\"https:\/\/www.tensorflow.org\/versions\/r2.0\/api_docs\/python\/tf\/pad\" rel=\"nofollow noreferrer\">padding<\/a> function from Keras's backend(typically TensorFlow).\n2. Define your own Layer class. See <a href=\"https:\/\/stackoverflow.com\/questions\/50677544\/reflection-padding-conv2d\">this question<\/a> for more details.<\/p>\n"}]}
{"QuestionId":41669489,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2017-01-16T04:32:59.380","AcceptedAnswerId":41670407.0,"OwnerUserId":5635517.0,"Title":"Dropout Implementation: Do all images in a batch get the same mask in Caffe?","Body":"<p>I just wanted to confirm whether all the images in a mini-batch get the same dropout mask or not.<\/p>\n\n<p>For clarification:  Suppose a 1000 sized vector is passed through the Dropout layer with a mini batch size of 100. \nNow, the 21st and 31st elements of the 1st image's vector gets dropped out.\nIs it necessary that the 21st and 31st elements will get dropped for all the remaining 99 images in the batch? Or is it the case that every image gets a separate mask?<\/p>\n","answers":[{"AnswerId":"41670407","CreationDate":"2017-01-16T06:11:09.483","ParentId":null,"OwnerUserId":"4221706","Title":null,"Body":"<p>No, each image in a batch gets an independent, completely random mask.<\/p>\n\n<p>Actually, the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/layers\/dropout_layer.cpp\" rel=\"nofollow noreferrer\">Dropout layer<\/a> doesn't even care about the shape of its input: It counts the number of elements in the bottom Blob (i.e. for a batch size of 10 images, 3 channels and 224x224, it would be <code>10 * 3 * 224 * 224 = 1505280<\/code>), and generates the same number of independent random numbers.<\/p>\n"}]}
{"QuestionId":41670022,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><pycaffe><deeplearning4j>","CreationDate":"2017-01-16T05:31:44.437","AcceptedAnswerId":41670660.0,"OwnerUserId":6494707.0,"Title":"Is there any example of using weighted loss for pixel-wise segmentation\/classification tasks?","Body":"<p>I am doing FCN32 semantic segmentation on my data. I ran the algorithm to fine-tune for my data (grayscale images with only one channel), till 80,000 iterations; however, the loss and accuracy are fluctuating and the output image completely black. Even, the loss is so high after 80,000 iterations. I thought the classifier cannot do training well on my data. So, I am going to train from scratch. \nOn the other hand, my data has imbalanced class members. The background pixels are more than the other four classes. Some researchers are suggesting using weighted loss. Does anyone have any idea? Am I doing the right way? How can I add this weighted loss to train_val.prototxt? <\/p>\n\n<p>I will be thankful if you know any resources\/examples related to training with weighted loss, please share with me here.<\/p>\n\n<p>Thanks again<\/p>\n","answers":[{"AnswerId":"41670660","CreationDate":"2017-01-16T06:33:50.290","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can tackle class imbalance using <a href=\"https:\/\/stackoverflow.com\/a\/30497907\/1714410\"><code>\"InfogainLoss\"<\/code><\/a>. This loss can be viewed as an extension to <code>\"SoftmaxWithLoss\"<\/code> that enables you to \"pay\" different loss value per label.<br>\nIf you want to use <code>\"InfogainLoss\"<\/code> for pixel-wise predictions, you might need to use <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3855\" rel=\"nofollow noreferrer\">BVLC\/caffe PR#3855<\/a>.<\/p>\n"}]}
{"QuestionId":41670665,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-16T06:34:07.133","AcceptedAnswerId":null,"OwnerUserId":847877.0,"Title":"Batch norm causes Inception model overfitting?","Body":"<p>I was posting this as a issue in <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/890\" rel=\"nofollow noreferrer\">tensorflow\/models<\/a> repo.<\/p>\n\n<p>I was training <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/nets\/inception_v4.py\" rel=\"nofollow noreferrer\">inception model<\/a> from scratch with my own dataset which contains two classes only, and I noticed that when I start to use batch normalization with use_batch_norm parameter, the validation accuracy was around 50% ~ 60% while training accuracy went to nearly 100%, this bad accuray was almost just guessing. It looked like overfitting, and when I turned off use_batch_norm, the validation accuracy went to nearly 100% soon.<\/p>\n\n<p>So my question is how exactly batch normalization could be possible to cause that? Any ideas on how to deal with it? I tried bunches of hyper-parameters of batch norm, it seems not helping.<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":41670773,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-01-16T06:42:30.377","AcceptedAnswerId":41671894.0,"OwnerUserId":6430408.0,"Title":"tf.summary.image gets the error InvalidArgumentError","Body":"<p>I have train my data with cnn model and shuffle images. The first convolution layer is defined:<\/p>\n\n<pre><\/pre>\n\n<p>if I remove the line\" tf.summary.image('conv1_img',conv1)\", program can run successfully. When I add this line ,the error:<\/p>\n\n<p>v<\/p>\n\n<p>happens,why?<\/p>\n","answers":[{"AnswerId":"41671894","CreationDate":"2017-01-16T08:07:56.593","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>The summary of you define with <code>tf.summary.image<\/code> is automatically added to a collection of summaries.<\/p>\n\n<p>You surely run the op <code>summaries = tf.summary.merge_all(key='summaries')<\/code> to collect all the summaries added to the collection named <code>summaries<\/code> (the default sumamry collection).<\/p>\n\n<p>Then, once you run this op within a session with <code>sess.run(summaries)<\/code>, every summary previously defined is executed.<\/p>\n\n<p>The histogram summaries depend on the model parameters values only, thus they don't need any external data to be computed.<\/p>\n\n<p>Instead, the <code>tf.summary.image('conv1_img',conv1)<\/code> draws the output of the <code>conv1<\/code> op that needs a placeholder (<code>image<\/code>) to be computed.<\/p>\n\n<p>Thus, you should execute the summary op feeding the graph with the image placeholder:<\/p>\n\n<p><code>sess.run(summaries, feed_dict{image: &lt;your image here&gt;})<\/code>.<\/p>\n\n<p>A suggestion: <\/p>\n\n<p>Let the placeholder be a placehoder. With the statement<\/p>\n\n<pre><code>image = tf.placeholder(tf.float32, [FLAGS.batch_size, 32, 32, 1])\nimage = tf.reshape(image, [FLAGS.batch_size, 32, 32, 1])\n<\/code><\/pre>\n\n<p>you're<\/p>\n\n<ol>\n<li>defining <code>image<\/code> as a placeholder<\/li>\n<li>overwriting the python <code>image<\/code>  variable to a tensorflow operation.<\/li>\n<\/ol>\n\n<p>Thus, when you run use the <code>feed_dict<\/code> parameter to inject into the computational graph a value for the image placeholder, you're in reality, overriding a tensorflow op (and thus you have to feed an already reshaped value to make it work).<\/p>\n\n<p>Thus, it's better to let the placeholder be a placeholder:<\/p>\n\n<pre><code>image_placeholder = tf.placeholder(tf.float32, [FLAGS.batch_size, 32, 32, 1])\nimage = tf.reshape(image_placeholder, [FLAGS.batch_size, 32, 32, 1])\n<\/code><\/pre>\n\n<p>and then use it to correctly feed the graph:<\/p>\n\n<pre><code>sess.run(&lt;your ops&gt;, feed_dict:{image_placeholder: &lt;your image&gt;})\n<\/code><\/pre>\n"}]}
{"QuestionId":41670995,"AnswerCount":3,"Tags":"<keras><loss>","CreationDate":"2017-01-16T06:59:13.760","AcceptedAnswerId":null,"OwnerUserId":2899869.0,"Title":"Keras Implementation of Customized Loss Function that need internal layer output as label","Body":"<p>in keras, I want to customize my loss function which not only takes (y_true, y_pred) as input but also need to use the output from the internal layer of the network as the label for an output layer.This picture shows the <a href=\"https:\/\/i.stack.imgur.com\/UCHFa.png\" rel=\"noreferrer\">Network Layout<\/a><\/p>\n\n<p>Here, the internal output is xn, which is a 1D feature vector. in the upper right corner, the output is xn', which is the prediction of xn. In other words, xn is the label for xn'. <\/p>\n\n<p>While [Ax, Ay] is traditionally known as y_true, and [Ax',Ay'] is y_pred.<\/p>\n\n<p>I want to combine these two loss components into one and train the network jointly. <\/p>\n\n<p>Any ideas or thoughts are much appreciated!<\/p>\n","answers":[{"AnswerId":"56913095","CreationDate":"2019-07-06T10:08:54.760","ParentId":null,"OwnerUserId":"6144641","Title":null,"Body":"<p>I have my reservations regarding this implementation. The loss computed at the merged layer is propagated back to both the merged branches. Generally you would like to propagate it through just one layer.<\/p>\n"},{"AnswerId":"41672221","CreationDate":"2017-01-16T08:33:10.650","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>First of all you should be using the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">Functional API<\/a>. Then you should define the network output as the output plus the result from the internal layer, merge them into a single output (by concatenating), and then make a custom loss function that then splits the merged output into two parts and does the loss computations on its own.<\/p>\n\n<p>Something like:<\/p>\n\n<pre><code>def customLoss(y_true, y_pred):\n    #loss here\n    internalLayer = Convolution2D()(inputs) #or other layers\n    internalModel = Model(input=inputs, output=internalLayer)\n    tmpOut = Dense(...)(internalModel)\n    mergedOut = merge([tmpOut, mergedOut], mode = \"concat\", axis = -1)\n    fullModel = Model(input=inputs, output=mergedOut)\n\n    fullModel.compile(loss = customLoss, optimizer = \"whatever\")\n<\/code><\/pre>\n"},{"AnswerId":"41798509","CreationDate":"2017-01-23T02:37:57.010","ParentId":null,"OwnerUserId":"2899869","Title":null,"Body":"<p>I have figured out a way out, in case anyone is searching for the same, I posted here (based on the network given in this post):<\/p>\n\n<p>The idea is to define the customized loss function and use it as the output of the network. (Notation: <code>A<\/code> is the true label of variable <code>A<\/code>, and <code>A'<\/code> is the predicted value of variable <code>A<\/code>)<\/p>\n\n<pre><code>def customized_loss(args):\n    #A is from the training data\n    #S is the internal state\n    A, A', S, S' = args \n    #customize your own loss components\n    loss1 = K.mean(K.square(A - A'), axis=-1)\n    loss2 = K.mean(K.square(S - S'), axis=-1)\n    #adjust the weight between loss components\n    return 0.5 * loss1 + 0.5 * loss2\n\n def model():\n     #define other inputs\n     A = Input(...) # define input A\n     #construct your model \n     cnn_model = Sequential()\n     ...\n     # get true internal state\n     S = cnn_model(prev_layer_output0)\n     # get predicted internal state output\n     S' = Dense(...)(prev_layer_output1)\n     # get predicted A output\n     A' = Dense(...)(prev_layer_output2)\n     # customized loss function\n     loss_out = Lambda(customized_loss, output_shape=(1,), name='joint_loss')([A, A', S, S'])\n     model = Model(input=[...], output=[loss_out])\n     return model\n\n  def train():\n      m = model()\n      opt = 'adam'\n      model.compile(loss={'joint_loss': lambda y_true, y_pred:y_pred}, optimizer = opt)\n      # train the model \n      ....\n<\/code><\/pre>\n"}]}
{"QuestionId":41671631,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-16T07:46:38.343","AcceptedAnswerId":41671945.0,"OwnerUserId":2558506.0,"Title":"Tensorflow: Save and restore the output of a tensor","Body":"<p>I would like to achieve something like this using tensorflow.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XOf4U.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XOf4U.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I can only find documentation on saving and restoring variables (weights). However, like #2-2, I want to utilize the output of a hidden layer (tensor) as input of another model. Can this be done?<\/p>\n","answers":[{"AnswerId":"41671945","CreationDate":"2017-01-16T08:13:05.057","ParentId":null,"OwnerUserId":"1150961","Title":null,"Body":"<p>As far as I'm aware, it is not possible to chain different computation graphs after they have been created however, you have a few options.<\/p>\n\n<p>Option 2: Create one large graph and use a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\" rel=\"nofollow noreferrer\">control flow op<\/a><\/p>\n\n<pre><code>output_layer, placeholder = build_my_model()\nsomething = tf.where(output_layer &lt; 0, do_something_1(), do_something_2())\n<\/code><\/pre>\n\n<p>where all function calls above should return tensorflow operations.<\/p>\n\n<p>Option2: Create two graphs and perform the conditional statement inside python<\/p>\n\n<pre><code># Build the first graph\nwith tf.Graph().as_default() as graph:\n    output_layer, placeholder = build_my_model()\n\n# Build the second two graphs\nwith tf.Graph().as_default() as graph_1:\n    something_1 = do_something_1()\nwith tf.Graph().as_default() as graph_2:\n    something_2 = do_something_2()\n<\/code><\/pre>\n\n<p>As a result, you will also end up with three different sessions and you need to feed the output from the first session to one of the other two<\/p>\n\n<pre><code># Get the output\n_output_layer = sess.run(output_layer, {placeholder: ...})\nif _output_layer &lt; 0:\n    something = sess1.run(something_1, {...})\nelse:\n    something = sess2.run(something_2, {...})\n<\/code><\/pre>\n\n<p>As you can see, if you can get away with the control flow op, your code will be significantly simpler. Another advantage of having everything in one graph is that the entire graph is differentiable and you can train parameters of the first stage of your model conditional on the loss at a later stage.<\/p>\n"}]}
{"QuestionId":41671791,"AnswerCount":1,"Tags":"<caffe><matrix-multiplication><cblas><inner-product>","CreationDate":"2017-01-16T07:59:06.043","AcceptedAnswerId":41726711.0,"OwnerUserId":2693257.0,"Title":"How to do dot product between matrices in caffe?","Body":"<p>In inner product layer, I need to multiply . First we calculate () as matrix multiplication in  and then do a  between  and .<\/p>\n\n<p>More explanation is defined as follow:<\/p>\n\n<pre><\/pre>\n\n<p>For more understanding, I checked . It is implemented as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I think I should perform multiplication () in  and after that do  with . how should I do?!<\/p>\n\n<p>Many thanks!!!! Any advice would be appreciated!<\/p>\n","answers":[{"AnswerId":"41726711","CreationDate":"2017-01-18T18:20:31.023","ParentId":null,"OwnerUserId":"628538","Title":null,"Body":"<p>If you just want to perform dot product between two matrices, you can use the following function to multiply matrices on CPU,<\/p>\n\n<p><code>void caffe_mul&lt;float&gt;(const int n, const float* a, const float* b, float* y)<\/code><\/p>\n\n<p>If you want to do the same operation on a GPU, use this template<\/p>\n\n<pre><code>void caffe_gpu_mul&lt;float&gt;(const int N, const float* a, const float* b, float* y)\n<\/code><\/pre>\n\n<p>a and b are you matrices and c will contain the final result. N is the total number of elements in your matrix.<\/p>\n\n<p>You can also use the 'Eltwise' layer, which already does this.<\/p>\n"}]}
{"QuestionId":41671817,"AnswerCount":3,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-01-16T08:01:35.040","AcceptedAnswerId":41675152.0,"OwnerUserId":6334174.0,"Title":"How to visualize a tensor summary in tensorboard","Body":"<p>I'm trying to visualize a tensor summary in tensorboard. However I can't see the tensor summary at all in the board. Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>where out is a multi-dimensional tensor. I guess there must be something wrong with my code. Probably I used the  function incorrectly.<\/p>\n","answers":[{"AnswerId":"45433366","CreationDate":"2017-08-01T08:59:15.853","ParentId":null,"OwnerUserId":"7703834","Title":null,"Body":"<p>Not sure whether this is kinda obvious, but you could use something like<\/p>\n\n<pre><code>def make_tensor_summary(tensor, name='defaultTensorName'):\n    for i in range(tensor.get_shape()[0]:\n        for j in range(tensor.get_shape()[1]:\n            tf.summary.scalar(Name + str(i) + '_' + str(j), tensor[i, j])\n<\/code><\/pre>\n\n<p>in case you know it is a 'matrix-shaped' Tensor in advance.<\/p>\n"},{"AnswerId":"41675152","CreationDate":"2017-01-16T11:16:39.877","ParentId":null,"OwnerUserId":"5687374","Title":null,"Body":"<p>What you do is you create a summary op, but you don't invoke it and don't write the summary (see <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/train\/summary_operations\" rel=\"nofollow noreferrer\">documentation<\/a>).\nTo actually create a summary you need to do the following:<\/p>\n\n<pre><code># Create a summary operation\nsummary_op = tf.summary.tensor_summary('softmax_input', out)\n\n# Create the summary\nsummary_str = sess.run(summary_op)\n\n# Create a summary writer\nwriter = tf.train.SummaryWriter(...)\n\n# Write the summary\nwriter.add_summary(summary_str)\n<\/code><\/pre>\n\n<p>Explicitly writing a summary (last two lines) is only necessary if you don't have  a higher level helper like a <a href=\"https:\/\/www.tensorflow.org\/how_tos\/supervisor\/#writing_summaries\" rel=\"nofollow noreferrer\">Supervisor<\/a>. Otherwise you invoke <\/p>\n\n<pre><code>sv.summary_computed(sess, summary_str)\n<\/code><\/pre>\n\n<p>and the Supervisor will handle it.<\/p>\n\n<p>More info, also see:\n<a href=\"https:\/\/stackoverflow.com\/questions\/37902705\/how-to-manually-create-a-tf-summary\/37915182#comment70127530_37915182\">How to manually create a tf.Summary()<\/a><\/p>\n"},{"AnswerId":"50298443","CreationDate":"2018-05-11T18:43:49.207","ParentId":null,"OwnerUserId":"5585800","Title":null,"Body":"<p>Hopefully a workaround which achieves what you want. ..<\/p>\n\n<p>If you wish to view the tensor values, you can convert them using as_string, then use summary.text. The values will appear in the tensorboard text tab.<\/p>\n\n<p>Not tried with 3D tensors, but feel free to slice according to needs.<\/p>\n\n<p>code snippet, which includes use of inserting a print statement to get console output as well.<\/p>\n\n<pre><code>predictions = tf.argmax(reshaped_logits, 1)\ntxtPredictions = tf.Print(tf.as_string(predictions),[tf.as_string(predictions)], message='predictions', name='txtPredictions')\ntxtPredictions_op = tf.summary.text('predictions', txtPredictions)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/y6uSb.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/y6uSb.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":41672114,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorflow-serving>","CreationDate":"2017-01-16T08:25:29.047","AcceptedAnswerId":41922300.0,"OwnerUserId":2106858.0,"Title":"Add Tensorflow pre-processing to existing Keras model (for use in Tensorflow Serving)","Body":"<p>I would like to include my custom pre-processing logic in my exported Keras model for use in Tensorflow Serving.<\/p>\n\n<p><em>My pre-processing performs string tokenization and uses an external dictionary to convert each token to an index for input to the Embedding layer:<\/em><\/p>\n\n<pre><\/pre>\n\n<p><em>Model architecture and training:<\/em><\/p>\n\n<pre><\/pre>\n\n<p>Since the model will be used in Tensorflow Serving, I want to incorporate all pre-processing logic into the model itself (encoded in the exported model file).<\/p>\n\n<p><strong>Q: How can I do so using the Keras library only?<\/strong><\/p>\n\n<p>I found <a href=\"https:\/\/blog.keras.io\/keras-as-a-simplified-interface-to-tensorflow-tutorial.html\" rel=\"noreferrer\">this guide<\/a> explains how to combine Keras and Tensorflow. But I'm still unsure how to export everything as one model.<\/p>\n\n<p>I know Tensorflow has built-in string splitting, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/lookup\/lookup_ops.py#L320\" rel=\"noreferrer\">file I\/O<\/a>, and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/lookup\/lookup_ops.py#L195\" rel=\"noreferrer\">dictionary lookup operations<\/a>.<\/p>\n\n<p><em>Pre-processing logic using Tensorflow operations:<\/em><\/p>\n\n<pre><\/pre>\n\n<p><strong>Q: How can I use these Tensorflow pre-defined pre-processing operations and my Keras layers together to both train and then export the model as a \"black box\" for use in Tensorflow Serving?<\/strong><\/p>\n","answers":[{"AnswerId":"47805771","CreationDate":"2017-12-14T04:10:08.110","ParentId":null,"OwnerUserId":"3525027","Title":null,"Body":"<p>The accepted answer is super helpful, however it uses an outdated Keras API as @Qululu mentioned, and an outdated TF Serving API (Exporter), and it does not show how to export the model so that its input is the original tf placeholder (versus Keras model.input, which is post preprocessing). Following is a version that works well as of TF v1.4 and Keras 2.1.2:<\/p>\n\n<pre><code>sess = tf.Session()\nK.set_session(sess)\n\nK._LEARNING_PHASE = tf.constant(0)\nK.set_learning_phase(0)\n\nmax_features = 5000\nmax_lens = 500\n\ndict_table = tf.contrib.lookup.HashTable(tf.contrib.lookup.TextFileInitializer(\"vocab.txt\",tf.string, 0, tf.int64, TextFileIndex.LINE_NUMBER, vocab_size=max_features, delimiter=\" \"), 0)\n\nx_input = tf.placeholder(tf.string, name='x_input', shape=(None,))\nsparse_tokenized_input = tf.string_split(x_input)\ntokenized_input = tf.sparse_tensor_to_dense(sparse_tokenized_input, default_value='')\ntoken_idxs = dict_table.lookup(tokenized_input)\ntoken_idxs_padded = tf.pad(token_idxs, [[0,0],[0, max_lens]])\ntoken_idxs_embedding = tf.slice(token_idxs_padded, [0,0], [-1, max_lens])\n\nmodel = Sequential()\nmodel.add(InputLayer(input_tensor=token_idxs_embedding, input_shape=(None, max_lens)))\n\n ...REST OF MODEL...\n\nmodel.load_weights(\"model.h5\")\n\nx_info = tf.saved_model.utils.build_tensor_info(x_input)\ny_info = tf.saved_model.utils.build_tensor_info(model.output)\n\nprediction_signature = tf.saved_model.signature_def_utils.build_signature_def(inputs={\"text\": x_info}, outputs={\"prediction\":y_info}, method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)\n\nbuilder = saved_model_builder.SavedModelBuilder(\"\/path\/to\/model\")\n\nlegacy_init_op = tf.group(tf.tables_initializer(), name='legacy_init_op')\n\ninit_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())\nsess.run(init_op)\n\n\n# Add the meta_graph and the variables to the builder\nbuilder.add_meta_graph_and_variables(\n  sess, [tag_constants.SERVING],\n  signature_def_map={\n       signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:\n           prediction_signature,\n  },\n  legacy_init_op=legacy_init_op)\n\nbuilder.save()  \n<\/code><\/pre>\n\n<p><strong>UPDATE<\/strong> Doing pre-processing for inference with Tensorflow is a CPU op, and is not carried out efficiently if the model is deployed on a GPU server. The GPU stalls really bad, and the throughput is very low. Therefore, we ditched this for efficient pre-processing in the client process, instead.<\/p>\n"},{"AnswerId":"41922300","CreationDate":"2017-01-29T15:04:05.573","ParentId":null,"OwnerUserId":"2106858","Title":null,"Body":"<p>I figured it out, so I'm going to answer my own question here.<\/p>\n\n<p><strong>Here's the gist:<\/strong><\/p>\n\n<p>First, (in separate code file) I trained the model using Keras only with my own pre-processing functions, exported the Keras model weights file and my token-to-index dictionary.<\/p>\n\n<p>Then, I copied just the Keras model architecture, set the input as the pre-processed tensor output, loaded the weights file from the previously trained Keras model, and sandwiched it between the Tensorflow pre-processing operations and the Tensorflow exporter.<\/p>\n\n<p><strong>Final product:<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nfrom keras import backend as K\nfrom keras.models import Sequential, Embedding, LSTM, Dense\nfrom tensorflow.contrib.session_bundle import exporter\nfrom tensorflow.contrib.lookup import HashTable, TextFileInitializer\n\n# Initialize Keras with Tensorflow session\nsess = tf.Session()\nK.set_session(sess)\n\n# Token to index lookup dictionary\ntoken_to_idx_path = '...'\ntoken_to_idx_dict = HashTable(TextFileInitializer(token_to_idx_path, tf.string, 0, tf.int64, 1, delimiter='\\t'), 0)\n\nmaxlen = ...\n\n# Pre-processing sub-graph using Tensorflow operations\ninput = tf.placeholder(tf.string, name='input')\nsparse_tokenized_input = tf.string_split(input)\ntokenized_input = tf.sparse_tensor_to_dense(sparse_tokenized_input, default_value='')\ntoken_idxs = token_to_idx_dict.lookup(tokenized_input)\ntoken_idxs_padded = tf.pad(token_idxs, [[0,0],[0,maxlen]])\ntoken_idxs_embedding = tf.slice(token_idxs_padded, [0,0], [-1,maxlen])\n\n# Initialize Keras model\nmodel = Sequential()\ne = Embedding(max_features, 128, input_length=maxlen)\ne.set_input(token_idxs_embedding)\nmodel.add(e)\nmodel.add(LSTM(128, activation='sigmoid'))\nmodel.add(Dense(num_classes, activation='softmax'))\n\n# Load weights from previously trained Keras model\nweights_path = '...'\nmodel.load_weights(weights_path)\n\nK.set_learning_phase(0)\n\n# Export model in Tensorflow format\n# (Official tutorial: https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/g3doc\/serving_basic.md)\nsaver = tf.train.Saver(sharded=True)\nmodel_exporter = exporter.Exporter(saver)\nsignature = exporter.classification_signature(input_tensor=model.input, scores_tensor=model.output)\nmodel_exporter.init(sess.graph.as_graph_def(), default_graph_signature=signature)\nmodel_dir = '...'\nmodel_version = 1\nmodel_exporter.export(model_dir, tf.constant(model_version), sess)\n\n# Input example\nwith sess.as_default():\n    token_to_idx_dict.init.run()\n    sess.run(model.output, feed_dict={input: [\"this is a raw input example\"]})\n<\/code><\/pre>\n"}]}
{"QuestionId":41672219,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-16T08:33:08.617","AcceptedAnswerId":41682356.0,"OwnerUserId":7424270.0,"Title":"How to generate CNN model with changing frames in tensorflow?","Body":"<p>I want to build a conditional CNN model in Tensorflow, but I met some trouble with it.<\/p>\n\n<p>Suppose there is a matrix named <i><b>features<\/b><\/i> with shape [64, ?, 50, 1] and another matrix named <i><b>conditions<\/b><\/i> with shape [64, ?, 1, 130].<\/p>\n\n<p>The fisrt dim is the batch size, and the second dim in means the frame number (<b>the value in this dim of the two matrices is same in each batch, but different in different batchs<\/b>). The third one is the feature dimention, and the last dim is channel num.<\/p>\n\n<p>I want to concat the two matrix, which means we can get a matrix with shape [64, ?, 50, 131], and do <i><b>conv2d<\/b><\/i> in tensorflow.<\/p>\n\n<p>But the problem is that I cannot concat these two matrices because of the difference in the third dim. Then I did as follows:<\/p>\n\n<pre><\/pre>\n\n<p>But it doesn't work because the second dim is not known.<\/p>\n\n<p>I wonder is there any way to solve this problem?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41682356","CreationDate":"2017-01-16T17:58:14.543","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>Although it is not clear from your question, it looks like you want to broadcast the second tensor with shape <code>[64, ?, 1, 130]<\/code> on the 3rd dimension, the one you call feature dimension, before you concatenate. Note that <code>tf.concat<\/code> needs all the dimensions to match, except the one along which you are concatenating. From the documentation of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#concat\" rel=\"nofollow noreferrer\"><code>tf.concat<\/code><\/a> :<\/p>\n\n<blockquote>\n  <p>The number of dimensions of the input tensors must match, and all\n  dimensions except concat_dim must be equal.<\/p>\n<\/blockquote>\n\n<p>To do the broadcast along the feature dimension, it is much cheaper to use  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#tile\" rel=\"nofollow noreferrer\"><code>tf.tile<\/code><\/a> instead of multiplying with a all ones tensor like you are doing. Here is how you would use <code>tf.tile<\/code> :<\/p>\n\n<pre><code>tf.concat(3, [x, tf.tile(y, [1, 1, x_shapes.as_list()[2], 1])])\n<\/code><\/pre>\n\n<p>In your case, since the 3rd dimension of <code>y<\/code> is known statically, the above code works. If that's not the case, you can form the second argument to <code>tf.tile<\/code> as follows :<\/p>\n\n<pre><code>tf.tile(y, tf.concat(0, [[1], [1], tf.shape(x)[2:3], [1]]))\n<\/code><\/pre>\n"}]}
{"QuestionId":41673889,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><computer-vision><deep-learning><tf-slim>","CreationDate":"2017-01-16T10:09:04.273","AcceptedAnswerId":41676452.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: does tf.train.batch automatically load the next batch when the batch has finished training?","Body":"<p>For instance, after I have created my operations, fed the batch data through the operation and run the operation, does tf.train.batch automatically feed in another batch of data to the session?<\/p>\n\n<p>I ask this because tf.train.batch has an attribute of  which makes it possible for the final batch to be loaded as a size lesser than the indicated batch size. Does this mean even without a loop, the next batch could be automatically fed? From the tutorial codes I am rather confused. When I load a single batch, I get literally a single batch size of shape [batch_size, height, width, num_channels], but the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#batch\" rel=\"noreferrer\">documentation<\/a> says it  Also, when I read the tutorial code in the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/slim_walkthough.ipynb\" rel=\"noreferrer\">tf-slim walkthrough tutorial<\/a>, where there is a function called load_batch, there are only 3 tensors returned: . Where are 'batches' of data as explained in the documentation?<\/p>\n\n<p>Thank you for your help.<\/p>\n","answers":[{"AnswerId":"41676452","CreationDate":"2017-01-16T12:28:07.553","ParentId":null,"OwnerUserId":"1246987","Title":null,"Body":"<blockquote>\n  <p>... does tf.train.batch automatically feeds in another batch of data to the session?<\/p>\n<\/blockquote>\n\n<p>No. Nothing happens automatically. You must call <code>sess.run(...)<\/code> again to load a new batch.<\/p>\n\n<blockquote>\n  <p>Does this mean even without a loop, the next batch could be automatically fed?<\/p>\n<\/blockquote>\n\n<p>No. <code>tf.train.batch(..)<\/code> will always load <code>batch_size<\/code> tensors. If you have for example 100 images and a <code>batch_size=30<\/code> then you will have 3*30 batches as in you can call <code>sess.run(batch)<\/code> three times before the input queue will start from the beginning (or stop if <code>epoch=1<\/code>). This means that you miss out <code>100-3*30=10<\/code> samples from training. In case you do not want to miss them you can do <code>tf.train.batch(..., allow_smaller_final_batch=True)<\/code> so now you will have 3x 30-sample-batches and 1x 10-sample-batch before the input queue will restart.<\/p>\n\n<p>Let me also elaborate with a code sample:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>queue = tf.train.string_input_producer(filenames,\n        num_epochs=1) # only iterate through all samples in dataset once\n\nreader = tf.TFRecordReader() # or any reader you need\n_, example = reader.read(queue)\n\nimage, label = your_conversion_fn(example)\n\n# batch will now load up to 100 image-label-pairs on sess.run(...)\n# most tf ops are tuned to work on batches\n# this is faster and also gives better result on e.g. gradient calculation\nbatch = tf.train.batch([image, label], batch_size=100)\n\nwith tf.Session() as sess:\n    # \"boilerplate\" code\n    sess.run([\n        tf.local_variables_initializer(),\n        tf.global_variables_initializer(),\n    ])\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\n    try:\n        # in most cases coord.should_stop() will return True\n        # when there are no more samples to read\n        # if num_epochs=0 then it will run for ever\n        while not coord.should_stop():\n            # will start reading, working data from input queue\n            # and \"fetch\" the results of the computation graph\n            # into raw_images and raw_labels\n            raw_images, raw_labels = sess.run([images, labels])\n    finally:\n        coord.request_stop()\n        coord.join(threads)\n<\/code><\/pre>\n"},{"AnswerId":"54558049","CreationDate":"2019-02-06T16:18:29.803","ParentId":null,"OwnerUserId":"11023902","Title":null,"Body":"<p>You need to call sess.run and pass the batch to it everytime when you want to load the next batch. See the code below.<\/p>\n\n<pre><code>img = [0,1,2,3,4,5,6,7,8]\nlbl = [0,1,2,3,4,5,6,7,8]\nimages = tf.convert_to_tensor(img)\nlabels = tf.convert_to_tensor(lbl)\ninput_queue = tf.train.slice_input_producer([images,labels])\nsliced_img = input_queue[0]\nsliced_lbl = input_queue[1]\n\nimg_batch, lbl_batch = tf.train.batch([sliced_img,sliced_lbl], batch_size=3)\nwith tf.Session() as sess:\n    coord   = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n\n    for i in range(0,3): #batch size\n        image_batch,label_batch = sess.run([img_batch,lbl_batch ])\n        print(image_batch, label_batch)\n\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n\n<p>the answer would be something like this: <\/p>\n\n<p>[4,1,8] [4,1,8]<\/p>\n\n<p>[2,3,7] [2,3,7]<\/p>\n\n<p>[2,6,8] [2,6,8]<\/p>\n"}]}
{"QuestionId":41675030,"AnswerCount":2,"Tags":"<python><windows><tensorflow><anaconda><spyder>","CreationDate":"2017-01-16T11:09:41.567","AcceptedAnswerId":null,"OwnerUserId":7032323.0,"Title":"Anaconda+Spyder+Windows Tensorflow installation","Body":"<p>The topic describes the problem. I cannot install TF and use it in Anaconda withtin Spyder IDE on Windows. Installation of TF GPU v0.12.1 within this script successfully finishes:<\/p>\n\n<pre><\/pre>\n\n<p>However, I still get this error:<\/p>\n\n<pre><\/pre>\n\n<p>I ensured that Visual C++ x64 2015 Redist is installed. Also tried a lot of reinstallations, reboots, google solutions etc, but nothing helps.<\/p>\n\n<p>Try to run this script:<\/p>\n\n<pre><\/pre>\n\n<p>Can you please advice something?<\/p>\n\n<p>Thank you so much!\nArtemiy<\/p>\n","answers":[{"AnswerId":"43821979","CreationDate":"2017-05-06T15:06:39.847","ParentId":null,"OwnerUserId":"7973404","Title":null,"Body":"<p>Update TensorFlow with <\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-0.12.1-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n"},{"AnswerId":"42354561","CreationDate":"2017-02-20T21:29:56.177","ParentId":null,"OwnerUserId":"4847576","Title":null,"Body":"<p>The error message is clear, you need to initialize TensorFlow outside its directory. If that is not your problem check if you have <code>MSVCP140.dll<\/code> installed in your system (see <strong>NOTE<\/strong> under <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup#pip_installation_on_windows\" rel=\"nofollow noreferrer\">Pip Installation on Windows<\/a>) considering your Cuda and cuDNN are properly installed and their directories set on your <code>%PATH%<\/code>.<\/p>\n"}]}
{"QuestionId":41675482,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2017-01-16T11:34:07.480","AcceptedAnswerId":null,"OwnerUserId":3839572.0,"Title":"Difference between CUDA out of memory and misaligned address","Body":"<p>I found the error when training using Caffe. Seems both of the errors are connected with memory because when I reduce the batch size, both the errors are solved.<\/p>\n\n<p>Just, seems the misaligned address comes first before out of memory.\nFor example, when I put batch size = 32 (or more), I got out of memory error. When I put batch size = 16, I got misaligned address error. When I put batch size = 8 (or less), there is no error.<\/p>\n\n<p>Actually what is the difference between those errors and the meaning of the errors?<\/p>\n\n<p>I use CUDA 8.0 and CuDNN 5.1<\/p>\n","answers":[]}
{"QuestionId":41677509,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-16T13:28:15.607","AcceptedAnswerId":null,"OwnerUserId":7194615.0,"Title":"When using tf.get_variable instead of tf.Variable(), it seems that the parameters initialized in every epoch","Body":"<p>This is just a simple conv-net. I just used  before, then I found I need a shared variable to see the test result. So I use  instead of , but I found the loss could not be decreased in the training process. I think maybe there is something wrong in my code, but I could not fix it. It seems like the parameters are initialized in every epoch, how to fix it in every epoch?<\/p>\n\n<pre><\/pre>\n\n<p>and this is another part of the code<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41681249","CreationDate":"2017-01-16T16:52:33.993","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>Try this instead:<\/p>\n\n<pre><code>logits=lenet5.inference(images)\ntf.get_variable_scope().reuse_variables()\nlogits_test=lenet5.inference(images_test)\n<\/code><\/pre>\n\n<p>and then remove everything below scope.reuse_variables in your inference function. <\/p>\n"}]}
{"QuestionId":41678837,"AnswerCount":1,"Tags":"<python><theano><theano-cuda>","CreationDate":"2017-01-16T14:39:10.187","AcceptedAnswerId":41702662.0,"OwnerUserId":3426322.0,"Title":"`No module named nose_parameterized` during theano installation","Body":"<p>At the moment I am trying to install theano on my ubuntu 16.04 system with cuda support. The installation process itself went well until I tried to test the installation:  <\/p>\n\n<pre><\/pre>\n\n<p>The last few lines of the output were:<\/p>\n\n<pre><\/pre>\n\n<p>These were preceded by approximately 20 blocks with what appears to be different tests failing on exactly the same top 3 lines in the stacktrace. The code where the final occurs looks like this (with added line numbers):  <\/p>\n\n<pre><\/pre>\n\n<p>While searching for a solution I came across:<br>\n<a href=\"https:\/\/stackoverflow.com\/questions\/29925801\/error-running-nosetests\">Error running nosetests<\/a><br>\nI tried the solution proposed there:<br>\n<br>\nHowever this had no effect. Does anyone know what else could be causing this error?  <\/p>\n\n<p>edit: so I reinstalled theano to make the problem more reproducible. I installed theano using the instructions found <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_ubuntu.html#install-ubuntu\" rel=\"nofollow noreferrer\">here<\/a>, so I used:  <\/p>\n\n<pre><\/pre>\n\n<p>Then I used:<\/p>\n\n<pre><\/pre>\n\n<p>The last few lines of the output of this are almost the same as earlier:  <\/p>\n\n<pre><\/pre>\n\n<p>edit 2: I thought it might be a python-version issue. When I try:  <\/p>\n\n<pre><\/pre>\n\n<p>I get:  <\/p>\n\n<pre><\/pre>\n\n<p>And the following folders exist:<br>\n<br>\nSo, that shouldn't be a problem as far as I can see. However when I try:  <\/p>\n\n<pre><\/pre>\n\n<p>I get:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41702662","CreationDate":"2017-01-17T16:51:41.980","ParentId":null,"OwnerUserId":"3426322","Title":null,"Body":"<p>Found the problem, anaconda was looking in its own folder. So to solve, I did:<br>\n<code>conda install nose-parameterized<\/code><\/p>\n"}]}
{"QuestionId":41679110,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-16T14:54:01.390","AcceptedAnswerId":null,"OwnerUserId":2806040.0,"Title":"How to use tensorflow-wavenet","Body":"<p>I am trying to use the <a href=\"https:\/\/github.com\/ibab\/tensorflow-wavenet\" rel=\"noreferrer\">tensorflow-wavenet<\/a> program for text to speech.<\/p>\n\n<p>These are the steps:<\/p>\n\n<ol>\n<li>Download Tensorflow<\/li>\n<li>Download librosa<\/li>\n<li>Install requirements <\/li>\n<li>Download corpus and put into directory named \"corpus\"<\/li>\n<li>Train the machine <\/li>\n<li>Generate audio <\/li>\n<\/ol>\n\n<p>After doing this, how can I generate a voice read-out of a text file?<\/p>\n","answers":[{"AnswerId":"41681673","CreationDate":"2017-01-16T17:17:32.400","ParentId":null,"OwnerUserId":"3856879","Title":null,"Body":"<p>According to the <a href=\"https:\/\/github.com\/ibab\/tensorflow-wavenet\/blob\/master\/README.md#missing-features\" rel=\"nofollow noreferrer\">tensorflow-wavenet page<\/a>: <\/p>\n\n<blockquote>\n  <p>Currently there is no local conditioning on extra information which would allow context stacks or controlling what speech is generated.<\/p>\n<\/blockquote>\n\n<p>You can find more information about current development of the project by reading the issues on the repository (<a href=\"https:\/\/github.com\/ibab\/tensorflow-wavenet\/issues\/189\" rel=\"nofollow noreferrer\">local conditioning is a desired feature!<\/a>)<\/p>\n\n<p>The Wavenet paper compares Wavenet to two TTS baselines, one of which appears to have code for training available online: <a href=\"http:\/\/hts.sp.nitech.ac.jp\" rel=\"nofollow noreferrer\">http:\/\/hts.sp.nitech.ac.jp<\/a><\/p>\n"},{"AnswerId":"48302818","CreationDate":"2018-01-17T13:54:34.860","ParentId":null,"OwnerUserId":"8053554","Title":null,"Body":"<p>A recent paper by DeepMind describes one approach to going from text to speech using WaveNet, which I have not tried to implement but which at least states the method they use: they first train one network to predict a spectrogram from text, then train WaveNet to use the same sort of spectrogram as an additional conditional input to produce speech. It's a neat idea, especially since you can train the WaveNet part on some huge database of voice-only data, for which you can extract the spectrogram, and then train the text-to-spectrogram part using a different dataset where you have text. <\/p>\n\n<p><a href=\"https:\/\/google.github.io\/tacotron\/publications\/tacotron2\/index.html\" rel=\"nofollow noreferrer\">https:\/\/google.github.io\/tacotron\/publications\/tacotron2\/index.html<\/a> has the paper and some example outputs.<\/p>\n\n<p>There seems to be a bunch of unintuitive engineering around the spectrogram prediction part (no doubt because of the nature of text-to-time learning), but there's some detail in the paper at least. The dataset is proprietary so I've no idea how hard it would be to get any results using other datasets.<\/p>\n"}]}
{"QuestionId":41679182,"AnswerCount":2,"Tags":"<python><optimization><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-16T14:57:20.713","AcceptedAnswerId":41679319.0,"OwnerUserId":7421488.0,"Title":"Is there some way can accomplish stochastic gradient descent not from scratch","Body":"<p>For a standard machine learning problem, e.g, image classification on MNIST, the loss function is fixed, therefor the optimization process can be accomplished simply by calling functions and feed the input into them. There is no need to derive gradients and code the descent procedure by hand. <\/p>\n\n<p>But now I'm confused when I met some complicated formulation. Say we are solving a semi-supervised problem, and the loss function has two parts:. The first part is a normal classification formulation, e.g, cross entropy loss. And the second part varies. In my situation,  is a matrix factorization loss, which in specific is:. And the total loss function can be written as:<\/p>\n\n<p><\/p>\n\n<p>Where parameters are  and . The first part is a classification loss, and the input  is a raw of , i.e. , a vector of size . And the label  can be a one-hot vector of size , so parameters  and  map the input to the label size. And the second part is a matrix factorization loss.<\/p>\n\n<p>Now I'm confused when I'm going to optimize this function using sgd. It can be solved by write down the formulation derive gradients then accomplish a training procedure from scratch. But I'm wondering if there is a simpler way? Because it's easy to use a deep learning tool like Tensorflow or Keras to train a classification model, all u need to do is build a network and feed the data. <\/p>\n\n<p>So similarly, is there a tool that can automatically compute gradients after I defined the loss function? Because deriving gradients and achieve them from scratch is really annoying. Both the classification loss and matrix factorization loss is very common, so I think the combination can be achieved thoroughly.<\/p>\n","answers":[{"AnswerId":"41679319","CreationDate":"2017-01-16T15:04:22.900","ParentId":null,"OwnerUserId":"2320035","Title":null,"Body":"<p>Theano and Tensorflow will exactly do this for you if you can formulate your optimization-problem within their framework \/ language. These frameworks are also general enough to implement non-NN-based algorithms, like simple first-order-based optimizations like yours.<\/p>\n\n<p>If that's not possible you can try <a href=\"https:\/\/github.com\/HIPS\/autograd\" rel=\"nofollow noreferrer\">autograd<\/a>, which can do this on a subset of numpy. Just formulate your loss as numpy-function (while sticking to supported functions; read the docs) and let autograd build the gradients.<\/p>\n\n<p>Keep in mind, that the somewhat by-construction-approach used by Theano &amp; Tensorflow will be more efficient (because of the more defined input and because these two libraries are a bit more evolved).<\/p>\n"},{"AnswerId":"50435429","CreationDate":"2018-05-20T13:39:39.000","ParentId":null,"OwnerUserId":"9819093","Title":null,"Body":"<p>Both Theano &amp; Tensorflow have built-in differentiation for you. So you only need to form the loss.<\/p>\n"}]}
{"QuestionId":41679830,"AnswerCount":3,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-01-16T15:34:32.853","AcceptedAnswerId":43508402.0,"OwnerUserId":235472.0,"Title":"Can TensorFlow support spiking neurons?","Body":"<p>I looked around for tutorials\/articles\/examples\/... to use spiking neurons (e.g. of the SRM\/Spike Response Model type) in TensorFlow, but I could not find anything.   <\/p>\n\n<p>Is it possible to simulate these models in TensorFlow at all?<br>\nCan TensorFlow simulate models which explicitely depend on time?<br>\nAre there any plug-ins\/extensions\/data files which can add this capability?<br>\nIs the GPU supported?   <\/p>\n","answers":[{"AnswerId":"51554761","CreationDate":"2018-07-27T09:19:55.230","ParentId":null,"OwnerUserId":"4103741","Title":null,"Body":"<p>As pointed out by Steven, Tensorflow is a computation framework and as such allows implementing any algorithm.<\/p>\n\n<p>The main difference between Tensorflow and other computation framework like Matlab or numpy\/scipy is that it relies on computation graphs: you do not perform the operations directly, but instead build a graph of operations that is later evaluated inside a session.<\/p>\n\n<p>I was also interested in Spiking neurons and Tensorflow and found that question. As joti, I implemented the same Matlab exercise in Tensorflow (<a href=\"http:\/\/kaizou.org\/2018\/07\/simulating-spiking-neurons-with-tensorflow\/%20link%20to%20article\" rel=\"nofollow noreferrer\">link to my blog post<\/a>)<\/p>\n\n<p>Here are for instance two operations defining the membrane and recovery factor increments assuming you provide u, v and i:<\/p>\n\n<pre><code>n = 10\nSPIKING_THRESHOLD = 35.0\n\nv = tf.placeholder(tf.float32, shape=[n])\nu = tf.placeholder(tf.float32, shape=[n])\ni = tf.placeholder(tf.float32)\n\n# Evaluate which neurons have reached the spiking threshold\nhas_fired_op = tf.greater_equal(v, tf.constant(SPIKING_THRESHOLD, shape=v.shape))\n\n# Evaluate membrane potential increment for the considered time interval\n# dv = 0 if the neuron fired, dv = 0.04v*v + 5v + 140 + I -u otherwise\ndv_op = tf.where(has_fired_op,\n                 tf.zeros(v.shape),\n                 tf.subtract(tf.add_n([tf.multiply(tf.square(v), 0.04),\n                                       tf.multiply(v, 5.0),\n                                       tf.constant(140.0, shape=v.shape),\n                                       i]),\n                             self.u))\n\n# Evaluate membrane recovery decrement for the considered time interval\n# du = 0 if the neuron fired, du = a*(b*v -u) otherwise\ndu_op = tf.where(has_fired_op,\n                 tf.zeros([v.shape]),\n                 tf.multiply(A, tf.subtract(tf.multiply(B, v), u)))\n<\/code><\/pre>\n\n<p>And you evaluate them like that:<\/p>\n\n<pre><code>with tf.Session() as sess:\n\n    sess.run(tf.global_variables_initializer())\n\n    feed = {u: np.full((n), -13.0), v: np.full((n), -65.0), i : 7.0}\n\n    dv, du = sess.run([dv_op, du_op], feed_dict=feed)\n<\/code><\/pre>\n\n<p>Note that this is just an example to illustrate how Tensorflow works, and not an actual simulation of spiking neuron: usually you want to evaluate also u and v based on synaptic input (in that case, the placeholders will be the synapse inputs).<\/p>\n"},{"AnswerId":"41681132","CreationDate":"2017-01-16T16:44:56.317","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>Yes tensorflow can implement spiking neuron models. It is a general purpose computation framework.<\/p>\n\n<p>Is there an implementation available: I don't think so but I have a friend who is interested in this project.<\/p>\n\n<p>The GPU is supported for many\/most of the tensorflow operations. You'll have to check the docs to see which ones are not supported.<\/p>\n"},{"AnswerId":"43508402","CreationDate":"2017-04-20T00:50:32.120","ParentId":null,"OwnerUserId":"7892753","Title":null,"Body":"<p>I was also interested in this problem and have done exactly what Pietro mentioned. i.e. Took a matlab implementation of a simplified Hodgkin-Huxley model and converted it to Tensorflow. \nHave a look at <a href=\"https:\/\/github.com\/jotia1\/spiking-net-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jotia1\/spiking-net-tensorflow<\/a> <\/p>\n\n<p><s><a href=\"https:\/\/joshuaarnold.com.au\/simulating-spiking-nets-in-tensorflow\/\" rel=\"nofollow noreferrer\">https:\/\/joshuaarnold.com.au\/simulating-spiking-nets-in-tensorflow\/<\/a> for the blog post with some of my thoughts on the whole process.<\/s> <strong>broken link<\/strong><\/p>\n\n<p>Interested in hearing your thoughts on it. <\/p>\n"}]}
{"QuestionId":41683024,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-16T18:40:52.580","AcceptedAnswerId":null,"OwnerUserId":1183980.0,"Title":"resnet with show_and_tell instead of inception_v3","Body":"<p>I have a resnet model trained in caffe. I want to try that with show_and_tell instead of inception_v3 image model. <\/p>\n\n<ol>\n<li>What is the best way to convert the model form caffe to tensorflow\nmodel ? <\/li>\n<li>How would I go about switching the inception model to resnet\n?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"41688590","CreationDate":"2017-01-17T03:23:17.987","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>I'm not sure about the \"best\" way, but there are some tools that seem to address this, like <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ethereon\/caffe-tensorflow<\/a><\/p>\n"}]}
{"QuestionId":41683119,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-16T18:47:20.653","AcceptedAnswerId":null,"OwnerUserId":863543.0,"Title":"Performance: tf.read_file(filename) vs. tf.WholeFileReader()","Body":"<p>Is there a difference in performance (in sense of pre-fetching) between  vs. ?<\/p>\n\n<p>For example, is there a difference in the performance between these two code samples:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks in advance!<\/p>\n","answers":[]}
{"QuestionId":41683248,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network><time-series>","CreationDate":"2017-01-16T18:56:42.247","AcceptedAnswerId":null,"OwnerUserId":7426939.0,"Title":"How to segment\/weigh cost and training data differently based on importance, for neural networks","Body":"<p>I'm looking for ideas\/direction to address a problem I have:<\/p>\n\n<p><strong>Context:<\/strong> I have hourly time-series data (i.e. one row of features for each of the 24 hours in a day), for 1000s of days. For each hour, I'm trying to predict a target value. Now, <strong>I care more about the accuracy of certain hours of the day<\/strong> (e.g. 08:00-18:00) vs. the remaining hours of the day. <\/p>\n\n<p><strong>Problem\/Question:<\/strong> how can I adjust my cost function to be more concerned with the accuracy in these key hours? Or is it a matter of weighing the data set differently, to give more weight to hours 08:00-18:00?<\/p>\n\n<p><strong>Approach:<\/strong> I've coded up a MLP in Tensorflow and am using mean square error as my cost function, and Adam as the optimizer. And while this approach is good for increasing overall accuracy (i.e. >90%), the accuracy in our key hours is not high enough. <\/p>\n\n<p>Here's my cost function:<\/p>\n\n<pre><\/pre>\n\n<p>Do you have any ideas?\nThanks!! :)<\/p>\n","answers":[{"AnswerId":"41684231","CreationDate":"2017-01-16T20:08:07.513","ParentId":null,"OwnerUserId":"2008194","Title":null,"Body":"<p>You can add a placeholder (batch size by 1) that represents how important a particular training example is. Get cost for each row in your batch (batch size by 1), and pointwise multiply your placeholder by your cost.<\/p>\n"}]}
{"QuestionId":41683309,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><conv-neural-network>","CreationDate":"2017-01-16T19:00:43.970","AcceptedAnswerId":41686157.0,"OwnerUserId":7400630.0,"Title":"What is the expected convolution1d_1 to have shape","Body":"<p>I am currently using a covolutional layer to train a neural, but seem to have some problems feeding the input and output dimensions to it.. <\/p>\n\n<p>The input consist of (x,2050) \nand the output is (x,13)\nSo for each row with 2050 columns should a array of 13 elements be created.. <\/p>\n\n<pre><\/pre>\n\n<p>I seem to have problems feeding in the data, i read somewhere that that the  expected a 3d  as input, and output, which is why i reshape it to a 3d.. But for some reason isn't it still not fitting. <\/p>\n\n<p>I am getting this error with the current code: <\/p>\n\n<pre><\/pre>\n\n<p>Why am i not able to start a training session?... <\/p>\n","answers":[{"AnswerId":"41686157","CreationDate":"2017-01-16T22:38:02.237","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So basically - in your network you have:<\/p>\n\n<ul>\n<li>an input which constist of sequences of 1-dimensional feature vectors with length of 2050 each.<\/li>\n<li>you are applying a set of 13 one-dimensional convolutional feature with <code>kernel_size = 3<\/code> what in a <code>same<\/code> border mode is producing sequences of 13-dimensional feature vectors with length of 2050 each. <\/li>\n<\/ul>\n\n<p>This explains why Keras shows you that it's expecting a <code>(None, 2050, 13)<\/code> target shape. <\/p>\n\n<p>There are several ways in order to make it works:<\/p>\n\n<ol>\n<li><p>You could flatten your output from a convolutional layer by adding:<\/p>\n\n<pre><code>from keras.layers import Flatten, Dense \nmodel.add(Flatten())\nmodel.add(Dense(13))\n<\/code><\/pre>\n\n<p>and then also flatten your target by <code>train_set_output_vstacked_normalized_reshaped[train].reshape((221286, 13))<\/code><\/p><\/li>\n<li><p>By adding yet another convolutional layers combined with pooling and different border modes in order to obtain an output sequence with length 13. There are so many possibilites here that I will skip this part.<\/p><\/li>\n<\/ol>\n\n<p>Please notice - that you are using here only activation functions. Whereas in a final <code>dense<\/code> layer it seems to be a good decision, in previous layers it's good to apply some non-linear activations (like <code>relu<\/code>) in order to make your network work better.<\/p>\n"}]}
{"QuestionId":41684446,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-16T20:25:17.783","AcceptedAnswerId":null,"OwnerUserId":7427246.0,"Title":"tensorflow initialize local variable with delay","Body":"<p>My database is a nested folder of images. I am using os.walk and slice input producer to create a filename queue which is fed to whole file readers.<\/p>\n\n<p>The code works fine on a small subset of my database (when the crawling is fast). But throws \"<strong>out of range<\/strong>\" error on a whole one. For a larger database the issue can be solved by setting epoch counter to None.<\/p>\n\n<p>Is there a way to initialize variable with delay\/dependencies ?<\/p>\n\n<p>The minimal example of code is below:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41684986,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-16T21:06:31.510","AcceptedAnswerId":null,"OwnerUserId":5947112.0,"Title":"Tensorflow Translation Tutorial fix for 0.12.1","Body":"<p>I am trying to run the translation tutorial (<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/translate\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/translate<\/a>) on an installation of tensorflow 0.12.1 from the current master branch on github.  <\/p>\n\n<p>I made the changes to seq2seq_model.py suggested by Kangmo here (<a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/853\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/issues\/853<\/a>) but am getting a different error.<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know a way to fix this error?<\/p>\n","answers":[{"AnswerId":"42773284","CreationDate":"2017-03-13T20:58:30.893","ParentId":null,"OwnerUserId":"7705277","Title":null,"Body":"<p>You need to convert your x tensor to float32 so use following code in your seq2seq_model.py:<\/p>\n\n<pre><code>def sampled_loss(inputs,labels):\n            inputs= tf.cast(inputs, tf.float32)\n            labels = tf.reshape(labels, [-1, 1])\n            return tf.cast(tf.nn.sampled_softmax_loss(w_t, b,  labels, inputs,num_samples,\n                    self.target_vocab_size),tf.float32)\n<\/code><\/pre>\n\n<p>You also need to replace this line<\/p>\n\n<pre><code>crossent = softmax_loss_function(target, logit)\n<\/code><\/pre>\n\n<p>with <\/p>\n\n<pre><code>crossent = softmax_loss_function(logit, target)\n<\/code><\/pre>\n\n<p>inside of your seq2seq.py<\/p>\n"}]}
{"QuestionId":41685279,"AnswerCount":1,"Tags":"<tensorflow><distributed-computing>","CreationDate":"2017-01-16T21:30:11.343","AcceptedAnswerId":41686009.0,"OwnerUserId":4394807.0,"Title":"Distributed TensorFlow: about the using of tf.train.Supervisor.start_queue_runners","Body":"<p>I'm looking into the code for <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">distributed inception model in TF<\/a>, in which I have below questions about the use of  in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/inception_distributed_train.py\" rel=\"nofollow noreferrer\">inception_distributed_train.py<\/a>:<\/p>\n\n<ol>\n<li><p>Why do we need to explicitly call  in line\n and line  in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/inception_distributed_train.py\" rel=\"nofollow noreferrer\">inception_distributed_train.py<\/a>? In <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/distributed_execution#Supervisor.start_queue_runners\" rel=\"nofollow noreferrer\">API\ndoc. of start_queue_runners<\/a>, I see there is no need for such\ncalls due to:<\/p>\n\n<blockquote>\n  <p>Note that the queue runners collected in the graph key QUEUE_RUNNERS\n  are already started automatically when you create a session with the\n  supervisor, so unless you have non-collected queue runners to start\n  you do not need to call this explicitly.<\/p>\n<\/blockquote><\/li>\n<li><p>I noticed the values of  in calling\n are different in line  and line\n in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/inception_distributed_train.py\" rel=\"nofollow noreferrer\">inception_distributed_train.py<\/a>. But aren't\n also in the collection of\n (all  are obtained in line )? If\nso, then there is no need for line  since the  has already\nbeen started in line .<\/p><\/li>\n<li><p>Besides, could you please explain to me or show me some references about what queues are created in ?<\/p><\/li>\n<\/ol>\n\n<p>Thanks for your time!<\/p>\n","answers":[{"AnswerId":"41686009","CreationDate":"2017-01-16T22:26:37.633","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Not an answer, but some general notes how to find an answer :)<\/p>\n\n<p>First of all, using github's blame, inception_distributed was checked in on April 13, while that comment in <code>start_queue_runners<\/code> was added on Apr 15th, so it's possible that functionality was changed but didn't get updated in all the places that use it.<\/p>\n\n<p>You could comment-out that line and see if things still work. And if not, you could add <code>import pdb; pdb.set_trace()<\/code> in the place where queue runner gets created (ie <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/dd01113c2b9efb16c3c3ea0a4aaa340a0494696f\/tensorflow\/python\/training\/queue_runner_impl.py#L47\" rel=\"nofollow noreferrer\">here<\/a>) and see who is creating those extra unattended queue runners.<\/p>\n\n<p>Also, Supervisor development seems to have slowed down and things are getting moved over to FooSession (from comment <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6604#issuecomment-270950324\" rel=\"nofollow noreferrer\">here<\/a>). Those provide a more robust training architecture (your workers won't crash because of temporary network error), but there are not many examples on how to use them yet.<\/p>\n"}]}
{"QuestionId":41685366,"AnswerCount":5,"Tags":"<tensorflow>","CreationDate":"2017-01-16T21:36:02.260","AcceptedAnswerId":null,"OwnerUserId":826983.0,"Title":"Getting 'undefined symbol: cudnnCreate'","Body":"<p>I am trying to run a quite simple Tensorflow graph, but as I run the script I am getting the following output:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea what the problem is here? <\/p>\n\n<p>The strange thing is that I am able to run the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_softmax.py\" rel=\"nofollow noreferrer\">MNIST softmax example<\/a> without any errors.<\/p>\n\n<p>This is the script which I am getting the error from:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<hr>\n\n<p><strong>Update:<\/strong> I followed <a href=\"https:\/\/alliseesolutions.wordpress.com\/2016\/09\/08\/install-gpu-tensorflow-from-sources-w-ubuntu-16-04-and-cuda-8-0-rc\/\" rel=\"nofollow noreferrer\">these instructions<\/a> to install CudNN but I am still getting the same error ..<\/p>\n","answers":[{"AnswerId":"46112191","CreationDate":"2017-09-08T08:38:06.630","ParentId":null,"OwnerUserId":"7718729","Title":null,"Body":"<p>All above answer is right, that you need to set the enviroment right. <\/p>\n\n<p>But if you install the wrong version of cudnn (even if it is the newer version), you will get this warning. You must install the right version of cuda and cudnn. You may follow the instructions on tensorflow.org, no TF re-compile is needed, just cp cudnn files to \/user\/local\/cuda\/*, and try again:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup#optional_install_cuda_gpus_on_linux\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup#optional_install_cuda_gpus_on_linux<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/erMbj.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/erMbj.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"43915267","CreationDate":"2017-05-11T12:21:21.977","ParentId":null,"OwnerUserId":"5295306","Title":null,"Body":"<p>I solved the same problem by setting the export path for CUDA and CUDA\/lib.<\/p>\n"},{"AnswerId":"52422117","CreationDate":"2018-09-20T09:43:41.367","ParentId":null,"OwnerUserId":"6573154","Title":null,"Body":"<p><strong>There are two ways to solve this problem (I faced it too)-<\/strong> <br\/>\n1)Either cudnn is not installed properly, to check this go to Path- \/usr\/local\/cuda\/lib64 if you are unable to find libcudnn.so then this working properly. Also check .bashrc file whether LD_LIBRARY_PATH is set to the above path. If its not set it. <br\/>\n2)The problem may be with Tensorflow version (occurs in most of the cases). Try to update tensorflow from tensorflow website. I faced the same issue and this was the reason.<\/p>\n"},{"AnswerId":"42979756","CreationDate":"2017-03-23T14:59:48.863","ParentId":null,"OwnerUserId":"1838346","Title":null,"Body":"<p>I got this error in windows environment<\/p>\n\n<pre><code>Check failed: s.ok() could not find cudnnCreate in cudnn DSO; dlerror: cudnnCreate not found\n<\/code><\/pre>\n\n<p><strong>Assumption<\/strong>: You have already downloaded tensorflow with GPU support and have installed CUDA toolkit from NVIDIA from <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-downloads<\/a><\/p>\n\n<p>I followed the below steps to resolve it.<\/p>\n\n<ol>\n<li>Download cuDNN libraries from NVIDIA from the URL. <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cudnn<\/a><\/li>\n<li>Create a free account, answer all the questions while creating the account.<\/li>\n<li><p>You should be able to download the zip file for respective platform. See below for the options\n<a href=\"https:\/\/i.stack.imgur.com\/srCbY.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/srCbY.png\" alt=\"enter image description here\"><\/a><\/p><\/li>\n<li><p>Since I am using Windows 8.1. I downloaded library for Windows 10. It is zip file which contains three folders -> bin, inlcude and lib folders.<\/p><\/li>\n<li>Extract the files to a folder<\/li>\n<li>With default CUDA toolkit installation would be something in the folder -> C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA<\/li>\n<li>Copy the files from step 6 to the location given in the step 6. All files from bin go into bin folder in step 6 and so on.<\/li>\n<\/ol>\n\n<p>These steps solved the error given above. Hopefully, this is helpful for someone.<\/p>\n"},{"AnswerId":"50370274","CreationDate":"2018-05-16T11:55:02.417","ParentId":null,"OwnerUserId":"4595522","Title":null,"Body":"<p>I solved it bu updating to tensorflow-gpu to 1.4. I was getting this in a conda environment. Updating tf-gpu solved the error.<\/p>\n"}]}
{"QuestionId":41685830,"AnswerCount":1,"Tags":"<tensorflow><computer-vision><image-recognition>","CreationDate":"2017-01-16T22:10:59.593","AcceptedAnswerId":null,"OwnerUserId":5997806.0,"Title":"Inception V3 Running Stats","Body":"<p>Has anyone created statistics on how fast and accurate Inception V3 can classify an image based on criteria such as: different models of GPUs\/CPUs, input image size, input image ratio, file format, etc?<\/p>\n\n<p>Example chart I'd hope to see:<\/p>\n\n<pre><\/pre>\n\n<p>X as in type of GPU\/ CPU<\/p>\n","answers":[{"AnswerId":"43854292","CreationDate":"2017-05-08T17:56:04.543","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>Recently, TensorFlow released some scaling benchmarks for convolutional neural networks including Inception at: <a href=\"https:\/\/www.tensorflow.org\/performance\/benchmarks\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/performance\/benchmarks<\/a><\/p>\n\n<p>Additionally, we have also released a performance guide to help you optimize your models and systems based on our current best practices: <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/performance\/performance_guide<\/a> and <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_models\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/performance\/performance_models<\/a> <\/p>\n\n<p>One important thing to note: depending on your system, you may achieve different results and optimal configurations. It's important to benchmark your own models on your hardware configurations to determine optimal configurations.<\/p>\n"}]}
{"QuestionId":41685954,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-16T22:21:50.497","AcceptedAnswerId":41687183.0,"OwnerUserId":1691278.0,"Title":"Tensorflow - show the gradients on all the neurons (not just the input variables)","Body":"<p>I am implementing a neural network using the Adam optimizer which utilizes gradients. Here is a brief look on my code<\/p>\n\n<pre><\/pre>\n\n<p>And this is how I feed in the data:<\/p>\n\n<pre><\/pre>\n\n<p>However, I would like to show gradients on all the neurons in the hidden layers (i.e. all the gradients used in the backprop). I tried to look up the relevant function, but all the functions that I found only compute the gradients with respect to the input variables. Can I please get a hint as to how to proceed?<\/p>\n","answers":[{"AnswerId":"41687183","CreationDate":"2017-01-17T00:26:22.197","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>See the builtin debugger of TensorFlow called <code>tfdbg<\/code> (available in versions 0.12+). It exposes all the intermediate Tensors (along with graph structure) during <code>Session.run()<\/code> calls.<\/p>\n\n<p>See a walk-through at: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/debugger\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/debugger\/<\/a><\/p>\n\n<p>The basic workflow is as follows. First you wrap your Session object right before the training Session.run() is called.<\/p>\n\n<pre><code>from tensorflow.python import debug as tf_debug\nsess = tf_debug.LocalCLIDebugWrapperSession(sess)\n<\/code><\/pre>\n\n<p>Then when the sess.run() occurs, you will drop to the <code>tfdbg&gt;<\/code> command line interface. Issue the <code>tfdbg&gt; run<\/code> command and you will then see a list of all intermediate Tensors generated during the <code>sess.run()<\/code> call, including the gradients.<\/p>\n\n<p>If you want to access the gradients programmatically, instead of interactively, <code>tfdbg<\/code> has a Python API available as well. See:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf_debug\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf_debug\/<\/a><\/p>\n"}]}
{"QuestionId":41685968,"AnswerCount":1,"Tags":"<python><generator><keras>","CreationDate":"2017-01-16T22:23:15.417","AcceptedAnswerId":null,"OwnerUserId":5133167.0,"Title":"What is an epoch when using keras generators?","Body":"<p>I want to use fit_generator to handle my data.<\/p>\n\n<p>I understand the generator must be running forever, and that samples_per_epoch defines the number of elements yielded from the generator before getting to the next epoch.<\/p>\n\n<p>But what is an epoch here ? When we run fit, an epoch is a run against the whole data set, split according to batch_size. But here, batch_size has no meaning since fit_generator works like calling train_on_batch on each element of the generator, and there is no such thing as \"the whole dataset\".<\/p>\n\n<p>My question is: do samples_per_epoch and nb_epoch have a meaning? Or in any case the network is trained the same way on samples_per_epoch*nb_epoch batches returned by the generator, and the epoch has no real meaning?<\/p>\n","answers":[{"AnswerId":"41686333","CreationDate":"2017-01-16T22:53:51.473","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p><code>fit_generator<\/code> is blind to what data is produced by the generator. Its the responsibility of the generator class to go over the whole dataset. Using <code>samples_per_epoch<\/code> the <code>fit_generator<\/code> simply keeps track of the epoch count. Check <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/engine\/training.py#L1473-L1477\" rel=\"nofollow noreferrer\">here<\/a> &amp; <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/engine\/training.py#L1532-LL1541\" rel=\"nofollow noreferrer\">here<\/a> &amp; <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/engine\/training.py#L1565\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":41686252,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-01-16T22:46:29.363","AcceptedAnswerId":null,"OwnerUserId":712997.0,"Title":"PyCaffe - changing script to take 10 images as input instead of one","Body":"<p>I have a project where I want to use Yahoo's OpenNSFW net in a python script, but as is, the script takes only a single image example and takes about 270ms to compute the forward pass (a little too slow). <\/p>\n\n<p>Amortizing it over say 50 images I think would be faster, but I wasn't sure if I could do this with just the deployprototxt doc. <\/p>\n\n<p>I changed the deploy.prototxt doc here by changing dim 1 -> 10 like so:<\/p>\n\n<pre><\/pre>\n\n<p>Now I need a way to <a href=\"https:\/\/github.com\/yahoo\/open_nsfw\/blob\/master\/classify_nsfw.py\" rel=\"nofollow noreferrer\">change it in the Python script it uses<\/a> code here:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a simple way to do this?<\/p>\n","answers":[{"AnswerId":"41690906","CreationDate":"2017-01-17T07:06:59.747","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can see an example of feeding multiple images to classifier in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/classifier.py#L47-L91\" rel=\"nofollow noreferrer\"><code>classifier.py<\/code><\/a> example script provided with caffe.<\/p>\n\n<p>You basically need to make your <code>transformed_image<\/code> a 4D array, with the different images stacked along the 0-th axis.<\/p>\n"}]}
{"QuestionId":41686280,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-16T22:48:38.327","AcceptedAnswerId":41687295.0,"OwnerUserId":3444294.0,"Title":"Subtract columns in Tensorflow tensor","Body":"<p>In Tensorflow, I want to subtract one column in a 2D tensor from the other column. I looked at using  or  to split into 2 tensors and then subtract, but this seemed unnecessarily complex. My current approach is to multiply one column by -1 and then  the columns:<\/p>\n\n<pre><\/pre>\n\n<p>Still seems needlessly complex. Is there a simpler way of doing this?<\/p>\n","answers":[{"AnswerId":"41687295","CreationDate":"2017-01-17T00:37:21.380","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>Lot of <code>numpy<\/code>'s array indexing work as expected in TensorFlow. The following works :<\/p>\n\n<pre><code>input = tf.constant(\n    [[5.8, 3.0],\n     [4.0, 6.0],\n     [7.0, 9.0]])\nsess = tf.InteractiveSession()\nans = input[:, :1] - input[:, 1:]\nprint(ans.eval())\n\narray([[ 2.80000019],\n   [-2.        ],\n   [-2.        ]], dtype=float32)\n<\/code><\/pre>\n"}]}
{"QuestionId":41688217,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-17T02:35:37.047","AcceptedAnswerId":41688506.0,"OwnerUserId":7428227.0,"Title":"How to load a graph with tensorflow.so and c_api.h in c++ language?","Body":"<p>I am not able to find any examples about how to load a graph with  and  in C++. I read the , however the  function was not in it. How can I load a graph without the  function?<\/p>\n","answers":[{"AnswerId":"41688506","CreationDate":"2017-01-17T03:12:26.870","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>If you're using C++, you might want to use the C++ API instead. The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/label_image\" rel=\"noreferrer\">label image example<\/a> would probably be a good sample to help you start.<\/p>\n\n<p>If you really want to use just the C API, use <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/e8f2aad\/tensorflow\/c\/c_api.h#L805\" rel=\"noreferrer\"><code>TF_GraphImportGraphDef<\/code><\/a> to load a graph. Note that the C API isn't particularly convenient to use (it is intended to build bindings in other languages such as Go, Java, Rust, Haskell etc.) For example:<\/p>\n\n<pre><code>#include &lt;stdio.h&gt;                                                                        \n#include &lt;stdlib.h&gt;                                                                       \n#include &lt;tensorflow\/c\/c_api.h&gt;                                                           \n\nTF_Buffer* read_file(const char* file);                                                   \n\nvoid free_buffer(void* data, size_t length) {                                             \n        free(data);                                                                       \n}                                                                                         \n\nint main() {                                                                              \n  \/\/ Graph definition from unzipped https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\n  \/\/ which is used in the Go, Java and Android examples                                   \n  TF_Buffer* graph_def = read_file(\"tensorflow_inception_graph.pb\");                      \n  TF_Graph* graph = TF_NewGraph();\n\n  \/\/ Import graph_def into graph                                                          \n  TF_Status* 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 1;\n  }       \n  fprintf(stdout, \"Successfully imported graph\");                                         \n  TF_DeleteStatus(status);\n  TF_DeleteBuffer(graph_def);                                                             \n\n  \/\/ Use the graph                                                                        \n  TF_DeleteGraph(graph);                                                                  \n  return 0;\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<\/code><\/pre>\n"},{"AnswerId":"41703148","CreationDate":"2017-01-17T17:14:20.233","ParentId":null,"OwnerUserId":"4367137","Title":null,"Body":"<p>The previous answer is your main option if you are wanting to use it outside of the TensorFlow project (and consequently not build with Bazel). You <strong>need<\/strong> to load it from the c_api.h with TF_GraphImportDef, I recommend training and doing testing in Python and then exporting the model\/graph for use with C++\/C Api when you have finished.<\/p>\n"}]}
{"QuestionId":41688330,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-17T02:50:53.100","AcceptedAnswerId":null,"OwnerUserId":6662040.0,"Title":"Is there an alternative method to apply tf.case, tf.cond, etc. on a tensor?","Body":"<p>I'm trying to use tf.case to use a index value in a tensor to direct to different network structure part, get different losses and then sum them up as the final loss for training. Take a simple example, I judge the value in a list and output a different value. For example\n[0,1,2,3] -> [0,7,10,13] where \ncase 0:\noutput 0\ncase 1:\noutput 7\ncase 2:\noutput 10\ncase 3:\noutput 13.\nHowever, the tf.cond, tf.case seem only to be used on a scalar. How to fulfill the goal?<\/p>\n","answers":[{"AnswerId":"43095070","CreationDate":"2017-03-29T13:34:24.700","ParentId":null,"OwnerUserId":"1232944","Title":null,"Body":"<p>The only operation I'm aware of that evaluates a condition separately on each element of a vector is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/where\" rel=\"nofollow noreferrer\">tf.where<\/a>.  You would leave <code>x=None, y=None<\/code>:<\/p>\n\n<pre><code>t_orig = tf.constant([0, 1, 2, 3, 1])\nt_filt = tf.where(tf.equal(t_orig, 1))\nwith tf.Session() as sess:\n    print sess.run(t_filt)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[[1]\n [4]]\n<\/code><\/pre>\n\n<p>However, this only evaluates the truth of a single condition.  If you want to evaluate the truth of multiple conditions, over each element of a vector, I think you'll have to use <code>tf.map_fn<\/code> combined with <code>tf.case<\/code>.  AFAIK, <code>tf.case<\/code> is the only operation that evaluates the truth of many conditions on a given value:<\/p>\n\n<pre><code>t_orig = tf.constant([0, 1, 2, 3])\nt_new = tf.map_fn(\n        lambda x: tf.case(\n            pred_fn_pairs=[\n                (tf.equal(x, 0), lambda: tf.constant(0)),\n                (tf.equal(x, 1), lambda: tf.constant(7)),\n                (tf.equal(x, 2), lambda: tf.constant(10)),\n                (tf.equal(x, 3), lambda: tf.constant(13))],\n            default=lambda: tf.constant(-1)),\n        t_orig)\nwith tf.Session() as sess:\n    print sess.run(t_new)\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[ 0  7 10 13]\n<\/code><\/pre>\n"},{"AnswerId":"57337229","CreationDate":"2019-08-03T09:56:07.123","ParentId":null,"OwnerUserId":"2820592","Title":null,"Body":"<p>try this<\/p>\n\n<pre><code>import tensorflow as tf\nvalue = [0, 1, 2, 3]\nones = tf.ones_like(value)\nout = tf.where(tf.equal(value, 0), ones * 0,\n               tf.where(tf.equal(value, 1), ones * 7,\n                        tf.where(tf.equal(value, 2), ones * 10,\n                                 tf.where(tf.equal(value, 3), ones * 13, ones * -1\n                                          )\n                                 )\n                        )\n               )\n\nwith tf.Session() as sess:\n    print(sess.run(out)) # [ 0  7 10 13]\n<\/code><\/pre>\n"}]}
{"QuestionId":41688532,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-17T03:14:48.413","AcceptedAnswerId":41707423.0,"OwnerUserId":6472284.0,"Title":"Efficient element-wise multiplication of a matrix and a matrix in Tensorflow","Body":"<p>What would be the most efficient way to multiply:\nMatrix A(shape: 4m * n, shape of A1: m * n)<\/p>\n\n<pre><\/pre>\n\n<p>by a matrix(shape: 4m * n, shape of B1: m * n):<\/p>\n\n<pre><\/pre>\n\n<p>to obtain a new matrix(shape: 4m * m):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41707423","CreationDate":"2017-01-17T21:35:03.347","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>This is how I am interpreting your question as follows :\nYou have matrices <code>A1, A2, A3, A4<\/code> and <code>B1, B2, B3, B4<\/code>, all of which have sizes <code>m x n<\/code>. You want to form a <code>4m x n<\/code> matrix which is formed by element-wise multiplication of the corresponding matrices, stacked together in the first dimension. Here is how you can do it :<\/p>\n\n<pre><code>tf.concat_v2([A1*B1, A2*B2, A3*B3, A4*B4], 0)\n<\/code><\/pre>\n"}]}
{"QuestionId":41688839,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><nlp><deep-learning><word-embedding>","CreationDate":"2017-01-17T03:56:33.140","AcceptedAnswerId":null,"OwnerUserId":7428399.0,"Title":"Large word embedding matrix update in tensorflow","Body":"<p>Recently I am using a CNN to do text classification which is described in <a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"nofollow noreferrer\">http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/<\/a>. <\/p>\n\n<p>My dataset is quite large and the vocabulary size is more than 1M words. My training algorithm becomes much slower when the vocabulary size gets bigger. There is a warning message saying that  \"....tensorflow\/python\/ops\/gradients.py:87: UserWarning: Converting sparse IndexedSlices to a dense Tensor with 145017088 elements. This may consume a large amount of memory.\" <\/p>\n\n<p>I am thinking this may be caused by a dense gradient update on the embedding matrix. Does anyone have any ideas on that? There is a post <a href=\"https:\/\/stackoverflow.com\/questions\/35803425\/update-only-part-of-the-word-embedding-matrix-in-tensorflow\">Update only part of the word embedding matrix in Tensorflow<\/a> discussing similar issue. The top answer states that tensorflow only updates specific rows. Why did it change in my case?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[]}
{"QuestionId":41689451,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-17T05:10:54.123","AcceptedAnswerId":null,"OwnerUserId":7428590.0,"Title":"ValueError: No gradients provided for any variable","Body":"<p>I'm facing a trouble with tensorFlow. Executing the following code <\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<p><strong>ValueError: No gradients provided for any variable, check your graph for ops that do not support gradients, between variables ['Tensor(\"Variable\/read:0\", shape=(784, 10), dtype=float32)', 'Tensor(\"Variable_1\/read:0\", shape=(10,), dtype=float32)'] and loss Tensor(\"Mean:0\", shape=(), dtype=float32).<\/strong><\/p>\n","answers":[{"AnswerId":"43578835","CreationDate":"2017-04-24T02:24:21.130","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>This problem is caused by the following line: <code>tf.nn.softmax_cross_entropy_with_logits(labels=activation, logits=Y)<\/code><\/p>\n\n<p>Based on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/softmax_cross_entropy_with_logits\" rel=\"noreferrer\">documentation<\/a> you should have<\/p>\n\n<blockquote>\n  <p>labels: Each row labels[i] must be a valid probability distribution.<\/p>\n  \n  <p>logits: Unscaled log probabilities.<\/p>\n<\/blockquote>\n\n<p>So logits suppose to be your hypothesis and thus equal to <code>activation<\/code> and valid probability distribution is <code>Y<\/code>. So just change it with <code>tf.nn.softmax_cross_entropy_with_logits(labels=Y, logits=activation)<\/code><\/p>\n"}]}
{"QuestionId":41689566,"AnswerCount":1,"Tags":"<python><session><machine-learning><tensorflow>","CreationDate":"2017-01-17T05:22:59.463","AcceptedAnswerId":41689694.0,"OwnerUserId":1691278.0,"Title":"Tensorflow - does keyboard interrupt close the current session","Body":"<p>I am using Tensorflow to test a neural network. This is an excerpt of my code:<\/p>\n\n<pre><\/pre>\n\n<p>If I use keyboard interrupt (such as ), the program stops but it seems like the session is closed as well. For example, if I pass , I would receive the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I suppose that means my session is closed? How can I interrupt the program without closing the session?<\/p>\n","answers":[{"AnswerId":"41689694","CreationDate":"2017-01-17T05:34:54.870","ParentId":null,"OwnerUserId":"6850023","Title":null,"Body":"<p>When you press ctrl-c an interrupt will be generated, which will cause execution to leave the with-block. This will cause the session to be closed, since the whole purpose of the with-block is to do automatic tear-down, see e.g. <a href=\"http:\/\/effbot.org\/zone\/python-with-statement.htm\" rel=\"nofollow noreferrer\">this explanation<\/a> (just the first hit on google for me).<\/p>\n\n<p>Thus replacing<\/p>\n\n<pre><code>with tf.Session() as sess:\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>sess = tf.Session()\n<\/code><\/pre>\n\n<p>should solve the issue.<\/p>\n"}]}
{"QuestionId":41689670,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-17T05:32:17.670","AcceptedAnswerId":null,"OwnerUserId":4482655.0,"Title":"Dropout in Tensorflow CNN network while using triplet based training","Body":"<p>I have implemented distance learning using triplets of data using a CNN in Tensorflow. For this I have a shared set of weights for each layer. I want to employ dropout  technique to compensate for over-fitting. I was thinking of doing this in the following way:<\/p>\n\n<pre><\/pre>\n\n<p>Here, <em>h_fci<\/em> are the outputs for the three data samples from a previous layer. This poses a potential problem in the sense that dropout is probabilistic hence even if I am providing the same dropout probability, it is possible that different set of neurons are dropped for each of the three outputs whereas ideally I would want the same set of neurons to be used for each. What would be the proper way to implement dropout in this context then ? <\/p>\n","answers":[{"AnswerId":"42683351","CreationDate":"2017-03-08T22:43:49.437","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Assuming everything has the same shape, setting the same seed should make them all identical.<\/p>\n\n<p>In practice though data samples are minibatched, and <code>tf.nn.dropout<\/code> will drop out the same units in the entire minibatch, so switching to stacking your examples is probably a good idea here.<\/p>\n"}]}
{"QuestionId":41690582,"AnswerCount":1,"Tags":"<tensorflow><anaconda><python-3.5><ubuntu-16.04><mnist>","CreationDate":"2017-01-17T06:43:11.047","AcceptedAnswerId":null,"OwnerUserId":4611374.0,"Title":"Python can't import existing TensorFlow module","Body":"<h1>Unable to import an existing module, using Ubuntu 16.04, python 3.5, Anaconda 4.2.15, tensorflow 0.12.<\/h1>\n\n<p>The TensorFlow package contains a helper <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\" rel=\"nofollow noreferrer\">module<\/a> called . I've verified that the file exists on my machine, and manually added the directory to . When I try to import the module, python can't seem to find it. I can  just fine, but python seems to be looking for the module in the wrong directory although I've explicitly given it the correct location:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Verifying the file<\/strong>  <\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Verifying the path<\/strong>  <\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Tensorflow Info<\/strong>  <\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Only 1 Python Environment<\/strong>  <\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>My source code comes from a tensorflow tutorial that runs perfectly on OSX, something about the Ubuntu install\/environment is wonky. \nI've been beating on this problem for a while without success. This related <a href=\"https:\/\/stackoverflow.com\/questions\/35236851\/tensorflow-no-module-named-example-tutorials-mnist-input-data\">question<\/a> doesn't address the issue either. Any suggestions?<\/p>\n","answers":[{"AnswerId":"41749291","CreationDate":"2017-01-19T18:38:25.643","ParentId":null,"OwnerUserId":"4611374","Title":null,"Body":"<p>Found a clue in the answer to this <a href=\"https:\/\/stackoverflow.com\/questions\/36854940\/tensorflow-gpu-is-cudnn-optional-couldnt-open-cuda-library-libcudnn-so\">related question<\/a>, and @gunan's comment.<\/p>\n\n<p>Although the machine in question doesn't have a python-gpu interface, python kept looking for cudnn files. <code>tensorflow 0.12.0rc0<\/code> (installed from a private <a href=\"https:\/\/anaconda.org\/jjhelmus\/tensorflow\" rel=\"nofollow noreferrer\">repo<\/a>) offers <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md\" rel=\"nofollow noreferrer\">gpu support by default<\/a>, and I couldn't disable it by resetting environment variables.<\/p>\n\n<p>The solution was to remove <code>tensorflow 0.12.0rc0<\/code> and replace it with <code>tensorflow 0.10.0rc0<\/code> from the default Anaconda repositories.<\/p>\n\n<pre><code>$ conda remove tensorflow\n$ conda install tensorflow\n$ conda list | grep tensorflow  \ntensorflow                0.10.0rc0           np111py35_0\n\n\n            ----------\n<\/code><\/pre>\n\n<h3>Warning<\/h3>\n\n<p>Several tensorflow functions were deprecated between versions 10 and 12. At present time, version 12 is the recommended tensorflow package for OSX Anaconda Python (installed from private repos with Anaconda pip). Any script that needs to run on both Mac and <strong>non-gpu<\/strong> Linux should use the deprecated version 10 functions. Mac users will need to suppress the resulting warnings.<\/p>\n"}]}
{"QuestionId":41690885,"AnswerCount":1,"Tags":"<neural-network><deep-learning><keras><gensim><word2vec>","CreationDate":"2017-01-17T07:05:39.923","AcceptedAnswerId":41696259.0,"OwnerUserId":1637673.0,"Title":"How to use word2vec with keras CNN (2D) to do text classification?","Body":"<p>There is Convolution1D example <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/imdb_cnn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/imdb_cnn.py<\/a> without word2vec.<\/p>\n\n<p>Currently, I am using gensim to train word2vec model.<\/p>\n\n<p>I want to use word2vec and keras cnn(2D not 1D) to do document classifacation(Chinese Text). I learned the basic flow of text classification in cnn and want to do a test.<\/p>\n\n<h2>For example(the steps I imagine):<\/h2>\n\n<ol>\n<li><p>Use a good Cinese Tokenized Text Set to train word2vec model<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Tokenize my sentences dataset to words lists dataset(the longest sentence has over 8000 words, shortest is less 50)<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Use a method to transform words lists dataset to word2vec dataset<\/p>\n\n<p>transform every word in every sencence to a vec by trained model.<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Pad  word2vec dataset (with size=200 zero array)<\/p>\n\n<pre><\/pre><\/li>\n<li><p>go to the  CNN (using Convolution2D)<\/p><\/li>\n<\/ol>\n\n<hr>\n\n<p>I search for a long time, but can't find any way to do the step 3 ( after step 3,  the parameter and layers setting in step 5 is hard to understand too). <\/p>\n","answers":[{"AnswerId":"41696259","CreationDate":"2017-01-17T11:43:53.630","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p><strong>Transforming a single sentence to a 2D vector<\/strong><\/p>\n\n<p>Assuming you have a list of words and a model you can do:<\/p>\n\n<pre><code>import numpy as np\nsentence_vec = None\nfor word in sentence:\n    word_vec = np.expand_dims(model[word], axis=0)\n    if sentence_vec is None:\n        sentence_vec = word_vec\n    else:\n        sentence_vec = np.concatenate((sentence_vec, word_vec), axis=0)\n<\/code><\/pre>\n\n<p>As for step 5 - it would be helpful if you listed what you are having trouble with. Basically you only need to do is change both 1D operations (Convolution1D, GlobalMaxPooling1D) to their 2D counter-parts .<\/p>\n"}]}
{"QuestionId":41691969,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-01-17T08:13:22.207","AcceptedAnswerId":null,"OwnerUserId":5257450.0,"Title":"Can the input layer of keras take customized input?","Body":"<p>In the documentation, the parameters of  are<\/p>\n\n<p><\/p>\n\n<p>My question is, can I have three different tensors as input therefore I can have something like ?<\/p>\n\n<p>PS:<\/p>\n\n<p>Sorry for the ambiguity. I believe the  function in keras acts similarly as , I am just wondering can I feed in more values I created in the model?<\/p>\n","answers":[{"AnswerId":"41692172","CreationDate":"2017-01-17T08:24:35.737","ParentId":null,"OwnerUserId":"5241407","Title":null,"Body":"<p>This X is a tensor\nFor example\n<code>\nX = [[1,2],[2,3],[3,4]]\ny = [3,5,7]\n<\/code>\n(The fitting line should be <code>y=X[0]+X[1]<\/code>)<\/p>\n\n<p>So if you clear which are features (for example, the x and y in your requirement) and which is the target class (z), you can define:<br>\n<code>_X = (x,y) # two nodes in the input layer<\/code><br>\n<code>_y = z<\/code><br>\n<code>fit(_X,_y)<\/code><br>\nDo you mean it?<\/p>\n"}]}
{"QuestionId":41692257,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><nvidia>","CreationDate":"2017-01-17T08:29:13.523","AcceptedAnswerId":null,"OwnerUserId":6825835.0,"Title":"High GPU Memory-Usage but zero volatile gpu-util","Body":"<p>My new training code occupied high GPU Memory-Usage but zero volatile gpu-util<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/omOcF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/omOcF.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><\/pre>\n\n<p>I'm trying a new data generator, and the old one run with volatile gpu-util of 10%, so the environment or the tensorflow version won't be the key problem <\/p>\n\n<p>so I wonder if anyone could tell me what element in the code could possibly cause this problem?<\/p>\n\n<p>thank you very much\uff01<\/p>\n","answers":[{"AnswerId":"41693933","CreationDate":"2017-01-17T09:54:17.900","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>Tensorflow by default pre-allocates all the available VRAM when you create a session, whether it is actually needed or not. 0% volatile GPU-util only shows that nothing is actually \"running\" on the GPUs(no active kernels).<\/p>\n\n<p>If you don't want this to happen, you can set <code>allow_growth<\/code> to <code>True<\/code> and pass it appropriately in a <code>config<\/code> object when you create your session.<\/p>\n\n<p>Read <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session.__init__\" rel=\"nofollow noreferrer\">tf.Session<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/protobuf\/config.proto\" rel=\"nofollow noreferrer\">config.proto<\/a> for more details.<\/p>\n\n<p>EDIT: This is how you set <code>allow_growth<\/code> to <code>True<\/code><\/p>\n\n<pre><code>session = tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True)))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6AdEL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6AdEL.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":41694157,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-17T10:04:56.983","AcceptedAnswerId":null,"OwnerUserId":7429725.0,"Title":"LSUV init in tensorflow","Body":"<p>I am trying to implement LSUV init as in :<\/p>\n\n<p><a href=\"https:\/\/github.com\/ducha-aiki\/LSUVinit\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ducha-aiki\/LSUVinit<\/a>\n<a href=\"https:\/\/github.com\/ducha-aiki\/LSUV-keras\/blob\/master\/lsuv_init.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ducha-aiki\/LSUV-keras\/blob\/master\/lsuv_init.py<\/a><\/p>\n\n<p>in tensorflow and I encounter some difficulties.\nBasically, the algorithm should be something like :<\/p>\n\n<pre><\/pre>\n\n<ol>\n<li>Is there any way to iterate over each layer in the network ? <\/li>\n<li>How to get the activation for each layer after a single pass ?<\/li>\n<li>I was trying to calculate the variance of each layer with tf.nn.moments but this function return a tensor, how to evaluate the value of this tensor ?<\/li>\n<li>I am wondering what is better to do : to build a different inference graph for initialization and training phase or to run some \"training\" in order to initialized the model and load it before the real training .Any advises ?<\/li>\n<\/ol>\n","answers":[]}
{"QuestionId":41695117,"AnswerCount":1,"Tags":"<deep-learning><keras><lstm><recurrent-neural-network><stateful>","CreationDate":"2017-01-17T10:48:57.187","AcceptedAnswerId":null,"OwnerUserId":6893857.0,"Title":"Understanding stateful LSTM","Body":"<p>I'm going through this <a href=\"http:\/\/machinelearningmastery.com\/understanding-stateful-lstm-recurrent-neural-networks-python-keras\/\" rel=\"noreferrer\">tutorial<\/a> on RNNs\/LSTMs and I'm having quite a hard time understanding stateful LSTMs. My questions are as follows : <\/p>\n\n<h1>1. Training batching size<\/h1>\n\n<p>In the Keras docs on <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"noreferrer\">RNNs<\/a>, I found out that the hidden state of the sample in -th position within the batch will be fed as input hidden state for the sample in -th position in the next batch. Does that mean that if we want to pass the hidden state from sample to sample we have to use batches of size 1 and therefore perform online gradient descent? Is there a way to pass the hidden state within a batch of size >1 and perform gradient descent on that batch ?<\/p>\n\n<h1>2. One-Char Mapping Problems<\/h1>\n\n<p>In the tutorial's paragraph 'Stateful LSTM for a One-Char to One-Char Mapping' were given a code that uses  and  to learn to predict the next letter of the alphabet given a letter of the alphabet. In the last part of the code (line 53 to the end of the complete code), the model is tested starting with a random letter ('K') and predicts 'B' then given 'B' it predicts 'C', etc. It seems to work well except for 'K'. However, I tried the following tweak to the code (last part too, I kept lines 52 and above): <\/p>\n\n<pre><\/pre>\n\n<p>Therefore, how does keeping the previous hidden state as initial hidden state for the current hidden state help us with the learning given that during test if we start with the letter 'K' for example, letters A to J will not have been fed in before and the initial hidden state won't be the same as during training ?  <\/p>\n\n<h1>3. Training an LSTM on a book for sentence generation<\/h1>\n\n<p>I want to train my LSTM on a whole book to learn how to generate sentences and perhaps learn the authors style too, how can I naturally train my LSTM on that text (input the whole text and let the LSTM figure out the dependencies between the words) instead of having to 'artificially' create batches of sentences from that book myself to train my LSTM on? I believe I should use stateful LSTMs could help but I'm not sure how.<\/p>\n","answers":[{"AnswerId":"41701750","CreationDate":"2017-01-17T16:03:17.733","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<ol>\n<li><p>Having a stateful LSTM in Keras means that a Keras variable will be used to store and update the state, and in fact you could check the value of the state vector(s) at any time (that is, until you call <code>reset_states()<\/code>). A non-stateful model, on the other hand, will use an initial zero state every time it processes a batch, so it is as if you always called <code>reset_states()<\/code> after <code>train_on_batch<\/code>, <code>test_on_batch<\/code> and <code>predict_on_batch<\/code>. The explanation about the state being reused for the next batch on stateful models is just about that difference with non-stateful; of course the state will always flow within the batch and you do <em>not<\/em> need to have batches of size 1 for that to happen. I see two scenarios where stateful models are useful:<\/p>\n\n<ul>\n<li>You want to train on split sequences of data because these are very long and it would not be practical to train on their whole length.<\/li>\n<li>On prediction time, you want to retrieve the output for each time point in the sequence, not just at the end (either because you want to feed it back into the network or because your application needs it). I personally do that in the models that I export for later integration (which are \"copies\" of the training model with batch size of 1).<\/li>\n<\/ul><\/li>\n<li><p>I agree that the example of an RNN for the alphabet does not really seem very useful in practice; it will only work when you start with the letter A. If you want to learn to reproduce the alphabet starting at any letter, you would need to train the network with that kind of examples (subsequences or rotations of the alphabet). But I think a regular feed-forward network could learn to predict the next letter of the alphabet training on pairs like (A, B), (B, C), etc. I think the example is meant for demonstrative purposes more than anything else.<\/p><\/li>\n<li><p>You may have probably already read it, but the popular post <a href=\"https:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/\" rel=\"noreferrer\">The Unreasonable Effectiveness of Recurrent Neural Networks<\/a> shows some interesting results along the lines of what you want to do (although it does not really dive into implementation specifics). I don't have personal experience training RNN with textual data, but there is a number of approaches you can research. You can build character-based models (like the ones in the post), where your input and receive one character at a time. A more advanced approach is to do some preprocessing on the texts and transform them into sequences of numbers; Keras includes some <a href=\"https:\/\/keras.io\/preprocessing\/text\/\" rel=\"noreferrer\">text preprocessing functions<\/a> to do that. Having one single number as feature space is probably not going to work all that well, so you could simply turn each word into a vector with one-hot encoding or, more interestingly, have the network learn the best vector representation for each for, which is what they call en <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"noreferrer\">embedding<\/a>. You can go even further with the preprocessing and look into something like <a href=\"http:\/\/www.nltk.org\" rel=\"noreferrer\">NLTK<\/a>, specially if you want to remove stop words, punctuation and things like that. Finally, if you have sequences of different sizes (e.g. you are using full texts instead of excerpts of a fixed size, which may or may not be important for you) you will need to be a bit more careful and use <a href=\"https:\/\/keras.io\/layers\/core\/#masking\" rel=\"noreferrer\">masking<\/a> and\/or <a href=\"https:\/\/keras.io\/models\/model\/#train_on_batch\" rel=\"noreferrer\">sample weighting<\/a>. Depending on the exact problem, you can set up the training accordingly. If you want to learn to generate similar text, the \"Y\" would be the similar to the \"X\" (one-hot encoded), only shifted by one (or more) positions (in this case you may need to use <code>return_sequences=True<\/code> and <a href=\"https:\/\/keras.io\/layers\/wrappers\/#TimeDistributed\" rel=\"noreferrer\">TimeDistributed layers<\/a>). If you want to determine the autor, your output could be a <a href=\"https:\/\/keras.io\/layers\/core\/#Dense\" rel=\"noreferrer\">softmax Dense layer<\/a>.<\/p><\/li>\n<\/ol>\n\n<p>Hope that helps.<\/p>\n"}]}
{"QuestionId":41695406,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-17T11:02:05.543","AcceptedAnswerId":null,"OwnerUserId":6809473.0,"Title":"Tensorflow topn_test fails on ppc64le","Body":"<p>With tensorflow 0.12, topn_test.py has failures for the tests written for TopN class on ppc64le.\nTaking an example of one of the failing tests in it, here is what is being observed -<\/p>\n\n<h2>FAIL: testSimple (<strong>main<\/strong>.TopNTest)<\/h2>\n\n<p>Traceback (most recent call last):<\/p>\n\n<pre><\/pre>\n\n<p>Expected is  whereas actual is .<\/p>\n\n<p>Code of this test is as below -<\/p>\n\n<pre><\/pre>\n\n<p>In above test, if we run ids into sess before , then the test passes which means running ids into session twice gives expected result in 2nd run. \nThe output of 1st sess.run(ids) gives  and then next run gives . This is weird. Ideally sess.run should always give fixed result in every run. Please correct if I'm wrong here. <\/p>\n\n<p>Same thing is observed even on x86, with the only difference is 1st run gives  and 2nd run gives . <\/p>\n\n<p>Could someone please tell me how sess.run(ids) gives different output in each run? Is this expected behavior or there is some other underlying problem with the implementation of TopN class? I've also checked with a constant tensor and this weird behavior is not seen with it.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[]}
{"QuestionId":41695844,"AnswerCount":6,"Tags":"<python><image><matplotlib><neural-network><keras>","CreationDate":"2017-01-17T11:23:02.743","AcceptedAnswerId":41697846.0,"OwnerUserId":2670775.0,"Title":"Keras showing images from data generator","Body":"<p>I am using image generator for keras like this:<\/p>\n\n<pre><\/pre>\n\n<p>This shows wrong colors:\n<a href=\"https:\/\/i.stack.imgur.com\/gwzou.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gwzou.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have two questions. <\/p>\n\n<ol>\n<li><p>How to fix the problem <\/p><\/li>\n<li><p>How to get file names of the files (so that I can read it myself from something like matplotlib) <\/p><\/li>\n<\/ol>\n\n<p>Edit : this is what my datagen looks like<\/p>\n\n<pre><\/pre>\n\n<p>Edit 2 : <\/p>\n\n<p>After following Marcin's answer : <\/p>\n\n<pre><\/pre>\n\n<p>I get normal colors , but there are still some weird colors:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wAFpa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wAFpa.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"43229486","CreationDate":"2017-04-05T11:10:49.663","ParentId":null,"OwnerUserId":"4773044","Title":null,"Body":"<p>I had the same problem as OP and solved it by rescaling the pixels from 0-255 to 0-1.<\/p>\n\n<p>Keras' ImageDataGenerator takes a 'rescale' parameter, which I set to (1\/255). This produced images with expected colors<\/p>\n\n<pre><code>image_gen = ImageDataGenerator(rescale=(1\/255))\n<\/code><\/pre>\n"},{"AnswerId":"56316327","CreationDate":"2019-05-26T18:33:11.577","ParentId":null,"OwnerUserId":"8801414","Title":null,"Body":"<pre><code>from skimage import io\n\ndef imshow(image_RGB):\n  io.imshow(image_RGB)\n  io.show()\n\nx,y = train_generator.next()\n\nfor i in range(0,11):\n    image = x[i]\n    imshow(image)\n<\/code><\/pre>\n\n<p>It works for me.<\/p>\n"},{"AnswerId":"49190401","CreationDate":"2018-03-09T09:21:09.993","ParentId":null,"OwnerUserId":"9466659","Title":null,"Body":"<p>The dtype of your image array is 'float32', just convert it into 'uint8':<\/p>\n\n<pre><code>plt.imshow(image.astype('uint8'))\n<\/code><\/pre>\n"},{"AnswerId":"41697846","CreationDate":"2017-01-17T12:57:39.333","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<ol>\n<li><p>There are at least three ways to have this twisted colors. So:<\/p>\n\n<ul>\n<li>one option is that you need to switch a color ordering like in this <a href=\"https:\/\/stackoverflow.com\/questions\/14387668\/invert-colors-when-plotting-a-png-file-using-matplotlib\">question<\/a>.<\/li>\n<li>second is that you might have your pictures made to be a negative (every channels gets transformed by <code>255 - x<\/code> transformation) this sometimes happens when it comes to using some GIS libraries.<\/li>\n<li>you could also use a <code>score\/255<\/code> transformation.<\/li>\n<\/ul>\n\n<p>You need to check which options happens in your case.<\/p><\/li>\n<li><p>In order to get the images on your own I usually use (when your folder has a format suitable for a Keras <code>flow_from_directory<\/code>) I usually use the mix of <code>os.listdir<\/code> and <code>os.path.join<\/code> by :<\/p>\n\n<pre><code>list_of_labels = os.listdir(path_to_dir_with_label_dirs)\nfor label in list_of_labels:\n    current_label_dir_path = os.path.join(path_to_dir_with_label_dirs, label\n    list_of_images = os.listdir(current_label_dir_path)\n    for image in list_of_images:\n        current_image_path = os.path.join(current_label_dir_path, image)\n        image = open(current_image_path) # use the function which you want.\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"41698322","CreationDate":"2017-01-17T13:22:05.973","ParentId":null,"OwnerUserId":"497600","Title":null,"Body":"<p>The color problem is rather strange.\nI'll try to reproduce it once I have access to my linux machine.<\/p>\n\n<p>For the filename part of the question, I would like to propose a small change to the Keras sourcecode:<\/p>\n\n<p>You might want to take a look at this file:\n<a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/preprocessing\/image.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/preprocessing\/image.py<\/a>\nIt contains the image preprocessing routines.<\/p>\n\n<p>Look at line 820, the <code>next()<\/code> function of the <code>DirectoryIterator<\/code>: this is called to get new images from the directory.<\/p>\n\n<p>Inside of that function, look at line 838, if <code>save_to_dir<\/code> has been set to a path, the generator will output the augmented images to this path, for debugging purposes.\nThe name of the augmented image is a mixture of an index and a hash. Not useful for you.<\/p>\n\n<p>But you can change the code quite easily:<\/p>\n\n<pre><code>filenames=[] #&lt;-------------------------------------------- new code\nfor i, j in enumerate(index_array):\n    fname = self.filenames[j]\n    img = load_img(os.path.join(self.directory, fname),\n                   grayscale=grayscale,\n                   target_size=self.target_size)\n    x = img_to_array(img, dim_ordering=self.dim_ordering)\n    x = self.image_data_generator.random_transform(x)\n    x = self.image_data_generator.standardize(x)\n\n    filenames.append(fname) # &lt;-----------------------------store the used image's name\n    batch_x[i] = x\n# optionally save augmented images to disk for debugging purposes\nif self.save_to_dir:\n    for i in range(current_batch_size):\n        img = array_to_img(batch_x[i], self.dim_ordering, scale=True)\n        #fname = '{prefix}_{index}_{hash}.{format}'.format(prefix=self.save_prefix,\n        #                                                  index=current_index + i,\n        #                                                  hash=np.random.randint(1e4),\n        #                                                  format=self.save_format)\n        fname=filenames[i] # &lt;------------------------------ use the stored code instead\n        img.save(os.path.join(self.save_to_dir, fname))\n<\/code><\/pre>\n\n<p><strong>Now the augmented image is saved with the original filename.<\/strong><\/p>\n\n<p>This should allow you to save the images under their original filenames.\nOk, how do you actually inject this into the Keras souce ?<\/p>\n\n<p>Do it like this:<\/p>\n\n<ol>\n<li>clone Keras: <code>git clone https:\/\/github.com\/fchollet\/keras<\/code><\/li>\n<li>go to the sourcefile I linked above. Make the change.<\/li>\n<li>Trick your python code to import the changed code instead of the version installed by pip.<\/li>\n<\/ol>\n\n<p>.<\/p>\n\n<pre><code># this is the path to the cloned repository\n# if you cloned it next to your script\n# then just use keras\/\n# if it's one folder above\n# then use ..\/keras\/\nsys.path.insert(0, os.getcwd() + \"\/path\/to\/keras\/\")\n\nimport keras\n<\/code><\/pre>\n\n<p>Now the <code>DirectoryIterator<\/code> is your patched version.<\/p>\n\n<p>I hope that this works, I'm currently on windows. My python stack is only on the linux machine. There might be a small syntax error. <\/p>\n"},{"AnswerId":"58461065","CreationDate":"2019-10-19T06:20:29.173","ParentId":null,"OwnerUserId":"10798917","Title":null,"Body":"<p>Just a bit of advice if you are using <code>test_batches=Imagedatagenerator().flow<\/code> from directory. If you use this to feed a predict generator make sure you set <code>shuffle=false<\/code> to maintain a correlation between the file and the associated prediction. If you have files numerically labelled in the directory for example as <em>1.jpg<\/em>, <em>2.jpg<\/em> etc. The images are not fetched as you might think. They are fetched in the order:\n<em>1.jpg<\/em>, <em>10.jpg<\/em>, <em>2.jpg<\/em>, <em>20.jpg<\/em> etc. This makes it hard to match a prediction to a specific file. You can get around this by using 0's padding for example <em>01.jpg<\/em>, <em>02.jpg<\/em> etc. <\/p>\n"}]}
{"QuestionId":41695855,"AnswerCount":2,"Tags":"<python><deep-learning><theano><keras>","CreationDate":"2017-01-17T11:23:28.290","AcceptedAnswerId":41696894.0,"OwnerUserId":4730816.0,"Title":"Modifying feature maps in keras","Body":"<p>The features maps can be obtained using:<\/p>\n\n<pre><\/pre>\n\n<p>This is good for visualisation of the data. But, I also intend to modify the output for each layer and then fed this output back to the network. Can anyone suggest me how I can do the same?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41696894","CreationDate":"2017-01-17T12:15:08.853","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>I'm restoring this answer with edits to reflect additional information.<\/p>\n\n<p>Assuming you have a model similar to this:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(1000, input_dim=1000))\nmodel.add(Dense(1000))\n<\/code><\/pre>\n\n<p>And you want to run a custom modification on the output of the first layer before passing it to the second layer you can use the <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\">lambda layer<\/a> as so:<\/p>\n\n<pre><code>f = K.function(\\* some function *\\)\n\nmodel = Sequential()\nmodel.add(Dense(1000, input_dim=1000))\nmodel.add(Lambda(lambda x: f(x))\nmodel.add(Dense(1000))\n<\/code><\/pre>\n"},{"AnswerId":"41696242","CreationDate":"2017-01-17T11:42:56.797","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>If you just want to do this once you can do something like this:<\/p>\n\n<pre><code>modified_layer_output = your_old_function([X]) * some_modification\nget_final_layer_output = K.function([model.layers[3].input],\n                                  [model.layers[-1].output])\nresult = get_final_layer_output(modified_layer_output)\n<\/code><\/pre>\n\n<p>You could also create a new model to learn on your modified layer output.<\/p>\n\n<p>Edit:\nYou could do your write your own keras layer to do whatever you want with the input and pass it to the next layer like shown here (<a href=\"https:\/\/keras.io\/layers\/writing-your-own-keras-layers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/writing-your-own-keras-layers\/<\/a>).<\/p>\n"}]}
{"QuestionId":41695893,"AnswerCount":2,"Tags":"<python><scope><tensorflow>","CreationDate":"2017-01-17T11:25:17.063","AcceptedAnswerId":41701881.0,"OwnerUserId":3990607.0,"Title":"Tensorflow: Conditionally add variable scope","Body":"<p>I would like to change variable scope conditionally in tensorflow.<\/p>\n\n<p>So for example, if  is either a string or :<\/p>\n\n<pre><\/pre>\n\n<p>But I don't want to have to write the ,  part in double. I just want the  to determine the scope and then do everything else the same from there.<\/p>\n\n<p>Any ideas on how I could do that? <\/p>\n","answers":[{"AnswerId":"41702895","CreationDate":"2017-01-17T17:02:25.150","ParentId":null,"OwnerUserId":"3990607","Title":null,"Body":"<p>Thanks for @keveman for putting me on the right track. Though I couldn't make his answer work he put me on the right track: what I needed was an empty scope, thus the following works:<\/p>\n\n<pre><code>class empty_scope():\n     def __init__(self):\n         pass\n     def __enter__(self):\n         pass\n     def __exit__(self, type, value, traceback):\n         pass\n\ndef cond_scope(scope):\n    return empty_scope() if scope is None else tf.variable_scope(scope)\n<\/code><\/pre>\n\n<p>After which I can do:<\/p>\n\n<pre><code>with cond_scope(scope):\n\n    a = tf.get_Variable(....)\n    b = tf.get_Variable(....)\n<\/code><\/pre>\n\n<p>For more on the <code>with<\/code> in python see:\n<a href=\"http:\/\/preshing.com\/20110920\/the-python-with-statement-by-example\/\" rel=\"nofollow noreferrer\">The Python \"with\" Statement by Example<\/a><\/p>\n"},{"AnswerId":"41701881","CreationDate":"2017-01-17T16:10:54.277","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>This is not specific to TensorFlow, but more of a general Python language question, just FYI. In any case, you can achieve what you want to do with a wrapper context manager as follows :<\/p>\n\n<pre><code>class cond_scope(object):\n  def __init__(self, condition, contextmanager):\n    self.condition = condition\n    self.contextmanager = contextmanager\n  def __enter__(self):\n    if self.condition:\n      return self.contextmanager.__enter__()\n  def __exit__(self, *args):\n    if self.condition:\n      return self.contextmanager.__exit__(*args)\n\nwith cond_scope(scope is not None, scope):\n  a = tf.get_variable(....)\n  b = tf.get_variable(....)\n<\/code><\/pre>\n\n<p><strong>EDIT<\/strong>: Fixed the code.<\/p>\n"}]}
{"QuestionId":41696996,"AnswerCount":2,"Tags":"<java><tensorflow><deep-learning><tensorflow-serving>","CreationDate":"2017-01-17T12:20:28.697","AcceptedAnswerId":43877512.0,"OwnerUserId":2598910.0,"Title":"How to create a tensorflow serving client for the 'wide and deep' model?","Body":"<p>I've created a model based on the 'wide and deep' example (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/wide_n_deep_tutorial.py\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/wide_n_deep_tutorial.py<\/a>).<\/p>\n\n<p>I've exported the model as follows:<\/p>\n\n<pre><\/pre>\n\n<p>My question is, how do I create a client to make predictions from this exported model?  Also, have I exported the model correctly?<\/p>\n\n<p>Ultimately I need to be able do this in Java too.  I suspect I can do this by creating Java classes from proto files using gRPC.<\/p>\n\n<p>Documentation is very sketchy, hence why I am asking on here.<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"43877512","CreationDate":"2017-05-09T18:38:50.907","ParentId":null,"OwnerUserId":"7501011","Title":null,"Body":"<p>I wrote a simple tutorial <a href=\"https:\/\/github.com\/MtDersvan\/tf_playground\/blob\/master\/wide_and_deep_tutorial\/wide_and_deep_basic_serving.md\" rel=\"nofollow noreferrer\">Exporting and Serving a TensorFlow Wide &amp; Deep Model<\/a>.<\/p>\n\n<p>TL;DR<\/p>\n\n<p>To export an estimator there are four steps:<\/p>\n\n<ol>\n<li><p>Define features for export as a list of all features used during estimator initialization.<\/p><\/li>\n<li><p>Create a feature config using <code>create_feature_spec_for_parsing<\/code>.<\/p><\/li>\n<li><p>Build a <code>serving_input_fn<\/code> suitable for use in serving using <code>input_fn_utils.build_parsing_serving_input_fn<\/code>.<\/p><\/li>\n<li><p>Export the model using <code>export_savedmodel()<\/code>.<\/p><\/li>\n<\/ol>\n\n<p>To run a client script properly you need to do three following steps:<\/p>\n\n<ol>\n<li><p>Create and place your script somewhere in the \/serving\/ folder, e.g. \/serving\/tensorflow_serving\/example\/<\/p><\/li>\n<li><p>Create or modify corresponding BUILD file by adding a <code>py_binary<\/code>.<\/p><\/li>\n<li><p>Build and run a model server, e.g. <code>tensorflow_model_server<\/code>.<\/p><\/li>\n<li><p>Create, build and run a client that sends a tf.Example to our <code>tensorflow_model_server<\/code> for the inference.<\/p><\/li>\n<\/ol>\n\n<p>For more details look at the tutorial itself.<\/p>\n"},{"AnswerId":"42658289","CreationDate":"2017-03-07T21:03:43.153","ParentId":null,"OwnerUserId":"4338098","Title":null,"Body":"<p>Just spent a solid week figuring this out. First off, <code>m.export<\/code> is going to deprecated in a couple weeks, so instead of that block, use: <code>m.export_savedmodel(export_path, input_fn=serving_input_fn)<\/code>.<\/p>\n\n<p>Which means you then have to define <code>serving_input_fn()<\/code>, which of course is supposed to have a different signature than the <code>input_fn()<\/code> defined in the wide and deep tutorial. Namely, moving forward, I guess it's recommended that <code>input_fn()<\/code>-type things are supposed to return an <code>InputFnOps<\/code> object, defined <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/utils\/input_fn_utils.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>Here's how I figured out how to make that work:<\/p>\n\n<pre><code>from tensorflow.contrib.learn.python.learn.utils import input_fn_utils\nfrom tensorflow.python.ops import array_ops\nfrom tensorflow.python.framework import dtypes\n\ndef serving_input_fn():\n  features, labels = input_fn()\n  features[\"examples\"] = tf.placeholder(tf.string)\n\n  serialized_tf_example = array_ops.placeholder(dtype=dtypes.string,\n                                                shape=[None],\n                                                name='input_example_tensor')\n  inputs = {'examples': serialized_tf_example}\n  labels = None  # these are not known in serving!\n  return input_fn_utils.InputFnOps(features, labels, inputs)\n<\/code><\/pre>\n\n<p>This is probably not 100% idiomatic, but I'm pretty sure it works. For now.<\/p>\n"}]}
{"QuestionId":41697662,"AnswerCount":1,"Tags":"<python><multithreading><tensorflow>","CreationDate":"2017-01-17T12:48:50.143","AcceptedAnswerId":null,"OwnerUserId":2547462.0,"Title":"Tensorflow and threading","Body":"<p>Below is the simple mnist tutorial (i.e. single layer softmax) from the Tensorflow website, which I tried to extend with a multi-threaded training step:<\/p>\n\n<pre><\/pre>\n\n<p>I must be missing something, as 8 threads as below yield inconsistent results (accuracy approx. = 0.1), when with 1 thread only the expected accuracy is obtained (approx. 0.92). Does anybody have a clue about my mistake(s)? Thanks!<\/p>\n","answers":[{"AnswerId":"42979411","CreationDate":"2017-03-23T14:45:46.727","ParentId":null,"OwnerUserId":"1780735","Title":null,"Body":"<p>Note that unfortunately, <code>threading<\/code>with python doesn't create real parallelism because of the <a href=\"https:\/\/wiki.python.org\/moin\/GlobalInterpreterLock\" rel=\"nofollow noreferrer\">GIL<\/a>. So what happens here is that you will have multiple threads which are all running on the same CPU where in reality they are running sequentially. Therefore, I would suggest using Coordinator in Tensorflow. More information about Coordinator can be found here: <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/threading_and_queues\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/threading_and_queues<\/a><br>\n<a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/reading_data<\/a><\/p>\n\n<p>Finally, I would suggest you say:<\/p>\n\n<pre><code>with tf.device('\/cpu:0'):\n    your code should go here... 'for the first thread'\n<\/code><\/pre>\n\n<p>Then use another cpu for the other thread and so on...\nHope this answer finds you well!!<\/p>\n"}]}
{"QuestionId":41697969,"AnswerCount":1,"Tags":"<python><tensorflow><standard-deviation>","CreationDate":"2017-01-17T13:03:32.893","AcceptedAnswerId":41698153.0,"OwnerUserId":null,"Title":"error in calculating standard deviation value","Body":"<p>I get two different values using tensorflow and numpy<\/p>\n\n<pre><\/pre>\n\n<p>gives ; whereas:<\/p>\n\n<pre><\/pre>\n\n<p>gives .<\/p>\n\n<p>Did I declared  object correctly?<\/p>\n","answers":[{"AnswerId":"41698153","CreationDate":"2017-01-17T13:13:04.940","ParentId":null,"OwnerUserId":"512225","Title":null,"Body":"<p>The <a href=\"https:\/\/en.wikipedia.org\/wiki\/Moment_%28mathematics%29#Significance_of_the_moments\" rel=\"nofollow noreferrer\">second central moment<\/a> of a sample is the variance (4+1+0+1+4)\/5.=2.0 in your example. The std is its square root.<\/p>\n"}]}
{"QuestionId":41699037,"AnswerCount":2,"Tags":"<python><neural-network><keras><sequential>","CreationDate":"2017-01-17T13:58:13.923","AcceptedAnswerId":null,"OwnerUserId":7320845.0,"Title":"Getting 0 or 1 as probability values in neural network Keras sequential model","Body":"<p>I am running my training set of 30k on python library Keras for neural network (model \"sequential\"). <\/p>\n\n<p>When i am fitting the model and calculating probabilities on the test set (size = 13k) then I am getting 0 or 1 only. Whereas I am supposed to get some fractional values like 0.62, 0.42, 0.31 etc. <\/p>\n\n<p>I tried using class_weight and setting it to 'auto' or 'balanced' but I was not able to solve this issue. <\/p>\n\n<p>Output is not integer type, its float type only.<\/p>\n\n<p>Below is the code<\/p>\n\n<pre><\/pre>\n\n<p>And the values of proba are all 0s.\nThe data has only 2% defaulter cases, so its imbalanced class.<\/p>\n","answers":[{"AnswerId":"42188439","CreationDate":"2017-02-12T13:57:06.693","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Maybe I am wrong but I have never seen this argument to fit so far :<\/p>\n\n<pre><code> class_weight ='auto'\n<\/code><\/pre>\n\n<p>I would try instead<\/p>\n\n<pre><code>class_weight = {0: 1., 1: 50.}\n<\/code><\/pre>\n\n<p>Indeed your problem seems to be coming from inbalanced dataset.<\/p>\n"},{"AnswerId":"41699632","CreationDate":"2017-01-17T14:25:51.720","ParentId":null,"OwnerUserId":"6409943","Title":null,"Body":"<p>It sounds awfully like that you might be getting an integer return, which can only be 1 or 0, and not a float return. Thus you could try seeing whether <code>int<\/code> is in your code anywhere turning it into an integer, or you might need to turn it into a <code>float<\/code> somewhere. Impossible to tell without more context...<\/p>\n"}]}
{"QuestionId":41700763,"AnswerCount":1,"Tags":"<tensorflow><gradient>","CreationDate":"2017-01-17T15:18:04.273","AcceptedAnswerId":null,"OwnerUserId":3043749.0,"Title":"generating unrecognizable image to fool vggnet","Body":"<p>I am trying to generate an unrecognizable image which can fool Vggnet. I used the following vgg model for tensorflow. I add some modification for calculating the gradient. In the ending part, you can see my modification for calculating the gradient respect to the given image (is it correct? I am trying to generate an image to whom the vggnet assign high probability at class 1). With this gradient, I update the random image for fooling the vggnet. But this is not so successful. I can't generate an image with high probability. The maximum probability I got is around 0.001. How can I make it keep increasing? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Lv4gc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Lv4gc.png\" alt=\"enter image description here\"><\/a>.  <\/p>\n\n<p><strong>Vggnet model<\/strong><\/p>\n\n#\n\n<pre><\/pre>\n\n<p><strong>Create session<\/strong><\/p>\n\n#\n\n<pre><\/pre>\n\n<p><strong>Generate new image for fooling<\/strong><\/p>\n\n#\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41737132","CreationDate":"2017-01-19T08:31:01.367","ParentId":null,"OwnerUserId":"497600","Title":null,"Body":"<p>I'm surprised that the shapes of the gradient and the image match.\nYou are taking the derivative of the loss with respect to the parameters, is should be with respect to the image placeholder. Excuse me, if I'm missing something obvious, I can't run the code right now.<\/p>\n\n<p>The computation of the loss is based on <code>fc3l<\/code>, the final output is <code>probs<\/code>. I don't see where <code>probs<\/code> is computed in the VGG code. Maybe there are layers in between. You could plot the first component of <code>fc3l<\/code> instead, see if that goes up.<\/p>\n\n<p>You should probably base the loss on <code>probs<\/code>.<\/p>\n"}]}
{"QuestionId":41701783,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-17T16:05:18.537","AcceptedAnswerId":41703783.0,"OwnerUserId":216880.0,"Title":"Prune computation graph in tensorflow","Body":"<p>A computation graph in tensorflow has some node  and some node .  represents the input tensor to a neural network,  its prediction. Some other subgraph feeds to , some losses (and other computations) follow on . After training is finished, I would like to keep only the (forward) subgraph between  and . Is it possible to remove all other ops? <a href=\"https:\/\/stackoverflow.com\/questions\/37103002\/deleting-all-but-a-few-nodes-in-tensorflow-graph\">Deleting all but a few nodes in TensorFlow graph<\/a> indicates that per-se this is not possible. Maybe through the de-tour of exporting stuff using ?<\/p>\n","answers":[{"AnswerId":"41703783","CreationDate":"2017-01-17T17:51:01.550","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Although it is not possible to delete nodes from a <code>tf.Graph<\/code>, TensorFlow includes tools that enable you to prune a lower-level <code>tf.GraphDef<\/code> protocol buffer. The TensorFlow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/99fe61a8a8f3dd41b4e1e4dedfc53b45f67e88a7\/tensorflow\/tools\/graph_transforms\/README.md\" rel=\"nofollow noreferrer\">graph transformation tool<\/a> includes a transformation pass called <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/99fe61a8a8f3dd41b4e1e4dedfc53b45f67e88a7\/tensorflow\/tools\/graph_transforms\/README.md#strip_unused_nodes\" rel=\"nofollow noreferrer\"><code>strip_unused_nodes<\/code><\/a> that allows you to name set of input nodes and a set of output nodes, and returns a pruned <code>tf.GraphDef<\/code>.<\/p>\n"}]}
{"QuestionId":41702060,"AnswerCount":1,"Tags":"<deep-learning><caffe><pycaffe><lmdb><deeplearning4j>","CreationDate":"2017-01-17T16:20:56.900","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"Is it possible to reduce the number of images in batch in convert_imageset.cpp to tackle out of memory of GPU?","Body":"<p>I was trying to run fcn on my data in caffe. I was able to convert my image sets into  by  builtin function caffe. However, once I wanted to train the , it gave me the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I went through many online resources to solve the memory failure, but most of them suggesting reducing batch size. Even, I reduced the size of images to 256x256. I could not tackle this issue yet. \nI checked the memory of GPU by this command , and the model is  and the memory is . Since the batch size in  is 1, I can not do anythin in . So my questions are:<\/p>\n\n<ol>\n<li>By looking at log file in Terminal,I realized that whenever  converting the data into LMDB, it is taking 1000 image in a group. Is it possible I change this number in line  and  of <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/tools\/convert_imageset.cpp#L143\" rel=\"nofollow noreferrer\"><\/a> to a smaller (for example 2; to take two image at a time), recompile caffe, and then convert images to lmdb by using convert_imageset? Does it make sense?<\/li>\n<li>If the answer to question 1 is yes, how can I compile caffe again,\nshould I remove  folder and again do caffe installation from\nscratch? <\/li>\n<li>How caffe process the LMDB data? Is it like taking a batch of those 1000 images showing while running convert_imagenet?<\/li>\n<\/ol>\n\n<p>Your help is really appreciated.\nThanks...<\/p>\n","answers":[{"AnswerId":"41702410","CreationDate":"2017-01-17T16:38:55.973","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<ol>\n<li><p>AFAIK, there is no effect for the number of entries committed to lmdb at each transaction (<code>txn-&gt;Commit();<\/code>) on the cuda out of memory.  <\/p><\/li>\n<li><p>If you do want to re-compile caffe for whatever reason, simply run <code>make clean<\/code>. This will clear everything and let you re-compile from scratch.<\/p><\/li>\n<li><p>Again, AFAIK, caffe access lmdb <code>batch_size<\/code> images at a time regardless of the size of the transaction size used when writing the dataset.<\/p><\/li>\n<\/ol>\n\n<p>Are you sure <code>batch_size<\/code> is set to 1 for <strong>both<\/strong> TRAIN and TEST phases?<\/p>\n"}]}
{"QuestionId":41702081,"AnswerCount":2,"Tags":"<tensorflow><conda>","CreationDate":"2017-01-17T16:22:37.893","AcceptedAnswerId":null,"OwnerUserId":1018464.0,"Title":"Anaconda environment for CPU and GPU versions of TensorFlow","Body":"<p>I want to create two separate environments with TensorFlow in anaconda, one with CPU only support, which is compiled from sources, and one with GPU support using official Tensorflow binaries.<\/p>\n\n<p>I have set up two environments with conda, using . I activate my new environment using . However, whenever I install a TensorFlow version using pip, the old environment is affected as well.<\/p>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"42805791","CreationDate":"2017-03-15T09:32:47.370","ParentId":null,"OwnerUserId":"7713921","Title":null,"Body":"<p>Use <code>conda create -n gpu python=3.6<\/code> (or whatever python version you want) instead of <code>conda create -n gpu<\/code> when you create an environment.<\/p>\n\n<hr>\n\n<p>The idea of conda is to have an isolated environment that it has its own python, pip, etc. To create a new environment with isolated installed program, you need to explicitly declare it, otherwise, it uses the shared program between all environments. You can activate the environments and try <code>which python<\/code> or <code>which pip<\/code> to make sure they have their own version of python and pip.<\/p>\n"},{"AnswerId":"41708019","CreationDate":"2017-01-17T22:18:09.793","ParentId":null,"OwnerUserId":"2127762","Title":null,"Body":"<p>Conda uses hard links extensively. It's likely that pip is overwriting files that conda has already created, without removing the file first, and thereby affecting files in your other conda environments.<\/p>\n\n<p>Try adding the <code>--copy<\/code> flag to your <code>create --clone<\/code> command.<\/p>\n"}]}
{"QuestionId":41702226,"AnswerCount":1,"Tags":"<python><git><tensorflow><tensorflow-serving>","CreationDate":"2017-01-17T16:30:35.837","AcceptedAnswerId":null,"OwnerUserId":2598910.0,"Title":"Where should Tensorflow serving be cloned to?","Body":"<p>I've installed Tensorflow but now wish to productionize my model.<\/p>\n\n<p>So I'm trying to follow this guide:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/g3doc\/setup.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/g3doc\/setup.md<\/a><\/p>\n\n<p>Does Tensorflow serving run along side Tensor flow?<\/p>\n\n<p>If so where should I clone the the repo to so that the packages can be seen on the python library path please?<\/p>\n\n<p>Many thanks.<\/p>\n","answers":[{"AnswerId":"42773158","CreationDate":"2017-03-13T20:49:41.953","ParentId":null,"OwnerUserId":"6061661","Title":null,"Body":"<p>This has not been developed as of 03\/13\/2017.  See <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/52\" rel=\"nofollow noreferrer\">this issue<\/a>.  It says there has not been a binary release of tensorflow serving yet so the only way you can import the packages is by cloning the repo and doing development inside of the cloned project.<\/p>\n\n<p>To answer your question about serving running alongside Tensorflow:  the tensorflow module inside of serving is the tensorflow project, so basically tensorflow_serving comes with its own tensorflow.<\/p>\n"}]}
{"QuestionId":41703144,"AnswerCount":1,"Tags":"<python><deep-learning><theano><keras><objective-function>","CreationDate":"2017-01-17T17:14:06.987","AcceptedAnswerId":null,"OwnerUserId":3120870.0,"Title":"Applying element wise conditional functions on Theano TensorVariable","Body":"<p>Easiest thing might be for me to just post the numpy code that I'm trying to perform directly in Theano if it's possible:<\/p>\n\n<pre><\/pre>\n\n<p>If it's not immediately obvious, what I'm trying to do is use the values from one matrix of a tensor made up of 7 different matrices and apply that to the other matrices in the tensor. <\/p>\n\n<p>Really, the problem I'm solving is doing conditional statements in an objective function for a fully convoltional network in Keras. Basically the loss for some of the feature map values is going to be calculated (and subsequently weighted) differently from others depending on some of the values in one of the feature maps. <\/p>\n","answers":[{"AnswerId":"41706564","CreationDate":"2017-01-17T20:40:02.810","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can easily implement conditionals with <code>switch<\/code> statement.<\/p>\n\n<p>Here would be the equivalent code:<\/p>\n\n<pre><code>import theano\nfrom theano import tensor as T\nimport numpy as np\n\n\ndef _check_new(var):\n    shape =  var.shape[0]\n    t_1, t_2 = T.split(var, [1, shape-1], 2, axis=0)\n    ones = T.ones_like(t_1)\n    cond = T.gt(t_1, ones)\n    mask = T.repeat(cond, t_2.shape[0], axis=0)\n    out  = T.switch(mask, t_2, T.zeros_like(t_2))\n    output = T.join(0, cond, out)\n    return output\n\ndef _check_old(var):\n    tensor = var.eval()\n\n    tensor2 = tensor[0,:,:]\n    tensor2[tensor2 &lt; 1] = 0.0\n    tensor2[tensor2 &gt; 0] = 1.0\n    new_tensor = [tensor2]\n\n    for i in range(1, tensor.shape[0]):\n        new_tensor.append(np.multiply(tensor2, tensor[i,:,:]))\n\n    output = theano.shared(np.array(new_tensor).reshape(7,16,16))\n    return output\n\n\ntensor = theano.shared(np.random.randn(7, 16, 16))\nout1 =  _check_new(tensor).eval() \nout2 =  _check_old(tensor).eval()\nprint out1\nprint '----------------'\nprint ((out1-out2) ** 2).mean()\n<\/code><\/pre>\n\n<p>Note: since your masking on the first filter, I needed to use <code>split<\/code> and <code>join<\/code> operations.<\/p>\n"}]}
{"QuestionId":41703184,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2017-01-17T17:16:35.950","AcceptedAnswerId":null,"OwnerUserId":7349995.0,"Title":"Can not compile torch on x ubuntu 15.10","Body":"<p>This is the error I get. <\/p>\n\n<pre><\/pre>\n\n<p>What does it mean?<\/p>\n","answers":[{"AnswerId":"51038718","CreationDate":"2018-06-26T08:53:12.997","ParentId":null,"OwnerUserId":"9126117","Title":null,"Body":"<p>I fixed the problem with\n<code>sudo apt-get install libopenblas-dev<\/code>\nas suggested in \n<a href=\"https:\/\/stackoverflow.com\/questions\/36483054\/install-openblas-via-apt-get-sudo-apt-get-install-openblas-dev\">Install openblas via apt-get `sudo apt-get install openblas-dev`<\/a><\/p>\n"}]}
{"QuestionId":41703199,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-17T17:17:12.233","AcceptedAnswerId":null,"OwnerUserId":2228912.0,"Title":"Using tensorflow.argmax on a tensorflow.placeholder","Body":"<p>Why does  work on a  while  can't, even though both accept as an input a  and return a ?<\/p>\n\n<p>The following piece of code works normally:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>While this one (using  instead) doesn't:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>And returns the following error (in the  call):<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41703310,"AnswerCount":0,"Tags":"<multithreading><tensorflow>","CreationDate":"2017-01-17T17:23:14.877","AcceptedAnswerId":null,"OwnerUserId":7427246.0,"Title":"tensorflow workers start before the main thread could initialize variables","Body":"<p>I am trying to use tf.train.batch to run enqueue images in multiple threads. When the number of threads is 1, the code works fine. But when I set a higher number of threads I receive an error:<\/p>\n\n<pre><\/pre>\n\n<p>The main thread has to run for some time under one second to index the database of folders and put it into tensor. <\/p>\n\n<p>I tried to use sess.run([some_image]) before running tf.train.bath loop. In that case workers fail in the background first with the same error, and after that I receive my images.<\/p>\n\n<p>I tried to use time.sleep(), but it does not seem to be possible to delay the workers.<\/p>\n\n<p>I tried adding a dependency to the batch:<\/p>\n\n<pre><\/pre>\n\n<p>where init_one, and init_two are <\/p>\n\n<p>the most relevant issue I could find is at: <a href=\"https:\/\/github.com\/openai\/universe-starter-agent\/issues\/44\" rel=\"nofollow noreferrer\">https:\/\/github.com\/openai\/universe-starter-agent\/issues\/44<\/a><\/p>\n\n<p>Is there a way I could ask the synchronize worker threads with the main thread so that they don't race first and die out ?<\/p>\n\n<p>A similar and easy to reproduce error with variable initialization happens when set up the epoch counter to anything that is not  Are there any potential solutions ? I've added the code needed to reproduce the error below:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41703539,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-17T17:36:10.077","AcceptedAnswerId":null,"OwnerUserId":1861627.0,"Title":"Stable initialization in TensorFlow","Body":"<p>I have a strange issue with TensorFlow initialization. I have a computational graph that does some results and if I add a constant which is actually not connected to the input\/output flow I am testing, I have other results even if the  is the same. Any hint? Here follows the code and how to reproduce the issue:<\/p>\n\n<pre><\/pre>\n\n<p>Place the code in a  module an run:<\/p>\n\n<pre><\/pre>\n\n<p>now uncomment the line indicated above in the code and run:<\/p>\n\n<pre><\/pre>\n\n<p>you should see:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41703875,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-17T17:56:02.670","AcceptedAnswerId":41705347.0,"OwnerUserId":2494417.0,"Title":"Runtime dynamic recursion depth on variable size tensors","Body":"<p>In Tensorflow, what is the simplest way to recursively apply a set of ops where each recursive step takes as input the output of the last step? The critical feature here is that the number of recursion steps is specified at evaluation, and can vary between examples.<\/p>\n\n<p>For example, for input tensors , , and some function , I'd like to compute<\/p>\n\n<p><\/p>\n\n<p>But the same model should also be able to compute ,  as:<\/p>\n\n<p><\/p>\n\n<p>Unfortunately, while loops will not work, because they enforce strict shape invariants, meaning that all dimensions of tensor X have to be known at compile time, but I won't know the example size (first dimension of X) until runtime.<\/p>\n\n<p>Thanks for your help!<\/p>\n","answers":[{"AnswerId":"41705347","CreationDate":"2017-01-17T19:26:39.823","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>What you want is easily achieved in TensorFlow using while loops, namely the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#while_loop\" rel=\"nofollow noreferrer\"><code>tf.while_loop<\/code><\/a> construct. <code>tf.while_loop<\/code> has a <code>shape_invariants<\/code> parameter, which lets you specify a (possibly partial) shape for every loop  variable. That shape only has to be less specific than the shape that loop variable can take during the loop. Here is a complete example where <code>F<\/code> is appending a value to a tensor.<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\n\ni = tf.constant(0, dtype=tf.int32)\nx = tf.constant(0, shape=[0], dtype=tf.int32)\n\ni, x = tf.while_loop(lambda i, _: i &lt; 10,\n                     lambda i, x: (i + 1, tf.concat_v2([x, tf.expand_dims(i, 0)], 0)),\n                     [i, x],\n                     shape_variants=[i.get_shape(), tf.TensorShape([None])])\n\nprint(x.eval())\n\n[0 1 2 3 4 5 6 7 8 9]\n<\/code><\/pre>\n"}]}
{"QuestionId":41703901,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-17T17:57:42.067","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Batch Normalization - Tensorflow","Body":"<p>I have looked at a few BN examples but still am a bit confused. So I am currently using this function which calls the function here;<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard4\/tf.contrib.layers.batch_norm.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard4\/tf.contrib.layers.batch_norm.md<\/a><\/p>\n\n<pre><\/pre>\n\n<p>This following part is a toy run where I am just checking that the function reuses the means and variances calculated in a training step for two features. Running this part of the code in test mode i.e. , the running mean\/variances calculated in the training step are changing which can be seen when we print out the BN variables which I get from calling <\/p>\n\n<pre><\/pre>\n\n<p>The first three iterations of test loop look as follows;<\/p>\n\n<pre><\/pre>\n\n<p>I am not doing BP so beta and gamma won't change. However my running means\/variances are changing. Where am I going wrong?<\/p>\n\n<p>EDIT:\nIt would be good to know why these variables need\/do not need changing between test and train;<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41991858","CreationDate":"2017-02-01T23:36:57.930","ParentId":null,"OwnerUserId":"3128177","Title":null,"Body":"<p>Your bn function is wrong. Use this instead:<\/p>\n\n<pre><code>def bn(x,is_training,name):\n    return batch_norm(x, decay=0.9, center=True, scale=True,\n    updates_collections=None,\n    is_training=is_training,\n    reuse=None,\n    trainable=True,\n    scope=name)\n<\/code><\/pre>\n\n<p>is_training is bool 0-D tensor signaling whether to update running mean etc. or not. Then by just changing the tensor is_training you're signaling whether you're in training or test phase.<\/p>\n\n<p>EDIT:\nMany operations in tensorflow accept tensors, and not constant True\/False number arguments. <\/p>\n"},{"AnswerId":"47697246","CreationDate":"2017-12-07T14:26:44.683","ParentId":null,"OwnerUserId":"9067885","Title":null,"Body":"<p>When you use slim.batch_norm,be sure to use <code>slim.learning.create_train_op<\/code> instead of <code>tf.train.GradientDecentOptimizer(lr).minimize(loss)<\/code> or other optimizer. Try it to see if it works! <\/p>\n"}]}
{"QuestionId":41703944,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><neural-network>","CreationDate":"2017-01-17T17:59:58.213","AcceptedAnswerId":null,"OwnerUserId":6830219.0,"Title":"tensorflow, I want to change input image size","Body":"<p>I check tutorial of Tnesorflow. And now, I want to change input IMAGE_SIZE from 28x28 to 56x56 for improving accuracy. So, I change IMAGE_SIZE variable but that program throw error. Below is an original code and I want to change input image size. Where should I change? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41704066","CreationDate":"2017-01-17T18:08:03.260","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>You use some hardcoded numbers while building the model. Please change them to use <code>IMAGE_SIZE<\/code> instead as follows :<\/p>\n\n<pre><code># \u5165\u529b\u309228x28x3\u306b\u5909\u5f62\nx_image = tf.reshape(images_placeholder, [-1, IMAGE_SIZE, IMAGE_SIZE, 3])\n<\/code><\/pre>\n"},{"AnswerId":"41704107","CreationDate":"2017-01-17T18:09:42.463","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There are at least two other places in the code that depend on the image size:<\/p>\n\n<ol>\n<li><p>The definition of <code>x_image<\/code> hard-codes the image size:<\/p>\n\n<pre><code>x_image = tf.reshape(images_placeholder, [-1, 28, 28, 3])\n<\/code><\/pre>\n\n<p>Assuming you set <code>IMAGE_SIZE<\/code> to be 56, you should replace it with:<\/p>\n\n<pre><code>x_image = tf.reshape(images_placeholder, [-1, IMAGE_SIZE, IMAGE_SIZE, 3])\n<\/code><\/pre><\/li>\n<li><p>The number of neurons in the output fully connected layer depends on the image size (downsampled by the pooling layers), and will increase by 4x when you increase the number of pixels in the input by 4x. The following lines:<\/p>\n\n<pre><code>W_fc1 = weight_variable([7*7*64, 1024])\nb_fc1 = bias_variable([1024])\nh_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])\n<\/code><\/pre>\n\n<p>...should be replaced with:<\/p>\n\n<pre><code>W_fc1 = weight_variable([14 * 14 * 64, 1024])\nb_fc1 = bias_variable([1024])\nh_pool2_flat = tf.reshape(h_pool2, [-1, 14 * 14 *64])\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":41704484,"AnswerCount":3,"Tags":"<math><machine-learning><tensorflow>","CreationDate":"2017-01-17T18:32:57.587","AcceptedAnswerId":44585250.0,"OwnerUserId":5708247.0,"Title":"What is difference between tf.truncated_normal and tf.random_normal?","Body":"<p> outputs random values from a normal distribution.<\/p>\n\n<p> outputs random values from a truncated normal distribution.<\/p>\n\n<p>I tried googling 'truncated normal distribution'. But didn't understand much.<\/p>\n","answers":[{"AnswerId":"43726140","CreationDate":"2017-05-01T20:36:30.547","ParentId":null,"OwnerUserId":"6031995","Title":null,"Body":"<p><code>tf.truncated_normal()<\/code> selects random numbers from a normal distribution whose mean is close to 0 and values are close to 0. For example, from -0.1 to 0.1. It's called truncated because your cutting off the tails from a normal distribution.<\/p>\n\n<p><code>tf.random_normal()<\/code> selects random numbers from a normal distribution whose mean is close to 0, but values can be a bit further apart. For example, from -2 to 2.<\/p>\n\n<p>In machine learning, in practice, you usually want your weights to be close to 0.<\/p>\n"},{"AnswerId":"41704789","CreationDate":"2017-01-17T18:53:05.973","ParentId":null,"OwnerUserId":"6850023","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/truncated_normal\" rel=\"nofollow noreferrer\">API documentation for tf.truncated_normal()<\/a> describes the function as:<\/p>\n\n<blockquote>\n  <p>Outputs random values from a truncated normal distribution.<\/p>\n  \n  <p>The generated values follow a normal distribution with specified mean and standard deviation, except that values whose magnitude is\n  more than 2 standard deviations from the mean are dropped and\n  re-picked.<\/p>\n<\/blockquote>\n"},{"AnswerId":"44585250","CreationDate":"2017-06-16T09:08:35.933","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/truncated_normal\" rel=\"noreferrer\">documentation<\/a> says it all:\nFor truncated normal distribution:<\/p>\n\n<blockquote>\n  <p>The generated values follow a normal distribution with specified mean\n  and standard deviation, except that values whose magnitude is more\n  than 2 standard deviations from the mean are dropped and re-picked.<\/p>\n<\/blockquote>\n\n<p>Most probably it is easy to understand the difference by plotting the graph for yourself (%magic is because I use jupyter notebook):<\/p>\n\n<pre><code>import tensorflow as tf\nimport matplotlib.pyplot as plt\n\n%matplotlib inline  \n\nn = 500000\nA = tf.truncated_normal((n,))\nB = tf.random_normal((n,))\nwith tf.Session() as sess:\n    a, b = sess.run([A, B])\n<\/code><\/pre>\n\n<p>And now <\/p>\n\n<pre><code>plt.hist(a, 100, (-4.2, 4.2));\nplt.hist(b, 100, (-4.2, 4.2));\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fXqXs.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fXqXs.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<hr>\n\n<p>The point for using truncated normal is to overcome saturation of tome functions like sigmoid (where if the value is too big\/small, the neuron stops learning).<\/p>\n"}]}
{"QuestionId":41704617,"AnswerCount":2,"Tags":"<deep-learning><keras>","CreationDate":"2017-01-17T18:41:02.490","AcceptedAnswerId":42056276.0,"OwnerUserId":1934212.0,"Title":"Exception extending keras VGG16 model : self.assert_input_compatibility(x) throws","Body":"<p>In order to use a VGG16 network for a regression task, I extend it in the following way: <\/p>\n\n<pre><\/pre>\n\n<p>Which immediately throws the following exception: <\/p>\n\n<pre><\/pre>\n\n<p>Do I have to change the architecture, or is there a problem with the inference of input dimensions?<\/p>\n","answers":[{"AnswerId":"42056276","CreationDate":"2017-02-05T19:29:31.913","ParentId":null,"OwnerUserId":"54858","Title":null,"Body":"<p>I dug into this; I believe it's a bug in the latest version when you call set_learning_phase; I submitted <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/5268\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/5268<\/a> and <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/5269\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/pull\/5269<\/a><\/p>\n\n<p>You can roll back to 1.2.0 as a temporary fix.<\/p>\n"},{"AnswerId":"41722430","CreationDate":"2017-01-18T14:48:41.563","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Worked for me:<\/p>\n\n<pre><code>input_tensor = Input((3, 224, 224))\nmodel = keras.applications.vgg16.VGG16(input_tensor=input_tensor, include_top=False)\nx = model.output\nx = model.output\nx = Flatten()(x)\nx = Dense(1024, activation='relu')(x)\nx = Dropout(0.5)(x)\nx = Dense(512,activation='relu')(x)\nx = Dropout(0.5)(x)\nx = Dense(256,activation='relu')(x)\nx = Dropout(0.5)(x)\nx = Dense(1)(x)\nmodel2 = Model(model.input, x)\nimport numpy as np\nmodel2.predict(np.zeros((1, 3, 224,224)))\n<\/code><\/pre>\n\n<p>Output: array([[0.11504173]], dtype=float32)<\/p>\n"}]}
{"QuestionId":41704963,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-01-17T19:02:26.823","AcceptedAnswerId":null,"OwnerUserId":5257450.0,"Title":"How can I apply convolution over the undefined shape in Keras or Tensorflow?","Body":"<p>With Keras, I defined,<\/p>\n\n<p>, <\/p>\n\n<p>where I have the shape of  be <\/p>\n\n<p>And I met the problem with<\/p>\n\n<blockquote>\n  <p>TypeError: unsupported operand type(s) for *: 'NoneType' and 'int'<\/p>\n<\/blockquote>\n\n<p>Since the convolutional weight matrices applied on each window of the 2D image share parameters, it is valid at least for me to do it without knowing the true dimension of the image.<\/p>\n","answers":[{"AnswerId":"41714913","CreationDate":"2017-01-18T08:52:23.130","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>If you want to process input of arbitrary size in your CNN define first layer as follows:<\/p>\n\n<pre><code>Convolution2D(num_filters, ws, ws, input_shape=(nb_channels, None, None))\n<\/code><\/pre>\n\n<p>Note that it takes much longer time to train a network with variable input size compared to fixed. See my question <a href=\"https:\/\/stackoverflow.com\/questions\/41322286\/training-a-fully-convolutional-neural-network-with-inputs-of-variable-size-takes\">here<\/a>.<\/p>\n"}]}
{"QuestionId":41705315,"AnswerCount":1,"Tags":"<tensorflow><tflearn>","CreationDate":"2017-01-17T19:24:32.607","AcceptedAnswerId":41913521.0,"OwnerUserId":697065.0,"Title":"Using a custom layer to define connectivity in tensorflow \/ tflearn","Body":"<p>Instead of a fully connected layer, I would like to specify the connectivity between activation nodes using a matrix. For example:<\/p>\n\n<p>I have a 20 node layer that is connected to a 10 node layer. Using a typical fully connected layer, my  matrix is 20 x 10, with a  vector of size 10. <\/p>\n\n<p>My activation looks something like <\/p>\n\n<p>If I had a matrix of ones and zeros that was the same size as , lets call it , I could do a pairwise multiply between  and  to remove connections between my first layer (20 nodes) and my second layer (10 nodes)<\/p>\n\n<p>Here is my current code:<\/p>\n\n<pre><\/pre>\n\n<p>Where my_fun gives me:  and  is a pairwise multiplication<\/p>\n\n<p>How do I create this function? I cant seem to find tflearn examples on how it is done, but I also know that tflearn allows for base tensorflow functions as well<\/p>\n","answers":[{"AnswerId":"41913521","CreationDate":"2017-01-28T18:50:49.837","ParentId":null,"OwnerUserId":"697065","Title":null,"Body":"<p>Its hard to do this strictly with tflearn, but if you're willing to include base tensorflow operations, its trivial:<\/p>\n\n<pre><code>F.shape\n# (20, 10)\nimport tflearn\nimport tensorflow as tf\n\ninput = tflearn.input_data(shape=[None, num_input])\ntf_F = tf.constant(F, shape=[20, 10])\n\nfirst = tflearn.fully_connected(input, 20, activation='relu')\n# Here is where I want to use a custom function, that uses my F matrix\n# I want only connections that are ones (and not zeros) in F\n\n# Old:\n# second = tflearn.fully_connected(first, 10, activation='relu')\n# Solution:\nW = tf.Variable(tf.random_uniform([20, 10]), name='Weights')\nb = tf.Variable(tf.zeros([10]), name='biases')\nW_filtered = tf.mul(tf_F, W)\nsecond = tf.matmul( W_filtered, first) + b\n<\/code><\/pre>\n"}]}
{"QuestionId":41705377,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-17T19:28:31.293","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"What is the right initializer one should give to the tf.contrib.layers.convolution2d function in TensorFlow?","Body":"<p>I was reading the documentation for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.layers\/higher_level_ops_for_building_neural_network_layers_#convolution2d\" rel=\"nofollow noreferrer\">making 2d convolutional layers<\/a> in tensorflow from the contrib section and was wondering what was the right or best way to initialize the weights when using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.layers\/higher_level_ops_for_building_neural_network_layers_#convolution2d\" rel=\"nofollow noreferrer\">tf.contrib.layers.convolution2d<\/a> function. Unfortunately they don't really say explicitly nor provide an example, so it was unclear to me what is the intended way to use this. The function has a  parameter which can be set. I have tried setting it to both:<\/p>\n\n<ol>\n<li>tf.contrib.layers.xavier_initializer<\/li>\n<li>tf.contrib.layers.xavier_initializer_conv2d<\/li>\n<\/ol>\n\n<p>neither seem to return an error and the first one seems to train fine (as far as I can tell). However, it would be awesome to check if this is the right way of using this contrib layer (or maybe since it seems to be a contrib function, how does one check the \"official\" source code maybe to see their docs or test cases or maybe address the my question in their gitissues, if appropriate).<\/p>\n","answers":[{"AnswerId":"41707929","CreationDate":"2017-01-17T22:11:44.323","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>At the time of writing, tf.contrib.layers.xavier_initializer_conv2d is an alias for tf.contrib.layers.xavier_initializer (The source at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/initializers.py#L59\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/initializers.py#L59<\/a> has a line that is literally xavier_initializer_conv2d = xavier_initializer) so it would be very surprising if one worked and the other did not :) <\/p>\n\n<p>So, one is as good as the other (both are fine). <\/p>\n"}]}
{"QuestionId":41706218,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-01-17T20:20:02.797","AcceptedAnswerId":null,"OwnerUserId":893697.0,"Title":"Tensorflow forward propagation extremely slow","Body":"<p>I'm encountering an issue in using tensorflow (0.12.1 and 0.11.0rc1 on two different machines), namely manually performing forward propagation and gradient computation being extremely slow (training works fine on the other hand).<\/p>\n\n<p>In particular, calling  takes between 1 and 1.2 seconds for 1 image. For reference, the same architecture implemented in caffe' takes around 100-150 ms, which is coherent with the times reported in the paper I have reimplemented. A similar amount of time (~1.5s) is required for the backpropagation through the network via . Again, the same architecture in caffe takes around 1\/10 of that time.<\/p>\n\n<p>I checked the GPU usage during the forward prop to be sure that it was actually using the GPU and it correctly spikes during the  (using a GTX TITAN Z with cuda 8 and a TITAN X with cuda 7.5 on the two machines respectively) <\/p>\n\n<p>The fact that the same architecture (a fairly straightforward conv-deconv, <a href=\"https:\/\/arxiv.org\/abs\/1603.06041\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1603.06041<\/a>) in caffe works fine makes me think that there is something wrong in either my tensorflow code or configuration, but I have no clue on where to start. Any suggestions? <\/p>\n\n<p>Thanks!<\/p>\n\n<p>Edit: link at the code added.<\/p>\n\n<p>Tensorflow network and forward function: <a href=\"http:\/\/pastebin.com\/y9jKBdsV\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/y9jKBdsV<\/a> (disregard the fact that the forward function looks like a dcgan, the problem persists regardless)<\/p>\n\n<p>Caffe prototxt and forward function: <a href=\"http:\/\/pastebin.com\/DqWdbHU8\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/DqWdbHU8<\/a><\/p>\n","answers":[]}
{"QuestionId":41707385,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-01-17T21:32:43.633","AcceptedAnswerId":41707470.0,"OwnerUserId":5257450.0,"Title":"TypeError: 'Tensor' object does not support item assignment","Body":"<pre><\/pre>\n\n<blockquote>\n  <p>---> 15             output[i,j,k,:] = pooling(conv, 0)<\/p>\n  \n  <p>TypeError: 'Tensor' object does not support item assignment<\/p>\n<\/blockquote>\n\n<p>In the code I implemented above, each  return us a , how am I suppose to pack these tensors into a larger one with shape I defined in ?<\/p>\n","answers":[{"AnswerId":"41707470","CreationDate":"2017-01-17T21:38:35.817","ParentId":null,"OwnerUserId":"5257450","Title":null,"Body":"<pre><code>output = []\nfor i, atten_embed in enumerate(atten_embeds):\n    for j, ws in enumerate(wss):\n        conv_layer = conv_layers_A[j]\n        conv = conv_layer(atten_embed)\n        new_shape = (reduce(lambda x,y:x*y, conv.get_shape()[:-1]).value,num_filters)\n        conv = K.reshape(conv, new_shape)\n        for k, pooling in enumerate([K.max, K.min, K.mean]):\n            output.append(pooling(conv, 0))\n\noutput = tf.reshape(tf.pack(output), shape=(2, len(wss), 3, num_filters))\n<\/code><\/pre>\n"}]}
{"QuestionId":41707621,"AnswerCount":3,"Tags":"<python><deep-learning><theano><keras><mean-square-error>","CreationDate":"2017-01-17T21:48:58.767","AcceptedAnswerId":null,"OwnerUserId":3120870.0,"Title":"Keras mean squared error loss layer","Body":"<p>I am currently implementing a custom loss layer and in the process, I stumbled upon the implementation of mean squared error in the objectives.py file [1]. I know I'm missing something in my understanding of this loss calculation because I always thought that the average was done separately across the samples for each output in each mini-batch (axis 0 of the tensor) but it appears that the average is actually being done across the last axis, which in a single vector, would mean it's being done across the outputs. I found this by accident while working on my custom loss layer because it requires discounting the loss of a few of the outputs it a training output in a specific place is a specific value. Anyways, is my understanding of the mean squared error incorrect? Why would Keras be using the last axis and thus turning a a 1xn output vector into a 1x1 output vector?<\/p>\n\n<p>Thanks.<\/p>\n\n<p>[1] <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/objectives.py#L7\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/objectives.py#L7<\/a><\/p>\n","answers":[{"AnswerId":"48735647","CreationDate":"2018-02-11T19:39:46.487","ParentId":null,"OwnerUserId":"7248396","Title":null,"Body":"<p>Let's detail the steps of how the losses are computed in Keras to show that the <code>axis=-1<\/code> in all the loss computations are correct : <\/p>\n\n<ul>\n<li><p>So we pick a loss in <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/losses.py#L13\" rel=\"nofollow noreferrer\">losses.py<\/a> that we will pass to the <code>compile<\/code> method of our model. <\/p><\/li>\n<li><p>In <code>compile<\/code>, the total loss is computed. It happens in several steps : <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py#L639\" rel=\"nofollow noreferrer\">The first step<\/a> creates a list of losses, one for each output of the model. <\/p><\/li>\n<li>This first step calls <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py#L399\" rel=\"nofollow noreferrer\">_weighted_masked_objective<\/a> which according to the docs 'Adds support for masking and sample-weighting to an objective function'<\/li>\n<li>Basically, <code>_weighted_masked_objective<\/code> returns a new objective functions which take into account the <code>weights<\/code> and <code>mask<\/code> parameters which the user will provide when using the method <code>fit<\/code>. <\/li>\n<\/ul>\n\n<p>If I cut the code to have only the lines that matter for the question, we get to something like that. <\/p>\n\n<pre><code>def _weighted_masked_objective(fn):\n    def weighted(y_true, y_pred, weights, mask=None):\n          score_array = fn(y_true, y_pred) # Compute loss as in losses.py\n          return K.mean(score_array) # Average over all axis\n\nclass Model(Container):\n    def compile(self, optimizer, loss, metrics=None, loss_weights=None,\n                sample_weight_mode=None, weighted_metrics=None,\n                target_tensors=None, **kwargs):\n        weighted_losses = [_weighted_masked_objective(fn) for fn in loss_functions]\n<\/code><\/pre>\n\n<p>So at the end, the loss is indeed averaged over every dimension, and the use of <code>axis=-1<\/code> is just an elegant way to enable masking and weighting of the loss at another point in the code<\/p>\n\n<p>NB : I didn't explain the other steps because they don't contribute to answering the question. <\/p>\n"},{"AnswerId":"41728435","CreationDate":"2017-01-18T20:02:17.947","ParentId":null,"OwnerUserId":"3120870","Title":null,"Body":"<p>I believe, after some conversations with coworkers, that I understand this situation and have a proper solution to the problem. Though I knew that Theano was providing lazy-evaluated tensor functions that were running the matrix operations on the GPU, what I did not realize was that Keras's loss functions are actually written in a way where the compiled theano execution graph is smart enough to cache certain values in order to properly back-propagate the loss values back throughout the network. Because of the type of network I'm creating, I dived into writing my own custom loss function without a completely understanding of how Theano actually treats the loss after it's been calculated by the function.<\/p>\n\n<p>From what I can tell, my concern was correct that Keras' use of the last axis is a problem. In my case, I have a fully-convolutional deep neural network and the input to the loss function is (x, 7, 16, 16) where the x is the size of the mini-batch. Normally, neural networks output a matrix where the first dimension is the mini-batch size and the second (usually last) dimension is the actual size of the output vector. Because of this, using the last axis in the output tensor to do the actual \"mean\" portion of the mean squared error is not correct. Instead, the axis should be 1 (in the case of zero-based indexing) because it's the 7 actual regression output features that need to be differentiated for back-propagation. <\/p>\n\n<p>I originally knew that the axis = -1 may not be correct and the reason I posted this question was because I couldn't quite explain why. It's been a long time since I've had to dive into the math behind the neural networks but when I finally did, I was able to resolve the gaps (<em>I think<\/em>).  I'm posting this response here for future people who may experience this same problem or gap in their understanding of Theano's tensor framework. <\/p>\n"},{"AnswerId":"41708000","CreationDate":"2017-01-17T22:17:10.113","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The code in question for the MSE loss is this:<\/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>Here first y_pred and y_true are subtracted, then that result is passed to K.square, which as expected, returns the square of its parameter, and then that result is given to K.mean, which computes the mean.<\/p>\n\n<p>So the code clearly is doing what its supposed to do. About why the last axis is operated upon, this has nothing to do with classes, it is just a convention. Note that in general, there are no classes in the MSE definition.<\/p>\n"}]}
{"QuestionId":41708106,"AnswerCount":3,"Tags":"<tensorflow><metadata><embedding><tensorboard><tflearn>","CreationDate":"2017-01-17T22:25:00.720","AcceptedAnswerId":44912512.0,"OwnerUserId":1335050.0,"Title":"Linking Tensorboard Embedding Metadata to checkpoint","Body":"<p>I'm using the tflearn wrapper over tensorflow to build a model, and would like to add metadata (labels) to the resultant embedding visualization. Is there a way to link a metadata.tsv file to a saved checkpoint after the fact of running it?  <\/p>\n\n<p>I've created a projector_config.pbtxt file in the logdir of the checkpoint summaries, with the metadata.tsv being in the same folder. The config looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>and was created using the code from the docs - <a href=\"https:\/\/www.tensorflow.org\/how_tos\/embedding_viz\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/embedding_viz\/<\/a> <\/p>\n\n<p>I've commented out the tf.Session part in the hopes of creating the metadata link without the need of doing so directly within a Session object, but I'm not sure if that's possible.<\/p>\n\n<pre><\/pre>\n\n<p>Below is a snap of the current embedding visualization. Note the empty metadata. Is there a way to directly attach the desired metafile to this embedding?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ANcPY.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ANcPY.png\" alt=\"Embedding Visualization\"><\/a><\/p>\n","answers":[{"AnswerId":"44912512","CreationDate":"2017-07-04T18:53:13.553","ParentId":null,"OwnerUserId":"8255388","Title":null,"Body":"<p>I had the same problem and it is soloved now :)<\/p>\n\n<p>Essentially, all you need to do is following 3 steps:<\/p>\n\n<ol>\n<li>save model checkpoint, supposing ckeckpoint's directory is <code>ckp_dir<\/code>;<\/li>\n<li>place <em>projector_config.pbtxt<\/em> and <em>metadata.tsv<\/em> in <code>ckp_dir<\/code>;<\/li>\n<li>run <code>tensorboard --logdir=ckp_dir<\/code> and click the Embedding Tab<\/li>\n<\/ol>\n\n<p>the content of <em>projector_config.pbtxt<\/em> is :<\/p>\n\n<pre><code>    embeddings {\n      tensor_name: \"embedding_name\"\n      metadata_path: \"metatdata.tsv\"\n    }\n<\/code><\/pre>\n\n<p>This is the key to link the embedding to metadata.tsv. In tf.Session(), we often get the embedding's value like <code>sess.run('embedding_name:0')<\/code>. But in <em>projector_config.pbtxt<\/em>, we just type <code>tensor_name: \"embedding_name\"<\/code>.<\/p>\n\n<p>Generally, we can specify the checkpoint path and metadata_path in <em>projector_config.pbtxt<\/em> so that we can place <em>checkpoint<\/em>, <em>projector_config.pbtxt<\/em> and <em>metadata.tsv<\/em> in different directories. But i think it is too complicated. I just solved it as above.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ip6qB.png\" rel=\"nofollow noreferrer\">the result shown here<\/a><\/p>\n"},{"AnswerId":"42920936","CreationDate":"2017-03-21T07:40:00.160","ParentId":null,"OwnerUserId":"7664921","Title":null,"Body":"<p>Try this with your <code>projector_config.pbtxt<\/code>:<\/p>\n\n<pre><code>embeddings {\n  tensor_name: \"Embedding\/W\"\n  metadata_path: \"$LOGDIR\/metadata.tsv\"\n}\n<\/code><\/pre>\n\n<p>Make sure your <code>$LOGDIR<\/code> is the same path you use to call <code>tensorboard --logdir=$LOGDIR<\/code> on your terminal; that is, it should be relative to your current directory (so it probably shouldn't include <code>C:\/..<\/code>). Also include the filename in the <code>metadata_path<\/code>.<\/p>\n\n<p>Let me know if this works for you, too.<\/p>\n\n<hr>\n\n<p>I stumbled upon the same problem trying to display words instead of indices for the <a href=\"https:\/\/www.tensorflow.org\/code\/tensorflow_models\/tutorials\/embedding\/word2vec.py\" rel=\"nofollow noreferrer\">word2vec tutorial<\/a>. To achieve that your <code>projector_config.pbtxt<\/code> should look like this:<\/p>\n\n<pre><code>embeddings {\n  tensor_name: \"w_in\"\n  metadata_path: \"$LOGDIR\/vocab.txt\"\n}\n<\/code><\/pre>\n\n<p>You might also want to modify the <code>save_vocab<\/code> function in the code linked above since, as is, it converts unicode to hex.<\/p>\n"},{"AnswerId":"41964036","CreationDate":"2017-01-31T17:39:42.603","ParentId":null,"OwnerUserId":"1129760","Title":null,"Body":"<p>I'm having the same problem. [EDIT:] The way I can make it work, is by creating a subdir and putting all the checkpoint files there, using a full path to the metadata file. The 'trick' is that then if you give tensorboard --logdir that specific dir, it somehow parses the metadata and displays the words in the plot instead of just the indices. The downside is that then then embeddings aren't found anymore when you just use the base log dir as --logdir, so you always have to start a separate tensorboard instance to view the embeddings.<\/p>\n\n<p>It's really quite annoying and I can't imagine this is the only way to make it work, but I spent hours before that one finally worked...<\/p>\n"}]}
{"QuestionId":41708506,"AnswerCount":1,"Tags":"<neural-network><keras><keras-layer>","CreationDate":"2017-01-17T22:55:00.693","AcceptedAnswerId":41717406.0,"OwnerUserId":7318346.0,"Title":"Multi-Output Multi-Class Keras Model","Body":"<p>For each input I have, I have a 49x2 matrix associated. Here's what 1 input-output couple looks like<\/p>\n\n<pre><\/pre>\n\n<p>Where both Label1 and Label2 are LabelEncode and they have respectively 1200 and 1300 different classes.<\/p>\n\n<p>Just to make sure this is what we call a multi-output multi-class problem?<\/p>\n\n<p>I tried to flatten the output but I feared the model wouldn't understand that all similar Label share the same classes.<\/p>\n\n<p>Is there a Keras layer that handle output this peculiar array shape?<\/p>\n","answers":[{"AnswerId":"41717406","CreationDate":"2017-01-18T10:50:09.300","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Generally, multi-class problems correspond with models outputting a probability distribution over the set of classes (that is typically scored against the one-hot encoding of the actual class through cross-entropy). Now, independently of whether you are structuring it as one single output, two outputs, 49 outputs or 49 x 2 = 98 outputs, that would mean having 1,200 x 49 + 1,300 x 49 = 122,500 output units - which is not something a computer cannot handle, but maybe not the most convenient thing to have. You could try having each class output to be a single (e.g. linear) unit and round it's value to choose the label, but, unless the labels have some numerical meaning (e.g. order, sizes, etc.), that is not likely to work.<\/p>\n\n<p>If the order of the elements in the input has some meaning (that is, shuffling it would affect the output), I think I'd approach the problem through an RNN, like an LSTM or a bidirectional LSTM model, with two outputs. Use <code>return_sequences=True<\/code> and <code>TimeDistributed<\/code> <code>Dense<\/code> softmax layers for the outputs, and for each 118-long input you'd have 118 pairs of outputs; then you can just use temporal sample weighting to drop, for example, the first 69 (or maybe do something like dropping the 35 first and the 34 last if you're using a bidirectional model) and compute the loss with the remaining 49 pairs of labellings. Or, if that makes sense for your data (maybe it doesn't), you could go with something more advanced like <a href=\"http:\/\/www.gitxiv.com\/posts\/9hLxYvJLCu9Z9N4pK\/ctc-connectionist-temporal-classification\" rel=\"nofollow noreferrer\">CTC<\/a> <s>(although Keras does not have it, I'm trying to integrate <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/connectionist_temporal_classification__ctc_\" rel=\"nofollow noreferrer\">TensorFlow implementation<\/a> into it without much sucess)<\/s>, which is <a href=\"https:\/\/keras.io\/backend\/#ctc_batch_cost\" rel=\"nofollow noreferrer\">also implemented in Keras<\/a> (thanks <a href=\"https:\/\/stackoverflow.com\/users\/2217210\/indraforyou\">@indraforyou<\/a>)!.<\/p>\n\n<p>If the order in the input has no meaning but the order of the outputs does, then you could have an RNN where your input is the original 118-long vector plus a pair of labels (each one-hot encoded), and the output is again a pair of labels (again two softmax layers). The idea would be that you get one \"row\" of the 49x2 output on each frame, and then you feed it back to the network along with the initial input to get the next one; at training time, you would have the input repeated 49 times along with the \"previous\" label (an empty label for the first one).<\/p>\n\n<p>If there are no sequential relationships to exploit (i.e. the order of the input and the output do not have a special meaning), then the problem would only be truly represented by the initial 122,500 output units (plus all the hidden units you may need to get those right). You could also try some kind of middle ground between a regular network and a RNN where you have the two softmax outputs and, along with the 118-long vector, you include the \"id\" of the output that you want (e.g. as a 49-long one-hot encoded vector); if the \"meaning\" of each label at each of the 49 outputs is similar, or comparable, it may work.<\/p>\n"}]}
{"QuestionId":41708572,"AnswerCount":3,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-17T23:00:48.233","AcceptedAnswerId":null,"OwnerUserId":3993270.0,"Title":"TensorFlow: questions regarding tf.argmax() and tf.equal()","Body":"<p>I am learning the TensorFlow, building a multilayer_perceptron model. I am looking into some examples like the one at: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3_NeuralNetworks\/multilayer_perceptron.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3_NeuralNetworks\/multilayer_perceptron.ipynb<\/a><\/p>\n\n<p>I then have some questions in the code below:<\/p>\n\n<pre><\/pre>\n\n<p>I am wondering what do  and  mean and return (type and value) exactly? And is  a variable instead of real values?<\/p>\n\n<p>Finally, how do we get the  array (the prediction result when the input data is ) from the tf session? Thanks a lot!<\/p>\n","answers":[{"AnswerId":"41863099","CreationDate":"2017-01-25T22:46:35.670","ParentId":null,"OwnerUserId":"3950213","Title":null,"Body":"<pre><code>tf.argmax(input, axis=None, name=None, dimension=None)\n<\/code><\/pre>\n\n<p>Returns the index with the largest value across axis of a tensor.<\/p>\n\n<p>input is a Tensor and axis describes which axis of the input Tensor to reduce across. For vectors, use axis = 0.<\/p>\n\n<p>For your specific case let's use two arrays and demonstrate this <\/p>\n\n<pre><code>pred = np.array([[31, 23,  4, 24, 27, 34],\n                [18,  3, 25,  0,  6, 35],\n                [28, 14, 33, 22, 20,  8],\n                [13, 30, 21, 19,  7,  9],\n                [16,  1, 26, 32,  2, 29],\n                [17, 12,  5, 11, 10, 15]])\n\ny = np.array([[31, 23,  4, 24, 27, 34],\n                [18,  3, 25,  0,  6, 35],\n                [28, 14, 33, 22, 20,  8],\n                [13, 30, 21, 19,  7,  9],\n                [16,  1, 26, 32,  2, 29],\n                [17, 12,  5, 11, 10, 15]])\n<\/code><\/pre>\n\n<p>Evaluating <code>tf.argmax(pred, 1)<\/code> gives a tensor whose evaluation will give <code>array([5, 5, 2, 1, 3, 0])<\/code><\/p>\n\n<p>Evaluating <code>tf.argmax(y, 1)<\/code> gives a tensor whose evaluation will give <code>array([5, 5, 2, 1, 3, 0])<\/code><\/p>\n\n<pre><code>tf.equal(x, y, name=None) takes two tensors(x and y) as inputs and returns the truth value of (x == y) element-wise. \n<\/code><\/pre>\n\n<p>Following our example, <code>tf.equal(tf.argmax(pred, 1),tf.argmax(y, 1))<\/code> returns a tensor whose evaluation will give<code>array(1,1,1,1,1,1)<\/code>.<\/p>\n\n<p>correct_prediction is a tensor whose evaluation will give a 1-D array of 0's and 1's <\/p>\n\n<p>y_test_prediction can be obtained by executing <code>pred = tf.argmax(logits, 1)<\/code> <\/p>\n\n<p>The documentation for tf.argmax and tf.equal can be accessed by following the links below. <\/p>\n\n<p>tf.argmax() <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/sequence_comparison_and_indexing#argmax\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/sequence_comparison_and_indexing#argmax<\/a><\/p>\n\n<p>tf.equal() <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops\/comparison_operators#equal\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops\/comparison_operators#equal<\/a><\/p>\n"},{"AnswerId":"41709288","CreationDate":"2017-01-18T00:10:31.103","ParentId":null,"OwnerUserId":"3799743","Title":null,"Body":"<p><strong>tf.argmax(input, axis=None, name=None, dimension=None)<\/strong> <\/p>\n\n<blockquote>\n  <p>Returns the index with the largest value across axis of a tensor.<\/p>\n<\/blockquote>\n\n<p>For the case in specific, it receives <code>pred<\/code> as argument for it's <code>input<\/code> and <code>1<\/code> as <code>axis<\/code>. The axis describes which axis of the input Tensor to reduce across. For vectors, use axis = 0.<\/p>\n\n<p>Example: Given the list <code>[2.11,1.0021,3.99,4.32]<\/code> argmax will return <code>3<\/code> which is the index of the highest value.<\/p>\n\n<hr>\n\n<p><strong>correct_prediction<\/strong> is a tensor that will be evaluated later. It is not a regular python variable. It contains the necessary information to compute the value later.\nFor this specific case, it will be part of another tensor <code>accuracy = tf.reduce_mean(tf.cast(correct_prediction, \"float\"))<\/code> and will be evaluated by <code>eval<\/code> on <code>accuracy.eval({x: X_test, y: y_test_onehot})<\/code>.<\/p>\n\n<hr>\n\n<p><strong>y_test_prediction<\/strong> should be your <code>correct_prediction<\/code> tensor.<\/p>\n"},{"AnswerId":"43537345","CreationDate":"2017-04-21T08:11:58.987","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>Reading the documentation:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/argmax\" rel=\"noreferrer\"><code>tf.argmax<\/code><\/a> <\/p>\n\n<blockquote>\n  <p>Returns the index with the largest value across axes of a tensor.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/equal\" rel=\"noreferrer\"><code>tf.equal<\/code><\/a><\/p>\n\n<blockquote>\n  <p>Returns the truth value of (x == y) element-wise.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/cast\" rel=\"noreferrer\"><code>tf.cast<\/code><\/a><\/p>\n\n<blockquote>\n  <p>Casts a tensor to a new type.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_mean\" rel=\"noreferrer\"><code>tf.reduce_mean<\/code><\/a><\/p>\n\n<blockquote>\n  <p>Computes the mean of elements across dimensions of a tensor.<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>Now you can easily explain what it does. Your <code>y<\/code> is one-hot encoded, so it has one 1 and all other are zero. Your <code>pred<\/code> represents probabilities of classes. So argmax finds the positions of best prediction and correct value. After that you check whether they are the same.<\/p>\n\n<p>So now your <code>correct_prediction<\/code> is a vector of True\/False values with the size equal to the number of instances you want to predict. You convert it to floats and take the average.<\/p>\n\n<hr>\n\n<p>Actually this part is nicely explained in <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/pros\" rel=\"noreferrer\">TF tutorial<\/a> in the <strong>Evaluate the Model<\/strong> part<\/p>\n"}]}
{"QuestionId":41708865,"AnswerCount":1,"Tags":"<python-2.7><machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-01-17T23:28:25.020","AcceptedAnswerId":41723032.0,"OwnerUserId":6137334.0,"Title":"Keras convolution shape has dimensions out of order (Error when checking model input)","Body":"<p>I've checked all the similar posts, but my error isn't being fixed with the fixes suggested. Thanks in advance for any help!<\/p>\n\n<p>I'm using a tensorflow backend with Keras, and my images have dimensions 1185 by 676. Most of the code is from one of the Keras examples.<\/p>\n\n<p>I'm getting  This error disappears when I switch to dim_ordering=\"th\", which is odd, considering that I'm using tensorflow, not theano.<\/p>\n\n<p>The code up to this point:<\/p>\n\n<pre><\/pre>\n\n<p>And just in case the data generation is part of the issue:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41723032","CreationDate":"2017-01-18T15:17:58.890","ParentId":null,"OwnerUserId":"3111778","Title":null,"Body":"<p>The image dimension ordering has been mixed in your code. Multiple way to fix this. <\/p>\n\n<p>One way is to add<\/p>\n\n<pre><code>from keras import backend as K\nK.set_image_dim_ordering('tf')\n<\/code><\/pre>\n\n<p>at the beginning of your code.<\/p>\n\n<p>Other approaches are summarized <a href=\"https:\/\/stackoverflow.com\/questions\/40135370\/keras-getting-wrong-output-shape\/40137162#40137162\">in this answer<\/a><\/p>\n"}]}
{"QuestionId":41709330,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-18T00:16:31.100","AcceptedAnswerId":null,"OwnerUserId":6227294.0,"Title":"How register tf.nn.max_pool_with_argmax in tensorflow?","Body":"<p>I wanna use tf.nn.max_pool_with_argmax in tensorflow, but it seems its gradient is not registered <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/nn_grad.py#L399\" rel=\"nofollow noreferrer\">here<\/a>, while it has defined <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/ops\/nn_ops.cc#L1316\" rel=\"nofollow noreferrer\">here<\/a>. I am trying to register it myself from this <a href=\"https:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/\" rel=\"nofollow noreferrer\">tutorial<\/a>, but it is a little hard for me to follow what they are doing. (<a href=\"https:\/\/stackoverflow.com\/questions\/41384167\/no-gradient-defined-for-operation-maxpoolwithargmax-in-tensorflow\">Here<\/a> is my previous post about my problem with tf.nn.max_pool_with_argmax)  <\/p>\n\n<p>So, does anyone any idea how I can register it in more simple way? or can anyone give me any idea about the concept of registering op in tensorflow? <\/p>\n\n<p>thanks,<\/p>\n\n<p>Ali<\/p>\n","answers":[]}
{"QuestionId":41710046,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><vectorization><svd>","CreationDate":"2017-01-18T01:39:03.670","AcceptedAnswerId":41711327.0,"OwnerUserId":3704652.0,"Title":"SVD++ vectorization with numpy or tensorflow","Body":"<p>I want to implement SVD++ with numpy or tensorflow. <br>\n( <a href=\"https:\/\/pdfs.semanticscholar.org\/8451\/c2812a1476d3e13f2a509139322cc0adb1a2.pdf\" rel=\"nofollow noreferrer\">https:\/\/pdfs.semanticscholar.org\/8451\/c2812a1476d3e13f2a509139322cc0adb1a2.pdf<\/a> ) <br>\n(4p equation 4)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Ki2cv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Ki2cv.png\" alt=\"enter image description here\"><\/a>\nI want to implement above equation without any for loop. <br>\nBut summation of y_j with index set R(u) makes it hard. <br><\/p>\n\n<p><strong>So my question is... <br> \nI want to implement below equation (q_v multiply sum of y_j) without any for loop <br>\n1. Is it possible to implement it with numpy without for loop?! <br>\n2. Is it possible to implement it with tensorflow without for loop?!<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lB3fX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lB3fX.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>My implementation is below... but I want to remove for loop further more<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41720059","CreationDate":"2017-01-18T12:57:41.743","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>Simply use two matrix-multiplications there with <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.dot.html\" rel=\"nofollow noreferrer\"><code>np.dot<\/code><\/a> for the final output -<\/p>\n\n<pre><code>r_hat = r.dot(y.T).dot(q)\n<\/code><\/pre>\n\n<p>Sample run to verify results -<\/p>\n\n<p>OP's sample setup :<\/p>\n\n<pre><code>In [68]: import numpy as np\n    ...: \n    ...: num_users = 3\n    ...: num_items = 5\n    ...: latent_dim = 2\n    ...: p = 0.1\n    ...: \n    ...: r = np.random.binomial(1, 1 - p,(num_users, num_items))\n    ...: r_hat = np.zeros([num_users,num_items])\n    ...: \n    ...: q = np.random.randn(latent_dim,num_items)\n    ...: y = np.random.randn(latent_dim,num_items)\n    ...: \n\nIn [69]: ## Second Try from OP\n    ...: for user in range(num_users):\n    ...:     for item in range(num_items):\n    ...:         q_j = q[:,item]\n    ...:         user_item_list = [i for i, e in enumerate(r[user,:]) if e != 0] # R_u\n    ...:         sum_y_j = np.sum(y[:,user_item_list],axis=1) # to make sum of y_i\n    ...:         r_hat[user,item] = np.dot(np.transpose(q_j),sum_y_j)\n    ...:         \n<\/code><\/pre>\n\n<p>Let's print out the result from OP's solution -<\/p>\n\n<pre><code>In [70]: r_hat\nOut[70]: \narray([[  4.06866107e+00,   2.91099460e+00,  -6.50447668e+00,\n          7.44275731e-03,  -2.14857566e+00],\n       [  4.06866107e+00,   2.91099460e+00,  -6.50447668e+00,\n          7.44275731e-03,  -2.14857566e+00],\n       [  5.57369599e+00,   3.76169533e+00,  -8.47503476e+00,\n          1.48615948e-01,  -2.82792374e+00]])\n<\/code><\/pre>\n\n<p>Now, I am using my proposed solution -<\/p>\n\n<pre><code>In [71]: r.dot(y.T).dot(q)\nOut[71]: \narray([[  4.06866107e+00,   2.91099460e+00,  -6.50447668e+00,\n          7.44275731e-03,  -2.14857566e+00],\n       [  4.06866107e+00,   2.91099460e+00,  -6.50447668e+00,\n          7.44275731e-03,  -2.14857566e+00],\n       [  5.57369599e+00,   3.76169533e+00,  -8.47503476e+00,\n          1.48615948e-01,  -2.82792374e+00]])\n<\/code><\/pre>\n\n<p>Value check seems successful!<\/p>\n"},{"AnswerId":"41711327","CreationDate":"2017-01-18T04:24:56.083","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Try this. <\/p>\n\n<pre><code>sum_y = []\nfor user in range(num_users):\n    mask = np.repeat(r[user,:][None,:],latent_dim, axis=0)\n    sum_y.append(np.sum(np.multiply(y, mask),axis=1))\nsum_y = np.asarray(sum_y)\n\nr_hat = (np.dot(q.T,sum_y.T)).T\nprint r_hat\n<\/code><\/pre>\n\n<p>It eliminates the enumerate loop, and also the dot product can be done in single go. I don't think it can be reduced beyond this.<\/p>\n"}]}
{"QuestionId":41710484,"AnswerCount":0,"Tags":"<neural-network><computer-vision><deep-learning><caffe><lstm>","CreationDate":"2017-01-18T02:36:03.850","AcceptedAnswerId":null,"OwnerUserId":7433337.0,"Title":"how to achieve Bi-LSTM in caffe?","Body":"<p>i am doing a project and want to achieve a bilstm in caffe,but i don't know how to do .Can it be modified directly in the prototxt.Any ideas appreciated.<\/p>\n","answers":[]}
{"QuestionId":41710646,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><syntaxnet>","CreationDate":"2017-01-18T02:59:13.890","AcceptedAnswerId":null,"OwnerUserId":2161903.0,"Title":"Command line SyntaxNet demo not working: TypeError: __init__() takes at most 2 arguments (3 given)","Body":"<p>I installed SyntaxNet on OS X El Capitan based on <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/syntaxnet#installation\" rel=\"nofollow noreferrer\">their documentation<\/a>. I am using Python 2.7. When I run the demo, I get error that says TypeError: takes two arguments but 3 given. Any help is appreciated. <\/p>\n\n<p>I am invoking the following command form SyntaxNet's working directory:<\/p>\n\n<p><\/p>\n\n<p>The detail of the error message is shown below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41893040","CreationDate":"2017-01-27T11:33:13.180","ParentId":null,"OwnerUserId":"7478105","Title":null,"Body":"<p>I met the similar error 'TypeError: <strong>init<\/strong>() takes at most 2 arguments (3 given)' when compiling on Ubuntu.<\/p>\n\n<p>Before running your command below\necho \"This is test.\"|syntaxnet\/demo.sh\nI would like to suggest you to run the following test command\nbazel test --linkopt=-headerpad_max_install_names \\\n    syntaxnet\/... util\/utf8\/...<\/p>\n"}]}
{"QuestionId":41711190,"AnswerCount":10,"Tags":"<python><tensorflow><deep-learning><keras>","CreationDate":"2017-01-18T04:07:16.950","AcceptedAnswerId":41712013.0,"OwnerUserId":5046896.0,"Title":"Keras, How to get the output of each layer?","Body":"<p>I have trained a binary classification model with CNN, and here is my code<\/p>\n\n<pre><\/pre>\n\n<p>And here, I wanna get the output of each layer just like TensorFlow, how can I do that?<\/p>\n","answers":[{"AnswerId":"58200323","CreationDate":"2019-10-02T11:11:47.900","ParentId":null,"OwnerUserId":"4157666","Title":null,"Body":"<p>In case you have one of the following cases:  <\/p>\n\n<ul>\n<li>error: <code>InvalidArgumentError: input_X:Y is both fed and fetched<\/code>  <\/li>\n<li>case of multiple inputs<\/li>\n<\/ul>\n\n<p>You need to do the following changes:<\/p>\n\n<ul>\n<li>add filter out for input layers in <code>outputs<\/code> variable<\/li>\n<li>minnor change on <code>functors<\/code> loop<\/li>\n<\/ul>\n\n<p>Minimum example:<\/p>\n\n<pre><code>from keras.engine.input_layer import InputLayer\ninp = model.input\noutputs = [layer.output for layer in model.layers if not isinstance(layer, InputLayer)]\nfunctors = [K.function(inp + [K.learning_phase()], [x]) for x in outputs]\nlayer_outputs = [fun([x1, x2, xn, 1]) for fun in functors]\n<\/code><\/pre>\n"},{"AnswerId":"43135938","CreationDate":"2017-03-31T08:24:19.963","ParentId":null,"OwnerUserId":"7428659","Title":null,"Body":"<p>I wrote this function for myself (in Jupyter) and it was inspired by <a href=\"https:\/\/stackoverflow.com\/users\/2217210\/indraforyou\">indraforyou<\/a>'s answer. It will plot all the layer outputs automatically. Your images must have a (x, y, 1) shape where 1 stands for 1 channel. You just call plot_layer_outputs(...) to plot.<\/p>\n\n<pre><code>%matplotlib inline\nimport matplotlib.pyplot as plt\nfrom keras import backend as K\n\ndef get_layer_outputs():\n    test_image = YOUR IMAGE GOES HERE!!!\n    outputs    = [layer.output for layer in model.layers]          # all layer outputs\n    comp_graph = [K.function([model.input]+ [K.learning_phase()], [output]) for output in outputs]  # evaluation functions\n\n    # Testing\n    layer_outputs_list = [op([test_image, 1.]) for op in comp_graph]\n    layer_outputs = []\n\n    for layer_output in layer_outputs_list:\n        print(layer_output[0][0].shape, end='\\n-------------------\\n')\n        layer_outputs.append(layer_output[0][0])\n\n    return layer_outputs\n\ndef plot_layer_outputs(layer_number):    \n    layer_outputs = get_layer_outputs()\n\n    x_max = layer_outputs[layer_number].shape[0]\n    y_max = layer_outputs[layer_number].shape[1]\n    n     = layer_outputs[layer_number].shape[2]\n\n    L = []\n    for i in range(n):\n        L.append(np.zeros((x_max, y_max)))\n\n    for i in range(n):\n        for x in range(x_max):\n            for y in range(y_max):\n                L[i][x][y] = layer_outputs[layer_number][x][y][i]\n\n\n    for img in L:\n        plt.figure()\n        plt.imshow(img, interpolation='nearest')\n<\/code><\/pre>\n"},{"AnswerId":"56368448","CreationDate":"2019-05-29T21:00:50.243","ParentId":null,"OwnerUserId":"1361125","Title":null,"Body":"<p>Assuming you have:<\/p>\n\n<p>1- Keras pre-trained <code>model<\/code>. <\/p>\n\n<p>2- Input <code>x<\/code> as image or set of images. The resolution of image should be compatible with dimension of the input layer. For example <strong>80*80*3<\/strong> for 3-channels (RGB) image. <\/p>\n\n<p>3- The name of the output <code>layer<\/code> to get the activation. For example, \"flatten_2\" layer. This should be include in the <code>layer_names<\/code> variable, represents name of layers of the given <code>model<\/code>.<\/p>\n\n<p>4- <code>batch_size<\/code> is an optional argument.<\/p>\n\n<p>Then you can easily use <code>get_activation<\/code> function to get the activation of the output <code>layer<\/code> for a given input <code>x<\/code> and pre-trained <code>model<\/code>:<\/p>\n\n<pre><code>import six\nimport numpy as np\nimport keras.backend as k\nfrom numpy import float32\ndef get_activations(x, model, layer, batch_size=128):\n\"\"\"\nReturn the output of the specified layer for input `x`. `layer` is specified by layer index (between 0 and\n`nb_layers - 1`) or by name. The number of layers can be determined by counting the results returned by\ncalling `layer_names`.\n:param x: Input for computing the activations.\n:type x: `np.ndarray`. Example: x.shape = (80, 80, 3)\n:param model: pre-trained Keras model. Including weights.\n:type model: keras.engine.sequential.Sequential. Example: model.input_shape = (None, 80, 80, 3)\n:param layer: Layer for computing the activations\n:type layer: `int` or `str`. Example: layer = 'flatten_2'\n:param batch_size: Size of batches.\n:type batch_size: `int`\n:return: The output of `layer`, where the first dimension is the batch size corresponding to `x`.\n:rtype: `np.ndarray`. Example: activations.shape = (1, 2000)\n\"\"\"\n\n    layer_names = [layer.name for layer in model.layers]\n    if isinstance(layer, six.string_types):\n        if layer not in layer_names:\n            raise ValueError('Layer name %s is not part of the graph.' % layer)\n        layer_name = layer\n    elif isinstance(layer, int):\n        if layer &lt; 0 or layer &gt;= len(layer_names):\n            raise ValueError('Layer index %d is outside of range (0 to %d included).'\n                             % (layer, len(layer_names) - 1))\n        layer_name = layer_names[layer]\n    else:\n        raise TypeError('Layer must be of type `str` or `int`.')\n\n    layer_output = model.get_layer(layer_name).output\n    layer_input = model.input\n    output_func = k.function([layer_input], [layer_output])\n\n    # Apply preprocessing\n    if x.shape == k.int_shape(model.input)[1:]:\n        x_preproc = np.expand_dims(x, 0)\n    else:\n        x_preproc = x\n    assert len(x_preproc.shape) == 4\n\n    # Determine shape of expected output and prepare array\n    output_shape = output_func([x_preproc[0][None, ...]])[0].shape\n    activations = np.zeros((x_preproc.shape[0],) + output_shape[1:], dtype=float32)\n\n    # Get activations with batching\n    for batch_index in range(int(np.ceil(x_preproc.shape[0] \/ float(batch_size)))):\n        begin, end = batch_index * batch_size, min((batch_index + 1) * batch_size, x_preproc.shape[0])\n        activations[begin:end] = output_func([x_preproc[begin:end]])[0]\n\n    return activations\n<\/code><\/pre>\n"},{"AnswerId":"56387938","CreationDate":"2019-05-31T02:28:19.957","ParentId":null,"OwnerUserId":"4170095","Title":null,"Body":"<p>Based on all the good answers of this thread, I wrote a library to fetch the output of each layer. It abstracts all the complexity and has been designed to be as user-friendly as possible:<\/p>\n\n<p><a href=\"https:\/\/github.com\/philipperemy\/keract\" rel=\"noreferrer\">https:\/\/github.com\/philipperemy\/keract<\/a><\/p>\n\n<p>It handles almost all the edge cases<\/p>\n\n<p>Hope it helps!<\/p>\n"},{"AnswerId":"43860397","CreationDate":"2017-05-09T03:07:29.790","ParentId":null,"OwnerUserId":"2097240","Title":null,"Body":"<p>Well, other answers are very complete, but there is a very basic way to \"see\", not to \"get\" the shapes.<\/p>\n\n<p>Just do a <code>model.summary()<\/code>. It will print all layers and their output shapes. \"None\" values will indicate variable dimensions, and the first dimension will be the batch size. <\/p>\n"},{"AnswerId":"55417982","CreationDate":"2019-03-29T13:00:08.923","ParentId":null,"OwnerUserId":"5838189","Title":null,"Body":"<p>Wanted to add this as a comment (but don't have high enough rep.) to @indraforyou's answer to correct for the issue mentioned in @mathtick's comment. To avoid the <code>InvalidArgumentError: input_X:Y is both fed and fetched.<\/code> exception, simply replace the line <code>outputs = [layer.output for layer in model.layers]<\/code> with <code>outputs = [layer.output for layer in model.layers][1:]<\/code>, i.e.<\/p>\n\n<p>adapting indraforyou's minimal working example:<\/p>\n\n<pre><code>from keras import backend as K \ninp = model.input                                           # input placeholder\noutputs = [layer.output for layer in model.layers][1:]        # all layer outputs except first (input) layer\nfunctor = K.function([inp, K.learning_phase()], outputs )   # evaluation function\n\n# Testing\ntest = np.random.random(input_shape)[np.newaxis,...]\nlayer_outs = functor([test, 1.])\nprint layer_outs\n<\/code><\/pre>\n\n<p>p.s. my attempts trying things such as <code>outputs = [layer.output for layer in model.layers[1:]]<\/code> did not work. <\/p>\n"},{"AnswerId":"45430345","CreationDate":"2017-08-01T06:24:09.607","ParentId":null,"OwnerUserId":"7657597","Title":null,"Body":"<p>Following looks very simple to me:<\/p>\n\n<pre><code>model.layers[idx].output\n<\/code><\/pre>\n\n<p>Above is a tensor object, so you can modify it using operations that can be applied to a tensor object. <\/p>\n\n<p>For example, to get the shape <code>model.layers[idx].output.get_shape()<\/code><\/p>\n\n<p><code>idx<\/code> is the index of the layer and you can find it from <code>model.summary()<\/code><\/p>\n"},{"AnswerId":"47253167","CreationDate":"2017-11-12T19:45:14.907","ParentId":null,"OwnerUserId":"524444","Title":null,"Body":"<p>From: <a href=\"https:\/\/github.com\/philipperemy\/keras-visualize-activations\/blob\/master\/read_activations.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/philipperemy\/keras-visualize-activations\/blob\/master\/read_activations.py<\/a><\/p>\n\n<pre><code>import keras.backend as K\n\ndef get_activations(model, model_inputs, print_shape_only=False, layer_name=None):\n    print('----- activations -----')\n    activations = []\n    inp = model.input\n\n    model_multi_inputs_cond = True\n    if not isinstance(inp, list):\n        # only one input! let's wrap it in a list.\n        inp = [inp]\n        model_multi_inputs_cond = False\n\n    outputs = [layer.output for layer in model.layers if\n               layer.name == layer_name or layer_name is None]  # all layer outputs\n\n    funcs = [K.function(inp + [K.learning_phase()], [out]) for out in outputs]  # evaluation functions\n\n    if model_multi_inputs_cond:\n        list_inputs = []\n        list_inputs.extend(model_inputs)\n        list_inputs.append(0.)\n    else:\n        list_inputs = [model_inputs, 0.]\n\n    # Learning phase. 0 = Test mode (no dropout or batch normalization)\n    # layer_outputs = [func([model_inputs, 0.])[0] for func in funcs]\n    layer_outputs = [func(list_inputs)[0] for func in funcs]\n    for layer_activations in layer_outputs:\n        activations.append(layer_activations)\n        if print_shape_only:\n            print(layer_activations.shape)\n        else:\n            print(layer_activations)\n    return activations\n<\/code><\/pre>\n"},{"AnswerId":"46359250","CreationDate":"2017-09-22T07:35:02.683","ParentId":null,"OwnerUserId":"470184","Title":null,"Body":"<p>From <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer<\/a><\/p>\n\n<p>One simple way is to create a new Model that will output the layers that you are interested in:<\/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>Alternatively, you can build a Keras function that will return the output of a certain layer given a certain input, for example:<\/p>\n\n<pre><code>from keras import backend as K\n\n# with a Sequential model\nget_3rd_layer_output = K.function([model.layers[0].input],\n                                  [model.layers[3].output])\nlayer_output = get_3rd_layer_output([x])[0]\n<\/code><\/pre>\n"},{"AnswerId":"41712013","CreationDate":"2017-01-18T05:28:56.300","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can easily get the outputs of any layer by using: <code>model.layers[index].output<\/code><\/p>\n\n<p>For all layers use this:<\/p>\n\n<pre><code>from keras import backend as K\n\ninp = model.input                                           # input placeholder\noutputs = [layer.output for layer in model.layers]          # all layer outputs\nfunctors = [K.function([inp, K.learning_phase()], [out]) for out in outputs]    # evaluation functions\n\n# Testing\ntest = np.random.random(input_shape)[np.newaxis,...]\nlayer_outs = [func([test, 1.]) for func in functors]\nprint layer_outs\n<\/code><\/pre>\n\n<p>Note: To simulate Dropout use <code>learning_phase<\/code> as <code>1.<\/code> in <code>layer_outs<\/code> otherwise use <code>0.<\/code><\/p>\n\n<p><strong>Edit:<\/strong> (based on comments)<\/p>\n\n<p><code>K.function<\/code> creates theano\/tensorflow tensor functions which is later used to get the output from the symbolic graph given the input. <\/p>\n\n<p>Now <code>K.learning_phase()<\/code> is required as an input as many Keras layers like Dropout\/Batchnomalization depend on it to change behavior during training and test time. <\/p>\n\n<p>So if you remove the dropout layer in your code you can simply use:<\/p>\n\n<pre><code>from keras import backend as K\n\ninp = model.input                                           # input placeholder\noutputs = [layer.output for layer in model.layers]          # all layer outputs\nfunctors = [K.function([inp], [out]) for out in outputs]    # evaluation functions\n\n# Testing\ntest = np.random.random(input_shape)[np.newaxis,...]\nlayer_outs = [func([test]) for func in functors]\nprint layer_outs\n<\/code><\/pre>\n\n<p><strong>Edit 2: More optimized<\/strong><\/p>\n\n<p>I just realized that the previous answer is not that optimized as for each function evaluation the data will be transferred CPU->GPU memory and also the tensor calculations needs to be done for the lower layers over-n-over. <\/p>\n\n<p>Instead this is a much better way as you don't need multiple functions but a single function giving you the list of all outputs:<\/p>\n\n<pre><code>from keras import backend as K\n\ninp = model.input                                           # input placeholder\noutputs = [layer.output for layer in model.layers]          # all layer outputs\nfunctor = K.function([inp, K.learning_phase()], outputs )   # evaluation function\n\n# Testing\ntest = np.random.random(input_shape)[np.newaxis,...]\nlayer_outs = functor([test, 1.])\nprint layer_outs\n<\/code><\/pre>\n"}]}
{"QuestionId":41711973,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-01-18T05:25:06.320","AcceptedAnswerId":null,"OwnerUserId":2118051.0,"Title":"accessing the values from a Tensor Variable of Tensorflow","Body":"<p>I have modified the sample code of autoencoder to suit my data. I want to know a way in which i can access the layer_1 values of the encoder. \nI want to use the encoder encoded values as the features for the input data and do further analysis on that.\nThe input is of 200 dimensions and the encoder has 100 hidden nodes.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43854362","CreationDate":"2017-05-08T17:59:21.977","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>In order to peer into what's going on in your TensorFlow computation, I recommend using <a href=\"https:\/\/www.tensorflow.org\/get_started\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">TensorBoard<\/a>. TensorBoard includes <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/summary\" rel=\"nofollow noreferrer\">Summary Operations<\/a> that can then be plotted in the UI to visualize the values of Tensors as the model trains.<\/p>\n\n<p>For a high-level introduction, I highly recommend this YouTube video from the TensorFlow Dev Summit: <a href=\"https:\/\/www.youtube.com\/watch?v=eBbEDRsCmv4\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=eBbEDRsCmv4<\/a> <\/p>\n"}]}
{"QuestionId":41712618,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-18T06:19:12.380","AcceptedAnswerId":null,"OwnerUserId":6334174.0,"Title":"tensorflow while_loop: ValueError: None values not supported","Body":"<p>I'm trying to use slice a tensor object in my graph. Code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>where self.batchSize is a  of . Last_outputs is a 2-d tensor. However, when I run the code, tensorflow gives an error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"48787034","CreationDate":"2018-02-14T12:16:36.600","ParentId":null,"OwnerUserId":"3235501","Title":null,"Body":"<p>The function passed to the <code>body<\/code> parameter should return a tensor value, not <code>None<\/code>. Try changing the function to the following.<\/p>\n\n<pre><code>def add_inputs(idx):\n    utt_input = tf.slice(last_outputs, begin=[idx, 0],\n    size=[self.args.uttWindowSize, self.args.wordUnits])\n    utt_inputs.append(utt_input)\n    return tf.constant(0 , dtype=tf.int32 , name=\"ret_val\") #Dummy\n<\/code><\/pre>\n\n<p>From the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"nofollow noreferrer\">docs<\/a>, <\/p>\n\n<blockquote>\n  <p>body is a callable returning a (possibly nested) tuple, namedtuple or list of tensors of the same arity (length and structure) and types as loop_vars.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41713266,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-18T07:07:03.267","AcceptedAnswerId":null,"OwnerUserId":4447972.0,"Title":"What is the use of a \"reuse\" parameter of tf.contrib.layers functions?","Body":"<p>While using some  functions, I found that the 'reuse' parameter of the functions does not work as I thought.<\/p>\n\n<p>I thought that it would act similarly to the \"reuse\" parameter of .\nThus, I thought that it would raise an error if I try to use the function with  for the first time, or if I try to use it with  for the case it is already used before.<\/p>\n\n<p>However, it seems not to emit any exception in the below cases:<\/p>\n\n<pre><\/pre>\n\n<p>The same phenomena also occur at other functions that use  class internally.\nIt seems that it ignores the  parameter and always shares parameters if the name of a variable is same. Is this intended or a bug?<\/p>\n","answers":[{"AnswerId":"47367970","CreationDate":"2017-11-18T15:51:48.753","ParentId":null,"OwnerUserId":"5545260","Title":null,"Body":"<p>This problem was most likely due to a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8626\" rel=\"nofollow noreferrer\">bug that was fixed in early 2017<\/a>.  Updating to 1.1 or later should fix this problem.<\/p>\n"}]}
{"QuestionId":41713516,"AnswerCount":2,"Tags":"<java><python><tensorflow><deep-learning><image-recognition>","CreationDate":"2017-01-18T07:22:09.880","AcceptedAnswerId":null,"OwnerUserId":6947773.0,"Title":"Running a tensorflow model written in python for training and prediction from java","Body":"<p>I have retrained inception model for my own data set. Tho model is built in python and i now have the saved graph as .pb file and label file as .txt. Now i need to predict using this model for an image through java. Can anyone please help me<\/p>\n","answers":[{"AnswerId":"41716036","CreationDate":"2017-01-18T09:46:38.237","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The TensorFlow team is developing a Java interface, but it is not stable yet. You can find the existing code here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/java\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/java<\/a> and follow updates on its development here <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5<\/a>. You can take a look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/src\/test\/java\/org\/tensorflow\/GraphTest.java\" rel=\"nofollow noreferrer\">GraphTest.java<\/a>, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/src\/test\/java\/org\/tensorflow\/SessionTest.java\" rel=\"nofollow noreferrer\">SessionTest.java<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/src\/test\/java\/org\/tensorflow\/TensorTest.java\" rel=\"nofollow noreferrer\">TensorTest.java<\/a> to see how it is currently used (although, as explained, this may change in the future). Basically, you need to load the binary saved graph into a <code>Graph<\/code> object, create a <code>Session<\/code> with it and run it with the appropriate values (as <code>Tensor<\/code>s) to receive a <code>List&lt;Tensor&gt;<\/code> with the output. Put together from the examples in the source:<\/p>\n\n<pre><code>import java.nio.file.Files;\nimport java.nio.file.Paths;\nimport org.tensorflow.Graph;\nimport org.tensorflow.Session;\nimport org.tensorflow.Tensor;\n\ntry (Graph graph = new Graph()) {\n    graph.importGraphDef(Files.readAllBytes(Paths.get(\"saved_model.pb\"));\n    try (Session sess = new Session(graph)) {\n        try (Tensor x = Tensor.create(1.0f);\n             Tensor y = s.runner().feed(\"x\", x).fetch(\"y\").run().get(0)) {\n            System.out.println(y.floatValue());\n        }\n    }\n}\n<\/code><\/pre>\n"},{"AnswerId":"44791219","CreationDate":"2017-06-27T22:59:18.223","ParentId":null,"OwnerUserId":"3795036","Title":null,"Body":"<p>The code I used that worked read a <code>protobuf<\/code> file, ending with <code>.pb<\/code>.<\/p>\n\n<pre><code>try (SavedModelBundle b = SavedModelBundle.load(\"\/tmp\/model\", \"serve\")) {\n    Session sess = b.session();\n    ...\n    float[][]matrix = sess.runner()\n        .feed(\"x\", input)\n        .feed(\"keep_prob\", keep_prob)\n        .fetch(\"y_conv\")\n        .run()\n        .get(0)\n        .copyTo(new float[1][10]);\n    ...\n}\n<\/code><\/pre>\n\n<p>The python code I used to save it was:<\/p>\n\n<pre><code>  signature = tf.saved_model.signature_def_utils.build_signature_def(\n    inputs = {'x': tf.saved_model.utils.build_tensor_info(x)},\n    outputs = {'y_conv': tf.saved_model.utils.build_tensor_info(y_conv)},\n  )\n  builder = tf.saved_model.builder.SavedModelBuilder(\"\/tmp\/model\" )\n  builder.add_meta_graph_and_variables(sess, \n       [tf.saved_model.tag_constants.SERVING],\n       signature_def_map={tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature}\n   )\n  builder.save()\n<\/code><\/pre>\n"}]}
{"QuestionId":41714318,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-01-18T08:16:35.260","AcceptedAnswerId":42683604.0,"OwnerUserId":2324298.0,"Title":"Input parameters of tf.contrib.learn.read_batch_features","Body":"<p>I am working through these tensorflow <a href=\"https:\/\/github.com\/dennybritz\/chatbot-retrieval\/tree\/master\/models%20&#39;codes&#39;\" rel=\"nofollow noreferrer\">codes<\/a> which implement a LSTM in tensorflow. While going through the codes, I came across this function (in <a href=\"https:\/\/github.com\/dennybritz\/chatbot-retrieval\/blob\/master\/udc_inputs.py\" rel=\"nofollow noreferrer\">input_fn code<\/a> - line 38) . I looked up the documentation of  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.learn\/input_processing#read_batch_features\" rel=\"nofollow noreferrer\">here<\/a>. This is what I got - <\/p>\n\n<pre>\nfile_pattern: List of files or pattern of file paths containing Example records. \nbatch_size: An int or scalar Tensor specifying the batch size to use.\nfeatures: A dict mapping feature keys to FixedLenFeature or VarLenFeature values.\nrandomize_input: Whether the input should be randomized.\nnum_epochs: Integer specifying the number of times to read through the dataset. If None, cycles through the dataset forever. NOTE - If specified, creates a variable that must be initialized, so call tf.local_variables_initializer() as shown in the tests.\nqueue_capacity: Capacity for input queue.\nreader_num_threads: The number of threads to read examples.\nname: Name of resulting op.\n<\/pre>\n\n<p>There are few input parameters that I am not able to understand and was hoping someone could help me with it.<\/p>\n\n<ol>\n<li><p>The  parameter. Does it mean it will shuffle the entire dataset?<\/p><\/li>\n<li><p>For , if I specify  does it mean that my  will keep feeding to the . In that case the training wouldn't stop. This doesn't make sense to me. I guess I'm going wrong somewhere here.<\/p><\/li>\n<li><p> I am not sure what this means<\/p><\/li>\n<\/ol>\n\n<p>Would appreciate any help around these questions. Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42683604","CreationDate":"2017-03-08T23:02:54.093","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p><code>randomize_input<\/code> means both that the filenames are randomized and that a shuffling queue is used to shuffle examples once in memory.<\/p>\n\n<p>For <code>num_epochs<\/code>, training can stop due to other reasons (often training stops at a given <code>global_step<\/code> regardless of the input function).<\/p>\n\n<p>For <code>queue_capacity<\/code>, TensorFlow uses queues internally to buffer I\/O for maximum throughput. This parameter allows you to control the size of that queue.<\/p>\n"}]}
{"QuestionId":41714533,"AnswerCount":1,"Tags":"<c++><multithreading><session><tensorflow>","CreationDate":"2017-01-18T08:29:59.523","AcceptedAnswerId":null,"OwnerUserId":7341429.0,"Title":"C++ Tensorflow, how to make session->Run() with multithread, or spend less time","Body":"<p>I run all the below on CPU.\nI run the sample ensorflow\/examples\/label_image spends 7~8 seconds. As I know, python spends about 0.5 second to process the same sample and this is because \"The TensorFlow Session object is multithreaded, so multiple threads can easily use the same session and run ops in parallel. \".\nBUT, how can I set multithread on Session with C++.<\/p>\n\n<p>What I try ...\nI hard code line 81 in \"tensorflow\/tensorflow\/core\/common_runtime\/direct_session.cc\":\n\"const int32 num_threads = 16;\"\nHowever, it does not work.<\/p>\n\n<p>How can I set some configure or what should I do something?<\/p>\n","answers":[{"AnswerId":"41732030","CreationDate":"2017-01-19T00:55:47.183","ParentId":null,"OwnerUserId":"7437816","Title":null,"Body":"<p>How about usiung multiple threads?<\/p>\n\n<pre><code>std::vector&lt;std::thread&gt; threads;\n\nfor (std::size_t i = 0; i &lt; 10; ++i) {\n   threads.push_back(std::thread([&amp;]{ session-&gt;Run(); })); \n}\n\n\nfor (std::size_t i = 0; i &lt; 10; ++i) {\n   threads[i].join();\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":41714801,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-18T08:45:42.317","AcceptedAnswerId":41717150.0,"OwnerUserId":3952274.0,"Title":"Tensorflow: _variable_with_weight_decay(...) explanation","Body":"<p>at the moment I'm looking at the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">cifar10 example<\/a> and I noticed the function _variable_with_weight_decay(...) in the file <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10.py#L115\" rel=\"nofollow noreferrer\">cifar10.py<\/a>. The code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I'm wondering if this function does what it says. It is clear that when a weight decay factor is given (wd not None) the deacy value (weight_decay) is computed. But is it every applied? At the end the unmodified variable (var) is return, or am I missing something?<\/p>\n\n<p>Second question would be how to fix this? As I understand the value of the scalar weight_decay must be subtracted from each element in the weight matrix, but I'm unable to find a tensorflow op that can do that (adding\/subtracting a single value from every element of a tensor). Is there any op like this?\nAs a workaround I thought it might be possible to create a new tensor initialized with the value of weight_decay and use tf.subtract(...) to achieve the same result. Or is this the right way to go anyway?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"41717150","CreationDate":"2017-01-18T10:38:52.160","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>The code does what it says. You are supposed to sum everything in the <code>'losses'<\/code> collection (which the weight decay term is added to in the second to last line) for the loss that you pass to the optimizer. In the <code>loss()<\/code> function in that example:<\/p>\n\n<pre><code>tf.add_to_collection('losses', cross_entropy_mean)\n[...]\nreturn tf.add_n(tf.get_collection('losses'), name='total_loss')\n<\/code><\/pre>\n\n<p>so what the <code>loss()<\/code> function returns is the classification loss plus everything that was in the <code>'losses'<\/code> collection before.<\/p>\n\n<p>As a side note, weight decay does not mean you subtract the value of <code>wd<\/code> from every value in the tensor as part of the update step, it multiplies the value by <code>(1-learning_rate*wd)<\/code> (in plain SGD). To see why this is so, recall that l2_loss computes<\/p>\n\n<pre><code>output = sum(t_i ** 2) \/ 2\n<\/code><\/pre>\n\n<p>with <code>t_i<\/code> being the elements of the tensor. This means that the derivative of <code>l2_loss<\/code> with regard to each tensor element is the value of that tensor element itself, and since you scaled <code>l2_loss<\/code> with <code>wd<\/code> the derivative is scaled as well.<\/p>\n\n<p>Since the update step (again, in plain SGD) is (forgive me for omitting the time step indexes)<\/p>\n\n<pre><code>w := w - learning_rate * dL\/dw\n<\/code><\/pre>\n\n<p>you get, if you only had the weight decay term <\/p>\n\n<pre><code>w := w - learning_rate * wd * w\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>w := w * (1 - learning_rate * wd)\n<\/code><\/pre>\n"}]}
{"QuestionId":41714947,"AnswerCount":1,"Tags":"<python-2.7><scikit-learn><keras>","CreationDate":"2017-01-18T08:54:24.043","AcceptedAnswerId":41740409.0,"OwnerUserId":6313251.0,"Title":"How to do bagging using scikit BaggingClassifier with keras Convolutional Neural Network as base estimator via keras-scikit wrapper?","Body":"<p>I'm trying to make an ensemble learning, which is bagging using scikit-learn BaggingClassifier with 2D Convolutional Neural Networks (CNN) as the base estimators.<\/p>\n\n<p>Before this, i've tried bagging with scikit's Neural Network to test scikit's BaggingClassifier and it worked. Also i've tested scikit's GridSearchCV with keras-wrapper to search for 2D CNN's hyperparameters, it also worked. <\/p>\n\n<p>Just now, when i tried using scikit's BaggingClassifier with keras-wrapper to wrap then create ensemble learning with the 2D CNN model as base estimator, i got an error.<\/p>\n\n<p>Here is the code snippet :<\/p>\n\n<pre><\/pre>\n\n<p>Here is the create_cnn_model_A function looks like :<\/p>\n\n<pre><\/pre>\n\n<p>Here is the shape of the self.patternTraining &amp; self.targetTraining before reshaping : <\/p>\n\n<pre><\/pre>\n\n<p>And this is the error i got : <\/p>\n\n<pre><\/pre>\n\n<p>I thought this is some kind of array-shape-error but i don't know how to resolve this. Or maybe it is not yet possible to use scikit's BaggingClassifier with keras' 2D CNN via keras-wrapper? <\/p>\n\n<p>If more details are needed, i'm ready to provide.\nAny help is appreciated, thanks.<\/p>\n","answers":[{"AnswerId":"41740409","CreationDate":"2017-01-19T11:13:38.987","ParentId":null,"OwnerUserId":"6507214","Title":null,"Body":"<p>This is not supported currently in sklearn using Keras. You have to implement yourself or I raised a same issue in the community months ago. I got an apt response and currently, they are trying to implement it. Wait for the next version or check the issue to know more.<\/p>\n"}]}
{"QuestionId":41715025,"AnswerCount":4,"Tags":"<python><machine-learning><neural-network><keras>","CreationDate":"2017-01-18T08:58:59.923","AcceptedAnswerId":null,"OwnerUserId":2670775.0,"Title":"Keras flowFromDirectory get file names as they are being generated","Body":"<p>Is it possible to get the file names that were loaded using  ? \nI have :<\/p>\n\n<pre><\/pre>\n\n<p>I have a custom generator for my multi output model like:<\/p>\n\n<pre><\/pre>\n\n<p>Node that at the moment I am generating random numbers for  but for real training , I wish to load up a  file that contains the bounding box coordinates for my images. For that I will need to get the file names that were generated using  method. After I have that , I can load the file, parse the  and pass it instead of . It is also necessary that the ordering of the  variable and the list of the file names that I get is the same. <\/p>\n","answers":[{"AnswerId":"55386448","CreationDate":"2019-03-27T21:05:34.627","ParentId":null,"OwnerUserId":"7442271","Title":null,"Body":"<p>Here is an example that works with  <code>shuffle=True<\/code> as well. And also properly handles last batch. To make one pass:<\/p>\n\n<pre><code>datagen = ImageDataGenerator().flow_from_directory(...)    \nbatches_per_epoch = datagen.samples \/\/ datagen.batch_size + (datagen.samples % datagen.batch_size &gt; 0)\nfor i in range(batches_per_epoch):\n    batch = next(datagen)\n    current_index = ((datagen.batch_index-1) * datagen.batch_size)\n    if current_index &lt; 0:\n        if datagen.samples % datagen.batch_size &gt; 0:\n            current_index = max(0,datagen.samples - datagen.samples % datagen.batch_size)\n        else:\n            current_index = max(0,datagen.samples - datagen.batch_size)\n    index_array = datagen.index_array[current_index:current_index + datagen.batch_size].tolist()\n    img_paths = [datagen.filepaths[idx] for idx in index_array]\n    #batch[0] - x, batch[1] - y, img_paths - absolute path\n<\/code><\/pre>\n"},{"AnswerId":"54918559","CreationDate":"2019-02-28T04:46:28.597","ParentId":null,"OwnerUserId":"5058588","Title":null,"Body":"<p>You can make a pretty minimal subclass that returns the <code>image, file_path<\/code> tuple by inheriting the <a href=\"https:\/\/github.com\/keras-team\/keras-preprocessing\/blob\/master\/keras_preprocessing\/image\/directory_iterator.py\" rel=\"nofollow noreferrer\"><code>DirectoryIterator<\/code><\/a>:<\/p>\n\n<pre><code>import numpy as np\nfrom keras.preprocessing.image import ImageDataGenerator, DirectoryIterator\n\nclass ImageWithNames(DirectoryIterator):\n    def __init__(self, *args, **kwargs):\n        super().__init__(*args, **kwargs)\n        self.filenames_np = np.array(self.filepaths)\n        self.class_mode = None # so that we only get the images back\n\n    def _get_batches_of_transformed_samples(self, index_array):\n        return (super()._get_batches_of_transformed_samples(index_array),\n                self.filenames_np[index_array])\n<\/code><\/pre>\n\n<p>In the init, I added a attribute that is the numpy version of <code>self.filepaths<\/code> so that we can easily index into that array to get the paths on each batch generation.  <\/p>\n\n<p>The only other change to the base class is to return a tuple that is the image batch <code>super()._get_batches_of_transformed_samples(index_array)<\/code> and the file paths <code>self.filenames_np[index_array]<\/code>.<\/p>\n\n<p>With that, you can make your generator like so:<\/p>\n\n<pre><code>imagegen = ImageDataGenerator()\ndatagen = ImageWithNames('\/data\/path', imagegen, target_size=(224,224))\n<\/code><\/pre>\n\n<p>And then check with<\/p>\n\n<pre><code>next(datagen)\n<\/code><\/pre>\n"},{"AnswerId":"57505794","CreationDate":"2019-08-15T06:48:27.590","ParentId":null,"OwnerUserId":"4794555","Title":null,"Body":"<p>at least with version 2.2.4,you can do it like this<\/p>\n\n<pre><code>datagen = ImageDataGenerator()\ngen = datagen.flow_from_directory(...)\nfor file in gen.filenames:\n    print(file)\n<\/code><\/pre>\n\n<p>or get the file path<\/p>\n\n<pre><code>for filepath in gen.filepaths:\n    print(filepath)\n<\/code><\/pre>\n"},{"AnswerId":"44063083","CreationDate":"2017-05-19T06:39:02.903","ParentId":null,"OwnerUserId":"6173226","Title":null,"Body":"<p>Yes is it possible, at least with version 2.0.4 (don't know about earlier version).<\/p>\n\n<p>The instance of <code>ImageDataGenerator().flow_from_directory(...)<\/code> has an attribute with <code>filenames<\/code> which is a list of all the files in the order the generator yields them and also an attribute <code>batch_index<\/code>. So you can do it like this:<\/p>\n\n<pre><code>datagen = ImageDataGenerator()\ngen = datagen.flow_from_directory(...)\n<\/code><\/pre>\n\n<p>And every iteration on generator you can get the corresponding filenames like this:<\/p>\n\n<pre><code>for i in gen:\n    idx = (gen.batch_index - 1) * gen.batch_size\n    print(gen.filenames[idx : idx + gen.batch_size])\n<\/code><\/pre>\n\n<p>This will give you the filenames of the images in the current batch.<\/p>\n"}]}
{"QuestionId":41715511,"AnswerCount":2,"Tags":"<arrays><tensorflow><slice>","CreationDate":"2017-01-18T09:23:04.553","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Slicing a tensor by using indices in Tensorflow","Body":"<p>Basically I have a 2d array and I want to do this nice numpy-like thing<\/p>\n\n<pre><\/pre>\n\n<p>in Tensorflow. Here rows and cols are just two integers.<\/p>\n","answers":[{"AnswerId":"42823194","CreationDate":"2017-03-16T00:24:40.957","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Indeed, TensorFlow now has better support for slicing, so you can use the exact same syntax as NumPy:<\/p>\n\n<pre><code>result = noise_spec[:rows, :cols]\n<\/code><\/pre>\n"},{"AnswerId":"41716085","CreationDate":"2017-01-18T09:49:14.300","ParentId":null,"OwnerUserId":"5016028","Title":null,"Body":"<p>found out, it's <\/p>\n\n<pre><code>tf.slice(noise_spec, [0,0],[rows, cols])\n<\/code><\/pre>\n"}]}
{"QuestionId":41716001,"AnswerCount":0,"Tags":"<python><numpy><keras><keras-layer>","CreationDate":"2017-01-18T09:45:08.563","AcceptedAnswerId":null,"OwnerUserId":996366.0,"Title":"Indexerror: out of bounds for size error in keras convnet1D","Body":"<p>I am trying to use CNN 1D for time series prediction. I have a time series data set with 30 features, 3 targets and more than 3000 rows.<\/p>\n\n<p>This is my keras model<\/p>\n\n<pre><\/pre>\n\n<p>The model compiles without any error but when I tried to do model fit it gave this error<\/p>\n\n<pre><\/pre>\n\n<p>I saw this <a href=\"https:\/\/stackoverflow.com\/a\/36798335\/996366\">answer<\/a> but my  (feature) and  (target) are already in numpy array form. How to solve this?<\/p>\n\n<p><strong>EDITED<\/strong><\/p>\n\n<p>After some tinkering I have found out the problem is caused by my CNN model itself. I tried to train the same dataset with a simple neural network and it ran with out any issue.<\/p>\n\n<pre><\/pre>\n\n<p>Any one know what is wrong with my CNN model?<\/p>\n","answers":[]}
{"QuestionId":41716090,"AnswerCount":0,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-01-18T09:49:25.450","AcceptedAnswerId":null,"OwnerUserId":1093494.0,"Title":"Tensorflow: different fetches bring different results","Body":"<p>I am playing around with  and there is one thing I can't seem to understand: the role of the  argument for the .\nFrom the <a href=\"http:\/\/classroom.udacity.com\/courses\/ud730\/lessons\/6370362152\/concepts\/71191606550923#\" rel=\"nofollow noreferrer\">doc<\/a>:<\/p>\n\n<blockquote>\n  <p>To fetch the outputs of operations, execute the graph with a run()\n  call on the Session object and pass in the tensors to retrieve.<\/p>\n<\/blockquote>\n\n<p>What I understood is that the parameter is used to read certain values from the from the run step. I see, though, that different  bring to different results<\/p>\n\n<pre><\/pre>\n\n<p>So now I wonder what else the  parameter is used for, given that it seems to influence the way the training is run. I didn't really understand it from the documentation...<\/p>\n","answers":[]}
{"QuestionId":41716380,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-01-18T10:02:24.633","AcceptedAnswerId":41729210.0,"OwnerUserId":5046896.0,"Title":"Keras, output of model predict_proba","Body":"<p>In the <a href=\"https:\/\/keras.io\/models\/sequential\/#sequential-model-methods\" rel=\"nofollow noreferrer\">docs<\/a>, the  is <\/p>\n\n<blockquote>\n  <p>Generates class probability predictions for the input samples batch by batch.<\/p>\n<\/blockquote>\n\n<p>and returns <\/p>\n\n<blockquote>\n  <p>A Numpy array of probability predictions.<\/p>\n<\/blockquote>\n\n<p>Suppose my model is binary classification model, does the output is , for  is probability of , and  is the probability of ?<\/p>\n","answers":[{"AnswerId":"41717416","CreationDate":"2017-01-18T10:50:33.147","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>It depends on how you specify output of your model and your targets. It can be both. Usually when one is doing binary classification the output is a single value which is a probability of the positive prediction. One minus the output is probability of the negative prediction.<\/p>\n"},{"AnswerId":"41729210","CreationDate":"2017-01-18T20:51:33.840","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Here the situation is different and somehow misleading, especially when you are comparing <code>predict_proba<\/code> method to <code>sklearn<\/code> methods with the same name. In Keras (not sklearn wrappers) a method <code>predict_proba<\/code> is <strong>exactly the same<\/strong> as a <code>predict<\/code> method. You can even check it <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/models.py\" rel=\"noreferrer\">here<\/a>:<\/p>\n\n<pre><code>def predict_proba(self, x, batch_size=32, verbose=1):\n        \"\"\"Generates class probability predictions for the input samples\n        batch by batch.\n        # Arguments\n            x: input data, as a Numpy array or list of Numpy arrays\n                (if the model has multiple inputs).\n            batch_size: integer.\n            verbose: verbosity mode, 0 or 1.\n        # Returns\n            A Numpy array of probability predictions.\n        \"\"\"\n        preds = self.predict(x, batch_size, verbose)\n        if preds.min() &lt; 0. or preds.max() &gt; 1.:\n            warnings.warn('Network returning invalid probability values. '\n                          'The last layer might not normalize predictions '\n                          'into probabilities '\n                          '(like softmax or sigmoid would).')\n        return preds\n<\/code><\/pre>\n\n<p>So - in a binary classification case - the output which you get depends on the design of your network:<\/p>\n\n<ul>\n<li>if the final output of your network is obtained by a single sigmoid output - then the output of <code>predict_proba<\/code> is simply a probability assigned to class 1.<\/li>\n<li>if the final output of your network is obtained by a two dimensional output to which you are applying a <code>softmax<\/code> function - then the output of <code>predict_proba<\/code> is a pair where <code>[a, b]<\/code> where <code>a = P(class(x) = 0)<\/code> and <code>b = P(class(x) = 1)<\/code>.<\/li>\n<\/ul>\n\n<p>This second method is rarely used and there are some theorethical advantages of using the first method - but I wanted to inform you - just in case.<\/p>\n"}]}
{"QuestionId":41717502,"AnswerCount":1,"Tags":"<arrays><tensorflow><slice>","CreationDate":"2017-01-18T10:54:52.650","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Expanding width of a 2D tensor in Tensorflow, by repeating its columns","Body":"<p>I have a 2d tensor which must be expanded in the width (column number) direction. In the example below, I want to make B as wide as A by repeating its columns. <\/p>\n\n<p>This can be done in numpy by:<\/p>\n\n<pre><\/pre>\n\n<p>Yields:<\/p>\n\n<pre><\/pre>\n\n<p>How can I do this for two constant tensors A and B, in Tensorflow? <\/p>\n","answers":[{"AnswerId":"41719856","CreationDate":"2017-01-18T12:48:01.910","ParentId":null,"OwnerUserId":"5016028","Title":null,"Body":"<pre><code>A = tf.constant([[1,2,3,4,4,5,6,7],[4,5,6,6,4,5,6,7],[6,7,8,9,4,5,6,7]])\nB = tf.constant([[19,15],\n             [18,14],\n             [17,13]])\n\ndiff = A.get_shape()[1] - B.get_shape()[1]\n\nBt = tf.transpose(B)\nfor idx in range(diff):\n    col = tf.gather_nd(Bt, [[idx]])\n    Bt = tf.concat(0, [Bt, col])\n\nresult = tf.transpose(Bt)\n\nwith tf.Session() as sess:\n    res = sess.run(result)\n    print(res)\n<\/code><\/pre>\n\n<p>not the most beautiful code ever, but it works.<\/p>\n\n<p>Output: <\/p>\n\n<pre><code>[[19 15 19 15 19 15 19 15]\n [18 14 18 14 18 14 18 14]\n [17 13 17 13 17 13 17 13]]\n<\/code><\/pre>\n"}]}
{"QuestionId":41717773,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-18T11:05:57.157","AcceptedAnswerId":41817843.0,"OwnerUserId":1128171.0,"Title":"Ordering from tensorflow queues","Body":"<p>I am trying to understand queues in better detail. Using the code below I expect that since I am not shuffling the alphabetic list the output collection will be in alphabetic order. This seems to be the case for all but the initial epoch. Am I misunderstanding something?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41817843","CreationDate":"2017-01-24T00:01:21.093","ParentId":null,"OwnerUserId":"1128171","Title":null,"Body":"<p>Changing the above to the below clears up my confusion<\/p>\n\n<pre><code> try:\n      step = 0\n      while not coord.should_stop():\n        start_time = time.time()\n\n        # Run one step of the model.\n        integer =  sess.run(images)\n        #print(\"Integer val\", integer)\n\n        char =  sess.run(alphs)\n        collection.append(char[0][0])\n        print(\"String val\", char)\n\n\n        duration = time.time() - start_time\n\n    except tf.errors.OutOfRangeError:\n      print('Saving')\n      print('Done training for %d epochs, %d steps.' % (FLAGS.num_epochs, step))\n    finally:\n      # When done, ask the threads to stop.\n      coord.request_stop()\n    print(str(collection))\n<\/code><\/pre>\n"}]}
{"QuestionId":41717943,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-18T11:14:03.187","AcceptedAnswerId":null,"OwnerUserId":826983.0,"Title":"Why is the conv2d output (?, 6, 1, 1) instead of (?, 5, 1, 1)","Body":"<p>What I try to understand is how I can create a convolutional neural network for natural language processing (NLP). To start of, I am using a standard architecture where a filter  slides over an input matrix  in order to create a convolutional filter map:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/O2NPI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/O2NPI.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The code below is supposed to implement the convolution step, but I do not quite understand the output dimensions. Please do not get confused from the image. In my code <em>rows<\/em> are word embeddings of words in contrast to the image, where word embeddings are columns.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>What I expect for a  sized filter  sliding over a  input matrix, where 10 is the number of word-vectors and 300 their respective dimension, is a  convolutional feature vector. If the filter is being placed \"at the start\" of the input matrix is and sled over it 1 step at a time, the filter can only make 5 steps without exceeding the input matrix.<\/p>\n\n<p>As you can see from the comments in the code, I am getting a  convolutional feature vector. Because of that I am quite sure that I misunderstood what's happening here.<\/p>\n\n<p>Why is  dimensioned  and not ?<\/p>\n","answers":[{"AnswerId":"41734787","CreationDate":"2017-01-19T05:59:17.613","ParentId":null,"OwnerUserId":"2888347","Title":null,"Body":"<p>If you write all the combinations of consecutive intervals (indexing from 0), you will get 6 cases:<\/p>\n\n<p><code>[0,4], [1,5], [2,6], [3,7], [4,8], [5,9]<\/code><\/p>\n"}]}
{"QuestionId":41719812,"AnswerCount":1,"Tags":"<python><numpy><theano><keras><keras-layer>","CreationDate":"2017-01-18T12:45:34.103","AcceptedAnswerId":null,"OwnerUserId":4730816.0,"Title":"Error while calling eval() on Tensor variable in keras","Body":"<p>I am using keras and using a layer output for some modifications. Before, using the output ( a tensor variable ) I am converting it to numpy array and thus calling eval() on it, as below:<\/p>\n\n<pre><\/pre>\n\n<p>The code fails with following error:<\/p>\n\n<pre><\/pre>\n\n<p>I intend to invoke a conversion function on the output of the previous layer). The conversion function takes the input as a tensor variable and do the computation. However, I also want to visualize the data and I need to compute bit_length on each element of it.<\/p>\n\n<p>For eg If layer A gives an output Y1. This output is used by Lambda layer L1 and it invokes the conversion method. Sample code:<\/p>\n\n<pre><\/pre>\n\n<p>I used similar scheme for converting the layer weights using model.get_weights(). The get_weights() method returns numpy array list and thus could easily iterate on each element and calc the bit length. But while converting the output is an issue as the output is a tensor variable and when i call eval() on it, gives the error I mentioned in previous post. I hope I was able to clear my intentions.<\/p>\n","answers":[{"AnswerId":"41726990","CreationDate":"2017-01-18T18:36:44.700","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You don't need to call eval() at all, your conversion_method should be done using symbolic functions (the ones from keras.backend) and should be differentiable.<\/p>\n\n<p>It won't work otherwise, the network won't be able to be trained with Keras\/Theano.<\/p>\n"}]}
{"QuestionId":41720346,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-01-18T13:11:02.747","AcceptedAnswerId":null,"OwnerUserId":7149693.0,"Title":"tensorflow slim inception_v3 model error","Body":"<p>I am trying to use the tensorflow inception_v3 model for a transfer learning project.I get the following error on building the model.\n\nThe same error does not arise on using the same script for inception_v1 model.\nThe models are imported from slim.nets<\/p>\n\n<p>Running on CPU\nTensorflow version : 0.12.1<\/p>\n\n<p><strong>Script<\/strong> <\/p>\n\n<pre><\/pre>\n\n<p><strong>Full Error Message<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41724258","CreationDate":"2017-01-18T16:16:51.810","ParentId":null,"OwnerUserId":"7149693","Title":null,"Body":"<p>Found my mistake, i was importing the model from <code>https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/slim\/python\/slim\/nets<\/code> whereas the updated models are at <code>https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\/nets<\/code>.<\/p>\n\n<p>Still haven't understood why there are two different repositories for the same classes.Must be a valid reason.<\/p>\n"}]}
{"QuestionId":41720751,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-18T13:29:40.200","AcceptedAnswerId":null,"OwnerUserId":6667826.0,"Title":"ignoring warning: Unable to load cuDNN DSO , effecting fft computation?","Body":"<p>Recently installed the GPU version of tensorflow 0.12.1 and using python 3.5.2 for windows. <\/p>\n\n<p>I used to use tensorflow cpu version and I didnt get these warning for other mathematical computations. FFT didnt work in cpu version<\/p>\n\n<pre><\/pre>\n\n<p>And as suggested in other thread i changed to the gpu version. I used to get warning for other functionalities, but now when ever i am using fft i get this error does any one how to resolve it ? I dont understand the error.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41731737","CreationDate":"2017-01-19T00:20:57.093","ParentId":null,"OwnerUserId":"2888347","Title":null,"Body":"<p>You need to copy the <code>dll<\/code> associated with <code>cudnn<\/code>, into a folder that is an element of your <code>PATH<\/code> environment variable. Since you are finding the other CUDA libraries, just copy the <code>cudnn64_v5.dll<\/code> to the directory with the other CUDA libraries.<\/p>\n"}]}
{"QuestionId":41721058,"AnswerCount":1,"Tags":"<if-statement><boolean><tensorflow>","CreationDate":"2017-01-18T13:44:18.147","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Writing an if statement in tensorflow and comparing the values of two tensors","Body":"<p>I have the condition:<\/p>\n\n<pre><\/pre>\n\n<p>where n_spec and s_spec are two tensors (2d arrays) and I want to perform that concatenation if one is smaller than the other in the 0 dimension. When I try this, tensorflow throws an error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I reformulate the above?<\/p>\n","answers":[{"AnswerId":"41724319","CreationDate":"2017-01-18T16:19:04.233","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>From your code, it seems like the shape of <code>n_spec<\/code> is known statically (since you are using <code>n_spec.get_shape()<\/code>). In this case, you can simply do the following :<\/p>\n\n<pre><code>if n_spec.get_shape()[0] &lt; s_spec.get_shape()[0]:\n  n_spec = tf.concat(0, [tf.zeros([empty_cols, n_spec.get_shape()[1]], tf.int32), n_spec])\n<\/code><\/pre>\n"}]}
{"QuestionId":41721526,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-18T14:05:00.787","AcceptedAnswerId":42030009.0,"OwnerUserId":5780341.0,"Title":"Tensorflow Serving basic example","Body":"<p>I am working through the Tensorflow serving_basic example at:<\/p>\n\n<p><a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"nofollow noreferrer\">https:\/\/tensorflow.github.io\/serving\/serving_basic<\/a><\/p>\n\n<h2>Setup<\/h2>\n\n<p>Following: <a href=\"https:\/\/tensorflow.github.io\/serving\/setup#prerequisites\" rel=\"nofollow noreferrer\">https:\/\/tensorflow.github.io\/serving\/setup#prerequisites<\/a><\/p>\n\n<p>Within a docker container based off of ubuntu:latest, I have installed:<\/p>\n\n<h3>bazel:<\/h3>\n\n<pre><\/pre>\n\n<h3>grpcio:<\/h3>\n\n<pre><\/pre>\n\n<h3>all packages:<\/h3>\n\n<pre><\/pre>\n\n<h3>tensorflow serving:<\/h3>\n\n<pre><\/pre>\n\n<p>I've built the source with bazel and all tests ran successfully:<\/p>\n\n<pre><\/pre>\n\n<p>I can successfully export the mnist model with:<\/p>\n\n<pre><\/pre>\n\n<p>And I can serve the exported model with:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<h2>The problem<\/h2>\n\n<p>When I test the server and try to connect a client to the model server with:<\/p>\n\n<pre><\/pre>\n\n<p>I see this output:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41925665","CreationDate":"2017-01-29T20:37:07.963","ParentId":null,"OwnerUserId":"6027520","Title":null,"Body":"<p>The \"--use_saved_model\" model flag is set to default \"true\"; use the --use_saved_model=false when starting the server. This should work:<\/p>\n\n<pre><code>bazel-bin\/tensorflow_serving\/model_servers\/tensorflow_model_server --use_saved_model=false --port=9000 --model_name=mnist --model_base_path=\/tmp\/mnist_model\/\n<\/code><\/pre>\n"},{"AnswerId":"42030009","CreationDate":"2017-02-03T17:40:59.260","ParentId":null,"OwnerUserId":"5780341","Title":null,"Body":"<p>I mentioned this on the <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/308#issuecomment-277306134\" rel=\"nofollow noreferrer\">tensorflow github<\/a>, and the solution was to remove the original model that had been created.  If you're running into this, run \n<code>\nrm -rf \/tmp\/mnist_model\n<\/code>\nand rebuild it<\/p>\n"}]}
{"QuestionId":41722535,"AnswerCount":1,"Tags":"<windows><installation><tensorflow><gpu><amd-gpu>","CreationDate":"2017-01-18T14:54:13.220","AcceptedAnswerId":null,"OwnerUserId":2230045.0,"Title":"How can I use TensorFlow on Windows with AMD GPU?","Body":"<p>I want to use TensorFlow on Windows (Win 10) with a AMD GPU.<\/p>\n\n<p>If I google, there are a lot discussions and sources but I just couldn't figure out what's the best way to do this at the moment.<\/p>\n\n<p><strong>Could someone write a short installation instruction<\/strong> that he thinks is the best and most up-to-date way of doing so?<\/p>\n","answers":[{"AnswerId":"52694446","CreationDate":"2018-10-08T01:48:44.423","ParentId":null,"OwnerUserId":"4534466","Title":null,"Body":"<p>Tensorflow officially only supports CUDA, which is a proprietary NVIDIA technology. There is one unofficial implementation using openCL <a href=\"https:\/\/github.com\/hughperkins\/tf-coriander\" rel=\"nofollow noreferrer\">here<\/a> which could work, or you could try using Google colab<\/p>\n"}]}
{"QuestionId":41722882,"AnswerCount":0,"Tags":"<indexing><tensorflow><linear-algebra><idiomatic><submatrix>","CreationDate":"2017-01-18T15:10:46.397","AcceptedAnswerId":null,"OwnerUserId":3202904.0,"Title":"Update submatrix in Tensorflow","Body":"<p>Quite simply, what I want to do is the following<\/p>\n\n<pre><\/pre>\n\n<p>except in Tensorflow (where the matrices can be arbitrarily complicated tensor expressions). Neither  nor  is a Tensorflow Variable, but just a run-of-the-mill tensor.<\/p>\n\n<p>What I have gathered so far: tensors are immutable, so I cannot assign to a submatrix.  is the current option for sub-assignment, but does not appear to support sub-matrices, only slices.<\/p>\n\n<p>Methods that should work, but are perhaps not ideal: <\/p>\n\n<ul>\n<li>I could pad B with zeros, but I'm sure this leads to instantiation of\nan unnecessarily large B - can it be made sparse, maybe?<\/li>\n<li>I could use the padding idea, but write it as a low-rank decomposition, e.g. in Numpy:  where  is a stacked zero and identity matrix. I'm not sure this is actually advantageous. <\/li>\n<li>I could split  into submatrices, and stack them back together. Might be the most efficient, but sounds like the code would be convoluted.<\/li>\n<\/ul>\n\n<p>Ideally, I want to do this operation N times for progressively smaller matrices, resulting in one large final result, but this is tangential.<\/p>\n\n<p>I'll use one of the hacks for now, but I'm hoping someone can tell me what the idiomatic version is!     <\/p>\n","answers":[]}
{"QuestionId":41723117,"AnswerCount":1,"Tags":"<tensorflow><cudnn>","CreationDate":"2017-01-18T15:22:37.077","AcceptedAnswerId":41723762.0,"OwnerUserId":4262324.0,"Title":"How to report issues on Tensorflow website?","Body":"<p>How can I report an issue on Tensorflow website? I am not talking about the API, but everything else, e.g. installation instructions and tutorials. <\/p>\n\n<p>For instance, installation instructions indicate that Tensorflow for Python 3.5 and GPU requires CuDNN v5, but that is incorrect, as with CuDNN v5 it doesn't work, and it requires CuDNN v5.1 instead.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41723762","CreationDate":"2017-01-18T15:53:57.133","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>Please open an issue at the TensorFlow GitHub Issues page: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues<\/a><\/p>\n"}]}
{"QuestionId":41723247,"AnswerCount":2,"Tags":"<python><scikit-learn><keras>","CreationDate":"2017-01-18T15:28:54.553","AcceptedAnswerId":41724334.0,"OwnerUserId":1141351.0,"Title":"SKlearn Tfidfvectorizer with Keras: expected dense_input_1 to have shape","Body":"<p>I'm trying to combine SKLearn Tfidfvectorizer with Keras but I'm stuck at the following error: \nException: Error when checking model input: expected dense_input_1 to have shape (None, 126) but got array with shape (700, 116)<\/p>\n\n<p>I understand it has something to do with the shape of the matrix but I can't figure out how to solve it. <\/p>\n\n<pre><\/pre>\n\n<p>Build model:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41724334","CreationDate":"2017-01-18T16:19:42.057","ParentId":null,"OwnerUserId":"1141351","Title":null,"Body":"<p>The problem is the difference in dimensions in x_train and x_test. Changing the max features in the tfidfvectorizer solved the problem.<\/p>\n\n<pre><code>vectorizer = TfidfVectorizer(analyzer=self.identity, use_idf=True, max_features=100)\n<\/code><\/pre>\n"},{"AnswerId":"45684479","CreationDate":"2017-08-14T23:29:17.220","ParentId":null,"OwnerUserId":"1491900","Title":null,"Body":"<p>When you vectorize your test set, you need to call <code>transform<\/code> and not <code>fit_transform<\/code>:<\/p>\n\n<pre><code>x_train_vec = vectorizer.fit_transform(x_train).toarray()\nx_test_vec = vectorizer.transform(self.x_test[i]).toarray()\n<\/code><\/pre>\n"}]}
{"QuestionId":41724406,"AnswerCount":0,"Tags":"<python><optimization><tensorflow>","CreationDate":"2017-01-18T16:23:01.157","AcceptedAnswerId":null,"OwnerUserId":7436370.0,"Title":"How to implement an external optimizer in Tensorflow with apply_gradients method?","Body":"<p>I'm trying to define some new optimization algorithm. I found an interface about external optimizer <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/opt\/python\/training\/external_optimizer.py\" rel=\"nofollow noreferrer\">here<\/a>. It says that I need to subclass it and implement _minimize. However, I need to call apply_gradients when using the external optimizer. How can I implement it?<\/p>\n\n<p>For example, I want to implement a custom optimizer by myself, so I look into the external optimizer and it says:<\/p>\n\n<pre><\/pre>\n\n<p>Also there is an official example that how to use this external optimizer interface in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/opt\/python\/training\/external_optimizer_test.py#L38\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>And the Mock Optimizer are used at Line85 and 90:<\/p>\n\n<pre><\/pre>\n\n<p>It seems that we override the 'minimize' function of ExternalOptimizerInterface. But sometimes, I need to use Optimizer.apply_gradients(). What should I do?<\/p>\n","answers":[]}
{"QuestionId":41725214,"AnswerCount":1,"Tags":"<python><multithreading><tensorflow><queue>","CreationDate":"2017-01-18T16:58:55.407","AcceptedAnswerId":null,"OwnerUserId":5038048.0,"Title":"Does queue enqueue_many block unitll all items are on the queue?","Body":"<p>I'm working on a problem where I need to guarantee the order of items on the queue, however, I would also like to use multiple threads to do this.  Thus I'm wondering if a call to enqueue_many will ensure that those items are placed in order?<\/p>\n\n<p>One solution to this is to only use one thread to enqueue items, however that is not optimal.  <\/p>\n\n<p>The documentation only mentions blocking in the case where the queue is full.<\/p>\n\n<blockquote>\n  <p>tf.QueueBase.enqueue_many(vals, name=None)<\/p>\n  \n  <p>Enqueues zero or more elements to this queue.<\/p>\n  \n  <p>This operation slices each component tensor along the 0th dimension to\n  make multiple queue elements. All of the tensors in vals must have the\n  same size in the 0th dimension.<\/p>\n  \n  <p>If the queue is full when this operation executes, it will block until\n  all of the elements have been enqueued.<\/p>\n  \n  <p>At runtime, this operation may raise an error if the queue is closed\n  before or during its execution. If the queue is closed before this\n  operation runs, tf.errors.AbortedError will be raised. If this\n  operation is blocked, and either (i) the queue is closed by a close\n  operation with cancel_pending_enqueues=True, or (ii) the session is\n  closed, tf.errors.CancelledError will be raised. Args:<\/p>\n\n<pre><\/pre>\n  \n  <p>Returns:<\/p>\n  \n  <p>The operation that enqueues a batch of tuples of tensors to the queue.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/io_ops\/queues#QueueBase\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/io_ops\/queues#QueueBase<\/a><\/p>\n","answers":[{"AnswerId":"41796694","CreationDate":"2017-01-22T22:09:12.953","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Empirically they seem to be atomic. IE, here's an <a href=\"https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/enqueue_many_test_singlerun.py\" rel=\"nofollow noreferrer\">example<\/a> that enqueues a large block of <code>zeros<\/code> and <code>ones<\/code> simultaneously, and then checks if there has been any mixing.<\/p>\n"}]}
{"QuestionId":41725873,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network><mnist>","CreationDate":"2017-01-18T17:33:31.200","AcceptedAnswerId":null,"OwnerUserId":7433039.0,"Title":"Tensor Flow - low accuracy on CNN Mnist Data set \/ How to batch accuracy calculations","Body":"<p>The code is in python 3.5.2 with Tensor flow. The neural network returns an accuracy of between .10 and 5.00, with the higher value tending to be the accuracy of the training data by a factor of roughly 6. I cannot tell whether the neural network is legitimately doing worse than random guessing or if the accuracy code i am using has a serious fault i cannot see.<\/p>\n\n<p>The neural network consists of 5 layers:<\/p>\n\n<ol>\n<li>input <\/li>\n<li>conv1 (with max pooling relu and dropout)<\/li>\n<li>conv2 (with max pooling relu and dropout)<\/li>\n<li>fully connected (with relu)<\/li>\n<li>output <\/li>\n<\/ol>\n\n<p>uses default Adam optimizer <\/p>\n\n<p>I feel very suspicious of my accuracy calculations as i made them differently than what i have seen due to RAM constraints. The accuracy calculation does both the accuracy of the train and test data.<\/p>\n\n<pre><\/pre>\n\n<p>This is a sample of the outputs:<\/p>\n\n<ul>\n<li>Epoch 0 completed out of 20 loss: 10333239.3396       83.29 ts 429<\/li>\n<li>Epoch 0 current test set accuracy :  0.7072 <\/li>\n<li>Epoch 0 current train set accuracy :  3.8039<\/li>\n<li>Epoch 1 completed out of 20 loss: 1831489.40747       39.24 ts 858 <\/li>\n<li>Epoch 1 current test set accuracy :  0.7765 <\/li>\n<li>Epoch 1 current train set accuracy :  4.2239 <\/li>\n<li>Epoch 2 completed out of 20 loss: 1010191.40466       25.89 ts 1287 <\/li>\n<li>Epoch 2 current test set accuracy :  0.8069 <\/li>\n<li>Epoch 2 current train set accuracy :  4.3898 <\/li>\n<li>Epoch 3 completed out of 20 loss: 631960.809082       0.267 ts 1716 <\/li>\n<li>Epoch 3 current test set accuracy :  0.8277 <\/li>\n<li>Epoch 3 current train set accuracy :  4.4955 <\/li>\n<li>Epoch 4 completed out of 20 loss: 439149.724823       2.001 ts 2145 <\/li>\n<li>Epoch 4 current test set accuracy :  0.8374 <\/li>\n<li>Epoch 4 current train set accuracy :  4.5674<\/li>\n<\/ul>\n\n<p>The full code is as follows (sorry about the length i added a lot of comments for my own use ):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41754043","CreationDate":"2017-01-20T00:18:10.593","ParentId":null,"OwnerUserId":"7433039","Title":null,"Body":"<p>The CNN had low results because of 4 reasons:<\/p>\n\n<ol>\n<li>Improper (Lack of) feeding of dropout \n -the keep rate was not being fed into <code>accuracy.eval(feed_dict={x: test_x, y: test_y})<\/code> causing the network to underpreform in its accuracy evaluations<\/li>\n<li>Poor Initialization of weights\n\n<ul>\n<li>RELU neuron work significantly better with weights closer to zero than normal distribution.<\/li>\n<\/ul><\/li>\n<li>far to high learning rate\n\n<ul>\n<li>Learning rate of .03 even with decay was far far to high and stoped it from training effectively<\/li>\n<\/ul><\/li>\n<li>errors in accuracy function\n\n<ul>\n<li>The accuracy function of the training data was receiving the size of the data set form <code>mnist.test.num_examples<\/code> instead of the proper <code>mnist.train.num_examples<\/code> and caused nonsensical values of accuracy in excess of 100% <\/li>\n<\/ul><\/li>\n<\/ol>\n"}]}
{"QuestionId":41726804,"AnswerCount":2,"Tags":"<tensorflow><gradient>","CreationDate":"2017-01-18T18:26:56.033","AcceptedAnswerId":null,"OwnerUserId":3219076.0,"Title":"Tensorflow update only selected variables","Body":"<p><strong>Overview:<\/strong> I want to update only selected variables in a network. The network has parts -> (in forward direction) and each of them has separate losses  and . I want to train the weights  of  to optimize . While doing this, the weights  of  should be fixed. How can I do this?<\/p>\n\n<p><strong>Approach 1:<\/strong> Select only  as variables to minimize using  in .\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/834\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/834<\/a> . This crashes with an error . This actually works fine in other scenarios, but apparently it does not like that weights  are not in the var_list.<\/p>\n\n<p><strong>Edit 1:<\/strong> The line that causes the error: <\/p>\n\n<p><strong>Approach 2:<\/strong> Same as Approach 1, but also include  in the var_list. The problem is now that the network updates a and b, whereas it should just send the gradients through  and only update .<\/p>\n\n<p><strong>Edit 2:<\/strong> The line that works, but is not what I want: <\/p>\n\n<p><strong>Approach 3:<\/strong> Use  <a href=\"https:\/\/stackoverflow.com\/questions\/34477889\/holding-variables-constant-during-optimizer\/34478044#34478044\">Holding variables constant during optimizer<\/a> . From the documentation I infer that this only stops the gradients from flowing further to the left in the graph. I want the ignore variables on the right.<\/p>\n\n<p><strong>Approach 4:<\/strong> Set , but that looks very inflexible if I want to alternate training between A and B.<\/p>\n","answers":[{"AnswerId":"41767658","CreationDate":"2017-01-20T15:56:35.330","ParentId":null,"OwnerUserId":"3219076","Title":null,"Body":"<p>It turns out that the final op in A was non-differentiable (tf_argmax) and therefore obviously gradients could not be passed from B to A.<\/p>\n"},{"AnswerId":"42037249","CreationDate":"2017-02-04T06:21:49.677","ParentId":null,"OwnerUserId":"2015952","Title":null,"Body":"<p>I found that, for a better control of which variables to update during the optimization, it is better to use: 'compute_gradients' and 'apply_gradients' approach.<\/p>\n\n<p>The compute_gradients will return a list of tuple of gradients and variables tensors. You can modify the returning gradient tensors whatever you want and also be able to select the subset of variables for updating.<\/p>\n\n<p>Then, you pass a list of tuple of gradients and variables that you want to update to 'apply_gradients'<\/p>\n\n<p>Here are some examples:<\/p>\n\n<pre><code>optimizer = tf.train.AdamOptimizer(learning_rate=0.0001)\ngrads = optimizer.compute_gradients(your_cost_function)\n\n# You can update 'g' and exclude some v's \ngrad_lists = [(g, v) for g, v in grads]\n\ntrain_op = optimizer.apply_gradients(grad_lists)\n\nThen, run your session.\n\nsess.run(train_op, feed_dict={...})\n<\/code><\/pre>\n\n<p>Also, since you have 2 loss functions, you should create 2 train operations.<\/p>\n\n<p>Hope this help!<\/p>\n"}]}
{"QuestionId":41726954,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-18T18:34:53.913","AcceptedAnswerId":null,"OwnerUserId":953273.0,"Title":"Printing tensorflow variable seq2seq","Body":"<p>I am trying to print a tensor attns in tensorflow seq2seq code. <a href=\"https:\/\/github.com\/petewarden\/tensorflow_ios\/blob\/master\/tensorflow\/python\/ops\/seq2seq.py#L557\" rel=\"nofollow noreferrer\">Seq2Seq.py<\/a><\/p>\n\n<p>I tried:<\/p>\n\n<pre><\/pre>\n\n<p>but it prints nothing.<\/p>\n\n<p>I tried <\/p>\n\n<pre><\/pre>\n\n<p>I this case it throws:  InvalidArgumentError: You must feed a value for placeholder tensor <\/p>\n\n<p>I have also tried using sess.run()<\/p>\n\n<pre><\/pre>\n\n<p>In this case sess object is None.<\/p>\n","answers":[{"AnswerId":"41732219","CreationDate":"2017-01-19T01:21:20.287","ParentId":null,"OwnerUserId":"4385912","Title":null,"Body":"<p><code>tf.Print<\/code> is not a \"classic\" operational instruction, since those are not executed in symbolic, graph-based code. What is needed instead is a specific node in the computation graph that will then be triggered whenever your computations \"passes\" that node.<\/p>\n\n<p>This is exactly what tf.Print does. It creates a \"wrapper\" node around any other node by creating an identity operation that, when triggered, prints the value of a list of tensors. <\/p>\n\n<p>The first argument of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/debugging_operations#Print\" rel=\"nofollow noreferrer\">this print function<\/a>, <code>input_<\/code> (or <code>attns<\/code> in your case) is the wrapped node, and <code>data<\/code> (or <code>[attns]<\/code> in your case) is the list of tensors to be printed.<\/p>\n\n<p>What you therefore want to do is to add this line:<\/p>\n\n<pre><code>attns = tf.Print(attns, [attns])\n<\/code><\/pre>\n\n<p>Here, <code>attns<\/code> is assigned a print wrapper identity operation on <code>attns<\/code> - so the tensor <code>attns<\/code> has exactly the same behavior, except that when it is computed, it will also print <code>[attns]<\/code>.<\/p>\n"}]}
{"QuestionId":41728089,"AnswerCount":2,"Tags":"<lua><emulation><ubuntu-16.04><torch>","CreationDate":"2017-01-18T19:42:25.510","AcceptedAnswerId":null,"OwnerUserId":7437583.0,"Title":"Can not run lua script from Fceux emulator usnig torch","Body":"<p>I am using Fceux 2.2.3 emulator and i am trying to run a lua script using torch while playing super mario bros. When i load the script i face this error<\/p>\n\n<blockquote>\n  <p>Lua thread bombed out: error loading module 'libpaths' from file\n  '\/home\/antonis\/torch\/install\/lib\/lua\/5.1\/libpaths.so':\n    \/home\/antonis\/torch\/install\/lib\/lua\/5.1\/libpaths.so: undefined\n  symbol: lua_gettop<\/p>\n<\/blockquote>\n\n<p>I looked in the specific folder in the path and the libpaths.so is there.\nWhat this error means?<\/p>\n","answers":[{"AnswerId":"41728545","CreationDate":"2017-01-18T20:08:24.987","ParentId":null,"OwnerUserId":"7437081","Title":null,"Body":"<p>Are you using a different Lua version to compile your code and run it? If not, you can try the steps in <a href=\"https:\/\/stackoverflow.com\/a\/31160347\">this answer<\/a><\/p>\n"},{"AnswerId":"54227436","CreationDate":"2019-01-17T00:35:59.787","ParentId":null,"OwnerUserId":"834518","Title":null,"Body":"<p>You need to build tha latest FCEUX from source (<a href=\"https:\/\/github.com\/TASVideos\/fceux\" rel=\"nofollow noreferrer\">https:\/\/github.com\/TASVideos\/fceux<\/a>) using the option to use system libraries. Then you specify LuaJIT (install it first) in the SContruct file and install. After that, install torch on your system. It will be available to FCEUX through \"require 'torch'\". That is the way I did it (right now, tbh), but you could probably install Torch first and then link its own LuaJIT files to FCEUX.<\/p>\n"}]}
{"QuestionId":41729019,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-18T20:38:57.027","AcceptedAnswerId":41800414.0,"OwnerUserId":588495.0,"Title":"NotFoundError running TensorFlow XLA example (libdevice.compute_35.10.bc)","Body":"<p>I'm running the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/experimental\/xla\/jit\" rel=\"nofollow noreferrer\">tutorial example for XLA<\/a> using a TensorFlow compiled from source.  Running  results in the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have CUDA 8 installed with cuDNN 5.1.  The file  does exist on the machine:<\/p>\n\n<pre><\/pre>\n\n<p>My hunch is that this has something to do with the message , but I'm not sure what to do about it.<\/p>\n","answers":[{"AnswerId":"41864292","CreationDate":"2017-01-26T00:47:59.510","ParentId":null,"OwnerUserId":"1884421","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/7079\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/pull\/7079<\/a> should be able to fix this. Thanks for the bug report! <\/p>\n"},{"AnswerId":"41800414","CreationDate":"2017-01-23T06:29:11.693","ParentId":null,"OwnerUserId":"61394","Title":null,"Body":"<p>The key is this log message:<\/p>\n\n<pre><code>I tensorflow\/core\/platform\/default\/cuda_libdevice_path.cc:35] TEST_SRCDIR environment variable not set: using local_config_cuda\/cuda under this executable's runfiles directory as the CUDA root.\n<\/code><\/pre>\n\n<p>(I only noticed it in the logs later; I actually found that file by digging around in the sources and only then noticed the message in the logs.)<\/p>\n\n<p>For reasons that I do not currently understand, XLA does not look in \/usr\/local\/cuda (or whatever directory you gave when you ran <code>.\/configure<\/code>) for libdevice.  Per cuda_libdevice_path.cc [1], it's looking for a symlink that was created specifically to point it to libdevice.<\/p>\n\n<p>I'm going to loop in the person who wrote this code to figure out what it's supposed to be doing.  In the meantime, I was able to work around it myself as follows:<\/p>\n\n<pre><code>$ mkdir local_config_cuda\n$ ln -s \/usr\/local\/cuda local_config_cuda\/cuda\n$ TEST_SRCDIR=$(pwd) python my_program.py\n<\/code><\/pre>\n\n<p>The important thing is to set <code>TEST_SRCDIR<\/code> to the parent of the local_config_cuda directory.<\/p>\n\n<p>Sorry for the trouble, and sorry I don't have a less-hacky answer for you right now.<\/p>\n\n<p>[1] <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/e1f44d8\/tensorflow\/core\/platform\/cuda_libdevice_path.cc#L23\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/e1f44d8\/tensorflow\/core\/platform\/cuda_libdevice_path.cc#L23<\/a> <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1084748efa3234c7daa824718aeb7df7b9252def\/tensorflow\/core\/platform\/default\/cuda_libdevice_path.cc#L27\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1084748efa3234c7daa824718aeb7df7b9252def\/tensorflow\/core\/platform\/default\/cuda_libdevice_path.cc#L27<\/a><\/p>\n"}]}
{"QuestionId":41729878,"AnswerCount":0,"Tags":"<python><tensorflow><eval><binaryfiles>","CreationDate":"2017-01-18T21:34:40.710","AcceptedAnswerId":null,"OwnerUserId":4040030.0,"Title":"Infinite Loop When Reading TFRecord file","Body":"<p>I'm currently trying to read a TFRecord file but my terminal crashes whenever I try to evaluate the tensors I have just generated from the file data. I'm working with RNNs (recurrent neural networks) and so I am trying to do my work using sequence data. I've copied most of the code from <a href=\"http:\/\/www.wildml.com\/2016\/08\/rnns-in-tensorflow-a-practical-guide-and-undocumented-features\/\" rel=\"nofollow noreferrer\">here<\/a> but I have made my own additions to try and include a TFRecordReader.<\/p>\n\n<pre><\/pre>\n\n<p>I'm still getting familiar with tensorflow so I'd appreciate an explanation for what I'm doing wrong (not just code fixes) so I don't make a similar mistake in future. Thank You!<\/p>\n\n<p>--<\/p>\n\n<p>EDIT: Ok I've altered the above with the following:<\/p>\n\n<pre><\/pre>\n\n<p>But this is giving me the following error: <\/p>\n\n<p><strong>ERROR:tensorflow:Exception in QueueRunner: Attempted to use a closed Session.<\/strong>  <\/p>\n\n<p>I'm not sure what is wrong with this. Yaroslav's comments are correct according to the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow noreferrer\">directions on tensorflow<\/a> but I'm not sure what I've done wrong. I've tried several different formats of my code but I'm not sure what it is I'm missing.<\/p>\n\n<p>--<\/p>\n\n<p>EDIT2: Ok, I figured it out. Apparently I needed to generate a coordinator and pass that to the queue runners to get it to work. Here is my finalized code:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41730822,"AnswerCount":1,"Tags":"<tensorflow><machine-translation>","CreationDate":"2017-01-18T22:45:25.980","AcceptedAnswerId":null,"OwnerUserId":3399881.0,"Title":"Adding to vocab during Tensorflow seq2seq run","Body":"<p>I am using the Tensorflow seq2seq <a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/\" rel=\"nofollow noreferrer\">tutorial<\/a> to play with machine translation. Say I have trained the model for some time and determine that I want to supplement the original vocab with new words to enhance the quality of the model. Is there a way to pause training, add words to the vocabulary, and then resume training from the most recent checkpoint? I attempted to do so but when I began training again I got this error: <\/p>\n\n<pre><\/pre>\n\n<p>Obviously the new vocab is larger and so the tensor sizes do not match. Is there some way around this?<\/p>\n","answers":[{"AnswerId":"47348243","CreationDate":"2017-11-17T10:15:46.523","ParentId":null,"OwnerUserId":"6363961","Title":null,"Body":"<p>You can not update your vocab once its set but you can always use shared wordpiece model. It will help you to directly copy out of vocab words form source to target output.<\/p>\n"}]}
{"QuestionId":41731171,"AnswerCount":1,"Tags":"<python><ubuntu><tensorflow>","CreationDate":"2017-01-18T23:16:28.487","AcceptedAnswerId":null,"OwnerUserId":7437971.0,"Title":"Cannot import Tensorflow correctly when running a python code","Body":"<p>Help!!! I have been working on this for several days and am at wit's end!  I have been trying to run a python code that keeps giving me this error no matter what I do....<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/rmVus.jpg\" rel=\"nofollow noreferrer\">Tensorflow problem<\/a><\/p>\n","answers":[{"AnswerId":"41731229","CreationDate":"2017-01-18T23:21:47.890","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Your Numpy is older than expected. You can try to update numpy.<\/p>\n\n<p>I suggest to use the anaconda distribution and install the official tensorflow package and should work seamlessly. <\/p>\n"}]}
{"QuestionId":41731544,"AnswerCount":1,"Tags":"<tensorflow><classification><loss>","CreationDate":"2017-01-18T23:56:51.020","AcceptedAnswerId":null,"OwnerUserId":2994860.0,"Title":"Using Tensorflow for multi-label classification","Body":"<p>I have a multi-label classification problem with 5 labels. Is it correct if I have a my output layer to be of shape [n_samples, n_labels=5] and have the loss function as tf.nn.sigmoid_cross_entropy_with_logits? Or should I have the output layer in the shape of [n_samples, n_labels=5, 2] where 2 is to encode if each label is assigned to the sample or not?<\/p>\n","answers":[{"AnswerId":"41731963","CreationDate":"2017-01-19T00:47:00.517","ParentId":null,"OwnerUserId":"4922660","Title":null,"Body":"<p>I think the shape of output layer should be [n_samples, n_labels], and when you using tf.nn.softmax_cross_entropy_with_logits, make sure the label array is in the format of one-hot, another recommended interface is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard4\/tf.nn.sparse_softmax_cross_entropy_with_logits.md\" rel=\"nofollow noreferrer\">tf.nn.sparse_softmax_cross_entropy_with_logits<\/a>, which labels should be the entry of n_labels;<\/p>\n"}]}
{"QuestionId":41731564,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><theano><theano-cuda>","CreationDate":"2017-01-18T23:58:58.220","AcceptedAnswerId":41731607.0,"OwnerUserId":5605365.0,"Title":"If I train a neural network with CUDA do I need to run the outputted algorithm with CUDA?","Body":"<p>Let's say I used CUDA to train an object tracking program. Could I then put that program on another computer that didn't have a powerful gpu and run the object tracking program? Or is gpu support required to run the outputted algorithm as well as train it?<\/p>\n","answers":[{"AnswerId":"41731607","CreationDate":"2017-01-19T00:04:41.667","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>No, it does not matter how you trained your model. You can execute it in completely different scenario, using CPU, GPU, cloud or whatever you want. Since execution is usually much cheaper than training - you will usually need much less powerful hardware.<\/p>\n"}]}
{"QuestionId":41731743,"AnswerCount":1,"Tags":"<python><numpy><out-of-memory><theano><keras>","CreationDate":"2017-01-19T00:21:26.343","AcceptedAnswerId":null,"OwnerUserId":4524164.0,"Title":"understanding Keras LSTM ( lstm_text_generation.py ) - RAM memory issues","Body":"<p>I'm diving into LSTM RNN with <strong>Keras<\/strong> and <strong>Theano<\/strong> backend. While trying to use lstm examples from keras' repo <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/lstm_text_generation.py\" rel=\"nofollow noreferrer\">whole code of lstm_text_generation.py on github<\/a>, I've got one thing that isn't pretty clear to me: the way it's vectorizing the input data (text characters):<\/p>\n\n<pre><\/pre>\n\n<p>Here, as you can see, they generate lists of zeros with <strong>Numpy<\/strong> and then put '1' to particular position of each list defined by input characters encoding sequences in such way.<\/p>\n\n<p>The question is: why did they use that algorithm? is it possible to optimize it somehow? maybe it's possible to encode input data in some other way, not using huge lists of lists? The problem is that it has severe limits of input data: generating such vectors for >10 Mb text causes MemoryError of Python (dozens of Gbs RAM needed to process it!). <\/p>\n\n<p>Thanks in advance, guys.<\/p>\n","answers":[{"AnswerId":"41741672","CreationDate":"2017-01-19T12:19:49.600","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>There are at least two optimizations in Keras which you could use in order to decrease amount of memory which is need in this case:<\/p>\n\n<ol>\n<li><p>An <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\">Embedding layer<\/a> which makes it possible to accept only a single integer intead of full one hot vector. Moreover - this layer could be pretrained before the final stage of network training - so you could inject some prior knowledge into your model (and even finetune it during the network fitting).<\/p><\/li>\n<li><p>A <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\"><code>fit_generator<\/code><\/a> method makes it possible to train a network using a predefinied generator which would produce pairs <code>(x, y)<\/code> need in network fitting. You could e.g. save the whole dataset to disk and read it part by part using a generator interface.<\/p><\/li>\n<\/ol>\n\n<p>Of course - both of this methods could be mixed. I think that simplicity was the reason behind this kind of implementation in the example you provided. <\/p>\n"}]}
{"QuestionId":41732521,"AnswerCount":2,"Tags":"<python><ubuntu><ubuntu-14.04><virtualenv><keras>","CreationDate":"2017-01-19T01:58:26.320","AcceptedAnswerId":null,"OwnerUserId":2414957.0,"Title":"Keras can't be imported in Virtual Env","Body":"<p>Any idea how this could be fixed?<\/p>\n\n<pre><\/pre>\n\n<p>UPDATE:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41733409","CreationDate":"2017-01-19T03:48:36.530","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>Is it possible when creating <code>cv2<\/code> environment, you did not create it with <code>--system-site-packages<\/code> flag?\nIf not, you can always <code>pip install keras<\/code> inside the virtualenv.<\/p>\n"},{"AnswerId":"41755167","CreationDate":"2017-01-20T02:34:48.160","ParentId":null,"OwnerUserId":"4237316","Title":null,"Body":"<p>The problem is that you are running <code>pip<\/code> with <code>sudo<\/code> from the virtualenv. This temporarily makes you root, resets all the environment variables that are set by virtualenvwrapper and thus installs the packages system-wide instead in the virtualenv.<\/p>\n\n<p>This should work better:<\/p>\n\n<pre><code>workon cv2\npip install keras tensorflow\n<\/code><\/pre>\n"}]}
{"QuestionId":41732537,"AnswerCount":0,"Tags":"<c++><machine-learning><tensorflow>","CreationDate":"2017-01-19T02:00:26.180","AcceptedAnswerId":null,"OwnerUserId":5166390.0,"Title":"How do you use load data from a CSV in C++ TensorFlow?","Body":"<p>I'm trying to load a model trained in Python into C++ and classify some data from a CSV. I found this tutorial:<\/p>\n\n<p><a href=\"https:\/\/medium.com\/@hamedmp\/exporting-trained-tensorflow-models-to-c-the-right-way-cf24b609d183#.3bmbyvby0\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@hamedmp\/exporting-trained-tensorflow-models-to-c-the-right-way-cf24b609d183#.3bmbyvby0<\/a> <\/p>\n\n<p>Which lead me to this piece of example code:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/label_image\/main.cc\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/label_image\/main.cc<\/a><\/p>\n\n<p>Which is looking very hopeful for me. However, the data I want to load is in a CSV, and not an image file, so I'm trying to rewrite the ReadTensorFromImageFile function. I was able to find a class DecodeCSV, but it's  a little different than the DecodePNG and DecodeJpeg classes in the example code, and I end up with an OutputList instead of and Output. Using the [] operator on the list seems to crash my program. If anyone happens to know how to deal with this, it would be greatly appreciated. He are the relevant changes to the code:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41732627,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-19T02:14:02.917","AcceptedAnswerId":null,"OwnerUserId":7438526.0,"Title":"Tensorflow feed_dict key cannot be interpreted as Tensor","Body":"<p>I just starts to learn Tensorflow in python.  I got the following error when I starts with a simple AddTwo class.  the error messages are:<\/p>\n\n<blockquote>\n  <p>Cannot interpret feed_dict key as Tensor: Tensor\n  Tensor(\"Placeholder:0\", dtype=float32) is not an element of this\n  graph.<\/p>\n<\/blockquote>\n\n<p>Can anyone help me to point out the correct way for me?<\/p>\n\n<pre><\/pre>\n\n<p>main code that calls the AddTwo class:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41744241","CreationDate":"2017-01-19T14:28:13.473","ParentId":null,"OwnerUserId":"213550","Title":null,"Body":"<p>As I suggested in comment, you should open the session with created graph, so code should be like this:<\/p>\n\n<pre><code>with self.graph.as_default():\n    # no session here\n    self.X = tf.placeholder(tf.float32)\n    self.Y = tf.placeholder(tf.float32)\n\n# open session with providing the graph\nwith tf.Session(graph=self.graph) as sess:\n    pass\n<\/code><\/pre>\n"}]}
{"QuestionId":41733210,"AnswerCount":2,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-01-19T03:28:31.607","AcceptedAnswerId":41733822.0,"OwnerUserId":1060036.0,"Title":"What size should my image be to retrain Inception?","Body":"<p>I am following \n<a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/#0\" rel=\"nofollow noreferrer\">codelabs Tensorflow for poets<\/a> guide to re-train inception v3 with my own images. But there is no mention of what the size my image should be. I also watched some Youtube video that suggested cropping and filling in white spaces to make square images. But it didn't really mention the size. <\/p>\n\n<p>How should I resize my training images to so I get the best result re-training inception?<\/p>\n","answers":[{"AnswerId":"41733822","CreationDate":"2017-01-19T04:34:49.653","ParentId":null,"OwnerUserId":"6912318","Title":null,"Body":"<p>The code prepares images for you and automatically and feeds them into the network. All you need to do is to properly setup the folders and provide enough training images. In my experience the size of images doesn't matter too much. I did retraining following the instructions using 640x480 and 1280x1024 images. I got great results with training sizes of 400 - 1000 images per class. <\/p>\n"},{"AnswerId":"47133089","CreationDate":"2017-11-06T08:58:58.410","ParentId":null,"OwnerUserId":"521617","Title":null,"Body":"<p>The code does the resizing for you. Have a look at <a href=\"https:\/\/github.com\/googlecodelabs\/tensorflow-for-poets-2\/blob\/master\/scripts\/retrain.py\" rel=\"noreferrer\" title=\"retrain.py\">retrain.py<\/a>. I have listed the code responsible for deciding the size of the images depending on the network architecture.<\/p>\n\n<pre><code>if architecture == 'inception_v3':\n    # pylint: disable=line-too-long\n    data_url = 'http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz'\n    # pylint: enable=line-too-long\n    bottleneck_tensor_name = 'pool_3\/_reshape:0'\n    bottleneck_tensor_size = 2048\n    input_width = 299\n    input_height = 299\n    input_depth = 3\n    resized_input_tensor_name = 'Mul:0'\n<\/code><\/pre>\n"}]}
{"QuestionId":41733331,"AnswerCount":0,"Tags":"<merge><keras>","CreationDate":"2017-01-19T03:42:01.103","AcceptedAnswerId":null,"OwnerUserId":5863915.0,"Title":"keras TypeError: unsupported operand type(s) for *: 'NoneType' and 'int' ,when I use a self-defined merge function","Body":"<pre><\/pre>\n\n<p>The function scaleMult means that the one element tensor 'en' multiply the 8\u00d78\u00d79 tensor 'patch'.\nOf course, I should construct multiplication between 32*1 tensor 'en' and 32\u00d78\u00d78\u00d79 tensor 'patch' when training in 32 size batch.<\/p>\n\n<p>But there always shows the error 'fan_in = shape[1] * receptive_field_size TypeError: unsupported operand type(s) for *: 'NoneType' and 'int'' at the Convolution2D step.<\/p>\n","answers":[]}
{"QuestionId":41734275,"AnswerCount":1,"Tags":"<python><tensorflow><celery>","CreationDate":"2017-01-19T05:18:33.340","AcceptedAnswerId":41777375.0,"OwnerUserId":3620076.0,"Title":"No response from celery worker with TensorFlow","Body":"<p>I am trying to use tensorflow in a celery worker. I encountered timeout rather than receiving response from the worker.<\/p>\n\n<p>I used following code:<\/p>\n\n<p>tasks.py<\/p>\n\n<pre><\/pre>\n\n<p>test.py<\/p>\n\n<pre><\/pre>\n\n<p>I executed a worker with this command.<\/p>\n\n<p><\/p>\n\n<p>When I executed the worker,  was printed out, since  had .\nThis means that tensorflow works properly.<\/p>\n\n<p>Then, I ran:<\/p>\n\n<p><\/p>\n\n<p>Then, I got:<\/p>\n\n<p><\/p>\n\n<p>What was wrong?\nHow can I investigate what happened?<\/p>\n\n<p>My environment is:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"41777375","CreationDate":"2017-01-21T08:47:05.203","ParentId":null,"OwnerUserId":"2052105","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>import tensorflow as tf\nfrom celery import Celery\nfrom celery.utils.log import get_task_logger\nfrom celery.signals import worker_init, worker_process_init\nfrom models import Network, Extractor\nfrom celery.concurrency import asynpool\nasynpool.PROC_ALIVE_TIMEOUT = 100.0 #set this long enough\n\nlogger = get_task_logger(__name__)\n\nCELERY_BROKER_URL = 'redis:\/\/localhost:6379\/'\nCELERY_RESULT_BACKEND = 'redis:\/\/localhost:6379\/'\n\n# Celery: Distributed Task Queue\napp = Celery('tasks', backend=CELERY_RESULT_BACKEND, broker=CELERY_BROKER_URL)\napp.conf.task_serializer   = 'json'\napp.conf.result_serializer = 'json'\n\ntf_model = None\n\n@worker_process_init.connect()\ndef on_worker_init(**_):\n  global tf_model\n  # Create server with model\n  logger.info('model for worker: started init')\n  print(\"model for dsa\")\n  session = tf.Session()\n  model = Network(session, True)\n  #model.load_model('.\/models\/test_2')\n  extractor = Extractor(model)\n  tf_model = extractor\n  logger.info('model for worker: initialized')\n\n\n@app.task(name='process_single')\ndef process_single(image):\n    logger.info('process_single: started')\n    descriptor = tf_model.process_single(image)\n    logger.info('process_single: completed')\n\n    return descriptor\n<\/code><\/pre>\n\n<p>It looks for me that this is working:<\/p>\n\n<pre><code>[2017-01-21 09:41:18,892: INFO\/Worker-1] ???[???]: model for worker: started init\n[2017-01-21 09:41:18,893: WARNING\/Worker-1] model for dsa\n[2017-01-21 09:41:18,902: INFO\/MainProcess] Connected to redis:\/\/localhost:6379\/\/\n[2017-01-21 09:41:18,915: INFO\/MainProcess] mingle: searching for neighbors\n[2017-01-21 09:41:19,920: INFO\/MainProcess] mingle: all alone\n[2017-01-21 09:41:19,949: WARNING\/MainProcess] celery@cospelpc ready.\n[2017-01-21 09:41:20,930: INFO\/Worker-1] ???[???]: model for worker: initialized\n[2017-01-21 09:41:31,648: INFO\/MainProcess] Received task: process_single[024068ba-9ea2-4405-8aab-d3504a06aa55]\n[2017-01-21 09:41:31,658: INFO\/Worker-1] process_single[024068ba-9ea2-4405-8aab-d3504a06aa55]: process_single: started\n[2017-01-21 09:41:33,125: INFO\/Worker-1] process_single[024068ba-9ea2-4405-8aab-d3504a06aa55]: process_single: completed\n[2017-01-21 09:41:33,128: INFO\/MainProcess] Task process_single[024068ba-9ea2-4405-8aab-d3504a06aa55] succeeded in 1.470330449s: [153608.4375, 0.0, 0.0, 243285.75, 0.0, 155679.671875, 346120.625, 70663.265625, 0.0, 29445.03125, 0.0, 518396.25, 0.0,...\n<\/code><\/pre>\n"}]}
{"QuestionId":41734584,"AnswerCount":2,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2017-01-19T05:44:43.650","AcceptedAnswerId":43279791.0,"OwnerUserId":6862133.0,"Title":"TensorFlow CNN Tutorial: How to edit top layer to be locally connected?","Body":"<p>I have some background in machine learning and python, but I am just learning TensorFlow. I am going through the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">tutorial on deep convolutional neural nets<\/a> to teach myself how to use it for image classification. Along the way there is an exercise, which I am having trouble completing.<\/p>\n\n<p><strong>EXERCISE:<\/strong> The model architecture in inference() differs slightly from the CIFAR-10 model specified in cuda-convnet. In particular, the top layers of Alex's original model are locally connected and not fully connected. Try editing the architecture to exactly reproduce the locally connected architecture in the top layer.<\/p>\n\n<p>The exercise refers to the inference() function in the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10.py\" rel=\"nofollow noreferrer\">cifar10.py model<\/a>. The 2nd to last layer (called local4) has a shape=[384, 192], and the top layer has a shape=[192, NUM_CLASSES], where NUM_CLASSES=10 of course. I think the code that we are asked to edit is somewhere in the code defining the top layer:<\/p>\n\n<pre><\/pre>\n\n<p>But I don't see any code that determines the probability of connecting between layers, so I don't know how we can change the model from fully connected to locally connected. Does somebody know how to do this?<\/p>\n","answers":[{"AnswerId":"43279791","CreationDate":"2017-04-07T13:46:29.060","ParentId":null,"OwnerUserId":"4632380","Title":null,"Body":"<p>I'm also working on this exercise. I'll try and explain my approach properly, rather than just give the solution. It's worth looking back at the mathematics of a fully connected layer (<a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners<\/a>).<\/p>\n\n<p>So the linear algebra for a fully connected layer is:<\/p>\n\n<p><em>y = W * x + b<\/em><\/p>\n\n<p>where <em>x<\/em> is the n dimensional input vector, <em>b<\/em> is an <em>n<\/em> dimensional vector of biases, and <em>W<\/em> is an <em>n<\/em>-by-<em>n<\/em> matrix of weights. The <em>i<\/em> th element of <em>y<\/em> is the sum of the <em>i<\/em> th row of W multiplied element-wise with <em>x<\/em>.<\/p>\n\n<p>So....if you only want <em>y[i]<\/em> connected to <em>x[i-1]<\/em>, <em>x[i]<\/em>, and <em>x[i+1]<\/em>, you simply set all values in the <em>i<\/em> th row of <em>W<\/em> to zero, apart from the <em>(i-1)<\/em> th, <em>i<\/em> th and <em>(i+1)<\/em> th column of that row. Therefore to create a locally connected layer, you simply enforce <em>W<\/em> to be a banded matrix (<a href=\"https:\/\/en.wikipedia.org\/wiki\/Band_matrix\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Band_matrix<\/a>), where the size of the band is equal to the size of the locally connected neighbourhoods you want. Tensorflow has a function for setting a matrix to be banded (<code>tf.batch_matrix_band_part(input, num_lower, num_upper, name=None)<\/code>).<\/p>\n\n<p>This seems to me to be the simplest mathematical solution to the exercise.<\/p>\n"},{"AnswerId":"42233039","CreationDate":"2017-02-14T17:56:30.103","ParentId":null,"OwnerUserId":"933703","Title":null,"Body":"<p>I'll try to answer your question although I'm not 100% I got it right as well. <\/p>\n\n<p>Looking at the cuda-convnet <a href=\"https:\/\/github.com\/akrizhevsky\/cuda-convnet2\/blob\/master\/layers\/layers-cifar10-11pct.cfg\" rel=\"nofollow noreferrer\">architecture<\/a> we can see that the TensorFlow and cuda-convnet implementations start to differ after the second pooling layer.<\/p>\n\n<p>TensorFlow implementation implements two fully connected layers and softmax classifier.<\/p>\n\n<p>cuda-convnet implements two locally connected layers, one fully connected layer and softmax classifier.<\/p>\n\n<p>The code snippet you included refers only to the softmax classifier and is in fact shared between the two implementations. To reproduce the cuda-convnet implementation using TensorFlow we have to replace the existing fully connected layers with two locally connected layers and a fully connected one.<\/p>\n\n<p>Since Tensor doesn't have locally connected layers as part of the SDK we have to figure out a way to implement it using the existing tools. Here is my attempt to implement the first locally connected layers:<\/p>\n\n<pre><code>  with tf.variable_scope('local3') as scope:\n    shape = pool2.get_shape()\n    h = shape[1].value\n    w = shape[2].value\n\n    sz_local = 3 # kernel size\n    sz_patch = (sz_local**2)*shape[3].value\n    n_channels = 64\n\n    # Extract 3x3 tensor patches\n    patches = tf.extract_image_patches(pool2, [1,sz_local,sz_local,1], [1,1,1,1], [1,1,1,1], 'SAME')\n    weights = _variable_with_weight_decay('weights', shape=[1,h,w,sz_patch, n_channels], stddev=5e-2, wd=0.0)\n    biases = _variable_on_cpu('biases', [h,w,n_channels], tf.constant_initializer(0.1))\n\n    # \"Filter\" each patch with its own kernel \n    mul = tf.multiply(tf.expand_dims(patches, axis=-1), weights)\n    ssum = tf.reduce_sum(mul, axis=3)\n    pre_activation = tf.add(ssum, biases)\n    local3 = tf.nn.relu(pre_activation, name=scope.name)\n<\/code><\/pre>\n"}]}
{"QuestionId":41735132,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe><image-segmentation><pycaffe>","CreationDate":"2017-01-19T06:26:11.050","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"FCN32 Model is not converging and loss is fluctuating after some points. Why?","Body":"<p>I am trying to train fcn32. I am training <a href=\"https:\/\/github.com\/shelhamer\/fcn.berkeleyvision.org\/tree\/master\/voc-fcn32s\" rel=\"nofollow noreferrer\">voc-fcn32s<\/a> model for my own data that has imbalanced class number. This is the learning curve for 18,000 iterations:\n<a href=\"https:\/\/i.stack.imgur.com\/BISxi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BISxi.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>As you can see training is diminishing in some points and then it is fluctuating. I read some online recommendations that they are suggesting reducing the learning rate or changing the bias value in convolution layers for fillers. So, what I did, is that I changed the <a href=\"https:\/\/github.com\/shelhamer\/fcn.berkeleyvision.org\/blob\/master\/voc-fcn32s\/train.prototxt\" rel=\"nofollow noreferrer\">train_val.prototxt<\/a> as follows for these two layers:<\/p>\n\n<pre><\/pre>\n\n<p>And this the trend of model\n<a href=\"https:\/\/i.stack.imgur.com\/6BXzp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6BXzp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>It seems not much thing has changed in the behavior of the model.<\/p>\n\n<p>1) Am doing the right way to add these values to ?<\/p>\n\n<p>2) Should I change the learning policy in the solver from  to  by reducing by the factor of 10 each time? Will it help to tackle this issue?<\/p>\n\n<p>I am worried that I am doing the wrong things and my model cannot converge. Does anyone have any suggestion about this? What important things I should consider while training model? What kind of changes can I do on  and  that model to be converged? <\/p>\n\n<p>I really appreciate your help.<\/p>\n\n<p><strong>More details after adding BatchNorm layer<\/strong><\/p>\n\n<p>Thanks @Shai and @Jonathan for suggesting to add  layers. \nI added  before  layers, this one sample layer:<\/p>\n\n<pre><\/pre>\n\n<p>As far as I knew from docs, I can only add one parameter in Batch normalization  instead of three since I have single channel images. Is this my understanding true? as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Should I add more parameters to scale layer, as documentation is mentioning? What are the meaning of these parameters in  layer? like:<\/p>\n\n<pre><\/pre>\n\n<p>and this is <a href=\"https:\/\/i.stack.imgur.com\/sigzi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/sigzi.png\" alt=\"the drawing\"><\/a> of the network. I am not sure how much I am wrong\/right. Have I added correctly?\nThe other question is about <a href=\"https:\/\/stackoverflow.com\/questions\/40510706\/how-to-interpret-caffe-log-with-debug-info\">debug_info<\/a>. What is the meaning of these lines of log file after activating ? What does it mean of  and ? And why the values are 0? Is my net working correctly?<\/p>\n\n<pre><\/pre>\n\n<p>I really appreciate if someone knows, please share ideas\/links\/resources here. Thanks again<\/p>\n","answers":[{"AnswerId":"41737403","CreationDate":"2017-01-19T08:48:35.443","ParentId":null,"OwnerUserId":"249226","Title":null,"Body":"<p>I would not expect changing the bias values to help with the training. The first thing I would try is lowering the learning rate. You can do this manually by retraining the the weights that have reached a plateau and use a solver with a lower base_lr. Or you can change your solver.prototxt to use a different update policy. You can either set the method to step or you can use an update policy such as Adam. See:<\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/solver.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/tutorial\/solver.html<\/a><\/p>\n\n<p>As <a href=\"https:\/\/stackoverflow.com\/questions\/41735132\/fcn32-model-is-not-converging-and-loss-is-fluctuating-after-some-points-why#comment70665511_41735132\">@Shai recommends<\/a>, adding <code>\"BatchNorm\"<\/code> layers should help. Batch Normalization is similar to \"whitening\"\/normalizing the input data, but is applied to the intermediate layers. The paper on Batch normalization is on <a href=\"https:\/\/arxiv.org\/abs\/1502.03167\" rel=\"nofollow noreferrer\">arxiv<\/a>.<\/p>\n\n<p>You should also reserve some data for validation. Just looking at the training loss can be misleading.<\/p>\n"},{"AnswerId":"41807458","CreationDate":"2017-01-23T13:24:04.857","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Regarding <code>\"BatchNorm\"<\/code> parameters:<br>\nThis layer has three internal parameters: (0) mean, (1) variance, and (2) moving average factor, <strong>regardless<\/strong> of the number of channels or the shape of your blob. Therefore, if you wish to explicitly set <code>lr_mult<\/code> you need to define it for all three. <\/p>\n\n<p>Regarding the zeros in the log:<br>\nPlease <a href=\"https:\/\/stackoverflow.com\/q\/40510706\/1714410\">read this post<\/a> on how to read caffe's debug log.<br>\nIt seems like you are training your model from scratch (not fine tuning) and that all weights are set to zero. This is a very poor initialization strategy.<br>\nPlease consider defining <code>filler<\/code>s and <code>bias_filler<\/code>s for initializing the weights.<\/p>\n"}]}
{"QuestionId":41735145,"AnswerCount":1,"Tags":"<ios><machine-learning><tensorflow>","CreationDate":"2017-01-19T06:27:03.243","AcceptedAnswerId":null,"OwnerUserId":3378821.0,"Title":"Tensorflow iOS returns different results in simple and camera examples using retrained inception model","Body":"<p>I have retrained the inception model to recognize custom set of images. I have tried the simple and camera examples and replaced the .pb and .txt files with my own files. I am able to get a reliable result of probability of 0.99 when using the simple example (replaced the grace_hopper with my image). However, I got different results when using the camera example. I point my camera to the photo but the result seems different from that obtained from simple example. The config I am using is as follows, is there any config I need to make? Thank you.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41752867","CreationDate":"2017-01-19T22:25:47.057","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>The input\/output layers have changed between Inception v1 (the version packaged with the Android demo) and v3 (the one you retrained).<\/p>\n\n<p>You'll need to update the constants found in in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/ClassifierActivity.java\" rel=\"nofollow noreferrer\">ClassifierActivity.java<\/a> with the values for Inception v3 (also found in the comments there).<\/p>\n"}]}
{"QuestionId":41736552,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-19T07:55:22.887","AcceptedAnswerId":null,"OwnerUserId":6201311.0,"Title":"Adding new Op in Tensorflow: shape functions","Body":"<p>I'm trying to add a new Op (using this how-to: <a href=\"https:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/<\/a>). I have no problem with simple Op like in example, but I need to add more complex Op. It should have 2 inputs, and output that has same shape as matrix multiplication of inputs. \nHow can I write shape function for this case?\nHow can I allocate output with appropriate shape?\nThank you in advance.<\/p>\n","answers":[{"AnswerId":"41744414","CreationDate":"2017-01-19T14:35:59.057","ParentId":null,"OwnerUserId":"6201311","Title":null,"Body":"<p>Have partial answer to the question.\nI still don't know how to write shape function, but I wrote checks like this:<\/p>\n\n<pre><code>const Tensor&amp; input1 = context-&gt;input(0);\nconst Tensor&amp; input2 = context-&gt;input(1);\nTensorShape sh1 = input1.shape();\nTensorShape sh2 = input2.shape();\nOP_REQUIRES(context, sh1.dim_size(1)==sh2.dim_size(0),\n errors::InvalidArgument(\"Can't multiplicate!\"));\n<\/code><\/pre>\n\n<p>Allocating:<\/p>\n\n<pre><code>sh1.RemoveDim(1);\nsh2.RemoveDim(0);\nsh1.AppendShape(sh2);\nOP_REQUIRES_OK(context, context-&gt;allocate_output(0, sh1, &amp;output_tensor));\n<\/code><\/pre>\n\n<p>But seems like I'm reinventing a bicycle. Is it any easier way to do it?<\/p>\n"}]}
{"QuestionId":41736870,"AnswerCount":0,"Tags":"<tensorflow><concatenation><tile>","CreationDate":"2017-01-19T08:14:32.113","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Repeating a matrix's columns using the padding function in Tensorflow","Body":"<p>I have written the following function:<\/p>\n\n<pre><\/pre>\n\n<p>the idea is that A is a wider matrix than B. I want to widen B by repeated copies of itself (should also work when the row-length of A does not exactly divide the one of B).<\/p>\n\n<p>If you run the above code, you get:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea on how this can be done just by using tensorflow functions? <\/p>\n","answers":[]}
{"QuestionId":41737957,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2017-01-19T09:18:20.173","AcceptedAnswerId":null,"OwnerUserId":4929293.0,"Title":"Reusing a LSTM cell in tensor flow","Body":"<p>I want to train many case with only one LSTM cell.  <\/p>\n\n<p>I tried like below and it give me an error like this <\/p>\n\n<p>I think this error is occurred because I tried to reuse LSTM cell.\nI can't find the method to solve this error.   <\/p>\n\n<p>Someone saids using  function to reuse LSTM cell but I can't understand how to use it.\nBelow is my source code. How can I solve this problem<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42730197","CreationDate":"2017-03-11T01:11:17.310","ParentId":null,"OwnerUserId":"7693042","Title":null,"Body":"<p>You can use <\/p>\n\n<pre><code>tf.get_variable_scope().reuse_variables()\n<\/code><\/pre>\n"}]}
{"QuestionId":41737992,"AnswerCount":1,"Tags":"<tensorflow><machine-learning><ctc>","CreationDate":"2017-01-19T09:20:15.770","AcceptedAnswerId":42550408.0,"OwnerUserId":3187222.0,"Title":"Connectionist Temporal Classification (CTC) blank label","Body":"<p>I am trying to use the CTC loss function in my network, but don't quite understand when to feed the 'blank' label as a label.<\/p>\n\n<p>I use it in gesture recognition as described by<a href=\"http:\/\/www.cv-foundation.org\/openaccess\/content_cvpr_2016\/papers\/Molchanov_Online_Detection_and_CVPR_2016_paper.pdf\" rel=\"nofollow noreferrer\">Molchanov<\/a>, but what get's me confused that there is a 'no gesture' as well.<\/p>\n\n<p>In tensorflow docs, it is described that <\/p>\n\n<blockquote>\n  <p>The inputs Tensor's innermost dimension size, num_classes, represents  num_labels + 1 classes, where num_labels is the number of true labels, and the largest value (num_classes - 1) is reserved for the blank label.<\/p>\n<\/blockquote>\n\n<p>If I now use the 'blank' label, to indicate that there is no gesture, I am limited in my training, because of the error<\/p>\n\n<blockquote>\n  <p>Saw a non-null label (index >= num_classes - 1) following a null label<\/p>\n<\/blockquote>\n\n<p>I am assuming that null label is the same as the blank label.<\/p>\n\n<p>The problem is, when I want to feed data that starts with no gesture (mapped to null label) and has then a gesture, I get exactly this error.\nI can avoid it by adding two more labels, one for 'no gesture' and one for 'blank label\/null label' next to my existing labels. Then I only feed the 'no gesture' label but never the 'blank' label, but this doesn't seem quite right.<\/p>\n\n<p>So my question is, what should I use the 'blank\/null' label for?<\/p>\n\n<p>I can imagine in language processing, you would use the sentence ending dot usually as the 'null' label? But there is no ending gesture as it is one continuous stream.<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"42550408","CreationDate":"2017-03-02T08:35:02.367","ParentId":null,"OwnerUserId":"3187222","Title":null,"Body":"<p><strong>EDIT<\/strong>\nI highly recommend reading this <a href=\"https:\/\/distill.pub\/2017\/ctc\/\" rel=\"nofollow noreferrer\">distill article<\/a>.\n\"The \u03f5 (blank) token doesn\u2019t correspond to anything and is simply removed from the output.\" It is used to 'interrupt' the merging of repeating tokens.<\/p>\n\n<p>The blank label serves as a transitioning state between two classes.<\/p>\n\n<p><a href=\"https:\/\/gab41.lab41.org\/speech-recognition-you-down-with-ctc-8d3b558943f0\" rel=\"nofollow noreferrer\">Read more<\/a><\/p>\n\n<p>To answer my question itself, you don't assign the blank label to anything, but still have it as an existing class.\nIn my case, I had added two more labels, one for the no gesture class and  one for the blank.<\/p>\n\n<p>(That's at least how I did it and got some decent results)<\/p>\n"}]}
{"QuestionId":41738993,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-19T10:07:12.080","AcceptedAnswerId":41747290.0,"OwnerUserId":482601.0,"Title":"How bad is it to have unused operations floating about?","Body":"<p>It would make my code easier to generate a bunch of operations even though they are not used.<\/p>\n\n<p>Does having unused operations in the graph have a non-negligible impact on performance?<\/p>\n","answers":[{"AnswerId":"41747290","CreationDate":"2017-01-19T16:48:10.620","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>TensorFlow executes only the ops necessary to get result requested by user so it's safe to leave unused operations in graph.<\/p>\n\n<p>Note that there's a 2GB limit on size of graph, so if you add too many unused ops, you may hit this error.<\/p>\n"}]}
{"QuestionId":41739283,"AnswerCount":1,"Tags":"<c++><database><caffe><lmdb>","CreationDate":"2017-01-19T10:20:02.417","AcceptedAnswerId":42039048.0,"OwnerUserId":4428123.0,"Title":"Bad Retrieval in LMDB Get?","Body":"<p>I'm using lmdb database in my project, I have a problem in retrieving data with lmdb's mdb_get function...I am sure about the mdb_put is working correctly but the get function concatenate some garbage with my data...<\/p>\n\n<pre><\/pre>\n\n<p>the returned str contains some garbage like for example, str should contain \"w ah n\" but when I retrieved it mdb_get  returns \"w ah n\\0004\", it concatenates garbage after string null character.<\/p>\n","answers":[{"AnswerId":"42039048","CreationDate":"2017-02-04T10:06:11.003","ParentId":null,"OwnerUserId":"4232652","Title":null,"Body":"<p>The problem is this line: <\/p>\n\n<pre><code>return (char *)data.mv_data;\n<\/code><\/pre>\n\n<p>Because <code>data.mv_data<\/code> is actually just a bunch of bytes not a null-terminated string. If you just cast it to <code>char*<\/code> you can get garbage at the end of string.<\/p>\n\n<p>The solution is to copy <code>data.mv_size bytes<\/code> from <code>data.mv_data<\/code> to a newly allocated <code>char*<\/code>. Alternatively, you can pass the destination pointer as an <code>out<\/code> parameter of your <code>Get<\/code> method.<\/p>\n\n<p>At least this is my solution.<\/p>\n"}]}
{"QuestionId":41739445,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-19T10:27:30.047","AcceptedAnswerId":null,"OwnerUserId":3952274.0,"Title":"Tensorflow: strange behavior in cifar10 example using custom variable creation method","Body":"<p>This is a follow up question to <a href=\"https:\/\/stackoverflow.com\/questions\/41714801\/tensorflow-variable-with-weight-decay-explanation\/41717150#41717150\">this one<\/a>.<\/p>\n\n<p>I'm still working on the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">cifar10 example<\/a> on the file <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10.py#L115\" rel=\"nofollow noreferrer\">cifar10.py<\/a> and noticed some strange behavior regarding the creation of variables.<\/p>\n\n<p>But first a side-question: Why are the variables created with weight decay factor of  and not ? That way you would have less vertices in the computation graph.<\/p>\n\n<p>Next, the strange behavior. I added the following function to make it more convenient to create variables:<\/p>\n\n<pre><\/pre>\n\n<p>When using this function to create the variables (with the original parameters), the logits that are computed come from a range of +-1e13 for the first batch, gradually getting better reaching +-1.5.  The loss on the other hand starts at around 400000 and gets bigger until it hits NaN.<\/p>\n\n<p>When using the original functions to create the variables, the logits come from a range of +-1 right from the beginning and the loss start at around 4.5, gradually getting smaller.<\/p>\n\n<p>Can somebody explain to me what the difference between my and the provided functions for variable generation is, and why the effect is so huge? I don't see it.<\/p>\n\n<p>The full code of my modified cifar10.py can be found <a href=\"http:\/\/pastebin.com\/Cv3jVPPv\" rel=\"nofollow noreferrer\">here<\/a>. To test it out simple replace the original file with my version. To than switch between the original and my function simply change line 212 to <\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"41741967","CreationDate":"2017-01-19T12:35:57.303","ParentId":null,"OwnerUserId":"3952274","Title":null,"Body":"<p>Stupid me! I used my own function the wrong way and passed the values for stddev as the mean and used the default stddev of 1.<\/p>\n\n<p>The curse of not addressing the arguments by their name.<\/p>\n\n<p>Anyway, why does this cause such a huge loss; sometimes even NaN?<\/p>\n"}]}
{"QuestionId":41740101,"AnswerCount":2,"Tags":"<python><tensorflow><protocol-buffers>","CreationDate":"2017-01-19T10:57:50.430","AcceptedAnswerId":44211182.0,"OwnerUserId":6311759.0,"Title":"tensorflow difference between saving model via exporter and tf.train.write_graph()?","Body":"<p>What  is difference between saving a model by<\/p>\n\n<ol>\n<li>using exporter as specified in tensorflow serving:<\/li>\n<\/ol>\n\n<p>eg:<\/p>\n\n<pre><\/pre>\n\n<ol start=\"2\">\n<li>Using tf.train.write_graph() and tf.train.Saver() directly:<\/li>\n<\/ol>\n\n<p>eg:<\/p>\n\n<pre><\/pre>\n\n<p>Question is in continuation of <a href=\"https:\/\/stackoverflow.com\/questions\/38947658\/tensorflow-saving-into-loading-a-graph-from-a-file\">TensorFlow saving into\/loading a graph from a file<\/a><\/p>\n","answers":[{"AnswerId":"44211182","CreationDate":"2017-05-26T22:59:36.080","ParentId":null,"OwnerUserId":"985088","Title":null,"Body":"<p>Given the Exporter is now officially deprecated, the new protocol for saving graph and data is to use Saver.  Here is an excellent blog with sample code:  <a href=\"https:\/\/blog.metaflow.fr\/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc\" rel=\"nofollow noreferrer\">https:\/\/blog.metaflow.fr\/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc<\/a>.<\/p>\n"},{"AnswerId":"43200336","CreationDate":"2017-04-04T06:58:46.893","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>Exporter is used after you train a model and want to serve it (use it for inference).<\/p>\n\n<p>Saver is used for checkpointing when you train a model. <\/p>\n"}]}
{"QuestionId":41740846,"AnswerCount":1,"Tags":"<python><tensorflow><nvidia>","CreationDate":"2017-01-19T11:36:45.113","AcceptedAnswerId":null,"OwnerUserId":235472.0,"Title":"Built TensorFlow successfully; example fails","Body":"<p>After successfully building TensorFlow from source, the first test example passes, and the second one fails:   <\/p>\n\n<p>1) First test: OK<\/p>\n\n<pre><\/pre>\n\n<p>2) Second test: FAILED<br>\n   Test file: tensorflow\/models\/image\/mnist\/convolutional.py<\/p>\n\n<pre><\/pre>\n\n<p>Since the  is one of the tests reported in the installation guide, I would expect it to be fine.<\/p>\n\n<p>I ran all the tests in the project's directory, and these are the results:  <\/p>\n\n<ul>\n<li>MNIST: failed<\/li>\n<li>AlexNet: passed<\/li>\n<li>Cifar10: failed<\/li>\n<li>ImageNet: passed <\/li>\n<\/ul>\n\n<p>Could have I done anything wrong in the build, or am I doing anything wrong in running these tests? <\/p>\n\n<p>Reference: <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/os_setup<\/a><br>\nPlatform: Linux\/Ubuntu 16.04, nVidia Quadro M6000, CUDA enabled.<\/p>\n","answers":[{"AnswerId":"41748882","CreationDate":"2017-01-19T18:14:33.303","ParentId":null,"OwnerUserId":"3657941","Title":null,"Body":"<p>From the <a href=\"https:\/\/www.tensorflow.org\/resources\/#report_issues\" rel=\"nofollow noreferrer\">Tensorflow web site<\/a>:<\/p>\n\n<blockquote>\n  <p>Please report bugs, feature requests and installation \/ compatibility issues on the TensorFlow issues tracker on GitHub.<\/p>\n<\/blockquote>\n\n<p>This appears to be an installation issue and is more appropriately handled by opening an issue on GitHub.<\/p>\n"}]}
{"QuestionId":41742153,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-19T12:46:56.203","AcceptedAnswerId":41742232.0,"OwnerUserId":5963300.0,"Title":"Why tensorflow equal gives incorrect result","Body":"<p>The  function seems to be weired when we apply a CNN network. In the case below,  returns incorrect result. <\/p>\n\n<pre><\/pre>\n\n<p>In the code, inputs would read the images by  and  defines our CNN network. The output of  is the logits of the last fully connected layer. In my case, the label is one_hot encoded and there are 2 classes, so it should be [1,0] or [0,1].<\/p>\n\n<p>The result is below:<\/p>\n\n<pre><\/pre>\n\n<p>From the result we find out that labels and preds are not equal at index 1 (start from index 0). However, the  gives us . It is as same as index 2 and index 3.<\/p>\n\n<p>Then I test  on some other cases and the result is correct.<\/p>\n\n<p>So, how could it happen?<\/p>\n\n<p>(I used a file_queue to read the images and apply  or  to create the batch for training and testing. Just like the example of cifar 10 in tensorflow examples)<\/p>\n","answers":[{"AnswerId":"41742232","CreationDate":"2017-01-19T12:51:29.483","ParentId":null,"OwnerUserId":"5963300","Title":null,"Body":"<p>After the thinking, there is one possibility that each time I call <code>sess.run()<\/code> the <code>file_queue<\/code> would read shuffled images and labels so that the result is weired.<\/p>\n"}]}
{"QuestionId":41742341,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-19T12:56:52.430","AcceptedAnswerId":41789251.0,"OwnerUserId":7091875.0,"Title":"Tensorflow crashes on sess.run() with no logs in python console","Body":"<p>I tried to implement style transfer using vgg-19 model. I loaded weights through existing keras model of vgg-19 using tensorflow as backend. First I tried it on my notebook and it ran fine though very slow on tf 0.11.0. Then I switched to my AWS instance with gpus, which has version 0.12.1, and tried the very same script: it produces no errors in jupyter notebook, but instead freezes the kernel and forces it to restart when I try to compute an output of a specific layer, except the input of the first layer. I installed gpu version of 0.11.0 - the same problem, though both versions passed test of matrix multiplication.<\/p>\n\n<p>I've also tried running on 0.12.1 someone's script using vgg-16 written for 0.11.0 and session crashed because of error in layer dimensions, though it ran fine on 0.11.0 on cpu. GPU version of tf 0.11.0 - same freeze and restart.<\/p>\n\n<p>Are there any logs written by tensorflow except for those in python console? I have not managed to find any mentions about them<\/p>\n","answers":[{"AnswerId":"41789251","CreationDate":"2017-01-22T09:22:33.093","ParentId":null,"OwnerUserId":"7091875","Title":null,"Body":"<p>Got it fixed. The cudnn version mismatched the tensorflow version - it gave the error output only to python console, not to the jupyter notebook<\/p>\n"}]}
{"QuestionId":41742697,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-01-19T13:13:52.430","AcceptedAnswerId":null,"OwnerUserId":4931135.0,"Title":"Is there any advanced optimization methods in keras, if not then how can i use them?","Body":"<p>I know what the optimizers are in keras BUT I am talking about methods like stochastic average gradient (SAG) or Stochastic Variance-Reduced Gradient (SVRG)?\ndid anyone use it before in his projects? How?<\/p>\n","answers":[{"AnswerId":"41745768","CreationDate":"2017-01-19T15:34:11.053","ParentId":null,"OwnerUserId":"3111778","Title":null,"Body":"<p>I just tried RMSprop over SGD recently on a project. For my application, RMSprop trains the network much faster than SGD. Switching optimizers in Keras is really simple, for example switching from SGD to RMSprop<\/p>\n\n<pre><code>rmsprop = RMSprop(lr=0.0001)\n\nmodel.compile(loss='binary_crossentropy',\n              optimizer=rmsprop,\n              #optimizer = sgd,\n              metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>As pointed by Sergii in the comment, there is a list of optimizers from which you can choose.<\/p>\n"}]}
{"QuestionId":41743093,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-01-19T13:35:04.610","AcceptedAnswerId":41745533.0,"OwnerUserId":6938362.0,"Title":"Tensorflow fails to evaluate gradients","Body":"<p>I have a problem, when trying to use tensorflow's  function together with the gradient optimizer (tensorflow version 0.12.1 on Windows x64 with Anaconda 3.5).<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I get the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>The operation seem to work fine, when using only variables:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Producing the correct output:<\/p>\n\n<pre><\/pre>\n\n<p>Being a beginner in tensorflow, I would highly appreciate your help, what is going wrong with the gradient evaluation of the defined operation?<\/p>\n\n<p>Best Bastian<\/p>\n","answers":[{"AnswerId":"41745533","CreationDate":"2017-01-19T15:24:39.067","ParentId":null,"OwnerUserId":"6938362","Title":null,"Body":"<p>I managed to solve the problem. I forgot to add a parameter to optimize in the example.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\n# Test data\nx_dat = np.arange(1, 13).reshape(4, 3).astype(np.float32)\ny_dat = np.arange(1, 9).reshape(4, 2).astype(np.float32)\n\n# Test data\nx = tf.placeholder(tf.float32, [None, 3])\ny_ = tf.placeholder(tf.float32, [None, 2])\n\nparameter = tf.Variable(tf.ones([3, 1], dtype=tf.float32))\n\n# Model: Take the column product of the elements of x using the rows of map_x_on_y as a filter. \n# Map for projecting x on y\nmap_x_on_y = tf.Variable(np.array([[True, False], [False, False], [ True, True]]))\n\nintermediate = tf.map_fn(lambda t: tf.div(t[0], tf.add(t[1], t[0])),\n                       (tf.transpose(x), parameter), dtype=tf.float32)\n\nfn = lambda t: tf.reduce_prod(tf.boolean_mask(intermediate, t), axis=0)\ny = tf.transpose(tf.stack([fn(i) for i in tf.unstack(map_x_on_y, axis=1)]))\n\n\n# Train the model\nsum_of_squared_errors = tf.reduce_mean(tf.square(y-y_)) \ntrain_step = tf.train.GradientDescentOptimizer(0.1).minimize(sum_of_squared_errors)\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:   \n    sess.run(init)\n    feed_dict ={x: x_dat, y_: y_dat} \n    for i in range(1):   \n        sess.run([train_step], feed_dict)       \n        print(sess.run(intermediate, feed_dict))\n        print(sess.run(y, feed_dict))\n<\/code><\/pre>\n\n<p>This produced the correct output:<\/p>\n\n<pre><code>[[ 0.50838345  0.80531198  0.87862223  0.91182482]\n [ 0.66666669  0.83333331  0.8888889   0.91666669]\n [ 0.76236677  0.86516243  0.90587789  0.92770731]]\n[[ 0.38757464  0.76236677]\n [ 0.69672567  0.86516243]\n [ 0.79592443  0.90587789]\n [ 0.84590656  0.92770731]]\n<\/code><\/pre>\n\n<p>Though, I get still the following warning:<\/p>\n\n<pre><code>C:\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\ops\\gradients_impl.py:91: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.\n  \"Converting sparse IndexedSlices to a dense Tensor of unknown shape. \"\n<\/code><\/pre>\n\n<p>Is this warning of concern?<\/p>\n"}]}
{"QuestionId":41744717,"AnswerCount":0,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-19T14:48:13.770","AcceptedAnswerId":null,"OwnerUserId":1471972.0,"Title":"Feeding numpy array to Tensorflow gives TypeError","Body":"<p>I am trying to classify digits on MNIST using Tensorflow and Python.<br>\nI am getting an error while feeding numpy arrays to feed_dict in this line -  <\/p>\n\n<pre><\/pre>\n\n<p>The error is TypeError : UnhashableType: 'numpy.ndarray'  <\/p>\n\n<p>I confirmed the shapes of batch_xs and batch_ys and they are (100, 784) and (100, 10), 100 is my batch-size.<br>\nAlso, same code is running when I tried to run the example given in Tensorflow tutorials, and there batch_xs and batch_ys are 'numpy.ndarray'.  <\/p>\n\n<p>I would really appreciate some help here. Thanks.<\/p>\n","answers":[]}
{"QuestionId":41745022,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-19T15:02:09.680","AcceptedAnswerId":null,"OwnerUserId":1459980.0,"Title":"Tensorflow Inception v3 retraining - attach text\/labels to individual images","Body":"<p>I am using the inception v3 model to retrain my own dataset. I have few folder which represent the classes which contain images for each class. What i would like to do is to 'attach' some text ids to these images so when they are retrained and used to run classification\/similarity-detection those ids are retrieved too. (basically its image similarity detection)<\/p>\n\n<p>For instance, Image X is of class 'Teachers' and it belongs to John. When i retrain the model, and run a classification on the new model, i would like to get the Teachers class, but in addition to this i would like to know who is teacher (John).<\/p>\n\n<p>Any ideas how to go for it?<\/p>\n\n<p>Regards<\/p>\n","answers":[{"AnswerId":"41792826","CreationDate":"2017-01-22T15:35:26.110","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>You have 3 main options - multiply your classes, multi-label learning or training several models.<\/p>\n\n<p>The first option is the most straight forward - instead of having  teachers who belong to John and teachers who belong to Jane you can have teachers whose class is Teachers_John and class whose class is Teachers_John and learn to classify to those categories as you would learn any other set of categories, or use something like hierarchical softmax.<\/p>\n\n<p>The second option is to have a set of categories that includes Teachers as well as John and Jane - now your target is not to correctly predict the one most accurate class (Teachers) but several (Teachers and John).<\/p>\n\n<p>Your last option is to create a hierarchy of models where the first learns to differentiate between John and Jane and others to classify the inner classes for each of them.<\/p>\n"}]}
{"QuestionId":41745496,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-19T15:23:03.910","AcceptedAnswerId":41748737.0,"OwnerUserId":850760.0,"Title":"loops without unrolling in tensorflow","Body":"<p>I am trying to understand the concept of tensorflow's computational graph. More specifically, how to implement a loop. Imagine I wish to multiply a constant by 2 for 4 times:<\/p>\n\n<pre><\/pre>\n\n<p>The usual way to do it would be to unroll the graph four times and create an operation and evaluate it for at the end.<\/p>\n\n<pre><\/pre>\n\n<p>which returns the correct answer of 160. but if I check the operations, I see that  operation was created four times, which is a waste of operations.\nTo avoid the creation of these \"redundant\" operations, the only way that comes to my mind is to replace the previous block with the following:<\/p>\n\n<pre><\/pre>\n\n<p>this also gives the correct answer and creates the  operation only once, however, it is highly inefficient because I have to evaluate  four times even though I only care for the final answer. This has happened because I cannot feed in a tensor to the feed_dict.<\/p>\n\n<p>Now my question is if there is a more efficient way to implement this recurrent loop without extra evaluations? should I care about creating these extra operations?<\/p>\n","answers":[{"AnswerId":"41747230","CreationDate":"2017-01-19T16:45:18.677","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>I imagine this is a simpler example than what you actually intend so I'll answer for both this example and for the more general case.<\/p>\n\n<p>for this case one can do:<\/p>\n\n<pre><code>s=tf.placeholder_with_default(10.0,[])\nexponent=tf.placeholder_with_default(4,[])\ndef tf_mul(op, exponent):\n    return tf.pow(op * 2, exponent)\n<\/code><\/pre>\n\n<p>In general to not have to use for loops one should work with vector operations.<\/p>\n\n<p>Simple example adding elements in a list:<\/p>\n\n<pre><code>sum = 0\nfor i in list:\n    sum += i\n<\/code><\/pre>\n\n<p>Vectorized:<\/p>\n\n<pre><code>ones = tf.ones(shape=tf.get_shape(list))\nsum = ones * list # the dot product\n<\/code><\/pre>\n"},{"AnswerId":"41748737","CreationDate":"2017-01-19T18:07:00.283","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>TensorFlow provides various <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/\" rel=\"nofollow noreferrer\">constructs<\/a> for control flow, including conditionals and looping. In your case, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#while_loop\" rel=\"nofollow noreferrer\"><code>tf.while_loop<\/code><\/a> is the most relevant construct to use. Here is a complete working example :<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\ni = tf.constant(0)\nx = tf.constant(10.0)\ni, x = tf.while_loop(lambda i, _: i &lt; 4, lambda i, x: (i + 1, x * 2.0), [i, x])\nsess.run([i, x])\n[4, 160.0]\n<\/code><\/pre>\n"}]}
{"QuestionId":41745755,"AnswerCount":1,"Tags":"<macos><tensorflow><bazel><syntaxnet>","CreationDate":"2017-01-19T15:33:47.837","AcceptedAnswerId":null,"OwnerUserId":1975005.0,"Title":"compiling SyntaxNet on OS X fails with \"error loading package '@jpeg\/\/': Extension file not found\"","Body":"<p>I can't get SyntaxNet compiled on my MacBook Pro.<\/p>\n\n<p>I followed the SyntaxNet installation steps here: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/syntaxnet#installation\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/syntaxnet#installation<\/a><\/p>\n\n<p>the  line fails with this output:<\/p>\n\n<p><\/p>\n\n<p>Here's some system info:<\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>I pulled the sources as follows:<\/p>\n\n<p><\/p>\n","answers":[{"AnswerId":"41892841","CreationDate":"2017-01-27T11:22:23.760","ParentId":null,"OwnerUserId":"7478105","Title":null,"Body":"<p>I encountered a similar problem when compiling on Ubuntu.<\/p>\n\n<p>Find the file <code>tensorflow\/third_party\/jpeg\/jpeg.BUILD<\/code><\/p>\n\n<p>Then, change the code <code>load(\"@\/\/third_party:common.bzl\", \"template_rule\")<\/code> <\/p>\n\n<p>to   <\/p>\n\n<p><code>load(\"@org_tensorflow\/\/third_party:common.bzl\", \"template_rule\")<\/code><\/p>\n"}]}
{"QuestionId":41746576,"AnswerCount":1,"Tags":"<python><ubuntu><tensorflow>","CreationDate":"2017-01-19T16:14:01.113","AcceptedAnswerId":41748641.0,"OwnerUserId":7436888.0,"Title":"Tensorflow with CUDA: ImportError","Body":"<p>I've installed TensorFlow step by step like described in this tutorial from NVIDIA (Ubuntu 16.04 Desktop, GTX 970):<\/p>\n\n<p><a href=\"http:\/\/www.nvidia.com\/object\/gpu-accelerated-applications-tensorflow-installation.html\" rel=\"nofollow noreferrer\">http:\/\/www.nvidia.com\/object\/gpu-accelerated-applications-tensorflow-installation.html<\/a><\/p>\n\n<p>Every steps completes without errors, but if I try to import TensorFlow in Python afterwards I'm always gettin' this error message:<\/p>\n\n<pre><\/pre>\n\n<p>I've installed the CUDA-Toolkit to \/opt\/cuda\/toolkit (also cuDNN).\nI already tried to symlink:<\/p>\n\n<pre><\/pre>\n\n<p>Or this:<\/p>\n\n<pre><\/pre>\n\n<p>But the error remains. Can anybody give me a hint why this error occurs?\nThanks a lot :)<\/p>\n","answers":[{"AnswerId":"41748641","CreationDate":"2017-01-19T18:01:20.130","ParentId":null,"OwnerUserId":"7442510","Title":null,"Body":"<p>I had a similar problem on a machine with similar hardware (GTX 970M) and I solved it using the steps provided in the following post:\n<a href=\"https:\/\/askubuntu.com\/a\/278840\">https:\/\/askubuntu.com\/a\/278840<\/a><\/p>\n\n<p>The only difference in my case was that in the nvidia.conf file I added only the following line:\n\/usr\/local\/cuda-8.0\/lib64<\/p>\n\n<p>and then ran: sudo ldconfig<\/p>\n\n<p>Hope it helps :)<\/p>\n"}]}
{"QuestionId":41747417,"AnswerCount":1,"Tags":"<lua><neural-network><gnuplot><torch>","CreationDate":"2017-01-19T16:54:24.020","AcceptedAnswerId":null,"OwnerUserId":4492375.0,"Title":"real-time plotting in Torch","Body":"<p>I'm new to torch. I'm training a neural network model where I want to make a plot for training and validation error with each iteration, a quick search shows me plotting with gnuplot, <a href=\"http:\/\/torch.ch\/docs\/five-simple-examples.html\" rel=\"nofollow noreferrer\">http:\/\/torch.ch\/docs\/five-simple-examples.html<\/a> but it's not done in real-time rather it plots the whole thing when all data from each time step\/iteration is obtained. <\/p>\n\n<p>I have made real-time plotting in python, something similar to this <a href=\"https:\/\/stackoverflow.com\/questions\/11874767\/real-time-plotting-in-while-loop-with-matplotlib\">real-time plotting in while loop with matplotlib<\/a><\/p>\n\n<p>Anyway I can do it in Torch? <\/p>\n","answers":[{"AnswerId":"44383997","CreationDate":"2017-06-06T07:35:39.010","ParentId":null,"OwnerUserId":"1550099","Title":null,"Body":"<p>You could just move the plotting code into the training loop. Then you can either rewrite the same plot on every loop or create a new plot per X iteration. Alternatively you can look at <a href=\"https:\/\/github.com\/facebookresearch\/visdom\" rel=\"nofollow noreferrer\">https:\/\/github.com\/facebookresearch\/visdom<\/a><\/p>\n"}]}
{"QuestionId":41748396,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-19T17:47:32.333","AcceptedAnswerId":null,"OwnerUserId":7442475.0,"Title":"TensorFlow neural net outputs linear function","Body":"<p>I implemented a basic MLP and I want it to predict a user-generated set of data, but the prediction looks as follows: <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/e0Sif.png\" alt=\"prediction\">.<\/p>\n\n<p>I am not sure why... I have nonlinearities in the hidden layers, and I tried multiple activations (, , ), tried different optimisers, different learning rates, various architectures (more layers, fewer layers, dropout), but I never got this right.<\/p>\n\n<p>Please note that I do believe it may be because of how I compute the predictions at the end () as it may be incorrect, but I wouldn't know why. I also tried other methods like extracting the weights with  and then feeding them to the  function along with the input, but nothing worked.<\/p>\n\n<p>Also, when monitoring the error, the error decreases between epochs.<\/p>\n\n<p>Any ideas?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41748731","CreationDate":"2017-01-19T18:06:46.623","ParentId":null,"OwnerUserId":"7442475","Title":null,"Body":"<p>Forgot bias terms: <a href=\"https:\/\/i.stack.imgur.com\/uMC9o.png\" rel=\"nofollow noreferrer\">new graph<\/a> <\/p>\n\n<p>It works now but not sure if this fixed it?<\/p>\n\n<p>New code uses: <\/p>\n\n<pre><code>weights = {\n    'h1': tf.Variable(tf.random_normal(shape=(input_size, h1_size))),\n    'h2': tf.Variable(tf.random_normal(shape=(h1_size, h2_size))),\n    'h3': tf.Variable(tf.random_normal(shape=(h2_size, output_size))),\n\n    'b1': tf.Variable(tf.zeros(shape=[1])),\n    'b2': tf.Variable(tf.zeros(shape=[1])),\n    'b3': tf.Variable(tf.zeros(shape=[1])),\n}\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>def model(inputs, weights):\n    out1 = tf.nn.relu(tf.matmul(inputs, weights['h1']) + weights['b1'])\n    out2 = tf.nn.relu(tf.matmul(out1, weights['h2']) + weights['b2'])\n    return tf.matmul(out2, weights['h3'] + weights['b3'])\n<\/code><\/pre>\n"}]}
{"QuestionId":41749398,"AnswerCount":4,"Tags":"<machine-learning><neural-network><regression><keras>","CreationDate":"2017-01-19T18:44:52.947","AcceptedAnswerId":41774345.0,"OwnerUserId":1934212.0,"Title":"Using Keras ImageDataGenerator in a regression model","Body":"<p>I want to use the <\/p>\n\n<pre><\/pre>\n\n<p>method of the <\/p>\n\n<pre><\/pre>\n\n<p>to generate training data for a regression model, where the target value can be any float value between 1 and -1. <\/p>\n\n<pre><\/pre>\n\n<p>has a \"class_mode\" parameter with the descripton<\/p>\n\n<blockquote>\n  <p>class_mode: one of \"categorical\", \"binary\", \"sparse\" or None. Default:\n  \"categorical\". Determines the type of label arrays that are returned:\n  \"categorical\" will be 2D one-hot encoded labels, \"binary\" will be 1D\n  binary labels, \"sparse\" will be 1D integer labels.<\/p>\n<\/blockquote>\n\n<p>Which of these values should I take? None of them seems to really fit...<\/p>\n","answers":[{"AnswerId":"45989013","CreationDate":"2017-08-31T19:45:03.277","ParentId":null,"OwnerUserId":"1638996","Title":null,"Body":"<p>I think that organizing your data differently, using a DataFrame (without necessarily moving your images to new locations) will allow you to run a regression model.  In short, create columns in your DataFrame containing the file path of each image and the target value.  This allows your generator to keep regression values and images properly synced even when you shuffle your data at each epoch.<\/p>\n\n<p>Here is an example showing how to link images with binomial targets, multinomial targets and regression targets just to show that \"a target is a target is a target\" and only the model might change:<\/p>\n\n<pre><code>df['path'] = df.object_id.apply(file_path_from_db_id)\ndf\n\n       object_id   bi  multi                                    path     target\nindex                                                               \n0         461756  dog  white    \/path\/to\/imgs\/756\/61\/blah_461756.png   0.166831\n1        1161756  cat  black   \/path\/to\/imgs\/756\/61\/blah_1161756.png   0.058793\n2        3303651  dog  white   \/path\/to\/imgs\/651\/03\/blah_3303651.png   0.582970\n3        3367756  dog   grey   \/path\/to\/imgs\/756\/67\/blah_3367756.png  -0.421429\n4        3767756  dog   grey   \/path\/to\/imgs\/756\/67\/blah_3767756.png  -0.706608\n5        5467756  cat  black   \/path\/to\/imgs\/756\/67\/blah_5467756.png  -0.415115\n6        5561756  dog  white   \/path\/to\/imgs\/756\/61\/blah_5561756.png  -0.631041\n7       31255756  cat   grey  \/path\/to\/imgs\/756\/55\/blah_31255756.png  -0.148226\n8       35903651  cat  black  \/path\/to\/imgs\/651\/03\/blah_35903651.png  -0.785671\n9       44603651  dog  black  \/path\/to\/imgs\/651\/03\/blah_44603651.png  -0.538359\n10      49557622  cat  black  \/path\/to\/imgs\/622\/57\/blah_49557622.png  -0.295279\n11      58164756  dog   grey  \/path\/to\/imgs\/756\/64\/blah_58164756.png   0.407096\n12      95403651  cat  white  \/path\/to\/imgs\/651\/03\/blah_95403651.png   0.790274\n13      95555756  dog   grey  \/path\/to\/imgs\/756\/55\/blah_95555756.png   0.060669\n<\/code><\/pre>\n\n<p>I describe how to do this in great detail with examples here:<\/p>\n\n<p><a href=\"https:\/\/techblog.appnexus.com\/a-keras-multithreaded-dataframe-generator-for-millions-of-image-files-84d3027f6f43\" rel=\"nofollow noreferrer\">https:\/\/techblog.appnexus.com\/a-keras-multithreaded-dataframe-generator-for-millions-of-image-files-84d3027f6f43<\/a><\/p>\n"},{"AnswerId":"41774345","CreationDate":"2017-01-21T00:11:44.093","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>At this moment (newest version of Keras from January 21st 2017) the <code>flow_from_directory<\/code> could only work in a following manner:<\/p>\n\n<ol>\n<li><p>You need to have a directories structured in a following manner:<\/p>\n\n<pre><code>directory with images\\\n    1st label\\\n        1st picture from 1st label\n        2nd picture from 1st label\n        3rd picture from 1st label\n        ...\n    2nd label\\\n        1st picture from 2nd label\n        2nd picture from 2nd label\n        3rd picture from 2nd label\n        ...\n    ...\n<\/code><\/pre><\/li>\n<li><code>flow_from_directory<\/code> returns batches of a fixed size in a format of <code>(picture, label)<\/code>.<\/li>\n<\/ol>\n\n<p>So as you can see it could only be used for a classification case and all options provided in a documentation specify only a way in which the class is provided to your classifier. But, there is a neat hack which could make a <code>flow_from_directory<\/code> useful for a regression task:<\/p>\n\n<ol>\n<li><p>You need to structure your directory in a following manner: <\/p>\n\n<pre><code>directory with images\\\n    1st value (e.g. -0.95423)\\\n        1st picture from 1st value\n        2nd picture from 1st value\n        3rd picture from 1st value\n        ...\n    2nd value (e.g. - 0.9143242)\\\n        1st picture from 2nd value\n        2nd picture from 2nd value\n        3rd picture from 2nd value\n        ...\n   ...\n<\/code><\/pre><\/li>\n<li><p>You also need to have a list <code>list_of_values = [1st value, 2nd value, ...]<\/code>. Then your generator is defined in a following manner:<\/p>\n\n<pre><code>def regression_flow_from_directory(flow_from_directory_gen, list_of_values):\n    for x, y in flow_from_directory_gen:\n        yield x, list_of_values[y]\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>And it's crucial for a <code>flow_from_directory_gen<\/code> to have a <code>class_mode='sparse'<\/code> to make this work. Of course this is a little bit cumbersome but it works (I used this solution :) )<\/p>\n"},{"AnswerId":"53057526","CreationDate":"2018-10-30T04:38:20.027","ParentId":null,"OwnerUserId":"4082092","Title":null,"Body":"<p>There's just one glitch in the accepted answer that I would like to point out. The above code fails with an error message like:<\/p>\n\n<pre><code>TypeError: only integer scalar arrays can be converted to a scalar index\n<\/code><\/pre>\n\n<p>This is because y is an array. The fix is simple:<\/p>\n\n<pre><code>def regression_flow_from_directory(flow_from_directory_gen,\n            list_of_values):\n    for x, y in flow_from_directory_gen:\n        values = [list_of_values[y[i]] for i in range(len(y))]\n        yield x, values\n<\/code><\/pre>\n\n<p>The method to generate the list_of_values can be found in <a href=\"https:\/\/stackoverflow.com\/a\/47944082\/4082092\">https:\/\/stackoverflow.com\/a\/47944082\/4082092<\/a><\/p>\n"},{"AnswerId":"53161302","CreationDate":"2018-11-05T19:56:00.413","ParentId":null,"OwnerUserId":"6114916","Title":null,"Body":"<p>With Keras 2.2.4 you can use \".flow_from_dataframe\" that solves what you want to do, allowing you to flow images from a directory for regression problems. You should store all your images in a folder and load a dataframe containing in one column the image IDs and in the other column the regression score (labels) and set \"class_mode='other'\" in \".flow_from_dataframe\". <\/p>\n\n<p>Here you can find an example where the images are in \"image_dir\", the dataframe with the image IDs and the regression scores is loaded with pandas from \"train file\" <\/p>\n\n<pre><code>train_label_df = pd.read_csv(train_file, delimiter=' ', header=None, names=['id', 'score'])\n\ntrain_datagen = ImageDataGenerator(rescale = 1.\/255, horizontal_flip = True,\n                                   fill_mode = \"nearest\", zoom_range = 0.2,\n                                   width_shift_range = 0.2, height_shift_range=0.2,\n                                   rotation_range=30) \n\ntrain_generator = train_datagen.flow_from_dataframe(dataframe=train_label_df, directory=image_dir, \n                                              x_col=\"id\", y_col=\"score\", has_ext=True, \n                                              class_mode=\"other\", target_size=(img_width, img_height), \n                                              batch_size=bs)\n<\/code><\/pre>\n"}]}
{"QuestionId":41749674,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-19T19:02:34.750","AcceptedAnswerId":41770159.0,"OwnerUserId":2191652.0,"Title":"How to multiply input images with mask in tensorflow?","Body":"<p>I want to multiply every input image with a mask of the same size as the input image. How would I do that in tensorflow?  <\/p>\n\n<p>My image reading function looks like this so far:  <\/p>\n\n<pre><\/pre>\n\n<p>Is that possible?\nNot sure if img is already a tensor object and I can use that function here. I'm new to tensorflow...<\/p>\n","answers":[{"AnswerId":"41770159","CreationDate":"2017-01-20T18:27:22.307","ParentId":null,"OwnerUserId":"6912318","Title":null,"Body":"<p>Here is the code that works well for me. I'm using jupyter notebook to run the code.\n<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>%matplotlib inline\nimport tensorflow as tf\nfrom matplotlib.image import imread \nimport matplotlib.pyplot as plt\n\n# Loading test image from the local filesystem \nx = tf.Variable(imread(\"test_img.jpg\"),dtype='float32')\nx_mask = tf.Variable(imread(\"test_mask.jpg\"),dtype='float32')\nimg_mult = tf.multiply(x,x_mask)\n\nplt.imshow(imread(\"test_img.jpg\"))\nplt.show()\nplt.imshow(imread(\"test_mask.jpg\"))\nplt.show()\n\nsess = tf.Session() \nsess.run(tf.global_variables_initializer())\nres = sess.run(img_mult)\n\nplt.imshow(res)\nplt.show()\n<\/code><\/pre>\n\n<p>Also, Here is a good YouTube tutorial covering image manipulation with TF: <a href=\"https:\/\/www.youtube.com\/watch?v=bvHgESVuS6Q&amp;t=447s\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=bvHgESVuS6Q&amp;t=447s<\/a><\/p>\n"}]}
{"QuestionId":41750088,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-19T19:26:55.050","AcceptedAnswerId":41814013.0,"OwnerUserId":2191652.0,"Title":"Tensorflow: Finetune pretrained model on new dataset with different number of classes","Body":"<p>How can I finetune a pretrained model in tensorflow on a new dataset? In Caffe I can simply rename the last layer and set some parameters for random initialization. Is something similar possible in tensorflow?<\/p>\n\n<p>Say I have a checkpoint file () and some code that sets up the computational graph in which I can modify the last layer such that it has the same number of ouputs as the new dataset has classes. <\/p>\n\n<p>Then I try to start the session like this:  <\/p>\n\n<pre><\/pre>\n\n<p>However this gives me an error when calling the  function since it expects the exact same graph structure as the the one it was saved from. \nHow can I only load all weights except for the last layer from the  file?\nI also tried changing the  layer name but no luck there either...<\/p>\n\n<p>I'm using the <a href=\"https:\/\/github.com\/DrSleep\/tensorflow-deeplab-resnet\">tensorflow-deeplab-resnet model<\/a><\/p>\n","answers":[{"AnswerId":"41814013","CreationDate":"2017-01-23T19:21:32.287","ParentId":null,"OwnerUserId":"2708478","Title":null,"Body":"<p>You can specify the names of the variables that you want to restore. <\/p>\n\n<p>So, you can get a list of all of the variables in the model and filter out the variables of the last layer:<\/p>\n\n<pre><code>all_vars = tf.all_variables()\nvar_to_restore = [v for v in all_vars if not v.name.startswith('xxx')]\n\nsaver = tf.train.Saver(var_to_restore)\n<\/code><\/pre>\n\n<p>See the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variables\/#choosing_which_variables_to_save_and_restore\" rel=\"nofollow noreferrer\">documentation<\/a> for the details.<\/p>\n\n<p>Alternatively, you can try to load the whole model an create a new \"branch\" out of the layer before the last and use it in the cost function during the training.<\/p>\n"}]}
{"QuestionId":41751100,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-19T20:28:31.743","AcceptedAnswerId":null,"OwnerUserId":7383444.0,"Title":"FailedPreconditionError tensorflow","Body":"<p>Why is this simple code giving error?<\/p>\n\n<pre><\/pre>\n\n<p>It gives me :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41751363","CreationDate":"2017-01-19T20:45:13.307","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p><code>tf.global_variables_initializer()<\/code> returns an <strong>operation<\/strong> that you must execute. This operation, indeed, when initializes the global variables.<\/p>\n\n<p>Therefore, change that line with:<\/p>\n\n<pre><code>init_op = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p>and within the session, execute it.<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(init_op)\n<\/code><\/pre>\n\n<p>Moreover, it's logically better to do not mix the graph definition and the graph execution.<\/p>\n\n<p>Define the graph outside the session, then execute the operations. Here's an improved version of your code.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ninputs = tf.Variable(np.random.rand(2,2))\ninit_op =  tf.global_variables_initializer()     \n\nmean_op = tf.reduce_mean(inputs)\n\nwith tf.Session() as sess:\n    sess.run(init_op)\n    mean_value = sess.run(mean_op)\n    print(mean_value)\n<\/code><\/pre>\n\n<p>By the way, I suggest you reading the tensorflow getting started page: <a href=\"https:\/\/www.tensorflow.org\/get_started\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/<\/a><\/p>\n"}]}
{"QuestionId":41754247,"AnswerCount":2,"Tags":"<save><deep-learning><keras>","CreationDate":"2017-01-20T00:41:01.520","AcceptedAnswerId":41755194.0,"OwnerUserId":5399559.0,"Title":"Keras save model issue","Body":"<p>This is a variational autoencoder network, I have to define a sampling method to generate latent z, I thinks it might be something wrong with this. This py file is doing training, the other py file is doing predicting online, so I need to save the keras model, there is nothing wrong with saving model, but when I load model from 'h5' file, it shows an error:<\/p>\n\n<pre><\/pre>\n\n<p>The following is code:<\/p>\n\n<pre><\/pre>\n\n<p>It shows error:\n<\/p>\n","answers":[{"AnswerId":"54978662","CreationDate":"2019-03-04T07:44:16.493","ParentId":null,"OwnerUserId":"11028280","Title":null,"Body":"<p>If you load model (.h5) file in your new py file, you can use load_model('\/.h5', compile = False).\nBecause you do not need to any custom objects (i.e loss function or latent_dim, etc) in prediction step.<\/p>\n"},{"AnswerId":"41755194","CreationDate":"2017-01-20T02:38:24.433","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>For variational loss you are using many variable not known by Keras module. You need to pass them through <code>custom_objects<\/code> param of <code>load_model<\/code> function. <\/p>\n\n<p>In your case:<\/p>\n\n<pre><code>vae.save('.\/vae_'+str(cluster_num)+'.h5')\nvae.summary()\n\ndel vae\n\nfrom keras.models import load_model\nvae = load_model('.\/vae_'+str(cluster_num)+'.h5', custom_objects={'latent_dim': latent_dim, 'epsilon_std': epsilon_std, 'vae_loss': vae_loss})\nvae.summary()\n<\/code><\/pre>\n"}]}
{"QuestionId":41754649,"AnswerCount":1,"Tags":"<tensorflow><neural-network><keras>","CreationDate":"2017-01-20T01:31:38.260","AcceptedAnswerId":41755667.0,"OwnerUserId":762819.0,"Title":"Keras calculate accuracy as +\/- 1 from predicted","Body":"<p>I am using Keras (+TensorFlow) to build a deep neural network model. In the model, I need to define my own accuracy function.<\/p>\n\n<p>Let's say, the model predicts the time taken to do a work (in minutes, between 0 and 20). I want the model to print out accuracy if the predicted output is within +\/- 2. If the predicted output is x minutes, while the expected output is x+1, I want to consider this is a correct prediction, if the expected output is x+3, I want to consider this is a wrong prediction.<\/p>\n\n<p>This is slightly different from <\/p>\n","answers":[{"AnswerId":"41755667","CreationDate":"2017-01-20T03:35:42.673","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can easily implement the logic using Keras backend apis .. which will also ensure your metric working both on tensorflow and theano.<\/p>\n\n<p>Here with test:<\/p>\n\n<pre><code>import numpy as np\nimport keras\nfrom keras import backend as K\n\nshift = 2\ndef custom_metric(y_true,y_pred):\n    diff = K.abs(K.argmax(y_true, axis=-1) - K.argmax(y_pred, axis=-1))\n    return K.mean(K.lesser_equal(diff, shift))\n\n\nt1 = np.asarray([   [0,0,0,0,0,0,1,0,0,],\n                    [0,0,0,0,0,0,1,0,0,],\n                    [0,0,0,0,0,0,1,0,0,],\n                    [0,0,0,0,0,0,1,0,0,],\n                    [0,0,0,0,0,0,1,0,0,],\n                    [0,0,0,0,0,0,1,0,0,],\n                ])\np1 = np.asarray([   [0,0,0,0,0,1,0,0,0,],\n                    [0,0,0,0,1,0,0,0,0,],\n                    [0,0,0,0,0,0,0,1,0,],\n                    [0,0,0,0,0,0,0,0,1,],\n                    [1,0,0,0,0,0,0,0,0,],\n                    [0,0,0,0,0,0,1,0,0,],\n                ])\n\n\nprint K.eval(keras.metrics.categorical_accuracy(K.variable(t1),K.variable(p1)))\nprint K.eval(custom_metric(K.variable(t1),K.variable(p1)))\n<\/code><\/pre>\n\n<p>now in your <code>compile<\/code> statement use it: <code>metrics=custom_metric<\/code><\/p>\n"}]}
{"QuestionId":41754901,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-20T02:00:07.773","AcceptedAnswerId":41766555.0,"OwnerUserId":1214630.0,"Title":"Image over Image convolution in Tensorflow","Body":"<p>Assume, I have two set of images, A and B, each 11X5x5x3, where 11 is a number of examples and 5x5x3 is an image dimension.<br>\nIs there an easy way in Tensorflow to apply convolution for each image in A_i over B_i (i.e. B_i plays a filter role and A_i is an input in tf.conv2d)? For example, conv2d(A_1,B_1), conv2d(A_2,B_2),...,conv2d(A_11,B_11)<br>\nNo weight learning here just wanted to apply convolution on one image over another. \nI tried to do it as follow:<\/p>\n\n<pre><\/pre>\n\n<p>but the problem with this is that it applies convolution on every A_i over all B_i's. I don't want this, I want A_i only over B_j where i==j. Of course I can do it one by one, but it wouldn't be efficient and I need to do it in a batch mode.<\/p>\n\n<p>Any comment how to solve this problem?<\/p>\n\n<p>Thanks.\nJ<\/p>\n","answers":[{"AnswerId":"41766555","CreationDate":"2017-01-20T15:00:57.297","ParentId":null,"OwnerUserId":"2911687","Title":null,"Body":"<p>I am not sure that is what you need because it is not really batch mode but you could use a map function :<\/p>\n\n<pre><code>A = tf.placeholder(dtype=tf.float32, shape=[None, 5, 5, 3])\nB = tf.placeholder(dtype=tf.float32, shape=[None, 5, 5, 3])\n\noutput = tf.map_fn(\n    lambda inputs : tf.nn.conv2d(\n        tf.expand_dims(inputs[0], 0),  # H,W,C -&gt; 1,H,W,C\n        tf.expand_dims(inputs[1], 3),  # H,W,C -&gt; H,W,C,1\n        strides=[1,1,1,1],\n        padding=\"SAME\"\n    ),  # Result of conv is 1,H,W,1\n   elems=[A,B],\n   dtype=tf.float32\n)\nfinal_output = output[:, 0, :, :, 0]  # B,1,H,W,1 -&gt; B,H,W\n<\/code><\/pre>\n\n<p>Performance will depend on how the tiny separate convolutions will be parallelized I guess.<\/p>\n"}]}
{"QuestionId":41755383,"AnswerCount":1,"Tags":"<machine-learning><merge><neural-network><theano><keras>","CreationDate":"2017-01-20T03:03:29.430","AcceptedAnswerId":41779369.0,"OwnerUserId":1587730.0,"Title":"Keras weighted merge","Body":"<p>I am trying to compute a weighted output from multiple parallel models using Keras'  layer. I'm using  backend.<\/p>\n\n<p>I have  parallel models . Each of their output layer is a k-sized softmax.<\/p>\n\n<p>There is one model , its output is a L-sized softmax.<\/p>\n\n<p>Here is what I have so far:<\/p>\n\n<p>Parallel models  each with k dimension in the output layer:<\/p>\n\n<pre><\/pre>\n\n<p>The weighing model , output layer:<\/p>\n\n<pre><\/pre>\n\n<p>The merger is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>where \"model_group\" is a (L+1)-length list , and 's signature is:<\/p>\n\n<pre><\/pre>\n\n<p>Mathematically, I would like the output of the merged layer to be a weighted sum:<\/p>\n\n<blockquote>\n  <p><em>out = N[1]x([C11, C12, C13, .., C1k]) + N[2]x([C21, C22, C23, ..., C2k]) + ... + N[L]x([CL1, CL2, CL3, ..., CLk])<\/em>, <\/p>\n<\/blockquote>\n\n<p>where out is a vector of size <\/p>\n\n<p>How can I use the Merge layer to achieve this ? <\/p>\n\n<p>I know that the magic would probably have to happen in the 'merge_fun', but I am not sure how to perform matrix algebra in Keras. The tensor parameters don't have a \"shape\" parameter - they have a  - but I am not sure how to combine parallel models' output into a matrix. <\/p>\n\n<p>I tried using a local evaluation of the following expressions:<\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre>\n\n<p>both of which didn't throw an error, so I can't use this as a guide. After the multiplication, the result returned did not have the  variable, so I'm not sure what the shape of the result is.<\/p>\n\n<p>Any suggestions ?<\/p>\n","answers":[{"AnswerId":"41779369","CreationDate":"2017-01-21T12:30:13.240","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>What I advise you is to use a <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">functional API<\/a> and use this is in a following manner:<\/p>\n\n<ol>\n<li><p>Define the <code>L<\/code> output models:<\/p>\n\n<pre><code>softmax_1 = Dense(K, activation='softmax', ...))(input_to_softmax_1)\nsoftmax_2 = Dense(K, activation='softmax', ...))(input_to_softmax_2)\n...\nsoftmax_L = Dense(K, activation='softmax', ...))(input_to_softmax_L)\n<\/code><\/pre><\/li>\n<li><p>Define the merge softmax:<\/p>\n\n<pre><code>merge_softmax= Dense(L, activation='softmax', ...)(input_to_merge_softmax)\nmerge_softmax = Reshape((1, L))(merge_softmax)\n<\/code><\/pre><\/li>\n<li><p>Merge and reshape the bag of <code>L<\/code> models:<\/p>\n\n<pre><code>bag_of_models = merge([softmax_1, ..., softmax_L], mode = 'concat')\nbag_of_models = Reshape((L, K))(bag_of_models)\n<\/code><\/pre><\/li>\n<li><p>Compute the final merged softmax:<\/p>\n\n<pre><code>final_result = merge([bag_of_models, merge_softmax], mode = 'dot', dot_axes = [1, 2])\nfinal_result = Reshape((K, ))(final_result)\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Of course - depending on your topology - different tensor might be the same (e.g. input to different softmaxes). I tested this on my machine but due to extensive refactoring - I might made mistake - so if you fin one - please inform me. <\/p>\n\n<p>The solution with <code>Sequential<\/code> is much less clear and a little bit cumbersome - but if you want one - please write in the comment so I will update my answer.<\/p>\n"}]}
{"QuestionId":41755633,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2017-01-20T03:31:47.820","AcceptedAnswerId":41755634.0,"OwnerUserId":1093708.0,"Title":"WARNING:tensorflow:The default value of combiner will change from \"sum\" to \"sqrtn\" after 2016\/11\/01","Body":"<p>If you try to follow the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/wide\/\" rel=\"nofollow noreferrer\">TensorFlow Linear Model Tutorial<\/a> depending the version of tensorflow installed is a challenge complete the routine without errors.<\/p>\n\n<p>I receive \"WARNING:tensorflow:The default value of combiner will change from \"sum\" to \"sqrtn\" after 2016\/11\/01\"<\/p>\n\n<p>how is possible solve the warning? <\/p>\n\n<p><strong>tensorflow: 0.12.1<\/strong> <\/p>\n","answers":[{"AnswerId":"41755634","CreationDate":"2017-01-20T03:31:47.820","ParentId":null,"OwnerUserId":"1093708","Title":null,"Body":"<p>you need to pay special attention to the definition of columns:<\/p>\n\n<blockquote>\n  <p>tf.contrib.layers.sparse_column_with_hash_bucket(...)<\/p>\n  \n  <p>tf.contrib.layers.bucketized_column(...)<\/p>\n  \n  <p>tf.contrib.layers.crossed_column(...)<\/p>\n<\/blockquote>\n\n<p>If you use version 0.12.1 the simple way to solve is including a modification of the parameter <strong>combiner<\/strong><\/p>\n\n<blockquote>\n  <p>tf.contrib.layers.sparse_column_with_hash_bucket(..., combiner=\"sqrtn\")<\/p>\n  \n  <p>tf.contrib.layers.bucketized_column(..., combiner=\"sqrtn\")<\/p>\n  \n  <p>tf.contrib.layers.crossed_column(..., combiner=\"sqrtn\")<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41756054,"AnswerCount":1,"Tags":"<python><scope><tensorflow>","CreationDate":"2017-01-20T04:20:50.077","AcceptedAnswerId":null,"OwnerUserId":1001930.0,"Title":"Tensorflow VariableScope: original_name_scope vs name","Body":"<p>In TensorFlow, the  class has both a  and  attribute. What are their differences and when should I use one over the other? I can't seem to find much documentation on them.<\/p>\n\n<p><strong>Use case:<\/strong>\nI'm using the  method. Its second argument expects a string, but my variable  has type . I'm trying both<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>. Both seem to work, but I'm not sure which is \"right\" and won't give me problems later down the road.<\/p>\n","answers":[{"AnswerId":"41801241","CreationDate":"2017-01-23T07:34:46.660","ParentId":null,"OwnerUserId":"1001930","Title":null,"Body":"<p><code>foo.name<\/code> returns the name (String) of the scope. On the other hand, <code>foo.original_name_scope<\/code> returns the same string as <code>foo.name<\/code>, except when the scope is recreated. In that case, all sub-scopes are appended with a <code>_#<\/code> as needed to make all calls to <code>foo.original_name_scope<\/code> return something unique for each instance of a scope.<\/p>\n\n<p>For example, in this code:<\/p>\n\n<pre><code>with tf.variable_scope('a') as a:\n    print(a.name)\n    print(a.original_name_scope)\n    print(a.original_name_scope)\n\nwith tf.variable_scope('a') as b:\n    print(b.name)\n    print(b.original_name_scope)\n<\/code><\/pre>\n\n<p>Returns<\/p>\n\n<pre><code>a\na\/\na\/\na\na_1\/\n<\/code><\/pre>\n\n<p>Note that the calls to <code>original_name_scope<\/code> corresponding to different scope instances <code>a<\/code> return different values.<\/p>\n\n<p>Presumably, this lets you distinguish between different scope instances with the same name.<\/p>\n"}]}
{"QuestionId":41756328,"AnswerCount":1,"Tags":"<tensorflow><keras><kaggle>","CreationDate":"2017-01-20T04:50:53.427","AcceptedAnswerId":41765096.0,"OwnerUserId":4251282.0,"Title":"Why the same neural architecture works in Keras but not Tensorflow ( Leaf Classification )?","Body":"<p>Recently I am playing the leaf classification problem in Kaggle. I have seen a notebook <a href=\"https:\/\/www.kaggle.com\/alexanderlazarev\/leaf-classification\/simple-keras-1d-cnn-features-split\" rel=\"nofollow noreferrer\">Simple Keras 1D CNN + features split<\/a>. But when I tried to construct the same model with Tensorflow, it generate very low accuracy and loss change little. Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>The result is something like:<\/p>\n\n<pre><\/pre>\n\n<p>But when using the same data and model in Keras, it indeed generate very good result.\ncode:<\/p>\n\n<pre><\/pre>\n\n<p>result:<\/p>\n\n<pre><\/pre>\n\n<p>By the way, keras uses Tensorflow backend. Is there any suggestion?<\/p>\n","answers":[{"AnswerId":"41765096","CreationDate":"2017-01-20T13:42:49.863","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>There are multiple differences between both models, your TF model uses ADAM, while your Keras model uses SGD. The learning rates are different as well, and learning rate greatly affects model convergence.<\/p>\n\n<p>The loss functions also don't match, the Keras model uses categorical cross-entropy, while the TF model is using sigmoid cross-entropy with logits (which usually is used for multilabel classification). Also sigmoid_cross_entropy_with_logits takes logits as input (real numbers), and you are giving it the output of a softmax function.<\/p>\n\n<p>There are also differences in weight initialization, you are using truncated normal distribution for the TF model, while Keras by default uses glorot_uniform or uniform.<\/p>\n\n<p>These differences are of course the reason why one model trains, and the other does not.<\/p>\n"}]}
{"QuestionId":41757986,"AnswerCount":0,"Tags":"<image><tensorflow><dataset>","CreationDate":"2017-01-20T07:12:48.087","AcceptedAnswerId":null,"OwnerUserId":2017022.0,"Title":"How to create a CXR Dataset just like MNIST dataset?","Body":"<p>I have jpeg images of chest radiographs. If i want to feed the data to a deep network what do i need to do ? <\/p>\n\n<p>I am using tensorflow library in python . I want to provide labels as normal or abnormal.How can i create a dataset like MNIST ?<\/p>\n","answers":[]}
{"QuestionId":41758385,"AnswerCount":4,"Tags":"<image-processing><keras>","CreationDate":"2017-01-20T07:39:05.060","AcceptedAnswerId":41759831.0,"OwnerUserId":1934212.0,"Title":"Resizing images in Keras ImageDataGenerator flow methods","Body":"<p>The Keras  class provides the two flow methods  and  (<a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"noreferrer\">https:\/\/keras.io\/preprocessing\/image\/<\/a>).<\/p>\n\n<p>Why is the parameter <\/p>\n\n<blockquote>\n  <p>target_size: tuple of integers, default: (256, 256). The dimensions to which all images found will be resized<\/p>\n<\/blockquote>\n\n<p>Only provided by <em>flow_from_directory(directory)<\/em> ? And what is the most concise way to add reshaping of images to the preprocessing pipeline using <em>flow(X, y)<\/em> ?<\/p>\n","answers":[{"AnswerId":"56158707","CreationDate":"2019-05-15T23:03:09.240","ParentId":null,"OwnerUserId":"1651560","Title":null,"Body":"<p>For large training dataset, performing transformations such as resizing on the entire training data is very memory consuming. As Keras did in ImageDataGenerator, it's better to do it batch by batch. As far as I know, there're 2 ways to achieve this other than operating the whole dataset:<\/p>\n\n<ol>\n<li>You can use <a href=\"https:\/\/www.quora.com\/What-is-the-Lambda-layer-in-Keras\" rel=\"nofollow noreferrer\">Lambda Layer<\/a> to create a layer and then feed original training data to it. The output is the resized you need.<\/li>\n<\/ol>\n\n<p>Here is the sample code if you use TensorFlow as the backend of Keras:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>original_dim = (32, 32, 3)\ntarget_size = (64, 64)\ninput = keras.layers.Input(original_dim)\nx = tf.keras.layers.Lambda(lambda image: tf.image.resize(image, target_size))(input)\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>As @Retardust mentioned, maybe you can customize your own ImageDataGenerator as well as the preprocessing_function.<\/li>\n<\/ol>\n"},{"AnswerId":"41759831","CreationDate":"2017-01-20T09:10:18.903","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p><code>flow_from_directory(directory)<\/code> generates augmented images from directory with arbitrary collection of images. So there is need of parameter <code>target_size<\/code> to make all images of same shape. <\/p>\n\n<p>While <code>flow(X, y)<\/code> augments images which are already stored in a sequence in X which is nothing but numpy matrix and can be easily preprocessed\/resized before passing to <code>flow<\/code>. So no need for <code>target_size<\/code> parameter. As for resizing I prefer using <code>scipy.misc.imresize<\/code> over <code>PIL.Image resize<\/code>,  or cv2.resize as it can operate on numpy image data.<\/p>\n\n<pre><code>import scipy\nnew_shape = (28,28,3)\nX_train_new = np.empty(shape=(X_train.shape[0],)+new_shape)\nfor idx in xrange(X_train.shape[0]):\n    X_train_new[idx] = scipy.misc.imresize(X_train[idx], new_shape)\n<\/code><\/pre>\n"},{"AnswerId":"49990985","CreationDate":"2018-04-23T22:49:26.210","ParentId":null,"OwnerUserId":"7097044","Title":null,"Body":"<pre><code>X_data_resized = [skimage.transform.resize(image, new_shape) for image in X_data]\n<\/code><\/pre>\n\n<p>because of the above code is now depreciated...<\/p>\n"},{"AnswerId":"58814428","CreationDate":"2019-11-12T08:08:17.573","ParentId":null,"OwnerUserId":"8243797","Title":null,"Body":"<p>For anyone else who wants to do this, .flow method of <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">ImageDataGenerator<\/a> does not have a target_shape parameter and we cannot resize an image using preprocessing_function parameter as the documentation states <code>The function will run after the image is resized and augmented. The function should take one argument: one image (Numpy tensor with rank 3), and should output a Numpy tensor with the same shape.\n<\/code>\nSo in order to use .flow, you will have to pass resized images only otherwise use a custom generator that resizes them on the fly.<\/p>\n"}]}
{"QuestionId":41758693,"AnswerCount":0,"Tags":"<tensorflow><linear-regression>","CreationDate":"2017-01-20T07:59:40.310","AcceptedAnswerId":null,"OwnerUserId":1616677.0,"Title":"Is there a way to intersect real-valued column with a sparse column?","Body":"<p> is able to intersect a few sparse (categorical) columns. Is there a way to intersect a real-valued column with a sparse column in a  ? The mathematical meaning of this seems clear: I need different weights at continuous feature X, depending on a categorical feature Z. <\/p>\n","answers":[]}
{"QuestionId":41759332,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><recurrent-neural-network>","CreationDate":"2017-01-20T08:41:27.890","AcceptedAnswerId":null,"OwnerUserId":908744.0,"Title":"Calling a stateful LSTM as a functional model?","Body":"<p>I have a stateful LSTM defined as a Sequential model:<\/p>\n\n<pre><\/pre>\n\n<p>Later, I use it as a Functional model:<\/p>\n\n<pre><\/pre>\n\n<p>Is the state from   preserved when we apply  again on ? If yes, how can I reset the model state in between the calls? <\/p>\n","answers":[{"AnswerId":"41792464","CreationDate":"2017-01-22T15:03:23.673","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Following <strong>Note on using statefulness in RNNs<\/strong> from this <a href=\"https:\/\/keras.io\/layers\/recurrent\/#recurrent\" rel=\"nofollow noreferrer\">link<\/a> and <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/57429d1567b94955ac95afe87792411a420ff0ba\/keras\/layers\/recurrent.py\" rel=\"nofollow noreferrer\">Keras implementation<\/a> the answer is <strong>yes<\/strong> if:<\/p>\n\n<ol>\n<li>The <code>batch_size<\/code> in both models is the same (it's important due to the way Keras computes the inner states).<\/li>\n<li>You would first build and compile both models and then use them - for some reason Keras is resetting the inner states during the <code>build<\/code> of a layer (you can check it <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/57429d1567b94955ac95afe87792411a420ff0ba\/keras\/layers\/recurrent.py\" rel=\"nofollow noreferrer\">here<\/a> by looking for <code>reset_states<\/code> method).<\/li>\n<\/ol>\n\n<p>If you want to reset states you could call <code>reset_states<\/code> method on each recurrent layer you want ot reset states on.<\/p>\n"}]}
{"QuestionId":41759771,"AnswerCount":1,"Tags":"<python><tensorflow><prediction><softmax>","CreationDate":"2017-01-20T09:07:17.033","AcceptedAnswerId":null,"OwnerUserId":7445294.0,"Title":"Evaluating softmax (classification) accuracy in Tensorflow with round or argmax?","Body":"<h1>Problem<\/h1>\n\n<p>I tried to make a softmax classifier with Tensorflow and predict with . I found out that one of  is always higher than 0.5, and I've used  instead of .<\/p>\n\n<p>However, the gap of accuracy between those two methods is about 20% - the accuracy with  is higher than .<\/p>\n\n<p>I expected the accuracy of those two methods should be exactly the same or that  should be lower than , but it's not. Does anyone know why this is the case?<\/p>\n\n<h1>Y and y_<br><\/h1>\n\n<pre><\/pre>\n\n<h1>Correct prediction<\/h1>\n\n<pre><\/pre>\n\n<h1>Accuracy<\/h1>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"52615686","CreationDate":"2018-10-02T20:09:40.130","ParentId":null,"OwnerUserId":"8676953","Title":null,"Body":"<p>The question is a bit old but someone might stumble accross it like I did...<\/p>\n\n<p>In short:<\/p>\n\n<ul>\n<li><code>tf.argmax<\/code> reduces one dimension specified by the <em>axis<\/em> argument, it returns the index to the largest value along that <em>axis<\/em><\/li>\n<li><code>tf.round<\/code> rounds each each value, returning the same shape as the input<\/li>\n<li>the \"accuracy\" using <code>tf.round<\/code> will increase with length of the dimension specified by <em>axis<\/em><\/li>\n<\/ul>\n\n<p>e.g.<\/p>\n\n<pre><code>y = [\n    [1.0, 0.0, 0.0],\n    [0.0, 1.0, 0.0],\n    [0.0, 0.0, 1.0]\n]\n\ny_ = [\n    [0.92, 0.4, 0.4],\n    [0.2,0.6,0.2],\n    [0.2,0.4,0.2]\n]\n\nprint('argmax shape:', tf.argmax(y_, 1).shape)\nprint('round shape:', tf.round(y_).shape)\n\nwith tf.Session():\n    print('argmax value', tf.argmax(y_, 1).eval())\n    print('round value', tf.round(y_).eval())\n\n    correct_prediction_argmax = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))\n    correct_prediction_round = tf.equal(tf.round(tf.cast(y, tf.float32)), tf.round(y_))\n    print('argmax accuracy:', tf.reduce_mean(tf.cast(correct_prediction_argmax, tf.float32)).eval())\n    print('round accuracy:', tf.reduce_mean(tf.cast(correct_prediction_round, tf.float32)).eval())\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>argmax shape: (3,)\nround shape: (3, 3)\nargmax value [0 1 1]\nround value [[1. 0. 0.]\n [0. 1. 0.]\n [0. 0. 0.]]\nargmax accuracy: 0.6666667\nround accuracy: 0.8888889\n<\/code><\/pre>\n\n<p>You can see that using <code>tf.round<\/code> usually gets <em>points<\/em> for the two extra zeros on each row.<\/p>\n"}]}
{"QuestionId":41759872,"AnswerCount":0,"Tags":"<python><tensorflow><convolution>","CreationDate":"2017-01-20T09:13:34.267","AcceptedAnswerId":null,"OwnerUserId":2563522.0,"Title":"nan in filter matrix in cnn training tensorflow","Body":"<p>I am trying to implement a cnn for text classification in tensorflow using <a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"nofollow noreferrer\">this<\/a> link.\nHowever, rather than learning the embedding weights I am already feeding a predefined embeddings.\nMy input is of dimension 33*160. <\/p>\n\n<p>Text CNN<\/p>\n\n<pre><\/pre>\n\n<p><strong>code explained<\/strong><\/p>\n\n<p>vec_shape = dimension of input i.e 33 * 160.<\/p>\n\n<p>Training:<\/p>\n\n<pre><\/pre>\n\n<p>After some training steps ( around 48), variable <strong>weights<\/strong> becoming nan.<\/p>\n\n<p>at step 47, weights variable is <\/p>\n\n<pre><\/pre>\n\n<p>at step 48, weights variable is<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41761757,"AnswerCount":3,"Tags":"<keras>","CreationDate":"2017-01-20T10:46:07.723","AcceptedAnswerId":null,"OwnerUserId":3529361.0,"Title":"Keras model not saving","Body":"<p>When I save my model in Keras via  the file gets saved and there are no errors, although when I open the file this is what I get:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/d6RGv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/d6RGv.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I've searched online and no one seems to have this problem. Also my Console has no details. Any ideas?<\/p>\n","answers":[{"AnswerId":"43751712","CreationDate":"2017-05-03T05:27:06.913","ParentId":null,"OwnerUserId":"7712022","Title":null,"Body":"<p>To save model<\/p>\n\n<pre><code># serialize model to JSON\nmodel_json = model.to_json()\nwith open(\"model.json\", \"w\") as json_file:             \n     json_file.write(model_json) \n\n# serialize weights to HDF5\nmodel.save_weights(\"model.h5\")\nprint(\"Saved model to disk\")\n<\/code><\/pre>\n\n<p>To load model<\/p>\n\n<pre><code>from keras.models import load_model\nmodel.load_weights(\"model.h5\")\n<\/code><\/pre>\n"},{"AnswerId":"42101106","CreationDate":"2017-02-07T22:19:47.270","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>To load a model use this<\/p>\n\n<pre><code>from keras.models import load_model\nmodel = load_model(filepath)\n<\/code><\/pre>\n\n<p>You can find more documentation <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"46289565","CreationDate":"2017-09-18T23:28:27.860","ParentId":null,"OwnerUserId":"8328970","Title":null,"Body":"<p>Here is a YouTube video that explains exactly what you're wanting to do: <a href=\"https:\/\/youtu.be\/7n1SpeudvAE\" rel=\"nofollow noreferrer\">Save and load a Keras model<\/a><\/p>\n\n<p>There are three different saving methods that Keras makes available. These are described in the video link above (with examples), as well as below. <\/p>\n\n<p>First, the reason you're receiving the error is because the HDF5 file is not readable with a text editor. <\/p>\n\n<p>You used the correct <code>model.save(filepath)<\/code> function. This save function saves:<\/p>\n\n<ul>\n<li>The architecture of the model, allowing to re-create the model.<\/li>\n<li>The weights of the model.<\/li>\n<li>The training configuration (loss, optimizer).<\/li>\n<li>The state of the optimizer, allowing to resume training exactly where you left off.<\/li>\n<\/ul>\n\n<p>To load this saved model, you would use the following:<\/p>\n\n<pre><code>    from keras.models import load_model\n    new_model = load_model(filepath)\n<\/code><\/pre>\n\n<p>If you simply used <code>model.to_json()<\/code>, you would only be saving the architecture of the model. Additionally, if you used <code>model.save_weights()<\/code>, you would only be saving the weights of the model. With both of these alternative saving techniques, you would not be saving the training configuration (loss, optimizer), nor would you be saving the state of the optimizer.<\/p>\n"}]}
{"QuestionId":41761788,"AnswerCount":1,"Tags":"<tensorflow><dart>","CreationDate":"2017-01-20T10:47:56.050","AcceptedAnswerId":null,"OwnerUserId":4001247.0,"Title":"Tensorflow for Dart lang?","Body":"<p>I would like to use Tensorflow in Dart Language.<\/p>\n\n<p>I saw in Tensorflow documentation (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/<\/a>) that Dart does not cited.<\/p>\n\n<p>Is there any initiative to develop Tensorflow APIs for Dart Lang?<\/p>\n","answers":[{"AnswerId":"41769854","CreationDate":"2017-01-20T18:08:00.783","ParentId":null,"OwnerUserId":"269589","Title":null,"Body":"<p>Not officially, although many community activities have begun to implement TensorFlow APIs using <a href=\"http:\/\/www.grpc.io\/\" rel=\"nofollow noreferrer\">gRPC<\/a>. Native bindings are less popular due to the complexities involved. Even the official C++ API is behind Python currently.<\/p>\n\n<p>I would suggest first looking at JavaScript if you want to pursue Dart and TensorFlow, otherwise it's not likely you'll have a solution in the near future. <\/p>\n\n<p>I've been working on a Swift API for TensorFlow, and utilizing the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/go\" rel=\"nofollow noreferrer\">Go<\/a> API, but even though these are two very popular languages, outside of Python, in the TensorFlow community, the complexity of building the APIs are proving difficult to overcome.<\/p>\n"}]}
{"QuestionId":41762272,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-01-20T11:10:30.307","AcceptedAnswerId":42516032.0,"OwnerUserId":4176951.0,"Title":"use tensorflow.GPUOptions within Keras when using tensorflow backend","Body":"<p>In tensorflow I can do something like this when creating a session:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to do the same in keras with the tensorflow backend?<\/p>\n","answers":[{"AnswerId":"42516032","CreationDate":"2017-02-28T18:19:15.120","ParentId":null,"OwnerUserId":"1459254","Title":null,"Body":"<p>You can set the Keras global tensorflow session with <code>keras.backend.tensorflow_backend.set_session()<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport keras.backend.tensorflow_backend as ktf\n\n\ndef get_session(gpu_fraction=0.333):\n    gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=gpu_fraction,\n                                allow_growth=True)\n    return tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))\n\n\nktf.set_session(get_session())\n<\/code><\/pre>\n"}]}
{"QuestionId":41764041,"AnswerCount":2,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-01-20T12:45:51.123","AcceptedAnswerId":41791568.0,"OwnerUserId":766391.0,"Title":"Fine tuning pretrained model in keras","Body":"<p>I want to use a pretrained imagenet VGG16 model in keras and add my own small convnet on top. I am only interested in the features, not the predictions<\/p>\n\n<pre><\/pre>\n\n<p>load images from directory (the dir contains 4 images)<\/p>\n\n<pre><\/pre>\n\n<p>load the base model, preprocess input and get the features<\/p>\n\n<pre><\/pre>\n\n<p>this works, and I get the features for my images on the pretrained VGG.<\/p>\n\n<p>I now want to finetune the model and add some convolutional layers.\nI read <a href=\"https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html\" rel=\"noreferrer\">https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html<\/a> and <a href=\"https:\/\/keras.io\/applications\/\" rel=\"noreferrer\">https:\/\/keras.io\/applications\/<\/a> but cannot quite bring them together.<\/p>\n\n<p>adding my model on top:<\/p>\n\n<pre><\/pre>\n\n<p>building the complete model<\/p>\n\n<pre><\/pre>\n\n<p>stop base layers from being learned<\/p>\n\n<pre><\/pre>\n\n<p>new model<\/p>\n\n<pre><\/pre>\n\n<p>now fit the new model, the model is 4 images and [1,0,1,0] are the class labels.\nBut this is obviously wrong:<\/p>\n\n<pre><\/pre>\n\n<p>How is this done?<\/p>\n\n<p>How would I do it if I only wanted to replace the last convolutional block (conv block5 in VGG16) instead of adding something?<\/p>\n\n<p>How would I only train the bottleneck features?<\/p>\n\n<p>The features output  has shape (4, 512, 7, 7). There are four images, but what is in the other dimensions? How would I reduce that to a (1,x) array?<\/p>\n","answers":[{"AnswerId":"41781064","CreationDate":"2017-01-21T15:25:40.693","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Use <code>model.fit(X, y)<\/code> to train on your dataset as explained here: <a href=\"https:\/\/keras.io\/models\/model\/#fit\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/models\/model\/#fit<\/a><\/p>\n\n<p>Additionally you should add a Flatten layer and a dense layer with an ouput shape of 1 to get the correct result shape. <\/p>\n"},{"AnswerId":"41791568","CreationDate":"2017-01-22T13:37:59.160","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p><strong>Fitting model<\/strong><\/p>\n\n<p>The problem with your generator code is that the fit_generator method expects a generator function to generate the data for fitting which you don't provide.\nYou can either define a generator as done in the tutorial that you have linked to or create the data and labels yourself and fit your model yourself:<\/p>\n\n<pre><code>model_complete.fit(images, labels, batch_size=100, nb_epoch=2)\n<\/code><\/pre>\n\n<p>where images are your generated training images and labels are the corresponding labels.<\/p>\n\n<p><strong>Removing last layer<\/strong><\/p>\n\n<p>Assuming you have a model variable and the \"pop\" method described below, you can do <code>model = pop(model)<\/code> to remove the last layer.<\/p>\n\n<p><strong>Training only specific layers<\/strong>\nAs you have done in your code, you can do:<\/p>\n\n<pre><code>for layer in base_model.layers:\n    layer.trainable = False\n<\/code><\/pre>\n\n<p>Then you can \"unfreeze\" and layer that you want by changing their <code>trainable<\/code> property to <code>True<\/code>.<\/p>\n\n<p><strong>Changing dimensions<\/strong><\/p>\n\n<p>To change the output to a 1D array you can use the <a href=\"https:\/\/keras.io\/layers\/core\/#flatten\" rel=\"noreferrer\" title=\"Flatten layer\">Flatten layer<\/a><\/p>\n\n<hr>\n\n<p><strong><em>The pop method<\/em><\/strong><\/p>\n\n<pre><code>def pop(model):\n    '''Removes a layer instance on top of the layer stack.\n    This code is thanks to @joelthchao https:\/\/github.com\/fchollet\/keras\/issues\/2371#issuecomment-211734276\n    '''\n    if not model.outputs:\n        raise Exception('Sequential model cannot be popped: model is empty.')\n    else:\n        model.layers.pop()\n        if not model.layers:\n            model.outputs = []\n            model.inbound_nodes = []\n            model.outbound_nodes = []\n        else:\n            model.layers[-1].outbound_nodes = []\n            model.outputs = [model.layers[-1].output]\n        model.built = False\n\n    return model\n<\/code><\/pre>\n"}]}
{"QuestionId":41764199,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-01-20T12:53:32.843","AcceptedAnswerId":41802308.0,"OwnerUserId":482601.0,"Title":"Row-wise Histogram","Body":"<p>Given a 2-dimensional tensor , what's the fastest way to compute a tensor  where<\/p>\n\n<pre><\/pre>\n\n<p>I.e. where  is called per row of the input tensor ?<\/p>\n\n<p>It seems that  is missing an  parameter that works like, e.g., 's  parameter.<\/p>\n","answers":[{"AnswerId":"54170386","CreationDate":"2019-01-13T15:34:16.087","ParentId":null,"OwnerUserId":"2230045","Title":null,"Body":"<p>I would like to propose another implementation.\nThis implementation can also handle multi axes and unknown dimensions (batching).<\/p>\n\n<pre><code>def histogram(tensor, nbins=10, axis=None):\n    value_range = [tf.reduce_min(tensor), tf.reduce_max(tensor)]\n\n    if axis is None:\n        return tf.histogram_fixed_width(tensor, value_range, nbins=nbins)\n    else:\n        if not hasattr(axis, \"__len__\"):\n            axis = [axis]\n\n        other_axis = [x for x in range(0, len(tensor.shape)) if x not in axis]\n        swap = tf.transpose(tensor, [*other_axis, *axis])\n        flat = tf.reshape(swap, [-1, *np.take(tensor.shape.as_list(), axis)])\n\n        count = tf.map_fn(lambda x: tf.histogram_fixed_width(x, value_range, nbins=nbins), flat, dtype=(tf.int32))\n\n        return tf.reshape(count, [*np.take([-1 if a is None else a for a in tensor.shape.as_list()], other_axis), nbins])\n<\/code><\/pre>\n\n<p>The only slow part here is <code>tf.map_fn<\/code> but it is still faster than the other solutions mentioned.<\/p>\n\n<p>If someone knows a even faster implementation please comment since this operation is still very expensive.<\/p>\n"},{"AnswerId":"41768777","CreationDate":"2017-01-20T16:58:42.527","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/histogram_ops\/histograms#histogram_fixed_width\" rel=\"nofollow noreferrer\"><code>tf.histogram_fixed_width<\/code><\/a> works on the entire tensor indeed. You have to loop through the rows explicitly to compute the per-row histograms. Here is a complete working example using TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#while_loop\" rel=\"nofollow noreferrer\"><code>tf.while_loop<\/code><\/a> construct :<\/p>\n\n<pre><code>import tensorflow as tf\n\nt = tf.random_uniform([2, 2])\n\ni = 0\nhist = tf.constant(0, shape=[0, 5], dtype=tf.int32)\n\ndef loop_body(i, hist):\n  h = tf.histogram_fixed_width(t[i, :], [0.0, 1.0], nbins=5)\n  return i+1, tf.concat_v2([hist, tf.expand_dims(h, 0)], axis=0)\n\ni, hist = tf.while_loop(\n              lambda i, _: i &lt; 2, loop_body, [i, hist],\n              shape_invariants=[tf.TensorShape([]), tf.TensorShape([None, 5])])\n\nsess = tf.InteractiveSession()\nprint(hist.eval())\n<\/code><\/pre>\n"},{"AnswerId":"41802308","CreationDate":"2017-01-23T08:54:44.490","ParentId":null,"OwnerUserId":"482601","Title":null,"Body":"<p>Inspired by keveman's answer and because the number of rows of <code>t<\/code> is fixed and rather small, I chose to use a combination of <code>tf.gather<\/code> to split rows and <code>tf.pack<\/code> to join rows. It looks simple and works, will see if it is efficient...<\/p>\n\n<pre><code>t_histo_rows = [\n        tf.histogram_fixed_width(\n            tf.gather(t, [row]),\n            vals, nbins)\n        for row in range(t_num_rows)]\n\nt_histo = tf.pack(t_histo_rows, axis=0)\n<\/code><\/pre>\n"}]}
{"QuestionId":41764330,"AnswerCount":7,"Tags":"<python><tensorflow>","CreationDate":"2017-01-20T13:00:26.970","AcceptedAnswerId":44915373.0,"OwnerUserId":3659855.0,"Title":"Tensorflow GPU - Spyder","Body":"<p>I am having a problem with importing tensorflow GPU on spyder. \nThis is what I get when I type import tensorflow on iPython using Spyder 3 (on MacOS)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/5S8j5.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5S8j5.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>When I checked what packages I have within tensorflow I get this list, which I see that iPython, and Spyder are not included.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/a46yg.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/a46yg.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Also in anaconda, when I click on tensorflow in Environments, I have the option to open it with Python, with terminal, but not with iPython or Spyder (or Jupyter)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/T5v4k.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/T5v4k.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I would very much appreciate any help.<\/p>\n\n<p>UPDATE 1: <\/p>\n\n<p>Ok I managed to get Spyder on tensorflow's ENV but when I run the command 'import tensorflow' on iPython I get this error (when I run the same command on Python's tab everything seems to work just fine.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dMhNk.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dMhNk.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"47438304","CreationDate":"2017-11-22T15:23:51.940","ParentId":null,"OwnerUserId":"3789876","Title":null,"Body":"<p>I figured out how to get this working using the instruction on Tensorflow link. Once you create the tensorflow enviroment you can <strong>Spyder(Tensorflow)<\/strong> in your ananconda start menu folder. If you run your code by opening this the tensorflow should work.<\/p>\n"},{"AnswerId":"47463764","CreationDate":"2017-11-23T22:04:47.757","ParentId":null,"OwnerUserId":"3349973","Title":null,"Body":"<p>I faced the same issue and solved by performing following steps in order.<\/p>\n\n<p>Assuming that you have created <code>conda<\/code> <code>environment<\/code>, installed <code>tensorflow<\/code> and activated it and also installed <code>spyder<\/code>.<\/p>\n\n<p>Check the executable python of your conda environment<\/p>\n\n<pre><code>&gt;&gt;import sys\n&gt;&gt;sys.executable\n<\/code><\/pre>\n\n<p>Note the path of <code>python<\/code> executable.<\/p>\n\n<p>Go to Spyder preferences and set the interpreter path to the one noted above.<\/p>\n"},{"AnswerId":"46025590","CreationDate":"2017-09-03T16:34:43.633","ParentId":null,"OwnerUserId":"8249000","Title":null,"Body":"<p>I have seen another issue in anaconda. If you install Tensor flow in root or custom environment (like 'tensorflow') its not accessible from jupyter notebook or spyder. The best way to do this is to install it in the administrator mode. Follow these steps:<\/p>\n\n<ol>\n<li>Open \"Anaconda Prompt\" as an administrator. <\/li>\n<li>Verify the status on top written \"Administrator: Anaconda Prompt\"<\/li>\n<li>DON'T Activate any of the environments, root or tensorflow. <\/li>\n<li>Type in the command \"pip install --ignore-installed --upgrade tensorflow-gpu\" to install Tensorflow with GPU support.<\/li>\n<li>To install Keras type \"conda install -c conda-forge keras\"<\/li>\n<li>To verify installation, type 'python' and then inside python env. type 'import tensorflow as tf'. If all is well it will work without error.<\/li>\n<\/ol>\n"},{"AnswerId":"46559188","CreationDate":"2017-10-04T07:21:47.073","ParentId":null,"OwnerUserId":"3789876","Title":null,"Body":"<p>The instructions mentioned on the (<a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a>) link do NOT work for the Ananconda\/Spyder setups on windows. Having struggled through this for hours below is the easiest solution to get this working. Hope this helps! <\/p>\n\n<p>Basically you do NOT need to create a seperate tensorflow environment if you want to run this on spyder. Use the below commands to install tensorflow on the ananconda client.<\/p>\n\n<p>1) Open the Ananconda prompt from the installation folder in the start menu.\n2) Run below commands:<\/p>\n\n<pre><code>conda install pyqt\nconda install tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"41764458","CreationDate":"2017-01-20T13:06:48.787","ParentId":null,"OwnerUserId":"6348508","Title":null,"Body":"<p>Spyder is picking up the default Anaconda Env, whereas you have tensorflow install in a separate environment <code>tensorflow<\/code><\/p>\n\n<p>To work with Spyder &amp; tensorflow, install tensor flow and openssl in default ENV <\/p>\n"},{"AnswerId":"45327384","CreationDate":"2017-07-26T12:44:29.333","ParentId":null,"OwnerUserId":"2183380","Title":null,"Body":"<p>Using the Anaconda Navigator:\nIt works for me in a different way: \nAs tensorflow is separately created environment, install spyder in tensorflow envirnment. \n(Anaconda>Home>Applications on><em>tensorflow<\/em>>spyder>install ... wait for installation to complete and ten launch)<\/p>\n"},{"AnswerId":"44915373","CreationDate":"2017-07-05T00:04:16.313","ParentId":null,"OwnerUserId":"8249000","Title":null,"Body":"<p>Finally I solved this mystery. If you have installed Spyder from the Anaconda, go to the Anaconda launcher. There go to environments, you will see two of them: root and tensorflow. The latter one is created due to the instructions by tensorflow.org. Just run all those instructions on the root, don't activate tensorflow environment, it will work. Everything will be available in spyder.<\/p>\n"}]}
{"QuestionId":41764572,"AnswerCount":1,"Tags":"<python><tensorflow><variable-assignment>","CreationDate":"2017-01-20T13:13:27.837","AcceptedAnswerId":41768462.0,"OwnerUserId":6667826.0,"Title":"How to assign tensor value in to another function's tensor variable?","Body":"<p>I am calculating some features from a set of data. I made as function for each calculation. I want to put all of those values in an array in another function and I want to call the function. how to do it ?<\/p>\n\n<p>The 3 functions that I have are <\/p>\n\n<pre><\/pre>\n\n<p>I want these values to be assigned to a array inside function<\/p>\n\n<pre><\/pre>\n\n<p>then I want to call the function<\/p>\n\n<pre><\/pre>\n\n<p>How can i do it ? <\/p>\n\n<p>I tried<\/p>\n\n<pre><\/pre>\n\n<p>but it gives me error <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41768462","CreationDate":"2017-01-20T16:41:15.450","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/variable_helper_functions#assign\" rel=\"nofollow noreferrer\"><code>tf.assign<\/code><\/a> only lets you assign a value to the entire variable, and not to a slice of a variable. You would have to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sparse_variable_updates#scatter_update\" rel=\"nofollow noreferrer\"><code>tf.scatter_update<\/code><\/a> to assign to a slice of a variable. However, in your case, you can simply create the <code>pool<\/code> tensor by concatenating the three values. Here is the complete working program :<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef peak_value(x):\n    return tf.reduce_max(tf.abs(x))\ndef rootms(x):\n    return tf.sqrt(tf.reduce_mean(tf.square(x)))\ndef meanofabs(x):\n    return tf.reduce_mean(tf.abs(x))\n\n\ndef pooldata(x,size):\n    pool = tf.concat_v2([tf.expand_dims(peak_value(x), 0),\n                         tf.expand_dims(rootms(x), 0),\n                         tf.expand_dims(meanofabs(x), 0)], axis=0)\n    # i want to\n    # assign pool[0] with peak_value(x)\n    # assign pool[1] with rootms(x)\n    # assign pool[2] with meanofabs(x)\n    return pool\n\nsess = tf.Session()\nsize = 1\nx = [1.0]\n\nprint(sess.run(pooldata(x, size)))\n<\/code><\/pre>\n"}]}
{"QuestionId":41765486,"AnswerCount":1,"Tags":"<tensorflow><max-pooling>","CreationDate":"2017-01-20T14:03:49.547","AcceptedAnswerId":null,"OwnerUserId":2736559.0,"Title":"What is the equivalent of Caffe's Maxpooling in tensorflow?","Body":"<p>I am confused how we define  in . The documentation is vague and does not explain the parameters well.<br>\nIn the pooling documentation it only says: <\/p>\n\n<blockquote>\n  <p>ksize: A list of ints that has length >= 4. The size of the window for\n  each dimension of the input tensor. strides: A list of ints that has\n  length >= 4. The stride of the sliding window for each dimension of\n  the input tensor.<\/p>\n<\/blockquote>\n\n<p>and <\/p>\n\n<blockquote>\n  <p>Each pooling op uses rectangular windows of size ksize separated by\n  offset strides. For example, if strides is all ones every window is\n  used, if strides is all twos every other window is used in each\n  dimension, etc.<\/p>\n<\/blockquote>\n\n<p>What is the equivalent of the following 's  in ?<\/p>\n\n<pre><\/pre>\n\n<p>I'm not sure whether they mean overlapping pooling by all one strides [1,1,1,1] and non-overlapping [2,2,2,2] by saying  <\/p>\n\n<blockquote>\n  <p>if strides is all ones every window is\n  used, if strides is all twos every other window is used in each\n  dimension, etc.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"41765619","CreationDate":"2017-01-20T14:11:00.413","ParentId":null,"OwnerUserId":"6786718","Title":null,"Body":"<p>To do max-pooling in Tensor-Flow use:<\/p>\n\n<pre><code>tf.nn.max_pool(value, ksize, strides, padding, data_format='NHWC', name=None)\n<\/code><\/pre>\n\n<p>where ksize defines the window used for max-pooling. Note that you must specify the window size for each dimension of your input. This is the biggest difference to caffe, where caffe does all the dimension calculations for you. Note that you may have varying dimensions depending on your number of outputs that come from the previous convolutional layer.<\/p>\n\n<p>Stride has still the same effect has in caffe (\"skipping\" the inputs. However you must specify the stride again for each dimension of the input.<\/p>\n\n<p>The dimensions are at least 4 or larger. <\/p>\n\n<p>See the documentation here: <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/pooling\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/pooling<\/a><\/p>\n"}]}
{"QuestionId":41767003,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><pycharm><data-science>","CreationDate":"2017-01-20T15:22:04.287","AcceptedAnswerId":null,"OwnerUserId":5368816.0,"Title":"How to set restored weights to zero randomly in tensorflow?","Body":"<p>Good afternoon.\nI'm a tensorflow newbie and currently trying to solve this problem:\n1) get a simple neural network, train it, print accuracy (done)\n2) save it (done)\n3) restore it (done)\n4) set restored weights to zero randomly. (&amp;)<\/p>\n\n<p>I've read this topic: <a href=\"https:\/\/stackoverflow.com\/questions\/36986218\/dynamically-changing-weights-in-tensorflow\">Dynamically changing weights in TensorFlow<\/a> and tried several things from there, but to no avail.\nHere's my code:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to use random distribution multiplied by zeros, than simply 0, nothing changes, the accuracy is the same even when I try to put W=0.<\/p>\n\n<p>Would be very grateful for someone's advice.<\/p>\n","answers":[{"AnswerId":"41769541","CreationDate":"2017-01-20T17:46:35.047","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>The line<\/p>\n\n<pre><code>W = tf.mul(W, 0)\n<\/code><\/pre>\n\n<p>creates a new node in the graph that is not used by anyone - <code>accuracy<\/code> is still using the old <code>W<\/code>, that is why you are seeing no change. The way to change <code>W<\/code> is to use TensorFlow's assign and run it (see <a href=\"https:\/\/stackoverflow.com\/questions\/34220532\/how-to-assign-value-to-a-tensorflow-variable\">How to assign value to a tensorflow variable?<\/a>), something like<\/p>\n\n<pre><code>assign_op = W.assign(tf.mul(W,0))\nsess.run(assign_op)\n<\/code><\/pre>\n"}]}
{"QuestionId":41767005,"AnswerCount":1,"Tags":"<python><linux><unicode><pytorch>","CreationDate":"2017-01-20T15:22:08.063","AcceptedAnswerId":43824857.0,"OwnerUserId":2775876.0,"Title":"Python wheels: cp27mu not supported","Body":"<p>I'm trying to install pytorch (<a href=\"http:\/\/pytorch.org\/\" rel=\"noreferrer\">http:\/\/pytorch.org\/<\/a>) on Linux, and according to my machine configuration, the command I should run is:<\/p>\n\n<p><\/p>\n\n<p>On one machine (Linux distribution Slackware 14.1) the installation fails with error: , while on another one (Ubuntu 15.10) it succeeds. From what I understood, the problem seems to be the  in the wheel name.<\/p>\n\n<p>Using the command  from the Python shell, I get this from the Slackware machine:<\/p>\n\n<p><\/p>\n\n<p>and this from the Ubuntu machine:<\/p>\n\n<p><\/p>\n\n<p>From <a href=\"https:\/\/www.python.org\/dev\/peps\/pep-0513\/\" rel=\"noreferrer\">https:\/\/www.python.org\/dev\/peps\/pep-0513\/<\/a>, it seems to me that supporting  or  depends on an option passed at compile time, .<\/p>\n\n<p>Now, maybe at this point I shouldn't even be asking the question, but just to be sure, does this mean that I have to compile Python with the  on the Slackware machine in order to install that wheel?<\/p>\n","answers":[{"AnswerId":"43824857","CreationDate":"2017-05-06T20:00:38.377","ParentId":null,"OwnerUserId":"2451329","Title":null,"Body":"<p>This is exactly that. \nRecompile python under slack with --enable-unicode=ucs4 and you can then install the whl.<\/p>\n"}]}
{"QuestionId":41767411,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-20T15:43:02.267","AcceptedAnswerId":41769656.0,"OwnerUserId":2382483.0,"Title":"Tensorflow: feeding placeholder in loop within loop fails","Body":"<p>I am trying to repeatedly train a neural network at different sizes of hidden layer to determine how many neurons it should be. I wrote a net that works fine when doing one pass through. The code is:<\/p>\n\n<pre><\/pre>\n\n<p>When I put this call in a loop however, I make it through only the first iteration. It seems to fail during the second iteration the first time I hit this line:<\/p>\n\n<pre><\/pre>\n\n<p>I get the error: <\/p>\n\n<p>I logged both  and  right before feeding and they look exactly like they do on each preceding run. If I comment that second  part out it works great, but I kinda need my summaries...<\/p>\n\n<p>My outer loop looks like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41769656","CreationDate":"2017-01-20T17:54:27.993","ParentId":null,"OwnerUserId":"2878004","Title":null,"Body":"<p>Because you call the train function inside a loop then each time it runs it is creating a new copy of the placeholders. The first time it runs it is fine because there is only one copy. The second time it runs now you have duplicate placeholders. The solution is to separate out the code that builds the model from the code that runs the training.<\/p>\n"}]}
{"QuestionId":41768217,"AnswerCount":1,"Tags":"<gcc><cmake><caffe>","CreationDate":"2017-01-20T16:28:25.777","AcceptedAnswerId":null,"OwnerUserId":7447194.0,"Title":"Compile py-fast-rcnn using Cmake and met \"Link error\" when \"make runtest\"","Body":"<pre><\/pre>\n\n<p>This error appears at the last steps of Cmake compilation for py-fast-rcnn when I \"make runtest\". It looks like a linker problem but I don't know how to fix it. My guess is the version of gcc boost may be incompatible.<\/p>\n","answers":[{"AnswerId":"41833971","CreationDate":"2017-01-24T16:56:58.017","ParentId":null,"OwnerUserId":"7447194","Title":null,"Body":"<p>I solved this problem by remove the Anaconda, seems Anaconda has some disruption upon my manually set dependencies.<\/p>\n"}]}
{"QuestionId":41768341,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-01-20T16:35:26.720","AcceptedAnswerId":null,"OwnerUserId":3052835.0,"Title":"Hacky way to augment multichannel images in Keras","Body":"<p>I need to augment multichannel images and would like to use ImageDataGenerator\nUnfortunately, it supports 1,3 and 4 channel images only and I need much more. Would it be OK to directly edit site-packages\/Keras\/preprocessing\/image.py adding necessary number of channels?<\/p>\n\n<pre><\/pre>\n\n<p>where XXX - is a number of channels that I need. Would this break anything?\nThanks!<\/p>\n","answers":[{"AnswerId":"41822633","CreationDate":"2017-01-24T07:42:32.493","ParentId":null,"OwnerUserId":"3052835","Title":null,"Body":"<p>I tried that anyway and it seems to work without too many side effects. I didn't explore every possible problem, though. So if you have multichannel data (like satellite imagery, etc) you can try this hack. There are couple of places where you need to augment the augmentation code :)<\/p>\n"}]}
{"QuestionId":41768773,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-20T16:58:29.783","AcceptedAnswerId":null,"OwnerUserId":4929293.0,"Title":"I want to know why this tensor flow models cost doesn't decrease?","Body":"<p>I tried to write source code for tagging verb as B with LSTM. \nI mapped word as number and one hot encoding.<br>\nI thought my source code is correct but result is different from my expectation.<\/p>\n\n<pre><\/pre>\n\n<p>Source code is shared above. I am expecting cost to decrease after studying some cases. However cost doesn't decrease and result is always return 0\nI can't find why this is happening. <\/p>\n","answers":[]}
{"QuestionId":41769301,"AnswerCount":1,"Tags":"<python><tensorflow><runtime-error><bazel>","CreationDate":"2017-01-20T17:31:14.927","AcceptedAnswerId":41770554.0,"OwnerUserId":3529271.0,"Title":"Error thrown in Google's Inception-v3 fine-tuning script","Body":"<p>When following the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception#how-to-retrain-a-trained-model-on-the-flowers-data\" rel=\"nofollow noreferrer\">Readme<\/a> to fine-tune Google's Inception-v3 image classification model, I get the error: <\/p>\n\n<p><\/p>\n\n<p>This occurs after running the final command:<\/p>\n\n<pre><\/pre>\n\n<p>I have 0 idea whats going on here as this error gets thrown from a python file written by the TF team. Additionally, being a TF newbie, I do not know my way around enough to attempt a deep debugging session. Just by looking at the path from the error, there might be an issue with the script running TF slim code?<\/p>\n\n<p>Anyhow, I am running macOS Sierra with Python 3.6 and the TensorFlow Python API r0.12.<\/p>\n","answers":[{"AnswerId":"41770554","CreationDate":"2017-01-20T18:54:29.857","ParentId":null,"OwnerUserId":"3529271","Title":null,"Body":"<p>So turns out this error was thrown if the current installation of tensorflow did not have the most recent tensorflow-slim code. Install directions <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":41771965,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><keras>","CreationDate":"2017-01-20T20:36:47.287","AcceptedAnswerId":41773932.0,"OwnerUserId":2414957.0,"Title":"Error when checking model input: expected convolution2d_input_1 to have shape (None, 3, 32, 32) but got array with shape (50000, 32, 32, 3)","Body":"<p>Can someone please guide how to fix this error? I just started on Keras:<\/p>\n\n<pre><\/pre>\n\n<p>Output is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41773932","CreationDate":"2017-01-20T23:20:20.657","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>If you print <code>x_train.shape<\/code> you will see the shape being <code>(50000, 32, 32, 3)<\/code> whereas you have given <code>input_shape=(3, 32, 32)<\/code> in the first layer. The error simply says that the expected input shape and data given are different. <\/p>\n\n<p>All you need to do is give  <code>input_shape=(32, 32, 3)<\/code>. Also if you use this shape then you must use <code>tf<\/code> as your image ordering. <code>backend.set_image_dim_ordering('tf')<\/code>. <\/p>\n\n<p>Otherwise you can permute the axis of data. <\/p>\n\n<pre><code>x_train = x_train.transpose(0,3,1,2)\nx_test = x_test.transpose(0,3,1,2)\nprint x_train.shape\n<\/code><\/pre>\n"}]}
{"QuestionId":41772031,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><lstm>","CreationDate":"2017-01-20T20:40:30.970","AcceptedAnswerId":41774584.0,"OwnerUserId":7415539.0,"Title":"Deep Learning - CLDNN (multi-model neural network) dropout","Body":"<p>I am currently building a CLDNN (Convolutional, LSTM, Deep Neural Network) model for raw signal classification.<\/p>\n\n<p>Since the number of trainable parameters are easily over millions, I thought dropout would help prevent overfitting.<\/p>\n\n<p>My question is also applicable to other networks with multiple models stacked.<\/p>\n\n<p>If I have the network structured as <\/p>\n\n<p>input -> convolution -> LSTM -> DNN -> output<\/p>\n\n<p>Do I have to put a dropout after each layer or only right before the output?<\/p>\n\n<h2>input -> convolution -> dropout -> LSTM -> dropout -> DNN -> dropout -> output<\/h2>\n\n<p>or<\/p>\n\n<h2>input -> convolution -> LSTM -> DNN -> dropout -> output<\/h2>\n\n<p>So far, I've only seen dropout applied to convNets, but I don't see why it should be only restricted to convNets.\nDo other networks, such as LSTM and DNN, also use dropout to prevent overfitting?<\/p>\n","answers":[{"AnswerId":"41774584","CreationDate":"2017-01-21T00:44:35.773","ParentId":null,"OwnerUserId":"2708478","Title":null,"Body":"<p>Yes, you can use Dropout after each of your layers. <\/p>\n\n<p>It does not make sense to apply dropout to your last layer (the layer that produces the probability distribution over the classes), though.<\/p>\n\n<p>Do not forget that LSTM is a recurrent model, so you have to use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/rnn_cell\/rnn_cell_wrappers__rnncells_that_wrap_other_rnncells_#DropoutWrapper\" rel=\"nofollow noreferrer\">DropoutWrapper<\/a> class to apply dropout at each time step.<\/p>\n"}]}
{"QuestionId":41774190,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2017-01-20T23:51:09.330","AcceptedAnswerId":41774456.0,"OwnerUserId":2939108.0,"Title":"Tensorflow: Modify Calculation of Hidden State","Body":"<p>I would like to slightly modify the hidden state calculated (and consequently re-inputed) at each time step of an LSTM. For example, something as simple as adding a constant vector to the hidden state that is produced by the regular LSTM process. I don't intend to modify the regular gate mechanics -- I view this more as a calculation of the final hidden state that is normally created by an LSTM. I feel creating my own LSTM cell is overdoing it. Alternatively, the  seems like it could achieve this goal, though I'm not intending to create a new output, just modify the hidden state.<\/p>\n","answers":[{"AnswerId":"41774456","CreationDate":"2017-01-21T00:25:25.803","ParentId":null,"OwnerUserId":"2708478","Title":null,"Body":"<p>You can modify the state of the LSTM between time steps if you are using a for loop to process your sequences, since an RNN cell in TensorFlow takes two arguments: the new input and the hidden state:<\/p>\n\n<pre><code>lstm = rnn_cell.BasicLSTMCell(lstm_size)\n\ninitial_state = state = tf.zeros([batch_size, lstm.state_size])\n\nwith tf.variable_scope(\"RNN\"):\n  for time_step in range(num_steps):\n    if time_step &gt; 0: tf.get_variable_scope().reuse_variables()\n\n    # modify the state\n    modified_state = state + 1\n\n    output, state = lstm(words[:, i], modified_state)\n\n\nfinal_state = state\n<\/code><\/pre>\n"}]}
{"QuestionId":41774578,"AnswerCount":1,"Tags":"<python><r><tensorflow>","CreationDate":"2017-01-21T00:43:41.380","AcceptedAnswerId":null,"OwnerUserId":6537035.0,"Title":"Locate Python binary location in windows (Tensorflow)","Body":"<p>I'm trying to follow tensorflow instructions in R but am stuck on providing the Python binary path location on Windows. I believe the instructions are given to Mac environment? I would like to know where can I locate my Python binary on my windows environment.<\/p>\n\n<p>Link to Tutorial: <a href=\"https:\/\/github.com\/rstudio\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rstudio\/tensorflow<\/a><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41838397","CreationDate":"2017-01-24T20:28:12.543","ParentId":null,"OwnerUserId":"1573619","Title":null,"Body":"<p>Try this on the command prompt:<\/p>\n\n<pre><code>python -c \"import sys; print(sys.executable)\"\n<\/code><\/pre>\n\n<p>That will give you the full executable path for the python version that you are using. <\/p>\n\n<p><strong>Output: [Your output may be different]<\/strong><\/p>\n\n<pre><code>C:\\Users\\&lt;username&gt;\\AppData\\Local\\Programs\\Python\\Python35\\python.exe\n<\/code><\/pre>\n"}]}
{"QuestionId":41775438,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><non-linear-regression>","CreationDate":"2017-01-21T03:27:25.313","AcceptedAnswerId":null,"OwnerUserId":5844494.0,"Title":"nan session output from polynomial regression training in Tensorflow","Body":"<p>I am new to Tensorflow and am working through the examples of regression examples given here <a href=\"https:\/\/github.com\/pkmital\/tensorflow_tutorials\" rel=\"nofollow noreferrer\">tensorflow tutorials<\/a>. Speicifically, I am working on the 3rd: \"polynomial_regression.py\"<\/p>\n\n<p>I followed the linear regression example fine, and have now moved on to the polynomial regression. <\/p>\n\n<p>However, I wanted to try substituting another set of data instead of that made up in the example.  I did this by exchanging <\/p>\n\n<pre><\/pre>\n\n<p>for <\/p>\n\n<pre><\/pre>\n\n<p>I.e. the second was what was given in the example, and I wanted to try to run the same training with the new xs,ys, n_observation.  These were the only lines I changed.  I also tried changing the dtype of the array to be float64, but this did not change the output.  <\/p>\n\n<p>The output I am getting (which is from  is just a repeated .  When I switch back to the original data, the network runs fine, and generates a fitting funciton.  <\/p>\n\n<p>Thank you for any ideas!<\/p>\n","answers":[{"AnswerId":"47379285","CreationDate":"2017-11-19T16:42:40.303","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>NaNs can be caused by many things, usually some form of numerical instability. Lowering the learning rate or using a more stable optimizer are good things to try.<\/p>\n"}]}
{"QuestionId":41776188,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-01-21T05:53:16.100","AcceptedAnswerId":41783571.0,"OwnerUserId":628538.0,"Title":"How to access iter_size variable in Python (Caffe)?","Body":"<p>I am trying to read the parameters of the  file in Python. I can see variables like , ,  etc. using ,  etc. however, I cannot access the  parameter which performs multiple iterations of forward\/backward before updating the weights<\/p>\n","answers":[{"AnswerId":"41783571","CreationDate":"2017-01-21T19:29:37.797","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like <code>iter_size<\/code> is not exposed to <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/_caffe.cpp#L443-L446\" rel=\"nofollow noreferrer\">pycaffe interface<\/a>:<\/p>\n\n<pre><code>bp::class_&lt;SolverParameter&gt;(\"SolverParameter\", bp::no_init)\n    .add_property(\"max_iter\", &amp;SolverParameter::max_iter)\n    .add_property(\"display\", &amp;SolverParameter::display)\n    .add_property(\"layer_wise_reduce\", &amp;SolverParameter::layer_wise_reduce);\n<\/code><\/pre>\n\n<p>As you can see <code>max_iter<\/code> and <code>display<\/code> are exposed, but not <code>iter_size<\/code>.<br>\nI suppose adding a line<\/p>\n\n<pre><code>    .add_property(\"iter_size\", &amp;SolverParameter::iter_size)\n<\/code><\/pre>\n\n<p>After the <code>\"display\"<\/code> line should do the trick.<br>\nRemember to compile the pycaffe interface after this change with <code>make pycaffe<\/code>.<\/p>\n"}]}
{"QuestionId":41776797,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-21T07:26:10.317","AcceptedAnswerId":null,"OwnerUserId":4929293.0,"Title":"Tensorflow variable initialization","Body":"<pre><\/pre>\n\n<p>If source code is like above, rnn_cell and state is initialized in every steps in  loops?<br>\nIf I want to use state in other training case then I have to reuse it. So rnn_cell and state should be initialized at first only not after that.<br>\nI can't imagine how this code works. <\/p>\n","answers":[{"AnswerId":"42958382","CreationDate":"2017-03-22T17:11:32.693","ParentId":null,"OwnerUserId":"5424661","Title":null,"Body":"<p>I think the problem is you have to separate your computational graph part with the session running part. what you are doing now is not how tensorflow usually works. maybe try this:<\/p>\n\n<pre><code>rnn_cell = tf.nn.rnn_cell.BasicLSTMCell(rnn_size)\nstate = rnn_cell.zero_state(batch_size, tf.float32)\nx_split = tf.split(0, time_step_size, x)\noutputs, state = tf.nn.rnn(rnn_cell, x_split, state)\n\nprediction = tf.reshape(tf.concat(1, outputs), [-1, rnn_size])\nreal = tf.reshape(y, [-1])\nratio = tf.ones([time_step_size * batch_size])\n\nloss = tf.nn.seq2seq.sequence_loss_by_example([prediction], [real], [ratio])\ncost = tf.reduce_mean(loss)\/batch_size\ntrain = tf.train.AdamOptimizer(0.01).minimize(cost)\n\ninit = tf.global_variables_initializer()\nsess = tf.Session()\nsess.run(init)\n\nfor i in range(len(x_data)):\n    x = process_x(x_data[i])[:std_size]\n    y = word[i][:std_size]\n\nstep = 0\nwhile step &lt; 1000:\n    sess.run(train, feed_dict={x_split:x, real:y})\n    step+=1\n    result = sess.run(tf.arg_max(prediction, 1))\n    print result, [t for t in result] == y\n<\/code><\/pre>\n\n<p>your code may have some design problem, but the point is separating your graph design with your \"training\".<\/p>\n"}]}
{"QuestionId":41777019,"AnswerCount":1,"Tags":"<python><variables><tensorflow>","CreationDate":"2017-01-21T07:55:04.677","AcceptedAnswerId":41779193.0,"OwnerUserId":6667826.0,"Title":"how to select particular values defined in one tensor with data in another tensor","Body":"<p>I have a tensor variable y ( ) and another variable <\/p>\n\n<p>I want to select all rows and columns [2,4,5,7] as mentioned in x for data in y.<\/p>\n\n<p>In Matlab I can simply define  gives me the selected columns in x for y data. \nhow to do it in tensorflow ?<\/p>\n","answers":[{"AnswerId":"41779193","CreationDate":"2017-01-21T12:13:11.087","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>If you want to do <code>req_data = y[:,x]<\/code><br>\nfirst use <code>tf.transpose<\/code>, so the shape of tensor will be (8, 140)<br>\nthen use <code>tf.gather<\/code> to select the data<br>\nbecause <code>tf.gather<\/code> only work on axis=0, so transpose first\nthen transpose back<\/p>\n\n<pre><code>a = tf.constant([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10],\n                 [11, 12, 13, 14, 15, 16, 17, 18, 19, 20]])\na_trans = tf.transpose(a)\nb = tf.constant([2,4,5,7])\nc = tf.gather(a_trans, b)\nc_trans = tf.transpose(c)\n\nwith tf.Session() as sess:\n    print sess.run(c_trans)\n    #output [[3  5  6  8]\n    #        [13 15 16 18]]\n<\/code><\/pre>\n"}]}
{"QuestionId":41777236,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-21T08:27:57.003","AcceptedAnswerId":41777496.0,"OwnerUserId":1938262.0,"Title":"Order of addends in a sum affects the result?","Body":"<p>In the following code,  and  should be identical. The only difference is in the order the addends appear in the sum. However,  returns . Why does this happen? (I'm using TensorFlow version 0.12.1).<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41777496","CreationDate":"2017-01-21T09:04:09.410","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>It's happening because floating point operations are not accurate and changing the order can change the results albeit, the discrepancy is not significant.<\/p>\n\n<p>The problem here is with <code>np.array_equal<\/code>, the error between <code>res1<\/code> and <code>res2<\/code> is more than it can tolerate. Try printing <code>res1 - res2<\/code> to get a better idea. You might want to use <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.10.0\/reference\/generated\/numpy.allclose.html#numpy.allclose\" rel=\"nofollow noreferrer\">numpy.allclose<\/a> with properly chosen tolerance values.<\/p>\n"}]}
{"QuestionId":41777466,"AnswerCount":1,"Tags":"<python><tensorflow><keras><cudnn>","CreationDate":"2017-01-21T08:59:55.283","AcceptedAnswerId":null,"OwnerUserId":1105688.0,"Title":"Native TF vs Keras TF performance comparison","Body":"<p>I created the exact same network with native and backend tensorflow but after many hours of testing using number of different parameters, still couldn't figure out why keras outperforms the native tensorflow and produces better(slightly but better) results.<\/p>\n\n<p>Does Keras implement a different weight initializer method? or performs different weight decay approach other than tf.train.inverse_time_decay?<\/p>\n\n<p>P.s. the score difference is always like<\/p>\n\n<pre><\/pre>\n\n<p>My environment is:<\/p>\n\n<blockquote>\n  <p>Python 3.5.2 -Anaconda \/ Windows 10<br>\n  CUDA: 8.0 with cuDNN 5.1<br>\n  Keras 1.2.1<br>\n  Tensorflow 0.12.1<br>\n  Nvidia Geforce GTX 860M<\/p>\n<\/blockquote>\n\n<p>and <strong>keras.json<\/strong> file:<\/p>\n\n<pre><\/pre>\n\n<p>and you can also copy and execute following two files <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41792069","CreationDate":"2017-01-22T14:25:59.437","ParentId":null,"OwnerUserId":"1105688","Title":null,"Body":"<p>The problem was due to incorrect use of keep_prob parameter of the dropout layer as I should have fed this parameter with different values on train and test process.<\/p>\n"}]}
{"QuestionId":41778255,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2017-01-21T10:31:24.373","AcceptedAnswerId":null,"OwnerUserId":1509695.0,"Title":"cuda torch integration - CudaTensor class not found","Body":"<p>Probably, I'm not sufficiently familiar with the dependencies stack here, but I've installed a Deep Learning project that uses Torch and CUDA: <a href=\"https:\/\/github.com\/donglixp\/lang2logic\" rel=\"nofollow noreferrer\">https:\/\/github.com\/donglixp\/lang2logic<\/a>. <\/p>\n\n<p>Running the project (e.g. by issuing: ), I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>The last error by the way, is issued from this self-explanatory piece of lua code, which checks for the availability of the CudaTensor class:<\/p>\n\n<pre><\/pre>\n\n<p>After getting this error, I also installed Lua's  and  through the following commands, which seemed to have finished well.<\/p>\n\n<pre><\/pre>\n\n<p>But I still get the very same error.<\/p>\n\n<p>The only installation quirk I've noticed had been that  seemed to do nothing, and I'm not running on Scientific Linux as the original author, but rather on Ubuntu 16.04.<\/p>\n\n<p>I had installed CUDA itself through: <\/p>\n\n<p>How would you address this integration error?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42069356","CreationDate":"2017-02-06T13:47:36.653","ParentId":null,"OwnerUserId":"6073048","Title":null,"Body":"<p>I would have preferred leaving only a comment, but I do not have enough reputation for that yet so here goes.<\/p>\n\n<p>It seems like you have installation issues. The most obvious thing to try would be to uninstall torch entirely via <code>rm -rf ~\/torch<\/code> and reinstall it following your link's instructions (replacing ClassNLLCriterion.cu file with theirs). Usually, following the instruction <a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow noreferrer\">here<\/a> should install <code>cutorch<\/code> and <code>cunn<\/code> for you but you can always install them via <code>luarocks<\/code> as you mentioned. This is also why <code>luarocks install class<\/code> didn't do anything because it was already installed.<\/p>\n\n<p>In answer to your comment, <code>torch.CudaTensor<\/code> as the name suggests is a torch class defined in <a href=\"https:\/\/github.com\/torch\/cutorch\" rel=\"nofollow noreferrer\">cutorch<\/a>. It acts just like any other <code>Tensor<\/code> class from the torch library but it lives on GPU.<\/p>\n\n<p>More information on how you installed torch could help.<\/p>\n\n<p>You also of course need a CUDA compatible device in order to use cutorch.<\/p>\n"}]}
{"QuestionId":41778632,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><neural-network><artificial-intelligence><conv-neural-network>","CreationDate":"2017-01-21T11:13:21.007","AcceptedAnswerId":41778973.0,"OwnerUserId":4093621.0,"Title":"Why is the x variable tensor reshaped with -1 in the MNIST tutorial for tensorflow?","Body":"<p>I'm following the TensorFlow tutorial<\/p>\n\n<p>Initially x is defined as <\/p>\n\n<p><\/p>\n\n<p>Later on it reshapes x, I'm trying to understand why.<\/p>\n\n<blockquote>\n  <p>To apply the layer, we first reshape x to a 4d tensor, with the second and third dimensions corresponding to image width and height, and the final dimension corresponding to the number of color channels.<\/p>\n<\/blockquote>\n\n<p><\/p>\n\n<p>What does -1 mean in the reshaping vector and why is x being reshaped?<\/p>\n","answers":[{"AnswerId":"41778973","CreationDate":"2017-01-21T11:48:45.420","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>1)  <strong>What does -1 mean in the reshaping vector<\/strong><\/p>\n\n<p>From the documentation of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\">reshape<\/a>:<\/p>\n\n<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. In\n  particular, a shape of [-1] flattens into 1-D. At most one component\n  of shape can be -1.<\/p>\n<\/blockquote>\n\n<p>this is a standard feature and is available in numpy as well. Basically it means - I do not have time to calculate all the dimensions, so infer the one for me. In your case because <code>x * 28 * 28 * 1 = 784<\/code> so your -1 = 1<\/p>\n\n<p>2) <strong>Why is x being reshaped<\/strong><\/p>\n\n<p>They are planning to use convolution for image classification. So they need to use some spatial information. Current data is 1 dimensional. So they transform it to 4 dimensions. I do not know the point of the forth dimension because in my opinion they might have used only (x, y, color). Or even (x, y). Try to modify their reshape and convolution and most probably you will get similar accuracy.<\/p>\n"},{"AnswerId":"51610612","CreationDate":"2018-07-31T09:56:44.320","ParentId":null,"OwnerUserId":"4873087","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\">why 4 dimensions<\/a><\/p>\n\n<p>TensorFlow\u2019s convolutional conv2d operation expects a 4-dimensional tensor with dimensions corresponding to batch, width, height and channel.<\/p>\n\n<pre><code>[batch, in_height, in_width, in_channels]\n<\/code><\/pre>\n"}]}
{"QuestionId":41779307,"AnswerCount":1,"Tags":"<theano><keras><keras-layer>","CreationDate":"2017-01-21T12:24:35.247","AcceptedAnswerId":41782548.0,"OwnerUserId":2442740.0,"Title":"Keras - Theano - Test for division by zero","Body":"<p>I have a Layer that computes the mean of timesteps and supports masking. \nMy problem is that there may be the case that the mask is empty (no padded timesteps) but i don't know how to check for zeros when working with tensors.<\/p>\n\n<p>I have a few training examples for which the mask is empty so i get a NaN loss and the program crashes.<\/p>\n\n<p>This is my Layer:<\/p>\n\n<pre><\/pre>\n\n<p>This  becomes zero. In order to bypass this i have increased the input_length so it covers all my training examples, but this is not a solution. Also i tried adding a try\/except but this also didn't work.<\/p>\n","answers":[{"AnswerId":"41782548","CreationDate":"2017-01-21T17:49:36.160","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p><code>try\/except<\/code> wont work because all this piece of code does is create the symbolic tensor graph which has no exception .. the evaluation hence the division by 0 happens in the fit\/evaluate\/predict function. You need to include the logic\/decision in the symbolic graph. <\/p>\n\n<p>You can use <code>switch(condition, then_expression, else_expression)<\/code> to include if and else:<\/p>\n\n<pre><code>def call(self, x, mask=None):\n    if mask is not None:\n        sum = mask.sum(axis=1, keepdims=True)\n        cond = K.equal(sum,0)\n        _the_other_tensor_ = ....\n        div = K.switch(cond, _the_other_tensor_ ,sum)\n        return K.cast(x.sum(axis=1) \/ div, K.floatx()) # this may result to division by zero\n    else:\n        return K.mean(x, axis=1)\n<\/code><\/pre>\n\n<p>Or just use <code>clip(x, min_value, max_value)<\/code> to clip with a very small number epsilon to make the division numerically stable.<\/p>\n\n<pre><code>def call(self, x, mask=None):\n    if mask is not None:\n        sum = mask.sum(axis=1, keepdims=True)\n        div = K.clip(sum, K.epsilon, 1)\n        return K.cast(x.sum(axis=1) \/ div, K.floatx()) # this may result to division by zero\n    else:\n        return K.mean(x, axis=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41780344,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-21T14:11:59.577","AcceptedAnswerId":41782426.0,"OwnerUserId":6306884.0,"Title":"Gradient of tf.floor is None","Body":"<p> return None gradient<br>\nit means the weights before  operation won't be update, right?<br>\nbut I still need the gradient to update weights<br>\nIsn't it weird? sometimes we use , ... but they can't deliver gradient or this is right action<br>\nas well as , the gradient return none <\/p>\n\n<p>Here is an issue <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/897\" rel=\"nofollow noreferrer\">#897<\/a> to discuss the gradient, but I don't understand why return None is good<\/p>\n\n<pre><\/pre>\n\n<p>Can I modify  directly? need other action?<br>\nsomething like delete ?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41782426","CreationDate":"2017-01-21T17:36:51.547","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>TensorFlow uses <code>None<\/code> to represent <code>0<\/code> for implementation reasons. I don't think it would help you if the returned gradient was true gradient <code>0<\/code>, since you want to train. You could use <code>gradient_override_map<\/code> to substitute gradient of <code>Identity<\/code> op for <code>Floor<\/code> as follows<\/p>\n\n<pre><code>tf.reset_default_graph()\nx = tf.Variable(10.)\nwith tf.get_default_graph().gradient_override_map({\"Floor\": \"Identity\"}):\n    x2 = tf.floor(x)\nloss = tf.square(x2)\nopt = tf.train.GradientDescentOptimizer(0.1)\ntrain_op = opt.minimize(loss)\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nfor i in range(10):\n    print(sess.run([loss, train_op]))\n<\/code><\/pre>\n"}]}
{"QuestionId":41780655,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-21T14:44:13.177","AcceptedAnswerId":41782361.0,"OwnerUserId":7287271.0,"Title":"What is the difference between tf.group and tf.control_dependencies?","Body":"<p>Aside from  being a context manager (i.e. used with Python ), what's the difference between  and ? <\/p>\n\n<p>When should which be used? <\/p>\n\n<p>Is it that  doesn't have any particular order of operations? I'd assume  executes ops in the list's order, but maybe that's not the case? The docstring doesn't specify a behaviour.<\/p>\n","answers":[{"AnswerId":"49602463","CreationDate":"2018-04-01T20:46:51.320","ParentId":null,"OwnerUserId":"336527","Title":null,"Body":"<p>Just adding a few minor points to @Yaroslav-Bulatov answer.<\/p>\n\n<p>As you can see from Yaroslav's answer:<\/p>\n\n<ul>\n<li><code>tf.control_depenencies<\/code> creates no ops by itself, and adds dependencies to whatever ops your create inside its scope<\/li>\n<li><code>tf.group<\/code> creates a single op (of type <code>NoOp<\/code>), adds dependencies to that op.<\/li>\n<\/ul>\n\n<p>More importantly, if <code>tf.group<\/code> arguments belong to multiple devices, <code>tf.group<\/code> will insert an intermediate layer between its inputs and the node it returns. That layer will contain one node per device, so that the dependencies are organized by device. This could reduce the cross-device data flow.<\/p>\n\n<p>So if your dependencies are on multiple devices, <code>tf.group<\/code> adds a (possibly critical) optimization.<\/p>\n\n<p>On the other hand, <code>tf.control_dependencies<\/code> supports nesting: the inner context will add dependencies to the union of all the ops in the outer contexts.<\/p>\n"},{"AnswerId":"41782361","CreationDate":"2017-01-21T17:29:52.330","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>If you look at the graphdef, the <code>c=tf.group(a, b)<\/code> produces the same graph as <\/p>\n\n<pre><code>with tf.control_dependencies([a, b]):\n    c = tf.no_op() \n<\/code><\/pre>\n\n<p>There's no specific order in which ops will run, TensorFlow tries to execute operations as soon as it can (i.e. in parallel).<\/p>\n"}]}
{"QuestionId":41780900,"AnswerCount":2,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-01-21T15:06:55.877","AcceptedAnswerId":41784598.0,"OwnerUserId":2736559.0,"Title":"Does using Fp16 in deeplearning have adverse effect on the end result?","Body":"<p>I see tensorflow offers the use of  in training and testing, is it safe to use it or will it have an adverse effect on the final result? <\/p>\n","answers":[{"AnswerId":"41784598","CreationDate":"2017-01-21T21:15:55.847","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>It will affect the output while training, because of the extra math precision that float32 provides, but after training you can 'quantize' the operations in your network to float16 to have faster performances if your hardware supports the float16 natively. If the hardware does not support such operation you might likely have a slow down in terms of performances.<\/p>\n"},{"AnswerId":"43291162","CreationDate":"2017-04-08T06:59:12.313","ParentId":null,"OwnerUserId":"7836065","Title":null,"Body":"<p>according to wikipedia data,the half precision float point performance on nvidia cards is even worse than using double, and i am wondoring performance reports to confirm that idea(<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1300\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1300<\/a>,see the bottom of the page, a 10% performance was reported).\nhence we can conclude that using fp16 is centered on the video ram issue(or,scale of model),training was to be carried out normally, \nit is OK to use lower precision, binary connection was also investigated. fit would be archived anyway\nIn contrast, the amd cards can offer astonishing fp16 performance while software support is rare. e.g. theano and the libgpuarray backgroud on openCL.though...people use fp32 to calculate<\/p>\n"}]}
{"QuestionId":41781294,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><scikit-image>","CreationDate":"2017-01-21T15:49:13.753","AcceptedAnswerId":null,"OwnerUserId":1417053.0,"Title":"Using 32-bit Lab images in Tensorflow","Body":"<p>I'm using  to convert RGB images to Lab colorspace but it seems that skimage uses  datatype while Tensorflow uses . <\/p>\n\n<p>Is there any way to convert the 64-bit Lab image to 32-bit datatype? The documentation doesn't cover anything specific about this and I'm not sure if using  is the way to go since it might damage the data precision (or not).<\/p>\n\n<p>Here's a part of the code:<\/p>\n\n<pre><\/pre>\n\n<p>The last line gives an error:<\/p>\n\n<pre><\/pre>\n\n<p>and here's the Stack Trace:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41783649","CreationDate":"2017-01-21T19:37:19.140","ParentId":null,"OwnerUserId":"214686","Title":null,"Body":"<p>Using <code>x.astype(np.float32)<\/code> is perfectly acceptable.  You'll seldom if ever need that level of accuracy.<\/p>\n\n<p>If you are not careful, however, you may accidentally cast an integer image (e.g., unsigned bytes, going from 0 to 255) into float.  So the safest approach, that will rescale as necessary, is<\/p>\n\n<pre><code>from skimage import img_as_float\nimage = img_as_float(image).astype(np.float32)\n<\/code><\/pre>\n"},{"AnswerId":"54156169","CreationDate":"2019-01-12T01:59:44.663","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>This was a difficult bug to track down.<\/p>\n\n<p>If you're using float32 then you get your error as described; however, the error will go away if you use float64 values instead.  <\/p>\n\n<p>Hope that helps! :)<\/p>\n\n<p>Known bug source:\n<a href=\"https:\/\/github.com\/scikit-image\/scikit-image\/issues\/2546\" rel=\"nofollow noreferrer\">Scikit-image Issue<\/a><\/p>\n"}]}
{"QuestionId":41782812,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-21T18:18:22.253","AcceptedAnswerId":41783019.0,"OwnerUserId":2671627.0,"Title":"How do I find alternative methods in tensorflow latest release to deprecated one like tf.image_summary used in some tensorflow official tutorials?","Body":"<p>I am new to tensorflow. While I was reading the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">CNN tutorial<\/a>, I found a broken link to a deprecated method <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train#image_summary\" rel=\"nofollow noreferrer\">image_summary<\/a>.<\/p>\n\n<ul>\n<li>What is the best practice to follow in such situation? <\/li>\n<li>Shall I try to inform the tensorflow team about the broken links in their tutorials? \n\n<ul>\n<li>if so, what is the best channel to do so?<\/li>\n<\/ul><\/li>\n<li>How shall I find the best alternative to the deprecated method in their latest release?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"41783019","CreationDate":"2017-01-21T18:37:33.527","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>The safest way might be a quick search into the Tensorflow Github repo. E.g. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/search?utf8=%E2%9C%93&amp;q=image+summary\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/search?utf8=%E2%9C%93&amp;q=image+summary<\/a>, where you'll see it's been renamed to tf.summary.image.<\/p>\n\n<p>Surely it'll be much appreciated if you let the team know. I think the best way is to raise an issue here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues<\/a><\/p>\n"}]}
{"QuestionId":41783136,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-21T18:48:26.743","AcceptedAnswerId":41784379.0,"OwnerUserId":541106.0,"Title":"Tensorflow batch size in input placholder","Body":"<p>I am new to Tensorflow and I can't get why the input placeholder is often dimensioned with the size of the batches used for training. <\/p>\n\n<p>In this example I found <a href=\"https:\/\/martin-thoma.com\/tensor-flow-quick\/\" rel=\"nofollow noreferrer\">here<\/a> and in the Official Mnist tutorial it is not<\/p>\n\n<pre><\/pre>\n\n<p>So what is the best and right way to dimension and create the model input and train it?<\/p>\n","answers":[{"AnswerId":"41784379","CreationDate":"2017-01-21T20:52:34.597","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Here you are specifying the model input. You want to leave the Batch size, to <code>None<\/code>, that means that you can run the model with a variable number of inputs (one or more). Batching is important to efficiently use your computing resources.<\/p>\n\n<pre><code>x = tf.placeholder(\"float\", shape=[None, 784])\ny_ = tf.placeholder(\"float\", shape=[None, 10])\n<\/code><\/pre>\n\n<p>The next important line is:<\/p>\n\n<pre><code>batch = mnist.train.next_batch(50)\n<\/code><\/pre>\n\n<p>Here you are sending 50 elements as input but you can also change that to just one <\/p>\n\n<pre><code>batch = mnist.train.next_batch(1)\n<\/code><\/pre>\n\n<p>Without modifying the graph. If you specify the Batch size (some number instead of None in the first snippet), then you would have to change each time and that is not ideal, specially in production.<\/p>\n"}]}
{"QuestionId":41786853,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2017-01-22T02:56:44.740","AcceptedAnswerId":null,"OwnerUserId":7452223.0,"Title":"tensorflow porting to android devices","Body":"<p>I have a question about tensorflow porting to android devices. When I was in centos, using bazel compiler generated apk file, suggesting the following error:<\/p>\n\n<p>ERROR: \/data\/tensorflow-master\/WORKSPACE:37:1: syntax error at 'tf_workspace': expected,\nERROR: \/data\/tensorflow-master\/WORKSPACE:44:1: non-keyword arg after keyword arg.\nERROR: error loading package 'external': Failed to parse WORKSPACE file.\nINFO: Elapsed time: 0.242s<\/p>\n\n<p>The following is the 37-44 line of the WORKSPACE file:<\/p>\n\n<p>Please add all new TensorFlow dependencies in workspace.bzl.<\/p>\n\n<p>37 tf_workspace () <\/p>\n\n<p>38<\/p>\n\n<p>39 new_http_archive (<\/p>\n\n<p>40 name = \"inception5h\",<\/p>\n\n<p>41 build_file = \"models.BUILD\",<\/p>\n\n<p>42 url = \"<a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip<\/a>\",<\/p>\n\n<p>43 sha256 = \"d13569f6a98159de37e92e9c8ec4dae8f674fbf475f69fe6199b514f756d4364\"<\/p>\n\n<p>44)<\/p>\n\n<p>I ask you have encountered similar problems? please help me.<\/p>\n","answers":[{"AnswerId":"44917920","CreationDate":"2017-07-05T05:51:30.617","ParentId":null,"OwnerUserId":"8256998","Title":null,"Body":"<p>I was facing the same problem, this is how I fixed it:<\/p>\n\n<pre><code># Uncomment and update the paths in these entries to build the Android demo.\nandroid_sdk_repository(\n    name = \"androidsdk\",\n    api_level = 24,\n#    # Ensure that you have the build_tools_version below installed in the\n#    # SDK manager as it updates periodically.\n    build_tools_version = \"25.0.3\",\n#    # Replace with path to Android SDK on your system\n    path = \"\/home\/yours\/Android\/Sdk\/\",\n) #my mistake #) --&gt; )\n#\n# Android NDK r12b is recommended (higher may cause issues with Bazel)\nandroid_ndk_repository(\n    name=\"androidndk\",\n    path=\"\/home\/yours\/android-ndk-r12b\/\",\n#    # This needs to be 14 or higher to compile TensorFlow.\n#    # Note that the NDK version is not the API level.\n    api_level=14)\n<\/code><\/pre>\n"}]}
{"QuestionId":41787480,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-22T04:51:16.857","AcceptedAnswerId":null,"OwnerUserId":2454808.0,"Title":"Running distributed tensorflow on machines with different versions of Tensorflow","Body":"<p>In particular I'm interested in what happens if the tensorflow installed on the parameter server is different from the tensorflow installed on the worker. (E.G: different versions of tensorflow). How does distributed tensorflow reconcile potential differences in the execution code between workers and parameter servers?<\/p>\n","answers":[{"AnswerId":"41788776","CreationDate":"2017-01-22T08:18:55.580","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>I'd say it's highly discouraged. Depending on how many versions they differ, in good cases, you get some errors. In base cases, things seemingly run normally but e.g. nothing trains and you'll have a hard time figuring out why....<\/p>\n"}]}
{"QuestionId":41787559,"AnswerCount":1,"Tags":"<python><flask><keras>","CreationDate":"2017-01-22T05:08:23.993","AcceptedAnswerId":41787609.0,"OwnerUserId":5636645.0,"Title":"How to display dynamically and image?","Body":"<p>Hello I am making an small page to show my results I am working in a project about sentiment analysis first I have the following labels:<\/p>\n\n<pre><\/pre>\n\n<p>I show this labels depending of the result of a predict function that I performed using keras, at this moment all is working well I wish to show an image depending of the label of above I tried creating an array with the path of the images as follows, I am not sure how to write the image function, <\/p>\n\n<pre><\/pre>\n\n<p>This is the function that perform the predict, at this moment it is just showing a label of above, I would like to also display a distinct image, so I need to modify the following function:<\/p>\n\n<pre><\/pre>\n\n<p>Since I am a beginner at flask I would like to appreciate support or suggestions to overcome this situation thanks for the help, <\/p>\n\n<p>After a very useful feedback I tried:<\/p>\n\n<pre><\/pre>\n\n<p>I am not getting any error but the image is not displayed I am not so sure what is wrong, at this moment I am just trying with one image located at the same level of my file called app.py, smile.jpg <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41787609","CreationDate":"2017-01-22T05:17:03.060","ParentId":null,"OwnerUserId":"790387","Title":null,"Body":"<p>Just create a dictionary of your keys and the image values; and use that to return the image for the particular sentiment:<\/p>\n\n<pre><code>&gt;&gt;&gt; senti=[\"furious\",\"angry\",\"angry0\",\"Indiferent\",\"happy\",\"enthusiastic\",\"Euphoric\"]\n&gt;&gt;&gt; images=['home\/image0.jpg','home\/image1.jpg','home\/image2.jpg','home\/image3.jpg','home\/image4.jpg','home\/image5.jpg','home\/image6.jpg']\n&gt;&gt;&gt; dict(zip(senti, images))\n{'enthusiastic': 'home\/image5.jpg', 'Indiferent': 'home\/image3.jpg', 'furious': 'home\/image0.jpg', 'Euphoric': 'home\/image6.jpg', 'angry': 'home\/image1.jpg', 'happy': 'home\/image4.jpg', 'angry0': 'home\/image2.jpg'}\n&gt;&gt;&gt; lookup_values = dict(zip(senti, images))\n&gt;&gt;&gt; lookup_values['angry']\n'home\/image1.jpg'\n<\/code><\/pre>\n\n<p>You can use this in your view method, to get the right image path and the send it to the template:<\/p>\n\n<pre><code>@app.route(\"\/\", methods=['GET', 'POST'])\ndef index():\n    senti=[\"furious\",\"angry\",\"angry0\",\"Indiferent\",\"happy\",\"enthusiastic\",\"Euphoric\"]\n    images=['home\/image0.jpg','home\/image1.jpg','home\/image2.jpg','home\/image3.jpg','home\/image4.jpg','home\/image5.jpg','home\/image6.jpg']\n    lookup_keys = dict(zip(senti, images))\n    print(request.method)\n    if request.method == 'POST':\n        q=request.form['querytext']\n        prediction=predict(q)\n        image_path = lookup_keys[prediction] # get the path\n        return render_template(\"result.html\",\n                               prediction=prediction,\n                               text=q,\n                               image_url=image_path)\n    return render_template(\"main.html\")\n<\/code><\/pre>\n"}]}
{"QuestionId":41787873,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><theano><keras>","CreationDate":"2017-01-22T06:07:40.597","AcceptedAnswerId":41795508.0,"OwnerUserId":5287146.0,"Title":"How Adagrad works in Keras? What does self.weights mean in Keras Optimizer?","Body":"<p>For example, the implementation of Keras' Adagrad has been:<\/p>\n\n<pre><\/pre>\n\n<p>And the Function 'get_update()' seems one step update. However should the accumulators be stored the history information? Why it has been initialized to zeros at each step? How it can be an accumulator through the whole training process?<\/p>\n\n<p>What does this line do?<\/p>\n\n<pre><\/pre>\n\n<p>It seems self.weights is never been called anymore.<\/p>\n","answers":[{"AnswerId":"41795508","CreationDate":"2017-01-22T19:58:06.440","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You are correct.. for all optimizers in Keras <code>get_updates()<\/code> implements the tensor logic for one step of updates. This function is called once for each <code>model.fit()<\/code> from <code>_make_train_function()<\/code> <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/engine\/training.py#L759-L761\" rel=\"nofollow noreferrer\">here<\/a>, which is used to create the tensor function by passing the update rule as <code>update=<\/code> <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/engine\/training.py#L765-L768\" rel=\"nofollow noreferrer\">here<\/a>. This update rule is used iteration to iteration to update the model parameters and other parameters.<\/p>\n\n<p><code>self.weights<\/code> of an optimizer class is its internal parameters. This is not used for training. It just functions to keep the state of the optimizer (list of pointers to the param\/accumulators tensor) and when <code>model.save<\/code> is called they are also saved by calling <code>get_weights()<\/code> <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/optimizers.py#L118-L124\" rel=\"nofollow noreferrer\">here<\/a> and is loaded back when <code>model.load<\/code> is called by <code>set_weights()<\/code> <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/optimizers.py#L90-L116\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
{"QuestionId":41788319,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><deep-learning><caffe>","CreationDate":"2017-01-22T07:15:30.667","AcceptedAnswerId":41788607.0,"OwnerUserId":5576762.0,"Title":"What is a .net file in caffe?","Body":"<p>I want to reduce the size of my .caffemodel file.<\/p>\n\n<p>I came across Deep Compression (<a href=\"https:\/\/github.com\/songhan\/Deep-Compression-AlexNet\" rel=\"nofollow noreferrer\">songhan\/Deep-Compression-AlexNet<\/a>).<\/p>\n\n<p>This creates a .net file. How do I use it for predictions? Or how to I convert it to a .caffemodel file?<\/p>\n","answers":[{"AnswerId":"41788607","CreationDate":"2017-01-22T07:54:49.460","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>See the <a href=\"https:\/\/github.com\/songhan\/Deep-Compression-AlexNet#usage\" rel=\"nofollow noreferrer\">Usage<\/a> section in the git repository you are using:<\/p>\n\n<blockquote>\n<pre><code>python decode.py bvlc_alexnet_deploy.prototxt AlexNet_compressed.net $CAFFE_ROOT\/alexnet.caffemodel\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Use the provided <code>decode.py<\/code> script to convert the compressed <code>'.net'<\/code> binary file to a usable <code>'.caffemodel'<\/code> you can use for deploying your net.<\/p>\n"}]}
{"QuestionId":41788431,"AnswerCount":2,"Tags":"<python><variables><tensorflow><size><shape>","CreationDate":"2017-01-22T07:31:03.780","AcceptedAnswerId":41788745.0,"OwnerUserId":6667826.0,"Title":"TypeError: 'Tensor' object cannot be interpreted as an integer","Body":"<p>I am trying to run a loop based on the size of the array. how to do that in tensorflow ? For example <\/p>\n\n<pre><\/pre>\n\n<p>I wanted to print the array a for 5 times. but it says loop is a tensor object and cannot taken as integer. \nI have tried taking the loop variable as <\/p>\n\n<pre><\/pre>\n\n<p>it has the same error. <\/p>\n","answers":[{"AnswerId":"41788744","CreationDate":"2017-01-22T08:14:55.777","ParentId":null,"OwnerUserId":"6186255","Title":null,"Body":"<p><code>tf.size()<\/code> does not give you a value or list.<\/p>\n\n<pre><code>a = tf.Variable([1,2,3,4,5],dtype = tf.int32)\n\nv = a.get_shape()\nloop = v.num_elements()\n\n...\n<\/code><\/pre>\n\n<p>Perhaps, try this.<\/p>\n"},{"AnswerId":"41788745","CreationDate":"2017-01-22T08:15:08.700","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>Not really sure what you want to achieve here. <code>loop<\/code> is a <code>tf.Tensor<\/code> and <code>range<\/code> expects an <code>integer<\/code> as argument, hence the error. If you just want to print a 5 times, why don't you just set loop to the numerical value of 5?<\/p>\n\n<p>Otherwise, the following code should work, as <code>loop.eval()<\/code> returns the value of <code>loop<\/code> which is 5:<\/p>\n\n<pre><code>a = tf.Variable([1,2,3,4,5],dtype = tf.int32)\nloop = tf.size(a)\n....\nfor i in range(loop.eval()):\n    print(sess.run(a))\n<\/code><\/pre>\n\n<p>If you don't want to execute the TF graph multiple times, take a look at <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#while_loop\" rel=\"nofollow noreferrer\">tf.while_loop<\/a>.<\/p>\n"}]}
{"QuestionId":41788538,"AnswerCount":1,"Tags":"<flask><keras>","CreationDate":"2017-01-22T07:44:47.517","AcceptedAnswerId":41788897.0,"OwnerUserId":5636645.0,"Title":"how to display the following images?","Body":"<p>first I defined two functions in my app.py file as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Here I get the values of  image_url that is the path of my image as follows:<\/p>\n\n<pre><\/pre>\n\n<p>then depending of the result I would like to display and image I am just using two smile.jpg and angry.jpg that are contained in the static folder, \nafter that I render my template called: \"result.html\"<\/p>\n\n<p>this file contains the following tags to show me the text and prediction as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Everything works well until this point I want to show the image dynamically I tried:<\/p>\n\n<pre><\/pre>\n\n<p>However is not displaying nothing, If I put the following tag:<\/p>\n\n<pre><\/pre>\n\n<p>the image is being displayed with any problems I would like to appreciate support to modify the tag of above to display the image dinamically I mean this part:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to change dinamically this string 'static\/angry.jpg' with the value of my variable called image_url, I am a begginer using flask so I am not sure how to use this value, thanks for the support.<\/p>\n","answers":[{"AnswerId":"41788897","CreationDate":"2017-01-22T08:35:59.880","ParentId":null,"OwnerUserId":"5802335","Title":null,"Body":"<p>If you sure your image_url path is fine then:    <\/p>\n\n<pre><code>&lt;td&gt;&lt;img src=\"{{image_url}}\" alt=\"\"&gt;&lt;\/td&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":41788924,"AnswerCount":1,"Tags":"<python><c++><tensorflow>","CreationDate":"2017-01-22T08:39:07.413","AcceptedAnswerId":null,"OwnerUserId":3835345.0,"Title":"How to calculate ctc probability for given input and expected output?","Body":"<p>I'm doing my first tensorflow project.<\/p>\n\n<p>I need to get ctc probability (not ctc loss) for given input and my expected sequences.<\/p>\n\n<p>Is there any api or ways to do it in python or c++?<\/p>\n\n<p>I prefer python side, but c++ side is also okay.<\/p>\n","answers":[{"AnswerId":"47182528","CreationDate":"2017-11-08T14:41:46.420","ParentId":null,"OwnerUserId":"2679594","Title":null,"Body":"<p>according to Graves paper [1], the loss for a batch is defined as sum(log(p(z|x))) over all samples (x,z) in this batch.\nIf you use a batch size of 1, you get log(p(z|x)), that is the log-probability of seeing the labelling z given the input x. This can be achieved with the ctc_loss function from TensorFlow.<\/p>\n\n<p>You can also implement the relevant parts of the Forward-Backward Algorithm described in Section 4.1 of the paper [1] yourself.\nFor small input sequences it is feasible to use a naive implementation by constructing the paths shown in Figure 3 and then sum over all that paths in the RNN output.\nI did this for a sequence of length 16 and for a sequence of length 100. For the former one the naive approach was sufficient while for the latter one the presented dynamic programming approach was needed.<\/p>\n\n<p>[1] Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks<\/p>\n"}]}
{"QuestionId":41789133,"AnswerCount":4,"Tags":"<python><tensorflow><deep-learning><lstm>","CreationDate":"2017-01-22T09:04:46.397","AcceptedAnswerId":41928428.0,"OwnerUserId":5299052.0,"Title":"What are c_state and m_state in Tensorflow LSTM?","Body":"<p>Tensorflow r0.12's documentation for tf.nn.rnn_cell.LSTMCell describes this as the init:<\/p>\n\n<pre><\/pre>\n\n<p>where  is as follows:<\/p>\n\n<blockquote>\n  <p>state: if state_is_tuple is False, this must be a state Tensor, 2-D, batch x state_size. If state_is_tuple is True, this must be a tuple of state Tensors, both 2-D, with column sizes c_state and m_state.<\/p>\n<\/blockquote>\n\n<p>What aare  and  and how do they fit into LSTMs? I cannot find reference to them anywhere in the documentation.<\/p>\n\n<p><a href=\"https:\/\/web.archive.org\/web\/20170223030652\/https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/rnn_cell\/rnn_cells_for_use_with_tensorflow_s_core_rnn_methods\" rel=\"nofollow noreferrer\">Here is a link to that page in the documentation.<\/a><\/p>\n","answers":[{"AnswerId":"41928428","CreationDate":"2017-01-30T02:52:52.270","ParentId":null,"OwnerUserId":"2283175","Title":null,"Body":"<p>I've stumbled upon same question, here's how I understand it! Minimalistic LSTM example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nsample_input = tf.constant([[1,2,3]],dtype=tf.float32)\n\nLSTM_CELL_SIZE = 2\n\nlstm_cell = tf.nn.rnn_cell.BasicLSTMCell(LSTM_CELL_SIZE, state_is_tuple=True)\nstate = (tf.zeros([1,LSTM_CELL_SIZE]),)*2\n\noutput, state_new = lstm_cell(sample_input, state)\n\ninit_op = tf.global_variables_initializer()\n\nsess = tf.Session()\nsess.run(init_op)\nprint sess.run(output)\n<\/code><\/pre>\n\n<p>Notice that <code>state_is_tuple=True<\/code> so when passing <code>state<\/code> to this <code>cell<\/code>, it needs to be in the <code>tuple<\/code> form. <code>c_state<\/code> and <code>m_state<\/code> are probably \"Memory State\" and \"Cell State\", though I honestly am NOT sure, as these terms are only mentioned in the docs. In the code and papers about <code>LSTM<\/code> - letters <code>h<\/code> and <code>c<\/code> are commonly used to denote \"output value\" and \"cell state\".\n<a href=\"http:\/\/colah.github.io\/posts\/2015-08-Understanding-LSTMs\/\" rel=\"noreferrer\">http:\/\/colah.github.io\/posts\/2015-08-Understanding-LSTMs\/<\/a>\nThose tensors represent combined internal state of the cell, and should be passed together. Old way to do it was to simply concatenate them, and new way is to use tuples.<\/p>\n\n<p>OLD WAY:<\/p>\n\n<pre><code>lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(LSTM_CELL_SIZE, state_is_tuple=False)\nstate = tf.zeros([1,LSTM_CELL_SIZE*2])\n\noutput, state_new = lstm_cell(sample_input, state)\n<\/code><\/pre>\n\n<p>NEW WAY:<\/p>\n\n<pre><code>lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(LSTM_CELL_SIZE, state_is_tuple=True)\nstate = (tf.zeros([1,LSTM_CELL_SIZE]),)*2\n\noutput, state_new = lstm_cell(sample_input, state)\n<\/code><\/pre>\n\n<p>So, basically all we did, is changed <code>state<\/code> from being 1 tensor of length <code>4<\/code> into two tensors of length <code>2<\/code>. The content remained the same. <code>[0,0,0,0]<\/code> becomes <code>([0,0],[0,0])<\/code>. (This is supposed to make it faster)<\/p>\n"},{"AnswerId":"45233303","CreationDate":"2017-07-21T08:53:40.203","ParentId":null,"OwnerUserId":"8343303","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/python\/ops\/rnn_cell_impl.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/python\/ops\/rnn_cell_impl.py<\/a><\/p>\n\n<p>Line #308 - 314<\/p>\n\n<p>class LSTMStateTuple(_LSTMStateTuple):\n  \"\"\"Tuple used by LSTM Cells for <code>state_size<\/code>, <code>zero_state<\/code>, and output state.\n  Stores two elements: <code>(c, h)<\/code>, in that order.\n  Only used when <code>state_is_tuple=True<\/code>.\n  \"\"\"<\/p>\n"},{"AnswerId":"42360236","CreationDate":"2017-02-21T06:32:46.620","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>I agree that the documentation is unclear. Looking at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/66d5d1fa0c192ca4c9b75cde216866805eb160f2\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py#L291-L382\" rel=\"noreferrer\"><code>tf.nn.rnn_cell.LSTMCell.__call__<\/code><\/a> clarifies (I took the code from TensorFlow 1.0.0):<\/p>\n\n<pre><code>def __call__(self, inputs, state, scope=None):\n    \"\"\"Run one step of LSTM.\n\n    Args:\n      inputs: input Tensor, 2D, batch x num_units.\n      state: if `state_is_tuple` is False, this must be a state Tensor,\n        `2-D, batch x state_size`.  If `state_is_tuple` is True, this must be a\n        tuple of state Tensors, both `2-D`, with column sizes `c_state` and\n        `m_state`.\n      scope: VariableScope for the created subgraph; defaults to \"lstm_cell\".\n\n    Returns:\n      A tuple containing:\n\n      - A `2-D, [batch x output_dim]`, Tensor representing the output of the\n        LSTM after reading `inputs` when previous state was `state`.\n        Here output_dim is:\n           num_proj if num_proj was set,\n           num_units otherwise.\n      - Tensor(s) representing the new state of LSTM after reading `inputs` when\n        the previous state was `state`.  Same type and shape(s) as `state`.\n\n    Raises:\n      ValueError: If input size cannot be inferred from inputs via\n        static shape inference.\n    \"\"\"\n    num_proj = self._num_units if self._num_proj is None else self._num_proj\n\n    if self._state_is_tuple:\n      (c_prev, m_prev) = state\n    else:\n      c_prev = array_ops.slice(state, [0, 0], [-1, self._num_units])\n      m_prev = array_ops.slice(state, [0, self._num_units], [-1, num_proj])\n\n    dtype = inputs.dtype\n    input_size = inputs.get_shape().with_rank(2)[1]\n    if input_size.value is None:\n      raise ValueError(\"Could not infer input size from inputs.get_shape()[-1]\")\n    with vs.variable_scope(scope or \"lstm_cell\",\n                           initializer=self._initializer) as unit_scope:\n      if self._num_unit_shards is not None:\n        unit_scope.set_partitioner(\n            partitioned_variables.fixed_size_partitioner(\n                self._num_unit_shards))\n      # i = input_gate, j = new_input, f = forget_gate, o = output_gate\n      lstm_matrix = _linear([inputs, m_prev], 4 * self._num_units, bias=True,\n                            scope=scope)\n      i, j, f, o = array_ops.split(\n          value=lstm_matrix, num_or_size_splits=4, axis=1)\n\n      # Diagonal connections\n      if self._use_peepholes:\n        with vs.variable_scope(unit_scope) as projection_scope:\n          if self._num_unit_shards is not None:\n            projection_scope.set_partitioner(None)\n          w_f_diag = vs.get_variable(\n              \"w_f_diag\", shape=[self._num_units], dtype=dtype)\n          w_i_diag = vs.get_variable(\n              \"w_i_diag\", shape=[self._num_units], dtype=dtype)\n          w_o_diag = vs.get_variable(\n              \"w_o_diag\", shape=[self._num_units], dtype=dtype)\n\n      if self._use_peepholes:\n        c = (sigmoid(f + self._forget_bias + w_f_diag * c_prev) * c_prev +\n             sigmoid(i + w_i_diag * c_prev) * self._activation(j))\n      else:\n        c = (sigmoid(f + self._forget_bias) * c_prev + sigmoid(i) *\n             self._activation(j))\n\n      if self._cell_clip is not None:\n        # pylint: disable=invalid-unary-operand-type\n        c = clip_ops.clip_by_value(c, -self._cell_clip, self._cell_clip)\n        # pylint: enable=invalid-unary-operand-type\n\n      if self._use_peepholes:\n        m = sigmoid(o + w_o_diag * c) * self._activation(c)\n      else:\n        m = sigmoid(o) * self._activation(c)\n\n      if self._num_proj is not None:\n        with vs.variable_scope(\"projection\") as proj_scope:\n          if self._num_proj_shards is not None:\n            proj_scope.set_partitioner(\n                partitioned_variables.fixed_size_partitioner(\n                    self._num_proj_shards))\n          m = _linear(m, self._num_proj, bias=False, scope=scope)\n\n        if self._proj_clip is not None:\n          # pylint: disable=invalid-unary-operand-type\n          m = clip_ops.clip_by_value(m, -self._proj_clip, self._proj_clip)\n          # pylint: enable=invalid-unary-operand-type\n\n    new_state = (LSTMStateTuple(c, m) if self._state_is_tuple else\n                 array_ops.concat([c, m], 1))\n    return m, new_state\n<\/code><\/pre>\n\n<p>The key lines are:<\/p>\n\n<pre><code>c = (sigmoid(f + self._forget_bias) * c_prev + sigmoid(i) *\n         self._activation(j))\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>m = sigmoid(o) * self._activation(c)\n<\/code><\/pre>\n\n<p>and <\/p>\n\n<pre><code>new_state = (LSTMStateTuple(c, m) \n<\/code><\/pre>\n\n<p>If you compare the  code to compute <code>c<\/code> and <code>m<\/code> with the LSTM equations (see below), you can see it corresponds to the cell state (typically denoted with <code>c<\/code>) and hidden state (typically denoted with <code>h<\/code>), respectively:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/fFh9H.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/fFh9H.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><code>new_state = (LSTMStateTuple(c, m)<\/code> indicates that the first element of the returned state tuple is <code>c<\/code> (cell state a.k.a. <code>c_state<\/code>), and the second element of the returned state tuple is <code>m<\/code> (hidden state a.k.a. <code>m_state<\/code>).<\/p>\n"},{"AnswerId":"41796717","CreationDate":"2017-01-22T22:12:08.950","ParentId":null,"OwnerUserId":"2904506","Title":null,"Body":"<p>Maybe this excerpt from the code will help<\/p>\n\n<pre><code>def __call__(self, inputs, state, scope=None):\n  \"\"\"Long short-term memory cell (LSTM).\"\"\"\n  with vs.variable_scope(scope or type(self).__name__):  # \"BasicLSTMCell\"\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(1, 2, state)\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(1, 4, concat)\n\n    new_c = (c * sigmoid(f + self._forget_bias) + sigmoid(i) *\n             self._activation(j))\n    new_h = self._activation(new_c) * sigmoid(o)\n\n    if self._state_is_tuple:\n      new_state = LSTMStateTuple(new_c, new_h)\n    else:\n      new_state = array_ops.concat(1, [new_c, new_h])\n    return new_h, new_state\n<\/code><\/pre>\n"}]}
{"QuestionId":41789209,"AnswerCount":1,"Tags":"<python><if-statement><math><tensorflow>","CreationDate":"2017-01-22T09:16:12.387","AcceptedAnswerId":41789381.0,"OwnerUserId":6667826.0,"Title":"how to define condition statement in tensorflow?","Body":"<p>I have three functions in file fe_extraction.py<\/p>\n\n<pre><\/pre>\n\n<p>I want to simply apply a condition if y > thershold  perform rms_freq(x) or else peak_value(x) and return that value.\n y is the value obtained from another function. <\/p>\n\n<pre><\/pre>\n\n<p>I get error as TypeError:fn1 must be callable. <\/p>\n","answers":[{"AnswerId":"41789381","CreationDate":"2017-01-22T09:39:11.100","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>With <code>rms_freq(x)<\/code> and <code>peak_value(x)<\/code> you're <strong>calling<\/strong> the function <code>rms_freq<\/code> and <code>peak_value<\/code> respectively,  passing <code>x<\/code> as argument.<\/p>\n\n<p>Instead, you have to pass a callable or, in other words, a function, that <code>tf.cond<\/code> can execute.<\/p>\n\n<p>Since you want <code>x<\/code> as a parameter for your functions, you can wrap them in a <code>lambda<\/code> that defines a callable object that captures the outside scope and thus sees the parameter <code>x<\/code>.<\/p>\n\n<pre><code>z = tf.cond(y&gt;Thersh,lambda: rms_freq(x) ,lambda: peak_value(x))\n<\/code><\/pre>\n"}]}
{"QuestionId":41789267,"AnswerCount":1,"Tags":"<python><tensorflow><python-3.5><tensorboard>","CreationDate":"2017-01-22T09:24:31.310","AcceptedAnswerId":null,"OwnerUserId":3047249.0,"Title":"Issue with TensorFlow-TensorBoard","Body":"<p>I'm following <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">this tutorial<\/a> . I've copied the code literally, so I don't see any error. \nThe errors come when I try to add this line to create a file for TensorBoard:<\/p>\n\n<pre><\/pre>\n\n<p>When i run the entire code Python returns this as an error:<\/p>\n\n<pre><\/pre>\n\n<p>I don't see why it doesn't work as exxpected. Could enyone help me?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41922028","CreationDate":"2017-01-29T14:34:54.717","ParentId":null,"OwnerUserId":"7483498","Title":null,"Body":"<p>Making these changes the code runs without any problems<\/p>\n\n<pre><code>logs_path = '.\/tensor_board' \n\nsess.run(init)\n<\/code><\/pre>\n"}]}
{"QuestionId":41789317,"AnswerCount":1,"Tags":"<python><tensorflow><ubuntu-16.04>","CreationDate":"2017-01-22T09:31:21.333","AcceptedAnswerId":null,"OwnerUserId":7453033.0,"Title":"Getting error while installing TensorFlow on Ubuntu 16.04?","Body":"<p>I'm getting error  when I run .<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41790099","CreationDate":"2017-01-22T11:03:43.787","ParentId":null,"OwnerUserId":"5507055","Title":null,"Body":"<p>You should rely the command but with <code>--user<\/code> like:<\/p>\n\n<pre><code>pip install --user tensorflow-gpu\n<\/code><\/pre>\n\n<p>And please improve the design of your question!<\/p>\n"}]}
{"QuestionId":41789961,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-01-22T10:46:09.787","AcceptedAnswerId":41792194.0,"OwnerUserId":1934212.0,"Title":"Keras warning: Epoch comprised more than `samples_per_epoch` samples","Body":"<p>I have about 6200 training images and I  want to use the  method of the  class to augment the small dataset in the following way:<\/p>\n\n<pre><\/pre>\n\n<p>where  \/  contain about 6000 training images and labels and  \/  the validation data and model is an augmented VGG16 model.<\/p>\n\n<p>The documentation says<\/p>\n\n<blockquote>\n  <p>flow(X, y): Takes numpy data &amp; label arrays, and generates batches of augmented\/normalized data. Yields batches indefinitely, in an infinite loop.<\/p>\n<\/blockquote>\n\n<p>For a training setup with 10 epochs, 1920 samples per epoch and a batch_size of 32 I got the following training trace:<\/p>\n\n<pre><\/pre>\n\n<p>Why didnt the generator generate infinite batches as the documentation says?<\/p>\n","answers":[{"AnswerId":"41792194","CreationDate":"2017-01-22T14:37:45.680","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So basically there is a small mistake in a <code>KerasImageGenerator<\/code> class implementation. What is nice - that beside this annoying warning nothing wrong happens. So to clarrify:<\/p>\n\n<ol>\n<li><p>Both <code>flow<\/code>and <code>flow_from_directory<\/code> actually yields samples in an infinite loop. You could easly check that by testing the following code (warning - it will freeze your <code>Python<\/code>):<\/p>\n\n<pre><code>for x, y in train_generator:\n    x = None\n<\/code><\/pre><\/li>\n<li><p>The warning which you mentioned is rasied in a <code>fit_generator<\/code> method. It basically checks if the number of samples processed in one epoch is lesser or equal to <code>samples_per_epoch<\/code>. In your case - where <code>samples_per_epoch<\/code> is divisible by a <code>batch_size<\/code> - if implementation of Keras would be correct - this warning should be never raised... but..<\/p><\/li>\n<li><p>.. yeah, so why this warning was raised? It's a little bit tricky. If you went deeper into implementation of a generator you would notice that generator is getting batches in a following manner : if you have let's say - 10 examples and <code>batch_size = 3<\/code> then:<\/p>\n\n<ul>\n<li>it will first shuffle the order of this 10 examples,<\/li>\n<li>then it would take 3 first of shuffled examples, then next three and so on,<\/li>\n<li>after the 3rd batch - when there is only 1 example left - it will return a batch.. with only one sample.<\/li>\n<\/ul>\n\n<p>Don't ask me why - this is the way how generator is implemented. Good thing is that it barely affects the training process.<\/p><\/li>\n<\/ol>\n\n<p>So - to sum up - you could either ignore this warning or you could make the number of samples you pass to a generator divisible by a <code>batch_size<\/code>. I know that it's cumbersome and I hope it will be fixed in a next release.<\/p>\n"}]}
{"QuestionId":41790857,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-01-22T12:25:08.540","AcceptedAnswerId":null,"OwnerUserId":720484.0,"Title":"What is the difference between saving a summary and saving the model in the logdir?","Body":"<p>Using Tensorflow (tf.contrib.slim in particular) we are required to calibrate a few parameters to produce the graphs that we want at tensorboard.<\/p>\n\n<p>Saving a <strong>summary interval<\/strong> is more clear for us what it does. It saves the value (or an average of them?) of a particular point in the graph at the interval provided.<\/p>\n\n<p>Now <strong>checkpoints for saving the model<\/strong> itself why should be required at the training process? Does the model changes?.. Not sure how this works<\/p>\n","answers":[{"AnswerId":"41791672","CreationDate":"2017-01-22T13:48:55.690","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>You <strong>save the model<\/strong> to checkpoints because the Variables in the model, including neural network weights and biases and the global_step counter, keep changing during the training process. The structure of the model doesn't change. The saved checkpoints allow you to load the trained model for serving and to resume training later.<\/p>\n"}]}
{"QuestionId":41791171,"AnswerCount":2,"Tags":"<macos><python-3.x><tensorflow><bazel>","CreationDate":"2017-01-22T12:59:32.927","AcceptedAnswerId":null,"OwnerUserId":6576617.0,"Title":"Bazel build for TensorFlow inception model","Body":"<p>I am following tutorial for inception model build for TensorFlow at url (<a href=\"https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/<\/a>). When I am running command , I am getting error:<\/p>\n\n<blockquote>\n  <p>The 'build' command is only supported from within a workspace.<\/p>\n<\/blockquote>\n\n<p>I have set up tensorflow using normal  command and no Docker or Anaconda has been installed to run with TensorFlow. Bazel was set up by brewing and I am on Mac OSX v10.11.x.<\/p>\n","answers":[{"AnswerId":"41792501","CreationDate":"2017-01-22T15:05:57.607","ParentId":null,"OwnerUserId":"4367137","Title":null,"Body":"<p>When I did this, I git cloned the directory (git clone <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/<\/a>) to $HOME\/tensorflow, then from the root of the 'tensorflow' directory (the workspace it is referring to), I ran .\/configure and then executed \"bazel build tensorflow\/examples\/image_retraining:retrain\". The output is then generated in a bazel-bin directory.<\/p>\n"},{"AnswerId":"45451770","CreationDate":"2017-08-02T04:57:13.543","ParentId":null,"OwnerUserId":"5305512","Title":null,"Body":"<p>First run <code>touch WORKSPACE<\/code>, and then run the <code>bazel build<\/code> command.<\/p>\n"}]}
{"QuestionId":41791469,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2017-01-22T13:27:47.110","AcceptedAnswerId":41791644.0,"OwnerUserId":6667826.0,"Title":"What's the difference between tf.Session() and tf.InteractiveSession()?","Body":"<p>In which cases should  and  be considered for what purpose?<\/p>\n\n<p>When I tried to use the former one, some functions (for example, ) didn't work, and when I changed to the later one, it worked.<\/p>\n","answers":[{"AnswerId":"51825390","CreationDate":"2018-08-13T15:07:31.133","ParentId":null,"OwnerUserId":"8451329","Title":null,"Body":"<p>On the top of installing itself as default session as per <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/InteractiveSession\" rel=\"nofollow noreferrer\">official documentation<\/a>, from some tests on memory usage, it seems that the interactive session uses the gpu_options.allow_growth = True option - see [using_gpu#allowing_gpu_memory_growth] -  while tf.Session() by default allocates the whole GPU memory.<\/p>\n"},{"AnswerId":"41791644","CreationDate":"2017-01-22T13:45:29.163","ParentId":null,"OwnerUserId":"2833802","Title":null,"Body":"<p>Mainly taken from <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/InteractiveSession\" rel=\"noreferrer\">official<\/a> documentation:<\/p>\n\n<blockquote>\n  <p>The only difference with a regular Session is that an InteractiveSession installs itself as the default session on construction. The methods Tensor.eval() and Operation.run() will use that session to run ops.<\/p>\n<\/blockquote>\n\n<p>This allows to use interactive context, like shell, as it avoids having to pass an explicit Session object to run op:<\/p>\n\n<pre><code>sess = tf.InteractiveSession()\na = tf.constant(5.0)\nb = tf.constant(6.0)\nc = a * b\n# We can just use 'c.eval()' without passing 'sess'\nprint(c.eval())\nsess.close()\n<\/code><\/pre>\n\n<p>It is also possible to say, that <code>InteractiveSession<\/code> supports less typing, as allows to run variables without needing to constantly refer to the session object.<\/p>\n"},{"AnswerId":"43405286","CreationDate":"2017-04-14T04:38:41.190","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The only difference between <code>Session<\/code> and an <code>InteractiveSession<\/code> is that <code>InteractiveSession<\/code> makes itself the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/client\/session_management\" rel=\"noreferrer\">default session<\/a> so that you can call <code>run()<\/code> or <code>eval()<\/code> without explicitly calling the session. <\/p>\n\n<p>This can be helpful if you experiment with TF in python shell or in Jupyter notebooks, because it avoids having to pass an explicit Session object to run operations.<\/p>\n"},{"AnswerId":"44581875","CreationDate":"2017-06-16T06:03:53.003","ParentId":null,"OwnerUserId":"5536013","Title":null,"Body":"<p>Rather than above mentioned differences - the most important difference is with <code>session.run()<\/code> we can fetch values of multiple tensors in one step.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>num1 = tf.constant(5)\nnum2 = tf.constant(10)\nnum3 = tf.multiply(num1,num2)\nmodel = tf.global_variables_initializer()\n\nsession = tf.Session()\nsession.run(model)\n\nprint(session.run([num2, num1, num3]))\n<\/code><\/pre>\n"}]}
{"QuestionId":41792188,"AnswerCount":3,"Tags":"<python><keras>","CreationDate":"2017-01-22T14:37:24.263","AcceptedAnswerId":41958892.0,"OwnerUserId":6360367.0,"Title":"What does `layer.get_weights()` return?","Body":"<p>I'm using Keras to do some experiments and I just monitored the weight update for a simple mlp model:<\/p>\n\n<pre><\/pre>\n\n<p>For the first dense layer, I got a matrix of zeros. I thought the training doesn't change this weight. However, weights in other layers are changed. So I'm confused, why the first layer is unchanged? \nI checked the source code but still got no answer, then I tried monitoring:<\/p>\n\n<pre><\/pre>\n\n<p>This time, the weights did change. So I'm wondering which weight is the \"true\" weight that does working during training? Why there are two elements in the weight list?<\/p>\n\n<hr>\n\n<p>Definition of :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"57384584","CreationDate":"2019-08-06T22:03:24.807","ParentId":null,"OwnerUserId":"9325130","Title":null,"Body":"<p>There is a way to see exactly how the values of all weights and biases are changing over time. You can can use the Keras callback method which can be used to record the weights values at each training epoch. Using a model like this for example,<\/p>\n\n<pre><code>import numpy as np\nmodel = Sequential([Dense(16, input_shape=(train_inp_s.shape[1:])), Dense(12), Dense(6), Dense(1)])\n<\/code><\/pre>\n\n<p>add the callbacks **kwarg during fitting:<\/p>\n\n<pre><code>gw = GetWeights()\nmodel.fit(X, y, validation_split=0.15, epochs=10, batch_size=100, callbacks=[gw])\n<\/code><\/pre>\n\n<p>where the callback is defined by<\/p>\n\n<pre><code>class GetWeights(Callback):\n    # Keras callback which collects values of weights and biases at each epoch\n    def __init__(self):\n        super(GetWeights, self).__init__()\n        self.weight_dict = {}\n\n    def on_epoch_end(self, epoch, logs=None):\n        # this function runs at the end of each epoch\n\n        # loop over each layer and get weights and biases\n        for layer_i in range(len(self.model.layers)):\n            w = self.model.layers[layer_i].get_weights()[0]\n            b = self.model.layers[layer_i].get_weights()[1]\n            print('Layer %s has weights of shape %s and biases of shape %s' %(\n                layer_i, np.shape(w), np.shape(b)))\n\n            # save all weights and biases inside a dictionary\n            if epoch == 0:\n                # create array to hold weights and biases\n                self.weight_dict['w_'+str(layer_i+1)] = w\n                self.weight_dict['b_'+str(layer_i+1)] = b\n            else:\n                # append new weights to previously-created weights array\n                self.weight_dict['w_'+str(layer_i+1)] = np.dstack(\n                    (self.weight_dict['w_'+str(layer_i+1)], w))\n                # append new weights to previously-created weights array\n                self.weight_dict['b_'+str(layer_i+1)] = np.dstack(\n                    (self.weight_dict['b_'+str(layer_i+1)], b))\n<\/code><\/pre>\n\n<p>This callback builds a dictionary with all the layer weights and biases, labeled by the layer numbers, so you can see how they are changing over time as your model is being trained. You'll notice that the shape of each weight and bias array depends on the shape of the model layer. One weights array and one bias array are saved for each layer in your model. The third axis (depth) shows their evolution over time.<\/p>\n\n<p>Here we used 10 epochs and a model with layers of 16, 12, 6, and 1 neurons:<\/p>\n\n<pre><code>for key in gw.weight_dict:\n    print(str(key) + ' shape: %s' %str(np.shape(gw.weight_dict[key])))\n\nw_1 shape: (5, 16, 10)\nb_1 shape: (1, 16, 10)\nw_2 shape: (16, 12, 10)\nb_2 shape: (1, 12, 10)\nw_3 shape: (12, 6, 10)\nb_3 shape: (1, 6, 10)\nw_4 shape: (6, 1, 10)\nb_4 shape: (1, 1, 10)\n<\/code><\/pre>\n"},{"AnswerId":"41958892","CreationDate":"2017-01-31T13:30:48.123","ParentId":null,"OwnerUserId":"6360367","Title":null,"Body":"<p>For the question of <code>layer.get_weights()<\/code>:<\/p>\n\n<p>I ran some tests on this issue and checked the source codes. I found that the <code>Dense<\/code> layer is a subclass of <code>Layer<\/code> and its weights, which is a type of python <code>list<\/code> has two elements weight of the layer stored at <code>layer.get_weights()[0]<\/code> and the <strong><code>bias<\/code><\/strong> is stored at <code>layer.get_weights()[1]<\/code>. <\/p>\n\n<p>There's one thing to note that, <code>bias<\/code> can be disabled during defining the layer: <code>model.add(Dense(503,init='normal',activation='relu',<\/code><strong><code>bias=False<\/code><\/strong><code>))<\/code>. In that case, the list <code>layer.get_weights()<\/code> has only one element. If you set the <code>bias<\/code> attribute as <code>False<\/code> after defining it, there will still be an element for <code>bias<\/code> and it would be updated after you fitting the model.<\/p>\n\n<p>For the question of <strong>not updating<\/strong>:<\/p>\n\n<p>I set up a <strong>Sequential<\/strong> model with only one dense layer: <\/p>\n\n<p><code>def mlp_2():\n    model=Sequential()\n    model.add(Dense(10, input_dim=784, activation='softmax', bias =False))\n    return model<\/code><\/p>\n\n<p>Then I use the same way above to compile and fit it. This is what I got:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/w6z2F.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/w6z2F.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>It still seems not update the weight, however, we can tell the weight is definately changed. Because the accuracy is increasing. I think the only explanation is updates on the first <code>dense<\/code> layer (which you define <code>input_dim<\/code>) is too small for Keras to printout. I didn't check the more precise value of the weights, it would be great if someone could confrim it.<\/p>\n"},{"AnswerId":"50521848","CreationDate":"2018-05-25T05:03:43.247","ParentId":null,"OwnerUserId":"3546399","Title":null,"Body":"<p>Here is a working example. <\/p>\n\n<pre><code>import numpy as np\nfrom  keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Activation, Flatten\n\nX_train=np.random.rand(1,10)\nY_train=2*X_train\ninput_dim = X_train.shape[1]\nmodel=Sequential()\nmodel.add(Dense(20, input_dim=10))\nmodel.add(Dense(10, activation='softmax'))\nweight_origin_0=model.layers[0].get_weights()[0]\nweight_origin_1=model.layers[1].get_weights()[0]\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nmodel.fit(X_train, Y_train, batch_size=1, nb_epoch=10, verbose=1)\nprint(weight_origin_0-model.layers[0].get_weights()[0])  #the first layer\nprint(weight_origin_1-model.layers[1].get_weights()[0])  #the second layer\n<\/code><\/pre>\n"}]}
{"QuestionId":41793122,"AnswerCount":0,"Tags":"<python><machine-learning><nlp><keras><lstm>","CreationDate":"2017-01-22T16:05:23.857","AcceptedAnswerId":null,"OwnerUserId":7226273.0,"Title":"RCV1 dataset classification with Keras","Body":"<p>I'm having issues training a text classifier using LSTM on the RCV1-v2 dataset,and after a thorough search I still don't know how to input the data to my model.<\/p>\n\n<p>I'm using the fetch_rcv1() function from sklearn (<a href=\"http:\/\/scikit-learn.org\/stable\/datasets\/rcv1.html\" rel=\"nofollow noreferrer\">http:\/\/scikit-learn.org\/stable\/datasets\/rcv1.html<\/a>), which states the following:<\/p>\n\n<p>data: The feature matrix is a scipy CSR sparse matrix, with 804414 samples and 47236 features.<\/p>\n\n<p>target: The target values are stored in a scipy CSR sparse matrix, with 804414 samples and 103 categories.<\/p>\n\n<p>sample_id: Each sample can be identified by its ID, ranging (with gaps) from 2286 to 810596:<\/p>\n\n<p>Note that a document can belong to multiple categories,thus making the problem multilabel.According to the above ,train shape is (23149,47236) =>23149 documents with 47236 TF-IDF  vectors<\/p>\n\n<p>Unfortunately there's no clear explanation or any tutorial whatsoever(even after days of searching!!) on how to manipulate the data. I'm trying to insert the data to a LSTM layer to no success.The model keeps failing to compile,example given:<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone guide me on how I should reshape the data\/layers to get the model to compile?Here's my code attempt so far:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41793538,"AnswerCount":2,"Tags":"<python><machine-learning><scikit-learn><neural-network><keras>","CreationDate":"2017-01-22T16:46:10.143","AcceptedAnswerId":null,"OwnerUserId":4013873.0,"Title":"keras: issue using ImageDataGenerator and KFold for fit_generator","Body":"<p><a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">flow_from_directory<\/a>(directory): This takes in directory but does not take split training images.<\/p>\n\n<p>sklearn.model_selection.KFold: Provides the split indices of images. Those could be used in fit() but not in fit_generator()<\/p>\n\n<p>How can anyone use KFold along with ImageDataGenerator? Is it there?<\/p>\n","answers":[{"AnswerId":"41796453","CreationDate":"2017-01-22T21:40:22.437","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>At the moment one cannot split a dataset held in the folder using a <code>flow_from_directory<\/code> generator. This option is simply not implemented. To get the test \/ train split one need to split the main directory into set of train \/ test \/val directories using e.g. <code>os<\/code> library in Python. <\/p>\n"},{"AnswerId":"49546531","CreationDate":"2018-03-29T01:10:40.183","ParentId":null,"OwnerUserId":"9315332","Title":null,"Body":"<p>Assuming that you have a classification problem with 2 classes, I would do something like:<\/p>\n\n<pre><code>from keras.utils import to_categorical  \n\ntrain_y = to_categorical(train_y, num_classes=2)\ntest_y = to_categorical(test_y, num_classes=2)\n\naug = ImageDataGenerator(...) #your ImageDataGenerator\n\nModel = model.fit_generator(aug.flow(train_x,tain_y, batch_size=32), \n            validation_data=(test_x,test_y))\n<\/code><\/pre>\n"}]}
{"QuestionId":41793677,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-22T16:59:04.667","AcceptedAnswerId":41799293.0,"OwnerUserId":7243382.0,"Title":"Tensorflow: Regarding tensorflow functions","Body":"<p>I'm new to tensorflow. I have the following problem: <\/p>\n\n<p>input: list of floats (or a dynamic array. In python list is the datatype to be used)\nOutput: is a 2-d array of size  \u00d7  <\/p>\n\n<h1>Example1:<\/h1>\n\n<p>Input:<\/p>\n\n<pre><\/pre>\n\n<p>Output: <\/p>\n\n<pre><\/pre>\n\n<p>I tried to create the function using  loop and calculating each row independently and concatenating them, but my instructor asked me to explore other ways.<\/p>\n\n<p>Can you suggest me an idea on how to approach this problem?<\/p>\n","answers":[{"AnswerId":"44770461","CreationDate":"2017-06-27T00:07:45.677","ParentId":null,"OwnerUserId":"7806129","Title":null,"Body":"<p>This is probably a bit late for your class, but hopefully it will help someone.<\/p>\n\n<p>If your goal is to simply output a <code>len(input)xlen(input)<\/code> array, you can matrix multiply a <code>1xlen(input)<\/code> tensor with your input array after expanding its dimensions to <code>len(input)x1<\/code>:<\/p>\n\n<pre><code>input_ = tf.placeholder(tf.float32, [len(input)])\ninput_shape = input_.get_shape().as_list()\ntfvar = tf.Variable(tf.random_normal([1,input_shape[0]], mean=0.0,\n                                    stddev=.01, dtype=tf.float32))\n\ndef function(input_):\n    x = tf.expand_dims(input_, axis=1) # dims = len(input)x1\n    return tf.matmul(x,tfvar) # mtrx multiplication produces 3x3 mtrx\n<\/code><\/pre>\n\n<p>This function should generalize to any 1D <code>input_<\/code> tensor and produce a square <code>len(input_)xlen(input_)<\/code> tensor.<\/p>\n\n<p>If your goal is to train a tensorflow variable to produce the provided output exactly, you can then train <code>tfvar<\/code> with a loss function and optimizer:<\/p>\n\n<pre><code>desired_output = tf.constant([[0.09003057, 0.24472847, 0.66524096], \n                              [0.26894142, 0.73105858, 0.0       ], \n                              [1.0,        0.0,        0.0       ]],\n                              dtype=tf.float32)\n\nactual_output = function(input_)\nloss = tf.reduce_mean(tf.square(actual_output-desired_output))\noptimizer = tf.train.AdamOptimizer().minimize(loss)\n\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    cost, opt = sess.run([loss, optimizer], feed_dict={input_:input})\n<\/code><\/pre>\n\n<p>Note, if you want a more robust training session, add a bias, a non-linearity, and more layers.<\/p>\n"},{"AnswerId":"41799293","CreationDate":"2017-01-23T04:34:12.820","ParentId":null,"OwnerUserId":"2708478","Title":null,"Body":"<p>You can achieve this with the following approach:<\/p>\n\n<ol>\n<li>Repeat the input array to create a square matrix tiled the input data<\/li>\n<li>Create a mask with consists of ones in the left upper corner<\/li>\n<li>Do softmax using the mask. Note that we cannot use <code>tf.nn.softmax<\/code> here because it will give small probabilities to those zeros also<\/li>\n<\/ol>\n\n<p>Here is a TensorFlow (v0.12.1) code that does this: <\/p>\n\n<pre><code>def create_softmax(x):\n    x_len = int(x.get_shape()[0])\n\n    # create a tiled array\n    # [1, 2, 3] \n    # =&gt;\n    # [[1,2,3], [1,2,3], [1,2,3]]\n    x_tiled = tf.tile(tf.expand_dims(x, 0), [x_len, 1])\n\n    # get the mask to do element-wise multiplication\n    mask = tf.ones_like(x_tiled) # returns an array of the same size filled with 1\n    mask = tf.matrix_band_part(mask, 0, -1) # zeros everythings except from the upper triangular part \n    mask = tf.reverse(mask, [False, True]) # reverses the y dimension\n\n    # compute masked softmax\n    exp = tf.exp(x_tiled) * mask\n    sum_exp = tf.reshape(tf.reduce_sum(exp, reduction_indices=1), (-1, 1))\n\n    x_softmax = exp \/ sum_exp\n\n    return x_softmax\n<\/code><\/pre>\n"}]}
{"QuestionId":41793762,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2017-01-22T17:07:00.413","AcceptedAnswerId":null,"OwnerUserId":7387749.0,"Title":"DNN in TF learn, Error: \"indices are out-of-bounds\"","Body":"<p>I am new to TFLearn, and I am trying to implement this DNN:<\/p>\n\n<pre><\/pre>\n\n<p>When I run <em>model.fit<\/em> , I get the error <em>IndexError: indices are out-of-bounds<\/em>. I suppose this issue is regarding columns's import by \"name\", and the network tries to elaborate the strings in the header, but I'm not sure about it.\nDo I miss something obvious?\nThanks for any help!  <\/p>\n","answers":[]}
{"QuestionId":41794545,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2017-01-22T18:20:33.917","AcceptedAnswerId":null,"OwnerUserId":7454411.0,"Title":"Tensorflow: How to write clean code for multi GPU model parallelism?","Body":"<p>Currently i'm implementing a large custom model and referencing the multi gpu example of CIFAR 10 that comes along with tensorflow. However the code I ended up writing based on that was not clean and was error prone. For e.g. I had to find every trainable variable and add \"with tf.device('\/cpu:0')\".<\/p>\n\n<p>Are there more efficient\/cleaner ways of adapting for multi gpu execution?<\/p>\n\n<p>Many thanks for any support.<\/p>\n","answers":[{"AnswerId":"41795309","CreationDate":"2017-01-22T19:36:37.920","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's an <a href=\"https:\/\/github.com\/rafaljozefowicz\/lm\/blob\/master\/language_model.py\" rel=\"nofollow noreferrer\">example<\/a> from Rafal<\/p>\n\n<p>You make a loop over towers with the body constructing <code>i<\/code>th tower as <code>with tf.device(assign_to_gpu(i))<\/code>. The function <code>assign_to_gpu<\/code> treats variables differently and assigns them onto \"ps-device\".<\/p>\n\n<p>Note: we found that when GPUs are p2p connected, training was faster when variables were kept <code>gpu:0<\/code> rather than <code>cpu:0<\/code><\/p>\n"}]}
{"QuestionId":41794879,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><deep-learning><conv-neural-network>","CreationDate":"2017-01-22T18:53:29.480","AcceptedAnswerId":null,"OwnerUserId":541106.0,"Title":"Object Localization through CNN","Body":"<p>I am new to deep learning and tensor flow and I am trying to train a CNN at localizing digits in the Street View House Numbers data set. To this end I have an input set of 32x32 images and, since I want to recognize up to 5 digits, I am using as labels vectors of 20 elements like this<\/p>\n\n<p>[top_x_digit1,top_y_digit1,width_digit1,height_digit1,top_x_digit2, etc..]<\/p>\n\n<p>0,0,0,0 when there is no digit<\/p>\n\n<p>As far as I understand, after (let me say) 3 layers of convolution and pooling I can add 5 (parallel) fully connected layers aimed at extracting each the box features of a different digit (when present, 0 0 0 0 otherwise).<\/p>\n\n<p>is my approach correct?<\/p>\n","answers":[]}
{"QuestionId":41795683,"AnswerCount":1,"Tags":"<centos><tensorflow><install>","CreationDate":"2017-01-22T20:16:34.537","AcceptedAnswerId":null,"OwnerUserId":7400630.0,"Title":"How do install tensor flow gpu?","Body":"<p>I am currently having some problems installing tensor flow with gpu support. <\/p>\n\n<p>This is the guide i've followed.<\/p>\n\n<ul>\n<li>Install NVIDIA CUDA (preinstalled)<\/li>\n<li>Install NVIDIA cuDNN (preinstalled)<\/li>\n<li><p>Install bazel<\/p>\n\n<pre><\/pre><\/li>\n<li><p>install tensor flow from source<\/p>\n\n<p>git clone <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow<\/a><\/p>\n\n<p>cd tensorflow\/<\/p>\n\n<p>.\/configure<\/p>\n\n<p>bazel build -c opt --config=cuda \/\/tensorflow\/tools\/pip_package:build_pip_package<\/p><\/li>\n<\/ul>\n\n<p>the last one don't finish.. or it does.. it comes up with an error message       <\/p>\n\n<pre><\/pre>\n\n<p>after this should i be able to do this <\/p>\n\n<pre><\/pre>\n\n<p>according to the guide, but is not possible. <\/p>\n\n<p>i've previously had tensor flow cpu version running, but since the need of gpu were sincerely pressing i decided to install it... but didn't think it would be this troublesome.. <\/p>\n\n<p>any idea on why it is not possible to build it?    <\/p>\n\n<p>os:centOS 7.1\ngpu:nvidea <\/p>\n","answers":[{"AnswerId":"41796535","CreationDate":"2017-01-22T21:50:29.657","ParentId":null,"OwnerUserId":"7226273","Title":null,"Body":"<p>I've had problems attemping to install Tensorflow from source using Bazel as well.It could be that the current bazel build has problems,since that occured in the past with me.It would help if you provided us with information about your system(OS &amp; GPU),but you're probably better off using pip(or pip3) and doing sudo pip(3) install tensorflow-gpu.<\/p>\n"}]}
{"QuestionId":41796367,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-22T21:30:22.053","AcceptedAnswerId":null,"OwnerUserId":6227294.0,"Title":"How can I see which variable or functions are running in CPU tensorflow","Body":"<p>I am running a CNN in tensorflow. I am using tf.device(\/gpu:0) to put all of my variables in gpu, but it seems that some of them are still in cpu. I am looking at my gpu utils when I am running my code, it goes up to 100% and then goes down to 0%. <\/p>\n\n<p>I know that if I use config.log_device_placement = True I can see which variable are assigned to which device. but because the number of variable are a lot in my code, I couldn't find out which one are in cpu.<\/p>\n\n<p>So, is there any way that I just see which variables are pinned to cpu? \nOr, do you have any idea why some of my variable should be pinned to cpu while I am using tf.device to assign them to gpu? <\/p>\n\n<p>By the way, after I changed upsampler (simple interpolator tf.image.resize_images) by following code for upsampling this issue happened: <\/p>\n\n<pre><\/pre>\n\n<p>I got this code from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2169\" rel=\"nofollow noreferrer\">here<\/a> for unpooling. <\/p>\n","answers":[{"AnswerId":"41880748","CreationDate":"2017-01-26T18:46:48.317","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>You can log device placement via configuration of Session. Setting:<\/p>\n\n<pre><code>  sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n<\/code><\/pre>\n"}]}
{"QuestionId":41796618,"AnswerCount":3,"Tags":"<python><scikit-learn><regression><keras>","CreationDate":"2017-01-22T21:59:48.763","AcceptedAnswerId":null,"OwnerUserId":7454972.0,"Title":"Python Keras cross_val_score Error","Body":"<p>I am trying to do this little tutorial on keras about regression:\n<a href=\"http:\/\/machinelearningmastery.com\/regression-tutorial-keras-deep-learning-library-python\/\" rel=\"noreferrer\">http:\/\/machinelearningmastery.com\/regression-tutorial-keras-deep-learning-library-python\/<\/a><\/p>\n\n<p>Unfortunately I am running into an error I cannot fix. \nIf i just copy  and paste the code I get the following error when running this snippet:<\/p>\n\n<pre><\/pre>\n\n<p>The error says:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for any help.<\/p>\n\n<p>Here is the full traceback:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42693524","CreationDate":"2017-03-09T10:57:42.357","ParentId":null,"OwnerUserId":"7683841","Title":null,"Body":"<p>Had the same problem.\nAfter upgrading keras version to 1.2.2 the problem went away.<\/p>\n\n<p>If you use pip to manage your packages you can upgrade keras with the following command: <\/p>\n\n<pre><code>sudo pip install --upgrade keras\n<\/code><\/pre>\n"},{"AnswerId":"41832675","CreationDate":"2017-01-24T15:59:34.990","ParentId":null,"OwnerUserId":"6010635","Title":null,"Body":"<p><strong>EDIT (25\/01\/2017):<\/strong> <em>This solution works because with the conda environment the version of Keras that is installed is 1.1.1, not the one with the bug (1.2.1). Jason's solution is the correct one. I left here my solution in case it can helps but Jason's solution is the actual solution.<\/em><\/p>\n\n<p>I had the same problem after upgrading Keras (1.2.1). I think the problem is with the versions of the software. What I recommend you is to install <code>Anaconda<\/code> , then to create a new environment where you install <code>tensorflow<\/code>. Basically follow these steps: <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#anaconda_installation\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/os_setup#anaconda_installation<\/a><\/p>\n\n<p>Activate the environment and install using the <code>conda<\/code> option. Then you can install other libraries that you're going to need. With the environment <code>tensorflow<\/code> activated you install with <code>conda install name_of_the_package<\/code>.<\/p>\n\n<p>You can change between <code>theano<\/code> and <code>tensorflow<\/code> with the backends of Keras (<a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/backend\/<\/a>).<\/p>\n\n<p>Basically, with conda environments you're creating a protected, encapsulated area where you can install and uninstall what you want and it doesn't affect your other programs out of the environment. What you're doing is like a delete and re-install everything afresh, with the latest and working versions.<\/p>\n\n<p>Hope it helps.<\/p>\n"},{"AnswerId":"41841066","CreationDate":"2017-01-24T23:54:04.123","ParentId":null,"OwnerUserId":"78453","Title":null,"Body":"<p>The specific error reported is:<\/p>\n\n<pre><code>TypeError: get_params() got an unexpected keyword argument 'deep'\n<\/code><\/pre>\n\n<p>The fault was introduced by a bug in Keras version 1.2.1. It occurs when you use the Keras wrapper classes (e.g. KerasClassifier and KerasRegressor) and scikit-learn function cross_val_score().<\/p>\n\n<p>The bug has been <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/5103\">identified<\/a> and <a href=\"https:\/\/github.com\/fchollet\/keras\/commit\/72f1ce4ed4396a308f831c717d8c2856a7a4e86b\">patched<\/a> in the Keras GitHub project.<\/p>\n\n<p>There are two fixes that I have tried:<\/p>\n\n<p><strong>Fix 1: Roll-back to Keras version 1.2.0.<\/strong><\/p>\n\n<p>Type:<\/p>\n\n<pre><code>sudo pip install keras==1.2.0\n<\/code><\/pre>\n\n<p><strong>Fix 2: Monkey-patch Keras with the fix.<\/strong><\/p>\n\n<p>After your imports, but before your work type:<\/p>\n\n<pre><code>from keras.wrappers.scikit_learn import BaseWrapper\nimport copy\n\ndef custom_get_params(self, **params):\n    res = copy.deepcopy(self.sk_params)\n    res.update({'build_fn': self.build_fn})\n    return res\n\nBaseWrapper.get_params = custom_get_params\n<\/code><\/pre>\n\n<p>Both fixes work for me (Python 2 and 3\/sklearn 0.18.1).<\/p>\n\n<p>Some additional candidate fixes:<\/p>\n\n<ul>\n<li>Wait for the next version of Keras (1.2.2) to be released.<\/li>\n<li>Checkout Keras from Github then build and install manually.<\/li>\n<\/ul>\n"}]}
{"QuestionId":41796965,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-22T22:43:27.953","AcceptedAnswerId":null,"OwnerUserId":6480160.0,"Title":"Tensorflow: How to use a trained model in a application?","Body":"<p>I have trained a Tensorflow Model, and now I want to export the \"function\" to use it in my python program. Is that possible, and if yes, how? Any help would be nice, could not find much in the documentation. (I dont want to save a session!)<\/p>\n\n<p>I have now stored the session as you suggested. I am loading it now like this:<\/p>\n\n<pre><\/pre>\n\n<p>However, I get the error \"no Variables to save\" when I try to initialize the saver. <\/p>\n\n<p>What I want to do is this: I am writing a bot for a board game, and the input is the situation on the board formatted into a tensor. Now I want to return a tensor which gives me the best position to play next, i.e. a tensor with 0 everywhere and a 1 at one position.<\/p>\n","answers":[{"AnswerId":"41831621","CreationDate":"2017-01-24T15:10:42.107","ParentId":null,"OwnerUserId":"4637867","Title":null,"Body":"<p>I don't know if there is any other way to do it, but you can use your model in another Python program by saving your session:<\/p>\n\n<p>Your training code:<\/p>\n\n<pre><code># build your model\n\nsess = tf.Session()\n# train your model\nsaver = tf.train.Saver()\nsaver.save(sess, 'model\/model.ckpt')\n<\/code><\/pre>\n\n<p>In your application:<\/p>\n\n<pre><code># build your model (same as training)\nsess = tf.Session()\nsaver = tf.train.Saver()\nsaver.restore(sess, 'model\/model.ckpt')\n<\/code><\/pre>\n\n<p>You can then evaluate any tensor in your model using a feed_dict. This obviously depends on your model. For example:<\/p>\n\n<pre><code>session.run(y_pred, feed_dict={x: input_data})\n<\/code><\/pre>\n"}]}
{"QuestionId":41797136,"AnswerCount":1,"Tags":"<python><python-3.x><machine-learning><tensorflow><conv-neural-network>","CreationDate":"2017-01-22T23:02:49.517","AcceptedAnswerId":41798789.0,"OwnerUserId":6348198.0,"Title":"How to Concatenate \"Jagged\" Tensors","Body":"<p>I am trying to write an implementation of <a href=\"http:\/\/www.aclweb.org\/anthology\/P14-1062\" rel=\"nofollow noreferrer\">this<\/a> paper in TensorFlow and I have come across a bit of a snag. In my pooling layer, I have to concatenate everything together. This is the code I use:<\/p>\n\n<pre><\/pre>\n\n<p>When I run this code, it prints out this for :<\/p>\n\n<pre><\/pre>\n\n<p>I originally tried this code without the  line in there and I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>When I added in the reshape line, I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>The second error I know is because I passed a \"?\" for the new size, and the first error I think is because the tensors aren't the same size. <strong>How could I properly pad these Tensors so I can concatenate them with no problems?<\/strong><\/p>\n","answers":[{"AnswerId":"41798789","CreationDate":"2017-01-23T03:18:09.303","ParentId":null,"OwnerUserId":"2708478","Title":null,"Body":"<p>You can pass <code>-1<\/code> as one of the component of the shape to the <code>tf.reshape<\/code> method; it will be automatically inferred from the the shape of you tensor so the total size will be the same.<\/p>\n\n<p>So, try to change the problem line to<\/p>\n\n<pre><code>pooled_outputs = [tf.reshape(out, [-1, 94, 1, self.max_length]) for out in pooled_outputs]\n<\/code><\/pre>\n\n<p>See the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/shapes_and_shaping#reshape\" rel=\"nofollow noreferrer\">documentation<\/a> for details<\/p>\n"}]}
{"QuestionId":41797279,"AnswerCount":1,"Tags":"<ruby-on-rails><keras><rails-postgresql>","CreationDate":"2017-01-22T23:20:35.733","AcceptedAnswerId":null,"OwnerUserId":3676943.0,"Title":"In Rails 5, How to create a controller action which serves some binary data?","Body":"<p>Today I am working with Keras.js which has a repo here:<\/p>\n\n<p><a href=\"https:\/\/github.com\/transcranial\/keras-js\" rel=\"nofollow noreferrer\">https:\/\/github.com\/transcranial\/keras-js<\/a><\/p>\n\n<p>To operate this I need to serve some JS which looks something like this:<\/p>\n\n<pre><\/pre>\n\n<p>I want\n'url\/path\/to\/model_weights.buf'\nto route to a controller action which serves some binary data.<\/p>\n\n<p>Question 1:\nIn Rails 5, How to create a controller action which serves some binary data?<\/p>\n\n<p>After I get the answer to Q1, I will ask:<\/p>\n\n<p>Question 2:\nIn Rails 5, How to create a controller action which gets some (small) binary data from a model (backed by Postgres) and then serves the binary data?<\/p>\n\n<p>I'm tempted to use the info at the URL below but I'd welcome any comments or syntax examples:<\/p>\n\n<p><a href=\"http:\/\/api.rubyonrails.org\/classes\/ActionController\/DataStreaming.html#method-i-send_data\" rel=\"nofollow noreferrer\">http:\/\/api.rubyonrails.org\/classes\/ActionController\/DataStreaming.html#method-i-send_data<\/a><\/p>\n","answers":[{"AnswerId":"41802112","CreationDate":"2017-01-23T08:41:27.733","ParentId":null,"OwnerUserId":"3676943","Title":null,"Body":"<p>This was an easy task to figure out:<\/p>\n\n<pre><code>def nnmodel_buf\n  # \/kpredictions\/nnmodel_buf\n  send_data Nnmodel.first().nnmodel_buf\nend\n<\/code><\/pre>\n"}]}
{"QuestionId":41797528,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-23T00:01:02.283","AcceptedAnswerId":null,"OwnerUserId":7363573.0,"Title":"Tensorflow One-Hot","Body":"<p>I'm new to Tensorflow (And neural networks) and i am working on a simple classification problem. Would like to ask 2 questions.<\/p>\n\n<p>Say i have 120 labels of a permutation of [1,2,3,4,5]. Is it really necessary for me to One-Hot encode before feeding it into my graph? If yes, should i encode before feeding into tensorflow?<\/p>\n\n<p>And if i do One-Hot encode, the softmax prediction will give [0.001 0.202 0.321\u2026\u20260.002 0.0003 0.0004]. Running arg_max will produce the right index. How would i get tensorflow to return me the correct label instead of a one-hot result?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"41798535","CreationDate":"2017-01-23T02:40:49.613","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>So your input are 120 labels in {1, 2, 3, 4, 5} (each of which can be either digit from 1 to 5)?<\/p>\n\n<pre><code># Your input, a 1D tensor of 120 elements from 1-5.\n# Better shift your label space to 0-4 instead.\nlabels = labels - 1\n\n# Now convert to a 2D tensor of 120 x 5 onehot labels.\nonehot_labels = tf.one_hot(labels, 5)\n\n# Now some computations.\n....\n\n# You end up with some onehot_output\n# of the same shape as your labels (120x5).\n# As you said, arg_max will give you the index of the result,\n# which is a 1D index label of 120 elements.\noutput = tf.argmax(onehot_output, axis=1).\n\n# You might want to shift back to {1,2,3,4,5}.\noutput = output + 1\n<\/code><\/pre>\n"}]}
{"QuestionId":41797741,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow><blender><reinforcement-learning>","CreationDate":"2017-01-23T00:38:13.573","AcceptedAnswerId":null,"OwnerUserId":7455286.0,"Title":"Communications between Blender and Tensorflow","Body":"<p>I have a simulation inside Blender that I would like to control externally using the TensorFlow library. The complete process would go something like this:<\/p>\n\n<pre><\/pre>\n\n<p>I don't have much experience with the threading or subprocess modules and so am unsure as to how I should go about actually building something like this. <\/p>\n","answers":[{"AnswerId":"41841991","CreationDate":"2017-01-25T01:33:44.180","ParentId":null,"OwnerUserId":"7455286","Title":null,"Body":"<p>Thanks for your response. Unfortunately, trying to get Universe working with my current setup was a bit of a pain. I'm also on a fairly tight deadline so I just needed something that worked straight away.<\/p>\n\n<p>I did find a somewhat DIY solution that worked well for my situation using the pickle module. I don't really know how to convert this approach into proper pseudocode, so here is the general outline:<\/p>\n\n<pre><code>Process 1 - TensorFlow:\n    load up TF graph\n    wait until pickled state arrives\n\nwhile not terminated:\n\n    Process 2 - Blender: \n        run simulation until next action required\n        pickle state\n        wait until pickled action arrives\n\n    Process 1 - TensorFlow:\n        unpickle state\n        feedforward state through neural net, find action\n        pickle action\n        wait until next pickled state arrives\n\n    Process 2 - Blender:\n        unpickle action\n        take action\n<\/code><\/pre>\n\n<p>This approach worked well for me, but I imagine there are more elegant low level solutions. I'd be curious to hear of any other possible approaches that achieve the same goal.<\/p>\n"},{"AnswerId":"44741187","CreationDate":"2017-06-24T21:38:15.230","ParentId":null,"OwnerUserId":"5655164","Title":null,"Body":"<p>I did this to install tensorflow with the python version that comes bundled with blender. <\/p>\n\n<ul>\n<li>Blender version: 2.78<\/li>\n<li><p>Python version: 3.5.2<\/p>\n\n<ol>\n<li>First of all you need to install pip for blender's python. For that I followed instructions mentioned in this link: <a href=\"https:\/\/blender.stackexchange.com\/questions\/56011\/how-to-use-pip-with-blenders-bundled-python\">https:\/\/blender.stackexchange.com\/questions\/56011\/how-to-use-pip-with-blenders-bundled-python<\/a>. (What I did was drag the python3.5m icon into the terminal and then further type the command '~\/get-pip.py').<\/li>\n<li>Once you have pip installed you are all set up to install 3rd party modules and use them with blender.<\/li>\n<li>Navigate to the bin folder insider '\/home\/path to blender\/2.78\/' directory. To install tensorflow, drag the python3.5m icon into terminal, then drag pip3 icon into terminal and give the command install tensorflow.<\/li>\n<li>I got an error mentioning module lib2to3 not found. Even installing the module didn't help as I got the message that no such module exists. Fortunately, I have python 3.5.2 installed on my machine. So navigate to  \/usr\/lib\/python3.5 and copy the lib2to3 folder from there to \/home\/user\/blender\/2.78\/python\/lib\/python3.5 and then again run the installation command for tensorflow. The installation should complete without any errors. Make sure that you test the installation by importing tensorflow. <\/li>\n<\/ol><\/li>\n<\/ul>\n"},{"AnswerId":"41797987","CreationDate":"2017-01-23T01:19:49.190","ParentId":null,"OwnerUserId":"98975","Title":null,"Body":"<p>Rather than mess around with Tensorflow connections and APIs, I'd suggest you take a look at the Open AI Universe Starter Agent[1]. The advantage here is that as long as you have a VNC session open, you can connect a TF based system to do reinforcement learning on your actions.<\/p>\n\n<p>Once you have a model constructed via this, you can focus on actually building a lower level API system for the two things to talk to each other.<\/p>\n\n<p>[1] <a href=\"https:\/\/github.com\/openai\/universe-starter-agent\" rel=\"nofollow noreferrer\">https:\/\/github.com\/openai\/universe-starter-agent<\/a><\/p>\n"}]}
{"QuestionId":41797875,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-23T01:02:24.173","AcceptedAnswerId":null,"OwnerUserId":6227294.0,"Title":"Running tf.mod and tf.floor_div in tensorflow in GPU","Body":"<p>I wanna use tf.mod and tf.floor_div for some operation in middle of my CNN network, but it seems that they can not be run in GPU, so they are passed to CPU and it causes to reduce the speed of training significantly. <\/p>\n\n<p>For example in the following simplified code I get following error: <\/p>\n\n<pre><\/pre>\n\n<p>here I get this error:<\/p>\n\n<blockquote>\n  <p>Cannot assign a device to node 'Mod': Could not satisfy explicit device specification '\/device:GPU:0' because no supported kernel for GPU devices is available.\n       [[Node: Mod = Mod[T=DT_FLOAT, _device=\"\/device:GPU:0\"](a, b)]]<\/p>\n<\/blockquote>\n\n<p>I can use allow_soft_placement = True for solving this problem here, but the operation will be done in CPU. <\/p>\n\n<p>I am running my code in ubuntu 14, python 2.7 with tensorflow 0.12. <\/p>\n\n<p>I appreciate if anyone can help me how I can solve the problem.<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":41798010,"AnswerCount":1,"Tags":"<python-2.7><keras><conv-neural-network>","CreationDate":"2017-01-23T01:23:06.800","AcceptedAnswerId":null,"OwnerUserId":7400630.0,"Title":"Setting up a CNN network in keras?","Body":"<p>I am currently trying to implement a cnn network, which can map an input to an output.  <\/p>\n\n<p>The input consist of stft of audio files, and the output is a feature vector. <\/p>\n\n<p>Due to the different length of audio files, will the number of total samples always be different, but each sample has a frame length of 25 ms and 10 ms overlap. shape(x,2050)<\/p>\n\n<p>The output is a feature vector shape is (x,13). <\/p>\n\n<p>I thought the use of cnn seemed appropriate here as the stft as the each input contains some information of the previous  sample due to the overlap. <\/p>\n\n<p>Is it possible in keras to design a model, which make use of this, so the there will be calculated a convolutional sum for each row of the matrix, and somehow make it aware of the 25 frame length and the 10 overlap. <\/p>\n","answers":[{"AnswerId":"41798055","CreationDate":"2017-01-23T01:28:55.353","ParentId":null,"OwnerUserId":"98975","Title":null,"Body":"<p>Yes it is, see line 220 of this file [1]. This is an implementation of Wavenet in Keras using convolutions. Even though they've created wrapper layers, this should give you the intuition on how to model audio samples.<\/p>\n\n<p>[1] <a href=\"https:\/\/github.com\/basveeling\/wavenet\/blob\/master\/wavenet.py#L220\" rel=\"nofollow noreferrer\">https:\/\/github.com\/basveeling\/wavenet\/blob\/master\/wavenet.py#L220<\/a><\/p>\n"}]}
{"QuestionId":41798311,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2017-01-23T02:08:41.873","AcceptedAnswerId":null,"OwnerUserId":6527887.0,"Title":"Tensorflow : Graph is finalized and cannot be modified","Body":"<p>I am trying to save variables through checkpoints to introduce fault tolerance to my program. I am trying to achieve this by using the MonitoredTrainingSession function. The following is my configuration:-<\/p>\n\n<pre><\/pre>\n\n<p>I get the following <strong>RuntimeError<\/strong> which I am unable to resolve:-<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41798401","CreationDate":"2017-01-23T02:19:46.110","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>The root cause for your error seems to be that MonitoredTrainingSession has finalized (frozen) the graph and your <code>tf.global_variable_initializer()<\/code> is no longer able to modify it.<\/p>\n\n<p>Having said that, there are multiple things that require attention:<\/p>\n\n<p>1) Why do you try to repeatedly initialize all variables here? <\/p>\n\n<pre><code>while not sess.should_stop():\n    sess.run(tf.global_variables_initializer())\n<\/code><\/pre>\n\n<p>2) It seems some of your code is already included in <code>MonitoredTrainingSession<\/code>, e.g. <code>ChiefSessionCreator<\/code>. Can you please take another look at the code (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/monitored_session.py#L243\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/monitored_session.py#L243<\/a>) or search for its sample usage and see how <code>MonitoredTrainingSession<\/code> is supposed to be used?<\/p>\n"},{"AnswerId":"50615216","CreationDate":"2018-05-31T00:56:15.183","ParentId":null,"OwnerUserId":"7211137","Title":null,"Body":"<p>This may not be recommended for your use case, but it is possible to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/48d13f510b70e48b61a435011f78bfef4df55c9b\/tensorflow\/python\/framework\/ops.py#L3122-L3130\" rel=\"nofollow noreferrer\">unfinalize a Graph<\/a>:<\/p>\n\n<pre><code>sess.graph._unsafe_unfinalize()\n<\/code><\/pre>\n"},{"AnswerId":"45212556","CreationDate":"2017-07-20T10:48:44.137","ParentId":null,"OwnerUserId":"7762401","Title":null,"Body":"<p>If you want to initialize the graph on loop, you can use the function to create new graph on top of loop.<\/p>\n\n<pre><code>import tensorflow as tf\n\ntf.reset_default_graph()\ntf.Graph().as_default()\n<\/code><\/pre>\n"},{"AnswerId":"45268933","CreationDate":"2017-07-23T19:08:33.550","ParentId":null,"OwnerUserId":"890824","Title":null,"Body":"<p>Since your aim is to use <code>MonitoredTrainingSession<\/code> to get you checkpointing, the usage is much simpler than your example:<\/p>\n\n<pre><code>import tensorflow as tf\n\nglobal_step = tf.contrib.framework.get_or_create_global_step()\nx = tf.constant(2)\ny1 = x + 300\ny2 = x**2\ny3 = x * 5\ny0 = x - 66\ny = y0 + y1 + y2 + y3\nstep = tf.assign_add(global_step, 1)\n\nwith tf.train.MonitoredTrainingSession(checkpoint_dir='\/tmp\/checkpoints') as sess:\n    while not sess.should_stop():\n        result, i = sess.run([y, step])\n        print(result, i)\n<\/code><\/pre>\n\n<ul>\n<li>The hooks for saving\/restoring checkpoints are created by <code>MonitoredTrainingSession<\/code> for you.<\/li>\n<li>If you pass in <code>save_checkpoint_secs<\/code> you can change the frequency of checkpointing from the 10 minute default. I find a higher frequency isn't worth it: saving checkpoints isn't free, so very frequent checkpointing will end up slowing training down.<\/li>\n<li>The <code>ChiefSessionCreator<\/code> and gRPC config is only needed for distributed running (see <a href=\"https:\/\/www.tensorflow.org\/deploy\/distributed\" rel=\"nofollow noreferrer\">here<\/a> for a description of the concepts. Similarly with assigning ops to specific devices - make sure you really need to do this before using it as it can slow things down if you're not careful.<\/li>\n<li>You don't need to wrap the result of operations on tensors with <code>tf.Variable()<\/code> - they already are variables.<\/li>\n<li>You can pass <code>save_summaries_steps<\/code> for monitoring training with tensorboard, but by default that'll happen every 100 steps anyway.<\/li>\n<\/ul>\n"}]}
{"QuestionId":41798359,"AnswerCount":0,"Tags":"<neural-network><classification><deep-learning><keras><conv-neural-network>","CreationDate":"2017-01-23T02:13:49.060","AcceptedAnswerId":null,"OwnerUserId":1637673.0,"Title":"How to set proper arguments to build keras Convolution2D NN model [Text Classification]?","Body":"<p>I am trying to use 2D CNN to do text classification on Chinese Article and have trouble on setting arguments of keras . I know the basic flow of  to cope with image, but stuck by using my dataset with keras.<\/p>\n\n<h2>Input data<\/h2>\n\n<pre><\/pre>\n\n<h3>Code for building model<\/h3>\n\n<pre><\/pre>\n\n<p><strong>Question 1<\/strong>. I have problem to set  arguments. My reseach is below,<\/p>\n\n<p>The official docs do not contain an exmaple for 2D CNN text classifacation(though has 1D CNN). <\/p>\n\n<p> defination is here <a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/convolutional\/<\/a>:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>nb_filter: Number of convolution filters to use.<br>\n    nb_row: Number of rows in the convolution kernel. \n    nb_col: Number of columns in the convolution kernel. \n    border_mode: 'valid', 'same' or 'full'. ('full' requires the Theano backend.)<\/p>\n<\/blockquote>\n\n<p>Some research about the arguments:<\/p>\n\n<ul>\n<li><p>This issue <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/233\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/233<\/a> is about 2D CNN for text classification, I read all comments and pick:<\/p>\n\n<p>(1) <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/233#issuecomment-117427013\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/233#issuecomment-117427013<\/a><\/p>\n\n<pre><\/pre>\n\n<p>(2) <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/233#issuecomment-117700913\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/233#issuecomment-117700913<\/a><\/p>\n\n<pre><\/pre>\n\n<p>But it seems has some diference to current keras version, also the arguments naming by different people are in a mess (I hope keras has an easy understandable argument expanation).<\/p><\/li>\n<li><p>Another comment I see about current api:<\/p>\n\n<p><a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1665#issuecomment-181181000\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/1665#issuecomment-181181000<\/a><\/p>\n\n<blockquote>\n  <p>The current API is as below:<\/p>\n\n<pre><\/pre>\n  \n  <p>So (36,1,7,7) seems the reason, the correct arguments would be (36,7,7,...).<\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<p>By above research, on my understanding of convolution,  create a (nb_filter, nb_row, nb_col) filter , by sliding a stride to get one filter result, recurse sliding, finally combine the result into array with shape (1, one_sample_article_length[6810] \/ nb_filter), and go to the next layer, is that right? Is my code below set  and  correct ?<\/p>\n\n<p><strong>Question 2<\/strong>. What is the proper MaxPooling2D arguments?  (for my dateset or for commonm, either is OK)<\/p>\n\n<p>I refer this issue <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/233#issuecomment-117427013\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/233#issuecomment-117427013<\/a> to set the argument, there are two kinds:<\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<\/ul>\n\n<p>I have no idea why they calculate  argument like that.<\/p>\n\n<p><strong>Question 3<\/strong>. Any recommendation for  and  to do such text classification? I have no idea at all.<\/p>\n","answers":[]}
{"QuestionId":41798519,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><lstm><recurrent-neural-network>","CreationDate":"2017-01-23T02:39:08.800","AcceptedAnswerId":null,"OwnerUserId":2847699.0,"Title":"State resetting in LSTMs during training and testing","Body":"<p>I am trying to understand and implement LSTMs. I understand that they one needs to define a sequence length T, and the training is performed in batches. So we fed to the network several sequences of length T. Now the LSTM needs a previous state as input, which as I understand, it is initialized to zero. My question is, is the state reset to zero after every sequence? for example I have a sequence 1, the state vector is carried forward in this sequence, and then I put it to zero for the next sequence? Or is it carried it to the next sequence 2? If so, How is it done for unrelated sequences; for example I have samples from 2 different texts, and It would not make sense to carry the state from text 1 to text 2; how is this handled in practice?\nAbout the testing time, the state vector is initialized as zero and carried for the whole sequence, or is it reset after each sub-sequence?<\/p>\n\n<p>Note: I put tag this also in Tensorflow since is the framework that I am using, and maybe someone from there can help me also.<\/p>\n","answers":[{"AnswerId":"41798597","CreationDate":"2017-01-23T02:50:16.877","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>In Tensorflow, I am 95% sure the starting state for every sequence is reset to zero for every element in your batch and between batches. (5% because the \"Never say never\" rule :)<\/p>\n\n<p>EDIT:<\/p>\n\n<p>I should probably elaborate more. How Tensorflow works, it first constructs a graph and then pushes your data through. When you look at the recurrent graph you constructed, I believe that you'll see its head (first state) is connected to zero, which means every time you push data through the graph (e.g. through sess.run()), it will get a new zero from the zero generator, hence its old state from previous runs if forgotten.<\/p>\n"},{"AnswerId":"50431710","CreationDate":"2018-05-20T04:36:38.803","ParentId":null,"OwnerUserId":"802861","Title":null,"Body":"<p>It depends on how you implement batch processing of RNN.\nTheoretically, the state is very important in processing the series of the data so, you should not reset the state if the whole sequence of data is not finished.\nSo, usually you need to reset when the epoch finishes and you should not reset when just one batch cycle finishes.<\/p>\n"}]}
{"QuestionId":41798975,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><caffe><dimensionality-reduction>","CreationDate":"2017-01-23T03:44:59.493","AcceptedAnswerId":null,"OwnerUserId":4928920.0,"Title":"Reduce dimensions of model's fully connected layer for image retrieval task","Body":"<p>I'm working on a image retrieval task(not involving faces) and one of the things I am trying is to swap out the softmax layer in the CNN model and use the <a href=\"http:\/\/shogun-toolbox.org\/static\/notebook\/current\/LMNN.html\" rel=\"nofollow noreferrer\">LMNN<\/a> classifier. For this purpose I fine tuned the model and then extracted the features at fully connected layer. I have about 3000 images right now. The fully connected layer gives a 4096 dim vector. So my final vector is a 3000x4096 vector with about 700 classes(Each class has 2+ images). I believe this is an extremely large dimension size which the LMNN algorithm is going to take forever(it really did take forever). \nHow can I reduce the number of dimensions? I tried PCA but that didn't squeeze down the dimensions too much(got down to 3000x3000). I am thinking 256\/512\/1024 dim vector should be able to help. If I were to add another layer to reduce dimensions, say a new fully connected layer would I have to fine tune my network again? Inputs on how to do that would be great!\nI am also currently trying to augment my data to get more images per class and increase the size of my dataset.  <\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"41799785","CreationDate":"2017-01-23T05:30:58.293","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p>PCA should let you reduce the data further - you should be able to specify the desired dimensionality - see <a href=\"https:\/\/en.wikipedia.org\/wiki\/Principal_component_analysis#Dimensionality_reduction\" rel=\"nofollow noreferrer\">the wikipedia article<\/a>.<\/p>\n\n<p>As well as PCA you can try <a href=\"https:\/\/en.wikipedia.org\/wiki\/T-distributed_stochastic_neighbor_embedding\" rel=\"nofollow noreferrer\">t-distributed stochastic neighbor embedding<\/a> (t-SNE). I really enjoyed <a href=\"http:\/\/distill.pub\/2016\/misread-tsne\/\" rel=\"nofollow noreferrer\">Wattenberg, et al.'s article<\/a> - worth a read if you want to get an insight into how it works and some of the pitfalls.<\/p>\n\n<p>In a neural net the standard way to reduce dimensionality is by adding more, smaller layers, as you suggested. As they can only learn during training, you'll need to re-run your fine-tuning. Ideally you would re-run the entire training process if you make a change to the model structure but if you have enough data it may be OK still.<\/p>\n\n<p>To add new layers in TensorFlow, you would add a fully connected layer whose input is the output of your 3000 element layer, and output size is the desired number of elements. You may repeat this if you want to go down gradually (e.g. 3000 -> 1024 -> 512). You would then perform your training (or fine tuning) again.<\/p>\n\n<p>Lastly, I did a quick search and found this paper that claims to support LMNN over large datasets through random sampling. You might be able to use that to save a few headaches: <a href=\"http:\/\/dl.acm.org\/citation.cfm?id=2919395\" rel=\"nofollow noreferrer\">Fast LMNN Algorithm through Random Sampling<\/a><\/p>\n"}]}
{"QuestionId":41799593,"AnswerCount":1,"Tags":"<error-handling><tensorflow><install><anaconda>","CreationDate":"2017-01-23T05:11:32.937","AcceptedAnswerId":null,"OwnerUserId":7455921.0,"Title":"Im trying to install tensorflow on Anaconda that supports python 2.7 and I'm stuck","Body":"<p>Im trying to install tensorflow on Anaconda that supports python 2.7, I did as the manual given on the tensorflow homepage told me to do, and this error occurs every time.<\/p>\n\n<p>UnsatisfiableError: The following specifications were found to be in conflict:\n  - python 2.7*\n  - tensorflow -> protobuf ==3.1.0 -> python 3.5*\nUse \"conda info \" to see the dependencies for each package.<\/p>\n\n<p>could you help me out??<\/p>\n\n<p>and this is my conda info <\/p>\n\n<p>conda info\nCurrent conda install:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42597885","CreationDate":"2017-03-04T15:43:39.567","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>As said in the comments, on Microsoft Windows TensorFlow only runs on Python 3.5 64-bit.<\/p>\n\n<p>Python 2.7 is unsupported because TensorFlow uses many C++11 features (e.g., <code>auto<\/code>), but <a href=\"https:\/\/web.archive.org\/web\/20170304154430\/https:\/\/docs.python.org\/2.7\/using\/windows.html#compiling-python-on-windows\" rel=\"nofollow noreferrer\">the version of the Microsoft Visual C++ Compiler used for Python 2.7  is 2008<\/a>.<\/p>\n"}]}
{"QuestionId":41799692,"AnswerCount":1,"Tags":"<python><memory><neural-network><out-of-memory><keras>","CreationDate":"2017-01-23T05:23:23.043","AcceptedAnswerId":null,"OwnerUserId":1637673.0,"Title":"Any way to optimize large input's memory usage in keras?","Body":"<p>I am trying to use 2D CNN to do text classification on Chinese articles and have some trouble of keras . I know the basic flow of  to cope with image, but stuck by using my dataset with keras. This is one of my problems:<\/p>\n\n<h2>Dataset<\/h2>\n\n<ol>\n<li><p>9800 Chinese Article.<\/p>\n\n<p>negative article and non-negtive article[please note it may be positive or neutrality] , just a binary classification problem. I have a test on  NN, but the result is not good.<\/p><\/li>\n<li><p>Use tokenizer and word2vec to transform to a shape .<\/p>\n\n<p>longest article has 6810 word, shortest article has less 50 word\uff0c need padding all article to 6810\uff0c 200 is word2vec size(Seems some people call it embedding_size ? ). Format like:<\/p>\n\n<pre><\/pre><\/li>\n<\/ol>\n\n<p>Is the article max. word length 6810 too large? \nI have to reduce the 9800 samples to 6500 to avoid a , because 6500 already eats all my 32GB RAM. Any way to optimize the memory usage except trimming all the articles to a shorter length?<\/p>\n","answers":[{"AnswerId":"41799776","CreationDate":"2017-01-23T05:30:28.373","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>The <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-use-keras-with-datasets-that-dont-fit-in-memory\" rel=\"noreferrer\">Keras FAQ<\/a> already answers this question partly. You can load your data in chunks using <a href=\"https:\/\/keras.io\/models\/sequential\/#fit_generator\" rel=\"noreferrer\"><code>model.fit_generator()<\/code><\/a>. The generator runs in a separate thread and produces your mini-batches, possibly loading them from your archive one-by-one, avoiding loading everything into RAM at once.<\/p>\n\n<p>The code for using this would roughly look like this:<\/p>\n\n<pre><code>def train_generator():\n    while True:\n        chunk = read_next_chunk_of_data()\n        x,y = extract_training_data_from_chunk(chunk)\n        yield (x,y)\n\n model.fit_generator(generator=train_generator())\n<\/code><\/pre>\n"}]}
{"QuestionId":41799988,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2017-01-23T05:50:34.243","AcceptedAnswerId":null,"OwnerUserId":3975448.0,"Title":"attempt to call method 'random' (a nil value) in Lua","Body":"<p>Below is my code<\/p>\n\n<pre><\/pre>\n\n<p>I am running above using the following command<\/p>\n\n<pre><\/pre>\n\n<p>It gives the following error<\/p>\n\n<pre><\/pre>\n\n<p>I know that there is some error in the line<\/p>\n\n<pre><\/pre>\n\n<p>But I am not able to figure out.<\/p>\n","answers":[{"AnswerId":"41800510","CreationDate":"2017-01-23T06:36:32.627","ParentId":null,"OwnerUserId":"2858170","Title":null,"Body":"<blockquote>\n  <p>luajit: feedforwad.lua:13: attempt to call method 'random' (a nil\n  value)<\/p>\n  \n  <p>Is not \"some error\" and you should not have problems to figure out what is wrong because the error message tells you exactly what is wrong.<\/p>\n<\/blockquote>\n\n<p>You tried to call the a method named <code>random<\/code> which happens to be a <code>nil<\/code> value.\nThis means that there is no function with that name and therefor you can't call it.<\/p>\n\n<p>According to the reference documentation (which you should have checked befor coming here) the function is actually named <code>rand<\/code><\/p>\n"}]}
{"QuestionId":41800366,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2017-01-23T06:25:49.860","AcceptedAnswerId":null,"OwnerUserId":7456105.0,"Title":"Tensorflow Tensor object not iterable","Body":"<p>all:\n    I am new to tensorflow. I used to use theano as keras backend.\n    When I try to use tensorflow instead of theano, I got error saying:\n    <strong>'Tensor' object not iterable<\/strong>\n    when call keras.backend.shape() function. Further more, I found this\n    error is occurred by tensorflow.shape(). However, when I test         tensorflow.shape\n    function with keras.backend.variables() instance, it goes well.\n    Can anybody tell me what's going on ?\n    keras version: 1.2.1\n    tensorflow version: 0.12.1.<\/p>\n\n<p>Thanks you all guys in advance.<\/p>\n","answers":[]}
{"QuestionId":41800604,"AnswerCount":1,"Tags":"<deep-learning><caffe><logistic-regression><cross-entropy>","CreationDate":"2017-01-23T06:44:30.553","AcceptedAnswerId":null,"OwnerUserId":1363774.0,"Title":"Need help understanding the Caffe code for SigmoidCrossEntropyLossLayer for multi-label loss","Body":"<p>I need help in understanding the Caffe function, <strong><\/strong>, which is the cross-entropy error with logistic activation.<\/p>\n\n<p>Basically, the cross-entropy error for a single example with N independent targets is denoted as:<\/p>\n\n<pre><\/pre>\n\n<p>where  is the target, 0 or 1, and  is the output, indexed by . , of course goes through a logistic activation.<\/p>\n\n<p>An algebraic trick for quicker cross-entropy calculation reduces the computation to: <\/p>\n\n<pre><\/pre>\n\n<p>and you can verify that from Section 3 <a href=\"https:\/\/www.ics.uci.edu\/~pjsadows\/notes.pdf\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>The question is, how is the above translated to the loss calculating code below: <\/p>\n\n<pre><\/pre>\n\n<p>Thank you.<\/p>\n\n<p>The function is reproduced below for convenience.  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41802485","CreationDate":"2017-01-23T09:04:28.107","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>In the expression <code>log(1 + exp(x[i]))<\/code> you might encounter numerical instability in case <code>x[i]<\/code> is very large. To overcome this numerical instability, one scales the sigmoid function like this:<\/p>\n\n<pre><code> sig(x) = exp(x)\/(1+exp(x)) \n        = [exp(x)*exp(-x(x&gt;=0))]\/[(1+exp(x))*exp(-x(x&gt;=0))]\n<\/code><\/pre>\n\n<p>Now, if you plug the new and stable expression for <code>sig(x)<\/code> into the loss you'll end up with the same expression as caffe is using.<\/p>\n\n<p>Enjoy!<\/p>\n"}]}
{"QuestionId":41800643,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><convolution>","CreationDate":"2017-01-23T06:47:26.180","AcceptedAnswerId":null,"OwnerUserId":3102470.0,"Title":"ValueError: Dimensions must be equal, but are 1 and 3 for 'Conv2D' (op: 'Conv2D') with input shapes: [1,400,400,1], [1,3,3,1]","Body":"<p>i am trying to create a simple one layer convolution for a 2D input the idea is just to have an input image , kernel and an output the code is :<\/p>\n\n<pre><\/pre>\n\n<p>and i am getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>my input is 400x400x1 and the kernel is 3x3 <\/p>\n","answers":[{"AnswerId":"41880215","CreationDate":"2017-01-26T18:15:23.170","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>Based on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/convolution#conv2d\" rel=\"noreferrer\">conv2d doc<\/a>:<\/p>\n\n<pre><code>shape of input = [batch, in_height, in_width, in_channels]\nshape of filter = [filter_height, filter_width, in_channels, out_channels]\n<\/code><\/pre>\n\n<p>Last dimension of input and third dimension of filter represents the number of input channels. In your case they are not equal.<\/p>\n\n<pre><code>You can change the shape of filter to [3, 3, 1, 1].\n<\/code><\/pre>\n"}]}
{"QuestionId":41800983,"AnswerCount":1,"Tags":"<c++><windows><deep-learning><caffe>","CreationDate":"2017-01-23T07:13:12.273","AcceptedAnswerId":43039488.0,"OwnerUserId":3246146.0,"Title":"Input Layer type: ImageData in windows caffe cpp giving Blank Output","Body":"<p>I am working on Image Segmentation problem using caffe in windows using cpp. I am using the \"Imagedata\" input type to train the network but I am getting the Blank output while testing. Can anyone help me to analyse the issue.<\/p>\n\n<pre><\/pre>\n\n<p>File_Triangle.txt and File_label_triangle.txt has the absolute path of the image location and a dummy label.\ne.g D:\\00000032.png 0<\/p>\n\n<pre><\/pre>\n\n<p>Code snippet for training in cpp<\/p>\n\n<pre><\/pre>\n\n<p>After Training I am using .caffemodel to test the network.<\/p>\n\n<pre><\/pre>\n\n<p>The Code Snippet used for testing.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43039488","CreationDate":"2017-03-27T06:41:49.677","ParentId":null,"OwnerUserId":"3246146","Title":null,"Body":"<p>I got the issue. The network is giving proper output but the mistake is with dumping it. The Network is giving output in float (i.e at upsample layer) and it is not in a normalize form. Below modifications is giving proper output.<\/p>\n\n<pre><code>const shared_ptr&lt;Blob&lt;float&gt; &gt;&amp; concat_blob = net_-&gt;blob_by_name(\"upsample\");\nconst float* concat_out = concat_blob-&gt;cpu_data();\n\ncv::Mat matout(height, width, CV_32FC1);\nfor (int i = 0; i &lt; height; i++)\n{\n    for (int j = 0; j &lt; width; j++)\n    {\n         matout.at&lt;float&gt;(i, j) = (float)(concat_out[i*width + j]);\n    }\n}\ncv::normalize(matout, matout, 0, 255, CV_MINMAX);\ncv::imwrite(\"output image path\", matout);\n<\/code><\/pre>\n"}]}
{"QuestionId":41801084,"AnswerCount":1,"Tags":"<python><csv><tensorflow>","CreationDate":"2017-01-23T07:21:40.150","AcceptedAnswerId":null,"OwnerUserId":5074987.0,"Title":"Tensorflow text parsing","Body":"<p>I am trying to parse text with Tensorflow. Lets say I have CSV file that looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>And I want the output to be an array of the values:<\/p>\n\n<pre><\/pre>\n\n<p>How would you do it? I was thinking about creating some kind of hash for each word (value) for the input, but I have no idea what to do for the output.<\/p>\n\n<p>I am not parsing CSV files, I only used them as an example.<\/p>\n","answers":[{"AnswerId":"41880494","CreationDate":"2017-01-26T18:33:34.943","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<pre><code>reader = tf.TextLineReader()\n_, lines = reader.read(...)\n\n# if your data is ccv\nlist-of-tensors = tf.decode_csv(lines, ...)\n\n# if your data is just a line of text\na-sparse-tensor = tf.string_split(lines, ...)\n<\/code><\/pre>\n"}]}
{"QuestionId":41801918,"AnswerCount":1,"Tags":"<image-processing><computer-vision><deep-learning><caffe><image-segmentation>","CreationDate":"2017-01-23T08:28:46.597","AcceptedAnswerId":41802667.0,"OwnerUserId":3755376.0,"Title":"How fully convolutional layer can work for segmentation task?","Body":"<p>I am reading the paper \"Fully Convolutional Networks for Semantic Segmentation by Jonathan Long*, Evan Shelhamer*, and Trevor Darrell. CVPR 2015 and PAMI 2016\"\nI want to understand why it can work for semantic segmentation. Let look at the fcn-32s architecture, it includes two phases: feature extraction (conv1-1->pool5), and feature classification (fc6->score_fr). Comparison with a normal classification network, the main different is the second phase. The FCN-32s replaces the fully connected layer by fully convolution layers (1 x 1) in fc7 to retain the spatial map (as the caption in the figure 2 of the paper). Hence, I was confused something about this point:<\/p>\n\n<ol>\n<li>If we replace the fully connected layer by fully convolution layer,\nhow can it learn the weight as traditional classification\narchitecture?<\/li>\n<li>Why we can retain the spatial map (heatmap) by using fully convolution layer?<\/li>\n<\/ol>\n\n<p>Thank you in advance.<\/p>\n\n<p>Update: This is figure to show how convert from fully connected to fully convolution layers\n<a href=\"https:\/\/i.stack.imgur.com\/8WsAK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8WsAK.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41802667","CreationDate":"2017-01-23T09:15:26.070","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you look at the math, <code>\"Convolution\"<\/code> layer and <code>\"InnerProduct\"<\/code> (aka \"fully connected\") layer are basically quite similar: they perform a linear operation on their respective receptive fields. The only difference is that <code>\"InnerProduct\"<\/code> takes the entire input as its \"receptive field\" while <code>\"Convolution\"<\/code> layer only looks at <code>kernel_size<\/code> window in the input.<\/p>\n\n<p>What happens if the input size is changed?<br>\n<code>\"Convolution\"<\/code> layer cannot care less, it simply outputs the feature map with spatial dimensions corresponding to the new input shape.<br>\nOn the other hand, <code>\"InnerProduct\"<\/code> layer fails since the number of weights it has does not match the new size of the receptive field.<\/p>\n\n<p>Replacing the top fully-connected layers in a model with <code>\"Convolution\"<\/code> layers allow for \"sliding window\" classification of the image: thus achieving coarse semantic segmentation - a labeling per-pixel rather than a label per image.  <\/p>\n\n<p>There is still a big issue of the scale gap between the input scale and the coarse scale of the output labels, but there are <code>\"Deconvolution\"<\/code> layers to bridge that gap.<\/p>\n"}]}
{"QuestionId":41802810,"AnswerCount":1,"Tags":"<arrays><numpy><insert><tensorflow>","CreationDate":"2017-01-23T09:23:03.217","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Inserting average rows in tensorflow tensor","Body":"<p>I have the following function in numpy:<\/p>\n\n<pre><\/pre>\n\n<p>which basically takes the average of the above and below rows and inserts it between the two, at every n intervals. <\/p>\n\n<p>Any idea on how I can do this one in Tensorflow? Specifically what can I use instead of np.insert, since there does not seem to be an equivalent function.<\/p>\n","answers":[{"AnswerId":"41803127","CreationDate":"2017-01-23T09:40:16.453","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>You could use an initialization based approach to solve it, like so -<\/p>\n\n<pre><code>def insert_row_averages_iniitalization_based(A, n=1):\n\n    # Slice and compute averages   \n    slice2 = A[n::n]\n    v = (A[n-1::n][:slice2.shape[0]] + slice2)\/2.0\n\n    # Compute number of rows for o\/p array\n    nv = v.shape[0]\n    nrows = A.shape[0] + nv\n\n    # Row indices where the v values are the inserted in o\/p array\n    v_rows = (n+1)*np.arange(nv)+n\n\n    # Initialize o\/p array\n    out = np.zeros((nrows,A.shape[1]))\n\n    # Insert\/assign v in output array\n    out[v_rows] = v  # Assign v\n\n    # Create 1D mask of length equal to no. of rows in o\/p array, such that its\n    # TRUE at places where A is to be assigned and FALSE at places where values\n    # from v were assigned in previous step.\n    mask = np.ones(nrows,dtype=bool)\n    mask[v_rows] = 0\n\n    # Use the mask to assign A.\n    out[mask] = A\n    return out\n<\/code><\/pre>\n"}]}
{"QuestionId":41803853,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-23T10:15:13.723","AcceptedAnswerId":43677545.0,"OwnerUserId":7456923.0,"Title":"Extracting specific elements from a tensor in tensorflow","Body":"<p>I'm using tensorflow on python\nI have a data tensor of shape [?, 5, 37], and a idx tensor of shape [?, 5]<\/p>\n\n<p>I'd like to extract elements from data and get an output of shape [?, 5] such that:<\/p>\n\n<pre><\/pre>\n\n<p>It looks loke the tf.gather_nd() function is the closest to my needs, but I don't see how to use it it my case...<\/p>\n\n<p>Thanks !<\/p>\n\n<p>EDIT : I managed to do it with gather_nd as shown below, but is there a better option ? (it seems a bit heavy-handed)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43677545","CreationDate":"2017-04-28T10:05:34.333","ParentId":null,"OwnerUserId":"7456923","Title":null,"Body":"<p>I managed to do it with <code>gather_nd<\/code> as shown below<\/p>\n\n<pre><code>nRows = tf.shape(length_label)[0] # ==&gt; ?\nnCols = tf.constant(MAX_LENGTH_INPUT + 1, dtype=tf.int32) # ==&gt; 5\nm1 = tf.reshape(tf.tile(tf.range(nCols), [nRows]),\n                                       shape=[nRows, nCols])\nm2 = tf.transpose(tf.reshape(tf.tile(tf.range(nRows), [nCols]),\n                                        shape=[nCols, nRows]))\nindices = tf.pack([m2, m1, idx], axis=-1)\n# indices should be of shape [?, 5, 3] with indices[i,j]==[i,j,idx[i,j]]\noutput = tf.gather_nd(data, indices=indices)\n<\/code><\/pre>\n"}]}
{"QuestionId":41804206,"AnswerCount":1,"Tags":"<python><csv><tensorflow><neural-network><batch-processing>","CreationDate":"2017-01-23T10:33:02.777","AcceptedAnswerId":null,"OwnerUserId":7096238.0,"Title":"make batches with own csv file python","Body":"<p>I have a python code for the MNIST dataset using Tensorflow.\nSession is the following:<\/p>\n\n<pre><\/pre>\n\n<p>The line:<\/p>\n\n<pre><\/pre>\n\n<p>is making everytime new batches with the size 100.<\/p>\n\n<p>My question is, how to replace this line with a new line that makes for me new batches if I have my own CSV file (which is a list of lists)? \nMy current code looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Where the \"nextbatch\" is a function I defined. But I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I appreciate any suggestion :D<\/p>\n\n<p>By the way, \"training_data_list\" comes from:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41880681","CreationDate":"2017-01-26T18:43:31.727","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>You need to implement an object which handles the indexing.\nYou need to implement nextbatch function inside that object.\nYou can look at implementation of nextbatch in the mnist.<\/p>\n"}]}
{"QuestionId":41804380,"AnswerCount":1,"Tags":"<matrix><tensorflow><gpu><nvidia>","CreationDate":"2017-01-23T10:40:42.677","AcceptedAnswerId":41810634.0,"OwnerUserId":470184.0,"Title":"Testing GPU with tensorflow matrix multiplication","Body":"<p>As many machine learning algorithms rely to matrix multiplication(or at least can be implemented using matrix multiplication) to test my GPU is I plan to create matrices a , b , multiply them and record time it takes for computation to complete.<\/p>\n\n<p>Here is code that will generate two matrices of dimensions 300000,20000 and multiply them :<\/p>\n\n<pre><\/pre>\n\n<p>Is this a sufficient test to compare performance of GPU's ? What other factors should I consider ?<\/p>\n","answers":[{"AnswerId":"41810634","CreationDate":"2017-01-23T16:06:56.423","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's an <a href=\"https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/matmul_benchmark.py\" rel=\"noreferrer\">example<\/a> of a matmul benchmark which avoids common pitfalls, and matches the official 11 TFLOP mark on Titan X Pascal.<\/p>\n\n<pre><code>import os\nimport sys\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"1\"\nimport tensorflow as tf\nimport time\n\nn = 8192\ndtype = tf.float32\nwith tf.device(\"\/gpu:0\"):\n    matrix1 = tf.Variable(tf.ones((n, n), dtype=dtype))\n    matrix2 = tf.Variable(tf.ones((n, n), dtype=dtype))\n    product = tf.matmul(matrix1, matrix2)\n\n\n# avoid optimizing away redundant nodes\nconfig = tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))\nsess = tf.Session(config=config)\n\nsess.run(tf.global_variables_initializer())\niters = 10\n\n# pre-warming\nsess.run(product.op)\n\nstart = time.time()\nfor i in range(iters):\n  sess.run(product.op)\nend = time.time()\nops = n**3 + (n-1)*n**2 # n^2*(n-1) additions, n^3 multiplications\nelapsed = (end - start)\nrate = iters*ops\/elapsed\/10**9\nprint('\\n %d x %d matmul took: %.2f sec, %.2f G ops\/sec' % (n, n,\n                                                            elapsed\/iters,\n                                                            rate,))\n<\/code><\/pre>\n"}]}
{"QuestionId":41805754,"AnswerCount":1,"Tags":"<python><validation><tensorflow><neural-network><keras>","CreationDate":"2017-01-23T11:51:56.503","AcceptedAnswerId":null,"OwnerUserId":5760497.0,"Title":"Neural Networks gives different results when trained with a different data set permutation, why?","Body":"<p>I have a neural network using keras with a tensorflow backend:<\/p>\n\n<pre><\/pre>\n\n<p>I am testing drop-outs from a public colleges with their socio-economic parameters as variables, initially I have 8 csv files (named a,b,c,d,e,f,g and h) with 12 column headers and 300,000 rows. The result is binary, 0 for retained and 1 for dropped,I normalized the data before feeding it to the NN.<\/p>\n\n<p>My first training set was a,b,c,d,e and f, with g and h as hold out for testing. the Neural networks provided me with a good specificity, sensitivity and accuracy of :70%, 65% and 66%.<\/p>\n\n<p>With that I trained another NN of the same architecture as stated above this time my training datasets are c,d,e,f, g and h with a and b as my new hold-out for testing, but then the model provides a very bad result for specificity, sensitivity and accuracy: 42%, 48% and 47%, I am wondering why? Are there any published papers citing this kind of phenomenon in neural networks?<\/p>\n\n<p>thanks!<\/p>\n","answers":[{"AnswerId":"41805927","CreationDate":"2017-01-23T12:02:36.607","ParentId":null,"OwnerUserId":"1055118","Title":null,"Body":"<p>Many machine learning methods can suffer from a problem known as <strong>over-fitting<\/strong>.\n<a href=\"https:\/\/en.wikipedia.org\/wiki\/Overfitting\" rel=\"nofollow noreferrer\">Wikipedia<\/a> gives a variety of refernces to this.<\/p>\n\n<p>The reason you at least use a hold-out data set is to test how well your trained model fits unseen data. In theory you could be 100% accurate on one data set and yet perform very badly on new data.<\/p>\n\n<p>Some people use cross-validation rather than just one or tow held back data sets - this will try each data point in a test and a training set. For example with 10 data points, use 9 to train and try to fit the tenth one. Then do this for each permutation.<\/p>\n\n<p>This can be appropriate if the various patterns are not evenly distribtued in a data set.<\/p>\n\n<p>If one of your training sets has all drops outs, then a model predicting everyone drops out will fit this best, but will not generalise to any data with no drop outs.<\/p>\n\n<p>It is often worth doing some exploitary data analysis to see if some of your data sets are not representative.<\/p>\n"}]}
{"QuestionId":41806128,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-01-23T12:14:32.990","AcceptedAnswerId":49363189.0,"OwnerUserId":5107084.0,"Title":"Tensorflow: Is preprocessing on TFRecord files faster than real-time data preprocessing?","Body":"<p>In Tensorflow, it seems that preprocessing could be done on either during training time, when the batch is created from raw images (or data), or when the images are already static. Given that <strong><em>theoretically<\/em><\/strong>, the preprocessing should take roughly equal time (if they are done using the same hardware), is there any <strong><em>practical<\/em><\/strong> disadvantage in doing data preprocessing (or even data augmentation) before training than during training in real-time?<\/p>\n\n<p>As a side question, could data augmentation even be done in Tensorflow if was not done during training?<\/p>\n","answers":[{"AnswerId":"42111038","CreationDate":"2017-02-08T10:49:12.967","ParentId":null,"OwnerUserId":"7533845","Title":null,"Body":"<p>I have been wondering the same thing and have been disappointed with my during-training-time image processing performance. It has taken me a while to appreciate quite how big an overhead the image manipulation can be.<\/p>\n\n<p>I am going to make myself a nice fat juicy preprocessed\/augmented data file. Run it overnight and then come in the next day and be twice as productive!<\/p>\n\n<p>I am using a single GPU machine and it seems obvious to me that piece-by-piece model building is the way to go. However, the workflow-maths may look different if you have different hardware. For example, on my Macbook-Pro tensorflow was slow (on CPU) and image processing was blinding fast because it was automatically done on the laptop's GPU. Now I have moved to a proper GPU machine, tensorflow is running 20x faster and the image processing is the bottleneck.<\/p>\n\n<p>Just work out how long your augmentation\/preprocessing is going to take, work out how often you are going to reuse it and then do the maths.<\/p>\n"},{"AnswerId":"49363189","CreationDate":"2018-03-19T12:38:21.613","ParentId":null,"OwnerUserId":"6409572","Title":null,"Body":"<blockquote>\n  <p>Is there any practical disadvantage in doing data preprocessing (or\n  even data augmentation) before training than during training in\n  real-time?<\/p>\n<\/blockquote>\n\n<p>Yes, there are advantages (<strong>+++<\/strong>) and disadvantages (<strong>---<\/strong>):<\/p>\n\n<p>Preprocessing <strong>before training<\/strong>:<\/p>\n\n<ul>\n<li><strong>---<\/strong> preprocessed samples need to be stored: disk space consumption* <em>(1)<\/em><\/li>\n<li><strong>---<\/strong> only a \"finite\" amount of samples can be generated<\/li>\n<li><strong>+++<\/strong> no runtime during training<\/li>\n<li><strong>---<\/strong>... but samples always need be read from storage, i.e. maybe storage (disk) I\/O becomes bottleneck<\/li>\n<li><strong>---<\/strong> not flexible: changing datset\/augmentation requires generating a new augmented dataset<\/li>\n<li><strong>+++<\/strong> for Tensorflow: Easily work on numpy.ndarray or other dataformats with any high-level image API (<em>open-cv, PIL, ...<\/em>) to do augmentation or even use any other language\/tool you like.<\/li>\n<\/ul>\n\n<p>Preprocessing <strong>during training<\/strong> (\"<strong><em>real-time<\/em><\/strong>\"):<\/p>\n\n<ul>\n<li><strong>+++<\/strong> infinite amount of samples can be generated (as it is generated on-the-fly)<\/li>\n<li><strong>+++<\/strong> flexible: changing dataset\/augmentation only requires changing code<\/li>\n<li><strong>+++<\/strong> if dataset fits in memory, no disk I\/O needed for data after reading once<\/li>\n<li><strong>---<\/strong> adds runtime to your training* <em>(2)<\/em><\/li>\n<li><strong>---<\/strong> for Tensorflow: Building the preprocessing as part of the <em>graph<\/em> requires working with Tensors and restricts usage of APIs working on ndarrays or other formats.* <em>(3)<\/em><\/li>\n<\/ul>\n\n<p>Some specific aspects discussed in detail: <\/p>\n\n<ul>\n<li><p><em>(1)<\/em> <strong><em>Reproducing experiments<\/em><\/strong> \"with the same data\" is kind of straightforward with a dataset generated before training. However this can be solved (even more!) elegantly with storing a seed for real-time data generation.<\/p><\/li>\n<li><p><em>(2)<\/em>: <strong><em>Training runtime for preprocessing<\/em><\/strong>: There are ways to avoid an expensive preprocessing pipeline to get in the way of your actual training. Tensorflow itself recommends filling <code>Queues<\/code> with many (CPU-)<code>threads<\/code> so that data generation can independently keep up with GPU data consumption. You can read more about this in the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/performance\/datasets_performance\" rel=\"nofollow noreferrer\">input pipeline performance guide<\/a>.<\/p><\/li>\n<li><p><em>(3)<\/em>: <strong><em>Data augmentation in tensorflow<\/em><\/strong><\/p>\n\n<blockquote>\n  <p>As a side question, could data augmentation even be done in Tensorflow\n  if was not done <strike><strong><em>during<\/em><\/strong><\/strike> <em>(I think you mean)<\/em> <strong><em>before<\/em><\/strong> training?<\/p>\n<\/blockquote>\n\n<p>Yes, tensorflow offers some functionality to do augmentation. In terms of value augmentation of scalar\/vector (or also more dimensional data), you can easily build something yourself with <code>tf.multiply<\/code> or other basic math ops. For image data, there are several ops implemented (see <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.4\/api_docs\/python\/tf\/image\" rel=\"nofollow noreferrer\">tf.image<\/a> and <a href=\"http:\/\/%20https:\/\/www.tensorflow.org\/versions\/r1.4\/api_docs\/python\/tf\/contrib\/image\" rel=\"nofollow noreferrer\">tf.contrib.image<\/a>), which should cover a lot of augmentation needs.<\/p>\n\n<p>There are off-the-shelf <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim\/preprocessing\" rel=\"nofollow noreferrer\"><strong>preprocessing examples<\/strong><\/a> on github, one of which is used and described in the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn#model_inputs\" rel=\"nofollow noreferrer\">CNN tutorial<\/a> (cifar10).<\/p><\/li>\n<\/ul>\n\n<hr>\n\n<p>Personally, I would always try to use real-time preprocessing, as generating (potentially huge) datasets feels clunky. But it is perfectly viable, I've seen it done many times and (as you see above) it definitely has it's advantages.<\/p>\n"}]}
{"QuestionId":41806689,"AnswerCount":2,"Tags":"<arrays><indexing><tensorflow>","CreationDate":"2017-01-23T12:43:04.820","AcceptedAnswerId":41808882.0,"OwnerUserId":5016028.0,"Title":"Tensorflow : Get indices of array rows which are zero","Body":"<p>For a tensor <\/p>\n\n<pre><\/pre>\n\n<p>how can I get the indices of the 0 rows? I.e. the list [1,3], in Tensorflow? <\/p>\n","answers":[{"AnswerId":"49022001","CreationDate":"2018-02-28T04:29:25.590","ParentId":null,"OwnerUserId":"8291566","Title":null,"Body":"<p>It can be done in an easier way too:    <\/p>\n\n<pre><code>    g = tf.Graph()\n    with g.as_default():\n       a = tf.placeholder(dtype=tf.float32, shape=[3, 4])\n       b = tf.placeholder(dtype=tf.float32, shape=[1, 4])\n\n       res = tf.not_equal(a, b)\n       res = tf.reduce_sum(tf.cast(res, tf.float32), 1)\n       res = tf.where(tf.equal(res1, [0.0]))[0]\n\n    with tf.Session(graph=g) as sess:\n    sess.run(tf.global_variables_initializer())\n    dict_ = {\n        a:np.array([[2.0,6.0,3.0,2.0],\n                   [1.0,8.0,32.0,1.0],\n                   [1.0,8.0,3.0,11.0]]),\n        b:np.array([[1.0,8.0,3.0,11.0]])\n    }\n\n    print(sess.run(res, feed_dict=dict_))\n    :[2]\n<\/code><\/pre>\n"},{"AnswerId":"41808882","CreationDate":"2017-01-23T14:39:43.700","ParentId":null,"OwnerUserId":"3883940","Title":null,"Body":"<p>As far as I know, you can't really do that in one command like you would with a more advanced library like NumPy.\nIf you really want to use TF functions I could suggest a few like:<\/p>\n\n<pre><code>x = tf.Variable([\n    [1,2,3,1],\n    [0,0,0,0],\n    [1,3,5,7],\n    [0,0,0,0],\n    [3,5,7,8]])\n\ny = tf.Variable([0,0,0,0])\ncondition = tf.equal(x, y)\nindices = tf.where(condition)\n<\/code><\/pre>\n\n<p>This would result the following:<\/p>\n\n<pre><code>[[1 0]\n [1 1]\n [1 2]\n [1 3]\n [3 0]\n [3 1]\n [3 2]\n [3 3]]\n<\/code><\/pre>\n\n<p>Or you could use the following if you just want to get only the zero lines:<\/p>\n\n<pre><code>row_wise_sum = tf.reduce_sum(tf.abs(x),1)\nselect_zero_sum = tf.where(tf.equal(row_wise_sum,0))\n\nwith tf.Session() as sess:\n    tf.global_variables_initializer().run()\n    print(sess.run(select_zero_sum))\n<\/code><\/pre>\n\n<p>The result being: <\/p>\n\n<pre><code>[[1]\n [3]]\n<\/code><\/pre>\n"}]}
{"QuestionId":41807416,"AnswerCount":1,"Tags":"<neural-network><computer-vision><deep-learning><heatmap><caffe>","CreationDate":"2017-01-23T13:21:54.390","AcceptedAnswerId":41812285.0,"OwnerUserId":2614596.0,"Title":"How to extract memnet heat maps with the caffe model?","Body":"<p>I want to extract both memorability score and memorability heat maps by using the available memnet caffemodel by Khosla et al. at <a href=\"http:\/\/memorability.csail.mit.edu\/download.html\" rel=\"nofollow noreferrer\">link<\/a>\nLooking at the prototxt model, I can understand that the final inner-product output should be the memorability score, but how should I obtain the memorability map for a given input image? <a href=\"https:\/\/phys.org\/news\/2015-12-deep-learning-algorithm-photos-near-human.html\" rel=\"nofollow noreferrer\">Here<\/a> some examples.<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"41812285","CreationDate":"2017-01-23T17:36:06.320","ParentId":null,"OwnerUserId":"4221706","Title":null,"Body":"<p>As described in their paper [1], the CNN (MemNet) outputs a single, real-valued output for the memorability. So, the network they made <a href=\"http:\/\/memorability.csail.mit.edu\/download.html\" rel=\"nofollow noreferrer\">publicly available<\/a>, calculates this single memorability score, given an input image - and not a heatmap.<\/p>\n\n<p>In section 5 of the paper, they describe how to use this trained CNN to predict a memorability heatmap:<\/p>\n\n<blockquote>\n  <p>To generate memorability maps, we simply scale up the image and apply MemNet to overlapping regions of the image. We do this for multiple scales of the image and average the resulting memorability maps.<\/p>\n<\/blockquote>\n\n<p>Let's consider the two important steps here:<\/p>\n\n<h3>Problem 1: Make the CNN work with any input size.<\/h3>\n\n<p>To make the CNN work on images of any arbitrary size, they use the method presented in [2].\nWhile convolutional layers can be applied to images of arbitrary size - resulting in smaller or larger outputs - the inner product layers have a fixed input and output size.\nTo make an inner product layer work with any input size, you apply it just like a convolutional kernel. For an FC layer with 4096 outputs, you interpret it as a 1x1 convolution with 4096 feature maps.<\/p>\n\n<p>To do that in Caffe, you can directly follow the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow noreferrer\">Net Surgery tutorial<\/a>. You create a new <code>.prototxt<\/code> file, where you replace the <code>InnerProduct<\/code> layers with <code>Convolution<\/code> layers. Now, Caffe won't recognize the weights in the <code>.caffemodel<\/code> anymore, as the layer types don't match anymore. So, you load the old net and its parameters into Python, load the new net, and assign the old parameters to the new net and save it as a new <code>.caffemodel<\/code> file.<\/p>\n\n<p>Now, we can run images of any dimensions (larger or equal than 227x227) through the network.<\/p>\n\n<h3>Problem 2: Generate the heat map<\/h3>\n\n<p>As explained in the paper [1], you apply this fully-convolutional network from Problem 1 to the same image at different scales. The MemNet is a re-trained AlexNet, so the default input dimension is 227x227. They mention that a 451x451 input gives a 8x8 output, which implies a stride of 28 for applying the layers. So a simple example could be:<\/p>\n\n<ul>\n<li>Scale 1: 227x227 \u2192 1x1. (I guess they definitely use this scale.)<\/li>\n<li>Scale 2: 283x283 \u2192 2x2. (Wild guess)<\/li>\n<li>Scale 3: 339x339 \u2192 4x4. (Wild guess)<\/li>\n<li>Scale 4: 451x451 \u2192 8x8. (This scale is mentioned in the paper.)<\/li>\n<\/ul>\n\n<p>The results will look like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KqDeA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KqDeA.png\" alt=\"multi-res\"><\/a><\/p>\n\n<p>So, you'll just average these outputs to get your final 8x8 heatmap. From the image above, it should be clear how to average the different-scale outputs: you'll have to upsample the low-res ones to 8x8, and average then.<\/p>\n\n<p>From the paper, I assume that they use very high-res scales, so their heatmap will be around the same size as the image initially was. They write that it takes 1s on a \"normal\" GPU. This is a quite long time, which also indicates that they probably upsample the input images quite to quite high dimensions.<\/p>\n\n<h3>Bibliography:<\/h3>\n\n<p>[1]: A. Khosla, A. S. Raju, A. Torralba, and A. Oliva, \"Understanding and Predicting Image Memorability at a Large Scale\", in: ICCV, 2015. <a href=\"http:\/\/people.csail.mit.edu\/khosla\/papers\/iccv2015_khosla.pdf\" rel=\"nofollow noreferrer\">[PDF]<\/a><br>\n[2]: J. Long, E. Shelhamer, and T. Darrell, \"Fully convolutional networks for semantic segmentation\", in: CVPR, 2015. <a href=\"https:\/\/people.eecs.berkeley.edu\/~jonlong\/long_shelhamer_fcn.pdf\" rel=\"nofollow noreferrer\">[PDF]<\/a><\/p>\n"}]}
{"QuestionId":41808353,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-01-23T14:12:49.800","AcceptedAnswerId":41809153.0,"OwnerUserId":4482655.0,"Title":"Tensorflow Slim: 'module' object has no attribute 'sum_of_squares'","Body":"<p>I am following the tutorial at <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/slim_walkthough.ipynb\" rel=\"nofollow noreferrer\">TF Slim<\/a>. However at <\/p>\n\n<pre><\/pre>\n\n<p>I seem to be getting . I have installed TF version 0.12head running on Ubuntu 16.04, CPU version. Complete code I am running is follows:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41809153","CreationDate":"2017-01-23T14:51:43.170","ParentId":null,"OwnerUserId":"4482655","Title":null,"Body":"<p>Apparently for some reason, it has been removed in the latest build as can be seen at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/losses\/python\/losses\/loss_ops.py\" rel=\"nofollow noreferrer\">GitHub Repo<\/a>. I switched to <code>loss = slim.losses.mean_squared_error(predictions, targets)<\/code> which should serve the purpose I assume.<\/p>\n"}]}
{"QuestionId":41809130,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-01-23T14:50:39.363","AcceptedAnswerId":null,"OwnerUserId":5449615.0,"Title":"TensorFlow: slim training loop crashing \"No session factory registered\"","Body":"<p>I have a working TF installation and slim works fine, too.<\/p>\n\n<p>However, when I try to run a slim training loop, my app crashes.<\/p>\n\n<p>Minimal code:<\/p>\n\n<pre><\/pre>\n\n<p>When I run this, I get:<\/p>\n\n<pre><\/pre>\n\n<p>Conversely, the same model will train when  is called \"manually\":<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone have an idea where to start debugging?<\/p>\n\n<p>Thank you,\nPhilip<\/p>\n","answers":[{"AnswerId":"41810121","CreationDate":"2017-01-23T15:40:30.323","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It looks like this line is causing problems:<\/p>\n\n<pre><code>    master=FLAGS.master,\n<\/code><\/pre>\n\n<p>From the error message, it appears that Slim is trying to create a session as <code>sess = tf.Session(\"local\")<\/code>, which is not a valid session target. Try passing the flag <code>--master=\"\"<\/code> when you run your script, or explicitly set <code>master=\"\"<\/code> when you call <code>slim.learning.train()<\/code>.<\/p>\n"}]}
{"QuestionId":41809850,"AnswerCount":1,"Tags":"<python><arrays><multidimensional-array><tensorflow><concat>","CreationDate":"2017-01-23T15:27:34.497","AcceptedAnswerId":null,"OwnerUserId":7457101.0,"Title":"How to increase multi dimension of array in tensorflow?","Body":"<p>I have a txt file which has 8 columns and I am selecting 1 column for my feature extraction which gives me 13 features values, the shape of output array will be [1x13].\nSimilarly I have 5 txt files in a folder I want to run a loop so that the returned variable will have 5x13 data. <\/p>\n\n<pre><\/pre>\n\n<p>In the above function I am able to get [1x13] to the variable 'pool' and I am expecting the size of the variable poolfinal as [5x13] but i get it as [1x13].\nhow to concat in vertical direction ?\nWhat is the mistake i did in the loop ?<\/p>\n","answers":[{"AnswerId":"41816712","CreationDate":"2017-01-23T22:19:49.790","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>each loop creates pool and poolfinal from sctratch. That's why you see only one data in poolfinal.\ninstead please try following:<\/p>\n\n<pre><code>pools = []\nfor ...:\n  pools.append(...)\npoolfinal = tf.concat_v2(pools, axis=0)\n<\/code><\/pre>\n"}]}
{"QuestionId":41810521,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network><keras>","CreationDate":"2017-01-23T16:01:21.010","AcceptedAnswerId":null,"OwnerUserId":3233501.0,"Title":"How to use a tensorflow model extracted from a trained keras model","Body":"<p>I want to build and train a neural network using the keras framework. I configured keras that it will use Tensorflow as a backend. After I trained the model with keras I tried to use Tensorflow only. I can access the session and get the tensorflow graph. But I do not know how to use the tensorflow graph for example to make a prediction.<\/p>\n\n<p>I build a network with the following tutorial\n<a href=\"http:\/\/machinelearningmastery.com\/tutorial-first-neural-network-python-keras\/\" rel=\"noreferrer\">http:\/\/machinelearningmastery.com\/tutorial-first-neural-network-python-keras\/<\/a><\/p>\n\n<p>in the train() method i build and train a model using keras only and save the keras and tensorflow model<\/p>\n\n<p>in the eval() method <\/p>\n\n<p>Here is my Code:<\/p>\n\n<pre><\/pre>\n\n<p>I can access the tensorflow graph (sess.graph) which the keras framework built for me but I do not know how I can predict with the tensorflow graph. I know how I can build a tensorflow graph and predict with it in generell but not with the model keras build for me.<\/p>\n","answers":[{"AnswerId":"44393842","CreationDate":"2017-06-06T15:16:27.420","ParentId":null,"OwnerUserId":"1854664","Title":null,"Body":"<p>You need to get the input and output tensors from the Keras model definition and then the current TensorFlow session. Then you can evaluate it using TensorFlow only. Assuming <code>model<\/code> is your <code>loaded_model<\/code> and <code>x<\/code> is your training data.<\/p>\n\n<pre><code>sess = K.get_session()\ninput_tensor = model.input\noutput_tensor = model.output\n\noutput_tensor.eval(feed_dict={input_tensor: x}, session=sess)\n<\/code><\/pre>\n"}]}
{"QuestionId":41810692,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-01-23T16:09:57.450","AcceptedAnswerId":41813713.0,"OwnerUserId":2397315.0,"Title":"Custom Metric for ADR and FAR in Keras","Body":"<p>I need to define a custom metric in keras with a tensorflow  backend, to get ADR and FAR metrics, where:<\/p>\n\n<ul>\n<li><p>ADR is the ratio between the number of correctly detected elements of Class 1 and the total number of elements in Class 1.<\/p><\/li>\n<li><p>FAR is the ratio between the number of Class 0 elements that are incorrectly classified as Class 1 elements and the total number of Class 0 elements.<\/p><\/li>\n<\/ul>\n\n<p>For now, I have been able to compute the total number of elements on each class:<\/p>\n\n<pre><\/pre>\n\n<p>Now, for each index  in the tensor, I need to count how many:<\/p>\n\n<ul>\n<li> for ADR<\/li>\n<li> for FAR<\/li>\n<\/ul>\n\n<p>But I have no idea how to do this with tensorflow operations, the closest I got is this:<\/p>\n\n<pre><\/pre>\n\n<p>But it doesn't seem to return anything else than 0.<\/p>\n\n<p>Thank you for you help!<\/p>\n\n<hr>\n\n<p>Edit to include the complete code:<\/p>\n\n<pre><\/pre>\n\n<p>And the error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41813713","CreationDate":"2017-01-23T19:01:32.813","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>Do you just need to count?<\/p>\n\n<pre><code>y_true = tf.constant([1, 1, 1, 0, 0, 0, 1, 1, 1])\ny_pred = tf.constant([1, 1, 0, 1, 1, 0, 1, 0, 0])\n\nadr_idx = tf.equal(y_true, y_pred) &amp; tf.equal(y_true, 1)\nadr_idx = tf.reshape(tf.where(adr_idx), [-1]) # [0, 1, 6]\n\nfar_idx = tf.equal(y_true, 0) &amp; tf.equal(y_pred, 1) \nfar_idx = tf.reshape(tf.where(far_idx), [-1])  # [3, 4]\n\nnum_adr = tf.shape(adr_idx)[0]  # 3\nnum_far = tf.shape(far_idx)[0]  # 2\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    a, b = sess.run([num_adr, num_far])\n<\/code><\/pre>\n\n<p>use keras 1.2.1 with tensorflow version 0.11<br>\nI think the reason is different shape<br>\nThe shape of y_true is <code>(?, ?)<\/code> and shape of y_pred is <code>(?, 1)<\/code> ?   <\/p>\n\n<pre><code>def adr_and_far(y_true, y_pred):\n    total_adr = K.sum(K.tf.cast(K.tf.equal(y_true, 1), 'int32'))\n    total_far = K.sum(K.tf.cast(K.tf.equal(y_true, 0), 'int32'))\n\n    adr_idx = K.tf.equal(y_true, 1) &amp; K.tf.equal(y_pred, 1)\n    adr_idx = K.sum(K.tf.cast(K.reshape(adr_idx, [-1]), 'int32'))\n\n    far_idx = K.tf.equal(y_true, 0) &amp; K.tf.equal(y_pred, 1)\n    far_idx = K.sum(K.tf.cast(K.reshape(far_idx, [-1]), 'int32'))\n\n\n    return { 'total_adr': total_adr,\n             'adr': adr_idx \/ total_adr,\n             'total_far': total_far,\n             'far': far_idx \/ total_far\n           }\n<\/code><\/pre>\n"}]}
{"QuestionId":41811051,"AnswerCount":0,"Tags":"<python-2.7><neural-network><keras><conv-neural-network>","CreationDate":"2017-01-23T16:29:10.837","AcceptedAnswerId":null,"OwnerUserId":7400630.0,"Title":"Why am i getting poor results?","Body":"<p>I am currently working on a Neural network that maps stft audio files to MFCC features.. <\/p>\n\n<p>But keep for some reason keep getting a validation_loss around 200. \nWhich compared desired range is (-100,100)..   so the output the NN outputs is not desirable.. <\/p>\n\n<p>I tried different implementation.. <\/p>\n\n<p>Simple NN 1-layer, 2-layer, 3-layer => wiht no difference. <\/p>\n\n<p>CNN , 1-layer, 2-layer => same result<\/p>\n\n<p>tried with random forrest => never finished, memory allocation error. <\/p>\n\n<p>So... <\/p>\n\n<p>What am I doing wrong?\nI tried normalizing the input and output, but the same error came back when i unormalized the dataset.. so... <\/p>\n\n<p>the shape of the input is (x,2050) and the shape of the output (x,13)...<\/p>\n\n<p>Any idea on why I am getting so poor results?.. <\/p>\n\n<p>Normalization: <\/p>\n\n<pre><\/pre>\n\n<p>Currently network structure options tried: <\/p>\n\n<pre><\/pre>\n\n<p>Convolutional: <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41811925,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-23T17:15:09.583","AcceptedAnswerId":null,"OwnerUserId":6306884.0,"Title":"Handle Tensor in loop","Body":"<p>I have a feature map and some coordinate<br>\nI want to do something like below<br>\ncrop the data from feature map by the coordinate<br>\nThen do , (or ?)<br>\nHow can I do this in Tensorflow?<\/p>\n\n<pre><\/pre>\n\n<p>Four coordinate in x,y can constitute 9 bins\nI want to do max pool in each bin and each channel\nso that I can get pool result, shape is  in this case<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/21pny.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/21pny.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I think I can use  to do this<br>\nBut I don't know how to access external tensor and return the result<br>\nI did a small test, it seem doesn't access external tensor <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41819045","CreationDate":"2017-01-24T02:28:54.007","ParentId":null,"OwnerUserId":"2708478","Title":null,"Body":"<p>I think you can just use a regular for loop in this case (note how the lambda was able to access an external tensor, <code>feature_map<\/code> in this case):<\/p>\n\n<pre><code>batch_feature_map_max_fn = lambda xy : tf.reduce_max(feature_map[xy[0]:xy[1], xy[2]:xy[3], -1])\n\ncrop = []\nfor j in range(num_coor-1):\n    x1 = x_coor[:, j]\n    x2 = x_coor[:, j+1]\n    y1 = y_coor[:, j]\n    y2 = y_coor[:, j+1]\n\n    temp = tf.map_fn(fn, [x1,x2,y1,y2], dtype=tf.float32)\n\n    crop.append(temp)\n\ncrop = tf.pack(crop, axis=1)\n<\/code><\/pre>\n\n<p>As for the <code>tf.while_loop<\/code>, I believe you will be able to access all tensors that you are passing in the <code>loop_vars<\/code> argument. You can just return the extra tensors without any changes. I am not sure <code>tf.while_loop<\/code> is really suitable in this case, though.<\/p>\n"}]}
{"QuestionId":41811994,"AnswerCount":0,"Tags":"<python><neural-network><deep-learning><keras><lstm>","CreationDate":"2017-01-23T17:19:20.060","AcceptedAnswerId":null,"OwnerUserId":1467533.0,"Title":"Keras: Using a TimeDistributed or recurrent layer on inputs with different numbers of timesteps","Body":"<p>I'm trying to use TimeDistributed and LSTM in a way that apparently isn't supported; have any of you run into the same problem?  And if so, is there a way around this?<\/p>\n\n<p>Here's the problem:<\/p>\n\n<pre><\/pre>\n\n<p>Because the TimeDistributed layer was built when applied to the first input, it thinks it has a particular input shape, but this assumption fails when it gets applied to the second input, with a different number of timesteps, and it crashes.<\/p>\n\n<p>Where do you run into this problem?  I'm trying to re-implement a <a href=\"https:\/\/www.semanticscholar.org\/paper\/Bidirectional-Attention-Flow-for-Machine-Seo-Kembhavi\/7586b7cca1deba124af80609327395e613a20e9d\" rel=\"nofollow noreferrer\">reading comprehension model<\/a>, and it uses highway layers on top of word embeddings for both a question and a passage of text.  So my question tensor has shape , and my passage tensor has shape .  I just want a highway layer applied to the embeddings, and I want it to be the <em>same<\/em> highway layer for both the question and the passage.  The code above seems like a natural way to implement this (assuming  and  are actually previous layer outputs that are my word embeddings for the question and the passage).  However, it doesn't work.<\/p>\n\n<p>I can think of one work around, and that is to instantiate two separate TimeDistributed objects, both using the same underlying Highway layer.  Because TimeDistributed doesn't have any parameters, this actually works, but it's a bit ugly.  And it wouldn't be so bad if this only applied to TimeDistributed, but you get the same problem with any recurrent layer:<\/p>\n\n<pre><\/pre>\n\n<p>And I don't know a way around this problem.  Any ideas?<\/p>\n\n<p>The best (undocumented) solution that I've found to this is to replace the  and  in the input shapes above with , and that solves some of the problems, but then I need to compute a similarity matrix between the words in  and the words in , and I'm getting errors when trying to use  to construct this matrix, which would then have shape  (for ).<\/p>\n","answers":[]}
{"QuestionId":41813167,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-23T18:28:20.337","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Tensorflow: construct sparse matrix from its diagonals","Body":"<p>Could you tell me, is there an analog of scipy.sparse.diags in tensorflow?<\/p>\n","answers":[{"AnswerId":"41813658","CreationDate":"2017-01-23T18:57:46.047","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Here are all the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops\/matrix_math_functions\" rel=\"nofollow noreferrer\">tensorflow matrix operations<\/a>. There are some diagonal construction facilities, but there is not a 1:1 scipy.sparse.diags function. What you can do is to generate the matrix in scipy and then pass it to tensorflow.<\/p>\n"}]}
{"QuestionId":41813355,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-23T18:39:44.210","AcceptedAnswerId":41813519.0,"OwnerUserId":4790871.0,"Title":"Does tensorflow handle swapping GPU memory when a model exceeds memory capacity?","Body":"<p>I've noticed that a recent model warns that 2.37G of memory wasn't able to be allocated:<\/p>\n\n<pre><\/pre>\n\n<p>But my GPU is operating at nearly 100% (small input compared to a large model in this case).<\/p>\n\n<p>If I am reading this correctly, I assume that my model did not fit entirely in GPU memory. However since the GPU is running at 100% am I also to assume that tensorflow is intelligently swapping graph elements in and out of GPU memory asynchronously?<\/p>\n\n<p>I'm just curious to know what's going on under the hood there.<\/p>\n","answers":[{"AnswerId":"41813519","CreationDate":"2017-01-23T18:50:15.747","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>To know what is going on under the hood add this code to your run function:<\/p>\n\n<pre><code>run_metadata = tf.RunMetadata()\nsess = tf.Session(config=config) \nsess.run(train_step,\n           feed_dict={x: batch_xs,\n                      y_: batch_ys},\n            options=tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE),\n           run_metadata=run_metadata)\ntrace = timeline.Timeline(step_stats=run_metadata.step_stats)\nwith open('timeline.ctf.json', 'w') as trace_file:\n   trace_file.write(trace.generate_chrome_trace_format())\n<\/code><\/pre>\n\n<p>and then open the generated timeline.ctf.json from the <code>chrome:\/\/timeline<\/code> interface and you will see what is going on under the hood.<\/p>\n\n<p>It is very likely that is swapping GPU memory. <\/p>\n"}]}
{"QuestionId":41813665,"AnswerCount":4,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-01-23T18:58:11.273","AcceptedAnswerId":41929380.0,"OwnerUserId":4482655.0,"Title":"Tensorflow Slim: TypeError: Expected int32, got list containing Tensors of type '_Message' instead","Body":"<p>I am following <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/slim_walkthough.ipynb\" rel=\"noreferrer\">this<\/a> tutorial for learning TensorFlow Slim but upon running the following code for Inception:<\/p>\n\n<pre><\/pre>\n\n<p>I seem to be getting this set of errors:<\/p>\n\n<pre><\/pre>\n\n<p>This is strange because all of this code is from their official guide. I am new to TF and any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"43500699","CreationDate":"2017-04-19T16:00:42.497","ParentId":null,"OwnerUserId":"3954958","Title":null,"Body":"<p>explicitly writing the name of the arguments solves the problem.<\/p>\n\n<p>instead of <\/p>\n\n<pre><code>net = tf.concat(3, [branch_0, branch_1, branch_2, branch_3])\n<\/code><\/pre>\n\n<p>use<\/p>\n\n<pre><code>net = tf.concat(axis=3, values=[branch_0, branch_1, branch_2, branch_3])\n<\/code><\/pre>\n"},{"AnswerId":"51225494","CreationDate":"2018-07-07T17:34:29.977","ParentId":null,"OwnerUserId":"9982188","Title":null,"Body":"<p>I found most people answering wrong way. Its just due to the change in the tf.concat.\nIt works in the following way. <\/p>\n\n<p><code>net = tf.concat(3, [branch_0, branch_1, branch_2, branch_3])<\/code><\/p>\n\n<p>use the following<\/p>\n\n<p><code>net = tf.concat(values=[branch_0, branch_1, branch_2, branch_3],axis=3,)<\/code><\/p>\n\n<p>Remember while passing the keyword arguments should be before the others.<\/p>\n"},{"AnswerId":"41929380","CreationDate":"2017-01-30T05:14:03.997","ParentId":null,"OwnerUserId":"1844390","Title":null,"Body":"<p>I got the same problem when using the 1.0 released and I could make it work without having to roll back on a previous version.<\/p>\n\n<p>The problem is caused by change in the api. That discussion helped me to find the solution: <a href=\"https:\/\/groups.google.com\/a\/tensorflow.org\/forum\/#!msg\/discuss\/OePXmC9kJ7o\/SRErOoYCDQAJ\" rel=\"noreferrer\">Google group > \nRecent API Changes in TensorFlow<\/a> <\/p>\n\n<p>You just have to update all the line with tf.concat<\/p>\n\n<p>for example <\/p>\n\n<pre><code>net = tf.concat(3, [branch_0, branch_1, branch_2, branch_3])\n<\/code><\/pre>\n\n<p>should be changed to <\/p>\n\n<pre><code>net = tf.concat([branch_0, branch_1, branch_2, branch_3], 3)\n<\/code><\/pre>\n\n<p>Note:<\/p>\n\n<p>I was able to use the models without problem. But I still got error afterward when wanting to load the pretrained weight.\nSeems that the slim module got several changed since they made the checkpoint file. The graph created by the code and the one present in the checkpoint file were different.<\/p>\n\n<p>Note2:<\/p>\n\n<p>I was able to use the pretrain weights for inception_resnet_v2 by adding to all conv2d layer <code>biases_initializer=None<\/code><\/p>\n"},{"AnswerId":"42469228","CreationDate":"2017-02-26T13:42:49.763","ParentId":null,"OwnerUserId":"7625321","Title":null,"Body":"<p>I got same error when I did the work.<\/p>\n\n<p>I found that<\/p>\n\n<pre><code>logits = tf.nn.xw_plus_b(tf.concat(outputs, 0), w, b)\nloss = tf.reduce_mean(\n  tf.nn.softmax_cross_entropy_with_logits(\n    labels=tf.concat(train_labels, 0), logits=logits))\n<\/code><\/pre>\n\n<p>The output is <code>shape=(10, 64, 64)<\/code>.<\/p>\n\n<p>The code want concat outputs[0] to outputs[9] => get a new shape(640,64).<\/p>\n\n<p>But the \"tf.concat\" API may not allow to do this.<\/p>\n\n<p>(train_labels same to this)<\/p>\n\n<p>So I write to <\/p>\n\n<pre><code>A = tf.concat(0,[outputs[0],outputs[1]])\nA = tf.concat(0,[A,outputs[2]])\nA = tf.concat(0,[A,outputs[3]])\nA = tf.concat(0,[A,outputs[4]])\nA = tf.concat(0,[A,outputs[5]])\nA = tf.concat(0,[A,outputs[6]])\nA = tf.concat(0,[A,outputs[7]])\nA = tf.concat(0,[A,outputs[8]])\nA = tf.concat(0,[A,outputs[9]])\nB = tf.concat(0,[train_labels[0],train_labels[1]])\nB = tf.concat(0,[B,train_labels[2]])\nB = tf.concat(0,[B,train_labels[3]])\nB = tf.concat(0,[B,train_labels[4]])\nB = tf.concat(0,[B,train_labels[5]])\nB = tf.concat(0,[B,train_labels[6]])\nB = tf.concat(0,[B,train_labels[7]])\nB = tf.concat(0,[B,train_labels[8]])\nB = tf.concat(0,[B,train_labels[9]])\n\nlogits = tf.nn.xw_plus_b(tf.concat(0, A), w, b)\nloss = tf.reduce_mean(\n  tf.nn.softmax_cross_entropy_with_logits(\n    labels=tf.concat(0, B), logits=logits))\n<\/code><\/pre>\n\n<p>It can run!<\/p>\n"}]}
{"QuestionId":41813899,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-23T19:14:30.220","AcceptedAnswerId":null,"OwnerUserId":2081646.0,"Title":"Storing string-to-int dictionary with tensorflow","Body":"<p>I'm using Tensorflow to perform online learning and classification where the number of classes is not specified in advance i.e. the model has to support adding and removing classes in an online fashion. Furthermore, I plan to maintain a dictionary of class names (string) to corresponding indices in the output layer. This dictionary will be updated every time a class is added or deleted.<\/p>\n\n<p>Is there any way to store this dynamic dictionary with the tensorflow graph? I have looked at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/lookup\/lookup_ops.py\" rel=\"nofollow noreferrer\">lookup_ops in tensorflow<\/a>, but they do not support operations for adding or deleting classes.<\/p>\n","answers":[{"AnswerId":"43200181","CreationDate":"2017-04-04T06:49:58.327","ParentId":null,"OwnerUserId":"6490351","Title":null,"Body":"<p>How about storing names in a 1d variable of string type, and use a second 1d variable of int type to store corresponding indices? This would have O(n) lookup time; might suffice if the number of classes is small though.<\/p>\n"}]}
{"QuestionId":41814530,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-01-23T19:57:44.767","AcceptedAnswerId":null,"OwnerUserId":697065.0,"Title":"Having trouble introducing a constant matrix to multiplication in Tensorflow","Body":"<p>I am trying to implement a layer that is <strong>not<\/strong> fully connected. I have a matrix that specifies the connectivity I desire in the variable , which is a numpy array of ones and zeros. <\/p>\n\n<p>The way I am currently trying to impliment the layer is by pairwise multiplying the weights, by this connectivity matrix :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/IIZ5N.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IIZ5N.gif\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>Is this the correct way to do this in tensorflow? Here is what I have so far<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"46083215","CreationDate":"2017-09-06T19:44:52.337","ParentId":null,"OwnerUserId":"8262535","Title":null,"Body":"<p>Masking out unwanted connections in each iteration should work, but I am not sure what the convergence properties are like. It may okay for a small enough learning rate? <\/p>\n\n<p>Another approach would be to penalize unwanted weights in the cost function. You would use a mask matrix with 1's at unwanted connections, and 0's at wanted ones (or have a smoother transition). This would be multiplied by weights, squared\/scaled and added to the cost function. This should converge more smoothly. <\/p>\n\n<p>P.S.: If you've made progress on this, it would be great to hear your comments as I am also working on this problem. <\/p>\n"}]}
{"QuestionId":41815023,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2017-01-23T20:29:31.707","AcceptedAnswerId":null,"OwnerUserId":6246338.0,"Title":"ValueError: None values not supported Keras Custom Loss Function in Tensorflow","Body":"<p>I get the following error in the fit stage of a Keras sequential model using a custom objective function written using TensorFlow.<\/p>\n\n<pre><\/pre>\n\n<p>My custom function is this<\/p>\n\n<pre><\/pre>\n\n<p>I can run this when I have an interactive session. The function primarily relies on code from this Tensorflow issue <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/206<\/a> .<\/p>\n","answers":[{"AnswerId":"47423706","CreationDate":"2017-11-21T22:19:15.853","ParentId":null,"OwnerUserId":"225240","Title":null,"Body":"<p>Custom loss functions in Keras only build a graph when called (if using TensorFlow as the back end). The TensorFlow code is not actually run (the graph isn't executed), until the call to fit(), as you've noticed.<\/p>\n\n<p>So you can't use a typical debugger to step through the code and find the offending line, or look at the data.<\/p>\n\n<p>Some techniques for debugging:<\/p>\n\n<ul>\n<li>Print tensor values to the console<\/li>\n<\/ul>\n\n<p>Wrap the execution with console output using tf.Print, such as:<\/p>\n\n<pre><code>total = tf.Print(total, [total], 'total', summarize=10)\n<\/code><\/pre>\n\n<p>to print the first 10 values of the tensor <code>total<\/code>.<\/p>\n\n<ul>\n<li>Remove computational dependencies<\/li>\n<\/ul>\n\n<p>Note that the first value is the one used in the computation (assigned to the LHS of the equation). The 2nd argument is the one printed.<\/p>\n\n<p>So to test TF computations and see the output, pass them as the 2nd argument, so they do not affect the output during debugging.<\/p>\n\n<p>Eliminate variables from the value returned, to find the problematic computation. For example:<\/p>\n\n<pre><code>knownGoodValue = K.sum(tf.square(y_true - y_pred))  # any expr known to work\nprintExpr = tf.reshape(y_true, [-1])  # 1 x batch * numCells\nknownGoodValue = tf.Print(knownGoodValue, [printExpr], 'flatTruth', summarize=10)\n\nreturn knownGoodValue\n<\/code><\/pre>\n\n<p>This prints and returns some expression you know works, while allowing any TF expr to be tested\/printed by wrapping the test expression, without actaully using its result.<\/p>\n\n<ul>\n<li>Use TF Debugger<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/python\/debug\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/python\/debug<\/a><\/p>\n\n<p>The TensorFlow debugger may be invoked by adding the following code before your fit function:<\/p>\n\n<pre><code>import keras.backend as K\nfrom tensorflow.python import debug as tf_debug\nsess = K.get_session()\nsess = tf_debug.LocalCLIDebugWrapperSession(sess)\nK.set_session(sess)\n<\/code><\/pre>\n\n<p>It then presents a GDB-like cmdline interface when fit() is invoked.<\/p>\n"}]}
{"QuestionId":41815354,"AnswerCount":3,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-01-23T20:48:07.133","AcceptedAnswerId":41816481.0,"OwnerUserId":2954187.0,"Title":"keras flow_from_directory over or undersample a class","Body":"<p>I'm trying to do a binary classification problem with Keras, using the  method to generate batches. However, my classes are very imbalanced, like about 8x or 9x more in one class than the other, causing the model to get stuck predicting the same output class for every example. Is there a way to set  to either oversample from my small class or undersample from my large class during each epoch? For now, I've just created multiple copies of each image in my smaller class, but I'd like to have a bit more flexibility.<\/p>\n","answers":[{"AnswerId":"41816481","CreationDate":"2017-01-23T22:01:19.323","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>With current version of Keras - it's not possible to balance your dataset using only Keras built-in methods. The <code>flow_from_directory<\/code> is simply building a list of all files and their classes, shuffling it (if need) and then it's iterating over it. <\/p>\n\n<p>But you could do a different trick - by writting your own generator which would make the balancing inside the <code>python<\/code>:<\/p>\n\n<pre><code>def balanced_flow_from_directory(flow_from_directory, options):\n    for x, y in flow_from_directory:\n         yield custom_balance(x, y, options)\n<\/code><\/pre>\n\n<p>Here <code>custom_balance<\/code> should be a function that given a batch <code>(x, y)<\/code> is balancing it and returning a balanced batch <code>(x', y')<\/code>. For most of the applications the size of the batch doesn't need to be the same - but there are some weird use cases (like e.g. <a href=\"https:\/\/stackoverflow.com\/questions\/41759332\/calling-a-stateful-lstm-as-a-functional-model\/41792464#41792464\">stateful<\/a> RNNs) - where batch sizes should have a fixed size).<\/p>\n"},{"AnswerId":"48050531","CreationDate":"2018-01-01T14:35:51.173","ParentId":null,"OwnerUserId":"7136516","Title":null,"Body":"<p>You can also calculate the number of files in each class and normalize the class_weights <\/p>\n\n<pre><code>files_per_class = []\nfor folder in os.listdir(input_foldr):\n    if not os.path.isfile(folder):\n            files_per_class.append(len(os.listdir(input_foldr + '\/' + folder)))\ntotal_files = sum(files_per_class)\nclass_weights = {}\nfor i in xrange(len(files_per_class)):\n    class_weights[i] = 1 - (float(files_per_class[i]) \/ total_files)\nprint (class_weights)\n...\n...\n...\nmodel.fit_generator(... ,class_weight=class_weights)\n<\/code><\/pre>\n"},{"AnswerId":"51485928","CreationDate":"2018-07-23T19:26:09.893","ParentId":null,"OwnerUserId":"10123561","Title":null,"Body":"<p>One thing you can do is set <code>class_weight<\/code> parameter when calling <code>model.fit()<\/code> or <code>model.fit_generator()<\/code>. <\/p>\n\n<p>It also happens that you can easily compute your class_weights using <code>sklearn<\/code> and <code>numpy<\/code> libraries as follows:<\/p>\n\n<pre><code>from sklearn.utils import class_weight\nimport numpy as np\n\nclass_weights = class_weight.compute_class_weight(\n           'balanced',\n            np.unique(train_generator.classes), \n            train_generator.classes)\n<\/code><\/pre>\n\n<p>Afterwards, it becomes as simple as setting your <code>class_weights<\/code> equal to <code>class_weight<\/code> parameter: <\/p>\n\n<pre><code>model.fit_generator(..., class_weight=class_weights) \n<\/code><\/pre>\n"}]}
{"QuestionId":41816017,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-23T21:29:45.213","AcceptedAnswerId":41816431.0,"OwnerUserId":6156524.0,"Title":"How to print the execution time of the computation done in a tensorflow code?","Body":"<p>This is a generic question. <\/p>\n\n<p>I wrote a piece of code that uses tensorflow to do the computation.<\/p>\n\n<p>I want to print the time consumed during the execution of the code. <\/p>\n\n<p>First I used: <\/p>\n\n<pre><\/pre>\n\n<p>But I read that this is an inaccurate way to measure the execution time. <\/p>\n\n<p>How to accurately measure the execution time of my program. <\/p>\n","answers":[{"AnswerId":"41816431","CreationDate":"2017-01-23T21:57:56.697","ParentId":null,"OwnerUserId":"5754656","Title":null,"Body":"<p>Use <a href=\"https:\/\/docs.python.org\/3\/library\/time.html#time.perf_counter\" rel=\"noreferrer\"><code>time.<strong>perf_counter<\/strong>()<\/code><\/a>. <code>perf_counter<\/code> is a \"performance counter\". It is the highest resolution time from an undefined starting point (Usually since the program started running) available on a platform. It is used for things measuring performance when subtracted from a subsequent call. It is a float in seconds.<\/p>\n\n<p><a href=\"https:\/\/docs.python.org\/3\/library\/time.html#time.time\" rel=\"noreferrer\"><code>time.<strong>time<\/strong>()<\/code><\/a> is the seconds since the Unix epoch (1 January 1970), and may not be more precise than a second.<\/p>\n\n<p>You can safely replace calls to <code>time.time()<\/code> with <code>time.perf_counter()<\/code>, as you are subtracting them.<\/p>\n\n<pre><code>import time\nstart = time.perf_counter()\nmain()\nelapsed = time.perf_counter() - start\nprint('Elapsed %.3f seconds.' % elapsed)\n# The .3f is to round to 3 decimal places.\n<\/code><\/pre>\n"}]}
{"QuestionId":41816439,"AnswerCount":1,"Tags":"<deep-learning><keras><lstm>","CreationDate":"2017-01-23T21:58:24.930","AcceptedAnswerId":41816677.0,"OwnerUserId":7280706.0,"Title":"loss function in LSTM neural network","Body":"<p>I do not understand what is being minimized in these networks.\nCan someone please explain what is going on mathematically when the loss gets smaller in LSTM network?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41816677","CreationDate":"2017-01-23T22:16:45.270","ParentId":null,"OwnerUserId":"3902926","Title":null,"Body":"<p>From the <a href=\"https:\/\/keras.io\/objectives\/\" rel=\"noreferrer\" title=\"keras documentation\">keras documentation<\/a>, <code>categorical_crossentropy<\/code> is just the multiclass logloss. Math and theoretical explanation for log loss <a href=\"http:\/\/www.exegetic.biz\/blog\/2015\/12\/making-sense-logarithmic-loss\/\" rel=\"noreferrer\">here<\/a>. <\/p>\n\n<p>Basically, the LSTM is assigning labels to words (or characters, depending on your model), and optimizing the model by penalizing incorrect labels in word (or character) sequences. The model takes an input word or character vector, and tries to guess the next \"best\" word, based on training examples. Categorical crossentropy is a quantitative way of measuring how good the guess is. As the model iterates over the training set, it makes less mistakes in guessing the next best word (or character). <\/p>\n"}]}
{"QuestionId":41816912,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-23T22:36:13.390","AcceptedAnswerId":null,"OwnerUserId":2008598.0,"Title":"How can input operators be added to a tensorflow graph after distributed execution has begun?","Body":"<p>I would like to use distributed tensorflow to train on model in streaming fashion using a parameter server. The worker setup is something like this based on <a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/distributed\/<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>With this setup, I get a runtime error that the  when adding input operators in the second pass. How can I make this example work?<\/p>\n","answers":[{"AnswerId":"47370316","CreationDate":"2017-11-18T19:47:59.420","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You can try ._unsafe_unfinalize() on your graph object, although a more robust solution would be to create all needed operations ahead of time (there are performance\/thread-safety issues caused by modifying the graph while running it\/concurrently) <\/p>\n"}]}
{"QuestionId":41817101,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><deep-learning><keras><kaggle>","CreationDate":"2017-01-23T22:50:39.557","AcceptedAnswerId":41817669.0,"OwnerUserId":2414957.0,"Title":"line 596, in call_cpp_shape_fn raise ValueError(err.message) ValueError: Negative dimension size caused by subtracting 3 from 1","Body":"<p>I am trying to learn Keras by example and found this code on Kaggle. How can I make this work for Keras with TensorFlow backend?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41817669","CreationDate":"2017-01-23T23:42:57.980","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Your problem lies here:<\/p>\n\n<pre><code>input_shape=(1, img_rows, img_cols)\n<\/code><\/pre>\n\n<p>This input shape is in theano format, in Tensorflow the channels dimension goes at the end. You need to change the input shape to:<\/p>\n\n<pre><code>input_shape=(img_rows, img_cols, 1)\n<\/code><\/pre>\n\n<p>And also make sure that the input data (train and validation) are in the same format, with the channels dimension at the end of the shape tuple.<\/p>\n"}]}
{"QuestionId":41817519,"AnswerCount":0,"Tags":"<deep-learning><keras><kaggle><keras-layer><imagenet>","CreationDate":"2017-01-23T23:27:14.043","AcceptedAnswerId":null,"OwnerUserId":3715132.0,"Title":"Error when checking model target: expected predictions to have shape (None, 1000) but got array with shape (64, 2)","Body":"<p>Notebook so far: <a href=\"https:\/\/www.dropbox.com\/s\/h8gbliyt76bsurr\/vgg_own.ipynb?dl=0\" rel=\"nofollow noreferrer\">Notebook<\/a><\/p>\n\n<p>I am trying to reshape the standard   model to use for the classic cats vs dogs competition (<a href=\"https:\/\/www.kaggle.com\/c\/dogs-vs-cats-redux-kernels-edition\" rel=\"nofollow noreferrer\">Kaggle Cats vs Dogs<\/a>)<\/p>\n\n<p>I had to recreate the  and  functions of the  Keras model to remove the last  layer and replace it with  layer.<\/p>\n\n<p>However, when I try to use the  function, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>It sounds like my model is still expecting to output 1000 categories rather than 2. Why is this?<\/p>\n\n<p>Model summary below:<\/p>\n\n<pre><\/pre>\n\n<p>The  function sets the  variable to False, so I am wondering if it's anything to do with that. And if it is, how do I \"build\" the model? Any help is greatly appreciated.<\/p>\n","answers":[]}
{"QuestionId":41817828,"AnswerCount":1,"Tags":"<audio><deep-learning><theano><keras>","CreationDate":"2017-01-23T23:59:41.873","AcceptedAnswerId":41825393.0,"OwnerUserId":2280645.0,"Title":"Produce Audio with Convolutional neural networks","Body":"<p><strong>Were there any approaches on Producing Audio with Convolutional neural networks?<\/strong><\/p>\n\n<p>There are a lot of approaches on producing images through convnets. But i see no articles or posts about producing audio.<\/p>\n\n<p>According to this topic on <a href=\"https:\/\/stackoverflow.com\/questions\/22471072\/convolutional-neural-network-cnn-for-audio\">stackoverflow<\/a> , the post writer says :<\/p>\n\n<blockquote>\n  <p>\"I have found out the audio can be represented as spectrograms.\"<\/p>\n<\/blockquote>\n\n<p><strong>So why cant it be done?<\/strong> <\/p>\n\n<p><strong>To do this with Convnets, should I :<\/strong><\/p>\n\n<p><strong>a)<\/strong> Use LSTM with the conv layers?<\/p>\n\n<p><strong>B)<\/strong> What should be the output? Considering the spectogram...<\/p>\n","answers":[{"AnswerId":"41825393","CreationDate":"2017-01-24T10:11:43.240","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, of course there is, WaveNet by Deepmind produces audio that is very similar to \"human-like\" quality. It can also generate music.<\/p>\n\n<p><a href=\"https:\/\/deepmind.com\/blog\/wavenet-generative-model-raw-audio\/\" rel=\"nofollow noreferrer\">https:\/\/deepmind.com\/blog\/wavenet-generative-model-raw-audio\/<\/a><\/p>\n"}]}
{"QuestionId":41818043,"AnswerCount":1,"Tags":"<performance><plot><time><keras><conv-neural-network>","CreationDate":"2017-01-24T00:24:37.697","AcceptedAnswerId":41841049.0,"OwnerUserId":7412459.0,"Title":"Keras Plotting accuracy vs time","Body":"<p>I've been training a CNN in Keras and plotting the training and validation accuracy as a function of epochs. I was wondering if there is a way of plotting the accuracy as a function of processing time. <\/p>\n\n<p>Reason being that I want to demonstrate the speed of transfer learning as opposed to retraining a full network. When transfer learning is used, the network takes a similar number of epochs to train, give or take, but each epoch takes far less time (an order of magnitude faster) and I want to capture this graphically.<\/p>\n\n<p>Here is the code I've been using so far:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"41841049","CreationDate":"2017-01-24T23:52:12.567","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So it's actually really simple to implement a piece of code to perform your task in Keras. In order to do that - it's good to become familiar with a <a href=\"https:\/\/keras.io\/callbacks\/#lambdacallback\" rel=\"nofollow noreferrer\"><code>keras.callback<\/code><\/a>. It makes possible to call a custom functions on:<\/p>\n\n<p><code>on_epoch_begin<\/code>: called at the beginning of every epoch.<\/p>\n\n<p><code>on_epoch_end<\/code>: called at the end of every epoch,<\/p>\n\n<p><code>on_batch_begin<\/code>: called at the beginning of every batch,<\/p>\n\n<p><code>on_batch_end<\/code>: called at the end of every batch,<\/p>\n\n<p><code>on_train_begin<\/code>: called at the beginning of model training,<\/p>\n\n<p><code>on_train_end<\/code>: called at the end of model training.<\/p>\n\n<p>So now you might implement e.g. a new callback which will :<\/p>\n\n<ul>\n<li>register the start of training time <code>on_train_begin<\/code>,<\/li>\n<li>at the end of each epoch it will register the actual time of end of computations using <code>on_epoch_end<\/code> in a <code>dict<\/code> provided,<\/li>\n<li>register the end of training <code>on_train_end<\/code>.<\/li>\n<\/ul>\n\n<p>By using the data collected by this callback you could easily present the dependency between time and accuracy in a variety of ways. Of course - it could be easily extended to <code>batch\/iterations<\/code> time periods.<\/p>\n"}]}
{"QuestionId":41818288,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-01-24T00:54:20.257","AcceptedAnswerId":41818377.0,"OwnerUserId":2168245.0,"Title":"TensorFlow: Available C++ equivalents of Python functions","Body":"<p>I would like to know how I can find the C++ equivalents for the TensorFlow methods widely used in Python. For example, where can find the equivalent of ? Is it available at all?<\/p>\n\n<p>I was reading the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/label_image\/main.cc\" rel=\"nofollow noreferrer\">main.cc<\/a> in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/label_image\" rel=\"nofollow noreferrer\">label_image<\/a> example, and I noticed that it uses functions like  and , whose documentations I cannot find in the C++ API docs. I wonder where these functions come from.<\/p>\n\n<p>Thanks.<\/p>\n\n<hr>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>This confusion is very much solved with the good documentation of the new <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/\" rel=\"nofollow noreferrer\">TensorFlow r1.0 C++ API<\/a>.<\/p>\n","answers":[{"AnswerId":"41818377","CreationDate":"2017-01-24T01:05:27.723","ParentId":null,"OwnerUserId":"2050","Title":null,"Body":"<p><em>Update<\/em>: The answer is mostly meaningful for pre-1.0 versions of TensorFlow. MBA points to some new explanation provided by the TF team directly (official docs), and there was good refactoring around the ops :-)<\/p>\n\n<hr>\n\n<p>The Python API delegates execution to C++ implementation for many operations.<\/p>\n\n<p>The source code under <code>tensorflow\/core\/ops<\/code> contains kernels and ops. They are the actual operations that process commands invoked through the different APIs.<\/p>\n\n<p>Kernels and ops follow a pattern. The <code>REGISTER_OP<\/code> macro \"attaches\" a C++ function implementation to a handle (a string) that gets invoked at runtime by one of the APIs. Such kind of architecture has its merits and drawbacks---out of scope here :-)<\/p>\n\n<p>For example <code>tf.matmul<\/code>, the code can easily be found in <code>tensorflow\/core\/ops\/math_ops.cc<\/code> in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v0.12.0\/tensorflow\/core\/ops\/math_ops.cc\" rel=\"nofollow noreferrer\">repository<\/a>. The typical pattern looks like:<\/p>\n\n<pre><code>REGISTER_OP(\"MatMul\")\n    .Input(\"a: T\")\n    .Input(\"b: T\")\n    .Output(\"product: T\")\n    .Attr(\"transpose_a: bool = false\")\n    .Attr(\"transpose_b: bool = false\")\n    .Attr(\"T: {half, float, double, int32, complex64, complex128}\")\n    .SetShapeFn(shape_inference::MatMulShape)\n.Doc(\"[Edited: some doc string.]\");\n<\/code><\/pre>\n\n<p>Where the actual code run to execute <code>tf.matmul<\/code> is in the function <code>shape_inference::MatMulShape<\/code> (the \"functor\" namely, a C++ function). Wrappers (for example the Python API) call the <code>MatMul<\/code> handle to execute the code.<\/p>\n"}]}
{"QuestionId":41818387,"AnswerCount":3,"Tags":"<python><python-3.x><tensorflow><keras>","CreationDate":"2017-01-24T01:06:05.620","AcceptedAnswerId":null,"OwnerUserId":5574453.0,"Title":"Is it possible to export python and its necessary libraries into a environment independent file?","Body":"<p>I have a python script that uses keras and tensorflow libraries, which are incredibly time consuming to set up on every machines. Is it possible to export my python script and the keras and tensorflow libraries into a file like java projects --> .jar so that I don't need to set up the libraries every time I moved to a new machine?<\/p>\n","answers":[{"AnswerId":"41818620","CreationDate":"2017-01-24T01:34:41.353","ParentId":null,"OwnerUserId":"2344584","Title":null,"Body":"<p>Python 3 includes <a href=\"https:\/\/docs.python.org\/3\/library\/zipapp.html\" rel=\"nofollow noreferrer\">a tool called <code>zipapp<\/code><\/a> that allows you to build single archives out of Python projects, bundling in all (or some) dependencies. <strong>The tool is only bundled with Python 3, but the generated archive will work with any version of Python >= 2.6<\/strong> you are targeting.<\/p>\n\n<p>Supposing you have all your source files in <code>src\/<\/code>, including the mentioned Tensor Flow library:<\/p>\n\n<pre><code>$ python3 -m zipapp -o yourapp.pyz -m \"your.entry.point.module:main_function\" src\/\n<\/code><\/pre>\n\n<p>The <code>-m<\/code> flag allows you to specify a module and a function inside it, separed with a <code>:<\/code>, that will be executed when you run the <code>.pyz<\/code> file. About running it, it's just a matter of:<\/p>\n\n<pre><code>$ python .\/yourapp.pyz\n<\/code><\/pre>\n\n<p>Again, you will need Python >= 2.6 in order for this to work.<\/p>\n\n<p>If you are targeting unix platforms, you might also add a shebang:<\/p>\n\n<pre><code>$ echo \"#! \/usr\/bin\/python\" &gt; yourapp2.pyz\n$ cat yourapp.pyz &gt;&gt; yourapp2.pyz\n$ chmod +x yourapp2.pyz\n<\/code><\/pre>\n\n<p>so that you can run the file as:<\/p>\n\n<pre><code>$ .\/yourapp2.pyz\n<\/code><\/pre>\n\n<p>As a side note, a <code>.pyz<\/code> is nothing more than a <code>.zip<\/code> file. You might create one yourself without <code>zipapp<\/code>, just include a <code>__main__.py<\/code> file at top level. That will be the entry point of your bundle.<\/p>\n"},{"AnswerId":"41818413","CreationDate":"2017-01-24T01:09:21.293","ParentId":null,"OwnerUserId":"7442052","Title":null,"Body":"<p>Not file, but a folder. Use virtualenv (<a href=\"https:\/\/virtualenv.pypa.io\/en\/stable\/\" rel=\"nofollow noreferrer\">https:\/\/virtualenv.pypa.io\/en\/stable\/<\/a>). You can specify a python version and it will install it in the folder:<\/p>\n\n<pre><code>virtualenv -p \/usr\/bin\/python2.6 &lt;path\/to\/new\/virtualenv\/&gt;\n<\/code><\/pre>\n\n<p>You can then use pip to install packages and modules for that specific environment and then zip that folder up and move it around.<\/p>\n"},{"AnswerId":"42087740","CreationDate":"2017-02-07T10:43:29.847","ParentId":null,"OwnerUserId":"590335","Title":null,"Body":"<p>This is related to deploying python applications in general\nthis isn't trivial, especially for complex projects, but if the only issue is handling library installation the common solution is:<\/p>\n\n<p>include in you project a text file (typically requirements.txt) that includes the libraries and versions you need <\/p>\n\n<p>you can use <\/p>\n\n<pre><code>pip freeze &gt; requirements.txt\n<\/code><\/pre>\n\n<p>but you should probably remove libraries that were installed in the past but not actually used<\/p>\n\n<p>on the target machine run:<\/p>\n\n<pre><code>pip install -r requirements.txt\n<\/code><\/pre>\n\n<p>using a virtual env for setting up the environment is recommended, but using a virtual env from one machine and copying it as is to a different one isn't very safe or scalable (will only work with same os, same python version, and same directory structure)<\/p>\n"}]}
{"QuestionId":41818618,"AnswerCount":2,"Tags":"<macos><python-3.x><ipython><anaconda><pytorch>","CreationDate":"2017-01-24T01:34:19.060","AcceptedAnswerId":null,"OwnerUserId":1144431.0,"Title":"PyTorch doesn't import after installing Anaconda","Body":"<p>I just installed PyTorch after installing Anaconda, and when I run iPython using Anaconda Python, it won't find PyTorch.\n<a href=\"https:\/\/i.stack.imgur.com\/62L38.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/62L38.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/v6iy7.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/v6iy7.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However I can verify PyTorch is in the pkgs directory of my anaconda folder.\nWhat's going on?<\/p>\n","answers":[{"AnswerId":"43016636","CreationDate":"2017-03-25T12:58:42.380","ParentId":null,"OwnerUserId":"4385382","Title":null,"Body":"<p>While the framework name is Pytorch, it must be imported using the <em>torch<\/em> namespace like so <code>import torch<\/code>. If you look at the official documentation <a href=\"http:\/\/pytorch.org\/docs\/\" rel=\"nofollow noreferrer\">here<\/a>, you will see any pytorch related module import like nn or autograd is done through the <em>torch<\/em> namespace. I admit it is a bit confusing but I guess it was done to instill a sense of familiarity for old torch7 users. Also, there is a very good introduction tutorial which you could probably use to familiarize yourself better with the ins and outs of pytorch found <a href=\"http:\/\/pytorch.org\/tutorials\/beginner\/deep_learning_60min_blitz.html\" rel=\"nofollow noreferrer\">here<\/a>. Hope this helps!<\/p>\n"},{"AnswerId":"41818807","CreationDate":"2017-01-24T02:00:19.663","ParentId":null,"OwnerUserId":"2960511","Title":null,"Body":"<p>The Py Torch package might have installed to a different python environment that you might have on your system. <\/p>\n\n<p>You can see the entire list of python packages installed in your current environment using <\/p>\n\n<p><code>pip freeze<\/code><\/p>\n\n<p>Ensure that the <code>pip<\/code> command being executed belongs to Anaconda distribution. The following command will provide the path of the pip executable.<\/p>\n\n<pre><code>which pip\n<\/code><\/pre>\n"}]}
{"QuestionId":41818654,"AnswerCount":3,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-01-24T01:41:18.517","AcceptedAnswerId":null,"OwnerUserId":5431568.0,"Title":"Keras BatchNormalization uninitialized value","Body":"<p>I am trying to add batch norm to a vgg style model in Keras. When I add the batch norm layers I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>Without the batch layers the script runs without errors, only when I add the batchNormalization layers does it throw the error.<\/p>\n\n<pre><\/pre>\n\n<p>Is this the correct syntax to use batch norm in Keras? I followed the example in this <a href=\"https:\/\/stackoverflow.com\/questions\/34716454\/where-do-i-call-the-batchnormalization-function-in-keras\">thread<\/a>.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"55594926","CreationDate":"2019-04-09T14:14:10.133","ParentId":null,"OwnerUserId":"7218537","Title":null,"Body":"<p>If \nyour input image's data format is \"channels_last\" and the input_shape is Image_Height x Image_Width x Image_Channel\nthen \ntry using BatchNormalization(axis = 3)<\/p>\n"},{"AnswerId":"54817620","CreationDate":"2019-02-21T23:05:00.517","ParentId":null,"OwnerUserId":"4549682","Title":null,"Body":"<p>Try <code>keras.backend.get_session().run(tf.local_variables_initializer())<\/code>.  For me, the global initializer didn't work, but local did.  Although this is probably not an issue with the latest TF\/Keras versions.<\/p>\n"},{"AnswerId":"41893302","CreationDate":"2017-01-27T11:48:25.990","ParentId":null,"OwnerUserId":"7478440","Title":null,"Body":"<p>Try <code>keras.backend.get_session().run(tf.global_variables_initializer())<\/code> before fit. There is an issue <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/4623\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
{"QuestionId":41818906,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-24T02:11:44.017","AcceptedAnswerId":42680925.0,"OwnerUserId":7428227.0,"Title":"Does NNpack are implemented in tensorflow?","Body":"<p>Does NNpack are implemented in tensorflow? I find that it can be  to be leveraged by higher-level frameworks, such as Torch, Caffe, Tensorflow, Theano, and Mocha.jl, Howerver, I have not found any implementation in tensorflow.<\/p>\n","answers":[{"AnswerId":"42680925","CreationDate":"2017-03-08T20:05:55.807","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow does not currently use <a href=\"https:\/\/github.com\/Maratyszcza\/NNPACK\" rel=\"nofollow noreferrer\">NNPACK<\/a> to accelerate its math kernels; instead it uses a combination of <a href=\"http:\/\/eigen.tuxfamily.org\/index.php?title=Tensor_support\" rel=\"nofollow noreferrer\">Eigen::Tensor<\/a>, <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow noreferrer\">cuDNN<\/a>, <a href=\"http:\/\/docs.nvidia.com\/cuda\/cublas\/\" rel=\"nofollow noreferrer\">cuBLAS<\/a>, <a href=\"https:\/\/www.tensorflow.org\/performance\/xla\/\" rel=\"nofollow noreferrer\">XLA-generated<\/a>, and hand-written kernels. <\/p>\n\n<p>If you wanted to use NNPACK inside a TensorFlow op, you could do this by <a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op\" rel=\"nofollow noreferrer\">adding a custom op kernel<\/a> that used that library.<\/p>\n"}]}
{"QuestionId":41819049,"AnswerCount":1,"Tags":"<python><tensorflow><readfile>","CreationDate":"2017-01-24T02:29:38.847","AcceptedAnswerId":null,"OwnerUserId":7460588.0,"Title":"Unable to read properly a text file in tensorflow","Body":"<p>I have a txt file which has 2000 row and 10 columns, columns data separated by tab. To read file I simply  defined a function. <\/p>\n\n<pre><\/pre>\n\n<p>and when I am trying to read the file <\/p>\n\n<pre><\/pre>\n\n<p>but the output is empty <\/p>\n\n<p>some posts say we have to use sess.run to load, so, I tried with out using a function<\/p>\n\n<pre><\/pre>\n\n<p>I get out put as in single line(It is reading the tab and next line also)<\/p>\n\n<pre><\/pre>\n\n<p>where I expect to be in size of [2000x10], which is not loaded correctly.<\/p>\n\n<p>I need to load the file inside a function because I need to do some mathematical calculations on the column data. How to load file inside a function ? So that I will be able to call the function. <\/p>\n\n<p>I can do it using <\/p>\n\n<pre><\/pre>\n\n<p>I hope this isn't the right way to load in to tensor<\/p>\n","answers":[{"AnswerId":"53416656","CreationDate":"2018-11-21T16:35:34.200","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p><code>sess.run(tf.shape(x))<\/code> returns the shape of x, not the value of x. The content of a file is a scalar string, so its shape is <code>[]<\/code>, as expected. <code>sess.run(x)<\/code> OTOH should give you the string value.<\/p>\n"}]}
{"QuestionId":41819080,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-24T02:33:24.277","AcceptedAnswerId":null,"OwnerUserId":5703903.0,"Title":"How do I use Batch Normalization in a multi-gpu setting in TensorFlow?","Body":"<p>Referencing this post on <a href=\"https:\/\/stackoverflow.com\/questions\/33949786\/how-could-i-use-batch-normalization-in-tensorflow\">How could I use Batch Normalization in TensorFlow?<\/a>.<\/p>\n\n<p>I have a multi-gpu setup similar to the CIFAR10 example. When I insert tf.contrib.layers.batch_norm to my network definition, I get a NoneType object in average_gradients. Specifically, the  is the NoneType.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Some sample code on how to run Batch Normalization in a multi-gpu environment would help.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>Simply removing the \"batch_norm\" variables solves this bug. However, the pressing question here is that each Batch Normalization has a beta and gamma on each GPU, with their own moving averages. How are all these moving averages over all the GPUs resolved at inference?<\/p>\n","answers":[{"AnswerId":"51163741","CreationDate":"2018-07-03T22:51:02.577","ParentId":null,"OwnerUserId":"1416611","Title":null,"Body":"<p>Just use BN independently across GPUs, while using one of the tower means to update the moving mean.<\/p>\n\n<pre><code>with tf.device('..'):\n  x,y = iterator.get_next()\n\n  \/\/ NN with variables copied over to each of the GPUs\n  loss = tower_loss(..)\n\n  \/\/ use last tower statistics to update the moving mean\/variance \n  batchnorm_updates = tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope=scope)\n\napply_gradient_op = average_gradients(*grads)\nbatchnorm_updates_op = tf.group(*batchnorm_updates)\ntrain_op = tf.group(apply_gradient_op, batchnorm_updates_op)\n<\/code><\/pre>\n\n<p>As gleaned from multiple comments <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7439#issuecomment-279193381\" rel=\"nofollow noreferrer\">here<\/a>, this simple asynchronous approach works well in practice for most domains, with the exception of problems like semantic segmentation, action video recognition etc. where the batch size is extremely small and async BN isn't able to afford the speed boost that it normally does.<\/p>\n"}]}
{"QuestionId":41819397,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-24T03:12:10.350","AcceptedAnswerId":41825234.0,"OwnerUserId":3251207.0,"Title":"In Tensorflow, how can I rename a certain operation name?","Body":"<p>I think this question may sound weird. Let's say, there is a tensor .<\/p>\n\n<pre><\/pre>\n\n<p>Is there an efficient way to refer  with a different name, like ?\nI am thinking a dummy operation like<\/p>\n\n<pre><\/pre>\n\n<p>I am asking this because I am trying different network architectures, and regardless of architectures, I would like to access the output tensor with a consistent name like<\/p>\n\n<pre><\/pre>\n\n<p>For now, output tensors are depending on architectures like  or . How can I wrap the final op with a certain name?<\/p>\n","answers":[{"AnswerId":"41825234","CreationDate":"2017-01-24T10:03:39.683","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/a\/34235413\/2542516\">This<\/a> answer suggests that <code>tf.Graph<\/code> is append only, you cannot modify it.<\/p>\n\n<p>That being said, a slightly better way to rename a tensor is to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/identity\" rel=\"noreferrer\">tf.identity<\/a> like this:<\/p>\n\n<pre><code>tf.identity(a, name=\"out\")\n<\/code><\/pre>\n\n<p>EDIT: After figuring out answer to one of my own questions, I can see a slightly better way to do this:<\/p>\n\n<pre><code>graph_def.node[&lt;index of the op you want to change&gt;].name = \"out\"\n<\/code><\/pre>\n\n<p>the <code>index<\/code> is usually <code>-1<\/code> because final output tensors are in the end of the <code>graph_def<\/code><\/p>\n"}]}
{"QuestionId":41819968,"AnswerCount":0,"Tags":"<python><theano><cudnn>","CreationDate":"2017-01-24T04:21:47.600","AcceptedAnswerId":null,"OwnerUserId":7002574.0,"Title":"Using Theano with CUDA but no cudNN","Body":"<p>I need to use Theano with my GPU (Nvidia 820M). However, cudNN requires Compute Capability 3.0, at least. Mine is 2.1. Therefore, I do not want to (actually, cannot) use cudNN, but I still would like to use my GPU with CUDA and Theano.<\/p>\n\n<p>I am using Theano 0.9 (dev) and when i disable cudNN in .theanorc file, only cpu is used. Is there a way to use Theano with my GPU (CUDA), but no cudNN?<\/p>\n\n<p>Thanks in advance!<\/p>\n\n<p>This is what I get:<\/p>\n\n<pre><\/pre>\n\n<p>Edit:<\/p>\n\n<p>I just saw on this link (<a href=\"http:\/\/philipperemy.github.io\/configure_gpu_for_deep_learning\/\" rel=\"nofollow noreferrer\">http:\/\/philipperemy.github.io\/configure_gpu_for_deep_learning\/<\/a>) that the author is not using cudnn as well but gpu was used instead of cpu. What could I have possibly done?<\/p>\n\n<p>Edit 2:<\/p>\n\n<p>I have totally forgotten to paste the code that generates this output. My apologies for this. Here it is:<\/p>\n\n<pre><\/pre>\n\n<p>I have news. I noticed that even though it prints 'Used the cpu' when I set device=cuda in .theanorc, the execution time is almost 30x shorter than when using device=cpu. Then I performed another experience, this time with a CNN that processes some images and I noticed that when I enter 'nvidia-smi' command while the algorithm is running, the amount of gpu memory used increases from 330 MB to a bit more than 415 MB. When I close the process, it goes back to the initial 330 MB. It clearly seems that theano is using GPU but the test code above prints 'Used the cpu' for some reason.<\/p>\n\n<p>Another information. When I enter 'import theano' in python interpreter, it shows the following message:<\/p>\n\n<pre><\/pre>\n\n<p>I do not understand why is the test saying that it is using CPU while I have evidences that my GPU is the one being used.<\/p>\n\n<p>Finally, this is my .theanorc file:<\/p>\n\n<pre><\/pre>\n\n<p>I had to assign None to optimizer option, otherwise the program would not work, since the following exception is shown:<\/p>\n\n<pre><\/pre>\n\n<p>This is probably because my gpu does not support cuDNN. Isn't it?<\/p>\n","answers":[]}
{"QuestionId":41821058,"AnswerCount":1,"Tags":"<python><tensorflow><linear-regression>","CreationDate":"2017-01-24T06:03:22.077","AcceptedAnswerId":null,"OwnerUserId":1616677.0,"Title":"Does tensorflow support online training?","Body":"<p>I try to feed the data sample-by-sample. The result is either completely wrong or very approximate (25-50% absolute error) on different datasets. The result is fine for all datasets, if training in one go.<\/p>\n\n<pre><\/pre>\n\n<p>P.S.  and  are the same according to the source<\/p>\n","answers":[{"AnswerId":"41864574","CreationDate":"2017-01-26T01:26:01.870","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>This looks like learning_rate and\/or optimization. Please try with them as follows:<\/p>\n\n<p><code>model = tf.contrib.learn.LinearRegressor(..., optimizer=tf.train.YOUR_OPTIMIZER(YOUR_LEARNING_RATE)))<\/code><\/p>\n"}]}
{"QuestionId":41821502,"AnswerCount":1,"Tags":"<machine-learning><initialization><tensorflow><bias-neuron><zero-initialization>","CreationDate":"2017-01-24T06:34:08.277","AcceptedAnswerId":41864521.0,"OwnerUserId":1552048.0,"Title":"Zero initialiser for biases using get_variable in tensorflow","Body":"<p>A code I'm modifying is using  for weight variables, and  for bias initialisation. After some searching, it seems that  should always be favoured due to its portability in regards to sharing. So I tried to change the bias variable to  but can't seem to get it to work.<\/p>\n\n<p>Original: <\/p>\n\n<p>My attempt: <\/p>\n\n<p>I get an error saying that the shape should not be specified for constants. But removing the shape then throws an error for no arguments.<\/p>\n\n<p>I'm very new to  so I'm probably misunderstanding something fundamental here. Thanks for the help in advance :) <\/p>\n","answers":[{"AnswerId":"41864521","CreationDate":"2017-01-26T01:18:00.987","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>Following should work:\n<code>tf.get_variable(name=\"b1\", shape=[128], initializer=tf.zeros_initializer())<\/code><\/p>\n"}]}
{"QuestionId":41821975,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><theano><keras>","CreationDate":"2017-01-24T07:03:38.130","AcceptedAnswerId":41841173.0,"OwnerUserId":3052835.0,"Title":"What Keras's Resnet expects as an input?","Body":"<p>I'm trying to use Keras's implementation of resnet for a transfer learaning task with a quite different set of images (B&amp;W 16 bit). So what Keras expects as an input? Image with 3 channels and -127-128 range (that's what I assume zero centered 8 bit image)? 0-255? What would happen if I pass something outside this range?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"41841173","CreationDate":"2017-01-25T00:05:03.737","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>According to the <a href=\"https:\/\/arxiv.org\/pdf\/1512.03385v1.pdf\" rel=\"nofollow noreferrer\">paper<\/a> provided in Keras documentation you should provide a <code>224 x 224 RGB [0 - 225]<\/code> image. The actual dimension ordering depends on the backend you use in your Keras installation. <\/p>\n\n<p>The data preparation was performed as in <a href=\"https:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf\" rel=\"nofollow noreferrer\">AlexNet<\/a> so the mean activation was subtracted from each color channel. The mean vector for RGB is <code>103.939, 116.779, 123.68<\/code>.<\/p>\n\n<p>If your color values would extend <code>-255, 255<\/code> range - it could harm your training because of the magnitude of data unknown for the network. But still - network could adapt to this changes, but it usually makes more time and make training more chaotic.<\/p>\n\n<p>In case of monochromatic images - a commonly used technique is a repeating the same channel 3 times in order to make dimensions plausible for network architecture.<\/p>\n"}]}
{"QuestionId":41822308,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><linear-gradients>","CreationDate":"2017-01-24T07:22:37.143","AcceptedAnswerId":41823646.0,"OwnerUserId":3139545.0,"Title":"How tf.gradients work in TensorFlow","Body":"<p>Given I have a linear model as the following I would like to get the gradient vector with regards to W and b.<\/p>\n\n<pre><\/pre>\n\n<p>However if I try something like this where cost is a function of  and I only want to gradients with respect to :<\/p>\n\n<pre><\/pre>\n\n<p>My placeholders will also be included (X and Y). \nEven if I do get a gradient with  how do I know which element in the gradient that belong to each parameter since it is just a list without names to which parameter the derivative has be taken with regards to?<\/p>\n\n<p>In this question I'm using parts of this <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/2_BasicModels\/linear_regression.ipynb\" rel=\"noreferrer\">code<\/a> and I build on <a href=\"https:\/\/stackoverflow.com\/questions\/39167070\/implementing-gradient-descent-in-tensorflow-instead-of-using-the-one-provided-wi\">this<\/a> question. <\/p>\n","answers":[{"AnswerId":"41823646","CreationDate":"2017-01-24T08:47:01.903","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>Quoting the docs for <code>tf.gradients<\/code><\/p>\n\n<blockquote>\n  <p>Constructs symbolic partial derivatives of sum of ys w.r.t. x in xs.<\/p>\n<\/blockquote>\n\n<p>So, this should work:<\/p>\n\n<pre><code>dc_dw, dc_db = tf.gradients(cost, [W, b])\n<\/code><\/pre>\n\n<p>Here, <code>tf.gradients()<\/code> returns the gradient of <code>cost<\/code> wrt each tensor in the second argument as a list in the same order.<\/p>\n\n<p>Read <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gradients\" rel=\"noreferrer\">tf.gradients<\/a> for more information.<\/p>\n"}]}
{"QuestionId":41823068,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><keras><one-hot-encoding>","CreationDate":"2017-01-24T08:12:06.447","AcceptedAnswerId":41824511.0,"OwnerUserId":7085724.0,"Title":"Train multi-class image classifier in Keras","Body":"<p>I was following a tutorial to learn train a classifier using Keras <\/p>\n\n<p><a href=\"https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html\" rel=\"noreferrer\">https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html<\/a><\/p>\n\n<p>Specifically, from the <a href=\"https:\/\/gist.github.com\/fchollet\/f35fbc80e066a49d65f1688a7e99f069\" rel=\"noreferrer\">second script<\/a> given by the author, I wanted to transform the script into a one that can train multi-class classifier(was a binary for cat and dog). I have 5 classes in my train folder so I did the following change: <\/p>\n\n<p>In the function of <strong>train_top_model():<\/strong><\/p>\n\n<p>I changed<\/p>\n\n<pre><\/pre>\n\n<p>into<\/p>\n\n<pre><\/pre>\n\n<p>After done the training, the model reached training accuracy of near 99% but only for like 70% accuracy of the validation accuracy. Thus I started thinking maybe it's not that simple to convert 2 classes training to 5 classes. Maybe I need to use one-hot encoding when labeling the classes(but I don't know how)<\/p>\n\n<h1>EDIT:<\/h1>\n\n<p>I attached my fine-tunning script as well. Another problem: the accuracy did not effectively increase when fine-tunning starts.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41824511","CreationDate":"2017-01-24T09:28:33.893","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<ol>\n<li><p>Use <code>softmax<\/code> as activation function of the output layer, it is a generalization of the logistic function for a multi class case. Read more about it <a href=\"https:\/\/en.wikipedia.org\/wiki\/Softmax_function\" rel=\"noreferrer\">here<\/a>.<\/p><\/li>\n<li><p>If validation error is much greater than the training one, as in your case, it is an indicator of overfitting. You should do some regularization, which is defined as any changes of the learning algorithm, that are intended to reduce the test error but not the training one. You can try things like data augmentation, early stopping, noise injection, more aggressive dropout, etc.<\/p><\/li>\n<li><p>If you have the same set-up as in the linked tutorial, change the <code>class_mode<\/code>of the <code>train_generator<\/code>and <code>validation_generator<\/code> to <code>categorical<\/code> and it will one-hot encode your classes. <\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":41823555,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><keras><conv-neural-network>","CreationDate":"2017-01-24T08:41:31.893","AcceptedAnswerId":41828123.0,"OwnerUserId":147019.0,"Title":"Multi scale CNN Network Python Keras","Body":"<p>I create a multi-scale CNN in Python Keras. The network architecture is similar to the diagram. Here, same image is fed to 3 CNN's with different architectures. The weights are NOT shared.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/2H4xD.png\"><img src=\"https:\/\/i.stack.imgur.com\/2H4xD.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The code I wrote is available below. The issue is that when I run this even with 10 images in  the network takes about 40GB RAM and finally is killed by the OS. This is <strong>\"Out of memory ERROR\"<\/strong>. I am running this on CPU. Any idea why this happens in Keras?<\/p>\n\n<p>I am using Theano-0.9.0.dev5 | Keras-1.2.1 | Python 2.7.12 | OSX Sierra 10.12.3 (16D32) <\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9HZC7.png\"><img src=\"https:\/\/i.stack.imgur.com\/9HZC7.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41828123","CreationDate":"2017-01-24T12:19:43.487","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>The number of nodes in <code>lower_model1<\/code> and <code>lower_model2<\/code> after flattening is\n<code>32 * 112 * 112 = 401 408<\/code>. Followed by a fully connected layer with 64 nodes this gives <code>401 408 * 2 * 64 = 51 380 224<\/code> parameters, which is quite a big number. I would suggest to reconsider size of the images fed to your \"lower\" models. Do you really need <code>224 x 224<\/code> size there? Take a closer look at the diagram that you attached. There you see that the first step in the second and the third model is subsampling: <code>8:1<\/code> and <code>4:1<\/code>. This is the step that you have missed in your implementation.<\/p>\n\n<p>Your <code>main_model<\/code> is fine because you have enough max pooling layers there that reduce the number of parameters.  <\/p>\n"}]}
{"QuestionId":41823931,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-24T09:00:14.170","AcceptedAnswerId":47383437.0,"OwnerUserId":7461815.0,"Title":"splitting and joining parameters in Tensorflow","Body":"<p>I would like to create a simple MLP. I have a parameter set [W1, b1, W2, b2, W3, b3] where W is 2-d and b is 1-d. I want to concatenate the parameter set to one vector and I did this:<\/p>\n\n<pre><\/pre>\n\n<p>This seems to work. Now I want to update learnable_param_list with a single vector. I tried<\/p>\n\n<pre><\/pre>\n\n<p>This does not work. I get many errors like \"variable not iterable\" and zero size tensors. How can I split a vector and assign it to multiple tensors of different shapes? <\/p>\n","answers":[{"AnswerId":"47383437","CreationDate":"2017-11-20T00:10:51.907","ParentId":null,"OwnerUserId":"7461815","Title":null,"Body":"<p>I figure out the answer. The split_size can be replaced [reduce(mul, param.get_shape().as_list(), 1) for param in learnable_param_list]<\/p>\n"}]}
{"QuestionId":41825245,"AnswerCount":1,"Tags":"<python><numpy><neural-network><caffe><resnet>","CreationDate":"2017-01-24T10:04:23.223","AcceptedAnswerId":41825869.0,"OwnerUserId":7462150.0,"Title":"How to reshape the res5c layer of ResNet (3D to 2D)?","Body":"<p>I extract the features of an image with ResNet of the  layer, resulting of a numpy array of shape <\/p>\n\n<p>I have trouble manipulating these dimensions. I understand there is 14*14 features of size 2048. I would like to iterate over to access every feature at a time.<\/p>\n\n<p>Therefore, how I can reshape this to an array of (14*14, 2048) without mistakes and then easily iterate over it with a for loop?<\/p>\n","answers":[{"AnswerId":"41825869","CreationDate":"2017-01-24T10:33:14.690","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can read the features after <code>net.forward()<\/code>:<\/p>\n\n<pre><code>feat = net.blobs['res5c'].data.cop() # copy to be on the safe side.\n<\/code><\/pre>\n\n<p>As you describe, <code>feat<\/code> is an <code>np.array<\/code> with <code>shape = (2048, 14, 14)<\/code>.<br>\nYou can <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html\" rel=\"nofollow noreferrer\"><code>reshape<\/code><\/a> it:<\/p>\n\n<pre><code>feat.reshape((2048,-1)) # fix the first dimension to 2048, -1 set the number of features to match that of `feat`.\n<\/code><\/pre>\n\n<p>Now you can iterate over features:<\/p>\n\n<pre><code>for fi in xrange(feat.shape[1]):\n    f = feat[:,fi] # get the fi-th feature\n    # do somethinf to the feature f\n<\/code><\/pre>\n"}]}
{"QuestionId":41826531,"AnswerCount":1,"Tags":"<python><image-processing><tensorflow><python-3.5><keras>","CreationDate":"2017-01-24T11:03:17.130","AcceptedAnswerId":null,"OwnerUserId":7462399.0,"Title":"TensorFlow: AttributeError: 'Tensor' object has no attribute 'log10'","Body":"<p>I have been using tensorflow with python3 support. There is line in my code which throws me an error.<\/p>\n\n<p>The error here : <\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>AttributeError: 'Tensor' object has no attribute 'log10'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42169735","CreationDate":"2017-02-10T22:35:15.987","ParentId":null,"OwnerUserId":"7548241","Title":null,"Body":"<p>you can do like this: <\/p>\n\n<pre><code>return 10.0 * K.log(1.0 \/ (K.mean(K.square(y_pred - y_true)))) \/ K.log(10.0)\n<\/code><\/pre>\n"}]}
{"QuestionId":41826744,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-24T11:12:07.750","AcceptedAnswerId":null,"OwnerUserId":7461815.0,"Title":"Jacobian in Tensorflow","Body":"<p>I see many people asking this question here but I didn't see a code that I can execute. I am trying to make two operations, to get dOuput\/dInput and to get dOutput\/dParameters. I tried<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>here state is input and action is output. Both methods give None... What did I do wrong?<\/p>\n","answers":[]}
{"QuestionId":41826753,"AnswerCount":0,"Tags":"<python><tensorflow><lstm><encoder><decoder>","CreationDate":"2017-01-24T11:12:25.627","AcceptedAnswerId":null,"OwnerUserId":7462417.0,"Title":"Encoder and Decoder size in seq2seq.py tensorflow","Body":"<p>I have gone through the code for English-French translation which was in tensorflow. I have some questions. <\/p>\n\n<ol>\n<li><p>Is it necessary that the number of layers and neurons in both encoder and decoder must be same?<\/p><\/li>\n<li><p>What is the relation between embedding size of the words and the number of hidden layers? (In , embedding size is set equal to the number of hidden neurons). Can we change the size?<\/p><\/li>\n<li><p>Where exactly is the code for calculating ? (I have gone through the function in , but where exactly is the error calculated?)<\/p><\/li>\n<\/ol>\n","answers":[]}
{"QuestionId":41828064,"AnswerCount":4,"Tags":"<machine-learning><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-01-24T12:17:19.023","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"How can I plot the log file?","Body":"<p>Does anyone know any function for plotting the obtained measures in Caffe? I would like to plot train loss, test loss, and accuracy, train moving average and etc. in one plot. Is there any function except Caffe built-in function that is available online?<\/p>\n\n<p>Edited:\nFirst, I ran  file (the following command):<\/p>\n\n<pre><\/pre>\n\n<p>Two files are created based on the log file ( and ). After that,I ran  file. It has options like: <\/p>\n\n<pre><\/pre>\n\n<p>Whenever, I chose option <strong>3<\/strong>, it is showing the following graph:\n<a href=\"https:\/\/i.stack.imgur.com\/dipuk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dipuk.png\" alt=\"enter image description here\"><\/a>\nand by choosing option <strong>0<\/strong> :\n<a href=\"https:\/\/i.stack.imgur.com\/HVBs4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HVBs4.png\" alt=\"enter image description here\"><\/a>\nHowever, whenever, I want to plot train-loss figure, it is giving error: <\/p>\n\n<pre><\/pre>\n\n<p>My question can be folded into three parts:<\/p>\n\n<ol>\n<li>Are the plots correct? Is the network behaving well?<\/li>\n<li><p>From which point this error stem from? I have the following columns in  (#Iters|Seconds |TrainingLoss |LearningRate). <\/p><\/li>\n<li><p>How can I show all chart types in one plot? I tried to include them by comma, like 0,2,3,6, however, it is showing error.<\/p><\/li>\n<\/ol>\n\n<p>Many thanks in advance.<\/p>\n","answers":[{"AnswerId":"41828303","CreationDate":"2017-01-24T12:30:48.230","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Take a look at <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/tools\/extra\/parse_log.py\" rel=\"nofollow noreferrer\"><code>parse_log.py<\/code><\/a> found in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/tools\/extra\/\" rel=\"nofollow noreferrer\"><code>$CAFFE_ROOT\/tools\/extra<\/code><\/a>.<br>\nThis python utility helps parsing and distilling information from caffe running log. <\/p>\n"},{"AnswerId":"41853174","CreationDate":"2017-01-25T13:42:16.773","ParentId":null,"OwnerUserId":"6573448","Title":null,"Body":"<p>start training your model by executing the command below: <\/p>\n\n<pre><code>\/home\/ubuntu\/caffe\/build\/tools\/caffe train --solver \/home\/ubuntu\/yourpath\/solver.prototxt 2&gt;&amp;1 | tee \/home\/ubuntu\/yourpath\/model_train.log\n<\/code><\/pre>\n\n<p>The training logs will be stored under yourpath\/model_train.log.<\/p>\n"},{"AnswerId":"41910995","CreationDate":"2017-01-28T14:36:30.677","ParentId":null,"OwnerUserId":"4039874","Title":null,"Body":"<p>I haven't looked at <code>caffe<\/code>'s built in plot scripts, but I use the script from <a href=\"https:\/\/github.com\/yassersouri\/omgh\/blob\/master\/src\/scripts\/vis_finetune.py\" rel=\"nofollow noreferrer\">here<\/a>. This only plots your train\/test loss, but you can add moving average calculation.<\/p>\n"},{"AnswerId":"41924235","CreationDate":"2017-01-29T18:09:16.263","ParentId":null,"OwnerUserId":"3409864","Title":null,"Body":"<p>Consider also installing <a href=\"https:\/\/github.com\/NVIDIA\/DIGITS\" rel=\"nofollow noreferrer\">DIGITS<\/a>, that provides a real-time plot showing all that kind of stuff.<\/p>\n"}]}
{"QuestionId":41828263,"AnswerCount":2,"Tags":"<python><numpy><matplotlib><tensorflow>","CreationDate":"2017-01-24T12:28:36.950","AcceptedAnswerId":41829035.0,"OwnerUserId":1190882.0,"Title":"Image blurs after assignment","Body":"<p>I just started learning Tensorflow and Numpy concepts. I am reshaping differently shaped images using Tensorflow to one fixed shape for which I am making use of loop. At the end of the loop, I am accumulating this reshaped images into one array. Now, if I plot the image from this array, I get blurred image. But if I plot the instance of reshaped image using Tensorflow, I get the correct image. Please can anybody explain me where I am going wrong in this?<\/p>\n\n<p><strong>Code:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/p1Udk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/p1Udk.png\" alt=\"Correct image\"><\/a><\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/eKKeq.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/eKKeq.png\" alt=\"Blurred incorrect image\"><\/a><\/p>\n\n<p>Please can anybody explain me where I am going wrong in this and what concept am I missing here in my understanding of this assignment.<\/p>\n","answers":[{"AnswerId":"41833560","CreationDate":"2017-01-24T16:37:49.140","ParentId":null,"OwnerUserId":"764322","Title":null,"Body":"<p>If converting the data to <code>uint8<\/code> solves the problem, the following might solve it aswell (without having to declare the array as <code>uint8<\/code>):<\/p>\n\n<pre><code>plt.imshow(X_data[2], vmin=0, vmax=255)\n<\/code><\/pre>\n\n<p>From your problem, I do understand that your float image is actually in the range of <code>[0, 255]<\/code> (which is what the conversion to uint ends up doing). <\/p>\n\n<p>The real problem then seems to be that matplotlib's <code>imshow<\/code> automatically adjusts the <code>vmin<\/code> and <code>vmax<\/code> of the image to the <code>X_data[2].min()<\/code> and <code>X_data[2].max()<\/code> by default if they are not provided.<\/p>\n\n<p>If you provide <code>vmin=0<\/code> and <code>vmax=255<\/code> the plotting should be identical (and maybe more adequate) to rounding your data to unsigned bytes.<\/p>\n\n<p>However, it might not be the case, since the code cannot be replicated and thus, I cannot test it :P.<\/p>\n"},{"AnswerId":"41829035","CreationDate":"2017-01-24T13:06:58.730","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>I have got the solution. The problem was with X_data type. The function <code>imshow<\/code> takes values of type uint8 or float32 only and moreover the values have to be in the range of 0.0 to 1.0.<\/p>\n\n<p>My variable X_data was accepting types of float32 but the values were greater than 1. So converting the type of X_data to uint8 solved the problem.<\/p>\n\n<p>Here is the solution line:<\/p>\n\n<pre><code>X_data = np.zeros((3, fixedW, fixedH, 3), dtype = np.uint8)\n<\/code><\/pre>\n"}]}
{"QuestionId":41829865,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2017-01-24T13:46:32.263","AcceptedAnswerId":41843003.0,"OwnerUserId":1469241.0,"Title":"When to apply zero-centered data in an convolutional neural network","Body":"<p>I am building a cnn right now to classify grayscale images. The network did train until 97% accuracy on the test set until it got stuck. \nNow i discovered that i forgot to zero-center the data before feeding it into the network. While thinking about where to best apply it i came up with more than one answer and like to have some thoughts on them. <\/p>\n\n<p>First a small overview over my architecture: \nI am loading the datasets (train, validation, test) completely into the RAM. During runtime i load batches of images into placeholders which i then distort by randomly manipulating the brightness and contrast and then feed them into the network. <\/p>\n\n<p>Now, when i'd like to apply zero-centering, i could either take the mean of the training dataset and subtract it from every picture i feed into the network or i could subtract the mean of every single picture by using the tf.image.per_image_standardization() method from tensorflow. <\/p>\n\n<p>But if i calculate the mean of the training images, do i subtract it before or after i randomly distort the images? \nAnd also if i use the network later to label real images, do i need to save the mean value and subtract it also from every picture i want to classify?<\/p>\n\n<p>Same goes for the per_image_standardization, do i apply it before or after distorting the images and do i also apply it on the real world data i want to classify later?<\/p>\n","answers":[{"AnswerId":"41843003","CreationDate":"2017-01-25T03:39:23.563","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>Distortions should be applied before and normalisation\/standardisation. The idea of distorting the training set is to simulate having more data than you actually have, so you should do everything you would do to a test example on the distorted training data. This means saving the dataset statistics if you want to use dataset-wide normalisation later for test images.<\/p>\n\n<p>Whether you standardise based on each image individually or as an entire dataset depends on the data. If there are large variations in lighting between images that aren't important - say, a classification problem for vehicles with pictures taken under different illumination levels - I'd normalise per image. If you're looking to classify something dependent on illumination levels - e.g. daytime\/nighttime, indoor\/outdoor classifications - I'd do it as an entire dataset.<\/p>\n"}]}
{"QuestionId":41830206,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-01-24T14:03:43.487","AcceptedAnswerId":41834225.0,"OwnerUserId":6408452.0,"Title":"How to share a queue containing variable length sequences batches between multiple gpus?","Body":"<p>According to <a href=\"https:\/\/stackoverflow.com\/questions\/34273951\/tensorflow-multi-gpu-single-input-queue\">Tensorflow: Multi-GPU single input queue<\/a>, it might be better to have a single queue shared by multiple GPUs.\nAnd the link suggested that we could increase the batch size and then split batches by ourselves. However, when the input data are variable length sequences, increasing the batch size could result in many zero-padded values. <\/p>\n\n<p>For example, If we create a 4-sequence batch and split the batch, it could be<\/p>\n\n<p><\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<\/ul>\n\n<hr>\n\n<p><\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<\/ul>\n\n<p>My question is: How to produce batches like:<\/p>\n\n<p><\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<\/ul>\n\n<hr>\n\n<p><\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<\/ul>\n\n<p>Following <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/train_image_classifier.py\" rel=\"nofollow noreferrer\">slim<\/a>, I tried using a  to create batches, putting batches into a , and then calling  on different GPUs. However, it seems that all GPUs got the same data on the latest tensorflow (master).<\/p>\n\n<p>The following code demonstrate the issue:<\/p>\n\n<pre><\/pre>\n\n<p>On the latest tensorflow, the output is \n<\/p>\n\n<p>On the older tensorflow (0.11), the output is\n\n, which is what I want.<\/p>\n\n<p>It seems <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/train_image_classifier.py\" rel=\"nofollow noreferrer\">slim<\/a> also fetches the same data across all GPUs with the latest tensorflow.<\/p>\n\n<p>Is there any better way to share a queue containing variable length sequences between multiple GPUs?<\/p>\n","answers":[{"AnswerId":"41834225","CreationDate":"2017-01-24T17:09:32.927","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Try running with<\/p>\n\n<pre><code>config = tf.ConfigProto(graph_options=tf.GraphOptions(optimizer_options=tf.OptimizerOptions(opt_level=tf.OptimizerOptions.L0)))\n<\/code><\/pre>\n\n<p>This is a bit counter-intuitive:, filed <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7038\" rel=\"nofollow noreferrer\">7038<\/a><\/p>\n"}]}
{"QuestionId":41830723,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><batch-normalization>","CreationDate":"2017-01-24T14:28:44.437","AcceptedAnswerId":null,"OwnerUserId":1436907.0,"Title":"Batch normalization with 3D convolutions in TensorFlow","Body":"\n\n<p>I'm implementing a model relying on 3D convolutions (for a task that is similar to action recognition) and I want to use batch normalization (see <a href=\"https:\/\/arxiv.org\/abs\/1502.03167\" rel=\"noreferrer\">[Ioffe &amp; Szegedy 2015]<\/a>). I could not find any tutorial focusing on 3D convs, hence I'm making a short one here which I'd like to review with you.<\/p>\n\n<p>The code below refers to TensorFlow r0.12 and it explicitly instances variables - I mean I'm not using tf.contrib.learn except for the tf.contrib.layers.batch_norm() function. I'm doing this both to better understand how things work under the hood and to have more implementation freedom (e.g., variable summaries).<\/p>\n\n<p>I will get to the 3D convolution case smoothly by first writing the example for a fully-connected layer, then for a 2D convolution and finally for the 3D case. While going through the code, it would be great if you could check if everything is done correctly - the code runs, but I'm not 100% sure about the way I apply batch normalization. I end this post with a more detailed question.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<h2>Fully-connected (FC) case<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<h2>2D convolutional (CNN) layer case<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<h2>3D convolutional (CNN3D) layer case<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>What I would like to make sure is whether the code above exactly implements batch normalization as described in <a href=\"https:\/\/arxiv.org\/abs\/1502.03167\" rel=\"noreferrer\">[Ioffe &amp; Szegedy 2015]<\/a> at the end of Sec. 3.2:<\/p>\n\n<blockquote>\n  <p>For convolutional layers, we additionally want the normalization to obey the convolutional property \u2013 so that different elements of the same feature map, at different locations, are normalized in the same way. To achieve this, we jointly normalize all the activations in a minibatch, over all locations. [...] Alg. 2 is modified similarly, so that during inference the BN transform applies the same linear transformation to each activation in a given feature map.<\/p>\n<\/blockquote>\n\n<p><strong>UPDATE<\/strong>\nI guess the code above is also correct for the 3D conv case. In fact, when I define my model if I print all the trainable variables, I also see the expected numbers of beta and gamma variables. For instance:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This looks ok to me since due to BN, one pair of beta and gamma are learned for each feature map (256 in total).<\/p>\n\n<hr>\n\n<p>[Ioffe &amp; Szegedy 2015]: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift<\/p>\n","answers":[{"AnswerId":"46620505","CreationDate":"2017-10-07T13:07:41.883","ParentId":null,"OwnerUserId":"712995","Title":null,"Body":"<p>That is a great post about 3D batchnorm, it's often unnoticed that batchnorm can be applied to any tensor of rank greater than 1. Your code is correct, but I couldn't help but add a few important notes on this:<\/p>\n\n<ul>\n<li><p>A \"standard\" 2D batchnorm (accepts a 4D tensor) can be significantly faster in tensorflow than 3D or higher, because it supports <code>fused_batch_norm<\/code> implementation, which applies <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide#common_fused_ops\" rel=\"nofollow noreferrer\">one kernel operation<\/a>:<\/p>\n\n<blockquote>\n  <p>Fused batch norm combines the multiple operations needed to do batch\n  normalization into a single kernel. Batch norm is an expensive process\n  that for some models makes up a large percentage of the operation\n  time. Using fused batch norm can result in a 12%-30% speedup.<\/p>\n<\/blockquote>\n\n<p>There is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5694\" rel=\"nofollow noreferrer\">an issue on GitHub<\/a> to support 3D filters as well, but there hasn't been any recent activity and at this point the issue is closed unresolved.<\/p><\/li>\n<li><p>Although the original paper prescribes using batchnorm before ReLU activation (and that's what you did in the code above), there is evidence that it's probably better to use batchnorm <em>after<\/em> the activation. Here's a comment on <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1802#issuecomment-187966878\" rel=\"nofollow noreferrer\">Keras GitHub<\/a> by Francois Chollet:<\/p>\n\n<blockquote>\n  <p>... I can guarantee that recent code written by Christian [Szegedy] \n  applies relu\n  before BN. It is still occasionally a topic of debate, though.<\/p>\n<\/blockquote><\/li>\n<li><p>For anyone interested to apply the idea of normalization in practice, there's been recent research developments of this idea, namely <a href=\"https:\/\/arxiv.org\/abs\/1602.07868\" rel=\"nofollow noreferrer\">weight normalization<\/a> and <a href=\"https:\/\/arxiv.org\/abs\/1607.06450\" rel=\"nofollow noreferrer\">layer normalization<\/a>, which fix certain disadvantages of original batchnorm, for example they work better for LSTM and recurrent networks.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":41831214,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-24T14:51:16.063","AcceptedAnswerId":41877617.0,"OwnerUserId":6871950.0,"Title":"What is SYCL 1.2?","Body":"<p>I am trying to install tensorflow<\/p>\n\n<pre><\/pre>\n\n<p>What should I do?What is SYCL 1.2?<\/p>\n","answers":[{"AnswerId":"53908834","CreationDate":"2018-12-24T03:16:16.273","ParentId":null,"OwnerUserId":"7907965","Title":null,"Body":"<p>It is optional step so you can skip if you want.<\/p>\n\n<p>OpenCL (Open Computing Language) is a framework for writing programs that execute across heterogeneous platforms consisting of central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), field-programmable gate arrays (FPGAs) and other processors or hardware accelerators.<\/p>\n\n<p>so if you want to intall you have to Setting Up TensorFlow With OpenCL Using SYCL bellow link provide step by step information about it\n<a href=\"https:\/\/developer.codeplay.com\/computecppce\/latest\/getting-started-with-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/developer.codeplay.com\/computecppce\/latest\/getting-started-with-tensorflow<\/a><\/p>\n"},{"AnswerId":"41877617","CreationDate":"2017-01-26T15:58:14.293","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><a href=\"https:\/\/www.khronos.org\/sycl\" rel=\"noreferrer\">SYCL<\/a> is a C++ abstraction layer for OpenCL. TensorFlow's <a href=\"https:\/\/www.codeplay.com\/portal\/tensorflow%E2%84%A2-for-opencl%E2%84%A2-using-sycl%E2%84%A2\" rel=\"noreferrer\">experimental support<\/a> for OpenCL uses SYCL, in conjunction with a SYCL-aware C++ compiler.<\/p>\n\n<p>As Yaroslav pointed out in <a href=\"https:\/\/stackoverflow.com\/questions\/41831214\/what-is-sycl-1-2#comment70853775_41831214\">his comment<\/a>, SYCL is only required if you are building TensorFlow with OpenCL support. The following question during the execution of <code>.\/configure<\/code> asks about OpenCL support:<\/p>\n\n<pre><code>Do you wish to build TensorFlow with OpenCL support? [y\/N]\n<\/code><\/pre>\n\n<p>If you answer <code>N<\/code>, you will not have to supply a SYCL path.<\/p>\n"}]}
{"QuestionId":41832779,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-24T16:03:53.023","AcceptedAnswerId":52225065.0,"OwnerUserId":2911687.0,"Title":"tensorflow 2.5x slower than pytorch on vgg16 architecture","Body":"<p>So I'm trying to get into tensorflow and liking it so far.<\/p>\n\n<p>Today I upgraded to cuda 8, cudnn 5.1 and tensorflow 0.12.1. Using a Maxwell Titan X GPU.<\/p>\n\n<p>Using the following short code of loading the pretrained vgg16:<\/p>\n\n<pre><\/pre>\n\n<p>I then time the forward pass :<\/p>\n\n<pre><\/pre>\n\n<p>I get 160ms per batch (forward pass only), which seems 2.5x slower than the respective configuration in torch according to <a href=\"https:\/\/github.com\/jcjohnson\/cnn-benchmarks\" rel=\"nofollow noreferrer\" title=\"this benchmark\">this benchmark<\/a> (and also slower than <a href=\"http:\/\/www.vlfeat.org\/matconvnet\/pretrained\/\" rel=\"nofollow noreferrer\">MatconvNet<\/a>).<\/p>\n\n<p>The operations seem correctly assigned to the gpu, and the cuda libraries properly found, what else am I missing?<\/p>\n\n<p>Edit : cudnn and cuda properly found <\/p>\n\n<pre><\/pre>\n\n<p>Also the feeding does not seem to be the problem since replacing  by does not change the timing.<\/p>\n\n<p>Edit 2: So I compared to the pytorch version running on the same machine and I get (batches of 16x224x224x3) :<\/p>\n\n<ul>\n<li>Resnet-50 : <strong>pytorch 48ms<\/strong> vs <strong>tf 58 ms<\/strong> (OK)<\/li>\n<li>VGG16 : <strong>pytorch 65ms<\/strong> vs <strong>tf 160ms<\/strong> (not OK)<\/li>\n<\/ul>\n","answers":[{"AnswerId":"52225065","CreationDate":"2018-09-07T14:53:32.303","ParentId":null,"OwnerUserId":"2911687","Title":null,"Body":"<p>Tested recently on cuda 9.0, tensorflow 1.9 and pytorch 0.4.1, the differences are now negligible for the same operations.<\/p>\n\n<p>See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7065#issuecomment-419461999\" rel=\"nofollow noreferrer\">the proper timing here<\/a>.<\/p>\n"}]}
{"QuestionId":41835789,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-24T17:57:24.533","AcceptedAnswerId":41836684.0,"OwnerUserId":7459941.0,"Title":"tensorflow ValueError: Dimension 0 in both shapes must be equal","Body":"<p>I am currently studying TensorFlow. I am trying to create a NN which can accurately assess a prediction model and assign it a score. My plan right now is to combine scores from already existing programs run them through a mlp while comparing them to true values. I have played around with the MNIST data and I am trying to apply what I have learnt to my project. Unfortunately i have a problem<\/p>\n\n<pre><\/pre>\n\n<p>The code gives me this error<\/p>\n\n<pre><\/pre>\n\n<p>This error occurs when running the line for cross_entropy. I don't understand why this is happing, if you need any more information I would be happy to give it to you.<\/p>\n","answers":[{"AnswerId":"41836684","CreationDate":"2017-01-24T18:48:20.497","ParentId":null,"OwnerUserId":"7460616","Title":null,"Body":"<p>in your case, y has shape [9517, 1] while y_ has shape [9517]. they are not campatible. Please try to reshape y_ using tf.reshape(y_, [-1, 1])<\/p>\n"},{"AnswerId":"51500214","CreationDate":"2018-07-24T13:43:40.390","ParentId":null,"OwnerUserId":"10128322","Title":null,"Body":"<p>This was caused by the weights.hdf5 file being incompatible with the new data in the repository. I have updated the repo and it should work now.<\/p>\n"}]}
{"QuestionId":41835919,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-24T18:05:43.037","AcceptedAnswerId":null,"OwnerUserId":7241513.0,"Title":"TensorFlow: print output value from each iteration","Body":"<p>I'm experimenting with TensorFlow to get the hang of it. Trying to create a version of basic MNIST example by reading lines from a CSV file instead of images. I'm piecing the code together based on the MNIST example and the CSV feeder example. <\/p>\n\n<p>The first 100 cells in each line of my CSV represent an \"image\" of a number, and next 10 cells in the same row represent a \"label\" that identifies that image, similarly to the label in the MNIST example.<\/p>\n\n<p>I'm trying to print what TensorFlow actually estimates the number to be during each iteration over the CSV lines by printing the actual label and the \"guessed\" output side-by-side for each iteration. But it looks like what I'm actually printing is the true label both times, since the  variable requires me to feed the current  values to replace the x placeholder.<\/p>\n\n<p>How can I print out the guess, rather than the actual label for each iteration?<\/p>\n\n<p>My code looks like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41840855","CreationDate":"2017-01-24T23:30:41.597","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>this line returns you the index of the highest value inside the softmax y<\/p>\n\n<pre><code>print(sess.run(tf.argmax(y, 1), feed_dict={x: f1}))\n<\/code><\/pre>\n\n<p>if you want to print such value then you have to fetch the value at that index.<\/p>\n\n<pre><code>print(sess.run(y[tf.argmax(y, 1)], feed_dict={x: f1}))\n<\/code><\/pre>\n\n<p>if I understood correctly your queston. <\/p>\n"}]}
{"QuestionId":41835984,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-24T18:09:26.360","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Tensorflow multiple inputs to graph","Body":"<p>This might be trivial. I have a model which takes some dequeued tensor ;<\/p>\n\n<pre><\/pre>\n\n<p>Now I want to feed another tensor  at times;<\/p>\n\n<pre><\/pre>\n\n<p>How do I do this without redefining the subgraph?<\/p>\n","answers":[{"AnswerId":"41839413","CreationDate":"2017-01-24T21:36:34.627","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The feed mechanism in TensorFlow allows you to feed a value for <em>any<\/em> tensor (not just <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/placeholders#placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a> tensors), as long as they match in shape and element type.<\/p>\n\n<p>Therefore, if <code>x<\/code> and <code>z<\/code> have the same shape, you should be able to write:<\/p>\n\n<pre><code>x = ...  # Some dequeued `tf.Tensor`. \nyPred = model(x, ...)\n\n# ...\n\nsess.run(yPred, feed_dict={x: ...})\n<\/code><\/pre>\n\n<p>In some cases, you might want to feed a tensor with a different shape to <code>x<\/code> (usually a less specific shape, such as with a different batch size dimension). In those cases, you can use <code>tf.placeholder_with_default()<\/code> to create a placeholder whose value defaults to <code>x<\/code> when you <em>don't<\/em> feed it:<\/p>\n\n<pre><code>x = ...  # Some dequeued `tf.Tensor`.\n\n# For example, a shape of `None` means that the shape is completely unconstrained.\n# In practice, you will probably want to constrain at least the rank of the\n# placeholder to match the rank of `x`.\nx_placeholder = tf.placeholder_with_default(x, shape=None)\n\nyPred = model(x, ...)\n\n# ...\n\nsess.run(yPred, feed_dict={x_placeholder: ...})\n<\/code><\/pre>\n"}]}
{"QuestionId":41836015,"AnswerCount":1,"Tags":"<macos><tensorflow>","CreationDate":"2017-01-24T18:11:21.553","AcceptedAnswerId":null,"OwnerUserId":3304971.0,"Title":"Tensorflow Installation : Both python2 and python3 are installed on macOS.","Body":"<p>I have both python2 and python3 installed on my system(macos sierra). <\/p>\n\n<p>I tried installing latest version of tensrorflow using pip. It's working for python 3 but not for python2. <\/p>\n\n<p>I am getting following error \"tensorflow-0.12.1-py2-none-any.whl is not a supported wheel on this platform.\" <\/p>\n","answers":[{"AnswerId":"41864466","CreationDate":"2017-01-26T01:12:02.710","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>If you have pip version 8.1 or higher, <code>pip install tensorflow<\/code> command should work. Otherwise, you may install using this command:\n<code>pip install https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.1-py2-none-any.whl<\/code> If neither of these do not work, you may need to update your pip version as suggested by @RobR<\/p>\n"}]}
{"QuestionId":41836602,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-01-24T18:43:57.137","AcceptedAnswerId":41841375.0,"OwnerUserId":3822261.0,"Title":"Why does shuffling my validation set in Keras change my model's performance?","Body":"<p><strong>Why I'm confused:<\/strong><\/p>\n\n<p>If I test my model on examples [A, B, C], it will obtain a certain accuracy. If I test the same model on examples [C, B, A], it should obtain the same accuracy. In other words, shuffling the examples shouldn't change my model's accuracy. But that's what seems to be happening below:<\/p>\n\n<p><strong>Step-by-step:<\/strong><\/p>\n\n<p>Here is where I train the model:<\/p>\n\n<pre><\/pre>\n\n<p>Here is where I test the model, without shuffling the validation set:<\/p>\n\n<pre><\/pre>\n\n<p>Here are the results (val_loss, val_acc):<\/p>\n\n<pre><\/pre>\n\n<p>Notice that the validation accuracies are the same.<\/p>\n\n<p>Here is where I test the model, with a <strong>shuffled<\/strong> validation set:<\/p>\n\n<pre><\/pre>\n\n<p>Here are the results (val_loss, val_acc):<\/p>\n\n<pre><\/pre>\n\n<p>Notice that <strong>the validation accuracies are inconsistent<\/strong>, despite an unchanged validation set and an unchanged model. What's going on?<\/p>\n\n<hr>\n\n<p><strong>Note<\/strong>:<\/p>\n\n<p>I'm evaluating on the entire validation set each time. <a href=\"https:\/\/keras.io\/models\/sequential\/#evaluate_generator\" rel=\"nofollow noreferrer\">model.evaluate_generator<\/a> returns after evaluating the model on the number of examples equal to , which is the number of examples in the validation set.<\/p>\n","answers":[{"AnswerId":"41841375","CreationDate":"2017-01-25T00:24:06.737","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>This is a really interesting problem. The answer is that it's because of that neural networks are using a <code>float32<\/code> format which is not so accurate as <code>float64<\/code> - the fluctuation like this are simply the realisation of an underflow phenomenon.<\/p>\n\n<p>It case of your loss - you may notice that the differences are occuring after 7th decimal digit of a fractional part - what is exactly the precision of a <code>float32<\/code> format. So - basically - you may assume that all numbers presented in your example are equal in terms of a <code>float32<\/code> representation.<\/p>\n"}]}
{"QuestionId":41836614,"AnswerCount":1,"Tags":"<python><random><tensorflow>","CreationDate":"2017-01-24T18:44:28.353","AcceptedAnswerId":41838233.0,"OwnerUserId":7460588.0,"Title":"how to generate random shuffled number in tensorflow","Body":"<p>I want to generate numbers from 0 to a specific number and then randomly shuffle those numbers.\nI get some value from a tensor . For example if the value of total is 20 I want to generate numbers from 0 to 19. <\/p>\n\n<p>I tried to generate using \nbut it . <\/p>\n\n<p>I need to generate equal numbers and then shuffle with the value that comes as 'total' variable.\nI expect output randomly something like [8,14,1,3,5,11,6,9,10,19,12,13,2,15,7,16,17,4,18]<\/p>\n","answers":[{"AnswerId":"41838233","CreationDate":"2017-01-24T20:18:27.740","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>use <code>tf.range<\/code> to create list, then <code>tf.random_shuffle<\/code><\/p>\n\n<pre><code>idx = tf.constant(20)\nidx_list = tf.range(idx) # 0~19\nshuffle = tf.random_shuffle(idx_list)\n\nwith tf.Session() as sess:\n    a, b = sess.run([idx_list, shuffle])\n<\/code><\/pre>\n"}]}
{"QuestionId":41837767,"AnswerCount":1,"Tags":"<python><machine-learning><scikit-learn><keras><data-science>","CreationDate":"2017-01-24T19:53:26.890","AcceptedAnswerId":42873918.0,"OwnerUserId":5179470.0,"Title":"scikit-learn - Convert pipeline prediction to original value\/scale","Body":"<p>I've create a pipeline as follows (using the <a href=\"https:\/\/keras.io\/scikit-learn-api\/\" rel=\"noreferrer\">Keras Scikit-Learn API<\/a>)<\/p>\n\n<pre><\/pre>\n\n<p>and fit it with<\/p>\n\n<pre><\/pre>\n\n<p>If I predict with , I (believe) I get standardised predictions.<\/p>\n\n<p><strong>How do I predict on  so that  it at the same scale as the actual (untouched)  (i.e. NOT standardised prediction, but instead the actual values)?<\/strong> I see there is an <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.pipeline.Pipeline.html#sklearn.pipeline.Pipeline.inverse_transform\" rel=\"noreferrer\"> method for Pipeline<\/a>, however appears to be for only reverting a transformed .<\/p>\n","answers":[{"AnswerId":"42873918","CreationDate":"2017-03-18T11:40:48.590","ParentId":null,"OwnerUserId":"3065498","Title":null,"Body":"<p>Exactly. The StandardScaler() in a pipeline is only mapping the inputs (trainX) of pipeline.fit(trainX,trainY).<\/p>\n\n<p>So, if you fit your model to approximate trainY and you need it to be standardized as well, you should map your trainY as<\/p>\n\n<pre><code>scalerY = StandardScaler().fit(trainY)  # fit y scaler\npipeline.fit(trainX, scalerY.transform(trainY))  # fit your pipeline to scaled Y\ntestY = scalerY.inverse_transform(pipeline.predict(testX))  # predict and rescale\n<\/code><\/pre>\n\n<p>The inverse_transform() function maps its values considering the standard deviation and mean calculated in StandardScaler().fit().<\/p>\n\n<p>You can always fit your model without scaling Y, as you mentioned, but this can be dangerous depending on your data since it can lead your model to overfit. You have to test it ;)<\/p>\n"}]}
{"QuestionId":41838466,"AnswerCount":1,"Tags":"<python><deep-learning><theano>","CreationDate":"2017-01-24T20:32:17.717","AcceptedAnswerId":null,"OwnerUserId":3109123.0,"Title":"How can I change subsets value in theano tensor impeded into loop?","Body":"<p>layer_Fmaps of size (1,96,236,236)\nsitwches of size (1,96,708,708) <\/p>\n\n<p>in each 3 by 3 matrix in layer_Fmaps there is only cell having a value 1 which should be replace by the opposite value of sitwches<\/p>\n\n<p>I can't find a way to solve the problem by assigning a direct value into a certain location using a loop <\/p>\n\n<pre><\/pre>\n\n<p>knowing that switchs and layer_Fmaps are tensor4<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestion would be appreciated.<\/p>\n","answers":[{"AnswerId":"41887294","CreationDate":"2017-01-27T05:10:28.153","ParentId":null,"OwnerUserId":"3737792","Title":null,"Body":"<pre><code>def switchs(layer_Fmaps, switches, step=2):\n     for idx in range(96):     \n         for i in range(0, 708, step):\n             for j in range(0, 708, step):\n                 val = layer_Fmaps[0,idx,i\/2,j\/2]\n                 switches = T.set_subtensor(switches[0, idx, i:i + step, j:j + step],val)\n     return  switches\n<\/code><\/pre>\n\n<p>And:<\/p>\n\n<pre><code>img = np.zeros((1,96,236,236))\nsswitchs =  np.zeros((1,96,708,708))\n\ninp = T.tensor4('img')\nSW = T.tensor4('SW')\n\ntester = switchs(inp,SW, 3)\n\nf = theano.function([inp, SW], tester)    \n\nd = f(img,sswitchs)\n<\/code><\/pre>\n\n<p>I modified several things:<\/p>\n\n<ul>\n<li>Replaced 96 for 69. I guess the \"96\" was a typo.<\/li>\n<li>Uniform indexing<\/li>\n<li>Key arguments should be at the end.<\/li>\n<\/ul>\n"}]}
{"QuestionId":41838924,"AnswerCount":1,"Tags":"<python><tensorflow><mnist>","CreationDate":"2017-01-24T21:00:45.153","AcceptedAnswerId":41841167.0,"OwnerUserId":6201311.0,"Title":"Data changing between calls in Tensorflow","Body":"<p>I made a little change to Tensorflow MNIST tutorial. \nOriginal code (fully_connected_feed.py, lines 194-202):<\/p>\n\n<pre><\/pre>\n\n<p>I simply added one more evaluation:<\/p>\n\n<pre><\/pre>\n\n<p>Results of this evaluations are close, <strong>but not same<\/strong> (numbers vary from launch to launch):<\/p>\n\n<pre><\/pre>\n\n<p>How does it possible?\nUPD: added link to tensorflow github:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/tutorials\/mnist\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/tutorials\/mnist<\/a><\/p>\n","answers":[{"AnswerId":"41841167","CreationDate":"2017-01-25T00:03:33.830","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>do_eval()<\/code> function in fact does have a side effect, because <code>data_sets.train<\/code> is a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/624861052f70f6b07c1f714352f0bf7092ea237f\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L103\" rel=\"nofollow noreferrer\">stateful <code>DataSet<\/code> object<\/a> that contains a current <code>_index_in_epoch<\/code> member, which is advanced on each call to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/624861052f70f6b07c1f714352f0bf7092ea237f\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160\" rel=\"nofollow noreferrer\"><code>DataSet.next_batch()<\/code><\/a> (i.e. in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/624861052f70f6b07c1f714352f0bf7092ea237f\/tensorflow\/examples\/tutorials\/mnist\/fully_connected_feed.py#L78\" rel=\"nofollow noreferrer\"><code>fill_feed_dict()<\/code><\/a>).<\/p>\n\n<p>On its own, this fact shouldn't be enough to give non-deterministic results, but there are two other details about <code>DataSet.next_batch()<\/code> that lead to the non-determinism:<\/p>\n\n<ol>\n<li><p>Every time a new epoch is started, the examples are <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/624861052f70f6b07c1f714352f0bf7092ea237f\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L178\" rel=\"nofollow noreferrer\">randomly shuffled<\/a>.<\/p><\/li>\n<li><p>When the data set <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/624861052f70f6b07c1f714352f0bf7092ea237f\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L173\" rel=\"nofollow noreferrer\">reaches the end of an epoch<\/a>, the data set resets to the start and the last <code>num_examples % batch_size<\/code> examples are discarded. Thanks to the random shuffling, a random sub-batch of examples is discarded each time, leading to the non-deterministic results.<\/p><\/li>\n<\/ol>\n\n<p>Given the way the code is structured (with the <code>DataSet<\/code> shared between training and testing), it's tricky to make the code deterministic.\nThe <code>DataSet<\/code> class is sparsely documented, but this behavior is surprising, so I'd consider <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">filing a GitHub issue<\/a> about this problem.<\/p>\n"}]}
{"QuestionId":41839530,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-01-24T21:44:23.617","AcceptedAnswerId":41840977.0,"OwnerUserId":7387749.0,"Title":"Keras, IndexError: indices are out-of-bounds","Body":"<p>I'm trying to implement this simple neural network by Keras (Tensorflow beckend):<\/p>\n\n<pre><\/pre>\n\n<p>But when I run this part:<\/p>\n\n<pre><\/pre>\n\n<p>I get this error: <em>IndexError: indices are out-of-bounds<\/em>. I am supposing that it is about the arguments in <em>model.fit(x_train, y_train)<\/em>. I have tried to pass these as numpy arrays by <em>.values<\/em>, but I still have the same error.<\/p>\n","answers":[{"AnswerId":"41840977","CreationDate":"2017-01-24T23:44:04.767","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Keras expects numpy arrays not pandas, so you need to convert all of the data that you are feeding into Keras APIs.. not just <code>y_train<\/code> and <code>y_test<\/code><\/p>\n\n<p>So:<\/p>\n\n<pre><code>x_train = x_train.values\ny_train = y_train.values\nx_test = x_test.values\ny_test = y_test.values\n<\/code><\/pre>\n\n<p>Or<\/p>\n\n<pre><code>x_train = numpy.asarray(x_train)\ny_train = numpy.asarray(y_train)\nx_test = numpy.asarray(x_test)\ny_test = numpy.asarray(y_test)\n<\/code><\/pre>\n"}]}
{"QuestionId":41839648,"AnswerCount":1,"Tags":"<python><python-3.x><cuda><tensorflow><cudnn>","CreationDate":"2017-01-24T21:53:52.837","AcceptedAnswerId":null,"OwnerUserId":7466099.0,"Title":"Tensorflow GPU installation Ubuntu 16.04 Error (libcuda.so not found)","Body":"<p>I have installed the gpu version of tensorflow but when I try to import it I get the following error message: <\/p>\n\n<pre><\/pre>\n\n<p>I have followed the installation guide from the tensorflow website, and have made sure that I have installed Cuda and CudNN correctly. <\/p>\n\n<p>I have tried creating sym-links from a  that I found in my system to the cuda directory as suggested on other posts on the nvidia's website as well but they did not help.<\/p>\n\n<p>I have also set the parameters in ~\/.bashrc to and  but still no luck :( .<\/p>\n\n<p>It would be most helpful if anyone could help me get this sorted as it is a crucial part of my project to run the code on my gpu.<\/p>\n","answers":[{"AnswerId":"41840639","CreationDate":"2017-01-24T23:10:52.990","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>There is a typo in the <code>LD_LIBRARY_PATH<\/code> you typed, and tensorflow read. I see that you have <code>usr\/local\/cuda...<\/code> instead of <code>\/usr\/local\/cuda<\/code> (Note the <code>\/<\/code> in the beginning of the path)<\/p>\n\n<p>Usually, for me the following LD_LIBRARY_PATH is sufficient after installing cuda using the runfile installer from nvidia.\n<code>export LD_LIBRARY_PATH=\/usr\/local\/cuda:\/usr\/local\/cuda\/lib64:$LD_LIBRARY_PATH<\/code><\/p>\n"}]}
{"QuestionId":41840239,"AnswerCount":3,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-01-24T22:38:18.717","AcceptedAnswerId":41924386.0,"OwnerUserId":1097347.0,"Title":"Embedding vectors not being updated when using Tensorflow on window classification","Body":"<p>I am trying to implement a window based classifier with tensorflow,<\/p>\n\n<p>The word embedding matrix is called  and is initialized randomly (I tried Xavier also).<\/p>\n\n<p>And the  variable is the a vector of the indices of the word vectors from the matrix.<\/p>\n\n<p>The first layer is  (5) word vectors concatenated.<\/p>\n\n<pre><\/pre>\n\n<p>And this is how I run the graph:<\/p>\n\n<pre><\/pre>\n\n<p><strong>The dimensions work out, and the code runs<\/strong><\/p>\n\n<p>However, the accuracy is near zero, and I suspect that the the word vectors aren't being updated properly.<\/p>\n\n<p>How can I make tensorflow update the word vectors back from the concatenated window form ?<\/p>\n","answers":[{"AnswerId":"42017300","CreationDate":"2017-02-03T05:20:31.923","ParentId":null,"OwnerUserId":"811789","Title":null,"Body":"<p>Your starting of algorithm is fine. But I have some confidence this approach don't work. Actually word to vector trick is became working after estimation approximations are found applicable for NLP . For example techniques called <strong><em>Importance Sampling<\/em><\/strong> and <strong><em>Noise-Contrastive Estimation<\/em><\/strong>. <\/p>\n\n<p>So why straight approach doesn't work? I think, that to solve the task model must precisely find right 1 answer from large vocabulary, say 80000 words. 1 from 80000 - is too hard to optimize model, gradients didn't tell anything for most cases.<\/p>\n\n<p><strong>Update<\/strong>:\nI forgot to mention that main reason of estimation approximation is performance issues of straight approach were you have large output. Each iteration steps for all examples require calculating loss for each output unit (like 80000). Optimization will take long time to be intractable.<\/p>\n\n<p>How to implement right word2vec using sampling and NCE loss? Easily, following tutorial <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/udacity\/5_word2vec.ipynb\" rel=\"nofollow noreferrer\">here<\/a>, loss function looks like this:<\/p>\n\n<pre><code>loss = tf.reduce_mean(\n    tf.nn.sampled_softmax_loss(weights=softmax_weights, biases=softmax_biases, inputs=embed,\n                           labels=train_labels, num_sampled=num_sampled, num_classes=vocabulary_size))\n<\/code><\/pre>\n\n<p>Main idea is we need only few m negative samples and 1 positive. Where m is far less than actual vocabulary size.<\/p>\n\n<p>Tensorflow also has <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/candidate_sampling#nce_loss\" rel=\"nofollow noreferrer\">tf.nn.nce_loss<\/a><\/p>\n\n<p>You can read more about mathematical behind of approaches in online book <a href=\"http:\/\/www.deeplearningbook.org\/\" rel=\"nofollow noreferrer\">www.deeplearningbook.org<\/a> (I. Goodfellow et al)<\/p>\n"},{"AnswerId":"41885525","CreationDate":"2017-01-27T00:49:13.607","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>You need to initialize your <code>W<\/code> matrices to a random value.\nRight now <code>y1<\/code> is always 0 due to zero initialization.<\/p>\n"},{"AnswerId":"41924386","CreationDate":"2017-01-29T18:24:58.733","ParentId":null,"OwnerUserId":"2704763","Title":null,"Body":"<p>Your embeddings are initialised using <code>tf.Variable<\/code> which are by default trainable. They will be updated. The problem might be with the way you are calculating loss. Look at these following lines<\/p>\n\n<pre><code>y1 = tf.nn.softmax(tf.matmul(y0, W1) + b1)\ny_ = tf.placeholder(tf.float32, [None, out_layer])\ncross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y1), reduction_indices=[1])) \n<\/code><\/pre>\n\n<p>Here you are calculating the softmax function which converts the scores into probabilities <\/p>\n\n<p><img src=\"https:\/\/latex.codecogs.com\/gif.latex?softmax%28x_j%29%20%3D%20%5Cfrac%7Be%5E%7Bx_j%7D%7D%7B%5Csum_%7Bk%3D1%7D%5E%7BN_c%7De%5E%7Bx_k%7D%7D\" alt=\"softmax equation\"><\/p>\n\n<p>If the denominator here becomes too large or too small then this function can go for a toss. To avoid this numerical instability usually a small <code>epsilon<\/code> is added like below. This makes sure that there is numerical stability. <\/p>\n\n<p><img src=\"https:\/\/latex.codecogs.com\/gif.latex?softmax%28x_j%20&plus;%20%5Cepsilon%29%20%3D%20%5Cfrac%7Be%5E%7Bx_j%20&plus;%20%5Cepsilon%7D%7D%7B%5Csum_%7Bk%3D1%7D%5E%7BN_c%7De%5E%7Bx_k%20&plus;%20%5Cepsilon%7D%7D%20%5C%5C%20%5C%5C%20%5C%5C%20%3D%5Cfrac%7Be%5E%7Bx_j%7D.%20e%5E%7B%5Cepsilon%7D%7D%7B%5Csum_%7Bk%3D1%7D%5E%7BN_c%7D%20e%5E%7Bx_k%7D%20e%5E%7B%5Cepsilon%7D%20%7D%20%5C%5C%20%5C%5C%20%5C%5C%20%3D%20%5Cfrac%7Be%5E%7Bx_j%7D%7D%7B%5Csum_%7Bk%3D1%7D%5E%7BN_c%7De%5E%7Bx_k%20%7D%7D%20%5C%5C%20%5C%5C%20%5C%5C%20%3D%20softmax%28x_j%29\" alt=\"softmax_with_epsilon\"><\/p>\n\n<p>You can see that even after adding an <code>epsilon<\/code> the softmax functions value remains the same. If you don't handle this on your own then the gradients may not update properly due to vanishing or exploding gradients.<\/p>\n\n<p>Avoid the three lines of code and use the tensorflow version \n<code>tf.nn.sparse_softmax_cross_entropy_with_logits<\/code><\/p>\n\n<p>Note that this function will calculate the softmax function internally.\nIt is advisable to use this instead of calculating the loss manually. You can use this as follows <\/p>\n\n<pre><code>y1 = tf.matmul(y0, W1) + b1\ny_ = tf.placeholder(tf.float32, [None, out_layer])\ncross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y1, labels=y_))\n<\/code><\/pre>\n"}]}
{"QuestionId":41840513,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow><dataset>","CreationDate":"2017-01-24T23:01:38.630","AcceptedAnswerId":null,"OwnerUserId":6304774.0,"Title":"Tensorflow make Training Testing dataset from images","Body":"<p>I am making an image classifier I downloaded images of 2 classes to 2 folders:<\/p>\n\n<pre><\/pre>\n\n<p>How can I make a training and testing dataset of images in tensorflow by shuffling the images and spliting into training and testing sets.<\/p>\n","answers":[{"AnswerId":"57963845","CreationDate":"2019-09-16T20:15:09.350","ParentId":null,"OwnerUserId":"11621702","Title":null,"Body":"<p>Set the working directory to the folder with the 1st class: <\/p>\n\n<p>First you will need to download the library <code>split-folders<\/code> (<code>pip install split-folders<\/code>)<\/p>\n\n<pre><code>import os\nimport split_folders\n\ncwd = os.getcwd()  \n\nsplit_folders.ratio(cwd, output=\"output\", seed=1337, ratio=(.8, .2)) \n<\/code><\/pre>\n\n<p>Repeat for the other class folders.<\/p>\n"},{"AnswerId":"41864364","CreationDate":"2017-01-26T00:57:59.290","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>Following should work:\n<code>tf.split_v(tf.shuffle(images, ...), ...)<\/code><\/p>\n"}]}
{"QuestionId":41840619,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-24T23:09:06.517","AcceptedAnswerId":null,"OwnerUserId":3009947.0,"Title":"Pretrained embedding type error","Body":"<p>I am creating a computational graph in Tensorflow and I want to use the pretrained vectors. I have a method the preloads the vectors of all my words in the dataset into a matrix.<\/p>\n\n<pre>\n    def preload_vectors(word2vec_path, word2id, vocab_size, emb_dim):\n    if word2vec_path:\n        print('Load word2vec_norm file {}'.format(word2vec_path))\n        with open(word2vec_path,'r') as f:\n            header=f.readline()\n            print(vocab_size, emb_dim)\n            scale = np.sqrt(3.0 \/ emb_dim)\n            init_W = np.random.uniform(-scale, scale, [vocab_size, emb_dim])\n\n            print('vocab_size={}'.format(vocab_size))\n            while True:\n                line=f.readline()\n                if not line:break\n                word=line.split()[0]\n                if word in word2id:\n                    init_W[word2id[word]] = np.array(line.split()[1:], dtype = np.float32)\n    return init_W\n\n    init_W = preload_vectors(\"data\/GoogleNews-vectors-negative300.txt\", word2id, word_vocab_size, FLAGS.word_embedding_dim)\n\n<\/pre>\n\n<p>Output:<\/p>\n\n<pre>\n    Load word2vec_norm file data\/GoogleNews-vectors-negative300.txt\n    2556 300\n    vocab_size=2556\n<\/pre>\n\n<p>In the computational graph, I have this:<\/p>\n\n<pre>\n    W = tf.Variable(tf.constant(0.0, shape = [word_vocab_size,FLAGS.word_embedding_dim]),trainable = False, name='word_embeddings')\n    embedding_placeholder = tf.placeholder(tf.float32, shape = [word_vocab_size, FLAGS.word_embedding_dim])\n    embedding_init = W.assign(embedding_placeholder)\n<\/pre>\n\n<p>And finally, in the session I feed the init_W to embedding_placeholder:<\/p>\n\n<pre><\/pre>\n\n<p>But I get this error:<\/p>\n\n<pre><\/pre>\n\n<p><\/p>\n\n<p>I checked the values of the init_W array and they are float:<\/p>\n\n<pre><\/pre>\n\n<p>I used to be able to do this recently with no problems. I must have missed out something here? Please, I need your help. Thanks!<\/p>\n","answers":[{"AnswerId":"41848493","CreationDate":"2017-01-25T09:59:02.763","ParentId":null,"OwnerUserId":"3009947","Title":null,"Body":"<p>Apparently, the problem was not with the embedding_placeholder but with some other input to feed_dict. It was not clear though with the error message. <\/p>\n"}]}
{"QuestionId":41840772,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-24T23:22:43.947","AcceptedAnswerId":41842780.0,"OwnerUserId":2059708.0,"Title":"Tensorflow reading data from tfrecords into mini batches properly","Body":"<p>I am trying to convert data from csv to tfrecords then read it out in mini batches and do a simple MLP but I am getting some errors that I can't figure out. <\/p>\n\n<blockquote>\n  <p>RuntimeError: Attempted to use a closed Session.<\/p>\n<\/blockquote>\n\n<p>Followed by:<\/p>\n\n<blockquote>\n  <p>TypeError: The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, or numpy ndarrays.<\/p>\n<\/blockquote>\n\n<p>I'm guessing the shuffle batch queue is closing and no longer feeding the expected data. Also, I think I am missing a step going from the shuffle queue to the feed dict. Any ideas how to make this work or a better way of doing it?<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>Stack Trace:<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41842780","CreationDate":"2017-01-25T03:14:57.850","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>Placeholders are one way of getting data into your model. Queue's are another. You can override the value in a tensor generated by a queue runner like any other tensor (e.g. placeholders), but you can't feed the results from a tensor into a placeholder in the same graph\/session run.<\/p>\n\n<p>In other words, rather than creating placeholders, just use the output of your <code>tf.train.shuffle_batch<\/code> call:<\/p>\n\n<pre><code>X = avgs_batch\nY = pdiff_batch\n<\/code><\/pre>\n\n<p>(or replace all references to <code>X<\/code> and <code>Y<\/code> with <code>avgs_batch<\/code> and <code>pdiff_batch<\/code> respectively)<\/p>\n"}]}
{"QuestionId":41841050,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-24T23:52:15.447","AcceptedAnswerId":41841243.0,"OwnerUserId":null,"Title":"TensorFlow: adding regularization to LSTM","Body":"<p>Following <a href=\"https:\/\/stackoverflow.com\/questions\/37869744\/tensorflow-lstm-regularization\">Tensorflow LSTM Regularization<\/a> I am trying to add regularization term to the cost function when training parameters of LSTM cells.<\/p>\n\n<p>Putting aside some constants I have:<\/p>\n\n<pre><\/pre>\n\n<p>The problem I have is that adding the regularization term hugely slows the learning process and actually the regularization term  is increasing with each iteration visibly when the term associated with  pretty much stagnated i.e. the  seems not to be taken into account.<\/p>\n\n<p>As I suspect I am adding this term in completely wrong way. I did not know how to add this term in the cost function itself so I used a workaround with scalar tf.placeholder and \"manually\" calculated the regularization cost. How to do it properly?<\/p>\n","answers":[{"AnswerId":"41841243","CreationDate":"2017-01-25T00:12:07.257","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>compute the L2 loss only once: <\/p>\n\n<pre><code>tv = tf.trainable_variables()\nregularization_cost = tf.reduce_sum([ tf.nn.l2_loss(v) for v in tv ])\ncost = tf.reduce_sum(tf.pow(pred - y, 2)) + regularization_cost\noptimizer = tf.train.AdamOptimizer(learning_rate = 0.01).minimize(cost)\n<\/code><\/pre>\n\n<p>you might want to remove the variables that are <code>bias<\/code> as those should not be regularized.  <\/p>\n"},{"AnswerId":"41841253","CreationDate":"2017-01-25T00:12:37.210","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>It slows down because your code <strong>creates new nodes in every iteration<\/strong>. This is not how you code with TF. First, you create your whole graph, including regularization terms, then, in the while loop you only execute them, each \"tf.XXX\" operation creates <strong>new nodes<\/strong>.<\/p>\n"}]}
{"QuestionId":41841492,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-25T00:36:45.637","AcceptedAnswerId":41842186.0,"OwnerUserId":5106953.0,"Title":"Softmax Jacobian in Tensorflow","Body":"<p>Suppose I have a simple one-layer neural network:<\/p>\n\n<pre><\/pre>\n\n<p>Thus, the variable  is of dimension  by . I want to compute the Jacobian of  with respect to  for each sample in the batch, which will be of dimension  by  by . Now, mathematically, the Jacobian (dy\/da)_{i,j} = -y_i y_j for i != j and otherwise, (dy\/da)_{i,i} = y_i (1 - y_i). <\/p>\n\n<p>I was wondering how to compute this Jacobian of the softmax with respect to its input in TensorFlow? I know that  will compute the gradient of a scalar with respect to a tensor, so I figure some combination of looping in TensorFlow with either  or even just trying to implement the analytic form given above should work. But I am not sure how to do this in TensorFlow with its ops and would appreciate any code to do this!<\/p>\n","answers":[{"AnswerId":"41842186","CreationDate":"2017-01-25T01:58:31.487","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>It seems <code>tf.gradients<\/code> applies a sum over the <code>output_dim<\/code>. Solution: unstack then restack. Not sure how this affect efficiency...<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nbatch_size = 3\ninput_dim = 10\noutput_dim = 20\n\nW_vals = np.random.rand(input_dim, output_dim).astype(np.float32)\n\ngraph = tf.Graph()\nwith graph.as_default():\n    x = tf.placeholder(tf.float32, [batch_size, input_dim])\n    # Use a constant for easier checking\n    W = tf.constant(W_vals, dtype=tf.float32)\n    a = tf.matmul(x, W)\n    y = a\n    # remove softmax for easier checking\n    # y = tf.nn.softmax(a)\n\n    grads = tf.stack([tf.gradients(yi, x)[0] for yi in tf.unstack(y, axis=1)],\n                     axis=2)\n\nwith tf.Session(graph=graph) as sess:\n    x_vals = np.random.rand(batch_size, input_dim).astype(np.float32)\n    g_vals = sess.run(grads, feed_dict={x: x_vals})\n\n# check gradients match\ntol = 1e-10\nfor i in range(batch_size):\n    if np.max(np.abs(g_vals[i] - W_vals)) &gt;= tol:\n        raise Exception('')\nprint('Gradients seem to match!')\n<\/code><\/pre>\n"}]}
{"QuestionId":41841713,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-25T01:00:02.750","AcceptedAnswerId":null,"OwnerUserId":7204444.0,"Title":"tensorflow gradient: unsupported operand type","Body":"<p>I got the following error:<\/p>\n\n<p>anaconda2\/envs\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gradients.py:90: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.\n  \"Converting sparse IndexedSlices to a dense Tensor of unknown shape. \"\nTraceback (most recent call last):<\/p>\n\n<pre><\/pre>\n\n<p>File \"anaconda2\/envs\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 196, in minimize\n    grad_loss=grad_loss)\n  File \"anaconda2\/envs\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/training\/optimizer.py\", line 253, in compute_gradients\n    colocate_gradients_with_ops=colocate_gradients_with_ops)\n  File \"anaconda2\/envs\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gradients.py\", line 469, in gradients\n    in_grads = _AsList(grad_fn(op, *out_grads))\n  File \"anaconda2\/envs\/tensorflow\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/array_grad.py\", line 504, in _ExtractImagePatchesGrad\n    rows_out = int(ceil(rows_in \/ stride_r))\nTypeError: unsupported operand type(s) for \/: 'NoneType' and 'long'<\/p>\n\n<p>there is look like gather ops is wrong.<\/p>\n","answers":[{"AnswerId":"56265842","CreationDate":"2019-05-22T22:35:00.120","ParentId":null,"OwnerUserId":"11506976","Title":null,"Body":"<p>I see that this is an old issue, but I have found a quick work-around for some cases of this.  Chances are, you are feeding your input using a placeholder and one of the dimensions of the placeholder shape is \"None\".  If you set that dimension to your batch size, it will no longer be an unknown shape.<\/p>\n"}]}
{"QuestionId":41842299,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-25T02:11:47.273","AcceptedAnswerId":null,"OwnerUserId":1625695.0,"Title":"Loss function with argmax returns int","Body":"<p>I'm defining a loss function RMSE as follows:<\/p>\n\n<pre><\/pre>\n\n<p>model.pred is int64 but sqrt gives a compilation error because it requires float input. When I use tf.cast the loss function becomes not differentiable. How can I fix this ?<\/p>\n","answers":[{"AnswerId":"41842842","CreationDate":"2017-01-25T03:21:43.127","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>The <code>argmax<\/code> is non-differentiable, so most functions involving it will also be non-differentiable. Without knowing more about the problem it's difficult to give a solution. If it's a classification problem use the logits (<code>pred<\/code>) rather than the argmax.<\/p>\n"}]}
{"QuestionId":41842310,"AnswerCount":2,"Tags":"<image-processing><machine-learning><tensorflow><computer-vision>","CreationDate":"2017-01-25T02:13:05.347","AcceptedAnswerId":41842957.0,"OwnerUserId":6102784.0,"Title":"Does image size matter when training with TensorFlow?","Body":"<p>I was wondering if there is any benefit to training on high resolution images rather than low resolution. I understand that it will take longer to train on larger images and that the dimensions must be a multiple of 32. My current image set is 1440x1920. Would I be better off resizing to 480x640, or is bigger better?<\/p>\n","answers":[{"AnswerId":"41842926","CreationDate":"2017-01-25T03:30:28.677","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>This depends largely on the application. As a rule of thumb, I'd ask myself the question: can I complete the task myself on the resized images? If so, I'd downsize to the lowest resolution before it makes the task more difficult for you yourself. If not... you're going to have to be -very- patient using images 1440 * 1920. I imagine you'll almost always be better off experimenting with more varied architectures and hyper-parameter sets on smaller images compared to fewer models on full resolution images.<\/p>\n\n<p>Whatever size you choose, you'll have to design your network for the image size you have in mind. If you're using convolutional layers, a larger image will require larger strides, filter sizes and\/or layers. The number of parameters will stay the same for each convolution, though the number of features will grow (along with batch normalisation parameters if you're using it).<\/p>\n"},{"AnswerId":"41842957","CreationDate":"2017-01-25T03:34:49.123","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p>It's certainly not a requirement that your images be powers of two. There may be some cases where it speeds things up (e.g. GPU allocation) but it's not critical.<\/p>\n\n<p>Smaller images will train significantly faster, and possibly even converge quicker (all other factors held constant) as you will be able to train on bigger batches (e.g. 100-1000 images in one pass, which you might not be able to do on a single machine with high res imagery).<\/p>\n\n<p>As to whether to resize, you need to ask yourself if every pixel in that image is critical to your task. Often this is not the case - you can probably resize a photo of a bus down to say 128x128 and still recognize that it's a bus.<\/p>\n\n<p>Using smaller images can also help your network generalise better, too, as there is less data to overfit.<\/p>\n\n<p>A technique often used in image classification networks is to perform distortions (e.g. random cropping, scaling &amp; brightness adjustment) on images to (a) convert odd-sized images to a constant size, (b) synthesize more data and (c) encourage the network to generalise.<\/p>\n"}]}
{"QuestionId":41842440,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-25T02:31:32.043","AcceptedAnswerId":null,"OwnerUserId":3218279.0,"Title":"TensorFlow - Input 'split_dim' of 'Split' Op has type float32 that does not match expected type of int32","Body":"<p>I've installed tensorflow using pip on ubuntu 16.04 LTS, when running this code <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/recurrent_network.py\" rel=\"noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/recurrent_network.py<\/a> i am getting this error<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41868500","CreationDate":"2017-01-26T07:19:54.483","ParentId":null,"OwnerUserId":"3450793","Title":null,"Body":"<p>It looks like you are using an older version of Tensorflow, and need to update to Tensorflow v0.12.0 or above. The error you are getting is indicating that the <code>split_dim<\/code> value in your <code>tf.split<\/code> function is expecting an integer, but is receiving the tensor <code>x<\/code> which is of type float32.<\/p>\n\n<p>This is because in Tensorflow versions &lt; 0.12.0 the split function takes the arguments as:<\/p>\n\n<pre><code>x = tf.split(0, n_steps, x) # tf.split(axis, num_or_size_splits, value)\n<\/code><\/pre>\n\n<p>The tutorial you are working from was written for versions > 0.12.0, which has been changed to be consistent with Numpy's split syntax:<\/p>\n\n<pre><code>x = tf.split(x, n_steps, 0) # tf.split(value, num_or_size_splits, axis)\n<\/code><\/pre>\n\n<p>See the changelog for details: \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/64edd34ce69b4a8033af5d217cb8894105297d8a\/RELEASE.md\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/64edd34ce69b4a8033af5d217cb8894105297d8a\/RELEASE.md<\/a><\/p>\n"},{"AnswerId":"41842564","CreationDate":"2017-01-25T02:48:33.367","ParentId":null,"OwnerUserId":"7444549","Title":null,"Body":"<p>It is because the argument order has changed <\/p>\n\n<p>You can see the issue here : <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6501\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6501<\/a><\/p>\n"}]}
{"QuestionId":41842796,"AnswerCount":2,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2017-01-25T03:16:15.747","AcceptedAnswerId":41842847.0,"OwnerUserId":7466373.0,"Title":"Tensorflow- Recurrent Neural Networks","Body":"<p>I got the files reader.py and ptb_word_lm.py from google and tried running them as . Please help me in resolving these errors.<\/p>\n\n<p>Errors:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41842847","CreationDate":"2017-01-25T03:22:03.627","ParentId":null,"OwnerUserId":"7444549","Title":null,"Body":"<p>First, you must download and extract the data in the same directory with you python code <\/p>\n\n<p><code>$ wget http:\/\/www.fit.vutbr.cz\/~imikolov\/rnnlm\/simple-examples.tgz<\/code><\/p>\n\n<p><code>$ tar xvf simple-examples.tgz<\/code><\/p>\n\n<p>Then, you must specify the data directory path for the code to work.\nYou can read the documentation that give sample command to run the program <\/p>\n\n<p><code>$ python ptb_word_lm.py --data_path=simple-examples\/data\/<\/code><\/p>\n"},{"AnswerId":"55333416","CreationDate":"2019-03-25T08:00:12.490","ParentId":null,"OwnerUserId":"6412744","Title":null,"Body":"<p>Using the complete path should solve your problem. \nUse<\/p>\n\n<p><code>python models\/tutorials\/rnn\/ptb\/ptb_word_lm.py --data_path=simple-examples\/data\/ --model=small<\/code><\/p>\n"}]}
{"QuestionId":41843037,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-25T03:43:36.373","AcceptedAnswerId":null,"OwnerUserId":5703903.0,"Title":"How to feed same result from tf.train.shuffle_batch to different functions?","Body":"<p>Following code is in the context of Generative Adversarial Networks.<\/p>\n\n<p> is the standard input pipeline you would expect from CIFAR10. The key thing to note is that it returns a batch from . I'm doing this in my code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, since  is a placeholder that draws from the queue, calling  twice like that would result in a totally different set of images. Specifically, generator would receive Set A of images and discriminator would receive Set B of images. I want the discriminator to receive Set A as well.<\/p>\n\n<p>What's the best way to solve this?<\/p>\n","answers":[]}
{"QuestionId":41843600,"AnswerCount":1,"Tags":"<cuda><tensorflow><cudnn>","CreationDate":"2017-01-25T04:48:57.363","AcceptedAnswerId":null,"OwnerUserId":2118051.0,"Title":"how to install cudnn5.1 with cuda 8.0 support for tensorflow?","Body":"<p>I am unable to find the cudnn-8.0-linux-x64-v5.1.tgz file. The download link has only .deb file and when I install it using \n    sudo dpkg -i \/path\/to\/deb\/file\nI get libcudnn.so.5.1.5 file and not the headers (cudnn.h). Where can I get the .tgz file with all the .so and .h files? I am looking for libcudnn.so.5.1<\/p>\n","answers":[{"AnswerId":"41849058","CreationDate":"2017-01-25T10:24:05.303","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>You can download cuDNN from <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow noreferrer\">here<\/a>, you need to have an NVIDIA developer account which is free of cost. After downloading it, extract the contents and copy the files to appropriate locations:<\/p>\n\n<pre><code>$ sudo cp -P include\/cudnn.h \/usr\/include\n$ sudo cp -P lib64\/libcudnn* \/usr\/lib\/x86_64-linux-gnu\/\n$ sudo chmod a+r \/usr\/lib\/x86_64-linux-gnu\/libcudnn*\n<\/code><\/pre>\n"}]}
{"QuestionId":41844800,"AnswerCount":1,"Tags":"<tensorflow><distributed-computing><horizontal-scaling>","CreationDate":"2017-01-25T06:23:51.473","AcceptedAnswerId":null,"OwnerUserId":767632.0,"Title":"How does TensorFlow cluster distribute load across machines if not specified explicitly?","Body":"<p>I took <a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">\"Distributed TensorFlow\"<\/a> how-to and tried to apply it to the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">\"MNIST For ML Beginners\"<\/a> tutorial. I started three TensorFlow worker nodes locally (there are 8 cores in the PC) and ran the training script with replacing this line:<\/p>\n\n<pre><\/pre>\n\n<p>with the following:<\/p>\n\n<pre><\/pre>\n\n<p>where  is a port where node 0 is listening (e.g. master session is created on node 0). Note that I did not specify explicitly where computations should be performed.<\/p>\n\n<p>Looking at 's output, I can see that the job is indeed performed by the cluster - it consumes some CPU. However, the only consumer is node 0, remaining nodes do not perform any work. If I select node 1 as a place to create master session, picture changes: only ~2\/3 of the work is performed on node 0 (judging by CPU load), but the remaining 1\/3 of the work is performed on node 1. If I select node 2 as master, then that 1\/3 of the work is performed on node 2. If I run two processes in parallel, one using node 1 as master and another using node 2 as master, both nodes 1 and 2 get some load, but node 0 is loaded much more (like, 200% vs 60% vs 60% of CPU).<\/p>\n\n<p>So far it looks like \"default\" behavior of distributed TensorFlow is not great for parallelizing work automatically right now. I'm wondering what the behavior is and whether distributed TensorFlow is intended for data parallelization at all (as opposed to manual model parallelization)?<\/p>\n","answers":[{"AnswerId":"41845710","CreationDate":"2017-01-25T07:26:20.140","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>TF is great for data parallelization, e.g. when you need to sift through tons of data, which is then distributed to multiple GPUs.<\/p>\n\n<p>It's also great for weights parallelization. Using <code>tf.train.replica_device_setter<\/code>, weights are distributed among multiple devices for better IO.<\/p>\n\n<p>Now, it seems you are asking for parallelization within a single model. That's difficult to do automatically, since TF does not know what's the best way to distribute your computation of the same model to multiple devices. It would depend on too many factors, e.g. how fast is the connection between your devices.<\/p>\n"}]}
{"QuestionId":41845321,"AnswerCount":3,"Tags":"<tensorflow><computer-vision><deep-learning>","CreationDate":"2017-01-25T07:00:44.483","AcceptedAnswerId":41845563.0,"OwnerUserId":1928778.0,"Title":"Should deep learning classification be used to classify details such as liquid level in bottle","Body":"<p>Can  be used to precisely  both the object and one of its features. For example to identify the bottle <em>(like Grants Whiskey)<\/em> and liquid level in the bottle <em>(in 10 percent steps - like 50% full)<\/em>. Is this the problem that can be best solved utilizing some of  <em>(Tensorflow etc)<\/em> or some other approach is more effective?<\/p>\n","answers":[{"AnswerId":"41846046","CreationDate":"2017-01-25T07:47:23.270","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>The rule of thumb for deep learning is, if it is visible in the image, hence a expert can tell you the answer solely based on the image then the chances are very high that you can learn this with deep learning, given enough annotated data. <\/p>\n\n<p>However you are quite unlikely to have the required data needed for such a task, therefore I would ask myself the question if i can simplify the problem. For example you could take gin, vodka and so on and use SIFT to find the bottle again in a new scene. Then RANSAC for bottle detection and cut the bottle out of the image. <\/p>\n\n<p>Then I would try gradient features to find the edge with the liquid level. Finally you can calculate the percentage with (liquid edge - bottom) \/ (top bottle - bottom bottle). <\/p>\n"},{"AnswerId":"41851544","CreationDate":"2017-01-25T12:21:31.423","ParentId":null,"OwnerUserId":"7451758","Title":null,"Body":"<p>Identifying the bottle label should not be hard to do - it's even available \"on tap\" for cheap (<a href=\"https:\/\/www.baacco.com\/\" rel=\"nofollow noreferrer\">these guys<\/a> actually use it to identify wine bottle labels on their website): <a href=\"https:\/\/aws.amazon.com\/blogs\/aws\/amazon-rekognition-image-detection-and-recognition-powered-by-deep-learning\/\" rel=\"nofollow noreferrer\">https:\/\/aws.amazon.com\/blogs\/aws\/amazon-rekognition-image-detection-and-recognition-powered-by-deep-learning\/<\/a><\/p>\n\n<p>Regarding the liquid level, this might be a problem AWS wouldn't be able to solve straight away - but I'm sure it's possible to make a custom CNN for it. Alternatively, you could use good old humans on Amazon Mechanical Turk to do the work for you.<\/p>\n\n<p>(Note: I do not work for Amazon)<\/p>\n"},{"AnswerId":"41845563","CreationDate":"2017-01-25T07:16:05.450","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>Well, this should be well possible if the liquor is well colored. If not (e.g. gin, vodka), I'd say you have no chance with today's technologies when observing the object from a natural view angle and distance.<\/p>\n\n<p>For colored liquor, I'd train two detectors. One for detecting the bottle, and a second one to detect the liquor given the bottle. The ratio between the two will be your percentage.<\/p>\n\n<p>Some of the proven state-of-the-art deep learning-based object detectors (just Google them):<\/p>\n\n<ul>\n<li>Multibox<\/li>\n<li>YOLO<\/li>\n<li>Faster RCNN<\/li>\n<\/ul>\n\n<p>Or non-deep-learning-based:<\/p>\n\n<ul>\n<li>Deformable part model<\/li>\n<\/ul>\n\n<p>EDIT:<\/p>\n\n<p>I was ask to elaborate more. Here is an example:<\/p>\n\n<p>The box detector e.g. draws a box in the image at [0.1, 0.2, 0.5, 0.6] (min_height, min_width, max_height, max_width) which is the relative location of your bottle.<\/p>\n\n<p>Now you crop the bottle from the original image and feed it to the second detector. The second detector draws e.g. [0.2, 0.3, 0.7, 0.8] in your cropped bottle image, the location indicates the fluid it has detected. Now (0.7 - 0.2) * (0.8 - 0.3) = 0.25 is the relative area of the fluid with respect to the area of the bottle, which is what OP is asking for.<\/p>\n\n<p>EDIT 2:<\/p>\n\n<p>I entered this reply assuming OP wants to use deep learning. I'd agree other methods should be considered if OP is still unsure with deep learning. For bottle detection, deep learning-based methods have shown to outperform traditional methods by a large margin. Bottle detection happens to be one of the classes in the PASCAL VOC challenge. See results comparison here: <a href=\"http:\/\/rodrigob.github.io\/are_we_there_yet\/build\/detection_datasets_results.html#50617363616c20564f43203230313020636f6d7034\" rel=\"nofollow noreferrer\">http:\/\/rodrigob.github.io\/are_we_there_yet\/build\/detection_datasets_results.html#50617363616c20564f43203230313020636f6d7034<\/a><\/p>\n\n<p>For the liquid detection however, deep learning might be slightly overkill. E.g. if you know what color you are looking for, even a simple color filter will give you \"something\".... <\/p>\n"}]}
{"QuestionId":41845379,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving><horizontal-scaling>","CreationDate":"2017-01-25T07:04:14.750","AcceptedAnswerId":41858303.0,"OwnerUserId":767632.0,"Title":"Is it possible to use TensorFlow Serving with distributed TensorFlow cluster to improve throughput\/latency?","Body":"<p>I'm looking into ways to improve latency and\/or throughput of a <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"nofollow noreferrer\">TensorFlow Serving<\/a> instance. I've seen the <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_inception\" rel=\"nofollow noreferrer\">\"Serving Inception\"<\/a> manual and three GitHub Issues (<a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/10\" rel=\"nofollow noreferrer\">2<\/a>, <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/49\" rel=\"nofollow noreferrer\">3<\/a>, <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/60\" rel=\"nofollow noreferrer\">4<\/a>), but all of them seem to create a separate instance of TensorFlow Serving per server and then choosing server on client. Issue <a href=\"https:\/\/github.com\/tensorflow\/serving\/issues\/60\" rel=\"nofollow noreferrer\">4<\/a> is actually about adding some load balancer in front of that stuff, which is currently absent in TensorFlow Serving itself.<\/p>\n\n<p>However, there is also <a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">\"Distributed TensorFlow\"<\/a> tutorial which shows how to join a set of machines into a fixed cluster and then manually \"pin\" some computations to some machines, which can improve both latency and throughput if model is \"wide\" and can be parallelized well. However, I do not see any mentions of combining this with TensorFlow Serving in either documentation.<\/p>\n\n<p>Question is: is it possible to configure TensorFlow Serving to use distributed TensorFlow cluster?<\/p>\n\n<p>I was able to make it create and use gRPC sessions (instead of local) with some hacks:<\/p>\n\n<ol>\n<li>Make  target publicly visible (it's internal to tensorflow package by default) by modifying  file.<\/li>\n<li>Add it as a dependency to the  target.<\/li>\n<li>Add an extra flag to  called  which sets up  in .<\/li>\n<li>Run the binary with , where  is an arbitrary node which will be used to create master sessions.<\/li>\n<li>See my cluster performing some computations on behalf of TensorFlow Serving.<\/li>\n<\/ol>\n\n<p>However, this set of hacks does not look enough for \"real-world usage\" for three reasons:<\/p>\n\n<ol>\n<li> target is probably internal for a reason.<\/li>\n<li>As noticed in my <a href=\"https:\/\/stackoverflow.com\/questions\/41844800\/\">other question<\/a>, distributed TensorFlow works better when computations are manually \"pinned\" to specific machines. So, if we use TensorFlow Serving, we need a way to save those \"pins\" and model's structure becomes tied with cluster's structure. I'm not sure whether this information is exported with \/ at all.<\/li>\n<li> creates session once - during bootstrap. If master node of the cluster goes down and then restores, serving server still holds the \"old\" session and cannot process further requests.<\/li>\n<\/ol>\n\n<p>All in all, it looks like this scenario is not officially supported yet, but I'm not sure.<\/p>\n","answers":[{"AnswerId":"41858303","CreationDate":"2017-01-25T17:52:11.320","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>If your model fits into single machine, then it's hard to see how distributing it over many machines will improve throughput. Essentially you are taking computations which can be done independently and adding a dependency. If one of your machines is slow or crashes, instead of making some queries slow, it will make all queries sow.<\/p>\n\n<p>That said, it's worth benchmarking to see if it helps, in which case it would make sense to ask for this use-case to be officially supported.<\/p>\n\n<p>Regarding questions:<\/p>\n\n<ol start=\"2\">\n<li><p>Worker assignments are done through <code>device<\/code> field in graph <code>.pbtxt<\/code>. Some importers\/exporters clear those assignments and have <code>clear_devices<\/code> flag. You could open graph definition (<code>.pbtxt<\/code> file or equivalently, str(tf.get_default_graph().as_graph_def(), and grep for <code>device<\/code> strings to check)<\/p><\/li>\n<li><p>If any worker restarts, or there's some temporary network connectivity your sess.run fails with error (Unavailable) and you need  to recreate the session. This is handled automatically by <code>MonitoredTrainingSession<\/code> in <code>tf.train<\/code>, but you need to handle this yourself with serving.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"45206650","CreationDate":"2017-07-20T06:19:25.183","ParentId":null,"OwnerUserId":"3910238","Title":null,"Body":"<p>If your model is not using images, or is not entirely too large, you shouldn't need too much compute for each inference\/serve, and I'm saying this using Inception-v# which takes ~1 sec to serve a response to an image on a Google Cloud Platform n1-standard-1 machine.<\/p>\n\n<p>Now that being said, perhaps its the throughput that you need to scale up and that is a different problem. Your best option for scale at that point would be to use Docker Swarm &amp; Compose, as well as Kubernetes to help scale e up and serve your inference \"micro-service\". You could use flask to iterate over a sequence of requests also if your use-case warrants it.<\/p>\n"}]}
{"QuestionId":41845802,"AnswerCount":2,"Tags":"<python><statistics><tensorflow><probability-density>","CreationDate":"2017-01-25T07:31:29.243","AcceptedAnswerId":41848710.0,"OwnerUserId":7457101.0,"Title":"how to calculate PDF in tensorflow","Body":"<p>In Matlab, I can calculate the probability density function(PDF) of Gaussian distribution by using <\/p>\n\n<pre><\/pre>\n\n<p>output: <\/p>\n\n<pre><\/pre>\n\n<p>Using tensorflow, I have tried with this <\/p>\n\n<pre><\/pre>\n\n<p>I get an error <\/p>\n\n<p>how to input the mu and sigma values to this distribution ? to get the similar output. <\/p>\n","answers":[{"AnswerId":"47706680","CreationDate":"2017-12-08T02:09:15.393","ParentId":null,"OwnerUserId":"7819691","Title":null,"Body":"<p>This function <code>pdf<\/code> is updated as <code>prob<\/code>:<\/p>\n\n<pre><code>dist = tf.contrib.distributions.Normal(1.0, 2.5) \ny = dist.prob(x)\n<\/code><\/pre>\n"},{"AnswerId":"41848710","CreationDate":"2017-01-25T10:08:37.097","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>First create a <code>Normal distributions<\/code> object and then use the <code>pdf<\/code> method<\/p>\n\n<pre><code>dist = tf.contrib.distributions.Normal(1.0, 2.5)\ny = dist.pdf(x)\n<\/code><\/pre>\n\n<p>See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard7\/tf.contrib.distributions.Normal.md\" rel=\"nofollow noreferrer\">this<\/a> for more details.<\/p>\n"}]}
{"QuestionId":41845821,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2017-01-25T07:32:44.700","AcceptedAnswerId":41847445.0,"OwnerUserId":7351755.0,"Title":"Model input error","Body":"<p>Here is the error in full:<\/p>\n\n<p><\/p>\n\n<p>Everything works until the final  chunk of code. I am using a theano backend.<\/p>\n\n<p>I'm pretty new to keras, so I'm not sure exactly how to proceed. Checking the documentation I can see that the  in  corresponds to the number of batches (or samples)? Substituting  yielded the following error . Using 23000 instead of 20 yielded the same error.<\/p>\n\n<p>Any help is appreciated.<\/p>\n\n<p>Below is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41847445","CreationDate":"2017-01-25T09:09:51.723","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>There is a mismatch between the expected size of the images and the actual one. Your model expects images of size <code>224 x 224<\/code> and according to the attached error message actual size is <code>244 x 244<\/code>.<\/p>\n"}]}
{"QuestionId":41846380,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-01-25T08:10:12.807","AcceptedAnswerId":null,"OwnerUserId":3839572.0,"Title":"Tensorflow VGG16 (converted from caffe) got low evaluation accuracy","Body":"<p>I didn't convert the weights by myself, instead I used vgg16_weights.npz from www(dot)cs(dot)toronto(dot)edu\/~frossard\/post\/vgg16\/. There, it is mentioned <\/p>\n\n<blockquote>\n  <p>We convert the Caffe weights publicly available in the author\u2019s GitHub profile (gist(dot)github(dot)com\/ksimonyan\/211839e770f7b538e2d8#file-readme-md) using a specialized tool (github(dot)com\/ethereon\/caffe-tensorflow).<\/p>\n<\/blockquote>\n\n<p>But, in that page, there is no validation code, so I made it referring to tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\" rel=\"nofollow noreferrer\">MNIST<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\" rel=\"nofollow noreferrer\">inception<\/a> code.<\/p>\n\n<p><strong>How I create TFRecords of Imagenet<\/strong><\/p>\n\n<p>I use build_imagenet_data.py from inception. I changed the <\/p>\n\n<blockquote>\n  <p>label_index = 0 #originally label_index = 1<\/p>\n<\/blockquote>\n\n<p>because inception use label_index 0 as background class (so in total there are 1001 classes). Caffe format doesn't use that as the number of output is 1000. I prefer to use TFRecord format as I will change process the weight and retrain.<\/p>\n\n<p><strong>How I load the weights<\/strong><\/p>\n\n<p>inference function taken from MNIST's mnist.py was modified so the Variable is taken from the vgg16_weights.npz<\/p>\n\n<p>How I load the weights:<\/p>\n\n<pre><\/pre>\n\n<p>How I put the variable in conv1_1:<\/p>\n\n<pre><\/pre>\n\n<p><strong>How I read the TFRecords<\/strong><\/p>\n\n<p>I took inception's image_processing.py, dataset.py, and ImagenetData.py with no change. Then, I run inception's inception_eval.py evaluate function with changing in inference code and deleting the restoring moving variable from checkpoint (as I already restore manually in variable initialization). However, the accuracy is not same with the VGG-16 in caffe. Top-5 accuracy is around 9%.<\/p>\n\n<p><strong>Closing<\/strong><\/p>\n\n<p>What is the problem of this method? There are several part of code that I still don't understand though:<\/p>\n\n<ol>\n<li>How TFReader move to the next batch of images after processing 1 batch of images? The output of inception's image_processing.py size is only the number of batch size. To be complete, this is the output based on documentation: <\/li>\n<\/ol>\n\n<blockquote>\n  <p>images: Images. 4D tensor of size [batch_size, FLAGS.image_size,\n                                         image_size, 3].<\/p>\n  \n  <p>labels: 1-D integer Tensor of [FLAGS.batch_size].<\/p>\n<\/blockquote>\n\n<ol start=\"2\">\n<li>Do I need softmax the logits before tf.in_top_k ? (Well, I don't think it is matter as the value sequence is same)<\/li>\n<\/ol>\n\n<p>Thank you for the help. Sorry if the link is messy as I can only post 2 links in 1 post because of my reputation.<\/p>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I tried myself by changing the caffe weight. Reverse the channel input dimension of conv1_1 (because caffe receive BGR, so the weight is for BGR instead of RGB in tensorflow) and get the same accuracy with the weight from website: around 9% in top-5.<\/p>\n\n<p>I found out that there is no mean image subtraction in tensorflow inception's image_processing.py. I add mean subtraction (in eval_image function) with tf.reduce_mean and got 11% accuracy.<\/p>\n\n<p>Then I tried to change the eval_image function with <\/p>\n\n<pre><\/pre>\n\n<p>and I got 13%. Increased but still lack a lot. Seems it is one of the problem. I am not sure what is the other problems.<\/p>\n","answers":[{"AnswerId":"42683750","CreationDate":"2017-03-08T23:14:45.670","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>In general porting whole model weights across libraries will be hard. You pointed out some differences from caffe, but there could be others. It might be easier to retrain the model in TensorFlow.<\/p>\n"}]}
{"QuestionId":41846972,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-25T08:44:50.327","AcceptedAnswerId":null,"OwnerUserId":7467838.0,"Title":"TensorFlow - How to read variable-length parts of data?","Body":"<p>I want to read variable_length parts of data of a single file using data reader in TensorFlow.\nMy file is as follows: [data_length_1, data_bytes_1, data_length_2, data_bytes_2, ...]\nThe first byte represents the number of bytes to be read, and then this process repeats.<\/p>\n\n<p>The problem is that I want to read data into queue in TensorFlow. I have tried <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/readers#FixedLengthRecordReader\" rel=\"nofollow noreferrer\">tf.FixedLengthRecordReader<\/a> , but it accepts only fixed length data to be read. Is there any way to change it into variable-length data reader?<\/p>\n","answers":[{"AnswerId":"42068155","CreationDate":"2017-02-06T12:45:07.670","ParentId":null,"OwnerUserId":"7467838","Title":null,"Body":"<p>I managed to solve my problem by adding a <a href=\"https:\/\/www.tensorflow.org\/how_tos\/new_data_formats\/\" rel=\"nofollow noreferrer\">custom data reader<\/a> that reads the record length for each record from its first data byte, and it works.<\/p>\n"}]}
{"QuestionId":41847376,"AnswerCount":3,"Tags":"<python><unicode><keras>","CreationDate":"2017-01-25T09:06:49.720","AcceptedAnswerId":41952183.0,"OwnerUserId":1934212.0,"Title":"Keras model.to_json() error: 'rawunicodeescape' codec can't decode bytes in position 94-98: truncated \\uXXXX","Body":"<pre><\/pre>\n\n<p>for the model<\/p>\n\n<blockquote>\n  <p>____________________________________________________________________________________________________ Layer (type)                     Output Shape          Param #<br>\n  Connected to<br>\n  ==================================================================================================== lambda_1 (Lambda)                (None, 3, 160, 320)   0<br>\n  lambda_input_1[0][0]<br>\n  ____________________________________________________________________________________________________ convolution2d_1 (Convolution2D)  (None, 1, 40, 16)     327696<br>\n  lambda_1[0][0]<br>\n  ____________________________________________________________________________________________________ elu_1 (ELU)                      (None, 1, 40, 16)     0<br>\n  convolution2d_1[0][0]<br>\n  ____________________________________________________________________________________________________ convolution2d_2 (Convolution2D)  (None, 1, 20, 32)     12832<br>\n  elu_1[0][0]<br>\n  ____________________________________________________________________________________________________ elu_2 (ELU)                      (None, 1, 20, 32)     0<br>\n  convolution2d_2[0][0]<br>\n  ____________________________________________________________________________________________________ convolution2d_3 (Convolution2D)  (None, 1, 10, 64)     51264<br>\n  elu_2[0][0]<br>\n  ____________________________________________________________________________________________________ flatten_1 (Flatten)              (None, 640)           0<br>\n  convolution2d_3[0][0]<br>\n  ____________________________________________________________________________________________________ dropout_1 (Dropout)              (None, 640)           0<br>\n  flatten_1[0][0]<br>\n  ____________________________________________________________________________________________________ elu_3 (ELU)                      (None, 640)           0<br>\n  dropout_1[0][0]<br>\n  ____________________________________________________________________________________________________ dense_1 (Dense)                  (None, 512)           328192<br>\n  elu_3[0][0]<br>\n  ____________________________________________________________________________________________________ dropout_2 (Dropout)              (None, 512)           0<br>\n  dense_1[0][0]<br>\n  ____________________________________________________________________________________________________ elu_4 (ELU)                      (None, 512)           0<br>\n  dropout_2[0][0]<br>\n  ____________________________________________________________________________________________________ dense_2 (Dense)                  (None, 1)             513<br>\n  elu_4[0][0]<br>\n  ==================================================================================================== Total params: 720,497 Trainable params: 720,497 Non-trainable params:\n  0\n  ____________________________________________________________________________________________________ None<\/p>\n<\/blockquote>\n\n<p>throws the exception<\/p>\n\n<blockquote>\n  <p>'rawunicodeescape' codec can't decode bytes in position 94-98:\n  truncated \\uXXXX<\/p>\n<\/blockquote>\n\n<p>What could be the problem and how can I solve it? <\/p>\n","answers":[{"AnswerId":"41952183","CreationDate":"2017-01-31T07:44:12.707","ParentId":null,"OwnerUserId":"2178438","Title":null,"Body":"<p>I came across a similar problem when using keras 1.2.1 with a tensorflow-gpu backend.<\/p>\n\n<p>I found out it was caused because windows 10 anniversary edition was having problems encoding the forward slash character.<\/p>\n\n<p>Using the Lambda layer makes the <code>to_json()<\/code> call fail but switching to batch normalization works just fine. <\/p>\n\n<pre><code>model = Sequential()\n\n# model.add(Lambda(lambda x: x \/ 255. - .5, input_shape=INPUT_DIMENSIONS))\nmodel.add(BatchNormalization(input_shape=INPUT_DIMENSIONS, axis=1))\n. . . \n# POST PROCESSING, SAVE MODEL TO DISK\nwith open('model.json', 'w') as json_file:\n    json_file.write(model.to_json())\n<\/code><\/pre>\n\n<p>Not an ideal solution but hopefully it works for someone looking at this in the future.<\/p>\n"},{"AnswerId":"41926741","CreationDate":"2017-01-29T22:36:14.013","ParentId":null,"OwnerUserId":"5107488","Title":null,"Body":"<p>It seems like your code is in directory like this: \"C:\\Users\\python\\u{...}.py\". This kind of error is related to python 3 version, where we get special character \\u and can't decode it on Windows machines. You can change filename or full path to file, so that it did not contain special characters or make a patch to function 'func_dump' from file <code>generic_utils.py<\/code> (it can be reached by following path 'keras\/utils\/generic_utils.py'). You should replace line <code>code = marshal.dumps(func.__code__).decode('raw_unicode_escape')<\/code> by line <code>code = marshal.dumps(func.__code__).replace(b'\\\\',b'\/').decode('raw_unicode_escape')<\/code>. <\/p>\n"},{"AnswerId":"47744822","CreationDate":"2017-12-11T00:02:53.650","ParentId":null,"OwnerUserId":"2422398","Title":null,"Body":"<p>I believe that this is actually due to a bug in Keras, and was thankfully fixed by <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/8572\" rel=\"nofollow noreferrer\">Keras pull request #8572<\/a>. However, that PR was only merged very recently, and while it's incorporated in the latest pypi release, it's <em>not<\/em> in the latest conda release, which is why I was getting this error.<\/p>\n\n<p>Installing Keras via <code>pip install keras<\/code> instead of <code>conda install keras<\/code> made everything work for me (though I also had to precede this with <code>conda install pip<\/code> to get pip to work properly).<\/p>\n"}]}
{"QuestionId":41848069,"AnswerCount":1,"Tags":"<tensorflow><windows-xp><anaconda>","CreationDate":"2017-01-25T09:38:25.750","AcceptedAnswerId":41856820.0,"OwnerUserId":4312306.0,"Title":"Is it possible to install TensorFlow using Anaconda 2.1 on Windows XP(32 bit)?","Body":"<p>I understand that initially TensorFlow did not work on Windows.\nI have a Windows XP system with 2GB RAM and Anaconda 2.1 installed on it. Is there anyway to make TensorFlow work on this?<\/p>\n","answers":[{"AnswerId":"41856820","CreationDate":"2017-01-25T16:33:46.663","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Currently there is no official <em>pre-built<\/em> TensorFlow distribution that works on 32-bit Windows, and it is unlikely that we will ever support such an old operating system, because the versions of software that TensorFlow depends on (such as Python 3.5) do not support Windows XP.<\/p>\n\n<p>The existing TensorFlow packages for Windows require a 64-bit installation of Python 3.5 (which may be Anaconda, or the Python.org distribution), and <a href=\"https:\/\/docs.python.org\/3\/using\/windows.html\" rel=\"nofollow noreferrer\">Python 3.5 requires Windows Vista or newer<\/a>. Python on Windows <a href=\"https:\/\/wiki.python.org\/moin\/WindowsCompilers\" rel=\"nofollow noreferrer\">restricts what compilers may be used<\/a> to build extensions for each version of Python, and only Python 3.5 supports a new enough compiler (Visual C++ 14.0, from Visual Studio 2015) to build TensorFlow.<\/p>\n"}]}
{"QuestionId":41848315,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-01-25T09:50:46.977","AcceptedAnswerId":41848744.0,"OwnerUserId":1934212.0,"Title":"Problems understanding keras logging output","Body":"<p>Training my model model with <\/p>\n\n<pre><\/pre>\n\n<p>generates the following output: <\/p>\n\n<pre><\/pre>\n\n<p>what does <\/p>\n\n<pre><\/pre>\n\n<p>mean? It certainly has to do with the batch_size of 64, but what do these specific output lines represent and why are they printed? <\/p>\n","answers":[{"AnswerId":"41848744","CreationDate":"2017-01-25T10:09:52.650","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>In case when you set the <code>verbose=1<\/code> option - Keras is printing the information about <code>loss<\/code>, estimated epoch time remaining and a metrics provided during model compilation after each batch. Usually it's printed dynamically - so you could only see a one bar for each epoch - but for some reason your <code>Python<\/code> is printing it new line batch after batch.<\/p>\n"}]}
{"QuestionId":41848660,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-25T10:06:33.103","AcceptedAnswerId":41848962.0,"OwnerUserId":562769.0,"Title":"Why the negative reshape (-1) in MNIST tutorial?","Body":"<p>Reading <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/\" rel=\"noreferrer\">the Tensorflow MNIST tutorial<\/a>, I stumbled over the line<\/p>\n\n<pre><\/pre>\n\n<p> comes from ,  comes from the number of channels. But why ?<\/p>\n\n<p>I guess this is related to mini-batch training, but I wondered why  and not  (which seems to give the same result in numpy).\n(Probably related: Why does the reshape of numpy give the same results for , and )?<\/p>\n","answers":[{"AnswerId":"50256112","CreationDate":"2018-05-09T14:41:29.180","ParentId":null,"OwnerUserId":"8210067","Title":null,"Body":"<p>In the MNIST tutorial that you are reading, the desired shape for your input layer : [batch_size, 28, 28, 1]<\/p>\n\n<p>x_image = tf.reshape(x, [-1,28,28,1])<\/p>\n\n<p>Here -1 for input x specifies that this dimension should be dynamically computed based on the number of input values in x, holding the size of all other dimensions constant. This allows us to treat batch_size(parameter with value -1) as a hyperparameter that we can tune.<\/p>\n"},{"AnswerId":"41848962","CreationDate":"2017-01-25T10:19:59.453","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p><code>-1<\/code> means that the length in that dimension is inferred. This is done based on the constraint that the number of elements in an <code>ndarray<\/code> or <code>Tensor<\/code> when reshaped must remain the same. In the tutorial, each image is a row vector (784 elements) and there are lots of such rows (let it be <code>n<\/code>, so there are <code>784n<\/code> elements). So, when you write<\/p>\n\n<p><code>x_image = tf.reshape(x, [-1, 28, 28, 1])<\/code> <\/p>\n\n<p>TensorFlow can infer that <code>-1<\/code> is <code>n<\/code>.<\/p>\n"}]}
{"QuestionId":41848885,"AnswerCount":1,"Tags":"<python><random><tensorflow>","CreationDate":"2017-01-25T10:16:13.753","AcceptedAnswerId":41863257.0,"OwnerUserId":7457101.0,"Title":"unable to copy random generated values in to variable using tensorflow","Body":"<p>I have data in one array say size [1x9] , I am generating random number 1 to 9 and shuffling it, I want to arrange data in that order. <\/p>\n\n<pre><\/pre>\n\n<p>It because I am not able to copy rn value as constant , when I am running sess.run every time it is changing. \nBut I need the random values generated in 'rn' first time generated as i need for testing on another ones. \nHow many times i print rn it should show the same values with out regenerating again.\nHow to do it ? <\/p>\n\n<p>I tried by importing random <\/p>\n\n<pre><\/pre>\n\n<p>but it gives \n`<\/p>\n","answers":[{"AnswerId":"41863257","CreationDate":"2017-01-25T22:59:48.207","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/constant_op\/random_tensors#random_shuffle\" rel=\"nofollow noreferrer\"><code>tf.random_shuffle()<\/code><\/a> op (and in general the other <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/constant_op\/random_tensors\" rel=\"nofollow noreferrer\"><code>tf.random_*()<\/code><\/a> ops) will generate new random values on each call to <code>sess.run()<\/code>. If you want to capture a particular value for a random tensor and use it in multiple calls to <code>sess.run()<\/code>, you should assign it to a <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops\/variables#Variable\" rel=\"nofollow noreferrer\"><code>tf.Variable<\/code><\/a>. For example, you could restructure your program as follows to solve the problem:<\/p>\n\n<pre><code># generating an array of number\nBCI = tf.constant([1, 1, 1, 2, 2, 2, 3, 3, 3])\n\n# generating random numbers with length of BCI and shuffling it\nrn = tf.Variable(tf.random_shuffle(tf.range(9)))\n\nrna = tf.cast(rn,tf.int32)\n\n# rearranging data \nBCI2 = tf.gather(BCI1, rna)\n\nsess.run(tf.global_variables_initializer())\nprint(sess.run(BCI1))  # ==&gt; '[1 1 1 2 2 2 3 3 3]'\nprint(sess.run(rn))    # ==&gt; '[2 8 3 0 1 4 6 5 7]'\nprint(sess.run(BCI2))  # ==&gt; '[1 3 2 1 1 2 3 2 3]'\nprint(sess.run(BCI2))  # ==&gt; '[1 3 2 1 1 2 3 2 3]'\n<\/code><\/pre>\n"}]}
{"QuestionId":41848963,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-25T10:20:00.807","AcceptedAnswerId":41849545.0,"OwnerUserId":1876627.0,"Title":"stop_forward_pass in Tensorflow","Body":"<p>let's imagine that Tensorflow graph is consist of two parallel branches that have a common start and common end. In some point, I want to tell Tensorflow to stop disable computations on one branch. In other words stop forward and backward passes.  For the backward pass we have  stop_gradient op. But what for the forward pass, how to disable it? If this is not implemented now, please give me hints how can I create such op? <\/p>\n","answers":[{"AnswerId":"41849545","CreationDate":"2017-01-25T10:46:27.107","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>From the discussion we had you can do something like this:<\/p>\n\n<pre><code>out = f(branch1, tf.cond(should_disable, lambda: last_branch2_value, lambda: branch2))\n<\/code><\/pre>\n\n<p>Here <code>should_disbale<\/code> is a <code>bool<\/code> that determines if <code>branch2<\/code> is computed or not. Note that <code>tf.cond()<\/code> is \"differentiable\" meaning that TensorFlow is smart enough to not do backward pass through <code>branch2<\/code> if <code>should_disbale<\/code> is <code>True<\/code><\/p>\n\n<p>See the docs for <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard0\/tf.cond.md\" rel=\"nofollow noreferrer\">tf.cond <\/a> for more details.<\/p>\n"}]}
{"QuestionId":41849496,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-25T10:44:21.693","AcceptedAnswerId":null,"OwnerUserId":7251162.0,"Title":"Multi layer loss tensorflow","Body":"<p>I am trying to build a multi layer loss\nI am using AlexNet as my base network and I have 4 classes that have 3 possible labels, so I tried to build it as follow:<\/p>\n\n<pre><\/pre>\n\n<p>this is not my real output but this is his size,meaning the output layer on the alexNet is of size [4,3]\nI want to be able to view only the output that is relevant to the class I put as an input.so in the end for each image i will get [1,3] size output that came from the relevant  part of  the original output\nfor example<\/p>\n\n<pre><\/pre>\n\n<p>i will get<\/p>\n\n<pre><\/pre>\n\n<p>how can I use the labels and the output to get new_output\nI tried to use mask and I failed\ncan you help me?  <\/p>\n","answers":[{"AnswerId":"41863663","CreationDate":"2017-01-25T23:35:46.520","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>Please try following:\n<code>tf.gather(output, input)<\/code><\/p>\n"}]}
{"QuestionId":41849559,"AnswerCount":1,"Tags":"<python><tensorflow><tiling>","CreationDate":"2017-01-25T10:46:56.413","AcceptedAnswerId":41851449.0,"OwnerUserId":2169557.0,"Title":"Removing or masking the I'th sub-element of a tilted tensor?","Body":"<p>I am searching for a way using Tensorflow to extract all of the sub-elements except for the one corresponding to the tensor's index.<\/p>\n\n<p>(eg. if looking at index 1 then only sub-elements 0 and 2 are present)<\/p>\n\n<p>Very similar to <a href=\"https:\/\/stackoverflow.com\/a\/41780121\/3961748\">this approach<\/a> using Numpy.<\/p>\n\n<p>Here's some example code to create a tiled tensor and a boolean mask: <\/p>\n\n<pre><\/pre>\n\n<p>Example output tensors: <\/p>\n\n<pre><\/pre>\n\n<p>Desired output:<\/p>\n\n<pre><\/pre>\n\n<p>I am also curious if there are more efficient methods for doing this as opposed to creating a large tensor and then masking it?<\/p>\n\n<p>Thanks! <\/p>\n","answers":[{"AnswerId":"41851449","CreationDate":"2017-01-25T12:17:35.323","ParentId":null,"OwnerUserId":"2169557","Title":null,"Body":"<p>Turns out Tensorflow had exactly what I was looking for already built in :)<\/p>\n\n<pre><code>result = tf.boolean_mask(tile, mask)\nresult = tf.reshape(result,  [n, n-1, -1])\n\n&gt;&gt;&gt; print(result)\n[[[ 2.  7.  0.]\n  [ 3.  7.  0.]]\n\n [[ 1.  7.  0.]\n  [ 3.  7.  0.]]\n\n [[ 1.  7.  0.]\n  [ 2.  7.  0.]]]\n<\/code><\/pre>\n"}]}
{"QuestionId":41849567,"AnswerCount":1,"Tags":"<python><pandas><tensorflow>","CreationDate":"2017-01-25T10:47:21.350","AcceptedAnswerId":null,"OwnerUserId":7338784.0,"Title":"CSV chunks or off-line","Body":"<p>I want to generate a df from the chunks such as:<\/p>\n\n<pre><\/pre>\n\n<p>However the computer's RAM is only 3GB, and the CSV file is more than 3GB as below:<\/p>\n\n<pre><\/pre>\n\n<p>I want to generate a df from the file by chunks in order to seperate the data to train X,  test y on some fields, such as \"a-e\" to train X, \"g-h\" to test y after that. The seperate function's inputs type is df.\nCould you give some specific examples to generate a df from the chunks or a csv off-line? <\/p>\n","answers":[{"AnswerId":"41850376","CreationDate":"2017-01-25T11:23:58.783","ParentId":null,"OwnerUserId":"4496708","Title":null,"Body":"<p>You could try using <code>pandas.read_csv<\/code> (<a href=\"http:\/\/pandas.pydata.org\/pandas-docs\/stable\/generated\/pandas.read_csv.html#pandas.read_csv\" rel=\"nofollow noreferrer\">documentation<\/a>). What you're looking for is a <code>usecols<\/code> parameter, where you can specify, which columns you want to load, say: <code>['a', 'b', 'c', 'd']<\/code> for train set.<\/p>\n\n<p>Also specifying <code>chunksize<\/code> might help. Here's more info: <a href=\"https:\/\/pandas-docs.github.io\/pandas-docs-travis\/io.html#iterating-through-files-chunk-by-chunk\" rel=\"nofollow noreferrer\">https:\/\/pandas-docs.github.io\/pandas-docs-travis\/io.html#iterating-through-files-chunk-by-chunk<\/a><\/p>\n"}]}
{"QuestionId":41850432,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><entropy>","CreationDate":"2017-01-25T11:26:37.387","AcceptedAnswerId":null,"OwnerUserId":6182430.0,"Title":"Difference between two versions of cross_entropy calculations in TensorFlow","Body":"<p>I'm using TensorFlow in order to train binary classification neural network.<\/p>\n\n<p>In order to build the network, half-year ago I followed the tutorial in TensorFlow website - <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/\" rel=\"nofollow noreferrer\">Deep MNIST for Experts<\/a>.<\/p>\n\n<p>Today, when I comparing both codes (the one in the tutorial and the one I wrote), I can see a difference in the cross-entropy calculation. A difference that I can't tell why it's there.<\/p>\n\n<p>In the <strong>tutorial<\/strong> the cross-entropy calculated as follows:<\/p>\n\n<pre><\/pre>\n\n<p>While in <strong>my code<\/strong> the calculation is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I'm new in Tensorflow, and I feel that I'm missing something. Mabey the difference is between two versions of the TensorFlow tutorials? <strong>What is the actual difference between the two lines?<\/strong><\/p>\n\n<p>Really appreciate your help. Thanks!<\/p>\n\n<p>The relevant code from the tutorial:<\/p>\n\n<pre><\/pre>\n\n<p>My code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41851303","CreationDate":"2017-01-25T12:09:00.737","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The difference is small but quite significative.<\/p>\n\n<p>softmax_cross_entropy_with_logits takes logits (real numbers without any range limit), passes them through the softmax function and then computes the cross-entropy. Combining both into one function is to apply some optimizations to improve numerical accuracy.<\/p>\n\n<p>The second code just applies cross-entropy directly to y_conv, which seems to be the output of a softmax function. This is correct, and both should give similar but not the same results, softmax_cross_entropy_with_logits is superior because of numerical stability. Just remember to give it logits and not the output of a softmax.<\/p>\n"}]}
{"QuestionId":41850496,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><deep-learning><cntk>","CreationDate":"2017-01-25T11:29:34.117","AcceptedAnswerId":41851232.0,"OwnerUserId":7345564.0,"Title":"What is the best way of combining clinical and Image data in a Deep Learning approach?","Body":"<p>I have CT images from patients and applied a CNN to those images to predict diseases. I would like to combine my clinical data with my image data in a CNN approach, is that possible? My clinical data has information like age, sex, dates, smoker, all of them are numbers, like 1 for smoker and 0 not.<\/p>\n","answers":[{"AnswerId":"41850827","CreationDate":"2017-01-25T11:46:12.180","ParentId":null,"OwnerUserId":"5979993","Title":null,"Body":"<p>Have a look at, for example, <a href=\"http:\/\/www.kdd.org\/kdd2016\/subtopic\/view\/images-dont-lie-transferring-deep-visual-semantic-features-to-large-scale-m\" rel=\"noreferrer\">this paper<\/a> where they combine features from a CNN with text data. In that paper, the CNN is already pre-trained (i.e., the CNN is essentially a featurizer), but you could clearly learn all in one go. The idea would always be to <\/p>\n\n<ul>\n<li>Run the image in your input through the convolution\/subsampling layers<\/li>\n<li>Just before your final fully connected (decision) layer, concatenate the other features you have available<\/li>\n<li>Feed all (pre-processed image and other features) into the decision layer.<\/li>\n<\/ul>\n\n<p>So the answer is \"yes, certainly\", the details depend on which framework you are using.<\/p>\n"},{"AnswerId":"41851232","CreationDate":"2017-01-25T12:06:04.757","ParentId":null,"OwnerUserId":"7451758","Title":null,"Body":"<p>As far as I know CNN is extremely suitable for image data, but not for other data.<\/p>\n\n<p>A solution to your problem would be to 'color' your images with the clinical data. (In image recognition CNNs, usually an input image is split into 3 color layers: red, grey and blue. See: <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/\" rel=\"nofollow noreferrer\">http:\/\/cs231n.github.io\/convolutional-networks\/<\/a>)<\/p>\n\n<p>Let's say your input data is a 32x32 pixel 8-bit greyscale image (so 1 color layer). I propose to add each clinical data variable as a 'color' layer. All input values in the same color layer should be the same.<\/p>\n\n<p>Whether each layer should be the same size as the image, or if you can get away with a single pixel, I'm not sure, but at least you can treat the clinical data as an 'image' alongside the CT images.<\/p>\n"}]}
{"QuestionId":41851378,"AnswerCount":0,"Tags":"<arrays><multidimensional-array><tensorflow>","CreationDate":"2017-01-25T12:13:15.780","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Tensorflow, extending a tensor in a specified dimension","Body":"<p>I am using the scatter_update function and I run into a dimensionality problem. More precisely: <\/p>\n\n<pre><\/pre>\n\n<p>gives:<\/p>\n\n<pre><\/pre>\n\n<p>I'm guessing that I must expand v_rows to the shape [5,9] for this to work. Any idea on how I can extend it from [1,9] to [5,9] ? <\/p>\n","answers":[]}
{"QuestionId":41851619,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-25T12:25:16.437","AcceptedAnswerId":41854895.0,"OwnerUserId":5496253.0,"Title":"Applying a function element-wise to a tensor with different parameters in Tensorflow","Body":"<p>I have a list that consists of a mix of constants and variables that act as inputs for a function I am trying to optimize. Here is the complex part: I need to apply a function to each of the elements in that list each of them with different parameters. I believe that this must be done in the computation graph to be efficient.<\/p>\n\n<p>What I thought of doing was putting the parameters into tensors that match up element by element and applying the function to the three tensors such that each row of elements is evaluated together. Is this possible?<\/p>\n\n<p>Example:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/dSV13.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dSV13.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Could end up with something like:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you so much in advance.<\/p>\n","answers":[{"AnswerId":"41854895","CreationDate":"2017-01-25T15:05:28.457","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>Use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/functional_ops\/higher_order_operators#map_fn\" rel=\"nofollow noreferrer\">tf.map_fn<\/a>. <\/p>\n\n<pre><code>g = tf.map_fn(f, tf.transpose(tf.concat(0, [optimal, p1, p2])))\n<\/code><\/pre>\n\n<p>First concatenate along <code>dim 0<\/code>, then transpose it, so that row <code>i<\/code> is <code>o_i, p1_i, p2_i<\/code>. <code>map_fn<\/code> simply applies <code>f<\/code> to each row.<\/p>\n"}]}
{"QuestionId":41851657,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-25T12:27:13.010","AcceptedAnswerId":41867177.0,"OwnerUserId":5662769.0,"Title":"TensorFlow You must feed a value for placeholder tensor 'Placeholder_2' with dtype float","Body":"<p>my code that fails with the infamous:<\/p>\n\n<p>InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder_2' with dtype float\n     [[Node: Placeholder_2 = Placeholderdtype=DT_FLOAT, shape=[], _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41867177","CreationDate":"2017-01-26T04:59:54.890","ParentId":null,"OwnerUserId":"5662769","Title":null,"Body":"<p>I figured out the issue.<\/p>\n\n<pre><code>Logits = LeNet(x)\n<\/code><\/pre>\n\n<p>the definition of LeNet(x) uses a \"keep_prob\" variable which isn't being fed.<\/p>\n\n<p>Changing the code to:<\/p>\n\n<pre><code>output = sess.run(output, feed_dict={x: own_images, keep_prob:1.0})\n<\/code><\/pre>\n\n<p>solves the issue.<\/p>\n\n<p>However word of warning. If you try to comment out the keep_prob variable within the LeNet function definition, it may not fix the issue as you have to refresh the function definitions and calls in other cells also.<\/p>\n"}]}
{"QuestionId":41852436,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2017-01-25T13:07:36.840","AcceptedAnswerId":null,"OwnerUserId":965640.0,"Title":"Torch \/ cunn install error - file already has LC_RPATH","Body":"<p>I'm trying to install cunn for Torch but make install fails. Cuda is 7.5.27 because 8 doesn't support OSX 10.10 anymore.<\/p>\n\n<p>I've been trying to debug this for a while, but I'm running out of things to try.<\/p>\n\n<p>While searching solution I saw few others having a similar problem (<a href=\"https:\/\/github.com\/TRIQS\/cthyb\/issues\/8\" rel=\"nofollow noreferrer\">https:\/\/github.com\/TRIQS\/cthyb\/issues\/8<\/a>, <a href=\"https:\/\/github.com\/jcjohnson\/neural-style\/issues\/134\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jcjohnson\/neural-style\/issues\/134<\/a>) but didn't find any solution that would fix this. It seems that this is a general problem related to OSX\/Xcode and not necessarily cunn.<\/p>\n\n<p>All help is much appreciated.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41858155","CreationDate":"2017-01-25T17:43:15.790","ParentId":null,"OwnerUserId":"965640","Title":null,"Body":"<p>The issue went away when I did <code>.\/clean.sh<\/code> in <code>~\/torch<\/code> followed by <code>.\/update.sh<\/code>. I still don't have any idea what was the problem.<\/p>\n"}]}
{"QuestionId":41852601,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-25T13:16:05.603","AcceptedAnswerId":null,"OwnerUserId":4482655.0,"Title":"Batch training in Tensorflow Slim","Body":"<p>I am looking at <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/slim_walkthough.ipynb\" rel=\"nofollow noreferrer\">TF Slim<\/a> introductory document and from what I understand, it only takes in one batch of image data at each run(32 images). Obviously, one wants to loop through this and train for many different batches. The intro does not cover this. How can this be done properly. I imagine there should be some way to specify a load batch function which should be called automatically when starting a batch training event, but I can't seem to find a simple example for this on the intro. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41859080","CreationDate":"2017-01-25T18:36:21.640","ParentId":null,"OwnerUserId":"3821732","Title":null,"Body":"<p>The <code>slim.learning.train<\/code> function contains a training loop, so the code you've given does in fact train on multiple batches of images. <\/p>\n\n<p>See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/learning.py#L793\" rel=\"nofollow noreferrer\">here in the source code<\/a>, where <code>train_step_fn<\/code> is called within a while loop. <code>train_step<\/code> (the default value of <code>train_step_fn<\/code>) contains the line <code>sess.run([train_op, global_step]...)<\/code>, which actually runs the training operation on a single batch of images.<\/p>\n"}]}
{"QuestionId":41853542,"AnswerCount":1,"Tags":"<tensorflow><word2vec>","CreationDate":"2017-01-25T13:59:48.793","AcceptedAnswerId":null,"OwnerUserId":7383444.0,"Title":"How can I retrieve corresponding words of each embedding in word2vec.py (tensorflow)?","Body":"<p>What word2vec.py does is returns a tensor of shape (vocab_size,embedding_size). I want the corresponding word of each embedding,sorta like a dictionary :<\/p>\n\n<pre><\/pre>\n\n<p>How do I get it?<\/p>\n","answers":[{"AnswerId":"41948134","CreationDate":"2017-01-31T01:03:26.197","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>You can use word2vec._id2word to get index to id. <\/p>\n"}]}
{"QuestionId":41853704,"AnswerCount":1,"Tags":"<bash><caffe>","CreationDate":"2017-01-25T14:07:48.410","AcceptedAnswerId":null,"OwnerUserId":6786806.0,"Title":"Error in LeNet on MNIST example","Body":"<p>I followed the guidance under the Training LeNet on MNIST with Caffe but I've got a error when creating lmdb..<\/p>\n\n<pre><\/pre>\n\n<p>I'm using GIT Bash for run the shell script,\nso what is wrong?<\/p>\n","answers":[{"AnswerId":"41885233","CreationDate":"2017-01-27T00:10:27.913","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>It appears that the linker\/loader does not have access to that Python BOOST library.  Did you run the scripts needed to set your environment variables?  I suspect that there's a missing path in your LD_LIBRARY_PATH.<\/p>\n\n<p>As a stop-gap, you can locate the missing file with<\/p>\n\n<pre><code>find \/ -name boost_python-vc140-mt-1_61.dll\n<\/code><\/pre>\n\n<p>... and then manually add that path to LD_LIBRARY_PATH.  The problem is that you'll hit thesefiles\/caffe\/build\/examples\/mnist\/convert_mnist_data.exe one at a time until you get them all.<\/p>\n\n<p>You can also try<\/p>\n\n<pre><code>ldd files\/caffe\/build\/examples\/mnist\/convert_mnist_data.exe\n<\/code><\/pre>\n\n<p>... and see what you might be missing.<\/p>\n"}]}
{"QuestionId":41853767,"AnswerCount":2,"Tags":"<python><numpy><tensorflow><convolution>","CreationDate":"2017-01-25T14:11:05.350","AcceptedAnswerId":44093772.0,"OwnerUserId":7415539.0,"Title":"Difference between Tensorflow convolution and numpy convolution","Body":"<pre><\/pre>\n\n<p>I am trying to see the results of convolutions from Tensorflow to check if it is behaving as I intended.\nWhen I run the numpy convolution and compare it to the Tensorflow convolution, the answer is different.\nThe above code is how I ran the test.\nI was hoping the  and  would be equal.<\/p>\n\n<p>My final goal is the run 2D convolution on a matrix with a 1 dimensional filter that runs 1d-convolution on each row with the same filter. In order to make this work (which will be basically a loop of 1d convolution over the rows) I need to figure out why my  and  give me different answers.<\/p>\n","answers":[{"AnswerId":"44093772","CreationDate":"2017-05-21T06:16:30.983","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>The problem that you see is because TF does not really calculate the convolution. If you will take a look at the explanation of what <a href=\"https:\/\/en.wikipedia.org\/wiki\/Convolution\" rel=\"noreferrer\">convolution actually does<\/a> (check for <strong>Visual explanations of convolution<\/strong>), you will see that the second function is flipped:<\/p>\n\n<ol>\n<li>Express each function in terms of a dummy variable <\/li>\n<li>Reflect one of the functions <strong>(this is the flip)<\/strong><\/li>\n<li>..... Some other stuff which I will not copy here.<\/li>\n<\/ol>\n\n<p>TF does everything except of that flip. So all you need to do is to flip the kernel either in TF or in numpy. Flipping for 1d case is just kernel in a reverse order, for 2d you will need to flip both axis (rotate the kernel 2 times).<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nI = [1, 0, 2, 3, 0, 1, 1]\nK = [2, 1, 3]\n\ni = tf.constant(I, dtype=tf.float32, name='i')\nk = tf.constant(K, dtype=tf.float32, name='k')\n\ndata   = tf.reshape(i, [1, int(i.shape[0]), 1], name='data')\nkernel = tf.reshape(k, [int(k.shape[0]), 1, 1], name='kernel')\n\nres = tf.squeeze(tf.nn.conv1d(data, kernel, 1, 'VALID'))\nwith tf.Session() as sess:\n    print sess.run(res)\n    print np.convolve(I, K[::-1], 'VALID')\n<\/code><\/pre>\n"},{"AnswerId":"41858091","CreationDate":"2017-01-25T17:39:15.083","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>The order of filter is reversed. TensorFlow convolution is actually correlation. Numpy gets notation from math, TF gets notation from machine learning papers and somewhere the order got reversed.<\/p>\n\n<p>This prints <code>True<\/code><\/p>\n\n<pre><code>filter_np2=filter_np[::-1,0,0]\nnp.allclose(np.convolve(X[0,:,0],filter_np2,'SAME'),  Xconv_tf.flatten())    np.convolve(X[0,:,0],filter_np2,'SAME')\n<\/code><\/pre>\n"}]}
{"QuestionId":41853851,"AnswerCount":0,"Tags":"<python><web><tensorflow>","CreationDate":"2017-01-25T14:15:51.257","AcceptedAnswerId":null,"OwnerUserId":7160913.0,"Title":"Connection of Web Interface with Tensorflow","Body":"<p>I am currently working on a project where I get input data in a Web Interface and send it to a python script for processing.<\/p>\n\n<p>The python script opens a session, restores the desired TF model and process the data.<\/p>\n\n<p>This way, I have to restore the model every time the user sends new data. To restore the model takes though 1:40 min , which is a problem for me. <\/p>\n\n<p>The solution I came up is to use two python script, one that creates a session and set it as default, then loads the model in that session and another script that gets the default session and default graph and processes the data. That way the model will be loaded it once and I will be able to use it for all user requests. The code looks smth like below..<\/p>\n\n<pre><\/pre>\n\n<p>This solution did not work, The information for the loaded model is lost. Is like having a for inside another for and then split it, normal that it does not work! <\/p>\n\n<p>Can you suggest me another way to keep this model loaded always, so as I can avoid this delay every time a user wants to process smth new?<\/p>\n\n<p>Thank you!<\/p>\n","answers":[]}
{"QuestionId":41854012,"AnswerCount":1,"Tags":"<tensorflow><syntaxnet>","CreationDate":"2017-01-25T14:23:39.620","AcceptedAnswerId":41876936.0,"OwnerUserId":2161903.0,"Title":"How to interpret the output of a synaxnet when annotating a corpus","Body":"<p>I annotated a corpus using pre-trained syntaxnet model (i.e. using Parse McParseface). I am having a problem understanding the output. There are two metrics reproted in the output. Are those for POS tagging and dependency parsing? If yes, which one is POS tagging performance and which one is for dependency parsing performance?<\/p>\n\n<p>Here is the output:<\/p>\n\n<p><\/p>\n","answers":[{"AnswerId":"41876936","CreationDate":"2017-01-26T15:27:03.440","ParentId":null,"OwnerUserId":"6614832","Title":null,"Body":"<p>If you're using\n<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/syntaxnet\/syntaxnet\/demo.sh\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/syntaxnet\/syntaxnet\/demo.sh<\/a>\nthen the first metric is POS tag accuracy, the second UAS. They are only meaningful if the conll data you input contains gold POS tags and gold dependencies.<\/p>\n"}]}
{"QuestionId":41854964,"AnswerCount":1,"Tags":"<arrays><multidimensional-array><indexing><tensorflow>","CreationDate":"2017-01-25T15:08:37.837","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Applying predefined function to batch in tensorflow","Body":"<p>I have a function, written in tensorflow, which manipulates arrays. It takes in 2d arrays and performs some operations on a part of the array, then outputs an array of the same size and shape, but with modified elements. My problem is that, in my model, I don't get just one array at a time, I get a batch of shape (batch_size, array_size, channel). Since the channel is usually mono, the input batch is a 4D tensor. \nIs there any way in which I can apply my function to every layer (matrix) in this 4D tensor, without significantly modifying the function? <\/p>\n\n<p>The function is something like:<\/p>\n\n<pre><\/pre>\n\n<p>How could I apply this to a batch of 5 such matrices? <\/p>\n","answers":[{"AnswerId":"41855519","CreationDate":"2017-01-25T15:34:00.580","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/functional_ops\/higher_order_operators#map_fn\" rel=\"nofollow noreferrer\">tf.map_fn<\/a><\/p>\n\n<p>It applies <code>f<\/code> to each element, in your case to each 2D array. These 2D arrays are obtained by unpacking the second parameter.<\/p>\n"}]}
{"QuestionId":41855455,"AnswerCount":0,"Tags":"<c++><windows><deep-learning><caffe><lmdb>","CreationDate":"2017-01-25T15:30:57.403","AcceptedAnswerId":null,"OwnerUserId":3246146.0,"Title":"Caffe Stucking in infinite loop of blocking queue while using LMDB in windows cpp?","Body":"<p>I am trying to train and test the simple network for segmentation. I have created the LMDB file using windows cpp.<\/p>\n\n<p>But while doing training in windows cpp it get stuck in <\/p>\n\n<pre><\/pre>\n\n<p>after Debugging I found it is stucking in a infinte loop of blocking queue. the flow is as follows.  <\/p>\n\n<blockquote>\n  <p>layers_[layer_id]->SetUp(bottom_vecs_[layer_id],\n  top_vecs_[layer_id])    -->    DataLayerSetUp(bottom, top)    ---->  Datum&amp;\n  datum = *(reader_.full().peek())     ---->  Blocking Queue<\/p>\n<\/blockquote>\n\n<p>It is displaying this Logs in executable window at runtime.<\/p>\n\n<pre><\/pre>\n\n<p>***********  mycaffe.cpp  ***************<\/p>\n\n<pre><\/pre>\n\n<p>*****************  mymodel.prototxt  ********************<\/p>\n\n<pre><\/pre>\n\n<p>*******************  mysolver.prototxt  ************<\/p>\n\n<pre><\/pre>\n\n<p>Please help me to figure out the mistake. Thank you.<\/p>\n","answers":[]}
{"QuestionId":41855512,"AnswerCount":4,"Tags":"<python><machine-learning><tensorflow><neural-network><keras>","CreationDate":"2017-01-25T15:33:34.563","AcceptedAnswerId":41855937.0,"OwnerUserId":2221270.0,"Title":"How does data normalization work in keras during prediction?","Body":"<p>I see that the imageDataGenerator allows me to specify different styles of data normalization, e.g. featurewise_center, samplewise_center, etc.<\/p>\n\n<p>I see from the examples that if I specify one of these options, then I need to call the fit method on the generator in order to allow the generator to compute statistics like the mean image on the generator.<\/p>\n\n<pre><\/pre>\n\n<p>My question is, how does prediction work if I have specified data normalization during training? I can't see how in the framework I would even pass knowledge of the training set mean\/std deviation along to predict to allow me to normalize my test data myself, but I also don't see in the training code where this information is stored.<\/p>\n\n<p>Are the image statistics needed for normalization stored in the model so that they can be used during prediction?<\/p>\n","answers":[{"AnswerId":"42945707","CreationDate":"2017-03-22T07:54:22.607","ParentId":null,"OwnerUserId":"448357","Title":null,"Body":"<p>I also had the same issue and I solved it using the same functionality, that the <code>ImageDataGenerator<\/code> used:<\/p>\n\n<pre><code># Load Cifar-10 dataset\n(trainX, trainY), (testX, testY) = cifar10.load_data()\ngenerator = ImageDataGenerator(featurewise_center=True, \n                               featurewise_std_normalization=True)\n\n# Calculate statistics on train dataset\ngenerator.fit(trainX)\n# Apply featurewise_center to test-data with statistics from train data\ntestX -= generator.mean\n# Apply featurewise_std_normalization to test-data with statistics from train data\ntestX \/= (generator.std + K.epsilon())\n\n# Do your regular fitting\nmodel.fit_generator(..., validation_data=(testX, testY), ...)\n<\/code><\/pre>\n\n<p>Note that this is only possible if you have a reasonable small dataset, like CIFAR-10. Otherwise the <a href=\"https:\/\/stackoverflow.com\/a\/41855937\/448357\">solution proposed by Marcin<\/a> sounds good more reasonable.<\/p>\n"},{"AnswerId":"43069409","CreationDate":"2017-03-28T12:20:32.597","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>Use the <code>standardize<\/code> method of the generator for each element. Here is a complete example for CIFAR 10:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n\nimport keras\nfrom keras.datasets import cifar10\nfrom keras.preprocessing.image import ImageDataGenerator\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Flatten\nfrom keras.layers import Conv2D, MaxPooling2D\n\n# input image dimensions\nimg_rows, img_cols, img_channels = 32, 32, 3\nnum_classes = 10\n\nbatch_size = 32\nepochs = 1\n\n# The data, shuffled and split between train and test sets:\n(x_train, y_train), (x_test, y_test) = cifar10.load_data()\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()\n\nmodel.add(Conv2D(32, (3, 3), padding='same', activation='relu',\n                 input_shape=x_train.shape[1:]))\nmodel.add(Conv2D(32, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Conv2D(64, (3, 3), padding='same', activation='relu'))\nmodel.add(Conv2D(64, (3, 3), activation='relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\n\nmodel.add(Flatten())\nmodel.add(Dense(512, activation='relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(num_classes, activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy', optimizer='rmsprop',\n              metrics=['accuracy'])\n\nx_train = x_train.astype('float32')\nx_test = x_test.astype('float32')\nx_train \/= 255\nx_test \/= 255\n\ndatagen = ImageDataGenerator(zca_whitening=True)\n\n# Compute principal components required for ZCA\ndatagen.fit(x_train)\n\n# Apply normalization (ZCA and others)\nprint(x_test.shape)\nfor i in range(len(x_test)):\n    # this is what you are looking for\n    x_test[i] = datagen.standardize(x_test[i])\nprint(x_test.shape)\n\n# Fit the model on the batches generated by datagen.flow().\nmodel.fit_generator(datagen.flow(x_train, y_train,\n                                 batch_size=batch_size),\n                    steps_per_epoch=x_train.shape[0] \/\/ batch_size,\n                    epochs=epochs,\n                    validation_data=(x_test, y_test))\n<\/code><\/pre>\n"},{"AnswerId":"43930523","CreationDate":"2017-05-12T06:14:28.883","ParentId":null,"OwnerUserId":"3371265","Title":null,"Body":"<p>I am using the <code>datagen.fit<\/code> function itself. <\/p>\n\n<pre><code>from keras.preprocessing.image import ImageDataGenerator\n\ntrain_datagen = ImageDataGenerator(\n    featurewise_center=True,\n    featurewise_std_normalization=True)\ntrain_datagen.fit(train_data)\n\ntest_datagen = ImageDataGenerator(  \n    featurewise_center=True, \n    featurewise_std_normalization=True)\ntest_datagen.fit(train_data)\n<\/code><\/pre>\n\n<p>Ideally with this,  <code>test_datagen<\/code> fitted on training dataset will learn the training datasets statistics. Then it will use these statistics to normalize testing data.<\/p>\n"},{"AnswerId":"41855937","CreationDate":"2017-01-25T15:51:48.420","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Yes - this is a really huge downside of <code>Keras.ImageDataGenerator<\/code> that you couldn't provide the standarization statistics on your own. But - there is an easy method on how to overcome this issue.<\/p>\n\n<p>Assuming that you have a function <code>normalize(x)<\/code> which is normalizing an image <strong>batch<\/strong> (remember that generator is not providing a simple image but an array of images - a <strong>batch<\/strong> with shape <code>(nr_of_examples_in_batch, image_dims ..)<\/code> you could make your own generator with normalization by using:<\/p>\n\n<pre><code>def gen_with_norm(gen, normalize):\n    for x, y in gen:\n        yield normalize(x), y\n<\/code><\/pre>\n\n<p>Then you might simply use <code>gen_with_norm(datagen.flow, normalize)<\/code> instead of <code>datagen.flow<\/code>.<\/p>\n\n<p>Moreover - you might recover the <code>mean<\/code> and <code>std<\/code> computed by a <code>fit<\/code> method by getting it from appropriate fields in datagen (e.g. <code>datagen.mean<\/code> and <code>datagen.std<\/code>).<\/p>\n"}]}
{"QuestionId":41856373,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-25T16:11:54.467","AcceptedAnswerId":41856781.0,"OwnerUserId":1660762.0,"Title":"index() function in tensorflow?","Body":"<p>Is there a way in tensorflow to get the index of a value in the tensor?<\/p>\n\n<p>e.g. I have a one-hot matrix, and I want to get the coordinates of the 1.<\/p>\n\n<pre><\/pre>\n\n<p>preferably this would be done with a function, e.g. <\/p>\n","answers":[{"AnswerId":"41856781","CreationDate":"2017-01-25T16:32:00.860","ParentId":null,"OwnerUserId":"6408452","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/api_docs\/python\/math_ops\/sequence_comparison_and_indexing#where\" rel=\"nofollow noreferrer\"><code>tf.where<\/code><\/a>.<\/p>\n\n<p>For example,<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.constant([[0, 0, 0],\n                 [0, 1, 0],\n                 [0, 3, 0]])\n\nwith tf.Session() as sess:\n    coordinates = tf.where(tf.greater(x, 0))\n    print(coordinates.eval()) # [[1 1], [2 1]]\n    print(tf.gather_nd(x, coordinates).eval()) # [1, 3]\n<\/code><\/pre>\n"}]}
{"QuestionId":41856601,"AnswerCount":1,"Tags":"<python><cygwin><anaconda><theano><keras>","CreationDate":"2017-01-25T16:23:15.910","AcceptedAnswerId":null,"OwnerUserId":7469788.0,"Title":"Installing Theano and Keras on my Windows 10 Workstation","Body":"<p>I am trying to use Keras to develop a Neural Network in Python, after managing to install on my Windows 10 Workstation Anaconda3 (with all its libraries: numpy, scikit-learn, pandas, SciPy and matplotlib), I realized to need TensorFlow or Theano, too.<\/p>\n\n<p>After I failed intalling TensorFlow, I downloaded and was able to install Theano, but trying to import it from the Python prompt, I received the following:<\/p>\n\n<blockquote>\n  <p><\/p>\n<\/blockquote>\n\n<p>Hoping in this way to solve the problem, I downloaded the GNU compiler for C++ Cygwin64, but nothing has changed, at all! Acknowledge that this is really the right way to move forward, how should I access the \"Theano flags cxx\"?<\/p>\n","answers":[{"AnswerId":"41860919","CreationDate":"2017-01-25T20:18:21.103","ParentId":null,"OwnerUserId":"7438048","Title":null,"Body":"<p>first, its only performance issue to run theano without g++. it a warning and not exception when importing it.<\/p>\n\n<p>BUT probably you want performance when using deep learning lib like keras so lets try fix the theano installation.<\/p>\n\n<p>please follow the <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html\" rel=\"nofollow noreferrer\">theano docs<\/a> about installing theano on windows. you might want to clean previous installation of requirements.<\/p>\n\n<p>to install the gcc follow <a href=\"http:\/\/deeplearning.net\/software\/theano\/install_windows.html#gcc\" rel=\"nofollow noreferrer\">this section<\/a> which says:<\/p>\n\n<blockquote>\n  <p>Theano C code compiler currently requires a GCC installation. We have\n  used the <strong>build TDM GCC<\/strong> which is provided for both 32- and 64-bit\n  platforms...<\/p>\n<\/blockquote>\n\n<p>download from <a href=\"http:\/\/tdm-gcc.tdragon.net\/\" rel=\"nofollow noreferrer\">here<\/a> follow the installation instruction.<\/p>\n\n<h2>Tensorflow<\/h2>\n\n<p>I recommending working with <code>tensorflow<\/code> as <code>keras<\/code> recently changed the default backend from <code>theano<\/code> to tensorflow. <\/p>\n\n<p>using <code>anaconda<\/code> and <code>pip<\/code> you should easily do <code>pip install tensorflow<\/code> and it will work.<\/p>\n\n<p>actually today I just installed <code>keras<\/code> and <code>tensorflow<\/code> on windows 10 using anaconda by just running <code>pip install keras tensorflow<\/code> so I suggest you try fresh clean installation of anaconda and python and try this again.<\/p>\n\n<p>please update if you succeed or having another issues installing theano \/ tensorflow \/ keras<\/p>\n"}]}
{"QuestionId":41856749,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-01-25T16:30:37.367","AcceptedAnswerId":41954139.0,"OwnerUserId":2191652.0,"Title":"Tensorflow: How to access Variables from a model imported from Caffe?","Body":"<p>I'm using the <a href=\"https:\/\/github.com\/DrSleep\/tensorflow-deeplab-resnet\" rel=\"nofollow noreferrer\">tensorflow-deeplab-resnet model<\/a> which transfers the Resnet model implemented in Caffe to tensorflow using <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\" rel=\"nofollow noreferrer\">caffe-tensorflow<\/a>.<\/p>\n\n<p>I'd like to know how I can access individual variables from the model that was imported from Caffe so I can check what is going wrong.<\/p>\n\n<p>I tried<\/p>\n\n<pre><\/pre>\n\n<p>which outputs<\/p>\n\n<pre><\/pre>\n\n<p>The  layers are the interesting layers that need to be reinitialized randomly. But when I try to access them and add a logging to the variable like this  <\/p>\n\n<pre><\/pre>\n\n<p>I can't see that variable in tensorboard. The only thing that is displayed in tensorboard is the graph itself.  <\/p>\n\n<p>How can I access these variables in order to monitor them in tensorboard?<br>\nCan I infer the correct names of the variables that I'd like to monitor by just looking at the graph (see picture)?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/4TdTT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/4TdTT.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>Edit<\/strong><br>\nI edited the focus of my question a little since there was a bug which has been fixed by now by the author of the code.<\/p>\n","answers":[{"AnswerId":"41857638","CreationDate":"2017-01-25T17:15:24.453","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>To clarify my understanding: You're trying to restore some part of the model from an old version, and initialize the rest randomly.<\/p>\n\n<p>If that's true, you can use <code>tf.contrib.framework.init_from_checkpoint<\/code> to initialize body of your model from the old checkpoint. The rest of the model (output layer) should be initialized randomly based on how you created it.<\/p>\n"},{"AnswerId":"41954139","CreationDate":"2017-01-31T09:41:24.293","ParentId":null,"OwnerUserId":"2191652","Title":null,"Body":"<p>It seems that it's acutally working in the way described in the question. I just needed to completely shutdown tensorboard and restart tensorboard for every new log file that I created. <\/p>\n"}]}
{"QuestionId":41857498,"AnswerCount":1,"Tags":"<python><numpy><scipy><theano>","CreationDate":"2017-01-25T17:08:25.243","AcceptedAnswerId":41864626.0,"OwnerUserId":4567933.0,"Title":"Scipy convolve2d with subsampling like Theano's conv2d?","Body":"<p>I wish to perform 2D convolution on images of size 600 X 400 using a 10 X 10 filter. The filter is not separable.  works well for me currently but, I am expecting a lot bigger images soon. <\/p>\n\n<p>To counter that, I have two ideas <\/p>\n\n<ol>\n<li>resizing images<\/li>\n<li>subsampling (or striding)?<\/li>\n<\/ol>\n\n<p>Focusing on the subsampling part, theano has a function which does convolution the same way as scipy convolve2d, see <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/signal\/conv.html#theano.tensor.signal.conv.conv2d\" rel=\"nofollow noreferrer\">theano conv2d<\/a><\/p>\n\n<p>It also has the subsampling option too. But, installing theano on windows has been painful to me. How do I get subsampling work with ? Any other alternatives (which doesn't require me installing me some heavyweight library)?<\/p>\n","answers":[{"AnswerId":"41864626","CreationDate":"2017-01-26T01:33:12.573","ParentId":null,"OwnerUserId":"7207392","Title":null,"Body":"<p>You could implement subsampling by hand, I'll only sketch 1d for simplicity. Say you want to sample s = d * f on a regular subgrid with spacing k. Then your nth sample is s_nk = sum_i=0^10 f_i d_nk-i. The thing to observe here is that the indices of f and d always sum to a multiple of k. This suggests splitting it up into sub-sums s_nk = sum_j=0^k-1 sum_i=0^10\/k f_j+ik d_-j+(n-i)k. So what you need to do is: subsample d and f at grids with spacing k at all offsets 0, ..., k-1. Convolve all pairs of subsampled d and f whose offsets sum to 0 or k and add the results.<\/p>\n\n<p>Here's some code for 1d. It roughly implements the above, only the grids are placed slightly differently to make index management easier. The second function does it the stupid way, i.e. computes the full convolution and then decimates. It is for testing the first function against.<\/p>\n\n<pre><code>import numpy as np\nfrom scipy import signal\n\ndef ss_conv(d1, d2, decimate):\n    n = (len(d1) + len(d2) - 1) \/\/ decimate\n    out = np.zeros((n,))\n    for i in range(decimate):\n        d1d = d1[i::decimate]\n        d2d = d2[decimate-i-1::decimate]\n        cv = signal.convolve(d1d, d2d, 'full')\n        out[:len(cv)] += cv\n    return out\n\ndef conv_ss(d1, d2, decimate):\n    return signal.convolve(d1, d2, 'full')[decimate-1::decimate]\n<\/code><\/pre>\n\n<p>Edit: 2d version:<\/p>\n\n<pre><code>import numpy as np\nfrom scipy import signal\n\ndef ss_conv_2d(d1, d2, decy, decx):\n    ny = (d1.shape[0] + d2.shape[0] - 1) \/\/ decy\n    nx = (d1.shape[1] + d2.shape[1] - 1) \/\/ decx\n    out = np.zeros((ny, nx))\n    for i in range(decy):\n        for j in range(decx):\n            d1d = d1[i::decy, j::decx]\n            d2d = d2[decy-i-1::decy, decx-j-1::decx]\n            cv = signal.convolve2d(d1d, d2d, 'full')\n            out[:cv.shape[0], :cv.shape[1]] += cv\n    return out\n\ndef conv_ss_2d(d1, d2, decy, decx):\n    return signal.convolve2d(d1, d2, 'full')[decy-1::decy, decx-1::decx]\n<\/code><\/pre>\n"}]}
{"QuestionId":41857602,"AnswerCount":1,"Tags":"<python><arrays><tensorflow>","CreationDate":"2017-01-25T17:13:22.550","AcceptedAnswerId":41859635.0,"OwnerUserId":7457101.0,"Title":"sorting of 2d array min to max in tensorflow","Body":"<p>I have an array <\/p>\n\n<pre><\/pre>\n\n<p>I want to sort the elements in each row from min to max. Is there any function for doing such ?<\/p>\n\n<p>In the example here they are using <a href=\"https:\/\/stackoverflow.com\/questions\/40784941\/sorting-an-array-in-tensorflow\">2d array<\/a>,using this I can loop to create the max to min. <\/p>\n\n<pre><\/pre>\n\n<p>Is there any thing similar for finding the min to max or how to reverse the array in each row ?<\/p>\n","answers":[{"AnswerId":"41859635","CreationDate":"2017-01-25T19:06:40.790","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>As suggested by @Yaroslav you can just use the <code>top_k<\/code> values.<\/p>\n\n<pre><code>a = tf.Variable([[0.51, 0.52, 0.53, 0.94, 0.35],\n             [0.32, 0.72, 0.83, 0.74, 0.55],\n             [0.23, 0.72, 0.63, 0.64, 0.35],\n             [0.11, 0.02, 0.03, 0.14, 0.15],\n             [0.01, 0.72, 0.73, 0.04, 0.75]],tf.float32)\n\nrow_size = a.get_shape().as_list()[-1]\ntop_k = tf.nn.top_k(-a, k=row_size)\nsess.run(-top_k.values)\n<\/code><\/pre>\n\n<p>this prints for me<\/p>\n\n<pre><code>array([[ 0.34999999,  0.50999999,  0.51999998,  0.52999997,  0.94      ],\n       [ 0.31999999,  0.55000001,  0.72000003,  0.74000001,  0.82999998],\n       [ 0.23      ,  0.34999999,  0.63      ,  0.63999999,  0.72000003],\n       [ 0.02      ,  0.03      ,  0.11      ,  0.14      ,  0.15000001],\n       [ 0.01      ,  0.04      ,  0.72000003,  0.73000002,  0.75      ]], dtype=float32)\n<\/code><\/pre>\n"}]}
{"QuestionId":41858327,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2017-01-25T17:53:34.057","AcceptedAnswerId":null,"OwnerUserId":696913.0,"Title":"How to Define Custom Theano Types Allowing Differentiation","Body":"<p>I am trying to follow the documentation to <a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/type.html\" rel=\"nofollow noreferrer\">create a double type<\/a> in Theano and implement operations on this type as described <a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/op.html\" rel=\"nofollow noreferrer\">here<\/a>. The current state can be found below:<\/p>\n\n<pre><\/pre>\n\n<p>In order to test the code, I wrote a few unit tests:<\/p>\n\n<pre><\/pre>\n\n<p>However, when testing the gradient computation, I get the following error at the marked line:<\/p>\n\n<pre><\/pre>\n\n<p>It seems that this is a problem of the double type defined above. However, the type itself is scale, so I should be able to compute the gradients using . Unfortunately, I could not find an example demonstrating the gradient computation on custom types and wasn't able to learn more about the  attribute ...<\/p>\n\n<p>Any help is appreciated; thanks!<\/p>\n\n<p><strong>Update.<\/strong> When trying to trick , e.g. by explicitly setting , the problems go on, e.g.<\/p>\n\n<pre><\/pre>\n\n<p>So it seems that I am missing something fundamental here and the defined Double type is missing several different type-specific information that is not mentioned in the documentation.<\/p>\n\n<p><strong>Update.<\/strong> After re-reading the documentation and looking into Theano's source code, the right question to ask is: Is it possible to define custom (non-tensor) types in Theano that allow differentation?<\/p>\n\n<p><strong>Update.<\/strong> Based on nouiz' answer, I am running into the next problems - these are giving me the impression that gradient computation is not intended for non-TensorType types:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41862452","CreationDate":"2017-01-25T22:01:34.537","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The answer is yes. You can. We do that ourself for sparse variable and GPU variable.<\/p>\n\n<p>But you hit corner cases theano.grad() wasn't made to support. Mostly, it expect an ndim parameter and a dtype parameter. Adding the dtype=\"float64\" parameter should work around that.<\/p>\n\n<p>ndim one is easy to fix in Theano with this diff:<\/p>\n\n<pre><code>diff --git a\/theano\/gradient.py b\/theano\/gradient.py\nindex 6d6fbaf..3b4d706 100644\n--- a\/theano\/gradient.py\n+++ b\/theano\/gradient.py\n@@ -433,7 +433,7 @@ def grad(cost, wrt, consider_constant=None,\n                          \"cost is NaN because \" +\n                          cost.type.why_null)\n\n-    if cost is not None and cost.ndim != 0:\n+    if cost is not None and getattr(cost, 'ndim', 0) != 0:\n         raise TypeError(\"cost must be a scalar.\")\n\n     if isinstance(wrt, set):\n<\/code><\/pre>\n\n<p>For the dtype, it is more complicated as we use it at many places for verifycation purpose (you can't take the grad of integers for example) and also to init the chain of gradient (or you can pass it via the known_grad parameter)<\/p>\n\n<p>UPDATE: The new error can be fixed with this bigger diff:<\/p>\n\n<pre><code>diff --git a\/theano\/gradient.py b\/theano\/gradient.py\nindex 6d6fbaf..6a9ec03 100644\n--- a\/theano\/gradient.py\n+++ b\/theano\/gradient.py\n@@ -433,7 +433,7 @@ def grad(cost, wrt, consider_constant=None,\n                          \"cost is NaN because \" +\n                          cost.type.why_null)\n\n-    if cost is not None and cost.ndim != 0:\n+    if cost is not None and getattr(cost, 'ndim', 0) != 0:\n         raise TypeError(\"cost must be a scalar.\")\n\n     if isinstance(wrt, set):\n@@ -1341,7 +1341,7 @@ def _float_ones_like(x):\n     if dtype not in tensor.float_dtypes:\n         dtype = theano.config.floatX\n\n-    return tensor.ones_like(x, dtype=dtype)\n+    return x.ones_like(dtype=dtype)\n\n\n class numeric_grad(object):\ndiff --git a\/theano\/tensor\/var.py b\/theano\/tensor\/var.py\nindex 2ecb9f0..6b08a45 100644\n--- a\/theano\/tensor\/var.py\n+++ b\/theano\/tensor\/var.py\n@@ -727,6 +727,9 @@ class _tensor_py_operators(object):\n     def zeros_like(model, dtype=None):\n         return theano.tensor.basic.zeros_like(model, dtype=dtype)\n\n+    def ones_like(model, dtype=None):\n+        return theano.tensor.basic.ones_like(model, dtype=dtype)\n+\n     def cumsum(self, axis=None):\n         return theano.tensor.extra_ops.cumsum(self, axis)\n<\/code><\/pre>\n\n<p>You need to add the ones_like method to your Variable like this:\n    def my_ones_like(model, dtype=None):\n        return ...\n    double.ones_like = my_ones_like<\/p>\n"}]}
{"QuestionId":41858831,"AnswerCount":1,"Tags":"<python><python-3.x><machine-learning><tensorflow><mnist>","CreationDate":"2017-01-25T18:21:58.247","AcceptedAnswerId":41859197.0,"OwnerUserId":6837416.0,"Title":"Python - How to use MNIST to predict a np.darray with different shape","Body":"<pre><\/pre>\n\n<p>I want to use Tensorflow or whatever is needed to predict which number is in this cropped image so I can keep doing that until all numbers can be predicted.<\/p>\n\n<p>The array properties are pretty different from MNIST database since this is not a 28x28 image.<\/p>\n\n<p>Is there any way that I can figure out linear transformations to do it for me or something?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41859197","CreationDate":"2017-01-25T18:42:25.300","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You can add a resize operation before the first convolution:<\/p>\n\n<pre><code>x = tf.image.resize_images(x, [28, 28])\n<\/code><\/pre>\n\n<p>So that can accept different image sizes. Make sure your input array has shape:<\/p>\n\n<pre><code>x.get_shape().as_list() == [ None, width, height, channels] \n<\/code><\/pre>\n"}]}
{"QuestionId":41859280,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-25T18:46:25.507","AcceptedAnswerId":41862591.0,"OwnerUserId":3531976.0,"Title":"How create new graph starting from a previous model in tensorflow?","Body":"<p>My goal is save an tensorflow model and restore it, but with different shape variables (I want add new classes and new training examples).\nIn <a href=\"https:\/\/stackoverflow.com\/questions\/35585053\/restoring-from-checkpoint-when-variable-shape-changes\">this question<\/a> it's explained that the option reshape of Saver works if the shapes have the same total number of elements, and it's necessary to build a new graph.\nMy question is: Is possible to build a new graph with the old trained model? So that resume the training with the new class.<\/p>\n","answers":[{"AnswerId":"41862591","CreationDate":"2017-01-25T22:10:14.320","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>You can initialize shared part of the model from old checkpoint. Please take a look to: <code>tf.contrib.framework.init_from_checkpoint<\/code><\/p>\n"}]}
{"QuestionId":41859605,"AnswerCount":4,"Tags":"<tensorflow><cross-validation><training-data>","CreationDate":"2017-01-25T19:05:06.920","AcceptedAnswerId":41862548.0,"OwnerUserId":2985049.0,"Title":"Split tensor into training and test sets","Body":"<p>Let's say I've read in a textfile using a . Is there some way to split this into train and test sets in ? Something like:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43498355","CreationDate":"2017-04-19T14:17:22.107","ParentId":null,"OwnerUserId":"702986","Title":null,"Body":"<p>As elham mentioned, you can use <a href=\"http:\/\/scikit-learn.org\/stable\/\" rel=\"noreferrer\">scikit-learn<\/a> to do this easily. scikit-learn is an open source library for machine learning. There are tons of tools for data preparation including the <a href=\"http:\/\/scikit-learn.org\/stable\/model_selection.html\" rel=\"noreferrer\"><code>model_selection<\/code><\/a> module, which handles comparing, validating and choosing parameters.<\/p>\n\n<p>The <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.train_test_split.html\" rel=\"noreferrer\"><code>model_selection.train_test_split()<\/code><\/a> method is specifically designed to split your data into train and test sets <strong>randomly and by percentage<\/strong>.<\/p>\n\n<pre><code>X_train, X_test, y_train, y_test = train_test_split(features,\n                                                    labels,\n                                                    test_size=0.33,\n                                                    random_state=42)\n<\/code><\/pre>\n\n<p><code>test_size<\/code> is the percentage to reserve for testing and <code>random_state<\/code> is to seed the random sampling.<\/p>\n\n<p>I typically use this to provide train and validation data sets, and keep true test data separately. You could just run <code>train_test_split<\/code> twice to do this as well. I.e. split the data into (Train + Validation) and Test, then split Train + Validation into two separate tensors.<\/p>\n"},{"AnswerId":"49825457","CreationDate":"2018-04-13T21:37:27.620","ParentId":null,"OwnerUserId":"9144242","Title":null,"Body":"<p>I managed to have a nice result using the map and filter functions of the tf.data.Dataset api. Just use the map function to randomly select the examples between train and testing. In order to do that you can, for each example, get a sample from a uniform distribution and check if the sample value is below the rate division.<\/p>\n\n<pre><code>def split_train_test(parsed_features, train_rate):\n    parsed_features['is_train'] = tf.gather(tf.random_uniform([1], maxval=100, dtype=tf.int32) &lt; tf.cast(train_rate * 100, tf.int32), 0)\n    return parsed_features\n\ndef grab_train_examples(parsed_features):\n    return parsed_features['is_train']\n\ndef grab_test_examples(parsed_features):\n    return ~parsed_features['is_train']\n<\/code><\/pre>\n"},{"AnswerId":"41862548","CreationDate":"2017-01-25T22:06:42.850","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>Something like following should work:\n<code>tf.split_v(tf.random_shuffle(...<\/code><\/p>\n"},{"AnswerId":"43185018","CreationDate":"2017-04-03T12:44:58.297","ParentId":null,"OwnerUserId":"5959830","Title":null,"Body":"<pre><code>import sklearn.model_selection as sk\n\nX_train, X_test, y_train, y_test = \nsk.train_test_split(features,labels,test_size=0.33, random_state = 42)\n<\/code><\/pre>\n"}]}
{"QuestionId":41859997,"AnswerCount":1,"Tags":"<python><neural-network><keras>","CreationDate":"2017-01-25T19:26:45.337","AcceptedAnswerId":41860156.0,"OwnerUserId":7230498.0,"Title":"Keras model load_weights for Neural Net","Body":"<p>I'm using the Keras library to create a neural network in python. I have loaded the training data (txt file), initiated the network and \"fit\" the weights of the neural network. I have then written code to generate the output text. Here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>My problem is: on execution the following error is produced:<\/p>\n\n<pre><\/pre>\n\n<p>I have added the following but the error still persists: <\/p>\n\n<pre><\/pre>\n\n<p>Any help would be appreciated.<\/p>\n","answers":[{"AnswerId":"41860156","CreationDate":"2017-01-25T19:36:15.033","ParentId":null,"OwnerUserId":"7438048","Title":null,"Body":"<p>you need to first create the network object called <code>model<\/code>, compile it and only after call the <code>model.load_weights(fname)<\/code><\/p>\n\n<p>working example:<\/p>\n\n<pre><code>from keras.models import Sequential\nfrom keras.layers import Dense, Activation\n\n\ndef build_model():\n    model = Sequential()\n\n    model.add(Dense(output_dim=64, input_dim=100))\n    model.add(Activation(\"relu\"))\n    model.add(Dense(output_dim=10))\n    model.add(Activation(\"softmax\"))\n\n    # you can either compile or not the model\n    model.compile(loss='categorical_crossentropy', optimizer='sgd', metrics=['accuracy'])\n    return model\n\n\nmodel1 = build_model()\nmodel1.save_weights('my_weights.model')\n\n\nmodel2 = build_model()\nmodel2.load_weights('my_weights.model')\n\n# do stuff with model2 (e.g. predict())\n<\/code><\/pre>\n\n<h1>Save &amp; Load an Entire Model<\/h1>\n\n<p>in Keras we can save &amp; load the entire model like this (more info <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\">here<\/a>):<\/p>\n\n<pre><code>from keras.models import load_model\n\nmodel1 = build_model()\nmodel1.save('my_model.hdf5')\n\nmodel2 = load_model('my_model.hdf5')\n# do stuff with model2 (e.g. predict()\n<\/code><\/pre>\n"}]}
{"QuestionId":41860195,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-25T19:37:58.187","AcceptedAnswerId":null,"OwnerUserId":7470820.0,"Title":"TensorFlow Vanishing Gradients","Body":"<p>i'm trying to replicate Recurrent Spatial Transformer Network implemented here (<a href=\"https:\/\/github.com\/skaae\/recurrent-spatial-transformer-code\" rel=\"nofollow noreferrer\">https:\/\/github.com\/skaae\/recurrent-spatial-transformer-code<\/a>) , however the loss didn't decrease at all .<\/p>\n\n<p>the configuration of the network is as follow:<\/p>\n\n<p>1 - relu activations .<\/p>\n\n<p>2 - xavier weight initialization for weights , zero initialization for biases .<\/p>\n\n<p>3 - cost function is softmax_cross_entropy_with_logits .<\/p>\n\n<p>4 - optimizer is RMSProp (i tried 1e-6 ;1e-10 espilon) .<\/p>\n\n<p>5 - gradient clipping by value .<\/p>\n\n<p>so what should i try next ?<\/p>\n\n<p>Below is the detailed Code <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43247120","CreationDate":"2017-04-06T06:15:22.487","ParentId":null,"OwnerUserId":"5915270","Title":null,"Body":"<p>Well you can use drop out it will be very useful. If you are using lstm or rnn you can implement dropout very easily. `<\/p>\n\n<pre><code>def create_rnn_cell():\n        encoDecoCell = tf.contrib.rnn.BasicLSTMCell(  # Or GRUCell, LSTMCell(args.hiddenSize)\n            self.args.hiddenSize,\n        )\n        if not self.args.test:  # TODO: Should use a placeholder instead\n            encoDecoCell = tf.contrib.rnn.DropoutWrapper(                   #using the dropout\n                encoDecoCell,\n                input_keep_prob=1.0,\n                output_keep_prob=self.args.dropout\n            )\n        return encoDecoCell\n<\/code><\/pre>\n\n<p>Batch normalization is also effective. But I haven't seen any BN examples implemented if tf specially for RNN modules. But here's good example to learn about batch norm <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/36978798\/batch-normalization-in-tensorflow\">Batch Normalization in tensorflow<\/a><\/p>\n\n<p>Also read this article where you can apply batch normalization to RNN<\/p>\n\n<p><a href=\"http:\/\/olavnymoen.com\/2016\/07\/07\/rnn-batch-normalization\" rel=\"nofollow noreferrer\">http:\/\/olavnymoen.com\/2016\/07\/07\/rnn-batch-normalization<\/a><\/p>\n"}]}
{"QuestionId":41860817,"AnswerCount":1,"Tags":"<optimization><machine-learning><tensorflow><deep-learning><bayesian>","CreationDate":"2017-01-25T20:13:07.383","AcceptedAnswerId":46318446.0,"OwnerUserId":7415539.0,"Title":"Hyperparameter optimization for Deep Learning Structures using Bayesian Optimization","Body":"<p>I have constructed a CLDNN (Convolutional, LSTM, Deep Neural Network) structure for raw signal classification task.<\/p>\n\n<p>Each training epoch runs for about 90 seconds and the hyperparameters seems to be very difficult to optimize.<\/p>\n\n<p>I have been research various ways to optimize the hyperparameters (e.g. random or grid search) and found out about Bayesian Optimization.<\/p>\n\n<p>Although I am still not fully understanding the optimization algorithm, I feed like it will help me greatly.<\/p>\n\n<p>I would like to ask few questions regarding the optimization task.<\/p>\n\n<ol>\n<li>How do I set up the Bayesian Optimization with regards to a deep network?(What is the cost function we are trying to optimize?)<\/li>\n<li>What is the function I am trying to optimize? Is it the cost of the validation set after N epochs?<\/li>\n<li>Is spearmint a good starting point for this task? Any other suggestions for this task?<\/li>\n<\/ol>\n\n<p>I would greatly appreciate any insights into this problem.<\/p>\n","answers":[{"AnswerId":"46318446","CreationDate":"2017-09-20T09:36:43.050","ParentId":null,"OwnerUserId":"712995","Title":null,"Body":"<blockquote>\n  <p>Although I am still not fully understanding the optimization\n  algorithm, I feed like it will help me greatly.<\/p>\n<\/blockquote>\n\n<p>First up, let me briefly explain this part.\nBayesian Optimization methods aim to deal with exploration-exploitation trade off in the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Multi-armed_bandit\" rel=\"noreferrer\">multi-armed bandit problem<\/a>. In this problem, there is an <em>unknown<\/em> function, which we can evaluate in any point, but each evaluation costs (direct penalty or opportunity cost), and the goal is to find its maximum using as few trials as possible. Basically, the trade off is this: you know the function in a finite set of points (of which some are good and some are bad), so you can try an area around the current local maximum, hoping to improve it (exploitation), or you can try a completely new area of space, that can potentially be much better or much worse (exploration), or somewhere in between.<\/p>\n\n<p>Bayesian Optimization methods (e.g. PI, EI, UCB), build a model of the target function using a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Gaussian_process\" rel=\"noreferrer\">Gaussian Process<\/a> (GP) and at each step choose the most \"promising\" point based on their GP model (note that \"promising\" can be defined differently by different particular methods).<\/p>\n\n<p>Here's an example:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ksQFy.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ksQFy.png\" alt=\"sin(x)*x\"><\/a><\/p>\n\n<p>The true function is <code>f(x) = x * sin(x)<\/code> (black curve) on <code>[-10, 10]<\/code> interval. Red dots represent each trial, red curve is the GP <em>mean<\/em>, blue curve is the mean plus or minus one <em>standard deviation<\/em>. \nAs you can see, the GP model doesn't match the true function everywhere, but the optimizer fairly quickly identified the \"hot\" area around <code>-8<\/code> and started to exploit it.<\/p>\n\n<blockquote>\n  <p>How do I set up the Bayesian Optimization with regards to a deep\n  network?<\/p>\n<\/blockquote>\n\n<p>In this case, the space is defined by (possibly transformed) hyperparameters, usually a multidimensional unit hypercube. <\/p>\n\n<p>For example, suppose you have three hyperparameters: a learning rate <code>\u03b1 in [0.001, 0.01]<\/code>, the regularizer <code>\u03bb in [0.1, 1]<\/code> (both continuous) and the hidden layer size <code>N in [50..100]<\/code> (integer). The space for optimization is a 3-dimensional cube <code>[0, 1]*[0, 1]*[0, 1]<\/code>. Each point <code>(p0, p1, p2)<\/code> in this cube corresponds to a trinity <code>(\u03b1, \u03bb, N)<\/code> by the following transformation:<\/p>\n\n<pre><code>p0 -&gt; \u03b1 = 10**(p0-3)\np1 -&gt; \u03bb = 10**(p1-1)\np2 -&gt; N = int(p2*50 + 50)\n<\/code><\/pre>\n\n<blockquote>\n  <p>What is the function I am trying to optimize? Is it the cost of the\n  validation set after N epochs?<\/p>\n<\/blockquote>\n\n<p>Correct, the target function is neural network validation accuracy. Clearly, each evaluation is expensive, because it requires at least several epochs for training.<\/p>\n\n<p>Also note that the target function is <em>stochastic<\/em>, i.e. two evaluations on the same point may slightly differ, but it's not a blocker for Bayesian Optimization, though it obviously increases the uncertainty.<\/p>\n\n<blockquote>\n  <p>Is spearmint a good starting point for this task? Any other\n  suggestions for this task?<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/kuz\/caffe-with-spearmint\" rel=\"noreferrer\">spearmint<\/a> is a good library, you can definitely work with that. I can also recommend <a href=\"http:\/\/hyperopt.github.io\/hyperopt\/\" rel=\"noreferrer\">hyperopt<\/a>.<\/p>\n\n<p>In my own research, I ended up writing my own tiny library, basically for two reasons: I wanted to code exact Bayesian method to use (in particular, I found a <a href=\"https:\/\/arxiv.org\/pdf\/1009.5419.pdf\" rel=\"noreferrer\">portfolio strategy<\/a> of UCB and PI converged faster than anything else, in my case); plus there is another technique that can save up to 50% of training time called <a href=\"http:\/\/aad.informatik.uni-freiburg.de\/papers\/15-IJCAI-Extrapolation_of_Learning_Curves.pdf\" rel=\"noreferrer\">learning curve prediction<\/a> (the idea is to skip full learning cycle when the optimizer is confident the model doesn't learn as fast as in other areas). I'm not aware of any library that implements this, so I coded it myself, and in the end it paid off. If you're interested, the code is <a href=\"https:\/\/github.com\/maxim5\/hyper-engine\" rel=\"noreferrer\">on GitHub<\/a>.<\/p>\n"}]}
{"QuestionId":41860871,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><nlp><deep-learning><tflearn>","CreationDate":"2017-01-25T20:15:21.687","AcceptedAnswerId":null,"OwnerUserId":4750553.0,"Title":"How to use tflearn deep learning for document classification","Body":"<p>I am using tflearn and tensorflow to classify documents. However I am facing issue with the size of the document and training time, the length of my largest document is ~98000 words and using this for the building the network is going to be extremely time consuming. I was looking at different methods to reduce the dimensions of the data or some technique that will be fast enough to train the model. I saw doc2vec and word2vec and built the respective models but am not sure of how to use them with tflearn. I am also classifying the documents by creating doc2vec and word2vec model and getting the feature vector by averaging the feature vectors of all words that exist in the model vocabulary in the doc. Can I use this final feature vector for DNN or does someone have other suggestions that I can use. <\/p>\n","answers":[{"AnswerId":"41862567","CreationDate":"2017-01-25T22:08:26.297","ParentId":null,"OwnerUserId":"884962","Title":null,"Body":"<p>You can easily build Document Classification models with TensorFlow and integrated into it TF.Learn library.<\/p>\n\n<p>There are even examples of the various document classifications models in examples folder: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/learn#text-classification\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/learn#text-classification<\/a><\/p>\n\n<p>The fastest model for any length document will be Bag of Words model - one that averages embeddings of words. This is recommended baseline for any document classification problems too. Then you can try more complicated models like RNNs or CNNs.<\/p>\n\n<p>Here is example code for it:<\/p>\n\n<pre><code>def bag_of_words_model(features, target):\n  \"\"\"A bag-of-words model. Note it disregards the word order in the text.\"\"\"\n  target = tf.one_hot(target, 15, 1, 0)\n  features = tf.contrib.layers.bow_encoder(\n      features, vocab_size=n_words, embed_dim=EMBEDDING_SIZE)\n  logits = tf.contrib.layers.fully_connected(features, 15, activation_fn=None)\n  loss = tf.losses.softmax_cross_entropy(target, logits)\n  train_op = tf.contrib.layers.optimize_loss(\n      loss,\n      tf.contrib.framework.get_global_step(),\n      optimizer='Adam',\n      learning_rate=0.01)\n  return ({\n    'class': tf.argmax(logits, 1),\n    'prob': tf.nn.softmax(logits)\n    }, loss, train_op)\n<\/code><\/pre>\n\n<p>For more details how to run it, see here - <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/text_classification.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/text_classification.py<\/a><\/p>\n\n<p>You can easily extend more fully connected layers (e.g. DNN part) by adding <code>tf.contrib.layers.fully_connected<\/code> before <code>logits<\/code> are computed.<\/p>\n\n<p>You can also initialize embeddings from pre-trained checkpoint with word2vec or other embeddings by using <code>tf.contrib.framework.init_from_checkpoint<\/code>(<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/contrib.framework\/checkpoint_utilities\" rel=\"nofollow noreferrer\">see documentation<\/a>).<\/p>\n"}]}
{"QuestionId":41860960,"AnswerCount":1,"Tags":"<python><python-2.7><hdf5><keras><h5py>","CreationDate":"2017-01-25T20:21:13.227","AcceptedAnswerId":41880397.0,"OwnerUserId":4446624.0,"Title":"Error while attempting to use an HDF5 dataset with Keras","Body":"<p>I get the following error when trying to use an HDF5 dataset with keras. It would seem that Sequential.fit(), while making a validation data slice, encounters that a key of the slice does not have the attribute 'stop'. I don't know if this is a formatting issue of my HDF5 dataset or something else. Any help would be appreciated.<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):<br>\n  File \"autoencoder.py\", line 73, in module<\/p>\n  \n  <p>validation_split=0.2)   <\/p>\n  \n  <p>File \"\/home\/ben\/.local\/lib\/python2.7\/site-packages\/keras\/models.py\", line 672, in fit <\/p>\n  \n  <p>initial_epoch=initial_epoch)   <\/p>\n  \n  <p>File \"\/home\/ben\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 1143, in fit<\/p>\n  \n  <p>x, val_x = (slice_X(x, 0, split_at), slice_X(x, split_at))   <\/p>\n  \n  <p>File \"\/home\/ben\/.local\/lib\/python2.7\/site-packages\/keras\/engine\/training.py\", line 301, in slice_X<\/p>\n  \n  <p>return [x[start:stop] for x in X]   <\/p>\n  \n  <p>File \"\/home\/ben\/.local\/lib\/python2.7\/site-packages\/keras\/utils\/io_utils.py\", line 71, in <strong>getitem<\/strong><\/p>\n  \n  <p>if key.stop + self.start &lt;= self.end: <\/p>\n  \n  <p>TypeError: unsupported operand type(s) for +: 'NoneType' and 'int'<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41880397","CreationDate":"2017-01-26T18:26:33.477","ParentId":null,"OwnerUserId":"4446624","Title":null,"Body":"<p>I didn't fix the error, but I got around it by using validation_data instead of validation_split in my fit call.<\/p>\n"}]}
{"QuestionId":41861354,"AnswerCount":2,"Tags":"<python><pytorch><torch><pre-trained-model>","CreationDate":"2017-01-25T20:45:35.297","AcceptedAnswerId":54261158.0,"OwnerUserId":1237313.0,"Title":"Loading Torch7 trained models (.t7) in PyTorch","Body":"<p>I am using Torch7 library for implementing neural networks.  Mostly, I rely on pre-trained models. In Lua I use torch.load function to load a model saved as torch .t7 file. I am curious about switching to PyTorch( <a href=\"http:\/\/pytorch.org\" rel=\"noreferrer\">http:\/\/pytorch.org<\/a>) and I read the documents. I couldn't find any information regarding the mechanisms to load a pre-trained model. The only relevant information I was able to find is this page:<a href=\"http:\/\/pytorch.org\/docs\/torch.html\" rel=\"noreferrer\">http:\/\/pytorch.org\/docs\/torch.html<\/a><\/p>\n\n<p>But the function torch.load described in the page seems to load a file saved with pickle. If someone has additional information on loading .t7 models in PyTorch, please share it here. <\/p>\n","answers":[{"AnswerId":"54261158","CreationDate":"2019-01-18T20:40:31.663","ParentId":null,"OwnerUserId":"2838606","Title":null,"Body":"<p>As of PyTorch 1.0 <code>torch.utils.serialization<\/code> is completely removed. Hence no one can import models from Lua Torch into PyTorch anymore. Instead, I would suggest installing PyTorch 0.4.1 through <code>pip<\/code> in a <code>conda<\/code> environment (so that you can remove it after this) and use <a href=\"https:\/\/github.com\/clcarwin\/convert_torch_to_pytorch\" rel=\"nofollow noreferrer\">this repo<\/a> to convert your Lua Torch model to PyTorch model, not just the <code>torch.nn.legacy<\/code> model that you cannot use for training. Then use PyTorch 1.xx to do whatever with it. You can also train your converted Lua Torch models in PyTorch this way :)<\/p>\n"},{"AnswerId":"42109564","CreationDate":"2017-02-08T09:43:31.803","ParentId":null,"OwnerUserId":"1330293","Title":null,"Body":"<p>The correct function is <code>load_lua<\/code>:<\/p>\n\n<pre><code>from torch.utils.serialization import load_lua\n\nx = load_lua('x.t7')\n<\/code><\/pre>\n"}]}
{"QuestionId":41861781,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-25T21:13:58.350","AcceptedAnswerId":41862142.0,"OwnerUserId":1217616.0,"Title":"feed_dict can not convert int to tensor in tensorflow","Body":"<p>It seems that the toughest part of tensorflow is matching tensors to their destinations.<\/p>\n\n<p>I get the runtime error:\nTypeError: \"Cannot interpret feed_dict key as Tensor: Can not convert a int into a Tensor.\"<\/p>\n\n<p>I'm trying to solve the XOR problem, this is my main set up.<\/p>\n\n<pre><\/pre>\n\n<p>Here's the whole code if you need to see more or want tell me what I can do better <a href=\"http:\/\/pastebin.com\/GqrX21vf\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/GqrX21vf<\/a><\/p>\n\n<p>inp_vec is [4x2] and solutions is [4x1]. Both should match the placeholders. I've been having problems with matching ever since I started tensorflow so I'm starting to think I fundamentally don't understand the setup. Clarification on this would be much appreciated.<\/p>\n","answers":[{"AnswerId":"41862142","CreationDate":"2017-01-25T21:40:46.007","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You are overwriting the reference to your variable <code>i<\/code> with the <code>i<\/code> of the for loop. Change the name to <code>x_<\/code> should fix the problem.<\/p>\n\n<pre><code>x_ = tf.placeholder(tf.float32, shape=[None,2])\ny_ = tf.placeholder(tf.float32, shape=[None,1])\n\ninp_vec = [[1.,0.],[0.,1.],[1.,1.],[0.,0.]]\nsolutions = [[1.],[1.],[0.],[0.]]\n\nwith tf.Session() as sess:\n    sess.run( tf.initialize_all_variables() )\n\n    for i in range(1000):\n       sess.run(optim, feed_dict={x_:inp_vec, y_:solutions})\n<\/code><\/pre>\n"}]}
{"QuestionId":41861885,"AnswerCount":2,"Tags":"<cuda><tensorflow><osx-elcapitan><cudnn>","CreationDate":"2017-01-25T21:22:04.930","AcceptedAnswerId":42292332.0,"OwnerUserId":5336205.0,"Title":"CUDA linking error on OS X El Capitan","Body":"<p>I've been trying to install CUDA and cuDNN for TensorFlow as <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#prepare_environment_for_mac_os_x\" rel=\"nofollow noreferrer\">instructed<\/a> on the TensorFlow page. Everything works until I do the following:<\/p>\n\n<pre><\/pre>\n\n<p>I then get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>My question of course is what's causing the error, and how to fix it. I had previously installed CUDA 7.5, and then I installed 8.0. The whole of CUDA (and its installation) is just a black box to me; I'm just trying to get it to work with TensorFlow. <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42292332","CreationDate":"2017-02-17T07:52:28.450","ParentId":null,"OwnerUserId":"1692033","Title":null,"Body":"<p>Run the following in your terminal -<\/p>\n\n<pre><code>xcode-select --install\n<\/code><\/pre>\n\n<p>Then try to run <code>make<\/code> again.<\/p>\n\n<p>Saw the answer on an NVidia forum - <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/628298\/cuda-samples-on-mac-os-x-mavericks-10-9-matrixmuldrv\/\" rel=\"nofollow noreferrer\">https:\/\/devtalk.nvidia.com\/default\/topic\/628298\/cuda-samples-on-mac-os-x-mavericks-10-9-matrixmuldrv\/<\/a><\/p>\n\n<p>It's also documented in this Quick Start Guide - <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-quick-start-guide\/index.html#axzz4YvZYOZc0\" rel=\"nofollow noreferrer\">http:\/\/docs.nvidia.com\/cuda\/cuda-quick-start-guide\/index.html#axzz4YvZYOZc0<\/a><\/p>\n"},{"AnswerId":"41861931","CreationDate":"2017-01-25T21:24:55.040","ParentId":null,"OwnerUserId":"5010116","Title":null,"Body":"<p>Add the folder containing the folder <code>CUDA.framework<\/code> to the command line as such:<\/p>\n\n<pre><code>-F&lt;folder&gt;\n<\/code><\/pre>\n\n<p>It looks like the linker isn't able to find the framework and you have to point it where it is<\/p>\n"}]}
{"QuestionId":41862924,"AnswerCount":1,"Tags":"<python><caffe><labels>","CreationDate":"2017-01-25T22:33:16.110","AcceptedAnswerId":null,"OwnerUserId":7064751.0,"Title":"Where can I find the ground truth labels of a ILSVRC2012 high level category, for example animals?","Body":"<p>I'm using the ILSVRC2012 dataset and for that, I have the ground_truth_labels and the synset_words.txt. Now, I access <a href=\"http:\/\/image-net.org\/synset?wnid=n02119789\" rel=\"nofollow noreferrer\">this<\/a> and download just the images containing animals.<\/p>\n\n<p>I use a neural network to make a prediction and now, I want to calculate the classification error but I don't have the ground truth labels.<\/p>\n\n<p><strong>My question is<\/strong>: where can I find the ground truth labels for these images? The name of them appear like this <\/p>\n\n<p>n00015388_12.JPEG, <\/p>\n\n<p>n00015388_24.JPEG, <\/p>\n\n<p>n00015388_119.JPEG<\/p>\n\n<p>Is there any relationship between the name of the images and the files given by caffe? (synset_word.txt and val.txt)<\/p>\n\n<p>What's the meaning of the numbers used to name the images?<\/p>\n","answers":[{"AnswerId":"41879584","CreationDate":"2017-01-26T17:38:56.777","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>The main correlation is that the directory (the <strong>n00015388<\/strong> part) maps uniquely to the label: There will be ~1300 images that share that prefix.<\/p>\n\n<p>I haven't yet found a correlation between the suffix (such as _119) and the file use.  The suffix is unique within each class, of course.  However, the quantity of each suffix varies quite a bit.<\/p>\n\n<p>If you want to investigate on your own, there's a complete list of files in your generated <strong>train.txt<\/strong> file, generated by one of the setup scripts -- or easy enough to generate from a \"ls\" file listing and header information.<\/p>\n"}]}
{"QuestionId":41863284,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2017-01-25T23:02:38.360","AcceptedAnswerId":null,"OwnerUserId":7387749.0,"Title":"Keras - Stationary result using model.fit()","Body":"<p>I'm implementing this simple neural network, with these inputs data:<\/p>\n\n<pre><\/pre>\n\n<p>After have defined the network's structure:<\/p>\n\n<pre><\/pre>\n\n<p>I run this to train and evaluate the NN:<\/p>\n\n<pre><\/pre>\n\n<p>and I always get the same result from model.fit( ... ) :\n <em>32\/143 [=====>........................] - ETA: 0s<\/em>.\n It seems that doesn't work at all, despite I obtain consistent results on training and validation. How does I have to interpreter this stationary result about <em>model. fit<\/em> output?<\/p>\n","answers":[]}
{"QuestionId":41863814,"AnswerCount":7,"Tags":"<python><statistics><tensorflow><entropy>","CreationDate":"2017-01-25T23:51:36.277","AcceptedAnswerId":null,"OwnerUserId":5540330.0,"Title":"Is there a built-in KL divergence loss function in TensorFlow?","Body":"<p>I have two tensors,  and  with shape , and I want to compute the KL divergence from  to . Is there a built-in function for this in TensorFlow?  I tried using , but it gives:<\/p>\n\n<blockquote>\n  <p>NotImplementedError: No KL(dist_a || dist_b) registered for dist_a type Tensor and dist_b type Tensor<\/p>\n<\/blockquote>\n\n<p>If there is no built-in function, what would be a good workaround?<\/p>\n","answers":[{"AnswerId":"41864069","CreationDate":"2017-01-26T00:18:07.757","ParentId":null,"OwnerUserId":"4751105","Title":null,"Body":"<p>I'm not sure why it's not implemented, but perhaps there is a workaround. The KL divergence is defined as:<\/p>\n\n<p><code>KL(prob_a, prob_b) = Sum(prob_a * log(prob_a\/prob_b))<\/code><\/p>\n\n<p>The cross entropy H, on the other hand, is defined as:<\/p>\n\n<p><code>H(prob_a, prob_b) = -Sum(prob_a * log(prob_b))<\/code><\/p>\n\n<p>So, if you create a variable <code>y = prob_a\/prob_b<\/code>, you could obtain the KL divergence by calling negative <code>H(proba_a, y)<\/code>. In Tensorflow notation, something like:<\/p>\n\n<p><code>KL = tf.reduce_mean(-tf.nn.softmax_cross_entropy_with_logits(prob_a, y))<\/code><\/p>\n"},{"AnswerId":"49418266","CreationDate":"2018-03-21T23:41:08.257","ParentId":null,"OwnerUserId":"4970778","Title":null,"Body":"<p>Assuming that you have access to logits a and b:<\/p>\n\n<pre><code>prob_a = tf.nn.softmax(a)\ncr_aa = tf.nn.softmax_cross_entropy_with_logits(prob_a, a)\ncr_ab = tf.nn.softmax_cross_entropy_with_logits(prob_a, b)\nkl_ab = tf.reduce_sum(cr_ab - cr_aa)\n<\/code><\/pre>\n"},{"AnswerId":"54252592","CreationDate":"2019-01-18T11:00:03.817","ParentId":null,"OwnerUserId":"7750274","Title":null,"Body":"<p>I think this might work:<\/p>\n\n<pre><code>tf.reduce_sum(p * tf.log(p\/q))\n<\/code><\/pre>\n\n<p>where p is my actual probability distribution \nand q is my approximate probability distribution.<\/p>\n"},{"AnswerId":"54726928","CreationDate":"2019-02-16T19:37:04.410","ParentId":null,"OwnerUserId":"10940584","Title":null,"Body":"<p>I used the function from <a href=\"https:\/\/gist.github.com\/FelixMohr\/a83e57393647dd2971d5ae12bb01f504#file-vae-py\" rel=\"nofollow noreferrer\">this code<\/a> (from <a href=\"https:\/\/towardsdatascience.com\/teaching-a-variational-autoencoder-vae-to-draw-mnist-characters-978675c95776\" rel=\"nofollow noreferrer\">this<\/a> Medium post) to calculate the KL-divergence of any given tensor from a normal Gaussian distribution, where <code>sd<\/code> is the standard deviation and <code>mn<\/code> is the tensor.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>latent_loss = -0.5 * tf.reduce_sum(1.0 + 2.0 * sd - tf.square(mn) - tf.exp(2.0 * sd), 1)\n<\/code><\/pre>\n"},{"AnswerId":"44815939","CreationDate":"2017-06-29T03:34:03.750","ParentId":null,"OwnerUserId":"7433574","Title":null,"Body":"<p>For there is softmax_cross_entropy_with_logits, there is no need to optimize on KL.<\/p>\n\n<pre><code>KL(prob_a, prob_b)  \n  = Sum(prob_a * log(prob_a\/prob_b))  \n  = Sum(prob_a * log(prob_a) - prob_a * log(prob_b))  \n  = - Sum(prob_a * log(prob_b)) + Sum(prob_a * log(prob_a)) \n  = - Sum(prob_a * log(prob_b)) + const \n  = H(prob_a, prob_b) + const \n<\/code><\/pre>\n"},{"AnswerId":"51031305","CreationDate":"2018-06-25T20:23:34.437","ParentId":null,"OwnerUserId":"5235274","Title":null,"Body":"<p>Assuming that your input tensors <code>prob_a<\/code> and <code>prob_b<\/code> are probability tensors that sum to 1 along the last axis, you could do it like this:<\/p>\n\n<pre><code>def kl(x, y):\n    X = tf.distributions.Categorical(probs=x)\n    Y = tf.distributions.Categorical(probs=y)\n    return tf.distributions.kl_divergence(X, Y)\n\nresult = kl(prob_a, prob_b)\n<\/code><\/pre>\n\n<p>A simple example:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\na = np.array([[0.25, 0.1, 0.65], [0.8, 0.15, 0.05]])\nb = np.array([[0.7, 0.2, 0.1], [0.15, 0.8, 0.05]])\nsess = tf.Session()\nprint(kl(a, b).eval(session=sess))  # [0.88995184 1.08808468]\n<\/code><\/pre>\n\n<p>You would get the same result with<\/p>\n\n<pre><code>np.sum(a * np.log(a \/ b), axis=1) \n<\/code><\/pre>\n\n<p>However, this implementation is a bit buggy (checked in Tensorflow 1.8.0).<\/p>\n\n<p>If you have zero probabilities in <code>a<\/code>, e.g. if you try <code>[0.8, 0.2, 0.0]<\/code> instead of <code>[0.8, 0.15, 0.05]<\/code>, you will get <code>nan<\/code> even though by Kullback-Leibler definition <code>0 * log(0 \/ b)<\/code> should contribute as zero.<\/p>\n\n<p>To mitigate this, one should add some small numerical constant. It is also prudent to use <code>tf.distributions.kl_divergence(X, Y, allow_nan_stats=False)<\/code> to cause a runtime error in such situations.<\/p>\n\n<p>Also, if there are some zeros in <code>b<\/code>, you will get <code>inf<\/code> values which won't be caught by the <code>allow_nan_stats=False<\/code> option so those have to be handled as well.<\/p>\n"},{"AnswerId":"45177940","CreationDate":"2017-07-18T22:09:59.543","ParentId":null,"OwnerUserId":"433287","Title":null,"Body":"<p><code>tf.contrib.distributions.kl<\/code> takes instances of a <code>tf.distribution<\/code> not a <code>Tensor<\/code>.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>  ds = tf.contrib.distributions\n  p = ds.Normal(loc=0., scale=1.)\n  q = ds.Normal(loc=1., scale=2.)\n  kl = ds.kl_divergence(p, q)\n  # ==&gt; 0.44314718\n<\/code><\/pre>\n"}]}
{"QuestionId":41863921,"AnswerCount":1,"Tags":"<python><machine-learning><keras><one-hot-encoding>","CreationDate":"2017-01-26T00:03:22.140","AcceptedAnswerId":41865011.0,"OwnerUserId":3466910.0,"Title":"How can I use categorical one-hot labels for training with Keras?","Body":"<p>I have inputs that look like this: <\/p>\n\n<pre><\/pre>\n\n<p>of shape , and labels that look like this:<\/p>\n\n<pre><\/pre>\n\n<p>of shape .<\/p>\n\n<p>However, when I try to run the following Keras code, I get this error:\n. From what I've read, this appears to stem from the fact that my labels are 2D, and not simply integers. Is this correct, and if so, <em>how can I use one-hot labels with Keras?<\/em><\/p>\n\n<p>Here's the code:<\/p>\n\n<pre><\/pre>\n\n<p>Which promptly throws the error above:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"41865011","CreationDate":"2017-01-26T02:24:49.237","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>There are a lot of things that do not add up.<\/p>\n\n<p>I assume that you are trying to solve a sequential classification task, i.e. your data is shaped as <code>(&lt;batch size&gt;, &lt;sequence length&gt;, &lt;feature length&gt;)<\/code>. <\/p>\n\n<p>In your batch generator you create a batch consisting of one sequence of length 8038 and 463 features per sequence element. You create a matching Y batch to compare against, consisting of one sequence with 8038 elements, each of size 2. <\/p>\n\n<p>Your problem is that <code>Y<\/code> does not match up with the output of the last layer. Your <code>Y<\/code> is 3-dimensional while the output of your model is only 2-dimensional: <code>Y.shape = (1, 8038, 2)<\/code> does not match <code>dense_1.shape = (1,1)<\/code>. This explains the error message you get.<\/p>\n\n<p>The solution to this: you need to enable <code>return_sequences=True<\/code> in the LSTM layer to return a sequence instead of only the last element (effectively removing the time-dimension). This would give an output shape of <code>(1, 8038, 100)<\/code> at the LSTM layer. Since the <code>Dense<\/code> layer is not able to handle sequential data you need to apply it to each sequence element individually which is done by wrapping it in a <a href=\"https:\/\/keras.io\/layers\/wrappers\/#timedistributed\" rel=\"noreferrer\"><code>TimeDistributed<\/code><\/a> wrapper. This then gives your model the output shape <code>(1, 8038, 1)<\/code>.<\/p>\n\n<p>Your model should look like this:<\/p>\n\n<pre><code>from keras.layers.wrappers import TimeDistributed\n\nmodel = Sequential()\nmodel.add(LSTM(100, input_dim=num_features, return_sequences=True))\nmodel.add(TimeDistributed(Dense(1, activation='sigmoid')))\n<\/code><\/pre>\n\n<p>This can be easily spotted when examining the summary of the model:<\/p>\n\n<pre><code>print(model.summary()) \n<\/code><\/pre>\n"}]}
{"QuestionId":41864333,"AnswerCount":2,"Tags":"<python><image-processing><machine-learning><tensorflow><batch-processing>","CreationDate":"2017-01-26T00:54:08.970","AcceptedAnswerId":41869952.0,"OwnerUserId":7449707.0,"Title":"Tensorflow mixes up images and labels when making batch","Body":"<p>So I've been stuck on this problem for weeks. I want to make an image batch from a list of image filenames. I insert the filename list into a queue and use a reader to get the file. The reader then returns the filename and the read image file. <\/p>\n\n<p>My problem is that when I make a batch using the decoded jpg and the labels from the reader, tf.train.shuffle_batch() mixes up the images and the filenames so that now the labels are in the wrong order for the image files. Is there something I am doing wrong with the queue\/shuffle_batch and how can I fix it such that the batch comes out with the right labels for the right files? <\/p>\n\n<p>Much thanks!<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n","answers":[{"AnswerId":"41864611","CreationDate":"2017-01-26T01:31:13.690","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>From your code I'm unsure how your labels are encoded\/extracted from the jpeg images. I used to encode everything in the same file, but have since found a much more elegant solution. Assuming you can get a list of filenames, <code>image_paths<\/code> and a numpy array of labels <code>labels<\/code>, you can bind them together and operate on individual examples with <code>tf.train.slice_input_producer<\/code> then batch them together using <code>tf.train.batch<\/code>.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.framework import ops\n\nshuffle = True\nbatch_size = 128\nnum_threads = 8\n\ndef get_data():\n    \"\"\"\n    Return image_paths, labels such that label[i] corresponds to image_paths[i].\n\n    image_paths: list of strings\n    labels: list\/np array of labels\n    \"\"\"\n    raise NotImplementedError()\n\ndef preprocess_image_tensor(image_tf):\n    \"\"\"Preprocess a single image.\"\"\"\n    image = tf.image.convert_image_dtype(image_tf, dtype=tf.float32)\n    image = tf.image.resize_image_with_crop_or_pad(image, 300, 300)\n    image = tf.image.per_image_standardization(image)\n    return image\n\nimage_paths, labels = get_data()\n\nimage_paths_tf = ops.convert_to_tensor(image_paths, dtype=tf.string, name='image_paths')\nlabels_tf = ops.convert_to_tensor(image_paths, dtype=tf.int32, name='labels')\nimage_path_tf, label_tf = tf.train.slice_input_producer([image_paths_tf, labels_tf], shuffle=shuffle)\n\n# preprocess single image paths\nimage_buffer_tf = tf.read_file(image_path_tf, name='image_buffer')\nimage_tf = tf.image.decode_jpeg(image_buffer_tf, channels=3, name='image')\nimage_tf = preprocess_image_tensor(image_tf)\n\n# batch the results\nimage_batch_tf, labels_batch_tf = tf.train.batch([image_tf, label_tf], batch_size=batch_size, num_threads=num_threads)\n<\/code><\/pre>\n"},{"AnswerId":"41869952","CreationDate":"2017-01-26T09:04:36.970","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>Wait.... Isn't your tf usage a little weird?<\/p>\n\n<p>You are basically running the graph twice by calling:<\/p>\n\n<pre><code>  print image_path_tf.eval()\n  print label_tf.eval()\n<\/code><\/pre>\n\n<p>And since you are only asking for <code>image_path_tf<\/code> and <code>label_tf<\/code>, anything below this line is not even run:<\/p>\n\n<pre><code>image_path_tf, label_tf = tf.train.slice_input_producer([image_paths_tf, labels_tf], shuffle=False)\n<\/code><\/pre>\n\n<p>Maybe try this?<\/p>\n\n<pre><code>image_paths, labels = sess.run([images_batch_tf, labels_batch_tf])\nprint(image_paths)\nprint(labels)\n<\/code><\/pre>\n"}]}
{"QuestionId":41865165,"AnswerCount":1,"Tags":"<python><tensorflow><nlp><recurrent-neural-network>","CreationDate":"2017-01-26T02:47:45.380","AcceptedAnswerId":42742109.0,"OwnerUserId":1462047.0,"Title":"Tensorflow examples for RNN","Body":"<p>I am trying to implement basic NLP tasks in Tensorflow without using the build in modules as much as possible (just for learning sake)<\/p>\n\n<p>I have been trying to implement a Part of Speech tagger using data from <a href=\"http:\/\/www.cnts.ua.ac.be\/conll2000\/chunking\/\" rel=\"nofollow noreferrer\">http:\/\/www.cnts.ua.ac.be\/conll2000\/chunking\/<\/a><\/p>\n\n<p>I am having a little difficulty with implementing a RNN code from scratch using an embedding layer in front and was wondering if there are examples and implementations on the same.<\/p>\n\n<p>I have seen tons of examples using Theano and on MNIST data but havent been able to find a concrete implementation of RNNs from scratch in Tensorflow on textual data.<\/p>\n\n<p>Any suggestions?<\/p>\n","answers":[{"AnswerId":"42742109","CreationDate":"2017-03-11T23:57:06.623","ParentId":null,"OwnerUserId":"1462047","Title":null,"Body":"<p>after searching for a long time, I decided to do it on my own.\nHere are a couple of repos I created:<\/p>\n\n<p><a href=\"https:\/\/github.com\/roopalgarg\/recurrent-neural-nets-tensorflow\" rel=\"nofollow noreferrer\">recurrent-neural-nets-tensorflow<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/roopalgarg\/lstm-tensorflow\" rel=\"nofollow noreferrer\">lstm-tensorflow<\/a><\/p>\n"}]}
{"QuestionId":41865218,"AnswerCount":2,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-01-26T02:54:36.270","AcceptedAnswerId":null,"OwnerUserId":314710.0,"Title":"Interleaving slim.dropout and slim.fully_connected in slim.stack?","Body":"<p>In tf.slim, I'd like to create a stack of fully-connected layers with dropout.<\/p>\n\n<p>To the example from documentation:\n, I'd like to add dropout. <\/p>\n\n<p>Is it possible to use slim.stack or do I have to go back to the verbose approach?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41884626","CreationDate":"2017-01-26T23:10:10.103","ParentId":null,"OwnerUserId":"314710","Title":null,"Body":"<p>Based on @guinny's comment, I can create a local function:<\/p>\n\n<pre><code>layer = lambda inputs, layer_args, **kwargs:      \n    slim.dropout(slim.fully_connected(inputs, layer_args, **kwargs))\nwith slim.arg_scope([slim.dropout], keep_prob=dropout_keep_prob):\n    layers = slim.stack(inputs,\n                        layer,\n                        layer_sizes,\n                        activation_fn=nonlinearity,\n                        scope=scope)\n<\/code><\/pre>\n"},{"AnswerId":"41870484","CreationDate":"2017-01-26T09:36:24.250","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>Just look at the code: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L1976\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L1976<\/a>. You need a local function that has the following signature:<\/p>\n\n<pre><code>outputs = layer(outputs, *layer_args, **kwargs)\n<\/code><\/pre>\n"}]}
{"QuestionId":41865338,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-01-26T03:10:57.500","AcceptedAnswerId":null,"OwnerUserId":7472044.0,"Title":"Changing Inception-v4 architecture to do Multi-label classification in Tensorflow","Body":"<p>I am working on image tagging and annotation problem, simply an image may contain multiple objects. I want to train inception-v4 for multi-label classification. My training data will be an image and a vector of length equals the number of classes and has 1 in each index if the object exists in the image. For example, If I have four classes (Person, car, tree, buildings). If an image contains a person and car. Then my vector will be (1, 1, 0, 0).<\/p>\n\n<p>What changes do I need to make to train inception-v4 for the tagging and annotation problem?<\/p>\n\n<p>Do I only need to change the input format and change the loss function from softmax to sigmoid_cross_entropy_with_logits in the inception-v4 architecture?<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/nets\/inception_v4.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/nets\/inception_v4.py<\/a><\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"43854543","CreationDate":"2017-05-08T18:10:35.700","ParentId":null,"OwnerUserId":"6378372","Title":null,"Body":"<p>If you'd like to retrain a model to output different labels, check out the <code>image_retraining<\/code> example: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/examples\/image_retraining\/retrain.py<\/a> <\/p>\n\n<p>In that example, we retrain the standard inception v3 model to recognize flowers instead of the standard ImageNet categories.<\/p>\n"}]}
{"QuestionId":41865482,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-26T03:29:15.837","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"Why does the location of saver in the script matter when there is a graph object in TensorFlow?","Body":"<p>I was training some models and I noticed that when I explicitly defined a graph variable, then it mattered where my saver object was created. First my code looked like this:<\/p>\n\n<pre><\/pre>\n\n<p>then I decided to change it to something like this and it seemed very sensitive where I defined my variables and where I defined saver. If they were not defined exactly after the graph variable was created for example, it would have errors. Similarly, I noticed that saver had to be defined exactly after one single variable (note being after the definition of the graph was <strong>not<\/strong> enough) for all the variables to be captured together by the saver (which didn't make sense to me, it would make more sense to require that its behind the definition of <strong>all<\/strong> the variables rather than a single one for this to work). <\/p>\n\n<p>This is how the code looks now (with comments showing the locations of where I've defined saver):<\/p>\n\n<pre><\/pre>\n\n<p>the code above should work, but I am having a hard time understand why it behaves like this or why it makes sense that this would happen. Someone know what the right thing to do is?<\/p>\n","answers":[{"AnswerId":"41866873","CreationDate":"2017-01-26T04:26:02.180","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>I'm not entirely sure what's going on here, but I suspect the issue is related to variables not going into the wrong graph, or the session having an outdated version of the graph. You create a graph, but don't set it as the default, then create a session with that graph... but when you create variables, you don't specify which graph they should go into. Maybe the creation of the session sets the specified graph to the default, but that's not the way tensorflow was designed to be used, so I wouldn't be surprised if it hasn't been thoroughly tested in this regime.<\/p>\n\n<p>While I don't have an explanation or what's going on, I can suggest a simple solution: separate graph construction with session running.<\/p>\n\n<pre><code>graph = tf.Graph()\nwith graph.as_default():\n    build_graph()\n    saver = tf.train.Saver()\n\nwith tf.Session(graph=graph) as sess:\n    do_stuff_with(sess)\n    saver.save(sess, path)\n<\/code><\/pre>\n"}]}
{"QuestionId":41866745,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-01-26T04:09:31.570","AcceptedAnswerId":41918083.0,"OwnerUserId":2059708.0,"Title":"Tensorflow shuffle_batch speed","Body":"<p>I noticed a big difference in speed if I load my training data into memory and feed it into the graph as a numpy array vs using a shuffle batch of the same size, my data has ~1000 instances. <\/p>\n\n<p>Using memory 1000 iterations takes less than a few seconds but using a shuffle batch it takes almost 10 minutes. I get the shuffle batch should be a bit slower but this seems way too slow. Why is this?<\/p>\n\n<p><strong>Added a bounty. Any suggestions on how to make shuffled mini-batches faster?<\/strong><\/p>\n\n<blockquote>\n  <p>Here is the training data: <a href=\"http:\/\/pastebin.com\/dNcqL9gN\" rel=\"nofollow noreferrer\">Link to bounty_training.csv (pastebin)<\/a><\/p>\n<\/blockquote>\n\n<p>Here is my code:<\/p>\n\n<blockquote>\n  <p>shuffle_batch<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>Full batch via numpy array<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41866982","CreationDate":"2017-01-26T04:38:14.357","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>In this case, you're running a session 3 times per step - once in <code>avgs_batch.eval<\/code>, once for <code>pdiff_batch.eval<\/code>, and once for the actual <code>sess.run<\/code> call. That doesn't explain the magnitude of the slow down, but it's definitely something you should keep in mind. At the very least the first two eval calls should be combined to one <code>sess.run<\/code> call.<\/p>\n\n<p>I suspect most of the slow-down is coming from use of <code>TFRecordReader<\/code>. I don't pretend to understand the inner workings of tensorflow, but you might find my answer <a href=\"https:\/\/stackoverflow.com\/questions\/41864333\/tensorflow-mixes-up-images-and-labels-when-making-batch-how-to-fix\/41864611#41864611\">here<\/a> helpful. <\/p>\n\n<p>Summary<\/p>\n\n<ul>\n<li>create minimal data associated with each example, i.e. image filenames, ids rather than entire images;<\/li>\n<li>convert to tensorflow ops with <code>tensorflow.python.framework.ops.convert_to_tensor<\/code>;<\/li>\n<li>use <code>tf.train.slice_input_producer<\/code> to get a tensor for a single example;<\/li>\n<li>do some preprocessing on individual examples - e.g. load images from filenames;<\/li>\n<li>batch them together using <code>tf.train.batch<\/code> to group them up.<\/li>\n<\/ul>\n"},{"AnswerId":"42953302","CreationDate":"2017-03-22T13:38:18.150","ParentId":null,"OwnerUserId":"7704013","Title":null,"Body":"<p>When using queues to get the data, you shouldn't use feed_dict. Instead, make your graph depend directly on the input data, that is:<\/p>\n\n<ul>\n<li>remove the X and Y PlaceHolders<\/li>\n<li><p>use your feature batch directly<\/p>\n\n<pre><code>hidden1 = tf.matmul(avgs_batch,W) + b\n<\/code><\/pre><\/li>\n<li><p>similarly, use the label batch (pdiff_batch) instead of Y when computing the loss<\/p><\/li>\n<li><p>finally, just keep the second session.run to compute the loss directly, and without using feed_dict<\/p>\n\n<pre><code># x_, y_ = sess.run([avgs_batch,pdiff_batch])\n# _, loss_val = sess.run([optimizer,loss],\n      feed_dict={X: x_, Y: y_.reshape(batch_size,1)} )\n\n_, loss_val = sess.run([optimizer,loss])\n<\/code><\/pre><\/li>\n<\/ul>\n"},{"AnswerId":"41918083","CreationDate":"2017-01-29T05:57:00.833","ParentId":null,"OwnerUserId":"2059708","Title":null,"Body":"<p>The trick is instead of feeding single examples into shuffle_batch you feed an n+1 dimensional tensor of examples to it with enqueue_many=True. I found this thread that was very helpful:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/41647784\/tfrecordreader-seems-extremely-slow-and-multi-threads-reading-not-working\">TFRecordReader seems extremely slow , and multi-threads reading not working<\/a><\/p>\n\n<pre><code>def get_batch(batch_size):\n    reader = tf.TFRecordReader()\n    _, serialized_example = reader.read(filename_queue)\n\n    batch_list = []\n    for i in range(batch_size):\n        batch_list.append(serialized_example)\n\n    return [batch_list]\n\nbatch_serialized_example = tf.train.shuffle_batch(\n get_batch(batch_size), batch_size=batch_size,\n    capacity=100*batch_size,\n    min_after_dequeue=batch_size*10,\n    num_threads=1, \n    enqueue_many=True)\n\nfeatures = tf.parse_example(\n    batch_serialized_example,\n    features={\n        'label': tf.FixedLenFeature([], np.float32),\n        'features': tf.FixedLenFeature([14], np.float32)})\n\nbatch_pdiff = features['label']\nbatch_avgs = features['features']\n\n...\n<\/code><\/pre>\n"}]}
{"QuestionId":41867191,"AnswerCount":4,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-26T05:01:32.977","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"How does one inspect variables in a checkpoint file in TensorFlow when TensorFlow can't find the tools attribute?","Body":"<p>I was trying to inspect checkpoints using the code at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/inspect_checkpoint.py\" rel=\"noreferrer\">inspect_checkpoint.py<\/a>. However, I wasn't able to have it work because they didn't really provide an example. I tried the simplest thing I thought would work:<\/p>\n\n<pre><\/pre>\n\n<p>however I get that  has no attribute :<\/p>\n\n<pre><\/pre>\n\n<p>it seems like a (embarrassingly) trivial bug\/issue. Does someone know what is going on? Why can't it find tools? Also, even if it did find it, how would one run the function provided in that file?<\/p>\n\n<hr>\n\n<p>Unfortunately, the very related question did not really provide an answer of how to get around this issue. The question is here <a href=\"https:\/\/stackoverflow.com\/questions\/38218174\/how-can-find-the-variable-names-that-saved-in-tensorflow-checkpoint\">How can find the variable names that saved in tensorflow checkpoint?<\/a><\/p>\n","answers":[{"AnswerId":"54193944","CreationDate":"2019-01-15T06:49:40.867","ParentId":null,"OwnerUserId":"420867","Title":null,"Body":"<p>You could also use the <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model#cli_to_inspect_and_execute_savedmodel\" rel=\"nofollow noreferrer\">command line interface<\/a> that uses <code>inspect_checkpoint<\/code>.<\/p>\n"},{"AnswerId":"41870281","CreationDate":"2017-01-26T09:25:09.553","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>Well, isn't inspect_checkpoint.py a binary?<\/p>\n\n<p>Something like this might work:<\/p>\n\n<pre><code>bazel run tensorflow\/python\/tools:inspect_checkpoint -- --file_name=YOUR_CKPT\n<\/code><\/pre>\n\n<p>EDIT:<\/p>\n\n<p>Or without bazel:<\/p>\n\n<p>Find where tensorflow is installed and run the command with <code>python<\/code>:<\/p>\n\n<pre><code>python PATH_TO_VENV\/lib\/python3.6\/site-packages\/tensorflow\/python\/tools\/inspect_checkpoint.py --file_name=YOUR_CKPT\n<\/code><\/pre>\n\n<p>For all options, see the file itself: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/inspect_checkpoint.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/inspect_checkpoint.py<\/a><\/p>\n"},{"AnswerId":"41917378","CreationDate":"2017-01-29T03:36:47.037","ParentId":null,"OwnerUserId":"1297184","Title":null,"Body":"<p>Try this:\n<\/p>\n\n<pre><code>from tensorflow.python.tools.inspect_checkpoint import print_tensors_in_checkpoint_file\nprint_tensors_in_checkpoint_file(file_name='.\/tmp\/mdl_ckpt', tensor_name='', all_tensors=False)\n<\/code><\/pre>\n\n<p>The <code>all_tensors<\/code> argument was added since <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/e0344f1e7ceff9c192369000d244c4ca47d7a1cc#diff-fb7984719b22f01a7748ef847e9eb731\" rel=\"noreferrer\">Tensorflow 0.12.0-rc0<\/a>.<\/p>\n"},{"AnswerId":"55615188","CreationDate":"2019-04-10T14:33:45.307","ParentId":null,"OwnerUserId":"5786339","Title":null,"Body":"<p>As of latest stable TensorFlow version, 1.13, and in the upcoming TF 2.0, to most straightforward way to inspect a checkpoint is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>path = '.\/tmp\/mdl_ckpt' \nget_checkpoint = tf.train.latest_checkpoint(path) \n#this retrieves the latest checkpoin file form path, but it also can be set manually\n\ninspect_list = tf.train.list_variables(get_checkpoint) \n<\/code><\/pre>\n\n<p>This creates a list of all variable names in given checkpoint<\/p>\n"}]}
{"QuestionId":41867993,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-01-26T06:33:40.700","AcceptedAnswerId":null,"OwnerUserId":6069513.0,"Title":"tensorflow one of 20 parameter server is very slow","Body":"<p>I am trying to train DNN model using tensorflow, my script have two variables, one is dense feature and one is sparse feature, each minibatch will pull full dense feature and pull specified sparse feature using embedding_lookup_sparse, feedforward could only begin after sparse feature is ready. I run my script using 20 parameter servers and increasing worker count did not scale out. So I profiled my job using tensorflow timeline and found one of 20 parameter server is very slow compared to the other 19. there is not dependency between different part of all the trainable variables. I am not sure if there is any bug or any limitation issues like tensorflow can only queue 40 fan out requests, any idea to debug it? Thanks in advance.\n<a href=\"https:\/\/i.stack.imgur.com\/7qwHG.png\" rel=\"nofollow noreferrer\">tensorflow timeline profiling<\/a><\/p>\n","answers":[{"AnswerId":"41893630","CreationDate":"2017-01-27T12:06:56.273","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Some suggestions that were too big for a comment:<\/p>\n\n<p>You can't see data transfer in timeline that's because the tracing of Send\/Recv is currently turned off, some discussion here -- <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4809\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4809<\/a>  <\/p>\n\n<p>In the latest version (nightly which is 5 days old or newer) you can turn on verbose logging by doing <code>export TF_CPP_MIN_VLOG_LEVEL=1<\/code> and it shows second level timestamps (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2076#issuecomment-275451855\" rel=\"nofollow noreferrer\">here<\/a> about higher granularity).<\/p>\n\n<p>So with vlog perhaps you can use messages generated by this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0a4f5b6bda405814af59803829216762e030728d\/tensorflow\/core\/common_runtime\/executor.cc#L1103\" rel=\"nofollow noreferrer\">line<\/a> to see the times at which <code>Send<\/code> ops are generated. <\/p>\n"},{"AnswerId":"41870204","CreationDate":"2017-01-26T09:20:44.813","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>It sounds like you might have exactly 2 variables, one is stored at PS0 and the other at PS1. The other 18 parameter servers are not doing anything. Please take a look at variable partitioning (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops\/variable_partitioners_for_sharding\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops\/variable_partitioners_for_sharding<\/a>), i.e. partition a large variable into small chunks and store them at separate parameter servers.<\/p>\n"},{"AnswerId":"42823232","CreationDate":"2017-03-16T00:28:50.650","ParentId":null,"OwnerUserId":"3117069","Title":null,"Body":"<p>This is kind of a hack way to log <code>Send\/Recv<\/code> timings from <code>Timeline<\/code> object for each iteration, but it works pretty well in terms of analyzing JSON dumped data (compared to visualize it on <code>chrome:\/\/trace<\/code>).<\/p>\n\n<p>The steps you have to perform are:<\/p>\n\n<ul>\n<li>download TensorFlow source and checkout a correct branch (r0.12 for example)<\/li>\n<li>modify the only place that calls <code>SetTimelineLabel<\/code> method inside <code>executor.cc<\/code>\n\n<ul>\n<li>instead of only recording non-transferable nodes, you want to record <code>Send\/Recv<\/code> nodes also.<\/li>\n<li>be careful to call <code>SetTimelineLabel<\/code> once inside <code>NodeDone<\/code> as it would set the text string of a node, which will be parsed later from a python script<\/li>\n<\/ul><\/li>\n<li>build TensorFlow from modified source<\/li>\n<li>modify model codes (for example, <code>inception_distributed_train.py<\/code>) with correct way of using <code>Timeline<\/code> and graph meta-data<\/li>\n<\/ul>\n\n<p>Then you can run the training and retrieve JSON file once for each iteration! :)<\/p>\n"}]}
{"QuestionId":41868343,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-26T07:07:11.510","AcceptedAnswerId":null,"OwnerUserId":541106.0,"Title":"Using a `tf.Tensor` as a Python `bool` is not allowed.","Body":"<p>I am new to tensorflow and I am now trying to calculate this value out of two sensors box1 and box2<\/p>\n\n<pre><\/pre>\n\n<p>I get <em>Using a  as a Python  is not allowed.<\/em> <\/p>\n\n<p>I understand tensors do not contain values until the session is not executed. My loss depends on this calculation and I can't see how to correctly express it.<\/p>\n","answers":[{"AnswerId":"41878750","CreationDate":"2017-01-26T16:52:24.483","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The Python <code>min()<\/code> and <code>max()<\/code> functions do not work on TensorFlow <code>tf.Tensor<\/code> objects.  Instead, use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/basic_math_functions#minimum\" rel=\"nofollow noreferrer\"><code>tf.minimum()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/basic_math_functions#maximum\" rel=\"nofollow noreferrer\"><code>tf.maximum()<\/code><\/a> ops as follows:<\/p>\n\n<pre><code>loss = tf.minimum(box1[0] + box1[2], box2[0] + box2[2]) - tf.maximum(box1[0], box2[0])\n<\/code><\/pre>\n"}]}
{"QuestionId":41868871,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-26T07:48:45.210","AcceptedAnswerId":null,"OwnerUserId":4482655.0,"Title":"Tensorflow Slim debugging during training","Body":"<p>I am largely following <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/slim_walkthough.ipynb\" rel=\"nofollow noreferrer\">this<\/a> for the inception model. Let's say I want to print the image data being used in the batch on every training loop i.e. I want to print the value of 'images' variable for every training iteration in the \n'Fine-tune the model on a different set of labels.' code, how do I do this since the session is actually being created inside the slim.training.train function hence I can't do a sess.run([images]) without sess here ?<\/p>\n","answers":[{"AnswerId":"41870063","CreationDate":"2017-01-26T09:11:23.453","ParentId":null,"OwnerUserId":"1029614","Title":null,"Body":"<p>So you want to visualize the training images? Well, you can add an image summary anywhere you have a handle to them, e.g. under \"Train the model on the Flowers dataset.\":<\/p>\n\n<pre><code>images, _, labels = load_batch(dataset)\ntf.summary.image(\"input\/images\", images)\n<\/code><\/pre>\n"},{"AnswerId":"44317326","CreationDate":"2017-06-01T21:25:28.147","ParentId":null,"OwnerUserId":"6440983","Title":null,"Body":"<p>With the following script you can see your images:\n<code>summaries.add(tf.summary.image('image_tensor,max_images=15,name='images'))<\/code><\/p>\n\n<p><code>tf.summary.image<\/code> is the easiest method because as mentioned session op is created inside <code>slim.learning.train<\/code>. Although interactivesession() can be used in the middle of the code, but it may disturb the session created by slim!<\/p>\n"},{"AnswerId":"49796587","CreationDate":"2018-04-12T12:28:46.837","ParentId":null,"OwnerUserId":"5295755","Title":null,"Body":"<p>Both precedent answers are very good to visualize images precisely.<\/p>\n\n<p>However, if you are interested to print any other variable value or even the pixel intensities of the image, you can use the following code, right before calling slim.learning.train for instance :<\/p>\n\n<pre><code>    variable = tf.get_default_graph().get_tensor_by_name(\"variable_name\")\n    train_tensor = tf.Print(train_tensor, [variable], text_to_print_before_value, summarize=number_of_values_to_print_if_array)\n<\/code><\/pre>\n\n<p>The first line gets the variable by its name (you can get it via tensorboard or by printing slim.get_model_variables() for instance).<\/p>\n\n<p>Second line prints it when train_tensor is evaluated and returns the train_tensor itself so that this Print node is added to the graph.<\/p>\n\n<p>An example printing a variable from a Mobilenet model:<\/p>\n\n<pre><code>    beta = tf.get_default_graph().get_tensor_by_name(\"MobilenetV1\/Conv2d_12_pointwise\/BatchNorm\/beta:0\")\n    train_tensor = tf.Print(train_tensor, [beta], \"beta \", summarize=256)\n<\/code><\/pre>\n"}]}
{"QuestionId":41869984,"AnswerCount":1,"Tags":"<python><deep-learning><keras><convolution>","CreationDate":"2017-01-26T09:06:18.340","AcceptedAnswerId":41870210.0,"OwnerUserId":3083243.0,"Title":"Keras - CNN Model Summary Diemension Interpretation","Body":"<p>I am using Keras library to build this deep learning model: INPUT(depth=1, height=15, width=27) -> CONV[depth=8](height=4, width=27) -> POOL(height=2, width=1) -> (Regression) output. <\/p>\n\n<p>I expect the ouput shape from convolution2d_1 to be (None, 8, 12, 1) and thence, the ouput shape from pooling2d_1 to be (None, 8, 6, 1); while I am getting (None, 8, 15, 27) and (None, 8, 7, 27) respectively.<\/p>\n\n<p>What am I doing or interpreting wrong here?<\/p>\n\n<p>P.S.: Also, this setting gives a Baseline Error: 99.23%! <\/p>\n\n<pre><\/pre>\n\n<p>=====================================<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41870210","CreationDate":"2017-01-26T09:21:03.907","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>Change <code>border_mode='same'<\/code> to <code>border_mode='valid'<\/code>. Border mode <code>same<\/code> adds zero padding to the input to make sure that the output of the convolutional layer has the same shape as its input. With border mode <code>valid<\/code> convolution is performed only where the input and the filter fully overlap.<\/p>\n"}]}
{"QuestionId":41870138,"AnswerCount":2,"Tags":"<python><numpy><deep-learning><keras>","CreationDate":"2017-01-26T09:16:15.900","AcceptedAnswerId":null,"OwnerUserId":6649170.0,"Title":"Keras MemoryError: alloc failed on Windows","Body":"<p>I wanted to train my net (image classification, unfortunately on CPU) and I have 71.000 records: 48x48 (gray-scale) images. (when I save it to a numpy array it is 1.4 Gb)<\/p>\n\n<p>After a few minutes I get this error message:<\/p>\n\n<pre><\/pre>\n\n<p>I use my laptop which has <em>8GB RAM<\/em> and when I saw the performance when the training runs and after a while it uses 100% of it.<\/p>\n\n<p>I have no idea how could I train my net now.<\/p>\n\n<p>This is the <strong>model structure<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p>I use a batch size of <em>64<\/em> and the train image size is <\/p>\n\n<p>How can I fix this error and train my network?<\/p>\n","answers":[{"AnswerId":"41870439","CreationDate":"2017-01-26T09:34:31.820","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>To reduce the memory requirement oyu can either reduce the size of the model (Dropout also requires memory equal to the previous Convolution layer) or you can reduce the batchsize. Anyways to train such a model on the cpu might take you days if not weeks. I would strongly suggest to calculate this on a gpu if possible. <\/p>\n"},{"AnswerId":"41903121","CreationDate":"2017-01-27T21:15:35.533","ParentId":null,"OwnerUserId":"6649170","Title":null,"Body":"<p>It turns out that <em>Theano<\/em> has memory leak. When I tried to train it with <em>Tensorflow<\/em> as Keras backend it was successful.<\/p>\n\n<p>So If you has a problem like that, change the backend (watch out for <code>dim_ordering<\/code>)<\/p>\n"}]}
{"QuestionId":41872069,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><conv-neural-network>","CreationDate":"2017-01-26T11:03:22.020","AcceptedAnswerId":null,"OwnerUserId":2598910.0,"Title":"Mixing categorical and continuous data in pure Tensorflow","Body":"<p>How do I take categorical and continuous data from a CSV file (such as the <a href=\"https:\/\/www.kaggle.com\/c\/titanic\" rel=\"nofollow noreferrer\">Titanic dataset<\/a>) and get it into a good format for Tensorflow without using the tf.learn API (i.e. pure Tensorflow)?<\/p>\n\n<p>For example, <\/p>\n\n<ul>\n<li>Categorical data might be (in my case) 'vehicle_make', 'vehicle_model'<\/li>\n<li>Continuous data might be 'price', 'mileage'<\/li>\n<\/ul>\n\n<p>I then want to use this data as input to a multilayer perceptron neural network as in the following example:<\/p>\n\n<p><a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3_NeuralNetworks\/multilayer_perceptron.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3_NeuralNetworks\/multilayer_perceptron.ipynb<\/a><\/p>\n\n<p>Cheers,<\/p>\n\n<p>Baz<\/p>\n","answers":[{"AnswerId":"41872659","CreationDate":"2017-01-26T11:39:31.160","ParentId":null,"OwnerUserId":"3032312","Title":null,"Body":"<p>You can use one-hot encoding for categorical data.<\/p>\n\n<p>Basically you can map vehicle_model to numerical value <\/p>\n\n<p>Tesla S -> 1<\/p>\n\n<p>Ford Focus -> 2<\/p>\n\n<p>... -> i<\/p>\n\n<p>And then use vector with all zeros and 1 in i-th position to represent that value:<\/p>\n\n<p>Tesla S -> [1, 0, 0]<\/p>\n\n<p>Ford Focus -> [0, 1, 0]<\/p>\n\n<p>both <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/array_ops\/slicing_and_joining#one_hot\" rel=\"nofollow noreferrer\">tensorflow<\/a> and <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.OneHotEncoder.html\" rel=\"nofollow noreferrer\">scikit learn<\/a> have functions to do that.<\/p>\n\n<p>Include continious data as-is, as one of values of input vector.<\/p>\n"}]}
{"QuestionId":41872217,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-26T11:11:35.077","AcceptedAnswerId":41881853.0,"OwnerUserId":2950581.0,"Title":"Understanding OpKernel::Compute and how tensorflow sets up GPU execution","Body":"<p>I'm having problems understanding the internals of TF (or rather have slowed down on progress). For the last three days I've been digging through the code (from the \"top\", going down). I've grasped graph creation and most of the stuff that happens before the OpKernel::Compute gets called. Just a fast summary (please correct me if I've got something important wrong): <\/p>\n\n<ol>\n<li>Define a graph wherever<\/li>\n<li>Call DirectSession::Run()\nInternally:<\/li>\n<li>The graph gets processed (split\/optimized\/etc...)<\/li>\n<li>Executors get created for each subgraph<\/li>\n<li>A RunState gets created<\/li>\n<li>Inputs get sent (\/stored) in the Rendevouz<\/li>\n<li>Executor::RunAsync() for all executors<\/li>\n<\/ol>\n\n<p>There:<\/p>\n\n<ol start=\"8\">\n<li>FillContextMap assigns contexts to all nodes<\/li>\n<li>The Process(TaggedNode,..) function gets scheduled to the ThreadPool for all root nodes<\/li>\n<\/ol>\n\n<p>There:<\/p>\n\n<ol start=\"10\">\n<li>The input tensors and parameters get prepared (and an OpKernel and OpContext get created)<\/li>\n<li>Device::Compute(kernel,context,..) gets called (sync or async)<\/li>\n<li>Device::Compute really just calls OpKernel::Compute from what I've seen<\/li>\n<\/ol>\n\n<p>So the problem is partially that I don't know how to go about the rest of the code (efficiently), and mostly that I can't grasp how OpKernels schedule instances themselves on the GPU and when. Prior to examining TF I read about GPGPU (the openCL docs) and I think I get the gist of it - work-items , groups, compute queues , synchronizations, memory management, types of memory on the physical device. But I can't seem to map this knowledge to how TF uses the GPU. As I said - I reached the OpKernel::Compute and there (in several types of kernels) I see only memory allocation, in some cases (matmul) CUBlass gets called (so I can't see what's happening), or there is just no GPU implementation . I expected to see some mechanism to assign a variable number of threads (work-items) to a kernel based on the analysis of the graph, some synchronization points being setup and so on.\n I'd be very thankful for any clarification on the above topics.<\/p>\n","answers":[{"AnswerId":"41881853","CreationDate":"2017-01-26T19:56:08.253","ParentId":null,"OwnerUserId":"2950581","Title":null,"Body":"<p>One more day on it and I think I got it:<\/p>\n\n<ol start=\"12\">\n<li>Device::Compute, when the device is a GPU device, has a different implementation: it checks if the inputs are in a context different from the on for this node (and wait for their completion via an event hook if they're not)<\/li>\n<li>Then the current stream gets changed (global for CUDA I think) and <\/li>\n<li>The OpKernel::Compute method gets called where if there's any kernel to be executed it is practically queued on the stream (which is the equivalent of a command queue in OpenCL).<\/li>\n<\/ol>\n"}]}
{"QuestionId":41874096,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><keras><generative>","CreationDate":"2017-01-26T13:04:41.043","AcceptedAnswerId":null,"OwnerUserId":6616602.0,"Title":"GAN Converges in Just a Few Epochs","Body":"<p>I implemented a genrative adversarial network in Keras. My training data size is about 16,000, where each image is of 32*32 size. All of my training images are the resized versions of the imageds from the imagenet dataset with regard to the object detection task. I fed the image matrix directly into the network without doing the center crop. I used the AdamOptimizer with the learning rate being 1e-4, and beta1 being 0.5 and I also set the dropout rate to be 0.1. I first trained the discrimator on 3000 real images and 3000 fake images and it achieved a 93% accuracy. Then, I trained for 500 epochs with the batch size being 32. However, my model seemed to converge in only a few epochs(&lt;10), and the images it generated were ugly. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kHZEi.png\" rel=\"nofollow noreferrer\">Plot of the Loss Function<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/D5AJ7.png\" rel=\"nofollow noreferrer\">Random Samples Generated by the Generator<\/a><\/p>\n\n<p>I was wondering whether my training dataset is too small(compared to those in the paper of DCGAN, which are more than 300,000) or my model configuration is not correct. What's more, should I train the SGD on D for k iterations (where k is small, perhaps 1) and then training with SGD on G for one iteration as suggested by Ian Goodfellow in the original paper?(I have just tried to train them one at a time)<\/p>\n\n<p>Below is the configuration of the generator.<\/p>\n\n<pre><\/pre>\n\n<p>Below is the configuration of the discriminator\uff1a<\/p>\n\n<pre><\/pre>\n\n<p>Below is the configuration of GAN as a whole:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"46301040","CreationDate":"2017-09-19T12:54:21.247","ParentId":null,"OwnerUserId":"5681394","Title":null,"Body":"<p>I suspect that your generator is trainable while you training the gan. You can verify by using <code>generator.layers[-1].get_weights()<\/code> to see if the parameters changed during training process of gan.<\/p>\n\n<p>You should freeze discriminator before you assemble it to gan:<\/p>\n\n<pre><code>generator.trainnable = False\ngan_input = Input(shape=[100])\nH = generator(gan_input)\ngan_V = discriminator(H)\nGAN = Model(gan_input, gan_V)\nGAN.compile(loss='categorical_crossentropy', optimizer=opt)\nGAN.summary()\n<\/code><\/pre>\n\n<p>see this discussion:\n<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/4674\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/4674<\/a><\/p>\n"}]}
{"QuestionId":41875067,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2017-01-26T13:55:44.860","AcceptedAnswerId":41877446.0,"OwnerUserId":7474000.0,"Title":"TensorFlow tf.equal() operator doesn't work as expected","Body":"<p>I wrote the following code, \nbut it doesn't work like I expect.\nI expected that  would be printed, but instead  is printed.\nCan you explain why this happens?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41877446","CreationDate":"2017-01-26T15:50:37.740","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops\/comparison_operators#equal\" rel=\"noreferrer\"><code>tf.equal()<\/code><\/a> operator is an <em>elementwise<\/em> operator. Assuming <code>x<\/code> and <code>y<\/code> are the same shape (as they are in your example) <code>tf.equal(x, y)<\/code> will produce a tensor with the same shape, where each element indicates whether the corresponding elements in <code>x<\/code> and <code>y<\/code> are equal. Therefore, <code>sess.run(tf.equal(y, y))<\/code> in your program will return the array <code>[True, True, True]<\/code>.<\/p>\n\n<p>In Python, the <code>is<\/code> operator computes reference equality between two objects, and the array <code>[True, True, True]<\/code> is not the same object as the (built-in) object <code>True<\/code>, so the result of the test is <code>False<\/code>. <\/p>\n\n<p>The following program would achieve the behavior you're expecting,<sup>*<\/sup> by using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/reduction#reduce_all\" rel=\"noreferrer\"><code>tf.reduce_all()<\/code><\/a> operator on the result of <code>tf.equal()<\/code> to compute a single boolean value:<\/p>\n\n<pre><code>y  = tf.constant([0, 1, 0])\nall_elems_equal = tf.reduce_all(tf.equal(y, y))\n\nwith tf.Session() as sess:\n    b = sess.run(all_elems_equal)\n    if b:\n        print 'true'\n    else:\n        print 'false'\n<\/code><\/pre>\n\n<p><sup>*<\/sup> Note however that <code>tf.equal(x, y)<\/code> will <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"noreferrer\">broadcast<\/a> its arguments if they have different shapes, so you might get the unexpected result that two tensors with different shapes are \"equal\" using this program. For example, using this test on the column vector <code>[[11], [22]]<\/code> and the row vector <code>[11, 22]<\/code> would indicate that these are equal. If you also need to compare the shapes in your equality test, you should also compare the result of <code>tf.shape(x)<\/code> and <code>tf.shape(y)<\/code>.<\/p>\n"}]}
{"QuestionId":41875915,"AnswerCount":1,"Tags":"<python><python-2.7><ubuntu><tensorflow><pip>","CreationDate":"2017-01-26T14:38:48.810","AcceptedAnswerId":44072647.0,"OwnerUserId":1179925.0,"Title":"Install tensorflow on Ubuntu 14.04","Body":"<p>I'm trying to follow <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#pip_installation\" rel=\"noreferrer\">these instructions<\/a> <\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<pre><\/pre>\n\n<p>Here is pip log:<\/p>\n\n<pre><\/pre>\n\n<p>So as I understand it needs some other python version? my python version is \n Python 2.7.6<\/p>\n\n<p>Also I tried (and it works):<\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>But I wonder why installation via  not working?<\/p>\n","answers":[{"AnswerId":"44072647","CreationDate":"2017-05-19T14:29:28.827","ParentId":null,"OwnerUserId":"8037205","Title":null,"Body":"<p>It's typically because you invoked a pip version lower than 8.1, run<\/p>\n\n<pre><code>pip install --upgrade pip\npip install --upgrade tensorflow\n<\/code><\/pre>\n\n<p><strong>If above code doesn't work<\/strong>, try installing TensorFlow in the active virtualenv environment by issuing a command of the following format:<\/p>\n\n<pre><code> (tensorflow)$ pip install --upgrade TF_PYTHON_URL   # Python 2.7\n (tensorflow)$ pip3 install --upgrade TF_PYTHON_URL  # Python 3.N \n<\/code><\/pre>\n\n<p>where TF_PYTHON_URL identifies the URL of the TensorFlow Python package. The appropriate value of TF_PYTHON_URLdepends on the operating system, Python version, and GPU support. Find the appropriate value for TF_PYTHON_URL for your system <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#the_url_of_the_tensorflow_python_package\" rel=\"noreferrer\">here<\/a>. For example, if you are installing TensorFlow for Linux, Python 2.7, and CPU-only support, issue the following command to install TensorFlow in the active virtualenv environment:<\/p>\n\n<pre><code>(tensorflow)$ pip3 install --upgrade \\\n https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.1.0-cp34-cp34m-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>More details is in <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#installing_with_virtualenv\" rel=\"noreferrer\">Tensorflow official website<\/a>.<\/p>\n"}]}
{"QuestionId":41876535,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-26T15:08:19.353","AcceptedAnswerId":null,"OwnerUserId":7348956.0,"Title":"Tensorflow: Using a value in a tensor as a parameter","Body":"<p>I want to calculate the loss function in my DNN in a different way depending on the value of the label.<\/p>\n\n<p>Conceptually it's something like this:<\/p>\n\n<pre><\/pre>\n\n<p>Obviously this won't work because I can't do this comparison on a tensor object. I also can't use , because I get an error that the network is not defined. Do I have another option?<\/p>\n","answers":[{"AnswerId":"41877444","CreationDate":"2017-01-26T15:50:31.520","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#cond\" rel=\"nofollow noreferrer\"><code>tf.cond<\/code><\/a> construct for this :<\/p>\n\n<pre><code>tf.cond(labels[0] == 0, lambda: loss_function_1(logits, labels),\n                        lambda: loss_function_2(logits, labels))\n<\/code><\/pre>\n"}]}
{"QuestionId":41877354,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><conv-neural-network>","CreationDate":"2017-01-26T15:46:35.157","AcceptedAnswerId":41878146.0,"OwnerUserId":5896419.0,"Title":"Training a Keras model with video input on a Macbook Pro 13 results in \"Killed: 9\" status","Body":"<p>The end goal of the system is to classify video input. The training data consists of 1-2 second 1080x1920 video sequences, which are preprocessed by transforming RGB frames to grayscale and ensuring video sequences have the same length by padding with blank frames. A series of 3D convolution layers connected to a fully-connected network is used. The <strong>problem<\/strong> is that, when the model is trained by calling the <strong>fit<\/strong> method, <strong>\"Killed: 9\"<\/strong> is returned and the program stops executing. Here is the code being used:<\/p>\n\n<pre><\/pre>\n\n<p>From what I understand, \"Killed: 9\" means that the program is consuming too much memory, and I only have 8 GB of RAM with my 13-inch Macbook Pro. I have tried reducing the batch size to 1, but even that does not solve the problem. Is there another way to train a Keras model to deal with this memory constraint?<\/p>\n","answers":[{"AnswerId":"41878146","CreationDate":"2017-01-26T16:22:21.777","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>These images are just too big for classification, and recent research never goes beyond 512x512 images (mostly for object detection). For classification the typical sizes are around 256x256 or 224x224.<\/p>\n\n<p>So your only option is to downscale the images to a reasonable size. Also reducing the batch size helps to reduce memory usage, but it won't work in this case.<\/p>\n"}]}
{"QuestionId":41878035,"AnswerCount":2,"Tags":"<tensorflow><keras>","CreationDate":"2017-01-26T16:17:19.290","AcceptedAnswerId":41879616.0,"OwnerUserId":1934212.0,"Title":"swap tensor axis in keras","Body":"<p>I want to swap tensor axis of image batches from (batch_size, row, col, ch) to \n(batch_size, ch, row, col). <\/p>\n\n<p>in numpy, this can be done with <\/p>\n\n<pre><\/pre>\n\n<p>How would I do that in Keras? <\/p>\n","answers":[{"AnswerId":"41879616","CreationDate":"2017-01-26T17:41:13.607","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can use <code>K.permute_dimensions()<\/code> which is exactly similar to <code>np.transpose()<\/code>.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>import numpy as np \nfrom keras import backend as K \n\nA = np.random.random((1000,32,64,3))\n# B = np.moveaxis( A, 3, 1)\nC = np.transpose( A, (0,3,1,2))\n\nprint A.shape\nprint C.shape\n\nA_t = K.variable(A)\nC_t = K.permute_dimensions(A_t, (0,3,1,2))\n\nprint K.eval(A_t).shape\nprint K.eval(C_t).shape\n<\/code><\/pre>\n"},{"AnswerId":"59082147","CreationDate":"2019-11-28T05:16:01.573","ParentId":null,"OwnerUserId":"3226167","Title":null,"Body":"<p>Use <code>keras.layers.Permute(dims)<\/code> where <code>dims<\/code> <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">does not include the samples dimension<\/a><\/p>\n\n<pre><code>model.add(Permute((2, 1), input_shape=(10, 64)))\n<\/code><\/pre>\n"}]}
{"QuestionId":41878052,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><keras>","CreationDate":"2017-01-26T16:18:16.683","AcceptedAnswerId":null,"OwnerUserId":1774961.0,"Title":"How to add more data during the training of a deep neural network using Keras or Tensorflow","Body":"<p>I'm trying to write a neural network. But I need to do something to the data during the training. For example, I have 5 convolution layers, after the 3rd layer, I need to change the shape of the output, and add more data to one of its dimension. Then the manipulated data will be fed into the 4th layer. <\/p>\n\n<p>The problem I have is, the input data dimension is not fixed, it should be got from the data in the training phase. When doing reshape, the Tensorflow will always give errors  because I put  for the dimension.<\/p>\n\n<p>In this case, is there any solution for this problem? Or could we divide the network into two parts, after the first part finished we manipulate the data and then continue to the second part?<\/p>\n\n<p>Thank you for your time!<\/p>\n\n<p><strong>Updates:<\/strong>\nFor example, the output of 3rd layer is , in which  is derived from the input and is not a fixed number. If we define  is the vectors of each point in  with the features as that in depth , i.e. .  I want to change it to a new shape , in which  is the same one, and in the depth dimension the data is from the concatenation of  and  plus  more new data.<\/p>\n","answers":[{"AnswerId":"42683766","CreationDate":"2017-03-08T23:16:26.747","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>You can get the runtime value for the shape of the input dimension from <code>tf.shape<\/code>, and then pass this instead of <code>None<\/code> to <code>tf.reshape<\/code>.<\/p>\n"}]}
{"QuestionId":41879440,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-26T17:30:54.200","AcceptedAnswerId":41899888.0,"OwnerUserId":2382483.0,"Title":"How to implement a sliding window in tensorflow?","Body":"<p>I have created a sliding window algorithm using numpy that slides over a wav audio file and feeds slices of it to my NN in tensorflow, which detects features in the audio slices. Once tensorflow does its thing, it returns its output to numpy land, where I reassemble the slices into an array of predictions that match each sample position of the original file:<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, this algorithm is quite slow. I placed some logs along the way and by far the slowest portion is the part where I actually run my tensorflow graph. This could be due to the actual tensorflow calculations being slow (if so, I'm probably just SOL), but I am wondering if a large part of the slowness isn't because I am transferring large audio files repeatedly back and forth in and out of tensorflow. So my questions are:<\/p>\n\n<p>1) Is feeding a placeholder repeatedly like this going to be noticeably slower than feeding it once and calculating the values for  inside tensorflow?<\/p>\n\n<p>2) If yes, whats the best way to implement a sliding window algorithm inside tensorflow to do this calculation?<\/p>\n","answers":[{"AnswerId":"41899888","CreationDate":"2017-01-27T17:39:43.053","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The first issue is that your algorithm is has quadratic time complexity in <code>window_size<\/code>, because of calling <code>np.hstack()<\/code> in each iteration to build the <code>output<\/code> array, which copies both the current values of <code>output<\/code> and <code>prediction<\/code> into a new array:<\/p>\n\n<pre><code>for w in range(window_size):\n    # ...\n    output = prediction if (output is None) else np.hstack((output, prediction))\n<\/code><\/pre>\n\n<p>Instead of calling <code>np.hstack()<\/code> in every iteration, it would be more efficient to build a Python list of the <code>prediction<\/code> arrays, and call <code>np.hstack()<\/code> on them once, after the loop terminates:<\/p>\n\n<pre><code>output_list = []\nfor w in range(window_size):\n    # ...\n    prediction = sess.run(...)\n    output_list.append(prediction)\noutput = np.hstack(output_list)\n<\/code><\/pre>\n\n<p>The second issue is that feeding large values to TensorFlow can be inefficient, if the amount of computation in the <code>sess.run()<\/code> call is small, because those values are (currently) copied into C++ (and the results are copied out. One useful strategy for this is to try and move the sliding window loop into the TensorFlow graph, using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/functional_ops\/higher_order_operators#map_fn\" rel=\"noreferrer\"><code>tf.map_fn()<\/code><\/a> construct. For example, you could restructure your program as follows:<\/p>\n\n<pre><code># NOTE: If you call this function often, you may want to (i) move the `np.pad()`\n# into the graph as `tf.pad()`, and (ii) replace `X_t` with a placeholder.\nX = np.pad(X, (pad_amount, 0), 'constant')\nX_t  = tf.convert_to_tensor(X)\n\ndef window_func(w):\n    start = w\n    end = w - window_size\n    X_matrix = tf.reshape(X_t[start:end], (-1, window_size))\n    return nn.forward_prop(X_matrix, Theta1, bias1, Theta2, bias2)\n\noutput_t = tf.map_fn(window_func, tf.range(window_size))\n# ...\noutput = sess.run(output_t)\n<\/code><\/pre>\n"}]}
{"QuestionId":41879770,"AnswerCount":1,"Tags":"<tensorflow><mnist>","CreationDate":"2017-01-26T17:49:12.267","AcceptedAnswerId":null,"OwnerUserId":6228958.0,"Title":"how to use MNIST datast on linux using tensorflow","Body":"<p>I'm new  in machine learning and I am following tensorflow's tutorial to create some simple Neural Networks which learn the MNIST data.<\/p>\n\n<p>i wanna run a code that do the recognition hand writing digits using the MNIST data but i don't know how to run it ... should i dowload the data on my machine and extracted and put it on a file and then set the path on the code or did tensorflow contain the data ...but when i do import input_data i get \nNo module named 'input_data' also when i do \nfrom tensorflow.examples.tutorials.mnist import input_data ==> No module named 'tensorflow.examples' \nps:when i do import tensorflow as tf i get no erreur so it's fine with tensorflow i think <\/p>\n\n<p>could u help me plz for example i wanna run the code below what should i do \n <a href=\"https:\/\/github.com\/hwalsuklee\/tensorflow-mnist-cnn\" rel=\"nofollow noreferrer\">https:\/\/github.com\/hwalsuklee\/tensorflow-mnist-cnn<\/a><\/p>\n","answers":[{"AnswerId":"44682232","CreationDate":"2017-06-21T17:06:35.660","ParentId":null,"OwnerUserId":"1042409","Title":null,"Body":"<p>If you cannot import <em>tensorflow.examples<\/em> I'm guessing something went wrong with the installation. Try reinstalling tensorflow with the latest version. \nYou don't need to download the data on your own, tensorflow will put it in the path you provide. But first, try these steps:<\/p>\n\n<p>I'm currently using tf 1.2.0 and I'm not getting that error.<\/p>\n\n<p>If you want to know which version you have installed:<\/p>\n\n<pre><code>import tensorflow as tf\nprint(tf.__version__)\n<\/code><\/pre>\n\n<p>After everything is installed try:<\/p>\n\n<pre><code>from tensorflow.examples.tutorials.mnist import input_data\ninput_data.read_data_sets(\".\/data\/\", one_hot=True)\n<\/code><\/pre>\n\n<p>That should copy the data to a \"data\" folder inside your working folder (the \"data\" folder will be created and all the files will be available there).<\/p>\n\n<p>If the above lines of code run with no errors, you should be able to run the example.<\/p>\n"}]}
{"QuestionId":41881099,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><data-science>","CreationDate":"2017-01-26T19:10:08.773","AcceptedAnswerId":null,"OwnerUserId":5368816.0,"Title":"Tensorflow tf.get_variable() by index","Body":"<p>Good afternoon.\nI continue to study tensorflow, and now stuck on a problem with reusing variable \"W\"\nHere's the code snippet:<a href=\"http:\/\/pastebin.com\/VZETt2ud\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/VZETt2ud<\/a><\/p>\n\n<p>I want to avoid hardcode and get values from restored model (instead of 10 - get_value(), etc).\nI have read several threads here, but everywhere only the whole varible was needed. However, I don't understand how to fetch properly, for instance, number 784 from here: <\/p>\n\n<pre><\/pre>\n\n<p>I tried this:<\/p>\n\n<pre><\/pre>\n\n<p>This approach gave me this mistake:\n\"Shape of a new variable (W) must be fully defined, but instead was unknown.\"<\/p>\n\n<p>(again, I avoid hardcode and cannot write the shape like [number, number])<\/p>\n\n<p>I changed the variables' scope, having thought it has smth to do with the scope, adding the lines:<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>But got this mistake:\n\"ValueError: Variable my\/W does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope?\"\nAfter setting reuse=None, I still have the same problem.<\/p>\n\n<p>Would you be so kind and advice me how to fetch the value by index in this code?<\/p>\n","answers":[{"AnswerId":"41899722","CreationDate":"2017-01-27T17:29:22.260","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>You need to create variable with get_variable for the first access too. For the later ones you need to set reuse=True.\nSomething like following should work:<\/p>\n\n<pre><code>W = tf.get_variable(\"W\", initializer=tf.zeros([784, 10], dtype=YOUR_DTYPE)\n...\ntemp_var = tf.get_variable(\"W\", reuse=True)\ntf.gather(...)\n<\/code><\/pre>\n"}]}
{"QuestionId":41881220,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-01-26T19:17:15.703","AcceptedAnswerId":null,"OwnerUserId":7470369.0,"Title":"keras predict always output same value in multi-classification","Body":"<p>here is my code. I want to made a multi-classification using Keras.  the vcl_acc is better when training, But the predict value always is a same value. I confused,please help me <\/p>\n\n<p>train.py<\/p>\n\n<pre><\/pre>\n\n<p>training output as follow<\/p>\n\n<pre><\/pre>\n\n<p>predict.py<\/p>\n\n<pre><\/pre>\n\n<p>the output always is same even I use different picture to test as follow<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"57184648","CreationDate":"2019-07-24T13:55:38.423","ParentId":null,"OwnerUserId":"11830394","Title":null,"Body":"<p>Well, after 1 month looking for a solution, i tried everything:\nlowering the learning rate, changing the optimizer, using a bigger dataset, increasing and decreasing model complexity, changing the input shape to smaller and larger images, changin the imports from <code>from keras import<\/code> to <code>from tensorflow.keras import<\/code> and further to <code>from tensorflow.python.keras import<\/code>, changing the activation function of every layer, combining them, trying others datasets, etc... Nothing helped. Even if i used a net like VGG16\/19 my results would be the same.\nBut, yesterday, i was reading a book (Deep Learning with Keras - Antonio Gulli, Sujit Pal) and i realized that the autors use imports like this:<\/p>\n\n<pre><code>from keras.layers.core import Dense, Flatten, Dropout\n<\/code><\/pre>\n\n<p>and not like this:<\/p>\n\n<pre><code>from keras.layers import Dense, Flatten, Dropout\n<\/code><\/pre>\n\n<p>the same for Conv, i was using:<\/p>\n\n<pre><code>from keras.layers import Conv2D, MaxPooling2D, SeparableConv2D\n<\/code><\/pre>\n\n<p>and the autors use:<\/p>\n\n<pre><code>from keras.layers.convolutional import Conv2D, MaxPooling2D, SeparableConv2D\n<\/code><\/pre>\n\n<p>And when i changed the imports everything started to work finally!\nI don't know if this is a bug or something like this, cause now my models always works even in datasets that it was predicting the same class.\nSo now i'm using the imports like this, for example:<\/p>\n\n<pre><code>from keras.layers.core import Dense, Dropout, Flatten\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D, SeparableConv2D\n<\/code><\/pre>\n\n<p>Try this, if doesn't work, look if your dataset is balanced, for example if you problem is to classify images of cats and dogs and you have 200 cat images and 800 dog images, try to use a number of images not so different cause it can cause problem: your model can 'think' well if i say that 10\/10 images are dogs so i get 80% accuracy, but this is not what we want. You can use <code>class_weight<\/code> if you don't have more images to balance you dataset and everything will be ok, you can as well use data-augmentation. You can use callbacks too, like <code>ReduceLROnPlateau<\/code> which reduce your learning rate when your loss is not getting lower. You can increase your <code>batch_size<\/code>, don't forget to shuffle the data on your <code>ImageDataGenerator<\/code> and normalize your images, like this:<\/p>\n\n<pre><code>g2 = ImageDataGenerator(rescale=1.\/255)\n<\/code><\/pre>\n\n<p>All these things are very important, but nothing really helped me. The only thing that worked was importing <code>keras.layers.core<\/code> and <code>keras.layers.convolutional<\/code>, may this help you too!<\/p>\n"},{"AnswerId":"41881502","CreationDate":"2017-01-26T19:33:30.017","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>It seems that your problem is caused by a huge class inbalance in your dataset. One can see that assigning a <code>0<\/code> class to each example gives you over 90% of accuracy. In order to deal with that you may use following strategies:<\/p>\n\n<ol>\n<li><p><strong>Rebalance your dataset:<\/strong> by either upsampling the less frequent class or downsampling the more frequent one.<\/p><\/li>\n<li><p><strong>Adjust class weights:<\/strong> by setting the higher class weight for less frequent class you'll be promoting your network training for putting more attention on the downsampled class.<\/p><\/li>\n<li><strong>Increase the time of training:<\/strong> in many cases - after longer period of training time network starts concentrating more on less frequent classes.<\/li>\n<\/ol>\n"}]}
{"QuestionId":41881308,"AnswerCount":2,"Tags":"<python><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2017-01-26T19:22:46.500","AcceptedAnswerId":42025194.0,"OwnerUserId":1931605.0,"Title":"How to calculate perplexity of RNN in tensorflow","Body":"<p>I'm running the word RNN implmentation of tensor flow of <a href=\"https:\/\/github.com\/hunkim\/word-rnn-tensorflow\">Word RNN<\/a><\/p>\n\n<p>How to calculate the perplexity of RNN.<\/p>\n\n<p>Following is the code in training that shows training loss and other things in each epoch:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42025194","CreationDate":"2017-02-03T13:14:02.457","ParentId":null,"OwnerUserId":"2628369","Title":null,"Body":"<p>The project you are referencing uses <code>sequence_to_sequence_loss_by_example<\/code>, which returns the cross-entropy loss. So for calculating the training perplexity, you just need to exponentiate the loss like explained <a href=\"https:\/\/en.wikipedia.org\/wiki\/Perplexity\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<pre><code>train_perplexity = tf.exp(train_loss)\n<\/code><\/pre>\n\n<p>We have to use e instead of 2 as a base, because TensorFlow measures the cross-entropy loss with the natural logarithm (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sigmoid_cross_entropy_with_logits\" rel=\"noreferrer\">TF Documentation<\/a>). Thank you, @Matthias Arro and @Colin Skow for the hint.<\/p>\n\n<h1>Detailed Explanation<\/h1>\n\n<p>The cross-entropy of two probability distributions P and Q tells us the minimum average number of bits we need to encode events of P, when we develop a coding scheme based on Q. So, P is the true distribution, which we usually don't know. We want to find a Q as close to P as possible, so that we can develop a nice coding scheme with as few bits per event as possible.<\/p>\n\n<p>I shouldn't say bits, because we can only use bits as a measure if we use base 2 in the calculation of the cross-entropy. But TensorFlow uses the natural logarithm, so instead let's measure the cross-entropy in <strong>nats<\/strong>.<\/p>\n\n<p>So let's say we have a bad language model that says every token (character \/ word) in the vocabulary is equally probable to be the next one. For a vocabulary of 1000 tokens, this model will have a cross-entropy of <em>log(1000) = 6.9 nats<\/em>. When predicting the next token, it has to choose uniformly between 1000 tokens at each step.<\/p>\n\n<p>A better language model will determine a probability distribution Q that is closer to P. Thus, the cross-entropy is lower - we might get a cross-entropy of 3.9 nats. If we now want to measure the perplexity, we simply exponentiate the cross-entropy: <\/p>\n\n<p><em>exp(3.9) = 49.4<\/em><\/p>\n\n<p>So, on the samples, for which we calculated the loss, the good model was as perplex as if it had to choose uniformly and independently among roughly 50 tokens.<\/p>\n"},{"AnswerId":"46606291","CreationDate":"2017-10-06T12:57:12.603","ParentId":null,"OwnerUserId":"507062","Title":null,"Body":"<p>It depends whether your loss function gives you a log likelihood of the data in base 2 or base e. This model is using legacy_seq2seq.sequence_loss_by_example, which uses TensorFlow's binary crossentropy, which <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\" rel=\"nofollow noreferrer\">appears to use logs of base e<\/a>. Therefore, even though we're dealing with a discrete probability distribution (text), we should exponentiate with e, i.e. use tf.exp(train_loss) as Colin Skow suggested.<\/p>\n"}]}
{"QuestionId":41881879,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-01-26T19:57:20.293","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"Where does keras store its data sets when using a docker container?","Body":"<p>I was trying to run deep learning models using the cifar10 data set. When keras is installed on the system and one tries to use the data set, keras downloads it once in:<\/p>\n\n<pre><\/pre>\n\n<p>I say that because I can ls that directory and I see:<\/p>\n\n<pre><\/pre>\n\n<p>however when I install keras in a docker container  doesn't exist. I am sure its being saved somewhere because I can later load the data set when I log in to the docker image through a container with bash (with ). However, I can't find here the files might be saved. Where are they saved?<\/p>\n\n<p>Also once I find this location, I was planning to volume my local  (or whatever computer I'm logged into) with the containers, so that the data sets don't have to be donwloaded 30 times every time I try to train some model.<\/p>\n\n<hr>\n\n<p>First progress I found where  is using the  command in the container (its in ). However, it seems that even though I volumed the folder correctly it downloads it anyway (even though I've went to the location  in the container and see the stuff there as it is in my local computer). I did:<\/p>\n\n<pre><\/pre>\n\n<p>and it still downloads the data set. Why?<\/p>\n\n<p>Another thing that strikes me a super odd is that when I bash into the container it starts me inside of root which it has never done that before.<\/p>\n","answers":[{"AnswerId":"41883923","CreationDate":"2017-01-26T22:12:37.170","ParentId":null,"OwnerUserId":"3167448","Title":null,"Body":"<p>Seems like I solved it. To find where <code>.keras<\/code> is in the container do once logged in to the container with bash ( as in <code>docker run -it --rm -v \/local:\/container image_name bash<\/code> for example) :<\/p>\n\n<pre><code>find -type d -name .keras \n<\/code><\/pre>\n\n<p>to discover that the location is <\/p>\n\n<pre><code>.\/root\/.keras\n<\/code><\/pre>\n\n<p>Now just volume the location of the data set:<\/p>\n\n<pre><code>docker run -it --rm -v \/Users\/folder1:\/Users\/folder1 -v \/Users\/Pinocchio\/.keras:\/root\/.keras tf_cpu cifar10_cnn.py\n<\/code><\/pre>\n\n<p>make sure to not have any typos, specially since its <code>.keras<\/code> and not <code>keras<\/code>.<\/p>\n"}]}
{"QuestionId":41882131,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-26T20:13:06.523","AcceptedAnswerId":null,"OwnerUserId":4833046.0,"Title":"tensorflow convnet batch normalization example","Body":"<p>Is there any documentation, example from tensorflow on how to use batch normalization layers in ConvNets?<\/p>\n\n<p>I have seen some examples on the internet but some are incomplete (<a href=\"https:\/\/gist.github.com\/tomokishii\/0ce3bdac1588b5cca9fa5fbdf6e1c412\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/tomokishii\/0ce3bdac1588b5cca9fa5fbdf6e1c412<\/a> refers to unavailable my_nn_lib) and others deal with fully connected layers only.<\/p>\n","answers":[]}
{"QuestionId":41882280,"AnswerCount":1,"Tags":"<neural-network><keras><training-data>","CreationDate":"2017-01-26T20:21:32.000","AcceptedAnswerId":41882456.0,"OwnerUserId":2145297.0,"Title":"What's the purpose of nb_epoch in Keras's fit_generator?","Body":"<p>It seems like I could get the exact same result by making num_samples bigger and keeping nb_epoch=1. I thought the purpose of multiple epochs was to iterate over the same data multiple times, but Keras doesn't reinstantiate the generator at the end of each epoch. It just keeps going. For example training this autoencoder:<\/p>\n\n<pre><\/pre>\n\n<p>It seems like I get the same result with (samples_per_epoch=100, nb_epoch=2) as I do for (samples_per_epoch=200, nb_epoch=1). Am I using fit_generator as intended?<\/p>\n","answers":[{"AnswerId":"41882456","CreationDate":"2017-01-26T20:33:00.270","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Yes - you are right that when using <code>keras.fit_generator<\/code> these two approaches are equivalent. But - there are variety of reasons why keeping <code>epochs<\/code> is reasonable:<\/p>\n\n<ol>\n<li><strong>Logging:<\/strong> in this case <code>epoch<\/code> comprises the amount of data after which you want to log some important statistics about training (like e.g. time or loss at the end of the epoch).<\/li>\n<li><strong>Keeping directory structure<\/strong> when you are using generator to load data from your hard disk - in this case - when you know how many files you have in your directory - you may adjust the <code>batch_size<\/code> and <code>nb_epoch<\/code> to such values that epoch would comprise going through every example in your dataset.<\/li>\n<li><strong>Keeping the structure of data when using <code>flow<\/code> generator<\/strong> - in this case, when you have e.g. a set of pictures loaded to your <code>Python<\/code> and you want to use <code>Keras.ImageDataGenerator<\/code> to apply different kind of data transformations, setting <code>batch_size<\/code> and <code>nb_epoch<\/code> in such way that epoch comprises going through every example in your dataset might help you in keeping track of a progress of your trainning process.<\/li>\n<\/ol>\n"}]}
{"QuestionId":41883800,"AnswerCount":1,"Tags":"<python-2.7><keras>","CreationDate":"2017-01-26T22:04:36.667","AcceptedAnswerId":41892935.0,"OwnerUserId":7476145.0,"Title":"Keras inception v3 retraining and finetuning error","Body":"<p>I am trying to make the example from here (<a href=\"https:\/\/keras.io\/applications\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/applications\/<\/a>) working since some hours and I am kind of getting crazy because it is not working... I would be very thankful if someone would have an idea what I could try! This is my example code:<\/p>\n\n<pre><\/pre>\n\n<p>When I run this I get the following error. I already tried to update pillow to the newest version but still the same error:<\/p>\n\n#\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41892935","CreationDate":"2017-01-27T11:27:29.600","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>There is a mismatch between the expected number of classes and the actual one, which is clear from the error message:<\/p>\n\n<pre><code>ValueError: Error when checking model target: expected dense_2 to have shape     (None, 200) but got array with shape (16, 2)\n<\/code><\/pre>\n\n<p>Here you specify that your model expects 200 classes, but in fact you have only 2.<\/p>\n\n<pre><code># and a logistic layer -- let's say we have 200 classes\npredictions = Dense(200, activation='softmax')(x)\n<\/code><\/pre>\n\n<p>Change it to <code>predictions = Dense(2, activation='softmax')(x)<\/code><\/p>\n"}]}
{"QuestionId":41884001,"AnswerCount":1,"Tags":"<python><classification><keras><object-detection>","CreationDate":"2017-01-26T22:19:19.827","AcceptedAnswerId":41894540.0,"OwnerUserId":4508333.0,"Title":"Keras Classification - Object Detection","Body":"<p>I am working on a classification then object detection with Keras and Python. I have classified cats\/dogs with 80%+ accuracy, Im ok with the current result for now. My question is how do I detect cat or dog from an input image? I'm completely confused. I want to use my own heights and not pretrained ones from internet.<\/p>\n\n<p>Here is my code currently:<\/p>\n\n<pre><\/pre>\n\n<p>So now since i classified cat and dog, how and what do I need to do to input an image and go through it to find cat or a dog in it with a bounding box? I'm completely new to this nd not even sure if I'm tackling this in a correct way?\nThank you.<\/p>\n\n<p><strong>UPDATE<\/strong>\nHi, Sorry to post results so late, was unable to work on this for few days.\nI am importing an image and reshaping it to 1,3,150,150 shape as 150,150 shape brings error:<\/p>\n\n<pre><\/pre>\n\n<p>Importing image:<\/p>\n\n<pre><\/pre>\n\n<p>Then I have changed def predict_function(x) to:<\/p>\n\n<pre><\/pre>\n\n<p>Now when I run:<\/p>\n\n<pre><\/pre>\n\n<p>I get output as best bounding box: None<\/p>\n\n<p>So I ran just:<\/p>\n\n<pre><\/pre>\n\n<p>And get the following out:<\/p>\n\n<pre><\/pre>\n\n<p>So it is not checking at all if its a cat or a dog... Any ideas? <\/p>\n\n<p>NOTE: when def predict)function(x) is using:<\/p>\n\n<pre><\/pre>\n\n<p>I do get the output as , it check boxes and gives the best one.<\/p>\n","answers":[{"AnswerId":"41894540","CreationDate":"2017-01-27T12:59:18.483","ParentId":null,"OwnerUserId":"7438048","Title":null,"Body":"<p>the machine learning model you built and the task you are trying to achieve are not the same. the model tries to solve a classification task while your goal is to detect an object inside the image, which is an <a href=\"https:\/\/en.wikipedia.org\/wiki\/Object_detection\" rel=\"noreferrer\">object detection task<\/a>.<\/p>\n\n<p>classification has a boolean question while detection quesion has more than two answers answers.<\/p>\n\n<h2>What can you do?<\/h2>\n\n<p>I can suggest you three possibilities to try:<\/p>\n\n<p><br \/><\/p>\n\n<h3>1. use sliding window combined with your model<\/h3>\n\n<p>crop boxes of defined sizes (e.g. from 20X20 to 160X160) and use sliding window. for each window, try to predict the probability its a dog and finally take the maximum window you predicted on.<\/p>\n\n<p>this will generate multiple candidates for the bounding box and you will choose the bounding box using the highest probability you got.<\/p>\n\n<p>this might be slow as we need to predict on hundreds+ samples.<\/p>\n\n<p>another option is to try implement <a href=\"https:\/\/cs.stanford.edu\/people\/karpathy\/rcnn\/\" rel=\"noreferrer\">RCNN<\/a> (<a href=\"https:\/\/github.com\/rbgirshick\/rcnn\" rel=\"noreferrer\">another link<\/a>) or <a href=\"https:\/\/arxiv.org\/abs\/1506.01497\" rel=\"noreferrer\">Faster-RCNN<\/a> network on top of your network. These networks are basically reducing the number of bounding box windows candidates to use.<\/p>\n\n<h2>Update - computing sliding window example<\/h2>\n\n<p>the following code demonstrate how to do the sliding window algorithm. you can change the parameters.<\/p>\n\n<pre><code>import random\nimport numpy as np\n\nWINDOW_SIZES = [i for i in range(20, 160, 20)]\n\n\ndef get_best_bounding_box(img, predict_fn, step=10, window_sizes=WINDOW_SIZES):\n    best_box = None\n    best_box_prob = -np.inf\n\n    # loop window sizes: 20x20, 30x30, 40x40...160x160\n    for win_size in window_sizes:\n        for top in range(0, img.shape[0] - win_size + 1, step):\n            for left in range(0, img.shape[1] - win_size + 1, step):\n                # compute the (top, left, bottom, right) of the bounding box\n                box = (top, left, top + win_size, left + win_size)\n\n                # crop the original image\n                cropped_img = img[box[0]:box[2], box[1]:box[3]]\n\n                # predict how likely this cropped image is dog and if higher\n                # than best save it\n                print('predicting for box %r' % (box, ))\n                box_prob = predict_fn(cropped_img)\n                if box_prob &gt; best_box_prob:\n                    best_box = box\n                    best_box_prob = box_prob\n\n    return best_box\n\n\ndef predict_function(x):\n    # example of prediction function for simplicity, you\n    # should probably use `return model.predict(x)`\n    random.seed(x[0][0])\n    return random.random()\n\n\n# dummy array of 256X256\nimg = np.arange(256 * 256).reshape((256, 256))\n\nbest_box = get_best_bounding_box(img, predict_function)\nprint('best bounding box %r' % (best_box, ))\n<\/code><\/pre>\n\n<p>example output:<\/p>\n\n<pre><code>predicting for box (0, 0, 20, 20)\npredicting for box (0, 10, 20, 30)\npredicting for box (0, 20, 20, 40)\n...\npredicting for box (110, 100, 250, 240)\npredicting for box (110, 110, 250, 250)\nbest bounding box (140, 80, 160, 100)\n<\/code><\/pre>\n\n<p><br \/><\/p>\n\n<h3>2. train new network for object detection task<\/h3>\n\n<p>you can take a look at the <a href=\"http:\/\/host.robots.ox.ac.uk\/pascal\/VOC\/voc2012\/index.html#data\" rel=\"noreferrer\"><strong>pascal dataset<\/strong><\/a> (<a href=\"http:\/\/host.robots.ox.ac.uk\/pascal\/VOC\/voc2012\/examples\/index.html\" rel=\"noreferrer\">examples here<\/a>) which contains 20 classes and two of them are cats and dogs.<\/p>\n\n<p>the dataset contains the location of the objects as the Y target.<\/p>\n\n<p><br \/><\/p>\n\n<h3>3. use existing network for this task<\/h3>\n\n<p>last but not least, you can reuse existing network or even do \"knowledge transfer\" (keras example here) for your specific task. <\/p>\n\n<p>take a look at the following <a href=\"https:\/\/github.com\/heuritech\/convnets-keras\" rel=\"noreferrer\"><code>convnets-keras<\/code><\/a> lib.<\/p>\n\n<p>so choose your best method to go and update us with the results.<\/p>\n"}]}
{"QuestionId":41884411,"AnswerCount":2,"Tags":"<apache-spark><machine-learning><tensorflow><linear-regression>","CreationDate":"2017-01-26T22:52:18.227","AcceptedAnswerId":41907029.0,"OwnerUserId":4115902.0,"Title":"Puzzled by Linear Regression Results from Spark ML and Tensorflow","Body":"<p>I ran the this program: <a href=\"https:\/\/github.com\/backstopmedia\/tensorflowbook\/blob\/master\/chapters\/04_machine_learning_basics\/linear_regression.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/backstopmedia\/tensorflowbook\/blob\/master\/chapters\/04_machine_learning_basics\/linear_regression.py<\/a><\/p>\n\n<p>I added \"print(\"w=\", W.eval(), \"b=\", b.eval())\" after line 55 in above program. The result I got is:<\/p>\n\n<p>w= [[ 3.5245235 ]\n[ 1.50171268]] b= 1.14499<\/p>\n\n<p>So  y=3.5245235x1 + 1.50171268x2 +1.14499.  <\/p>\n\n<p>I used the same data as above program (the file format is attached later), and ran the program <a href=\"https:\/\/github.com\/apache\/spark\/blob\/master\/examples\/src\/main\/java\/org\/apache\/spark\/examples\/ml\/JavaLinearRegressionWithElasticNetExample.java\" rel=\"nofollow noreferrer\">https:\/\/github.com\/apache\/spark\/blob\/master\/examples\/src\/main\/java\/org\/apache\/spark\/examples\/ml\/JavaLinearRegressionWithElasticNetExample.java<\/a>\nthe result is:<\/p>\n\n<p>Coefficients: [0.3827266230806965,5.1690760222564425] Intercept: 82.22008153614573\nnumIterations: 6\nobjectiveHistory: [0.5,0.41583549697777683,0.15548328325638935,0.15439025905767773,0.15432368309706285,0.15432368309449543]<\/p>\n\n<p>so y=0.3827266230806965x1+5.1690760222564425x2 + 82.22008153614573.<\/p>\n\n<p>I am confused how the results are so different for the same problem? The data format I used the Spark program is:<\/p>\n\n<pre>\n<\/pre>\n","answers":[{"AnswerId":"41884572","CreationDate":"2017-01-26T23:05:05.527","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>The answer is quite simple, the spark model <strong>is not<\/strong> linear regression.<\/p>\n\n<p>Linear regression minimizes || y - Wx ||_2^2<\/p>\n\n<p>Spark model is <strong>elastic net<\/strong> which minimizes  || y - Wx ||_2^2\n + a1 || W ||_2^2 + a2 || W ||_1<\/p>\n\n<p>If you want this spark code to be a linear regression <strong>remove<\/strong> the regularization terms:<\/p>\n\n<pre><code>  .setRegParam(0.3)\n  .setElasticNetParam(0.8);\n<\/code><\/pre>\n\n<p>and increase amount of iterations to make sure it converges.<\/p>\n"},{"AnswerId":"41907029","CreationDate":"2017-01-28T06:27:41.153","ParentId":null,"OwnerUserId":"4115902","Title":null,"Body":"<p>See <a href=\"https:\/\/stackoverflow.com\/questions\/37281777\/tensorflow-on-simple-linear-regression?rq=1\">Tensorflow on simple linear regression<\/a>. The code has the same issue. The fix will be the same as the answer there. Also, the learning_rate is too small (set it to .001) and steps need to be 100000. Choosing these initial values are very technical (one expert talked to me about this). <\/p>\n"}]}
{"QuestionId":41885519,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-27T00:48:19.483","AcceptedAnswerId":41900401.0,"OwnerUserId":7476177.0,"Title":"Tensorflow dynamic_rnn parameters meaning","Body":"<p>I'm struggling to understand the cryptic RNN docs. Any help with the following will be greatly appreciated. <\/p>\n\n<pre><\/pre>\n\n<p>I'm struggling to understand how these parameters relate to the mathematical LSTM equations and RNN definition. Where is the cell unroll size? Is it defined by the 'max_time'  dimension of the inputs? Is the batch_size only a convenience for splitting long data or it's related to minibatch SGD? Is the output state passed across batches?<\/p>\n","answers":[{"AnswerId":"41900401","CreationDate":"2017-01-27T18:10:17.253","ParentId":null,"OwnerUserId":"5235369","Title":null,"Body":"<p><code>tf.nn.dynamic_rnn<\/code> takes in a batch (with the minibatch meaning) of unrelated sequences. <\/p>\n\n<ul>\n<li><code>cell<\/code> is the actual cell that you want to use (LSTM, GRU,...)<\/li>\n<li><code>inputs<\/code> has a shape of <code>batch_size x max_time x input_size<\/code> in which max_time is the number of steps in the longest sequence (but all sequences could be of the same length)<\/li>\n<li><code>sequence_length<\/code> is a vector of size <code>batch_size<\/code> in which each element gives the length of each sequence in the batch (leave it as default if all your sequences are of the same size. <strong>This parameter is the one that defines the cell unroll size.<\/strong><\/li>\n<\/ul>\n\n<h3>Hidden state handling<\/h3>\n\n<p>The usual way of handling hidden state is to define an initial state tensor before the <code>dynamic_rnn<\/code>, like this for instance : <\/p>\n\n<pre><code>hidden_state_in = cell.zero_state(batch_size, tf.float32) \noutput, hidden_state_out = tf.nn.dynamic_rnn(cell, \n                                             inputs,\n                                             initial_state=hidden_state_in,\n                                             ...)\n<\/code><\/pre>\n\n<p>In the above snippet, both <code>hidden_state_in<\/code> and <code>hidden_state_out<\/code> have the same shape <code>[batch_size, ...]<\/code> (<em>the actual shape depends on the type of cell you use but the important thing is that the first dimension is the batch size<\/em>). <\/p>\n\n<p>This way, <code>dynamic_rnn<\/code> has an initial hidden state for each sequence. <strong>It will pass on the hidden state from time step to time step for each sequence in the <code>inputs<\/code> parameter on its own<\/strong>, and <code>hidden_state_out<\/code> will contain the final output state for each sequence in the batch. No hidden state is passed between sequences of the same batch, but only between time steps of the same sequence.<\/p>\n\n<h3>When do I need to feed back the hidden state manually?<\/h3>\n\n<p>Usually, when you're training, every batch is unrelated so you don't have to feed back the hidden state when doing a <code>session.run(output)<\/code>.<\/p>\n\n<p>However, if you're testing, and you need the output at each time step, (i.e. you have to do a <code>session.run()<\/code> at every time step) you'll want to evaluate and feed back the output hidden state using something like this :<\/p>\n\n<pre><code>output, hidden_state = sess.run([output, hidden_state_out],\n                                feed_dict={hidden_state_in:hidden_state})\n<\/code><\/pre>\n\n<p>otherwise tensorflow will just use the default <code>cell.zero_state(batch_size, tf.float32)<\/code> at each time step which equates to reinitialising the hidden state at each time step.<\/p>\n"}]}
{"QuestionId":41885665,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-27T01:05:55.360","AcceptedAnswerId":41886530.0,"OwnerUserId":3592550.0,"Title":"Maximum likelihood linear regression tensorflow","Body":"<p>After I implemented a LS estimation with gradient descent for a simple linear regression problem, I'm now trying to do the same with Maximum Likelihood.\n<img src=\"https:\/\/wikimedia.org\/api\/rest_v1\/media\/math\/render\/svg\/8d2b89d3967913366cdccf60d3f7f07642f8bfb1\">\nI used this equation from <a href=\"https:\/\/de.wikipedia.org\/wiki\/Lineare_Regression#Maximum-Likelihood-Sch.C3.A4tzung\" rel=\"nofollow noreferrer\">wikipedia<\/a>. The maximum has to be found.<\/p>\n\n<pre><\/pre>\n\n<p>This code works, but I still have some questions about it:<\/p>\n\n<ul>\n<li>If I change the lhf function from  to  and minimize  (according to the equation), it does not work. But why?<\/li>\n<li>The value for lhf goes up and down during optimization. Shouldn't it only change in one direction?<\/li>\n<li>The value for lhf sometimes is a NaN during optimization. How can I avoid that?<\/li>\n<li>In the equation, \u03c3\u00b2 is the variance of the error (right?). My values are perfectly on a line. Why do I get a value of var above 100?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"41886530","CreationDate":"2017-01-27T03:20:40.197","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The symptoms in your question indicate a common problem: the learning rate or step size might be too high for the problem.<\/p>\n\n<p>The zig-zag behaviour, where the function to be maximized goes up and down, is usual when the learning rate is too high. Specially when you get NaNs.<\/p>\n\n<p>The simplest solution is to lower the learning rate, by dividing your current learning rate by 10 until the learning curve is smooth and there are no NaNs or up-down behavior.<\/p>\n\n<p>As you are using TensorFlow you can also try AdamOptimizer as this adjust the learning rate dynamically as you train.<\/p>\n"}]}
{"QuestionId":41885913,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2017-01-27T01:41:58.507","AcceptedAnswerId":null,"OwnerUserId":7476456.0,"Title":"How to use TensorFlow LSTM tutorial for character-level language modeling?","Body":"<p>I am trying to implement RNN character-level language model from Andrej Karpathy's blog <a href=\"http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/\" rel=\"nofollow noreferrer\">http:\/\/karpathy.github.io\/2015\/05\/21\/rnn-effectiveness\/<\/a>\nusing TensorFlow RNN. For starters, I took ptb_word_lm.py from the LSTM tutorial in TensorFlow, and changed the reader function, so that now instead of the original PTB dataset with 10000 distinct words it is trained on a stream of 36 distinct characters. (The shape of embedding was reduced accordingly: to 36x36 instead of 10000x200). The training\/validation set was the text of War and Peace (3.5 million lowercase chars).<\/p>\n\n<p>I expected a straightforward carryover. \nInstead, after about 20 runs RNN gets stuck in a state with over-represented T's in the beginning and end of words, A sample output starts looking like this after 5 or so epoch runs:\nOF TrTSONAL TY tO tHE soS OF TOERE  AHME  AND TOmSEDtik  aN TEpoUNCENG THE soSECTIOaELING OF THE sNTIFiNDENTE OF TNE S sWN srrSONAL TY  TUT TN tF T  IauGR  aHE seW sisWoeID Tt WS tHuE thaT th wONNOT seeL THa sorEMENT OF THE sMRTH  AUT te T VITTENG Tt  sNPENelITY OiRhREAVEDTN T OORD<\/p>\n\n<p>(here capital letters stand for higher likelyhood). No further progress is made over the next 100+ epoch runs. Perplexities are quickly frozen around 4 for train set, and 3.75 for validation set.<\/p>\n\n<p>It is true that among all the character pairs in a typical text, \"space-T\" and \"T-space\" are the most frequent, but RNN is supposed to resolve such issues,right?  I tried varying parameters (ADAM for gradient, longer step, forget-bias=1.0, etc) - nothing changes.\nI am at the end of my rope. What could be going wrong here?<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[{"AnswerId":"42444067","CreationDate":"2017-02-24T16:50:10.463","ParentId":null,"OwnerUserId":"6291063","Title":null,"Body":"<p>Did you check your initialization? Learning can get stuck if the gradient is at zero. This would happen if you initialize weights (or bias) to either zeroes or ones. There are different initialization alternatives, depending on your activation (non-linearity). A well rounded initialization would be normal or truncated normal.<\/p>\n\n<p>e.g. <code>tf.truncated_normal([size_in, size_out], stddev=0.1)<\/code> <\/p>\n\n<p>for bias, small constant also works <code>tf.constant(0.1, shape=[size_out])<\/code><\/p>\n"}]}
{"QuestionId":41886136,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-27T02:15:45.797","AcceptedAnswerId":41900893.0,"OwnerUserId":6815505.0,"Title":"cost not updating in tensorflow","Body":"<p>I am using the following code from Sirajalogy. <a href=\"https:\/\/github.com\/llSourcell\/How_to_use_Tensorflow_for_classification-LIVE\/blob\/master\/demo.ipynbIt\" rel=\"nofollow noreferrer\">https:\/\/github.com\/llSourcell\/How_to_use_Tensorflow_for_classification-LIVE\/blob\/master\/demo.ipynbIt<\/a> \nIt has been modified to accept my own .csv with different dimensions than the one used in his example.<\/p>\n\n<pre><\/pre>\n\n<p>The code is running, but producing the following cost updates.<\/p>\n\n<pre><\/pre>\n\n<p>Question: Why are the costs not updating with each training step?\nThanks!<\/p>\n","answers":[{"AnswerId":"41900893","CreationDate":"2017-01-27T18:41:49.533","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>Problem: your gradients are zero, therefore your weights doesn't change. You provide single dimension (batch_size, 1) to softmax. This makes output of softmax a constant (1). This makes it's gradient be zero.<\/p>\n\n<p>Solutions:<\/p>\n\n<p>If you're doing a logistic regression, please use <code>tf.nn.sigmoid_cross_entropy_with_logits(y_values, y_)<\/code><\/p>\n\n<p>If you're doing a linear regression, please use (i.e. don't use softmax):\n<code>cost = tf.reduce_sum(tf.pow(y_ - y_values, 2))\/(2*n_samples)<\/code><\/p>\n\n<p>If you insist on mixing softmax and MSE, please use following instead of softmax:\n<code>y = tf.reciprocal(1 + tf.exp(-y_values))<\/code><\/p>\n"}]}
{"QuestionId":41886506,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-27T03:16:55.543","AcceptedAnswerId":42047142.0,"OwnerUserId":3296050.0,"Title":"How do I access the value of a label in Tensorflow during training\/eval?","Body":"<p>I'm working on a neural network in Tensorflow with a custom loss function, which based on the label of the input image I sample a vector (from another data set corresponding to the label) and take the dot product of the input image's embedding (the softmax preactivation) and the sampled vector. I also negatively sample a mismatch vector that conflicts with the input label, along with another random training input image embedding whose label differs from the current input label. Obviously these are all constructed as same-dimensional tensors, and the custom loss is:<\/p>\n\n<pre><\/pre>\n\n<p>That was mainly for background, but the issue I'm having is how to access the values of labels corresponding to the input images? I need to be able to access these labels' values in order to sample the correct vectors and compute my loss. <\/p>\n\n<p>I read in documentation that you can use .eval() to get the value of a tensor by running in a session, but when I tried this my terminal just hung... Technically I'm already running a session when I'm training my neural network, so not sure if there is an issue with running a second session inside another session and trying to evaluate values that are technically part of the other running session. Anyway, I'm completely out of ideas as to how I can make this work. Any help would be greatly appreciated! <\/p>\n\n<p>Here's my original attempt that proved problematic:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42047142","CreationDate":"2017-02-05T00:05:51.903","ParentId":null,"OwnerUserId":"3296050","Title":null,"Body":"<p>The issue was I was not using <code>feed_dict<\/code> to pass in the sampled values. I was trying to identify the value of the tensor for the input's label at the time of loss computation, in order to <em>then<\/em> sample the other values needed for my loss function. <\/p>\n\n<p>I have worked around this issue by precomputing and organizing the <code>mismatching_sample<\/code> and <code>matching_sample<\/code> values for example, using a <code>tf.placeholder<\/code> for these values initially, and passing in the values using <code>feed_dict<\/code>dictionary object when I execute <code>sess.run([train_op, loss...], feed_dict=feed_dict)<\/code> to access the values for my loss calculation.<\/p>\n"}]}
{"QuestionId":41886703,"AnswerCount":2,"Tags":"<azure><deep-learning><theano><azure-dsvm>","CreationDate":"2017-01-27T03:51:49.080","AcceptedAnswerId":null,"OwnerUserId":5960199.0,"Title":"Theano on Azure DSVM","Body":"<p>I am trying to use Theano on the Azure DSVM which is a preconfigured VM for Data Science. Is anyone aware of such a VM and does it support Theano out of the box? <\/p>\n\n<blockquote>\n  <p>WARNING (theano.configdefaults): g++ not detected ! Theano will be unable to execute optimized C-implementations (for both CPU and GPU) and will default to Python implementations. Performance will be severely degraded. To remove this warning, set Theano flags cxx to an empty string.<\/p>\n<\/blockquote>\n\n<p>Anyone with some experience in this regard?<\/p>\n","answers":[{"AnswerId":"43664914","CreationDate":"2017-04-27T18:00:34.847","ParentId":null,"OwnerUserId":"7385308","Title":null,"Body":"<p>We recently released a new version of the Data Science Virtual Machine for Linux (Ubuntu) edition which has Theano (and several other popular deep learning tools) pre-installed that you can use on either NC-Series GPU Azure VM or any CPU only VM. <\/p>\n\n<p>More info: <\/p>\n\n<p><a href=\"http:\/\/aka.ms\/dsvm\/ubuntu\" rel=\"nofollow noreferrer\">http:\/\/aka.ms\/dsvm\/ubuntu<\/a><\/p>\n\n<p><a href=\"http:\/\/aka.ms\/dsvm\/ubuntu\/docs\" rel=\"nofollow noreferrer\">http:\/\/aka.ms\/dsvm\/ubuntu\/docs<\/a><\/p>\n"},{"AnswerId":"43244485","CreationDate":"2017-04-06T02:09:14.087","ParentId":null,"OwnerUserId":"757590","Title":null,"Body":"<p>I have experienced the same issue since Theano has dependency on g++. So what i did is the following:<\/p>\n\n<p>1) Install anaconda from <a href=\"https:\/\/www.continuum.io\/downloads\" rel=\"nofollow noreferrer\">here<\/a>.\n3) Open anaconda command prompt and write the following command. This will install Theano with all its dependencies including g++ and others.<\/p>\n\n<pre><code>conda install theano\n<\/code><\/pre>\n"}]}
{"QuestionId":41888085,"AnswerCount":2,"Tags":"<keras><embedding><word2vec>","CreationDate":"2017-01-27T06:28:06.513","AcceptedAnswerId":41954253.0,"OwnerUserId":4535284.0,"Title":"How to implement word2vec CBOW in keras with shared Embedding layer and negative sampling?","Body":"<p>I want to create a word embedding pretraining network which adds something on top of word2vec CBOW. Therefore, I'm trying to implement word2vec CBOW first. Since I'm very new to <a href=\"https:\/\/keras.io\/\" rel=\"nofollow noreferrer\">keras<\/a>, I'm unable to figure out how to implement CBOW in it.<\/p>\n\n<p><strong>Initialization<\/strong>:<\/p>\n\n<p>I have calculated the vocabulary and have the mapping of word to integers.<\/p>\n\n<p><strong>Input to the (yet to be implemented) network<\/strong>:<\/p>\n\n<p>A list of  integers (representing the central word and  words in context)<\/p>\n\n<p><strong>Network Specification<\/strong><\/p>\n\n<p>A shared  layer should take this list of integers and give their corresponding vector outputs. Further a mean of  context vector is to be taken (I believe this can be done using ).<\/p>\n\n<p>It will be very helpful if anyone can share a small code-snippet of this.<\/p>\n\n<p><strong>P.S.<\/strong>: I was looking at <a href=\"https:\/\/github.com\/niitsuma\/word2vec-keras-in-gensim\" rel=\"nofollow noreferrer\">word2veckeras<\/a>, but couldn't follow its code because it also uses a gensim.<\/p>\n\n<p><strong>UPDATE 1<\/strong>:<\/p>\n\n<p>I want to share the embedding layer in the network. The embedding layer should be able to take context words (2*k) and the current word as well. I can do this by taking all 2*k + 1 word indices in the input and write a custom lambda function which will do the needful. But, after that I also want to add negative sampling network for which I'll have to take embedding of more words and dot product with the context vector. Can someone provide with an example where Embedding layer is a shared node in the  network<\/p>\n","answers":[{"AnswerId":"41954253","CreationDate":"2017-01-31T09:47:33.137","ParentId":null,"OwnerUserId":"4535284","Title":null,"Body":"<p><code>Graph()<\/code> has been deprecated from <code>keras<\/code><\/p>\n\n<p>Any arbitrary network can be created by using <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">keras functional API<\/a>.\nFollowing is the demo code which created a word2vec cbow model with negative sampling tested on randomized inputs<\/p>\n\n<pre><code>from keras import backend as K\nimport numpy as np\nfrom keras.utils.np_utils import accuracy\nfrom keras.models import Sequential, Model\nfrom keras.layers import Input, Lambda, Dense, merge\nfrom keras.layers.embeddings import Embedding\n\nk = 3 # context windows size\ncontext_size = 2*k\nneg = 5 # number of negative samples\n# generate weight matrix for embeddings\nembedding = []\nfor i in range(10):\n    embedding.append(np.full(100, i))\nembedding = np.array(embedding)\nprint embedding\n\n# Creating CBOW model\nword_index = Input(shape=(1,))\ncontext = Input(shape=(context_size,))\nnegative_samples = Input(shape=(neg,))\nshared_embedding_layer = Embedding(input_dim=10, output_dim=100, weights=[embedding])\n\nword_embedding = shared_embedding_layer(word_index)\ncontext_embeddings = shared_embedding_layer(context)\nnegative_words_embedding = shared_embedding_layer(negative_samples)\ncbow = Lambda(lambda x: K.mean(x, axis=1), output_shape=(100,))(context_embeddings)\n\nword_context_product = merge([word_embedding, cbow], mode='dot')\nnegative_context_product = merge([negative_words_embedding, cbow], mode='dot', concat_axis=-1)\n\nmodel = Model(input=[word_index, context, negative_samples], output=[word_context_product, negative_context_product])\n\nmodel.compile(optimizer='rmsprop', loss='mse', metrics=['accuracy'])\n\ninput_context = np.random.randint(10, size=(1, context_size))\ninput_word = np.random.randint(10, size=(1,))\ninput_negative = np.random.randint(10, size=(1, neg))\n\nprint \"word, context, negative samples\"\nprint input_word.shape, input_word\nprint input_context.shape, input_context\nprint input_negative.shape, input_negative\n\noutput_dot_product, output_negative_product = model.predict([input_word, input_context, input_negative])\nprint \"word cbow dot product\"\nprint output_dot_product.shape, output_dot_product\nprint \"cbow negative dot product\"\nprint output_negative_product.shape, output_negative_product\n<\/code><\/pre>\n\n<p>Hope it helps!<\/p>\n\n<h2>UPDATE 1:<\/h2>\n\n<p>I've completed the code and uploaded it <a href=\"https:\/\/github.com\/abaheti95\/Deep-Learning\/tree\/master\/word2vec\/keras\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"41891287","CreationDate":"2017-01-27T10:02:33.687","ParentId":null,"OwnerUserId":"7046074","Title":null,"Body":"<p>You could try something like this. Here I've initialized the embedding matrix to a fixed value. For an input array of shape <code>(1, 6)<\/code> you'll get the output of shape <code>(1, 100)<\/code> where the <code>100<\/code> is the average of the 6 input embedding.<\/p>\n\n<pre><code>model = Sequential()\nk = 3 # context windows size\ncontext_size = 2*k\n# generate weight matrix for embeddings\nembedding = []\nfor i in range(10):\n    embedding.append(np.full(100, i))\nembedding = np.array(embedding)\nprint embedding\n\nmodel.add(Embedding(input_dim=10, output_dim=100, input_length=context_size, weights=[embedding]))\nmodel.add(Lambda(lambda x: K.mean(x, axis=1), output_shape=(100,)))\n\nmodel.compile('rmsprop', 'mse')\n\ninput_array = np.random.randint(10, size=(1, context_size))\nprint input_array.shape\n\noutput_array = model.predict(input_array)\nprint output_array.shape\nprint output_array[0] \n<\/code><\/pre>\n"}]}
{"QuestionId":41889147,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-27T07:51:41.603","AcceptedAnswerId":null,"OwnerUserId":4375328.0,"Title":"tensorflow execution freezes for a small CNN","Body":"<p>I am trying to run some simple convolutional NN in tensorflow in Windows 10 with 40GB memory, using the CPU version. However, so far I keep running into trouble with the execution either hangs right after initializing variables or after a few training iterations. Below is a summary of my code and what I want to achieve.<\/p>\n\n<p>I have in each image a five-letter sequence, and I would like to train the CNN to recognize the sequence in each image. To do this I have two convolution layer (height\/width\/channel: 4\/4\/5, 4\/4\/10) each fed into a relu layer, followed by two fully connected relu layers, and finally a softmax layer with cross-entropy loss function.<\/p>\n\n<pre><\/pre>\n\n<p>The CNN is rather simple and fairly small, so I was rather surprised when I saw it froze after initializing all variables, or at best after a few training runs. There is no output of any kind, and ctrl+c would not interrupt the execution. I wonder if it could have anything to do with the Windows version of Tensorflow, but am currently at a loss where to look for clues. <\/p>\n\n<p>Could someone share their suggestion\/opinion on what may have caused my issue? Thank you!<\/p>\n\n<p>EDIT:\nAs pointed out in the comments, there may be an issue with the way I feed data to the model. Therefore I also post that part of code below.<\/p>\n\n<pre><\/pre>\n\n<p>I am simply using a random input and its labels to test if the code runs. Unfortunately the execution again freezes within the first few steps of the training.<\/p>\n","answers":[{"AnswerId":"41901239","CreationDate":"2017-01-27T19:05:05.337","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p>Most likely it is due to queue-runners.\nCould you please try to run your training step within MonitoredSession which handles initialization and queue running under the hood. It will look like as follows:<\/p>\n\n<pre><code>with tf.train.MonitoredTrainingSession(...) as sess:\n  sess.run(optimizer)\n<\/code><\/pre>\n"}]}
{"QuestionId":41889453,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-27T08:14:54.873","AcceptedAnswerId":41890543.0,"OwnerUserId":2677164.0,"Title":"How to load tensorflow graph from memory address","Body":"<p>I'm using the TensorFlow C++ API to load a graph from a file and execute it. Everything is working great, but I'd like to load the graph from memory rather than from a file (so that I can embed the graph into the binary for better portability). I have variables that reference both the binary data (as an unsigned char array) and the size of the data.<\/p>\n\n<p>This how I am currently loading my graph.<\/p>\n\n<pre><\/pre>\n\n<p>Feels like this should be simple but most of the discussion is about the python API. I did try looking for the source of  but wasn't able to find it in the tensorflow repo.<\/p>\n","answers":[{"AnswerId":"41890543","CreationDate":"2017-01-27T09:23:15.360","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>The following should work:<\/p>\n\n<pre><code>GraphDef graph_def;\nif (!graph_def.ParseFromArray(data, len)) {\n    \/\/ Handle error\n}\n...\n<\/code><\/pre>\n\n<p>This is because <code>GraphDef<\/code> is a sub-class of <code>google::protobuf::MessageList<\/code>, and thus inherits <a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/reference\/cpp\/google.protobuf.message_lite#parsing\" rel=\"nofollow noreferrer\">a variety of parsing methods<\/a><\/p>\n\n<p>Edit: Caveat: As of January 2017, the snippet above works only when the serialized graph is &lt;64MB because of a <a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/reference\/cpp\/google.protobuf.io.coded_stream#CodedInputStream.SetTotalBytesLimit.details\" rel=\"nofollow noreferrer\">default protocol buffer setting<\/a>. For larger graphs, take inspiration from <code>ReadBinaryProto<\/code>s implementation<\/p>\n\n<p>FWIW, the code for <code>ReadBinaryProto<\/code> is in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a304537954a865752ad1b18461e6bd67b36082db\/tensorflow\/core\/platform\/env.cc#L321\" rel=\"nofollow noreferrer\"><code>tensorflow\/core\/platform\/env.cc<\/code><\/a><\/p>\n"}]}
{"QuestionId":41889531,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-01-27T08:20:18.677","AcceptedAnswerId":null,"OwnerUserId":6375313.0,"Title":"Slice of a variable returns gradient None","Body":"<p>I've been playing around with the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/train\/gradient_computation#gradients\" rel=\"nofollow noreferrer\">tf.gradients()<\/a> function and came across a behavior I didn't expect. Namely it seems to be unable to calculate the gradient of a sliced Variable. I put together a example, that hopefully shows what I mean:<\/p>\n\n<pre><\/pre>\n\n<p>My questions are:<\/p>\n\n<p>1) Do I use the tf.gradients() function the way it is intended?<\/p>\n\n<p>2) If so, is there a reason for this behavior? In my understanding slicing should not necessarily break the backpropagation.<\/p>\n\n<p>3) Does that mean I need to avoid slicing within my entire network (or at least for every path from a Variable to the loss)? For example this would mean, that I must not slice the results of a fully connected layer into numerous meaningful parts (Like estimating multiple scalars with one fc layer and then slicing the joint estimation into parts I want to use).<\/p>\n\n<p>I'm working with Tensorflow 0.11 RC0 build from source on Ubuntu 16 with python 3.5.<\/p>\n","answers":[{"AnswerId":"41901640","CreationDate":"2017-01-27T19:31:03.950","ParentId":null,"OwnerUserId":"1454804","Title":null,"Body":"<p><code>d = c[:, ]<\/code> creates a different tensor then <code>a, b, c<\/code>. If you consider dependency graph, d depends on c. then gradients doesn't work in this case. <code>grad(y, x)<\/code> works if x depends on y, not the other way around. <\/p>\n"}]}
{"QuestionId":41890181,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-01-27T09:02:11.267","AcceptedAnswerId":null,"OwnerUserId":2324298.0,"Title":"Using TensorFlow's TFRecordReader","Body":"<p>I have a  file and want to view the contents of the file using tensorflow's . I want to display the contents of the file in the command prompt but haven't gotten anywhere. Any directions would help<\/p>\n","answers":[{"AnswerId":"41960114","CreationDate":"2017-01-31T14:31:41.487","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>If you are looking for information on how to use the TFRecordReader class, this example has some code that uses it:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py<\/a>\nwhich is linked from\n<a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/reading_data\/<\/a>\n. The reference documentation is here:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/io_ops\/readers\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.12\/api_docs\/python\/io_ops\/readers<\/a>\n.<\/p>\n\n<p>If you just want to view the contents of a TFRecord file as bytes, the tf.python_io.tf_record_iterator class should do what you want. It is defined here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/lib\/io\/tf_record.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/lib\/io\/tf_record.py<\/a><\/p>\n\n<p>Here's an example of code that uses it:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/gcs_test\/python\/gcs_smoke.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/gcs_test\/python\/gcs_smoke.py<\/a><\/p>\n"}]}
{"QuestionId":41890404,"AnswerCount":0,"Tags":"<c++><deep-learning><benchmarking><caffe>","CreationDate":"2017-01-27T09:14:42.243","AcceptedAnswerId":null,"OwnerUserId":2927593.0,"Title":"Caffe Batch processing no speedup","Body":"<p>I would like to speedup the forward pass of classification of a CNN using caffe.<\/p>\n\n<p>I have tried batch classification in Caffe using code provided in here:\n<a href=\"https:\/\/stackoverflow.com\/questions\/32668699\/modifying-the-caffe-c-prediction-code-for-multiple-inputs\">Modifying the Caffe C++ prediction code for multiple inputs<\/a>\nThis solution enables me to give a vector of Mat, but it does not speed up anything. Even though the input layer is modified.<\/p>\n\n<p>I am processing pretty small images (3x64x64) on a powerful pc with two GTX1080, and there is no issue in terms of memory.\nI tried also changing the deploy.prototxt, but I get the same result.<\/p>\n\n<p>It seems that at one point the forward pass of the CNN becomes sequential. \nI have seen someone pointing this out here also:\n<a href=\"https:\/\/stackoverflow.com\/questions\/32522469\/batch-processing-mode-in-caffe-no-performance-gains\">Batch processing mode in Caffe - no performance gains<\/a><\/p>\n\n<p>Another similar thread, for python : <a href=\"https:\/\/stackoverflow.com\/questions\/37003575\/batch-size-does-not-work-for-caffe-with-deploy-prototxt?rq=1\">batch size does not work for caffe with deploy.prototxt<\/a><\/p>\n\n<p>I have seen some things about MemoryDataLayer, but I am not sure this will solve my problem.<\/p>\n\n<p>So I am kind of lost on what to do exactly... does anyone have any information on how to speedup classification time.\nThanks for any help !<\/p>\n","answers":[]}
{"QuestionId":41890549,"AnswerCount":3,"Tags":"<cuda><tensorflow><nvidia>","CreationDate":"2017-01-27T09:23:44.980","AcceptedAnswerId":42405657.0,"OwnerUserId":5016028.0,"Title":"Tensorflow cannot open libcuda.so.1","Body":"<p>I have a laptop with a GeForce 940 MX. I want to get Tensorflow up and running on the gpu. I installed everything from their tutorial page, now when I import Tensorflow, I get<\/p>\n\n<pre><\/pre>\n\n<p>after which I think it just switches to running on the cpu. <\/p>\n\n<p>EDIT: After I nuked everything , started from scratch. Now I get this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"52770816","CreationDate":"2018-10-12T01:04:37.953","ParentId":null,"OwnerUserId":"4549682","Title":null,"Body":"<p>In the case I just solved, it was updating the GPU driver to the latest and installing the cuda toolkit. First, the ppa was added and GPU driver installed:<\/p>\n\n<pre><code>sudo add-apt-repository ppa:graphics-drivers\/ppa\nsudo apt update\nsudo apt install nvidia-390\n<\/code><\/pre>\n\n<p>After adding the ppa, it showed options for driver versions, and 390 was the latest 'stable' version that was shown.<\/p>\n\n<p>Then install the cuda toolkit:<\/p>\n\n<pre><code>sudo apt install nvidia-cuda-toolkit\n<\/code><\/pre>\n\n<p>Then reboot:<\/p>\n\n<pre><code>sudo reboot\n<\/code><\/pre>\n\n<p>It updated the drivers to a newer version than the 390 originally installed in the first step (it was 410; this was a p2.xlarge instance on AWS).<\/p>\n"},{"AnswerId":"53307646","CreationDate":"2018-11-14T19:40:17.453","ParentId":null,"OwnerUserId":"6713869","Title":null,"Body":"<p>In case anyone still encounters this. First make sure to add the <code>--runtime=nvidia<\/code> parameter in order to run your container.<\/p>\n\n<pre><code>docker run --runtime=nvidia -t tensorflow\/serving:latest-gpu\n<\/code><\/pre>\n\n<p>where <code>tensorflow\/serving:latest-gpu<\/code> is the name of the docker image.<\/p>\n"},{"AnswerId":"42405657","CreationDate":"2017-02-23T01:53:39.463","ParentId":null,"OwnerUserId":"4058925","Title":null,"Body":"<p>libcuda.so.1 is a symlink to a file that is specific to the version of your NVIDIA drivers.  It may be pointing to the wrong version or it may not exist.<\/p>\n\n<pre><code># See where the link is pointing.  \nls  \/usr\/lib\/x86_64-linux-gnu\/libcuda.so.1 -la\n# My result:\n# lrwxrwxrwx 1 root root 19 Feb 22 20:40 \\\n# \/usr\/lib\/x86_64-linux-gnu\/libcuda.so.1 -&gt; .\/libcuda.so.375.39\n\n# Make sure it is pointing to the right version. \n# Compare it with the installed NVIDIA driver.\nnvidia-smi\n\n# Replace libcuda.so.1 with a link to the correct version\ncd \/usr\/lib\/x86_64-linux-gnu\nsudo ln -f -s libcuda.so.&lt;yournvidia.version&gt; libcuda.so.1\n<\/code><\/pre>\n\n<p>Now in the same way, make another symlink from libcuda.so.1 to a link of the same name in your <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-linux\/#environment-setup\" rel=\"noreferrer\">LD_LIBRARY_PATH directory<\/a>.<\/p>\n\n<p>You may also find that you need to create a link to libcuda.so.1 in \/usr\/lib\/x86_64-linux-gnu named libcuda.so<\/p>\n"}]}
{"QuestionId":41892027,"AnswerCount":0,"Tags":"<tensorflow><gradient>","CreationDate":"2017-01-27T10:40:36.453","AcceptedAnswerId":null,"OwnerUserId":4844184.0,"Title":"Modifying intermediate gradients Tensorflow","Body":"<p>You can get intermediate gradients with and you can create a new tensor by applying an op on this results (like clipping) but how to modify the backpropagation accordingly ?<\/p>\n\n<p>For instance to implement the <a href=\"https:\/\/en.wikipedia.org\/wiki\/Huber_loss\" rel=\"nofollow noreferrer\">Huber loss<\/a> (with delta=1). \nThe first method is to create a boolean mask on the batch dimension doing something like.<\/p>\n\n<pre><\/pre>\n\n<p>A simpler way to implement it would be to use the l2 loss and to clip its gradients wrt inputs to 1.<\/p>\n\n<pre><\/pre>\n\n<p>But when you train your network you have to use  and , which only give you gradients wrt variables. How to make your optimizer use the tensor  when doing the chain rule ?<\/p>\n\n<p>Do you have to use gradient_override_map as in this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1095\" rel=\"nofollow noreferrer\">github issue<\/a> ?\nIs there a simpler way without registering a new op\/gradients ?<\/p>\n","answers":[]}
{"QuestionId":41892191,"AnswerCount":1,"Tags":"<tensorflow><lstm><recurrent-neural-network>","CreationDate":"2017-01-27T10:49:35.847","AcceptedAnswerId":41960518.0,"OwnerUserId":5762613.0,"Title":"Tensorflow: tf.nn.dynamic_rnn() :Unable to gather output values from last dimension of dynamic time-major","Body":"<p>I am trying to classify news articles using RNN. Since the length of news articles are not fixed I am using tf.nn.dynamic_rnn().<\/p>\n\n<pre><\/pre>\n\n<p>The tensor shape of 'outputs' from the code above is (batch_size, ?, lstm_size).<\/p>\n\n<p>I want to gather the output at the end of sentences, which is dynamic. I am using the following command for that<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the following error, <\/p>\n\n<pre><\/pre>\n\n<p>I believe this error is because of dynamic shape of outputs across time_major(sentence_major).<\/p>\n\n<p>In other words, result of <strong>\"outputs.get_shape()[0]\" is \"?\"(None)<\/strong><\/p>\n\n<p>Above technique for getting last output works when we use fixed time_major(sentence length). <\/p>\n\n<p><strong>Is there a way to achieve it for dynamic time_major(sentence length)?<\/strong><\/p>\n\n<p>As of now, I am doing the following<\/p>\n\n<pre><\/pre>\n\n<p>But my understanding is, by doing the mean across the time_major(sentence length), I am not using the potential of RNN of finding the sequential pattern. <strong>Please let me know your view on the same.<\/strong><\/p>\n","answers":[{"AnswerId":"41960518","CreationDate":"2017-01-31T14:50:02.970","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>In general, <code>get_shape()<\/code> is best-effort. Tensorflow does not always know the shape of a Tensor before the graph runs.<\/p>\n\n<p>There are a number of things you could try. One is to compute the offset of the last index yourself in Python without using get_shape; if you know the sizes of the inputs, this should not be difficult.<\/p>\n\n<p>Another option would be to use Tensorflow's slice functionality, which supports a Numpy-style \"-1\" index to represent the last element. For example, if <code>x<\/code> is a 3D Tensor, <code>x[:, -1, :]<\/code> should slice out the last element of the middle dimension.<\/p>\n\n<p>For more documentation, see the <code>tf.Tensor.__getitem__<\/code> documentation here:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Tensor\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/core_graph_data_structures#Tensor<\/a><\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41892246,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-01-27T10:52:48.063","AcceptedAnswerId":null,"OwnerUserId":2439854.0,"Title":"Keras unexpected keyword argument 'dim_ordering'","Body":"<p>I am developing in Keras with tensorflow backend. My program runs smoothly on two machine but the third. It gives me the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>I updated keras and tensorflow to the most recent version but the error still occurs. Here is the corresponding code segment:<\/p>\n\n<pre><\/pre>\n\n<p>The weird thing is that from the error message it is clear that the line:<\/p>\n\n<pre><\/pre>\n\n<p>in keras\/engine\/topology.py calls the my_init method but there is no parameter dim_ordering which is passed the my_init method<\/p>\n\n<p>BTW the keras.json is the same on all three machin:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41892636,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-27T11:12:02.303","AcceptedAnswerId":41892878.0,"OwnerUserId":5674154.0,"Title":"Indexing subtensors in TensorFlow","Body":"<p>I want to get an output tensor by applying  to <strong>subtensors<\/strong> of an input tensor.<\/p>\n\n<p>For example in NumPy,<\/p>\n\n<pre><\/pre>\n\n<p>gives me desired output (I want symbolic version of this). <\/p>\n\n<p>Similiarly in Theano,<\/p>\n\n<pre><\/pre>\n\n<p>gives me a desired .<\/p>\n\n<p>How can I do this in <strong>TensorFlow<\/strong>?<\/p>\n\n<pre><\/pre>\n\n<p>Unlike the example with NumPy, the length of  (=1st dimension of ) is not fixed.<\/p>\n","answers":[{"AnswerId":"41892878","CreationDate":"2017-01-27T11:24:21.877","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can do the slicing with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#gather_nd\" rel=\"nofollow noreferrer\"><code>tf.gather_nd<\/code><\/a>:<\/p>\n\n<pre><code>output = tf.gather_nd(input, tf.stack((tf.range(tf.shape(index)[0]), index), -1))\n<\/code><\/pre>\n"}]}
{"QuestionId":41893565,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><conv-neural-network>","CreationDate":"2017-01-27T12:03:09.677","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"How to fix my accuracy function in `tensorflow`?","Body":"<p>I have error when evaluating .<\/p>\n\n<p>This is my whole code:<\/p>\n\n<pre><\/pre>\n\n<p>And  is:<\/p>\n\n<pre><\/pre>\n\n<p>When i call print(computer_accuracy) the error appears.<\/p>\n\n<p><strong>The error is<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I do not know why  is like that and prediction go wrong. can anyone help me to fix it ?<\/p>\n","answers":[{"AnswerId":"41899303","CreationDate":"2017-01-27T17:04:17.887","ParentId":null,"OwnerUserId":"5903959","Title":null,"Body":"<p>TensorFlow's philosophy is different than pure Python. First, you <strong>define<\/strong> <em>tensors<\/em>, then you <strong>run<\/strong> it.<\/p>\n\n<p>The first part, where you <strong>define<\/strong> <code>prediction<\/code>, nothing is running. Its declarative. You declare that <code>prediction<\/code> depends on different variables\/operation. <\/p>\n\n<p>The calculation only occurs when you call <code>sess.run<\/code>. Thus, <code>y_pres<\/code> is not a tensor, it is a python variable (numpy probably) holding the results. Therefore, you are not supposed to pass it to <code>TF<\/code> functions. <\/p>\n\n<p><strong>So what?<\/strong> So you should separate Tensorflow declarative part from the running part. Which means, only keep your second <code>sess.run<\/code> call, everything between this one and the first <code>sess.run<\/code> belongs to declarative part. It should look a bit like:<\/p>\n\n<p><strong>Disclaimer: the code below won't even work, cf Problem #1 and maybe other issues<\/strong><\/p>\n\n<p><strong>declarative part (defining tensors i.e. TF's Compute Graph):<\/strong><\/p>\n\n<pre><code>  v_ys = tf.Placeholder( &lt;type&gt; )  \n\n  W_fc1 = weight_variable([14*14*32,256])\n  b_fc1 = bias_variable([256])\n  h_pool4_flat = tf.reshape(h_pool4,[-1,14*14*32])\n  h_fc1 = tf.nn.relu(tf.matmul(h_pool4_flat,W_fc1) + b_fc1)\n  h_fc1_drop = tf.nn.dropout(h_fc1,keep_prob)\n  W_fc2 = weight_variable([256,8])\n  b_fc2 = bias_variable([8])\n  prediction = tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2) + b_fc2)\n\n  # Defining correct prediction as a Tensor as well\n  correct_prediction = tf.equal(tf.argmax(prediction,1), tf.argmax(v_ys,1))\n  accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n<\/code><\/pre>\n\n<p><strong>Running the Model:<\/strong><\/p>\n\n<pre><code>def compute_accuracy(imageTestName , imageTestLabel ):\n      global prediction\n      v_xs = np.empty([1,224,224], dtype=np.int)\n      g = imageTestName[0]\n      v_xs = np.array(Image.open(path+g))\n      v_xs = np.reshape(v_xs,[1,224,224])\n      v_ys = imageTestLabel[0] ## It is a list = ['0', '0', '0', '0', '0', '0', '1', '0']\n      v_ys = np.reshape(v_ys , [1,8]) ## After reshape it is = [['0' '0' '0' '0' '0' '0' '1' '0']]\n\n      # Directly evaluate `accuracy`\n      result = sess.run(accuracy, feed_dict={xs: v_xs, ys: v_ys, keep_prob: 1})\n      return result\n<\/code><\/pre>\n"}]}
{"QuestionId":41893813,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-27T12:15:57.553","AcceptedAnswerId":null,"OwnerUserId":7478608.0,"Title":"Im2txt: Using TFRecords of preprocessed MSCOCO dataset to perform captioning","Body":"<p>I'm trying to perform a captioning on the MSCOCO validation set, with the open source <strong><\/strong> model<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/im2txt\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/im2txt<\/a><\/p>\n<\/blockquote>\n\n<p>I was wondering if is it possible to exploit the validation set  shards generated by the  preprocessing script <\/p>\n\n<p>e.g.  to do this, since captioning a large set of JPEG files takes too long.<\/p>\n\n<p>How can I adapt a <strong><\/strong> to handle  instead of JPG or PNG files?<\/p>\n","answers":[]}
{"QuestionId":41894947,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-27T13:21:23.823","AcceptedAnswerId":41895214.0,"OwnerUserId":5674154.0,"Title":"How to get a inverse-cumulative sum of a vector in Tensorflow","Body":"<p>How to get a inverse-cumulative sum of a symbolic vector?<\/p>\n\n<pre><\/pre>\n\n<p>For example<\/p>\n\n<p><strong>Input<\/strong> <\/p>\n\n<pre><\/pre>\n\n<p><strong>Output<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41895214","CreationDate":"2017-01-27T13:34:16.250","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/scan#cumsum\" rel=\"nofollow noreferrer\"><code>tf.cumsum<\/code><\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\ninput = tf.placeholder('float32', [None])\noutput = tf.cumsum(input, reverse=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":41897212,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-27T15:15:29.670","AcceptedAnswerId":null,"OwnerUserId":6770703.0,"Title":"How to sort a multi-dimensional tensor using the returned indices of tf.nn.top_k?","Body":"<p>I have two multi-dimensional tensors  and . And I want to sort them by the values of .<\/p>\n\n<p>I found <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/api_docs\/python\/nn\/evaluation#top_k\" rel=\"noreferrer\"><\/a> is able to sort a tensor and return the indices which is used to sort the input. How can I use the returned indices from  to sort ?<\/p>\n\n<p>For example,<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<h3>Update<\/h3>\n\n<p>Combining <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#gather_nd\" rel=\"noreferrer\"><\/a> with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/shapes_and_shaping#meshgrid\" rel=\"noreferrer\"><\/a> can be one solution. For example, the following code is tested on python 3.5 with tensorflow :<\/p>\n\n<pre><\/pre>\n\n<p>However, I wonder if there is a solution which is more readable and doesn't need to create  above.<\/p>\n","answers":[{"AnswerId":"41898058","CreationDate":"2017-01-27T15:58:44.823","ParentId":null,"OwnerUserId":"5540516","Title":null,"Body":"<p>Your code have a problem.<\/p>\n\n<pre><code>b = tf.reshape(tf.range(60), (2, 5, 3, 7))\n<\/code><\/pre>\n\n<p>Because TensorFlow Cannot reshape a tensor with 60 elements to shape [2,5,3,7] (210 elements).\nAnd you can't sort a rank 4 tensor (b) using indices of rank 3 tensors.<\/p>\n"}]}
{"QuestionId":41898437,"AnswerCount":3,"Tags":"<machine-learning><neural-network><deep-learning><keras><conv-neural-network>","CreationDate":"2017-01-27T16:19:30.110","AcceptedAnswerId":41898769.0,"OwnerUserId":6311007.0,"Title":"Possible\/maybe category in deep learning","Body":"<p>I'm interested in taking advantage of some partially labeled data that I have in a deep learning task. I'm using a fully convolutional approach, not sampling patches from the labeled regions.<\/p>\n\n<p>I have masks that outline regions of definite positive examples in an image, but the unmasked regions in the images are not necessarily negative - they may be positive. Does anyone know of a way to incorporate this type of class in a deep learning setting?<\/p>\n\n<p>Triplet\/contrastive loss seems like it may be the way to go, but I'm not sure how to accommodate the \"fuzzy\" or ambiguous negative\/positive space.<\/p>\n","answers":[{"AnswerId":"41902093","CreationDate":"2017-01-27T20:00:53.387","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>You could treat this as a semi-supervised problem. Use the full dataset without labels to train a bottleneck autoencoder structure (or a GAN approach). This pretrained model can then be adjusted (e.g. removing the last layers, adding a better layer structure at the end on top of the bottleneck features) and finetuned on the labeled data.<\/p>\n"},{"AnswerId":"41909845","CreationDate":"2017-01-28T12:34:57.763","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>I'm assuming that you are struggling with a data segmantation task with a problem of a ill-definied background (e.g. you are not sure if all examples are correctly labeled). Recently I came across the similiar problem and this is what I came across during my research:<\/p>\n\n<ol>\n<li><p>In <em>old days<\/em> before deep learning and at the begining of deep learning era - the common way to deal with that is to <em>smooth<\/em> your output with some kind of a probability model which would take into account the possibility of a noisy labels (you could read about this in a <strong>Learning to Label from Noisy Data<\/strong> chapter from this <a href=\"https:\/\/www.cs.toronto.edu\/~vmnih\/docs\/Mnih_Volodymyr_PhD_Thesis.pdf\" rel=\"nofollow noreferrer\">book<\/a>. It's important to discriminate this probabilistic models from models used to smooth your labels w.r.t. to image or label structure like classical CRFs for bilateral smoothing.<\/p><\/li>\n<li><p>What we finally used (and worked really well) is the Channel Inhibited Softmax idea from this <a href=\"http:\/\/www.ingentaconnect.com\/content\/ist\/jist\/2016\/00000060\/00000001\/art00003\" rel=\"nofollow noreferrer\">paper<\/a>. In terms of a mathematical properties - it makes your network much more robust to some objects not labeled - because it makes your network to output much higher positive valued <code>logits<\/code> at correctly labeled objects. <\/p><\/li>\n<\/ol>\n"},{"AnswerId":"41898769","CreationDate":"2017-01-27T16:37:00.647","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>Try <em>label smoothing<\/em> as described in section 7.5.1 of <a href=\"http:\/\/www.deeplearningbook.org\/\" rel=\"nofollow noreferrer\">Deep Learning book<\/a>:<\/p>\n\n<blockquote>\n  <p>We can assume that for some small constant <code>eps<\/code>, the training set label y is correct with probability <code>1 - eps<\/code>, and otherwise any of the other possible labels might be correct.<\/p>\n  \n  <p>Label smoothing regularizes a model based on a softmax with <code>k<\/code> output values by replacing the hard 0 and 1 classification targets with targets of <code>eps \/ k and 1 - (k - 1) \/ k * eps<\/code>, respectively.<\/p>\n<\/blockquote>\n\n<p>See my question about <a href=\"https:\/\/stackoverflow.com\/questions\/39335535\/label-smoothing-soft-targets-in-pandas\">implementing label smoothing in Pandas<\/a>.<\/p>\n\n<p>Otherwise if you know for sure, that some areas are negative, other are positive while some are uncertain, then you can introduce a third <em>uncertain<\/em> class. I have worked with data sets that contained <em>uncertain<\/em> class, which corresponded to samples that could belong to any of the available classes.<\/p>\n"}]}
{"QuestionId":41899454,"AnswerCount":1,"Tags":"<theano><bayesian><hidden-markov-models><pymc3>","CreationDate":"2017-01-27T17:13:20.930","AcceptedAnswerId":42007237.0,"OwnerUserId":6210630.0,"Title":"What pymc3 Monte-Carlo stepper can I use for a custom categorical distribution?","Body":"<p>I am working on implementing hidden-Markov-Chains in pymc3.  I have gotten pretty far in implementing the hidden states.  Below, I am showing a simple 2-state Markov-chain:<\/p>\n\n<pre><\/pre>\n\n<p>I am now defining a class that describes the states.  As input, I need to know the probability P1 to move from state 0 to state 1, and P2 to move from 1->0.  I also need to know the probability PA for the first state being 0.<\/p>\n\n<pre><\/pre>\n\n<p>I am pretty proud of the advanced indexing ninja tricks that I learned on the theano google group.  You can also implement the same with a tt.switch.  Something I was not too sure about is the self.mode.  I just gave it 0 to avoid a testvalue error.  Here is how to use the class in a model that tests whether it works.  In this case the state is not hidden, but observed.<\/p>\n\n<pre><\/pre>\n\n<p>the output reproduces the data nicely.  In the next model, I will show the problem.  Here I do not directly observe the states but the state with some Gaussian noise added (thus hidden state).  If you run the model with a Metropolis stepper then it crashes with an index error, which I traced back to problems related to using the Metropolis stepper on Categorical Distributions.  Unfortunately, the only Stepper that would apply to my class is the CategoricalGibbsMetropolis stepper, but it refuses to work with my class since it is not explicitly a Categorial Distribution.<\/p>\n\n<pre><\/pre>\n\n<p>The ElemwiseCategorical makes it run, but does not assign the correct value for my states.  The states are either all 0, or all 1s.<\/p>\n\n<p>How can I tell the ElemwiseCategorial to assign a vector of states of 1s and 0s, or alternatively how can I get the CategorialGibbsMetropolis to recognize my distribution as categorical.  This must be a common problem with custom distributions.<\/p>\n","answers":[{"AnswerId":"42007237","CreationDate":"2017-02-02T16:21:21.347","ParentId":null,"OwnerUserId":"6210630","Title":null,"Body":"<p>Since I have not heard from anyone on my question, I answered it myself.  The trick that I used was suggested by Chris Fonnesbeck on pymc3 github where I opened the issue.  He suggested to subclass pm.Categorical.<\/p>\n\n<pre><code>class HMMStates(pm.Categorical):\n    \"\"\"\n    Hidden Markov Model States\n    Parameters\n    ----------\n    P1 : tensor\n        probability to remain in state 1\n    P2 : tensor\n        probability to move from state 2 to state 1\n\n    \"\"\"\n\n    def __init__(self, PA=None, P1=None, P2=None,\n                 *args, **kwargs):\n        super(pm.Categorical, self).__init__(*args, **kwargs)\n        self.PA = PA\n        self.P1 = P1\n        self.P2 = P2\n        self.k = 2 # two state model\n        self.mean = 0.\n        self.mode = tt.cast(0,dtype='int64')\n\n    def logp(self, x):\n        PA = self.PA\n        P1 = self.P1\n        P2 = self.P2\n\n        # now we need to create an array with probabilities\n        # so that for x=A: PA=P1, PB=(1-P1)\n        # and for x=B: PA=P2, PB=(1-P2)\n        PT = tt.stack((P1,P2))\n\n        P = PT[x[:-1]]\n\n        x_i = x[1:]\n\n        ou_like = pm.Categorical.dist(P, shape=(N_chain-1,2)).logp(x_i)\n        return pm.Categorical.dist(PA).logp(x[0]) + tt.sum(ou_like)\n<\/code><\/pre>\n\n<p>my HMMStates cannot really call the pm.Categorical super init, so I am calling the super class of pm.Categorical which is pm.Discrete.  That trick makes it pass the test of BinaryGibbsMetropolis and CategoricalGibbsMetropolis.<\/p>\n\n<p>If you are interested in implementing 2-state and multiple state HMM, I will implement all these cases <a href=\"https:\/\/github.com\/hstrey\/Hidden-Markov-Models-pymc3\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":41899610,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-27T17:21:59.793","AcceptedAnswerId":41934512.0,"OwnerUserId":5011934.0,"Title":"Inference results depend on order of images in training batch","Body":"<p>I've trained the same network two times on the same dataset of 5 images. For the first time, the images in a batch for each step were in the same order. For the second time, the batch was shuffled before every training step. Both models overfit. Both models were tested on shuffled images from training dataset.\nThe firs model shows 100% accuracy.\n```<\/p>\n\n<pre><\/pre>\n\n<p>```<\/p>\n\n<p>The second model shows accuracy close to a random guess.\n```<\/p>\n\n<pre><\/pre>\n\n<p>```<\/p>\n\n<p>Related issues - none<\/p>\n\n<h3>Environment:<\/h3>\n\n<ul>\n<li>Ubuntu 14.04 64-bit<\/li>\n<li>CUDA 8.0<\/li>\n<li>cuDNN 5.1<\/li>\n<li>tensorflow 0.12.1 from pip with GPU support<\/li>\n<\/ul>\n\n<h3>Dataset<\/h3>\n\n<p>5 images from MNIST<\/p>\n\n<h3>The attached archive<\/h3>\n\n<p>contains code for reproducible example. <\/p>\n\n<ul>\n<li>runner.py is responsible for preparing image batches and maintains the queue.<\/li>\n<li>network.py contains the code of simple neural network(2 convolutional layers with relu activation, softmax output layer)<\/li>\n<li>train_not_shuffled.py trains and tests the network using non-shuffled batch<\/li>\n<li>train_shuffled.py trains and tests the network using shuffled batch<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/files\/735318\/example.gz\" rel=\"nofollow noreferrer\">example.gz<\/a><\/p>\n\n<p>I've retrained the models several times, tried different datasets and network architectures. I've visualized graph to make sure inference uses the same variables. I've also visualized weights, they look slightly different for two cases but don't contain any anomalies that could explain such behavior.<\/p>\n","answers":[{"AnswerId":"41934512","CreationDate":"2017-01-30T11:14:51.443","ParentId":null,"OwnerUserId":"5011934","Title":null,"Body":"<p>The mistake was in the testing code. This lines<\/p>\n\n<pre><code>images_batch = images_batch.eval()\nlabels_batch = labels_batch.eval()\n<\/code><\/pre>\n\n<p>run separately, so images and labels were actually from different batches. If batches were identical, labels correspond to images and testing results were perfect. On the other hand, when batches were shuffled, test predictions were compared to wrong labels, so the results were poor.\nReplacing the two lines above with<\/p>\n\n<pre><code>images_batch, labels_batch = sess.run(train_runner.get_inputs())\n<\/code><\/pre>\n\n<p>fixed the problem.<\/p>\n"}]}
{"QuestionId":41899858,"AnswerCount":0,"Tags":"<python><python-2.7><tensorflow><redhat><ubuntu-16.04>","CreationDate":"2017-01-27T17:37:46.010","AcceptedAnswerId":null,"OwnerUserId":6528756.0,"Title":"Same code and same data run on Tensorflow on Ubuntu (16.04-16.10) or Redhat (7.3) gives very different results","Body":"<p>I'm running a Convolutional Neural Network script using tensorflow on the same python version using anaconda distribution (Python 2.7.12 :: Anaconda 4.2.0 (64-bit)) on the exact same dataset, on 2 different machines:<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>In both of the machines I have installed tensorflow using  (v0.12) and before doing that I ran a  to make sure I had all the packages with same versions.<\/p>\n\n<p>Here the fun part starts. Only on the RedHat mahcine the ADAM optimizer converges (I predict the 3d location of neurons and it goes below 10um in around 500 epochs and the final accuracy on validation is around 5um on validation set). Excatly same code and exactly same data (cloned as they are from git) on \nthe Ubuntu machine give much worse results: after 500 iteration the error is still around 48 um (starting 'random' accuracy is 69 um) and the final accuracy is around 56um on the validation set.<\/p>\n\n<p>Now, I checked that the data are exactly the same, they are shuffled with the same random seed and the training and validation set are the same. It only seems that the ADAM optimizer (or others as well) does not converge on the Ubuntu system, while it easily converges on the Redhat one. <\/p>\n\n<p>This are the first iterations:<\/p>\n\n<p>RedHat<\/p>\n\n<pre><\/pre>\n\n<p>Ubuntu<\/p>\n\n<pre><\/pre>\n\n<p>On the Redhat machine the CNN converges also using other optimizers. I tried the same code and the same data also on another machine with the same Redhat version installed and on Ubuntu 16.04 and I got the same weird results: working properly on Redhat, and not converging on Ubuntu.<\/p>\n\n<p>I have no idea how the results I get are so different, since I checked that all the installed packages version are the same.<\/p>\n","answers":[]}
{"QuestionId":41900139,"AnswerCount":0,"Tags":"<python-2.7><memory-leaks><out-of-memory><keras><lstm>","CreationDate":"2017-01-27T17:54:18.100","AcceptedAnswerId":null,"OwnerUserId":7476145.0,"Title":"Keras Neural Network LSTM multiple inputs out-of-memory","Body":"<p>I tried to re-create this example from Andrew (answer 3 here <a href=\"https:\/\/stackoverflow.com\/questions\/39674713\/neural-network-lstm-keras-multiple-inputs\">Neural Network LSTM input shape from dataframe<\/a>).<\/p>\n\n<p>What I want to do is if I have time series data that looks like this:\nDate - Observation1 - Observation2 - Observation3 ... n\nI want to use for example Observation1 and Observation2 to predict Observation3. <\/p>\n\n<p>Based on the example I linked before this works so far but the problem is that if I have too much datapoints then the script gets after a long time a KILL signal at this line:<\/p>\n\n<pre><\/pre>\n\n<p>I checked the memory usage and it just fills up until the memory and the swap is full and then it gets the script stops with KILL signal (I tried it also on a machine with 16GB of memory). Does anyone have an idea how I could avoid that or if there is a way to split up the data for the pad_seqences function?<\/p>\n","answers":[]}
{"QuestionId":41900775,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-01-27T18:35:05.817","AcceptedAnswerId":41961734.0,"OwnerUserId":1628635.0,"Title":"TensorFlow: Value Error Shape and Rank Do Not Match: ValueError: Shape (?, 128, 128, 2) must have rank 2","Body":"<p>I'm getting this exception when running a prediction using a deconv neural network. The rank and shape seem to be the same so I'm not sure what the issue is.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41961734","CreationDate":"2017-01-31T15:51:01.667","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>Shape <code>(?, 128, 128, 2)<\/code> is a rank 4 tensor, but the softmax operator apparently expected a rank 2 tensor.<\/p>\n\n<p>It looks to me like you are using an old version of Tensorflow. In version 0.10, <code>tf.nn.softmax<\/code> required that its input be rank 2:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn\/classification#softmax\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn\/classification#softmax<\/a><\/p>\n\n<p>In current versions of Tensorflow, <code>tf.nn.softmax<\/code> accepts inputs of any rank:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#softmax\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#softmax<\/a><\/p>\n\n<p>Try using a newer version of Tensorflow.<\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41900924,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-27T18:43:37.110","AcceptedAnswerId":null,"OwnerUserId":6842947.0,"Title":"Reading from tf.train.string_input_producer","Body":"<p>I have a dataset with about 1.6 million images. I have stored the full path to these images in a text file (each full path in one line, thus with 1.6 million lines).\nI have stored this list in a  using the  function. I would like to read the name of a file from this, and then read it in a function and then perform some preprocessing on it before feeding it for training. <\/p>\n\n<p>A small snippet from my code file is the following :<\/p>\n\n<pre><\/pre>\n\n<p>This gives the key as follows :<\/p>\n\n<pre><\/pre>\n\n<p>and the value as follows :<\/p>\n\n<pre><\/pre>\n\n<p>At this point I feel lost.<\/p>\n\n<ol>\n<li>Why  ?<\/li>\n<li>How do I obtain the actual filename which I can use inside a function ?<\/li>\n<\/ol>\n","answers":[]}
{"QuestionId":41901862,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><keras>","CreationDate":"2017-01-27T19:45:47.077","AcceptedAnswerId":41903161.0,"OwnerUserId":3165156.0,"Title":"Can anyone tell me what is the error in this CNN code?","Body":"<p>The size of image is 32*32*3<\/p>\n\n<pre><\/pre>\n\n<p>The error in the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41903161","CreationDate":"2017-01-27T21:17:57.493","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>I think that you use \"tf\" dim_ordering. Change this in keras.json to \"th\" so the 2nd parameter is the filter size instead of the last.<\/p>\n\n<p>Also your problem is partly to the usage of <code>'border_mode='valid'<\/code>. Use <code>border_mode='same'<\/code> to preserve dimension for convolutions. <\/p>\n"},{"AnswerId":"41901969","CreationDate":"2017-01-27T19:52:42.553","ParentId":null,"OwnerUserId":"5003873","Title":null,"Body":"<p>Here I've copied and pasted the relevant section of your error code.<\/p>\n\n<p><code>tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative dimension size caused by subtracting 5 from 3 for 'Conv2D' (op: 'Conv2D') with input shapes: [?,3,32,32], [5,5,32,6]<\/code><\/p>\n\n<p>The part after that is caused by this one and will go away once you fix this.\nHowever, w\/o seeing any code I cannot tell you what exactly is causing this.<\/p>\n"}]}
{"QuestionId":41901978,"AnswerCount":1,"Tags":"<python><theano><keras>","CreationDate":"2017-01-27T19:53:28.737","AcceptedAnswerId":41902115.0,"OwnerUserId":4896417.0,"Title":"Keras\/Theano output","Body":"<p>I'm struggling with theano types.<\/p>\n\n<p>I created some  model in Keras, where the last layer is<\/p>\n\n<pre><\/pre>\n\n<p>When I try<\/p>\n\n<pre><\/pre>\n\n<p>How can I get get the actual output number?\nI tried using  resp. \nbut that results only in <\/p>\n\n<p><\/p>\n","answers":[{"AnswerId":"41902115","CreationDate":"2017-01-27T20:02:24.953","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Don't drown on a glass of water.<\/p>\n\n<p>You can just use final_model.predict(input) to get an array of predictions, given an array of input values. Just note that this function can be used to predict more than set of input values.<\/p>\n"}]}
{"QuestionId":41903051,"AnswerCount":3,"Tags":"<keras><keras-layer>","CreationDate":"2017-01-27T21:10:48.260","AcceptedAnswerId":41905306.0,"OwnerUserId":1934212.0,"Title":"Change input tensor shape for VGG16 application","Body":"<p>I want to feed images with the shape (160,320,3) to <\/p>\n\n<pre><\/pre>\n\n<p>How can I include a layer that reshapes the images to the shape expected by the VGG16 model, which is (224,224,3) ?<\/p>\n","answers":[{"AnswerId":"58099966","CreationDate":"2019-09-25T13:49:33.983","ParentId":null,"OwnerUserId":"12119431","Title":null,"Body":"<p>You can use resize() function of Opencv library.<\/p>\n\n<pre><code> import cv2\n    width = int(224)\n    height = int(224)\n    dim = (width, height)\n    '''images contains original dimension image array'''\n    resized_images=[]\n    for i in range(0,images.shape[0]):\n           resized = cv2.resize(images[i], dim, interpolation = cv2.INTER_AREA)\n           resized_images.append(resized)\n<\/code><\/pre>\n"},{"AnswerId":"48747535","CreationDate":"2018-02-12T13:23:09.607","ParentId":null,"OwnerUserId":"2998566","Title":null,"Body":"<p>There are two things that you need to do:<\/p>\n\n<ol>\n<li>Explicitly declare the input shape to have variable sized inputs by defining None for the image width and height.<\/li>\n<li>Donot use flatten() as it relies on the fixed input shape. Instead use GlobalMaxPooling that will not only do the adaptive pooling, but also flatten the input tensor for the FC to work on.<\/li>\n<\/ol>\n\n<p>I hope this will help you achieve what you what.<\/p>\n"},{"AnswerId":"41905306","CreationDate":"2017-01-28T00:51:39.527","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p><code>VGG16<\/code> model in itself is just a set of weights of the fixed sequence of layers and fixed convolution kernel sizes etc. That doesn't mean that those convolution kernels cannot be applied to images of other sizes. <\/p>\n\n<p>For example in your case: <\/p>\n\n<pre><code>from keras.models import Model\nfrom keras.layers import Dense,Flatten\nfrom keras.applications import vgg16\nfrom keras import backend as K\n\nmodel = vgg16.VGG16(weights='imagenet', include_top=False, input_shape=(160,320,3))\nmodel.summary(line_length=150)\n\nflatten = Flatten()\nnew_layer2 = Dense(10, activation='softmax', name='my_dense_2')\n\ninp2 = model.input\nout2 = new_layer2(flatten(model.output))\n\nmodel2 = Model(inp2, out2)\nmodel2.summary(line_length=150)\n<\/code><\/pre>\n\n<p>According to <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/applications\/vgg16.py#L77\" rel=\"noreferrer\">here<\/a> the minimum image size can be <code>48x48x3<\/code> anything above than that is fine. <\/p>\n\n<p>Now its true the original weights were learnt on <code>224,224,3<\/code> shaped images but the filters weights act as very good starting point for new tasks with new set of images. You do need to re-train the network but the network would converge very quickly. This is the basis of transfer learning. <\/p>\n"}]}
{"QuestionId":41903062,"AnswerCount":1,"Tags":"<validation><tensorflow><deep-learning>","CreationDate":"2017-01-27T21:11:49.113","AcceptedAnswerId":null,"OwnerUserId":7479868.0,"Title":"Tensorflow: How the validation set improves the learning curve","Body":"<p>In the tutorial for <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/tf\/\" rel=\"nofollow noreferrer\">MNIST<\/a>, it is shown that the dataset is split into three different sets: (1) <em>data_sets.train<\/em>, (2) <em>data_sets.validation<\/em> and (3) <em>data_sets.test<\/em>. However, in the training loop just the <strong>data_sets.train<\/strong> is used for training.<\/p>\n\n<pre><\/pre>\n\n<p>And every  the model is evaluated so three different evaluations are gotten: (1) Training Data Eval, (2) Validation Data Eval and (3) Test Data Eval.<\/p>\n\n<p>Usually, in Machine Learning, the validation set is used for fine-tuning the parameters of the model, and it improves the learning curve.<\/p>\n\n<p>How does the validation set is employed in Tensorflow so that the learning curve improves?<\/p>\n","answers":[{"AnswerId":"41907265","CreationDate":"2017-01-28T07:08:11.317","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Because training in deep learning takes so long it takes way too long to use something like grid search or random search for hyperparameter optimization. <\/p>\n\n<p>Validation set is mostly used to look out for overfitting on the train dataset during training. Then the model \/ hyperparameter are tweaked according to the validation set. This is done manually as far as I know. <\/p>\n\n<p>The test set is something that has no influence in designing the algorithm to estimate the performance of the classifier on new unseen data.<\/p>\n\n<p>Edit:<\/p>\n\n<p>There are some techniques that use the validation set like EarlyStopping (<a href=\"https:\/\/keras.io\/callbacks\/#earlystopping\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/#earlystopping<\/a>) or reducing the learning rate when the validation error did not decreasing for x epochs (<a href=\"https:\/\/keras.io\/callbacks\/#reducelronplateau\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/callbacks\/#reducelronplateau<\/a>).<\/p>\n"}]}
{"QuestionId":41903394,"AnswerCount":2,"Tags":"<android><ios><machine-learning><react-native><tensorflow>","CreationDate":"2017-01-27T21:35:55.967","AcceptedAnswerId":44083450.0,"OwnerUserId":3765706.0,"Title":"How can I use Tensorflow with react-native?","Body":"<p>I've looked around the web and cant seem to find a way to use <a href=\"https:\/\/facebook.github.io\/react-native\/\" rel=\"noreferrer\">react native<\/a> with tensorflow.<\/p>\n\n<p>I dont think TF supports react-native (at least not officially) integration but i hope someone in the community has found a way.<\/p>\n\n<p>How can one use tensorflow in a react-native project?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"47655136","CreationDate":"2017-12-05T13:45:56.143","ParentId":null,"OwnerUserId":"3054025","Title":null,"Body":"<p>Came across the following <a href=\"https:\/\/github.com\/tensorflow\/tfjs\/tree\/master\/tfjs-react-native\" rel=\"nofollow noreferrer\">Platform Adapter for React Native<\/a>.<\/p>\n\n<blockquote>\n  <p>This package provides a TensorFlow.js platform adapter for react\n  native. It provides GPU accelerated execution of TensorFlow.js\n  supporting all major modes of tfjs usage<\/p>\n<\/blockquote>\n\n<hr>\n\n<p>There is also a library <a href=\"https:\/\/github.com\/reneweb\/react-native-tensorflow\" rel=\"nofollow noreferrer\">react-native-tensorflow<\/a> bridging tensorflow with react-native.<\/p>\n\n<blockquote>\n  <p>A TensorFlow inference library for react native.<\/p>\n<\/blockquote>\n\n<p>They have recently added support for iOS, so it's available for both Android and iOS platforms.<\/p>\n\n<p>To get started:<\/p>\n\n<pre><code>$ npm install react-native-tensorflow --save\n<\/code><\/pre>\n\n<p>Perform library link:<\/p>\n\n<pre><code>$ react-native link react-native-tensorflow\n<\/code><\/pre>\n\n<p>For iOS you will also need to use CocoaPods with content:<\/p>\n\n<pre><code>target '&lt;ProjectName&gt;'\n       pod 'TensorFlow-experimental'\n<\/code><\/pre>\n\n<p>There are performance limitations due to the nature of react-native bridge when transferring large data between native side and JS. For other known issues and usage check out the library's documentation.<\/p>\n"},{"AnswerId":"44083450","CreationDate":"2017-05-20T08:07:53.423","ParentId":null,"OwnerUserId":"7242275","Title":null,"Body":"<p>This article explains using tensorflow with native android development.\n<a href=\"https:\/\/omid.al\/posts\/2017-02-20-Tutorial-Build-Your-First-Tensorflow-Android-App.html\" rel=\"noreferrer\">https:\/\/omid.al\/posts\/2017-02-20-Tutorial-Build-Your-First-Tensorflow-Android-App.html<\/a><\/p>\n\n<p>I also needs a tf supporting library for react-native. But it seems as you said there is no library yet.<\/p>\n\n<p>Here they have built an adroid app with react native with tensorflow support, have a look,\n<a href=\"https:\/\/github.com\/NorthFoxz\/whatsthis\" rel=\"noreferrer\">https:\/\/github.com\/NorthFoxz\/whatsthis<\/a><\/p>\n"}]}
{"QuestionId":41903928,"AnswerCount":4,"Tags":"<keras><keras-layer>","CreationDate":"2017-01-27T22:17:38.317","AcceptedAnswerId":41905696.0,"OwnerUserId":1934212.0,"Title":"Add a resizing layer to a keras sequential model","Body":"<p>How can I add a resizing layer to <\/p>\n\n<pre><\/pre>\n\n<p>using <\/p>\n\n<pre><\/pre>\n\n<p>To resize an image from shape (160, 320, 3) to (224,224,3) ?<\/p>\n","answers":[{"AnswerId":"44900216","CreationDate":"2017-07-04T07:59:56.673","ParentId":null,"OwnerUserId":"1983378","Title":null,"Body":"<p>I think you should consider using tensorflow's resize_images layer. <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/resize_images\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/resize_images<\/a><\/p>\n\n<p>It appears keras does not include this, and perhaps because the feature does not exist in theano. I have written a custom keras layer that does the same. It's a quick hack, so it might not work well in your case.<\/p>\n\n<pre><code>import keras\nimport keras.backend as K\nfrom keras.utils import conv_utils\nfrom keras.engine import InputSpec\nfrom keras.engine import Layer\nfrom tensorflow import image as tfi\n\nclass ResizeImages(Layer):\n    \"\"\"Resize Images to a specified size\n\n    # Arguments\n        output_size: Size of output layer width and height\n        data_format: A string,\n            one of `channels_last` (default) or `channels_first`.\n            The ordering of the dimensions in the inputs.\n            `channels_last` corresponds to inputs with shape\n            `(batch, height, width, channels)` while `channels_first`\n            corresponds to inputs with shape\n            `(batch, channels, height, width)`.\n            It defaults to the `image_data_format` value found in your\n            Keras config file at `~\/.keras\/keras.json`.\n            If you never set it, then it will be \"channels_last\".\n\n    # Input shape\n        - If `data_format='channels_last'`:\n            4D tensor with shape:\n            `(batch_size, rows, cols, channels)`\n        - If `data_format='channels_first'`:\n            4D tensor with shape:\n            `(batch_size, channels, rows, cols)`\n\n    # Output shape\n        - If `data_format='channels_last'`:\n            4D tensor with shape:\n            `(batch_size, pooled_rows, pooled_cols, channels)`\n        - If `data_format='channels_first'`:\n            4D tensor with shape:\n            `(batch_size, channels, pooled_rows, pooled_cols)`\n    \"\"\"\n    def __init__(self, output_dim=(1, 1), data_format=None, **kwargs):\n        super(ResizeImages, self).__init__(**kwargs)\n        data_format = conv_utils.normalize_data_format(data_format)\n        self.output_dim = conv_utils.normalize_tuple(output_dim, 2, 'output_dim')\n        self.data_format = conv_utils.normalize_data_format(data_format)\n        self.input_spec = InputSpec(ndim=4)\n\n    def build(self, input_shape):\n        self.input_spec = [InputSpec(shape=input_shape)]\n\n    def compute_output_shape(self, input_shape):\n        if self.data_format == 'channels_first':\n            return (input_shape[0], input_shape[1], self.output_dim[0], self.output_dim[1])\n        elif self.data_format == 'channels_last':\n            return (input_shape[0], self.output_dim[0], self.output_dim[1], input_shape[3])\n\n    def _resize_fun(self, inputs, data_format):\n        try:\n            assert keras.backend.backend() == 'tensorflow'\n            assert self.data_format == 'channels_last'\n        except AssertionError:\n            print \"Only tensorflow backend is supported for the resize layer and accordingly 'channels_last' ordering\"\n        output = tfi.resize_images(inputs, self.output_dim)\n        return output\n\n    def call(self, inputs):\n        output = self._resize_fun(inputs=inputs, data_format=self.data_format)\n        return output\n\n    def get_config(self):\n        config = {'output_dim': self.output_dim,\n                  'padding': self.padding,\n                  'data_format': self.data_format}\n        base_config = super(ResizeImages, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n<\/code><\/pre>\n"},{"AnswerId":"41905696","CreationDate":"2017-01-28T02:09:03.350","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>Normally you would use the <a href=\"https:\/\/keras.io\/layers\/core\/#reshape\" rel=\"nofollow noreferrer\"><code>Reshape<\/code><\/a> layer for this:<\/p>\n\n<pre><code>model.add(Reshape((224,224,3), input_shape=(160,320,3))\n<\/code><\/pre>\n\n<p>but since your target dimensions don't allow to hold all the data from the input dimensions (<code>224*224 != 160*320<\/code>), this won't work. You can only use <code>Reshape<\/code> if the number of elements does not change.<\/p>\n\n<p>If you are fine with losing some data in your image, you can specify your own lossy reshape:<\/p>\n\n<pre><code>model.add(Reshape(-1,3), input_shape=(160,320,3))\nmodel.add(Lambda(lambda x: x[:50176])) # throw away some, so that #data = 224^2\nmodel.add(Reshape(224,224,3))\n<\/code><\/pre>\n\n<p>That said, often these transforms are done before applying the data to the model because this is essentially wasted computation time if done in every training step.<\/p>\n"},{"AnswerId":"53656223","CreationDate":"2018-12-06T16:55:08.393","ParentId":null,"OwnerUserId":"2191065","Title":null,"Body":"<p>The accepted answer uses the <a href=\"https:\/\/keras.io\/layers\/core\/#reshape\" rel=\"nofollow noreferrer\">Reshape<\/a> layer, which works like <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html\" rel=\"nofollow noreferrer\">NumPy's reshape<\/a>, which can be used to reshape a 4x4 matrix into a 2x8 matrix, but that will result in the image loosing locality information:<\/p>\n\n<pre><code>0 0 0 0\n1 1 1 1    -&gt;    0 0 0 0 1 1 1 1\n2 2 2 2          2 2 2 2 3 3 3 3\n3 3 3 3\n<\/code><\/pre>\n\n<p>Instead, image data should be rescaled \/ \"resized\" using, e.g., <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/resize_images\" rel=\"nofollow noreferrer\">Tensorflows <code>image_resize<\/code><\/a>.\nBut beware about the correct <a href=\"https:\/\/hackernoon.com\/how-tensorflows-tf-image-resize-stole-60-days-of-my-life-aba5eb093f35\" rel=\"nofollow noreferrer\">usage<\/a> and the bugs!\nAs shown in the <a href=\"https:\/\/stackoverflow.com\/questions\/42260265\/resizing-an-input-image-in-a-keras-lambda-layer\">related question<\/a>, this can be used with a lambda layer:<\/p>\n\n<pre><code>model.add( keras.layers.Lambda( \n    lambda image: tf.image.resize_images( \n        image, \n        (224, 224), \n        method = tf.image.ResizeMethod.BICUBIC,\n        align_corners = True, # possibly important\n        preserve_aspect_ratio = True\n    )\n))\n<\/code><\/pre>\n\n<p>In your case, as you have a 160x320 image, you also have to decide whether to keep the aspect ratio, or not. If you want to use a pre-trained network, then you should use the same kind of resizing that the network was trained for.<\/p>\n"},{"AnswerId":"56266748","CreationDate":"2019-05-23T00:53:42.303","ParentId":null,"OwnerUserId":"4619958","Title":null,"Body":"<p>A modification of @KeithWM 's answer, adding <em>output_scale<\/em>, e.g. <em>output_scale=2<\/em> means the output is 2 times the input shape :)<\/p>\n\n<pre><code>class ResizeImages(Layer):\n    \"\"\"Resize Images to a specified size\n    https:\/\/stackoverflow.com\/questions\/41903928\/add-a-resizing-layer-to-a-keras-sequential-model\n\n    # Arguments\n        output_dim: Size of output layer width and height\n        output_scale: scale compared with input\n        data_format: A string,\n            one of `channels_last` (default) or `channels_first`.\n            The ordering of the dimensions in the inputs.\n            `channels_last` corresponds to inputs with shape\n            `(batch, height, width, channels)` while `channels_first`\n            corresponds to inputs with shape\n            `(batch, channels, height, width)`.\n            It defaults to the `image_data_format` value found in your\n            Keras config file at `~\/.keras\/keras.json`.\n            If you never set it, then it will be \"channels_last\".\n\n    # Input shape\n        - If `data_format='channels_last'`:\n            4D tensor with shape:\n            `(batch_size, rows, cols, channels)`\n        - If `data_format='channels_first'`:\n            4D tensor with shape:\n            `(batch_size, channels, rows, cols)`\n\n    # Output shape\n        - If `data_format='channels_last'`:\n            4D tensor with shape:\n            `(batch_size, pooled_rows, pooled_cols, channels)`\n        - If `data_format='channels_first'`:\n            4D tensor with shape:\n            `(batch_size, channels, pooled_rows, pooled_cols)`\n    \"\"\"\n\n    def __init__(self, output_dim=(1, 1), output_scale=None, data_format=None, **kwargs):\n        super(ResizeImages, self).__init__(**kwargs)\n        data_format = normalize_data_format(data_format)  # does not have\n        self.naive_output_dim = conv_utils.normalize_tuple(output_dim,\n                                                           2, 'output_dim')\n        self.naive_output_scale = output_scale\n        self.data_format = normalize_data_format(data_format)\n        self.input_spec = InputSpec(ndim=4)\n\n    def build(self, input_shape):\n        self.input_spec = [InputSpec(shape=input_shape)]\n        if self.naive_output_scale is not None:\n            if self.data_format == 'channels_first':\n                self.output_dim = (self.naive_output_scale * input_shape[2],\n                                   self.naive_output_scale * input_shape[3])\n            elif self.data_format == 'channels_last':\n                self.output_dim = (self.naive_output_scale * input_shape[1],\n                                   self.naive_output_scale * input_shape[2])\n        else:\n            self.output_dim = self.naive_output_dim\n\n    def compute_output_shape(self, input_shape):\n        if self.data_format == 'channels_first':\n            return (input_shape[0], input_shape[1], self.output_dim[0], self.output_dim[1])\n        elif self.data_format == 'channels_last':\n            return (input_shape[0], self.output_dim[0], self.output_dim[1], input_shape[3])\n\n    def _resize_fun(self, inputs, data_format):\n        try:\n            assert keras.backend.backend() == 'tensorflow'\n            assert self.data_format == 'channels_last'\n        except AssertionError:\n            print(\"Only tensorflow backend is supported for the resize layer and accordingly 'channels_last' ordering\")\n        output = tf.image.resize_images(inputs, self.output_dim)\n        return output\n\n    def call(self, inputs):\n        output = self._resize_fun(inputs=inputs, data_format=self.data_format)\n        return output\n\n    def get_config(self):\n        config = {'output_dim': self.output_dim,\n                  'padding': self.padding,\n                  'data_format': self.data_format}\n        base_config = super(ResizeImages, self).get_config()\n        return dict(list(base_config.items()) + list(config.items()))\n<\/code><\/pre>\n"}]}
{"QuestionId":41904521,"AnswerCount":2,"Tags":"<dataset><deep-learning><caffe><conv-neural-network><mnist>","CreationDate":"2017-01-27T23:11:30.540","AcceptedAnswerId":41905566.0,"OwnerUserId":1730503.0,"Title":"How to convert an arbitrary dataset to the siamese network dataset format in caffe?","Body":"<p>I have a dataset that I have created of gray scale images which i want to use with the siamese network example in caffe in which the documentation uses mnist dataset. I want to replace the mnist dataset with my own dataset <\/p>\n\n<p>I see that for doing this I need my dataset to be in the format required by the siamese netwrk. This can be created using the 'create_mnist_siamese.sh' which loads the mnist dataset in the idx3-ubyte format and creates a dataset lmdb database with two images and a matching\/non matching label in each location of the lmdb database.<\/p>\n\n<p>So I figured for me to use the 'create_mnist_siamese.sh' script, my dataset also needs to be in the idx-ubyte format. I tried to convert my dataset to the idx-ubyte format using 'mnisten'. However i get the error 'error:total images are less than num_tests'. I guess the script is not identifying my images. The folder structure of the dataset is like this:<\/p>\n\n<pre><\/pre>\n\n<p>parent directory name - 'generated dataset'<br>\nsubfolders - 1 ,2 ,3 ... (the subfolders are titled 1 - 30 as I want to label the data in each subfolder by the name of the subfolder)<br>\nThe txt file contains image title on each row with the class label.<\/p>\n\n<p>How do I work with my dataset on the siamese network in caffe? Is there a direct way to convert my dataset to the lmdb format for the siamese network? Or do I have to use mnisten? If I do then how do I fix my error ? Anu help will be much appreciated. Thanks.<\/p>\n","answers":[{"AnswerId":"41905566","CreationDate":"2017-01-28T01:40:34.760","ParentId":null,"OwnerUserId":"4221706","Title":null,"Body":"<p>You don't need to use the exact same format - this is just a tutorial.... All you need to do is provide one or multiple data layers, with a total of three top Blobs: <code>data<\/code>, <code>data_p<\/code>, and <code>sim<\/code>. You can do that in any way you'd like, e.g. LMDB (like in the MNIST example), HDF5, or whatever. <\/p>\n\n<h3>General explanation<\/h3>\n\n<p>In the tutorial, they further show and easy way to load the image pairs: you concatenate two images in the channel dimension. For gray-scale, you take two input images, where each has for example the dimension <code>[1, 1, 28, 28]<\/code> (i.e. 1 image, 1 channel, 28x28 resolution). Then you concatenate them to be one image of size <code>[1, 2, 28, 28]<\/code> and save them e.g. to an LMDB.<\/p>\n\n<p>In the network, the first step after loading the data is a \"Slice\" layer, which takes this image, and slices it (i.e. it splits it up) along that axis, thus creating two Top blobs, <code>data<\/code> and <code>data_p<\/code>.<\/p>\n\n<h3>\u00a0How to create the data files?<\/h3>\n\n<p>There is no single right way to do that. The <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/siamese\/convert_mnist_siamese_data.cpp\" rel=\"nofollow noreferrer\">code<\/a> from the tutorial is only for the MNIST set, so unless you have the <em>exact<\/em> same format, you can't use it without changes. You have a couple of possibilities:<\/p>\n\n<ol>\n<li><p>Convert your images to the MNIST-format. Then, the code from the Caffe tutorial works out-of-the-box. It appears that you are trying this - if you need help on that, please be specific: what is \"mnisten\", include your code, and so on.<\/p><\/li>\n<li><p>Write your own script to convert the images.\nThis is actually very simple: all you need to do is read the images in your favorite programming language, select the pairs, calculate the labels, and re-save as LMDB.\nThis is definitely the more flexible way.<\/p><\/li>\n<li><p>Create HDF5 files with multiple Top blobs. This is very simple to do, but will probably be a bit slower than using LMDB.<\/p><\/li>\n<\/ol>\n\n<p>What you use is up to you - I'd probably go with HDF5, as this is an easy and very flexible way to start.<\/p>\n\n<h3>\u00a0How to generate the pairs?<\/h3>\n\n<p>Now, this is the difficult question here. The code from the tutorial just selects random pairs, which is not really optimal, and will make learning rather slow. You don't just need random pairs, you needmeaningful, difficult, but still solvable pairs. How to do that depends entirely on your dataset.<\/p>\n\n<p>A very sophisticated example is presented, in (Radenovi\u0107, 2016): they use a Siamese network for learning a representation for image retrieval on buildings. They use a Structure-from-Motion (SfM) algorithm to create a 3-D reconstruction of a building, and then sample image pairs from those reconstructions. <\/p>\n\n<p>How exactly you create the pairs depends on your data - maybe you are fine with random pairs - maybe you need a sophisticated method. <\/p>\n\n<p><strong>Literature:<\/strong> <\/p>\n\n<p>F. Radenovi\u0107, G. Tolias, and O. Chum. \"CNN Image Retrieval Learns from BoW: Unsupervised Fine-Tuning with Hard Examples\". In: European Conference on Computer Vision (ECCV), 2016. arXiv: <a href=\"http:\/\/arxiv.org\/abs\/1604.02426\" rel=\"nofollow noreferrer\">1604.02426<\/a>.<\/p>\n"},{"AnswerId":"43166763","CreationDate":"2017-04-02T09:29:03.900","ParentId":null,"OwnerUserId":"6241664","Title":null,"Body":"<p>Generating pairs is the most important step in a Siamese network. However, there is a simple way to do that using caffe.<\/p>\n\n<p><strong>Load the data as separate lmdbs<\/strong><\/p>\n\n<p>Create 2 lmdbs <code>data_1<\/code> and <code>data_2<\/code> using the <code>create_imagenet.sh<\/code> or <code>convert_imageset.cpp<\/code> script. Use the same data for both the sets except <code>data_2<\/code> containing one image less that <code>data_1<\/code>.<\/p>\n\n<p>This will ensure that at each epoc, a different pair of images will be compared, thus allowing us to cover all <code>nC2<\/code> combinations(<code>n^2<\/code>actually)<\/p>\n\n<pre><code>layer {\n  name: \"data\"\n  type: \"Data\"\n  top: \"data\"\n  top: \"label\"\n  data_param{\n      source: \"\/home\/subho\/SSD2\/data_1\/\"\n      batch_size: 8\n      backend: LMDB\n     }\n  }\nlayer {\nname: \"data_p\"\ntype: \"Data\"\ntop: \"data_p\"\ntop: \"label_p\"\ndata_param {\n    source: \"\/home\/subho\/SSD2\/data_2\/\"\n    batch_size: 8\n    backend: LMDB\n   }\n}\n<\/code><\/pre>\n\n<p><strong>Introduce a Similarity layer in the prototxt<\/strong><\/p>\n\n<pre><code>layer {\n  name: \"sim_check\"\n  type: \"Similarity\"\n  bottom: \"label\"\n  bottom: \"label_p\"\n  top: \"sim_check\"\n  propagate_down: false # for each bottom_blob\n  propagate_down: false\n  }\nlayer {\n    name: \"loss\"\n    type: \"ContrastiveLoss\"\n    contrastive_loss_param {\n    margin: 1.0\n    }\n    bottom: \"feat\"\n    bottom: \"feat_p\"\n    bottom: \"sim_check\"\n    top: \"loss\"\n  }\n<\/code><\/pre>\n\n<p><strong>Create files for Similarity layer<\/strong><\/p>\n\n<p><a href=\"https:\/\/drive.google.com\/open?id=0B1GbmILdKGiSeV9janltdFBSYzA\" rel=\"nofollow noreferrer\">Download files for similarity layer<\/a><\/p>\n\n<p>Place the <code>similarity_layer.cpp<\/code> in <code>caffe\/src\/caffe\/layers\/<\/code> and <code>similarity_layers.hpp<\/code> in <code>caffe\/include\/caffe\/layers\/<\/code> and rebuild caffe.<\/p>\n\n<pre><code>cd build\ncmake ..\nmake -j12\n<\/code><\/pre>\n\n<p><strong>Note<\/strong>\nIf your network does not converge using the above technique, then you should have a look at the following :<\/p>\n\n<ul>\n<li>Selection of image pairs using hard negatives<\/li>\n<li>Ensuring balance of positive and negative pairs(dissimilar pairs)<\/li>\n<\/ul>\n"}]}
{"QuestionId":41905488,"AnswerCount":0,"Tags":"<python><python-3.x><machine-learning><tensorflow><conv-neural-network>","CreationDate":"2017-01-28T01:23:19.113","AcceptedAnswerId":null,"OwnerUserId":6348198.0,"Title":"How to Get the input_x Tensor to be the Correct Shape","Body":"<p>I am implementing a dynamic-k max pooling convolutional neural network and I can't seem to get the  tensor, which I am using in the  to be the correct shape. I always get the error: <\/p>\n\n<pre><\/pre>\n\n<p>The trace points to the line where I actually run the session. Here is the code for my <\/p>\n\n<pre><\/pre>\n\n<p>Those 4 prints print out the following: <\/p>\n\n<pre><\/pre>\n\n<p><strong>Is there some mistake that I have made in my actual implementation or is there an easy way to reshape these tensors so that I don't get an error (I tried reshaping  to just , but that obviously didn't work)?<\/strong><\/p>\n\n<p><strong>Edit<\/strong> Here is the full code.<\/p>\n\n<pre><\/pre>\n\n<p>And here is the second file (sorry about the indenting).<\/p>\n\n<pre><\/pre>\n\n<p>The training data I use is the Rotten Tomatoes data set of labeled positive and negative movie reviews.<\/p>\n","answers":[]}
{"QuestionId":41908379,"AnswerCount":3,"Tags":"<keras>","CreationDate":"2017-01-28T09:46:13.640","AcceptedAnswerId":41909089.0,"OwnerUserId":7387749.0,"Title":"Keras - Plot training, validation and test set accuracy","Body":"<p>I want to plot the output of this simple neural network: <\/p>\n\n<pre><\/pre>\n\n<p>I have plotted <em>accuracy<\/em> and <em>loss<\/em> of training and validation:<\/p>\n\n<pre><\/pre>\n\n<p>Now I want to add and plot test set's accuracy from , but from  I obtain the same value <em>'acc'<\/em> utilized for plotting accuracy on training data . How could I plot test set's accuracy?<\/p>\n","answers":[{"AnswerId":"41909089","CreationDate":"2017-01-28T11:11:48.000","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>It is the same because you are training on the test set, not on the train set. Don't do that, just train on the training set:<\/p>\n\n<pre><code>history = model.fit(x_test, y_test, nb_epoch=10, validation_split=0.2, shuffle=True)\n<\/code><\/pre>\n\n<p>Change into:<\/p>\n\n<pre><code>history = model.fit(x_train, y_train, nb_epoch=10, validation_split=0.2, shuffle=True)\n<\/code><\/pre>\n"},{"AnswerId":"53078430","CreationDate":"2018-10-31T07:36:31.863","ParentId":null,"OwnerUserId":"8068733","Title":null,"Body":"<p>Validate the model on the test data as shown below and then plot the accuracy and loss<\/p>\n\n<pre><code>model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\nhistory = model.fit(X_train, y_train, nb_epoch=10, validation_data=(X_test, y_test), shuffle=True)\n<\/code><\/pre>\n"},{"AnswerId":"56807595","CreationDate":"2019-06-28T13:11:45.653","ParentId":null,"OwnerUserId":"10368811","Title":null,"Body":"<pre><code>import keras\nfrom matplotlib import pyplot as plt\nhistory = model1.fit(train_x, train_y,validation_split = 0.1, epochs=50, batch_size=4)\nplt.plot(history.history['acc'])\nplt.plot(history.history['val_acc'])\nplt.title('model accuracy')\nplt.ylabel('accuracy')\nplt.xlabel('epoch')\nplt.legend(['train', 'val'], loc='upper left')\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FmcMs.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FmcMs.png\" alt=\"Model Accuracy\"><\/a><\/p>\n\n<pre><code>plt.plot(history.history['loss'])\nplt.plot(history.history['val_loss'])\nplt.title('model loss')\nplt.ylabel('loss')\nplt.xlabel('epoch')\nplt.legend(['train', 'val'], loc='upper left')\nplt.show()\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3HL0C.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3HL0C.png\" alt=\"Model Loss\"><\/a>       <\/p>\n"}]}
{"QuestionId":41908745,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><rank>","CreationDate":"2017-01-28T10:33:33.013","AcceptedAnswerId":null,"OwnerUserId":6725958.0,"Title":"Tensorflow: How to convert a rank 1 tensor into a rank 2 tensor","Body":"<p>I'm trying to make a simple neural network and I have a simple question: How do I convert a tensor which is rank 1 to a tensor which is rank 2?<\/p>\n","answers":[{"AnswerId":"41914985","CreationDate":"2017-01-28T21:24:32.953","ParentId":null,"OwnerUserId":"5235369","Title":null,"Body":"<p>You might be looking for <code>tf.expand_dims()<\/code> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/expand_dims\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/expand_dims<\/a><\/p>\n"}]}
{"QuestionId":41909111,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-28T11:13:45.717","AcceptedAnswerId":42058378.0,"OwnerUserId":2542516.0,"Title":"How to \"append\" Op at the beginning of a TensorFlow graph?","Body":"<p>I have a  proto file which I am importing using . Ops can be added at the end of the graph like this:<\/p>\n\n<pre><\/pre>\n\n<p>But I want to add Ops at the beginning of the graph, so essentially the first Op in the graph_def needs to take the output of my Op as input, how do I do it?<\/p>\n","answers":[{"AnswerId":"55806303","CreationDate":"2019-04-23T07:32:08.483","ParentId":null,"OwnerUserId":"4715661","Title":null,"Body":"<p>The script @Priyatham gives in the link is a good example how to add node in tf graph_def. <code>name<\/code>, <code>op<\/code>, <code>input<\/code>, <code>attr<\/code> are 4 required elements. <code>name<\/code> and <code>op<\/code> could be assigned, whereas <code>input<\/code> should use <code>extend<\/code> and <code>attr<\/code> should use <code>CopyFrom<\/code> method for assignment, like: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>new_node = graph_def.node.add()\nnew_node.op = \"Cast\"\nnew_node.name = \"To_Float\"\nnew_node.input.extend([\"To_Float\"])\nnew_node.attr[\"DstT\"].CopyFrom(attr_value_pb2.AttrValue(type=types_pb2.DT_FLOAT))\nnew_node.attr[\"SrcT\"].CopyFrom(attr_value_pb2.AttrValue(type=types_pb2.DT_FLOAT))\nnew_node.attr[\"Truncate\"].CopyFrom(attr_value_pb2.AttrValue(b=True))\n<\/code><\/pre>\n"},{"AnswerId":"42058378","CreationDate":"2017-02-05T23:09:08.453","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>Finally found a way to do this. I am sure the function Yarolsav mentioned in the comments does something similar internally.<\/p>\n\n<pre><code>new_input = graph_def.node.add()\nnew_input.op = 'new_op_name'  # eg: 'Const', 'Placeholder', 'Add' etc\nnew_input.name = 'some_new_name'\n# set any attributes you want for new_input here\nold_input.input[0] = 'some_new_name'  #  must match with the name above\n<\/code><\/pre>\n\n<p>For details about how to set the attributes, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/graph_transforms\/python\/transform_graph_test.py\" rel=\"noreferrer\">this<\/a> file.<\/p>\n"}]}
{"QuestionId":41909172,"AnswerCount":1,"Tags":"<python><numpy><ipython><ipython-notebook><torch>","CreationDate":"2017-01-28T11:20:27.370","AcceptedAnswerId":41914619.0,"OwnerUserId":5831205.0,"Title":"Unable to import anything in itorch notebook","Body":"<p>I have installed IPython notebook,torch and Itorch on my Ubuntu 16.04.\nI open an iTorch notebook using:<\/p>\n\n<pre><\/pre>\n\n<p>In the notebook I am able to use:<\/p>\n\n<pre><\/pre>\n\n<p>Basically I am able to use the functions of torch.\nBut when I import any python library as numpy or pandas:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Even when i run itorch in terminal I get the same error.\nPlz help.<\/p>\n","answers":[{"AnswerId":"41914619","CreationDate":"2017-01-28T20:41:33.587","ParentId":null,"OwnerUserId":"1993206","Title":null,"Body":"<p>Multilingual notebooks turned out to be very complicated to implement, so Jupyter notebooks can only use one language at a time. See e.g. <a href=\"https:\/\/github.com\/jupyterhub\/jupyterhub\/issues\/202\" rel=\"nofollow noreferrer\">here<\/a>. Unless <code>iTorch<\/code> defines a <a href=\"https:\/\/ipython.org\/ipython-doc\/3\/interactive\/magics.html\" rel=\"nofollow noreferrer\">magic<\/a> or something for using Python too, there's no way to use both in the same notebook. Sorry.<\/p>\n"}]}
{"QuestionId":41909369,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-01-28T11:40:37.823","AcceptedAnswerId":41909605.0,"OwnerUserId":1934212.0,"Title":"Interpreting training trace of a deep neural network: very low training loss and even lower validation loss","Body":"<p>I am a bit sceptical towards the following log , which I get when training a deep neural network for regression target values between -1.0 and 1.0,  with a learning rate of 0.001 and 19200\/4800 training\/validation samples:<\/p>\n\n<pre><\/pre>\n\n<p>Both training an validation loss decrease, which is good news at first sight. But how can the training loss be so low already during the first epoch? And how can the validation loss be even lower? Is that an indication of a systematic error somewhere in my model or training setup?<\/p>\n","answers":[{"AnswerId":"41909605","CreationDate":"2017-01-28T12:06:46.900","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Actually - the validation loss which is smaller than training loss is not so rare phenomenon as one can think. It may occur e.g. when all examples in a validation data are well <em>covered<\/em> by examples from your training set and your network simply learnt the actual structure of your dataset.<\/p>\n\n<p>It happens very often when the structure of your data is not very complexed. Actually - the small value of a loss after first epoch which suprised you might be a clue that this happened in your case.<\/p>\n\n<p>In terms of loss being to small - you haven't specified what your loss is, but assuming that your task is a regression - I guessed its <code>mse<\/code> - and in this case a mean squared error at the level of <code>0.01<\/code> means that a mean Euclidean distance between the true value and the actual value is equal to <code>0.1<\/code> what is <code>5%<\/code> of a diameter of your values set <code>[-1, 1]<\/code>. So - is this error actually so small? <\/p>\n\n<p>You also haven't specified the number of batches which are analysed during one epoch. Maybe if the structure of your data is not so complexed and the batch size was small - one epoch was a sufficient amount of time to learn your data well.<\/p>\n\n<p>In order to check if your model is trained well I advise you to plot a <code>correlation plot<\/code>when you plot <code>y_pred<\/code> on e.g. X-axis and <code>y_true<\/code> on Y-axis. Then you'll actually see how your model is actually trained.<\/p>\n\n<p><strong>EDIT<\/strong> : As Neil mentioned - there might be even more reasons behind the small validation error - like not a good separation of cases. I would also add - that beacuse of the fact - that 5 epoch comprised no more than 90 minutes - maybe it's good to check the result of the model by using a classic crossvalidation schema with e.g. 5 folds. This would assure you that in case of your dataset - your model is performing well.<\/p>\n"}]}
{"QuestionId":41909915,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-28T12:42:20.973","AcceptedAnswerId":null,"OwnerUserId":6842947.0,"Title":"tf.train.string_input_producer behavior in a loop","Body":"<p>The following snippet has been taken from the TensorFlow 0.12 API documentation<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The question I have might be very basic for a regular TensorFlow user, but I am an absolute beginner. The question is the following :<\/p>\n\n<ul>\n<li> creates a queue for holding the filenames. As the  is called over and over again during training, how will it be ensured that everytime the same queue is used ? I guess, it is important since, if different calls to  result in a creation of a new queue, there does not seem to be a way to ensure that different images are picked everytime  and epoch counter and shuffling can be properly maintained.<\/li>\n<\/ul>\n","answers":[{"AnswerId":"41915871","CreationDate":"2017-01-28T23:06:41.237","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>The <code>input_pipeline<\/code> function only creates the part of a (usually larger) graph that is responsible for producing batches of data. If you were to call <code>input_pipeline<\/code> twice - for whatever reason - you would be creating two different queues indeed.<\/p>\n\n<p>In general, the function <code>tf.train.string_input_producer<\/code> actually creates a queue <em>node<\/em> (or <em>operation<\/em>) in the currently active graph (which is the default graph unless you specify something different). <code>read_my_file_format<\/code> then reads from that queue and sequentially produces single \"example\" tensors, while <code>tf.train.shuffle_batch<\/code> then batches these into bundles of length <code>batch_size<\/code> each.<\/p>\n\n<p>However, the output of <code>tf.train.shuffle_batch<\/code>, two Tensors here that are returned from the <code>input_pipeline<\/code> function, only really takes on a (new) value when it is evaluated under a session. If you evaluate these tensors multiple times, they will contain different values - taken, through <code>read_my_file_format<\/code>, from files listed in the input queue.<\/p>\n\n<p>Think of it like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>X_batch, Y_batch = input_pipeline(..., batch_size=100)\n\nwith tf.Session() as sess:\n   sess.run(tf.global_variable_initializer())   \n   tf.train.start_queue_runners()\n\n   # get the first 100 examples and labels\n   X1, Y1 = sess.run((X_batch, Y_batch))\n\n   # get the next 100 examples and labels\n   X2, Y2 = sess.run((X_batch, Y_batch))\n\n   # etc.\n<\/code><\/pre>\n\n<p>The boilerplate code to get it running is a bit more complex, e.g. because queues need to actually be started and stopped in the graph, because they will throw a <code>tf.errors.OutOfRangeError<\/code> when they run dry, etc.\nA more complete example could look like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>with tf.Graph().as_default() as graph:\n   X_batch, Y_batch = input_pipeline(..., batch_size=100)\n\n   prediction = inference(X_batch)\n   optimizer, loss = optimize(prediction, Y_batch)\n\ncoord = tf.train.Coordinator()\nwith tf.Session(graph=graph) as sess:\n   init = tf.group(tf.local_variable_initializer(),\n                   tf.global_variable_initializer())\n   sess.run(init)\n\n   # start the queue runners\n   threads = tf.train.start_queue_runners(coord=coord)\n\n   try:\n       while not coord.should_stop():\n\n           # now you're really indirectly querying the\n           # queue; each iteration will see a new batch of\n           # at most 100 values.\n           _, loss = sess.run((optimizer, loss))\n\n           # you might also want to do something with\n           # the network's output - again, this would\n           # use a fresh batch of inputs\n           some_predicted_values = sess.run(prediction)\n\n   except tf.errors.OutOfRangeError:\n       print('Training stopped, input queue is empty.')\n   finally:\n       coord.request_stop()\n\n   # stop the queue(s)\n   coord.request_stop()\n   coord.join(threads)\n<\/code><\/pre>\n\n<p>For a deeper understanding, you might want to look at the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow noreferrer\">Reading data<\/a> documentation.<\/p>\n"}]}
{"QuestionId":41910073,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-28T13:01:19.477","AcceptedAnswerId":null,"OwnerUserId":6539009.0,"Title":"TensorFlow: tf.case() & f_default should do nothing","Body":"<p>I'm working on a network with 4 class-specific autoencoders (3 layer feedforward), and within the training iteration, there is a case check to decide, which autoencoder has to be updated: <\/p>\n\n<pre><\/pre>\n\n<p>In the case, that no condition is fulfilled, nothing should be done. In this scenario, one of the four cases will be realized, so it's no practical problem here yet, but I want to modify the code, and then it will be important, that the training sample will be skipped in the default case. The problem is, that the return type(s) of f_default and all other return types have to be the same, because sess.run([optimizer, update, cost]) is expecting a certain type. How can I do this, that really nothing happens in the default case? I have already tried to use tf.no_op() but that's not working... <\/p>\n\n<p>Thanks, <\/p>\n\n<p>Meridius<\/p>\n","answers":[{"AnswerId":"41914633","CreationDate":"2017-01-28T20:42:46.097","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To make the signatures match, you could define <code>g()<\/code> as follows:<\/p>\n\n<pre><code>def g():\n  return tf.no_op(), tf.no_op(), tf.constant(0.0)\n<\/code><\/pre>\n\n<p>Note that it would be slightly more efficient to pass <code>g<\/code> directly as <code>f_default<\/code> (as opposed to passing <code>g()<\/code> as the current code does) but the behaviour should be the same.<\/p>\n"}]}
{"QuestionId":41910617,"AnswerCount":2,"Tags":"<numpy><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-01-28T13:58:17.710","AcceptedAnswerId":41910709.0,"OwnerUserId":5051564.0,"Title":"Keras: ValueError: decode_predictions expects a batch of predictions","Body":"<p>I'm using keras' pre-trained model VGG16, following this link: <a href=\"https:\/\/keras.io\/applications\/#vgg16\" rel=\"nofollow noreferrer\">Keras VGG16<\/a>\nI'm trying to decode the prediction output into word of what's in the image:<\/p>\n\n<pre><\/pre>\n\n<p>The full error is:<\/p>\n\n<blockquote>\n  <p>ValueError:  expects a batch of predictions (i.e.\n  a 2D array of shape (samples, 1000)). Found array with shape: (1, 7,\n  7, 512)<\/p>\n<\/blockquote>\n\n<p>Any comments or suggestion is highly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"41910709","CreationDate":"2017-01-28T14:07:36.997","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>You should change a first line to:<\/p>\n\n<pre><code>model = VGG16(weights='imagenet', include_top=True)\n<\/code><\/pre>\n\n<p>Without this line your model is producing a 512 feature maps with size of 7 x 7 pixels. This the reason behind your error.<\/p>\n"},{"AnswerId":"49447760","CreationDate":"2018-03-23T10:39:14.617","ParentId":null,"OwnerUserId":"7769076","Title":null,"Body":"<p>Just to add on the correct answer by @Marcin Mo\u017cejko<\/p>\n\n<p>The same applies to the other available models, so you must always include the top three layers:<\/p>\n\n<pre><code>vgg19 &lt;- application_vgg19(include_top = TRUE, weights = \"imagenet\")\n\nmodel_resnet50 &lt;- application_resnet50(include_top = TRUE, weights = \"imagenet\")\n\nmodel_inception_v3 &lt;- application_inception_v3(include_top = TRUE, weights = \"imagenet\")\n\nmodel_xception &lt;- application_xception(include_top = TRUE, weights = \"imagenet\")\n<\/code><\/pre>\n"}]}
{"QuestionId":41910822,"AnswerCount":0,"Tags":"<c++><opencv><tensorflow><deep-learning>","CreationDate":"2017-01-28T14:20:16.960","AcceptedAnswerId":null,"OwnerUserId":7483011.0,"Title":"How to use a tensorflow graph in opencv c++?","Body":"<p>I want to load a pre-trained tensorflow graph into OpenCV and use to detect the class name of a new image provided by the user.\nThis example seems to work but gives the same class name for every image.(strange)\n<a href=\"https:\/\/github.com\/opencv\/opencv_contrib\/blob\/master\/modules\/dnn\/samples\/tf_inception.cpp\" rel=\"noreferrer\">https:\/\/github.com\/opencv\/opencv_contrib\/blob\/master\/modules\/dnn\/samples\/tf_inception.cpp<\/a><\/p>\n","answers":[]}
{"QuestionId":41910940,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2017-01-28T14:32:16.300","AcceptedAnswerId":null,"OwnerUserId":5050706.0,"Title":"Feed non-placeholder variables in seq2seq","Body":"<p>I'm playing with Tensorflow seq2seq model and I'm wondering how I can feed a trained seq2seq decoder with an arbitrary initial decoder memory (during the training, this initial decoder memory is an output of the encoder).<\/p>\n\n<p>I figured that I need to use feed_dict and TF forces me to feed input sequence because a placeholder is defined for it, same for decoder input.<\/p>\n\n<p>But if I do so, I can't actually force the initial decoder memory to be what I want it to be because it's defined from the decoder input. I still tried to do it but it ignores my third line.<\/p>\n\n<pre><\/pre>\n\n<p>Do you have any clue on how to do what I want ? ie feeding the model variables states even though they aren't designed to be fed by placeholders.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42119418","CreationDate":"2017-02-08T17:08:53.843","ParentId":null,"OwnerUserId":"907954","Title":null,"Body":"<p>Shouldn't you be feeding random numbers, but rather 0's? That's a suggestion in the karpathy lesson on rnns: <a href=\"https:\/\/youtu.be\/cO0a0QYmFm8?list=PLlJy-eBtNFt6EuMxFYRiNRS07MCWN5UIA\" rel=\"nofollow noreferrer\">Recurrent Neural Networks, Image Captioning<\/a><\/p>\n"}]}
{"QuestionId":41911883,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><keras><conv-neural-network>","CreationDate":"2017-01-28T16:08:11.333","AcceptedAnswerId":null,"OwnerUserId":5051564.0,"Title":"TypeError: 'Tensor' object is not callable","Body":"<p>I'm trying to display the output of each layer of the convolutions neural network.\nThe backend I'm using is TensorFlow.\nHere is the code:<\/p>\n\n<pre><\/pre>\n\n<p>The full Error is:<\/p>\n\n<blockquote>\n  <p>convout1_f = K.function([model.input(train=False)],\n  convout1.get_output(train=False)) TypeError: 'Tensor' object is not\n  callable<\/p>\n<\/blockquote>\n\n<p>Any comment or suggestion is highly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"41912305","CreationDate":"2017-01-28T16:51:06.993","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Both  <code>get_output<\/code> and <code>get_input<\/code> methods return either <code>Theano<\/code> or <code>TensorFlow<\/code> tensor. It's not callable because of the nature of this objects.<\/p>\n\n<p>In order to compile a function you should provide only layer tensors and a special Keras tensor called  <code>learning_phase<\/code> which sets in which option your model should be called.<\/p>\n\n<p>Following this <a href=\"https:\/\/stackoverflow.com\/questions\/36745821\/keras-1-0-getting-intermediate-layer-output\/36789699#36789699\">answer<\/a> your function should look like this:<\/p>\n\n<pre><code>convout1_f = K.function([model.input, K.learning_phase()], convout1.get_output)\n<\/code><\/pre>\n\n<p>Remember that you need to pass either <code>True<\/code> or <code>False<\/code> when calling your function in order to make your model computations in either learning or training phase mode.<\/p>\n"}]}
{"QuestionId":41912334,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-28T16:54:54.820","AcceptedAnswerId":41914213.0,"OwnerUserId":4474386.0,"Title":"How to omit zeroes in a 1-D tensorflow tensor","Body":"<p>If I have a tensor like:<\/p>\n\n<p>[    0.   235.     0.   213.  1006.     0.     0.     0.     0.    77.\n     0.   234.]<\/p>\n\n<p>How could one obtain the same vector without zeroes?<\/p>\n\n<p>I want my result to be: [235. 213. 1006. 77. 234]<\/p>\n","answers":[{"AnswerId":"41914213","CreationDate":"2017-01-28T19:59:19.607","ParentId":null,"OwnerUserId":"2911687","Title":null,"Body":"<p>Yes there is a way with pure tensorflow functions :<\/p>\n\n<pre><code>a = tf.Variable([0.235, 0.213, 1006.0, 0, 0, 0, 0, 77, 0,234.0])\nsession = tf.Session()\ninitialiser_op = tf.global_variables_initializer()\nsession.run(initialiser_op)\na_without_zeroes = tf.boolean_mask(a, tf.not_equal(a, 0))\n<\/code><\/pre>\n\n<p><code>tf.py_func<\/code> is another solution to make it work with numpy but I'd expect drawbacks compared to a pure approach, mainly that you would not get the gradient computation among other things (no serialization in the graphdef, no distributed, no gpu?, probably slower performance?)<\/p>\n"}]}
{"QuestionId":41912622,"AnswerCount":1,"Tags":"<tensorflow><jupyter-notebook><tensorboard>","CreationDate":"2017-01-28T17:22:20.057","AcceptedAnswerId":41919896.0,"OwnerUserId":7483498.0,"Title":"tensorflow error utf-8 OS X Sierra","Body":"<p>I've installed TensorFlow with Anaconda on OS X Sierra. I didn't\u00b4t have any problems during installation. \nWriting the tipical example:<\/p>\n\n<pre><\/pre>\n\n<p>The file is created in the appropriate folder, but when I try to pen it with TensorBoard there are not graph.\nWhen I tryed to open the file with Jupyter-Notebook I receive the following text.<\/p>\n\n<p>Error! \/Users\/Pancho\/tf-notebooks\/my_graph\/events.out.tfevents.xxxxxxxxx.MacBook-Pro-xxx.local is not UTF-8 encoded<\/p>\n\n<p>Saving disabled.<\/p>\n\n<p>See console for more details.<\/p>\n\n<p>No additional info on console.<\/p>\n","answers":[{"AnswerId":"41919896","CreationDate":"2017-01-29T10:31:17.700","ParentId":null,"OwnerUserId":"7483498","Title":null,"Body":"<p>I\u00b4ve found the problem.\nI wasn\u00b4t able to open the file using tensorboard because I wasn\u00b4t using the full path to the directory.\nThe UTF-8 error was produced when I tried to open the file with Jupyter notebook.<\/p>\n"}]}
{"QuestionId":41913121,"AnswerCount":1,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2017-01-28T18:14:22.253","AcceptedAnswerId":41945079.0,"OwnerUserId":6433279.0,"Title":"tflearn custom loss function for cosine similarity","Body":"<p>I have an lstm network in tflearn that predicts the next word in a sequence given a context of preceding words. The words are fed into the network as indices of a certain-sized vocabulary, and are output in binary classes, for example:<\/p>\n\n<p>context: [45, 243, 1, 1906, 4, 2, 0, 0, 0, 0]<\/p>\n\n<p>label:   [0,0,0.......1,0,0,......0,0,0] (length of VOCAB_SIZE)<\/p>\n\n<p>However, the model learns to predict nearly the same word every time because I'm using the \"categorical_crossentropy\" objective function in the regression layer.<\/p>\n\n<p>I'd like to try evaluating the loss based on word embeddings (word2vec), which I have already generated for the dataset. Thus, the model predicting 'hello' where ground truth is 'hi' would have a much lower loss than predicting 'pizza'. The plan is to calculate the cosine between two embedding vectors to obtain the \"similarity\" between words.<\/p>\n\n<p>I'm adding this custom loss function to the objectives.py file in the tflearn installation, but have hit a roadblock when trying to convert the predictions into the embedding vectors.<\/p>\n\n<p>tflearn\/objectives.py:<\/p>\n\n<pre><\/pre>\n\n<p>The error returned is:<\/p>\n\n<pre><\/pre>\n\n<p>This indicates that I can't use a tensor to index into vocab (a numpy array). However, I can't use eval() to grab the value of the tensor since this isn't run in a session. So I need a way to set up the transformation of a 1-D tensor of indices to a tensor containing the corresponding word vectors, in order to calculate the loss.<\/p>\n\n<p>Any help with this issue, or another way to evaluate my model, is greatly appreciated.<\/p>\n","answers":[{"AnswerId":"41945079","CreationDate":"2017-01-30T20:43:34.137","ParentId":null,"OwnerUserId":"6433279","Title":null,"Body":"<p>I used tf.gather to resolve this issue. Additionally, I added scaling by the confidence of the algorithm's classification, to fix a bug I ran into where the gradients could not be computed. Below is the code:<\/p>\n\n<p>code at the header of objectives.py:<\/p>\n\n<pre><code>import numpy as np\nimport gensim\nvocab = np.genfromtxt('\/home\/vocab.txt',dtype='str')\nembedding_model = gensim.models.Word2Vec.load('\/home\/embedding')\ni2v = []\nfor v in vocab:\ni2v.append(embedding_model[v])\n<\/code><\/pre>\n\n<p>embedded_similarity(y_pred, y_true):<\/p>\n\n<pre><code>global i2v\nwith tf.name_scope(\"Similarity\"):\n\n    i2v_tensors = [ tf.cast(tf.constant(iv), tf.float32) for iv in i2v ]\n    i2v_tensorarray = tf.pack(i2v_tensors)\n\n    #convert one-hot to indices\n    pred_max = tf.cast(tf.argmax(y_pred,dimension=1), tf.int32)\n    true_max = tf.cast(tf.argmax(y_true,dimension=1), tf.int32)\n\n    #extract probabilities for scaling later\n    pred_iter = tf.concat(tf.constant(1),[y_pred,tf.pack([tf.cast(pred_max,tf.float32)],axis=1)])\n    confidence_scaler = 1 \/ tf.map_fn(lambda x: tf.gather(x, tf.cast(tf.gather(x,tf.constant(5002)),tf.int32)), pred_iter, tf.float32)\n\n    #convert indices into vectors (of type tensor)\n    pred_vectors = tf.map_fn(lambda x: tf.gather(i2v_tensorarray, x), pred_max, tf.float32)\n    true_vectors = tf.map_fn(lambda x: tf.gather(i2v_tensorarray, x), true_max, tf.float32)\n\n    #calc dot product\n    dot_products = tf.reduce_sum(tf.mul(pred_vectors,true_vectors),axis=1)\n\n    #divide by magnitudes\n    pred_magnitudes = tf.sqrt(tf.reduce_sum(tf.mul(pred_vectors,pred_vectors),axis=1))\n    true_magnitudes = tf.sqrt(tf.reduce_sum(tf.mul(true_vectors,true_vectors),axis=1))\n\n    cosines = dot_products \/ tf.mul(pred_magnitudes,true_magnitudes)\n    loss = -1*tf.cast(cosines, tf.float32) + 2 #min loss is 1, max is 3\n    scaled_loss = tf.multiply(loss, confidence_scaler)\n    # loss = -1*cosines + 1\n\n    # return inverse sum of dot products\n    return tf.reduce_mean(scaled_loss)\n<\/code><\/pre>\n\n<p>However, I am encountering a strange bug. When I try to fit the model, the code runs perfectly fine until it lists the number of training and validation samples, as below:<\/p>\n\n<pre><code>---------------------------------\nTraining samples: 800\nValidation samples: 200\n<\/code><\/pre>\n\n<p>Then the output simply freezes, but not the whole computer. I am unable to Ctrl-C the code, and have to start another terminal. There doesn't seem to be a significant slowdown in the system either, and I have tried decreasing both the training set sizes and batch sizes to laughably low numbers, with no results.<\/p>\n\n<p>I'm going to mark this issue as resolved because I answered the main problem I was having, but if anyone has encountered this type of behavior before please comment. Thanks!<\/p>\n"}]}
{"QuestionId":41914939,"AnswerCount":1,"Tags":"<python><theano><keras>","CreationDate":"2017-01-28T21:19:38.420","AcceptedAnswerId":41915669.0,"OwnerUserId":4896417.0,"Title":"Keras input dimension","Body":"<p>I'm struggling with formatting my data.<\/p>\n\n<p>I designed a simple model in Keras.\nIt consists of two separate models ( layer and  layer).\nThose two models are then merges with . They are then combined together by some convolution layers.<\/p>\n\n<p>Those are my embedded example sentences.<\/p>\n\n<pre><\/pre>\n\n<p>Using  I get desired single output for the combined sentences.<\/p>\n\n<pre><\/pre>\n\n<p>I can even use<\/p>\n\n<pre><\/pre>\n\n<p>to train the model, where<\/p>\n\n<pre><\/pre>\n\n<p><strong>The question is<\/strong> how can I format examples  and  so that I can use it in fit function.  didn't work, yields<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41915669","CreationDate":"2017-01-28T22:42:20.980","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So, following the Keras <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"nofollow noreferrer\">Embedding<\/a> documentation you need to provide as an input to Embedding layer a <code>numpy.array<\/code> of dimension 2. Because you have a multiple inputs - you need to provide a list of numpy arrays in order to feed them to multiple inputs.<\/p>\n\n<p>In the last case however your are feeding a list of lists of inputs. This caused a problem. In order to solve this you need to <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.concatenate.html\" rel=\"nofollow noreferrer\"><code>numpy.concatenate<\/code><\/a> these two arrays by:<\/p>\n\n<pre><code>x = numpy.concatenate((x1, x2), axis = 0)\ny = numpy.concatenate((y1, y2), axis = 0)\n<\/code><\/pre>\n"}]}
{"QuestionId":41915056,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-28T21:33:30.950","AcceptedAnswerId":null,"OwnerUserId":7484188.0,"Title":"Saving LSTM RNN state between runs in Tensorflow","Body":"<p>What's the best way to save the LSTM state between runs in Tensorflow? For the prediction phase, I need to pass in data one timestep at a time because the input of the next timestep relies on the output of the previous timestep.<\/p>\n\n<p>I used the suggestion from this post: <a href=\"https:\/\/stackoverflow.com\/questions\/37969065\/tensorflow-best-way-to-save-state-in-rnns\">Tensorflow, best way to save state in RNNs?<\/a> and tested it by passing in the same input over and over again without running the optimizer. If I understand correctly, if the output changes each time then it is saving the state but if it stays the same then it isn't. The result was that it saves the state the first time but then stays the same.<\/p>\n\n<p>Here's my code:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the error over time:<\/p>\n\n<ul>\n<li>0: 419.861<\/li>\n<li>1: 419.756<\/li>\n<li>2: 419.756<\/li>\n<li>3: 419.756\n...<\/li>\n<\/ul>\n","answers":[{"AnswerId":"46971322","CreationDate":"2017-10-27T09:04:22.383","ParentId":null,"OwnerUserId":"8842526","Title":null,"Body":"<p>I recommend you <a href=\"https:\/\/stackoverflow.com\/questions\/44703593\/tensorflow-how-to-access-all-the-middle-states-of-an-rnn-not-just-the-last-sta\">this answer<\/a> which i tried few days ago. It works well.<\/p>\n\n<p>By the way, there's a way avoid setting <code>state_is_tuple<\/code> to <code>false<\/code>:<\/p>\n\n<pre><code>class CustomLSTMCell(tf.contrib.rnn.LSTMCell):\n    def __init__(self, *args, **kwargs):\n        # kwargs['state_is_tuple'] = False # force the use of a concatenated state.\n        returns = super(CustomLSTMCell, self).__init__(\n            *args, **kwargs)  # create an lstm cell\n        # change the output size to the state size\n        self._output_size = np.sum(self._state_size)\n        return returns\n\n    def __call__(self, inputs, state):\n        output, next_state = super(\n            CustomLSTMCell, self).__call__(inputs, state)\n        # return two copies of the state, instead of the output and the state\n        return tf.reshape(next_state, shape=[1, -1]), next_state\n<\/code><\/pre>\n"}]}
{"QuestionId":41915076,"AnswerCount":0,"Tags":"<tensorflow><gpu><ubuntu-16.04><bazel>","CreationDate":"2017-01-28T21:35:11.810","AcceptedAnswerId":null,"OwnerUserId":5241351.0,"Title":"bazel failing to install tensorflow with this error java installation could't find\/access rt.jar in \/usr\/lib\/jvm\/java-9-oracle","Body":"<p>While trying to install tensorflow with gpu on my ubuntu 16.04 box with python 3.5 and cuda 7.5 and cudann 5.1, on configuring I am getting below error. Can some one suggest how to fix this?\n<a href=\"https:\/\/i.stack.imgur.com\/zO8LZ.png\" rel=\"nofollow noreferrer\">Error seen<\/a><\/p>\n\n<p>By the way tried with  lower version of bazel aswell 0.3.2 but still no help. Seems this has been reported by more people... Need google to resolve this issue with Bazel or tensorflow.<\/p>\n","answers":[]}
{"QuestionId":41915775,"AnswerCount":1,"Tags":"<python><python-3.x><optimization><machine-learning><tensorflow>","CreationDate":"2017-01-28T22:54:32.897","AcceptedAnswerId":null,"OwnerUserId":6348198.0,"Title":"ValueError Due to the Input Being the Wrong Size for AdamOptimizer.compute_gradients","Body":"<p>I'm not sure why, but I am getting the following error when calling  on my  with the learning rate set at :<\/p>\n\n<pre><\/pre>\n\n<p>Here is the relevant code segment:<\/p>\n\n<pre><\/pre>\n\n<p>And that print statement prints out:<\/p>\n\n<pre><\/pre>\n\n<p>I understand what this error means (loss is the wrong size). However, <strong>how do I fix this problem.<\/strong> <\/p>\n\n<p>Here is my full code (I am attempting to build a dynamic convolutional neural network):<\/p>\n\n<pre><\/pre>\n\n<p>And here is the second file (sorry about the indenting).<\/p>\n\n<pre><\/pre>\n\n<p>The training data I use is the Rotten Tomatoes data set of labeled positive and negative movie reviews.<\/p>\n","answers":[{"AnswerId":"41916633","CreationDate":"2017-01-29T01:10:09.680","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>It sounds like there's a problem here self.embedded_chars = tf.nn.embedding_lookup(W, self.input_x) <\/p>\n\n<p>The input from rotten tomatoes dataset might not fit one of your placeholder variables.  I would double check that.<\/p>\n"}]}
{"QuestionId":41915862,"AnswerCount":1,"Tags":"<theano><keras><keras-layer>","CreationDate":"2017-01-28T23:05:32.033","AcceptedAnswerId":41916751.0,"OwnerUserId":2442740.0,"Title":"Theano \/ Keras: Set the K-first values of Tensor to a value","Body":"<p>I have a custom Keras layer and i use Theano as the backend and i want to do the following operation: <\/p>\n\n<p>Suppose we have a tensor with shape (N,). I want to set the K first values to a fixed value  (3 or whatever...). How do i do that? I assume i have to use argsort but i don't know how to implement it using Theano.<\/p>\n\n<p>For example in a simple FF layer, how can i set the first N values of the tensor  to a value ? <\/p>\n\n<pre><\/pre>\n\n<p>For example using numpy i can set the 10 first values of  to 1 like this:<\/p>\n\n<pre><\/pre>\n\n<p>I want to do the same using Keras backend functions or Theano functions.<\/p>\n","answers":[{"AnswerId":"41916751","CreationDate":"2017-01-29T01:31:21.710","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>For <code>a[a.argsort() &lt;= 10] = 1<\/code> the equivalent code will be:<\/p>\n\n<pre><code>import numpy as np\nfrom keras import backend as K\n\na = np.asarray([[3,4,2,44,22,4,5,6,77,86,3,2,3,23,44,21],\n                [3,4,22,44,2,4,54,6,77,8,3,2,36,23,4,2]], dtype=np.float)\na_t = K.variable(a)\n\na[a.argsort() &lt;= 10] = 1\nprint a\n\narg_sort = K.T.argsort(a_t)\n_cond = K.lesser_equal(arg_sort,10)\na_new = K.switch(_cond, 1, a_t)\nprint K.eval(a_new)\n<\/code><\/pre>\n\n<p>In a single statement it will be:<\/p>\n\n<pre><code>a_new = K.switch(K.lesser_equal(K.T.argsort(a_t),10), 1, a_t)\nprint K.eval(a_new)\n<\/code><\/pre>\n\n<p>I am not sure if this is exactly what you need.<\/p>\n"}]}
{"QuestionId":41916634,"AnswerCount":2,"Tags":"<python><python-3.x><machine-learning><tensorflow><conv-neural-network>","CreationDate":"2017-01-29T01:10:12.297","AcceptedAnswerId":null,"OwnerUserId":6348198.0,"Title":"How to Properly Implement a Dynamic Convolutional Neural Network using TensorFlow","Body":"<p>I know that this is a very broad question, but I have asked many other questions and I have still been unable to properly implement a <em>simple<\/em> dynamic-k max pooling convolutional neural network as described in <a href=\"http:\/\/www.aclweb.org\/anthology\/P14-1062\" rel=\"nofollow noreferrer\">this<\/a> paper. Currently, I am trying to modify the code from <a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"nofollow noreferrer\">this<\/a> tutorial. I believe I have successfully implemented the dynamic-k part. However, my main problem is because the k value is different for each input, the tensors that are produced are different shapes. I have tried countless things to try and fix this (which is why you may see some funny reshaping), but I can't figure out how. I think that you'd need to pad each tensor to get them all to be the size of the biggest one, but I can't seem to get that to work. Here is my code (I am sorry, it is generally rather sloppy). <\/p>\n\n<pre><\/pre>\n\n<p>And here is the actual DCNN class:<\/p>\n\n<pre><\/pre>\n\n<p>I am using the Rotten Tomatoes sentiment labeled data set. The current error I am getting is this:<\/p>\n\n<pre><\/pre>\n\n<p><strong>How can I fix this code so that all of the tensors are normalized to the same size after pooling (while keeping pooling dynamic) and so that the code runs to completion?<\/strong><\/p>\n\n<p>Sorry about all of the random commented out lines and prints and stuff, but I have tried extensively to make this work. <\/p>\n","answers":[{"AnswerId":"45214488","CreationDate":"2017-07-20T12:18:22.993","ParentId":null,"OwnerUserId":"7765065","Title":null,"Body":"<p>Although tensorflow doesn't provide k-max pooling directly, I think tf.nn.top_k might help you build that op.<\/p>\n"},{"AnswerId":"42574629","CreationDate":"2017-03-03T09:04:47.990","ParentId":null,"OwnerUserId":"3480922","Title":null,"Body":"<p>There are three things to note here.<\/p>\n\n<ol>\n<li><code>max-pooling<\/code> and <code>k-max pooling<\/code> are two different operations. <\/li>\n<\/ol>\n\n<blockquote>\n  <p>max-pooling retrieves the maximum valued activation out of the pooling window while k-max pooling retrieves k maximum values from the pooling window.<\/p>\n<\/blockquote>\n\n<ol start=\"2\">\n<li><p>Tensorflow doesn't provide API for k-max pooling as of now. The one\nwhich you are trying now is max-pooling operation and not k-max\npooling operation.<\/p><\/li>\n<li><p>As per my knowledge, tensorflow does not provide functionality to handle pooling resulting in different size of matrices. So, you may use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/contrib.training\/bucketing\" rel=\"nofollow noreferrer\">bucketing to create batches of sentences of similar length<\/a> and the use k-max pooling.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":41917424,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-29T03:48:16.320","AcceptedAnswerId":null,"OwnerUserId":5674154.0,"Title":"n-cumulative sum in Tensorflow","Body":"<p>I want to get a n-cumulative sum of a vector in <strong>tensorflow<\/strong>.<\/p>\n\n<pre><\/pre>\n\n<p>That is,<\/p>\n\n<p><strong>INPUT<\/strong><\/p>\n\n<p> = [1, 3, 5, 8]<\/p>\n\n<p> = 2<\/p>\n\n<p><strong>OUTPUT<\/strong><\/p>\n\n<p> = [1+3, 3+5, 5+8, 8]<\/p>\n\n<p>For another example,<\/p>\n\n<p><strong>INPUT<\/strong><\/p>\n\n<p> = [1,5,6,2,8,7,9]<\/p>\n\n<p> = 3<\/p>\n\n<p><strong>OUTPUT<\/strong><\/p>\n\n<p> = [1+5+6, 5+6+2, 6+2+8, 2+8+7, 8+7+9, 7+9, 9]<\/p>\n\n<p>What should I use for ?<\/p>\n","answers":[{"AnswerId":"41928330","CreationDate":"2017-01-30T02:37:56.753","ParentId":null,"OwnerUserId":"691733","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#while_loop\" rel=\"nofollow noreferrer\"><code>tf.while_loop<\/code><\/a> is a handy function for such things. Here is the complete working code :<\/p>\n\n<pre><code>import tensorflow as tf\ninput = tf.placeholder('float32', [None])\nn = tf.placeholder('int32', ())\nsess = tf.InteractiveSession()\n\ntmp = tf.concat_v2([input,\n                    tf.zeros(tf.expand_dims(n-1, 0),\n                             dtype='float32')], axis=0)\ni = tf.constant(0, dtype='int32')\noutput = tf.zeros([0], dtype='float32')\n\ndef body(i, output):\n  return i + 1, tf.concat_v2([output,\n                              tf.expand_dims(tf.reduce_sum(tmp[i:i+n]), 0)],\n                             axis=0)\n\ni, output = tf.while_loop(lambda i, _: i &lt; tf.shape(input)[0],\n                          body,\n                          [i, output],\n                          shape_invariants=[tf.TensorShape([]),\n                                            tf.TensorShape([None])])\n\noutput.eval(feed_dict={n: 2, input:[1, 3, 5, 8]})\n# array([  4.,   8.,  13.,   8.], dtype=float32)\n\noutput.eval(feed_dict={n: 3, input:[1,5,6,2,8,7,9]})\n# array([ 12.,  13.,  16.,  17.,  24.,  16.,   9.], dtype=float32)\n<\/code><\/pre>\n"}]}
{"QuestionId":41917491,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-29T04:02:01.050","AcceptedAnswerId":41918273.0,"OwnerUserId":4704204.0,"Title":"Predict with TensorFlow MNIST for Experts","Body":"<p>I am following the TensorFlow MNIST for Experts tutorial, but I do not know how to get the trained network to predict a new set of data. <\/p>\n\n<p>Below is my code: I have all the lines in <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/\" rel=\"nofollow noreferrer\">TensorFlow MNIST for Experts tutorial<\/a> and I imported a csv file with pandas as dataframe testt. <\/p>\n\n<pre><\/pre>\n\n<p>I get this error<\/p>\n\n<pre><\/pre>\n\n<p>Please help. I am not sure how to properly call the trained network. <\/p>\n","answers":[{"AnswerId":"41918273","CreationDate":"2017-01-29T06:29:49.277","ParentId":null,"OwnerUserId":"2059708","Title":null,"Body":"<p>You just need to get the output of y with a <code>sess.run<\/code> call after your training loop:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    for i in range(1000):\n        batch = mnist.train.next_batch(100)\n        train_step.run(feed_dict={x: batch[0], y_: batch[1]})\n\n    logits = sess.run(y, feed_dict={x: test_data})\n    pred = tf.nn.softmax(logits)\n    print(pred)  # or print(tf.argmax(pred, dimension=1)) for the index\n<\/code><\/pre>\n"}]}
{"QuestionId":41918795,"AnswerCount":2,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2017-01-29T07:58:11.950","AcceptedAnswerId":41948131.0,"OwnerUserId":2150703.0,"Title":"Minimize a function of one variable in Tensorflow","Body":"<p>I am new to Tensorflow and was wondering whether it would be possible to minimize a function of one variable using Tensorflow.<\/p>\n\n<p>For example, can we use Tensorflow to minimize 2*x^2 - 5^x + 4 using an initial guess (say x = 1)?<\/p>\n\n<p>I am trying the following:<\/p>\n\n<pre><\/pre>\n\n<p>But this gives the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Please help me understand what am I doing wrong here.<\/p>\n","answers":[{"AnswerId":"41948131","CreationDate":"2017-01-31T01:02:34.717","ParentId":null,"OwnerUserId":"2888347","Title":null,"Body":"<p>If you want to minimize a single parameter you could do the following (I've avoided using a placeholder since you are trying to train a parameter - placeholders are often used for hyper-parameters and input and aren't considered trainable parameters):<\/p>\n\n<pre><code>import tensorflow as tf\n\nx = tf.Variable(10.0, trainable=True)\nf_x = 2 * x* x - 5 *x + 4\n\nloss = f_x\nopt = tf.train.GradientDescentOptimizer(0.1).minimize(f_x)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    for i in range(100):\n        print(sess.run([x,loss]))\n        sess.run(opt)\n<\/code><\/pre>\n\n<p>This will output the following list of pairs (x,loss):<\/p>\n\n<pre><code>[10.0, 154.0]\n[6.5, 56.0]\n[4.4000001, 20.720001]\n[3.1400001, 8.0192013]\n[2.3840001, 3.4469128]\n[1.9304, 1.8008881]\n[1.65824, 1.2083197]\n[1.494944, 0.99499512]\n[1.3969663, 0.91819811]\n[1.3381798, 0.89055157]\n[1.3029079, 0.88059855]\n[1.2817447, 0.87701511]\n[1.2690468, 0.87572551]\n[1.2614281, 0.87526155]\n[1.2568569, 0.87509394]\n[1.2541142, 0.87503386]\n[1.2524685, 0.87501216]\n[1.2514811, 0.87500429]\n[1.2508886, 0.87500143]\n[1.2505331, 0.87500048]\n[1.2503198, 0.875]\n[1.2501919, 0.87500024]\n[1.2501152, 0.87499976]\n[1.2500691, 0.875]\n[1.2500415, 0.875]\n[1.2500249, 0.87500024]\n[1.2500149, 0.87500024]\n[1.2500089, 0.875]\n[1.2500054, 0.87500024]\n[1.2500032, 0.875]\n[1.2500019, 0.875]\n[1.2500012, 0.87500024]\n[1.2500007, 0.87499976]\n[1.2500005, 0.875]\n[1.2500002, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n[1.2500001, 0.87500024]\n<\/code><\/pre>\n"},{"AnswerId":"57560560","CreationDate":"2019-08-19T16:21:04.993","ParentId":null,"OwnerUserId":"3968093","Title":null,"Body":"<p>Here is the same solution with the same output using tensorflow 2.0<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.training import gradient_descent\n\nx = tf.Variable(10.0, trainable=True)\n\n\n@tf.function\ndef f_x():\n    return 2 * x * x - 5 * x + 4\n\n\nfor _ in range(100):\n    print([x.numpy(), f_x().numpy()])\n    opt = gradient_descent.GradientDescentOptimizer(0.1).minimize(f_x)\n<\/code><\/pre>\n"}]}
{"QuestionId":41918854,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2017-01-29T08:08:17.320","AcceptedAnswerId":41919139.0,"OwnerUserId":1934212.0,"Title":"Visualizing results of preprocessing layers in keras","Body":"<p>I have several image preprocessing steps built into my model:<\/p>\n\n<pre><\/pre>\n\n<p>The benefit of that approach is that the pipeline for training and testing is exactly the same. The disadvantage is that its rather difficult to verify the results of the preprocessing steps, e.g. by watching the images after the cropping step.<\/p>\n\n<p>Is there a good way to do that in keras ?<\/p>\n","answers":[{"AnswerId":"41919139","CreationDate":"2017-01-29T08:51:28.727","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can easily get the outputs of any layer by using: <code>model.layers[index].output<\/code> and using <code>K.function<\/code> create a tensor function out of it and then evaluate the function and visualize the output. <\/p>\n\n<p>Check:\n<a href=\"https:\/\/stackoverflow.com\/questions\/41711190\/keras-how-to-get-the-output-of-each-layer\/41712013#41712013\">Keras, How to get the output of each layer?<\/a>  <\/p>\n"}]}
{"QuestionId":41919899,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-29T10:31:33.660","AcceptedAnswerId":41964487.0,"OwnerUserId":3113510.0,"Title":"Tensorflow slim and assertion error","Body":"<p>I am new to tensorflow and have been experimenting with slim experiment. I have tried to translate the MNIST tutorial in the tensorflow tutorials to the slim syntax. It had been working ok with an unbatched set of images fed to the model. I then added a tf.train_batch thread to the code and it stopped working when i run the entire file. giving the error <\/p>\n\n<pre><\/pre>\n\n<p>However, i can selectively re run create_train_op line and then train the model, although the loss function does not decrease here and essentially it doesn't work. This still allows me to get the graph visualisation from tensorboard (attached below) and I cannot see any errors in this. <\/p>\n\n<p>I know I am doing something wrong, but am not sure where it is.   <\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wHHbn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wHHbn.png\" alt=\"enter image description here\"><\/a> <\/p>\n","answers":[{"AnswerId":"41964487","CreationDate":"2017-01-31T18:05:04.067","ParentId":null,"OwnerUserId":"6785153","Title":null,"Body":"<p>You need to create all ops under the same graph, that includes the input data<\/p>\n\n<pre><code>with tf.Graph().as_default():\n  mnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=False)\n  batch = mnist.train.next_batch(20000)\n\n  x_image = tf.reshape(batch[0], [-1,28,28,1])\n  label = tf.one_hot(batch[1], 10)\n\n  image, labels = tf.train.batch([x_image[0], label[0]], batch_size= 100)\n\n  tf.logging.set_verbosity(tf.logging.DEBUG)\n  logits, _ = model(image)\n  predictions = tf.nn.softmax(logits)\n  loss = slim.losses.softmax_cross_entropy(predictions, labels)\n  config = tf.ConfigProto()\n  optimiser = tf.train.AdamOptimizer(1e-4)\n  train_op = slim.learning.create_train_op(loss, optimiser)\n  thisloss = slim.learning.train(train_op, \"C:\/temp\/test2\", number_of_steps=100, save_summaries_secs=30, session_config=config)\n<\/code><\/pre>\n"}]}
{"QuestionId":41919912,"AnswerCount":1,"Tags":"<python><machine-learning><keras><theano-cuda>","CreationDate":"2017-01-29T10:33:21.780","AcceptedAnswerId":null,"OwnerUserId":7346966.0,"Title":"Keras unsymmetrical data Diabetic Retinopathy Detection","Body":"<p>I'm trying to make a predictive model for <a href=\"https:\/\/www.kaggle.com\/c\/diabetic-retinopathy-detection\" rel=\"nofollow noreferrer\">Diabetic Retinopathy Detection<\/a>. The competition's trainig dataset includes hy-res images are unsymmetricaly divided in 5 classes: Normal-25807 images-73.48%; Mild-2442 images-6.96%; Moderate-5291 images-15.07%; Severe-873 images-2.48% and Proliferative-708 images - 2.01%.\nFor this purpose I use Keras framework with Theano backend (for CUDA comutations). <\/p>\n\n<p>For image augmentation I used the ImageDataGenerator (the code is below). I've resized images to 299x299 and divided them into 5 folders accordingly their classes:<\/p>\n\n<pre><\/pre>\n\n<p>At first, just for testing, I desided to use a simple convolutional model:<\/p>\n\n<pre><\/pre>\n\n<p>In fitting Image generator, I pointed the class_weights in order to fix the asymmetry of data: class_weight ={0: 25807., 1:2442., 2:5291., 3:873., 4:708.};<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/KVQr9.png\" rel=\"nofollow noreferrer\">My folders with images<\/a><\/p>\n\n<p>Problems:<\/p>\n\n<ol>\n<li>The model outputs with high loss and high accuracy. Why?<\/li>\n<\/ol>\n\n<p>Epoch 1\/50\n110s - loss: 5147.2669 - acc: 0.7366<\/p>\n\n<p>Epoch 2\/50\n105s - loss: 5052.3844 - acc: 0.7302<\/p>\n\n<p>Epoch 3\/50\n105s - loss: 5042.0261 - acc: 0.7421<\/p>\n\n<p>Epoch 4\/50\n105s - loss: 4986.3544 - acc: 0.7361<\/p>\n\n<p>Epoch 5\/50\n105s - loss: 4999.4177 - acc: 0.7361<\/p>\n\n<ol start=\"2\">\n<li>Every image model predict as '0' class:<\/li>\n<\/ol>\n\n<p>datagen_2=ImageDataGenerator(rescale=1.\/255)<\/p>\n\n<pre><\/pre>\n\n<p>the output of it is:<\/p>\n\n<pre><\/pre>\n\n<p>without argmax(partly)<\/p>\n\n<pre><\/pre>\n\n<p>If I've tried to use class_weight ='auto'. In this case, model showed 'predictable' output:<\/p>\n\n<p>Epoch 1\/50\n107s - loss: 0.9036 - acc: 0.7381<\/p>\n\n<p>Epoch 2\/50\n104s - loss: 0.9333 - acc: 0.7321<\/p>\n\n<p>Epoch 3\/50\n105s - loss: 0.8865 - acc: 0.7351<\/p>\n\n<p>Epoch 4\/50\n106s - loss: 0.8978 - acc: 0.7351<\/p>\n\n<p>Epoch 5\/50\n105s - loss: 0.9158 - acc: 0.7302<\/p>\n\n<p>But, it still doesn't work:<\/p>\n\n<pre><\/pre>\n\n<p>What I've done wrong?<\/p>\n\n<hr>\n\n<p>After answer of Sergii Gryshkevych, I fixed my model: I've changed class_weight to {0:1, 1:10.57, 2:4.88, 3:29, 4:35} (I divided images in each classes to maximum images (in first class)). Next, I changed metrics to categorical_accuracy. And I inctreased the number of layers in model (like <a href=\"https:\/\/deepsense.io\/diagnosing-diabetic-retinopathy-with-deep-learning\/\" rel=\"nofollow noreferrer\">here<\/a>). \nSo, the output after 5 epochs is:<\/p>\n\n<p>Epoch 1\/5\n500\/500 [==============================] - 52s - loss: 5.6944 - categorical_accuracy: 0.1840<br>\nEpoch 2\/5\n500\/500 [==============================] - 52s - loss: 6.7357 - categorical_accuracy: 0.2040<br>\nEpoch 3\/5\n500\/500 [==============================] - 52s - loss: 6.7373 - categorical_accuracy: 0.0800<br>\nEpoch 4\/5\n500\/500 [==============================] - 52s - loss: 6.0311 - categorical_accuracy: 0.0180<br>\nEpoch 5\/5\n500\/500 [==============================] - 51s - loss: 4.9924 - categorical_accuracy: 0.0560   <\/p>\n\n<p>Is it correct?<\/p>\n\n<p>Is there any way to make assign a quadratic weighted kappa as metrics in keras?<\/p>\n","answers":[{"AnswerId":"41920547","CreationDate":"2017-01-29T11:48:04.257","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>\"High\" accuracy around 73-74% comes from the fact that all images are classified as <code>0<\/code> class. Your data set is imbalanced, since the majority class accounts for 73% of samples. So accuracy does not say much in this case, you need to use other metrics derived from confusion matrix like precision, recall, F1 score, etc. <\/p>\n\n<p>Multiclass log loss function punishes wrong predictions extremely. Your predictions are almost zero for all classes except the <code>0<\/code>, so there is nothing surprising in such high loss values.<\/p>\n\n<p>To sum up, you are facing classic Class Imbalance Problem. Two most common ways to mitigate it are<\/p>\n\n<ol>\n<li>Adjust class weights. Make minority classes \"more important\", so learning algorithm does not ignore them. You can provide your custom class weights as an argument to the <code>fit<\/code> method:<\/li>\n<\/ol>\n\n<blockquote>\n  <p><code>class_weight<\/code>: dictionary mapping classes to a weight value, used for scaling the loss function (during training only).<\/p>\n<\/blockquote>\n\n<ol start=\"2\">\n<li>Oversampling\/Undersampling. Simply oversample examples of minority classes to make your data set balanced, or combine it with undersampling, when some randomly selected examples of the majority class are dropped at the beginning of each learning epoch. <\/li>\n<\/ol>\n\n<p>The Class Imbalance problem is nothing new, so there is plenty of reading on this topic like <a href=\"http:\/\/www.chioka.in\/class-imbalance-problem\/\" rel=\"nofollow noreferrer\">this<\/a> and <a href=\"http:\/\/machinelearningmastery.com\/tactics-to-combat-imbalanced-classes-in-your-machine-learning-dataset\/\" rel=\"nofollow noreferrer\">this<\/a> introductory posts. <\/p>\n"}]}
{"QuestionId":41920166,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-29T11:03:52.933","AcceptedAnswerId":41920594.0,"OwnerUserId":3102085.0,"Title":"AtributeError:int object has no attribute name","Body":"<p>Following is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Here I am applying thresholding on result. If its value is greater than 0.2 then assign 1 otherwise assign 0. But I keep getting this error. Kindly tell what I am doing wrong.<\/p>\n\n<p>Traceback:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41920594","CreationDate":"2017-01-29T11:53:42.107","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>The callbacks should return tensors, not ints. Try:<\/p>\n\n<pre><code>one = tf.constant(1, dtype=tf.int32, name='one')\nzero = tf.constant(0, dtype=tf.int32, name='zero')\n<\/code><\/pre>\n\n<p>and inside the class:<\/p>\n\n<pre><code>def truef(self):\n    return one\n\ndef falsef(self):\n    return zero\n<\/code><\/pre>\n"}]}
{"QuestionId":41920371,"AnswerCount":2,"Tags":"<multithreading><tensorflow>","CreationDate":"2017-01-29T11:27:05.453","AcceptedAnswerId":41935825.0,"OwnerUserId":2547570.0,"Title":"Tensorflow Multi-threaded QueueRunner","Body":"<p>I have a graph:<\/p>\n\n<pre><\/pre>\n\n<p>which is a typical Data I\/O pipeline.<\/p>\n\n<p>Problem: The multiple threads for tf.train.batch() have a race condition.\ne.g. Thread1 fetches sample1_img and sample2_label because Thread2 already took sample1_label, making a pair (sample1_img, sample2_label). I guess this is because slice_input_producer has 2 separate queues for imgs and labels and two queues work independently.<\/p>\n\n<p>Q1. Does each of n enqueueing threads run its own replica of the subgraph? If yes, setting num_threads=n requires n times more memory for the corresponding subgraph in runtime? If no, do threads run the different part of the subgraph for one enqueue op?<\/p>\n\n<p>Q2 (solved). If I create a FIFOQueue and enqueue a tuple of (img, label), the pair will be atomically dequeued and multi-threading will actually help. Is this correct? (although it's not 100% utilization because label tensor waits for the preprocessing of img tensor)<\/p>\n\n<p>Q3 (solved). Is there a function like tuple_input_producer() which takes a list of tensors and internally use only one queue?<\/p>\n\n<hr>\n\n<p><strong>Update<\/strong> (Q2,Q3)<\/p>\n\n<p>I was wrong about slice_input_tensor.\nThe problem only happens with two queues, not with slice_input_producer.<\/p>\n\n<p>So just use slice_input_producer, and if two tensors need to go into different queues, I can use a single-threaded bottleneck (QueueRunner) to bundle them together.<\/p>\n\n<p>Example code (0.11):<\/p>\n\n<pre><\/pre>\n\n<p>Result (see the first line):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41935825","CreationDate":"2017-01-30T12:20:37.723","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You have parallel run calls, called \"steps\", and each step maintains their own copy of tensors produced during execution. So in the worst case you need N times more memory for N parallel steps. In practice it tends to be better than times N because memory is released as soon as tensor is not needed anymore. Stateful objects like queues and variables are shared across steps.<\/p>\n\n<p>What's happening in your case is the following scenario:<\/p>\n\n<pre><code>step1: dequeue queue1\nstep2: dequeue queue1\nstep2: dequeue queue2\nstep1: dequeue queue2\n<\/code><\/pre>\n\n<p>You can see that queues get out of sync for both steps. Two ways to avoid it:<\/p>\n\n<ol>\n<li>Don't issue parallel run calls (num_threads=1)<\/li>\n<li>Combine two queues into a single queue with images\/labels, and dequeue from that queue in parallel.<\/li>\n<\/ol>\n\n<p>In the second example you would have a single <code>dequeue<\/code> op return a pair of <code>image\/label<\/code>, and things should stay synchronized because <code>dequeues<\/code> are atomic<\/p>\n"},{"AnswerId":"45161383","CreationDate":"2017-07-18T08:29:31.367","ParentId":null,"OwnerUserId":"2281593","Title":null,"Body":"<p>I've tried your Update's example code with tf=0.12.1<\/p>\n\n<p>when the iteration changes from range(10) to range(100), the mismatch also happend.<\/p>\n\n<p>If I modified the <code>tf.train.batch<\/code> parameter with num_threads=1, it solved.<\/p>\n\n<p>I've not tried <code>Combine two queues into a single queue with images\/labels, and dequeue from that queue in parallel.<\/code> yet.<\/p>\n"}]}
{"QuestionId":41920834,"AnswerCount":1,"Tags":"<python><machine-learning><keras>","CreationDate":"2017-01-29T12:21:25.520","AcceptedAnswerId":41921584.0,"OwnerUserId":6391711.0,"Title":"Error in fitting the model","Body":"<p>I am trying to code digit Recognizer. I am having the dataset that contains pixel data of image having dimensions 60000*28*28 where 60000 is number of images and 28 is width and height in pixels.<\/p>\n\n<pre><\/pre>\n\n<p>I am getting following error.<\/p>\n\n<pre><\/pre>\n\n<p>I am not getting what's the problem. Please Help.<\/p>\n","answers":[{"AnswerId":"41921584","CreationDate":"2017-01-29T13:45:14.730","ParentId":null,"OwnerUserId":"6148139","Title":null,"Body":"<p>Seems like error related to output shape.\nAs I can see by NN code \n<code>\nclassifier.add(Dense(output_dim = 10, activation= 'softmax'))\n<\/code>\nOutput must have shape <code>[recordCount, 10]<\/code>.\nBut when I run python console and typed next - I see wrong <code>y_train<\/code> shape<\/p>\n\n<pre><code>&gt;&gt;&gt; from keras.datasets import mnist\nUsing Theano backend.\nUsing gpu device 0: GeForce GT 730 (CNMeM is enabled with initial size: 70.0% of memory, cuDNN not available)\n&gt;&gt;&gt; (x_train, y_train), (x_test, y_test) = mnist.load_data()\nDownloading data from https:\/\/s3.amazonaws.com\/img-datasets\/mnist.pkl.gz\n15253504\/15296311 [============================&gt;.] - ETA: 0s&gt;&gt;&gt;\n&gt;&gt;&gt; x_train.shape\n(60000, 28, 28)\n&gt;&gt;&gt; y_train.shape\n(60000,)\n<\/code><\/pre>\n\n<p>And values of y_train in <code>0..9<\/code> range. So seems like I can suggest yoy to make next transform:<\/p>\n\n<pre><code>&gt;&gt;&gt; import numpy\n&gt;&gt;&gt; y_train_new = numpy.zeros([60000, 10])\n&gt;&gt;&gt; for i in range(0, 10):\n...     y_train_new[:, i] = (y_train == i).astype(numpy.int32)\n<\/code><\/pre>\n"}]}
{"QuestionId":41921620,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2017-01-29T13:49:51.467","AcceptedAnswerId":41925047.0,"OwnerUserId":1934212.0,"Title":"Verifying the output of an isolated layer in keras","Body":"<p>Is it possible to verify the output of<\/p>\n\n<pre><\/pre>\n\n<p>Without constructing and training a model? I.e. I just want to pass and image to Cropping and get and display the output image. <\/p>\n","answers":[{"AnswerId":"41925047","CreationDate":"2017-01-29T19:30:55.527","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>Yes, this is described in the Keras <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">FAQ<\/a>. To quote:<\/p>\n\n<pre><code>from keras import backend as K\n\n# with a Sequential model\nget_3rd_layer_output = K.function([model.layers[0].input],\n                                  [model.layers[3].output])\nlayer_output = get_3rd_layer_output([X])[0]\n<\/code><\/pre>\n\n<p>This example assumes that your <code>Cropping<\/code> layer has index 3. You need to substitute this index with the correct number according to your model.<\/p>\n"}]}
{"QuestionId":41921746,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-29T14:04:55.050","AcceptedAnswerId":42603692.0,"OwnerUserId":2452761.0,"Title":"Tensorflow VarLenFeature vs FixedLenFeature","Body":"<p>I was trying to save images of different sizes into tf-records. I found that even though the images have different sizes, I can still load them with . <\/p>\n\n<p>By checking the docs on  and ,  I found the difference seems to be that  returns a sparse tensor.<\/p>\n\n<p>Could anyone illustrate some situations one should use  or ?<\/p>\n","answers":[{"AnswerId":"42603692","CreationDate":"2017-03-05T02:13:17.400","ParentId":null,"OwnerUserId":"3970726","Title":null,"Body":"<p>You can load images probably beacause you saved them using feature type <code>tf.train.BytesList()<\/code> and whole image data is one big byte value inside a list.<\/p>\n\n<p>If I'm right you're using <code>tf.decode_raw<\/code> to get the data out of the image you load from TFRecord.<\/p>\n\n<p>Regarding example use cases:\nI use <code>VarLenFeature<\/code> for saving datasets for object detection task:\nThere's variable amount of bounding boxes per image (equal to object in image) therefore I need another feature <code>objects_number<\/code> to track amount of objects (and bboxes).\nEach bounding box itself is a list of 4 float coordinates<\/p>\n\n<p>I'm using following code to load it:<\/p>\n\n<pre><code>features = tf.parse_single_example(\n    serialized_example,\n    features={\n        # We know the length of both fields. If not the\n        # tf.VarLenFeature could be used\n        'height': tf.FixedLenFeature([], tf.int64),\n        'width': tf.FixedLenFeature([], tf.int64),\n        'depth': tf.FixedLenFeature([], tf.int64),\n        # Label part\n        'objects_number': tf.FixedLenFeature([], tf.int64),\n        'bboxes': tf.VarLenFeature(tf.float32),\n        'labels': tf.VarLenFeature(tf.int64),\n        # Dense data\n        'image_raw': tf.FixedLenFeature([],tf.string)\n\n    })\n\n# Get metadata\nobjects_number = tf.cast(features['objects_number'], tf.int32)\nheight = tf.cast(features['height'], tf.int32)\nwidth = tf.cast(features['width'], tf.int32)\ndepth = tf.cast(features['depth'], tf.int32)\n\n# Actual data\nimage_shape = tf.parallel_stack([height, width, depth])\nbboxes_shape = tf.parallel_stack([objects_number, 4])\n\n# BBOX data is actually dense convert it to dense tensor\nbboxes = tf.sparse_tensor_to_dense(features['bboxes'], default_value=0)\n\n# Since information about shape is lost reshape it\nbboxes = tf.reshape(bboxes, bboxes_shape)\nimage = tf.decode_raw(features['image_raw'], tf.uint8)\nimage = tf.reshape(image, image_shape)\n<\/code><\/pre>\n\n<p>Notice that \"image_raw\" is fixed length Feature (has one element) and holds values of type \"bytes\", however a value of \"bytes\" type can itself have variable size (its a string of bytes, and can have many symbols within it).\nSo \"image_raw\" is a list with ONE element of type \"bytes\", which can be super big.<\/p>\n\n<p><strong>To further elaborate on how it works:<\/strong>\n<strong>Features<\/strong> are lists of  values, those values have specific \"type\".<\/p>\n\n<p>Datatypes for <strong>features<\/strong> are subset of data types for tensors, you have:<\/p>\n\n<ul>\n<li>int64 (64 bit space in memory)<\/li>\n<li>bytes (occupies as many bytes in memory as you want)<\/li>\n<li>float (occupies 32-64 bits in memory idk how much)<\/li>\n<\/ul>\n\n<p>You can check <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/dims_types#data_types\" rel=\"noreferrer\">here<\/a>  <strong>tensors<\/strong> data types.<\/p>\n\n<p>So you can store variable length data without <code>VarLenFeatures<\/code> at all (actually well you do it), but first you would need to convert it into bytes\/string feature, and then decode it. \nAnd this is most common method.<\/p>\n"}]}
{"QuestionId":41923418,"AnswerCount":1,"Tags":"<image-processing><machine-learning><classification><deep-learning><theano>","CreationDate":"2017-01-29T16:51:58.860","AcceptedAnswerId":null,"OwnerUserId":1525479.0,"Title":"Why test accuracy remains constant and do not increase in binary classification when test and train dataset are from different source","Body":"<p>I have <strong><em>train<\/em><\/strong> dataset and <strong><em>test<\/em><\/strong> dataset from two different sources. I mean they are from two different experiments but the results of both of them are same biological images. I want to do binary classification using <strong><em>deep CNN<\/em><\/strong> and I have following results on test accuracy and train accuracy. The blue line shows train accuracy and the red line shows test accuracy after almost <em>250<\/em> epochs. Why the test accuracy is almost constant and not raising?  Is that because <strong><em>Test<\/em><\/strong> and <strong><em>Train<\/em><\/strong> dataset are come from different distributions?<a href=\"https:\/\/i.stack.imgur.com\/gqcdU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gqcdU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Edited:\nAfter I have add dropout layer, reguralization terms and mean subtraction I still get following strange results which says the model is overfitting from the beginning!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iexkF.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iexkF.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/BpviO.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BpviO.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"41923659","CreationDate":"2017-01-29T17:12:33.920","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>There could be 2 reasons. First you overfit on the training data. This can be validated by using the validation score as a comparison metric to the test data. If so you can use standard techniques to combat overfitting, like weight decay and dropout. <\/p>\n\n<p>The second one is that your data is too different to be learned like this. This is harder to solve. You should first look at the value spread of both images. Are they both normalized. Matplotlib normalizes automatically for plotted images. If this still does not work you might want to look into augmentation to make your training data more similar to the test data. Here I can not tell you what to use, without seeing both the trainset and the testset.<\/p>\n\n<p>Edit:<\/p>\n\n<p>For normalization the test set and the training set should have a similar value spread. If you do dataset normalization you calculate mean and std on training set. But you also need to use those calculated values on the test set and not calculate the test set values from the test set. This only makes sense if the value spread is similar for both the training and test set. If this is not the case you might want to do per sample normalization first. <\/p>\n\n<p>Other augmentation that are commonly used for every dataset are oversampling, random channel shifts, random rotations, random translation and random zoom. This makes you invariante to those operations.<\/p>\n"}]}
{"QuestionId":41923478,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-29T16:58:02.243","AcceptedAnswerId":41966959.0,"OwnerUserId":4829166.0,"Title":"Dynamically tile a tensor depending on the batch size","Body":"<p>I have a 1D tensor a that I want to stack\/pack\/tile into a 2D tensor like . If I knew how many times I wanted it repeated, I could use  along with .<\/p>\n\n<p>But I don't because the size is dependent on the batch size. The placeholder value is  which isn't a valid input. I know for  one can input  and let tensorflow figure it out, but I don't see how tensorflow could infer the correct size. I do have a tensor  that would be equal in shape to , but I don't see a  function.<\/p>\n\n<p>Any suggestions?<\/p>\n","answers":[{"AnswerId":"57144947","CreationDate":"2019-07-22T11:18:40.013","ParentId":null,"OwnerUserId":"1283063","Title":null,"Body":"<p>Thanks, @Peter Hawkins for your great answer. In many cases, an additional reshape step is required in order to add the batch_size as the first dimension. I have added an optional extra step that reshapes the tensor:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nbatch_dependent_tensor = tf.placeholder(tf.float32, shape=[None, 3])\nother_tensor = tf.constant([2, 3])\n\ny = tf.tile(other_tensor, tf.shape(batch_dependent_tensor)[0:1])\n\n# Reshaping the y tensor by adding the batch_size as the first dimension\nnew_shape = tf.concat([tf.shape(batch_dependent_tensor)[0:1], tf.shape(other_tensor)[0:1]], axis=0)\ny_reshaped = tf.reshape(y, new_shape)\n\nsess = tf.Session()\ny_val, y_reshaped_val = sess.run([y, y_reshaped], feed_dict={batch_dependent_tensor: np.zeros([11, 3])})\nprint(\"y_val has shape %s, and value: %s\" %(y_val.shape, y_val))\nprint(\"y_reshaped_val has shape %s, and value: %s\" %(y_reshaped_val.shape, y_reshaped_val))\n\n\"\"\"\n# print output:\n\ny_val has shape (22,), and value: [2 3 2 3 2 3 2 3 2 3 2 3 2 3 2 3 2 3 2 3 2 3]\ny_reshaped_val has shape (11, 2), and value: [[2 3]\n [2 3]\n [2 3]\n [2 3]\n [2 3]\n [2 3]\n [2 3]\n [2 3]\n [2 3]\n [2 3]\n [2 3]]\n\"\"\"\n<\/code><\/pre>\n\n<p>Please note that if other_tensor has a rank > 1 (it is a matrix or a tensor with higher dimension) some modifications to the code are required.<\/p>\n"},{"AnswerId":"41966959","CreationDate":"2017-01-31T20:32:02.940","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>You can use <code>tf.shape<\/code> to find out the runtime shape of a tensor, and use it as the basis for the argument to <code>tf.tile<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nx = tf.placeholder(tf.float32, shape=[None, 3])\n\ny = tf.tile([2, 3], tf.shape(x)[0:1])\n\nsess = tf.Session()\nprint(sess.run(y, feed_dict={x: np.zeros([11, 3])}))\n<\/code><\/pre>\n\n<p>I verified this code works with the Tensorflow 1.0 release candidiate. Hope that helps!<\/p>\n"}]}
{"QuestionId":41923716,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-29T17:17:55.753","AcceptedAnswerId":null,"OwnerUserId":2820379.0,"Title":"Is it normal for dynamic_rnn() and rnn() to produce different results?","Body":"<p>In the documentation of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard8\/tf.nn.dynamic_rnn.md\" rel=\"nofollow noreferrer\">dynamic_rnn()<\/a> it says: <\/p>\n\n<blockquote>\n  <p>This function is functionally identical to the function  above,\n  but   performs fully dynamic unrolling of .<\/p>\n<\/blockquote>\n\n<p>When I use the <em>exact<\/em> same parameters for both dynamic_rnn() and rnn() I get slightly differing results (leaving aside that the input has to be provided in differing formats - the parameter content is the same). I noticed that for the <em>first<\/em> batch the first 16 batch rows give identical results (for both state and output). Everything else of the first and all other batches is unequal.<\/p>\n\n<p>Given the suspicious base 2 number 16 I suspect that it has to do with numerical issues but I would like some re-assurance on that if possible.<\/p>\n","answers":[]}
{"QuestionId":41924453,"AnswerCount":3,"Tags":"<python><torch><pytorch>","CreationDate":"2017-01-29T18:31:24.687","AcceptedAnswerId":42054194.0,"OwnerUserId":5831205.0,"Title":"PyTorch: How to use DataLoaders for custom Datasets","Body":"<p>How to make use of the  and  on your own data (not just the )?<\/p>\n\n<p>Is there a way to use the inbuilt  which they use on  to be used on any dataset?<\/p>\n","answers":[{"AnswerId":"42054194","CreationDate":"2017-02-05T16:04:41.183","ParentId":null,"OwnerUserId":"6589616","Title":null,"Body":"<p>Yes, that is possible. Just create the objects by yourself, e.g.<\/p>\n\n<pre><code>import torch.utils.data as data_utils\n\ntrain = data_utils.TensorDataset(features, targets)\ntrain_loader = data_utils.DataLoader(train, batch_size=50, shuffle=True)\n<\/code><\/pre>\n\n<p>where <code>features<\/code> and <code>targets<\/code> are tensors. <code>features<\/code> has to be 2-D, i.e. a matrix where each line represents one training sample, and <code>targets<\/code> may be 1-D or 2-D, depending on whether you are trying to predict a scalar or a vector.<\/p>\n\n<p>Hope that helps!<\/p>\n\n<hr>\n\n<p><strong>EDIT<\/strong>: response to @sarthak's question<\/p>\n\n<p>Basically yes. If you create an object of type <code>TensorData<\/code>, then the constructor investigates whether the first dimensions of the feature tensor (which is actually called <code>data_tensor<\/code>) and the target tensor (called <code>target_tensor<\/code>) have the same length:<\/p>\n\n<pre><code>assert data_tensor.size(0) == target_tensor.size(0)\n<\/code><\/pre>\n\n<p>However, if you want to feed these data into a neural network subsequently, then you need to be careful. While convolution layers work on data like yours, (I think) all of the other types of layers expect the data to be given in matrix form. So, if you run into an issue like this, then an easy solution would be to convert your 4D-dataset (given as some kind of tensor, e.g. <code>FloatTensor<\/code>) into a matrix by using the method <code>view<\/code>. For your 5000xnxnx3 dataset, this would look like this:<\/p>\n\n<pre><code>2d_dataset = 4d_dataset.view(5000, -1)\n<\/code><\/pre>\n\n<p>(The value <code>-1<\/code> tells PyTorch to figure out the length of the second dimension automatically.)<\/p>\n"},{"AnswerId":"42202068","CreationDate":"2017-02-13T10:41:49.633","ParentId":null,"OwnerUserId":"3693922","Title":null,"Body":"<p>You can easily do this be extending the <code>data.Dataset<\/code> class.\nAccording to the <a href=\"http:\/\/pytorch.org\/docs\/torchvision\/datasets.html\" rel=\"noreferrer\">API<\/a>, all you have to do is implement two function: <code>__getitem__<\/code> and <code>__len__<\/code>.<\/p>\n\n<p>You can then wrap the dataset with the DataLoader as shown in the API and in @pho7 's answer.<\/p>\n\n<p>I think the <code>ImageFolder<\/code> class is a reference. See code <a href=\"https:\/\/github.com\/pytorch\/vision\/blob\/master\/torchvision\/datasets\/folder.py\" rel=\"noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"58265626","CreationDate":"2019-10-07T08:06:48.810","ParentId":null,"OwnerUserId":"8044858","Title":null,"Body":"<p>In addition to <a href=\"https:\/\/stackoverflow.com\/a\/42202068\/8044858\">user3693922's answer<\/a> and the <a href=\"https:\/\/stackoverflow.com\/a\/42054194\/8044858\">accepted answer<\/a>, which respectively <a href=\"https:\/\/pytorch.org\/docs\/master\/torchvision\/datasets.html\" rel=\"nofollow noreferrer\">link<\/a> the \"quick\" PyTorch documentation example to create custom dataloaders for custom datasets, and create a custom dataloader in the \"simplest\" case, <strong>there is a much more detailed dedicated official PyTorch tutorial<\/strong> on how to create a custom dataloader with the associated preprocessing: <a href=\"https:\/\/pytorch.org\/tutorials\/beginner\/data_loading_tutorial.html\" rel=\"nofollow noreferrer\">\"writing custom datasets, dataloaders and transforms\" official PyTorch tutorial<\/a><\/p>\n"}]}
{"QuestionId":41924639,"AnswerCount":1,"Tags":"<deep-learning><caffe><object-detection>","CreationDate":"2017-01-29T18:49:21.357","AcceptedAnswerId":null,"OwnerUserId":3516826.0,"Title":"Faster RCNN make fail","Body":"<p>I have been trying for some days to make faster RCNN work in a laptop with GPU(Quadro K5100M). \nI was able to run the demo.py in a laptop with only CPU.\nI am using CUDA 8 and CuDnn 4 and caffe build works but the build process of caffe in faster rcnn does not. I have CuDnn commented out now because of some other errors.\nCan anyone give me some suggestions.<\/p>\n\n<p>I get this error during make -j8 &amp;&amp; make pycaffe. <\/p>\n\n<p>int feature_extraction_pipeline(int, char**)':\nextract_features.cpp:(.text._Z27feature_extraction_pipelineIfEiiPPc[_Z27feature_extraction_pipelineIfEiiPPc]+0x162): undefined reference to test()':\ncaffe.cpp:(.text+0x1157): undefined reference to train()':\ncaffe.cpp:(.text+0x2a3e): undefined reference to caffe::P2PSync::run(std::vector > const&amp;)'\ncaffe.cpp:(.text+0x2a73): undefined reference to caffe::P2PSync::~P2PSync()'\ncaffe.cpp:(.text+0x3f45): undefined reference to time()':\ncaffe.cpp:(.text+0x4136): undefined reference to caffe::Layer::Lock()'\ncaffe.cpp:(.text+0x45ed): undefined reference to <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44124362","CreationDate":"2017-05-23T01:24:59.000","ParentId":null,"OwnerUserId":"1819423","Title":null,"Body":"<p>I had the same error because my library path (\/usr\/lib\/) contained a previous build of libcaffe.so. Removing it fixed the problem.<\/p>\n"}]}
{"QuestionId":41924812,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><computer-vision>","CreationDate":"2017-01-29T19:05:14.897","AcceptedAnswerId":41924977.0,"OwnerUserId":2476373.0,"Title":"Tensorflow queue runner - is it possible to queue a specific subset?","Body":"<p>In tensorflow, I plan to build some model and compare it to other baseline models with respect to different subsets of the training data. I.e. I would like to train my model and the baseline models with the same subsets of training data.<\/p>\n\n<p>In the naive way queue-runner and TFreaders are implemented (e.g. <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/im2txt\" rel=\"nofollow noreferrer\">im2txt<\/a>), this requires duplicating the data per each selection of subsets, which is my case, will require to use very large amounts of disk space.<\/p>\n\n<p>It will be best, if there would be a way to tell the queue to fetch only samples from a specified subset of ids, or to ignore samples if they are not part of a given subset of ids. <\/p>\n\n<p>If I understand correctly ignoring samples is not trivial, because it will require to stitch samples from different reads to a single batch.<\/p>\n\n<p>Does anybody knows a way to do that? Or can suggest an alternative approach which does not requires pre-loading all the training data into the RAM?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41924977","CreationDate":"2017-01-29T19:22:52.603","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You could encode your condition as part of <code>keep_input<\/code> parameter of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/io_ops\/input_pipeline#maybe_batch\" rel=\"nofollow noreferrer\">tf.train.maybe_batch<\/a><\/p>\n"}]}
{"QuestionId":41925765,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2017-01-29T20:47:03.110","AcceptedAnswerId":null,"OwnerUserId":1934212.0,"Title":"Keras Cropping2D changes color channel","Body":"<p>I tried to visualize the effect of keras cropping 2D using the following code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the plotting function:<\/p>\n\n<pre><\/pre>\n\n<p>The result is<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/oL4Tp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/oL4Tp.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Obviously, the color channel has been changed. But why? Is there an error in my code or could that be a keras bug?<\/p>\n","answers":[{"AnswerId":"42023251","CreationDate":"2017-02-03T11:31:29.887","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Its not a Keras bug. Tensors are usually of <code>float32<\/code> type so when the output is evaluated they are also of <code>float32<\/code> type. You need to convert the image data to <code>uint8<\/code> type before displaying.<\/p>\n\n<pre><code>ax2.imshow(np.uint8(right_image))\n<\/code><\/pre>\n\n<p>in <code>compare_images<\/code> should display the image correctly.<\/p>\n"}]}
{"QuestionId":41926679,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-01-29T22:28:59.070","AcceptedAnswerId":41931266.0,"OwnerUserId":1914161.0,"Title":"logistic regression with tensorflow not working","Body":"<p>Learning something new here so I'm hoping the community can help. I'm mostly a ruby guy but trying to transition to python so I can tackle Machine learning with tensorflow.  I'm having the most difficult time getting this logistic regression script to work using housing data I collected. <\/p>\n\n<p>Links to data:<br>\n<a href=\"https:\/\/storage.googleapis.com\/datastorage_machinelearning\/first1500.csv\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/datastorage_machinelearning\/first1500.csv<\/a><\/p>\n\n<p>logistic regression script:<br>\n<a href=\"https:\/\/gist.github.com\/Nick-Harvey\/404b605423b3c19710eb2a1de6cb5880\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/Nick-Harvey\/404b605423b3c19710eb2a1de6cb5880<\/a><\/p>\n\n<p>Script Output:<br>\n<a href=\"https:\/\/gist.github.com\/Nick-Harvey\/3eab9262770bfb690730cad1fbadf9eb\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/Nick-Harvey\/3eab9262770bfb690730cad1fbadf9eb<\/a><\/p>\n\n<p>The error is somewhat obvious as it's saying that there are incompatible shapes. This is most likely due to the encoding I'm doing to city names adding additional columns. However, I can't seem to figure out a way to be able to fit the data so I can predict house price by sqft and be able to plot it all. Eventually, I'd like to be able to plot the data so you could compare sqft to price and sort it by the city. <\/p>\n","answers":[{"AnswerId":"41931266","CreationDate":"2017-01-30T08:06:24.950","ParentId":null,"OwnerUserId":"2251840","Title":null,"Body":"<p><strong>Note:<\/strong> I think you're not doing logistic regression. Logistic regression requires the dependent variable be binary and is estimated by maximum likelihood, such as Fischer scoring. Your error function is average least squares and your dependent variable is numerical. It looks like multiple linear regression to me.<\/p>\n\n<hr>\n\n<p>I studied your code and was able to train a model. Your code crashed later when you calculated the least square loss. But I'll leave that to you.<\/p>\n\n<p>Your problem is the dimension of your training data. Your dependent variable <code>y_train<\/code> has shape: <code>(1176)<\/code>.<\/p>\n\n<p>Try:<\/p>\n\n<blockquote>\n  <p>y_train = y_train.reshape((y_train.shape[0], 1))<\/p>\n<\/blockquote>\n\n<p>after<\/p>\n\n<blockquote>\n  <p>X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.2)<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41926806,"AnswerCount":2,"Tags":"<flask><tensorflow><neural-network><keras>","CreationDate":"2017-01-29T22:43:24.413","AcceptedAnswerId":null,"OwnerUserId":5051564.0,"Title":"ValueError: Tensor 'A' must be from the same graph as Tensor 'B'","Body":"<p>I am using keras' pre-trained model and the error came up when calling ResNet50(weights='imagenet').\nI have the following code in flask server:<\/p>\n\n<pre><\/pre>\n\n<p>when calling in in <strong>main<\/strong>, it works fine<\/p>\n\n<pre><\/pre>\n\n<p>however, the ValueError rises when I call it from the flask POST function:<\/p>\n\n<pre><\/pre>\n\n<p>The full error is as follow:<\/p>\n\n<blockquote>\n  <p>ValueError: Tensor(\"cond\/pred_id:0\", dtype=bool) must be from the same\n  graph as Tensor(\"batchnorm\/add_1:0\", shape=(?, 112, 112, 64),\n  dtype=float32)<\/p>\n<\/blockquote>\n\n<p>Any comment or suggestion is highly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"51528973","CreationDate":"2018-07-25T23:23:15.810","ParentId":null,"OwnerUserId":"9319024","Title":null,"Body":"<p>you'll need open different session and specify which graph goes with each session, else Keras will replace each graph as default.<\/p>\n\n<pre><code>from tensorflow import Graph, Session, load_model\nfrom Keras import backend as K\n<\/code><\/pre>\n\n<p>Loading the graphs:<\/p>\n\n<pre><code>graph1 = Graph()\n    with graph1.as_default():\n        session1 = Session()\n        with session1.as_default():\n            model = load_model(foo.h5)\n\ngraph2 = Graph()\n    with graph2.as_default():\n        session2 = Session()\n        with session2.as_default():\n            model2 = load_model(foo2.h5)\n<\/code><\/pre>\n\n<p>Predicting\/Using the graphs:<\/p>\n\n<pre><code>K.set_session(session1)\n    with graph1.as_default():\n        result = model.predict(data)\n<\/code><\/pre>\n"},{"AnswerId":"41927983","CreationDate":"2017-01-30T01:43:46.530","ParentId":null,"OwnerUserId":"1667840","Title":null,"Body":"<p>The problem here is your loop. You're trying to generate a new graph in each iteration.<\/p>\n\n<p>This line <\/p>\n\n<pre><code>model = ResNet50(weights='imagenet')\n<\/code><\/pre>\n\n<p>Should only be called once. So either define it as a global variable or create it before and pass it as a parameter to getResNet50Prediction()<\/p>\n"}]}
{"QuestionId":41927426,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-30T00:14:35.500","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"Why do we need to explicitly update the moving_mean and moving_variance in TensorFlow's Batch normalization in tf.contrib.layers.batch_norm?","Body":"<p><strong>To Long To Read<\/strong>: How can I use Batch Normalization with  without having to explicitly tell session to update the moving_statistics ( and ) or not?<\/p>\n\n<p>A few months ago I provided an answer to <a href=\"https:\/\/stackoverflow.com\/questions\/33949786\/how-could-i-use-batch-normalization-in-tensorflow\/38320613#38320613\">How could I use Batch Normalization in TensorFlow?<\/a> and noticed a few weird details that I wanted to address. First it seems that the implementation that I provide seems repetitive with respect to the  variable. Recall my suggested code:<\/p>\n\n<pre><\/pre>\n\n<p>in it I have a  variable that just holds a tf boolean . As you can see, it is used to decide if the batch norm layer should be in inference mode or not. However, the variable seemed a little redundant, since it seems I have two variables that specify the same thing twice. i.e. once in  and another in . Is that really necessary?<\/p>\n\n<p>I thought about it a bit and it seems I might to be able to remove the hard coded () with the (pseudo)code:<\/p>\n\n<pre><\/pre>\n\n<p>which seems to make the  variable completely redundant\/silly. This actually highlights my most important point, is the  variable and  even necessary? Which actually brings up my biggest complaint about the code (though I think this code might not even run because when defining the graph the placeholder train_phase might not even have a value but you get the idea).<\/p>\n\n<p>Honestly I find having to even explicitly define  very dangerous because it seems very unnecessary for users to have to handle the inference\/training mode of Batch Norm this explicitly. Though, \"normal\" users of Batch Norm should <strong>always<\/strong> update the , with the train data and any standard user of Batch Norm should not be updating , with test statistics at any time. Since the user is required to do:<\/p>\n\n<pre><\/pre>\n\n<p>it can bring cause really bad bugs for users that shouldn't even exist in the first place (at least in my opinion). Furthermore, it seems weird to have to explicitly say what the  is because whenever one trains, one uses an optimizer, so it should be incredibly clear when that code is called that it should be true. Maybe this is a terrible idea but it feels like the optimizer or the session should be setting that to true automatically rather than relying on the user to do it right.<\/p>\n\n<p>I understand that sometimes users are allowed more flexibility to be more creative but I can't really appreciate how this (even for a researcher) be a good feature. Maybe I am just using the library incorrectly or being paranoic, but should the user really be forced to be so explicit when using batch norm? Is there some way around this?<\/p>\n\n<p>As a side point, having the  be part of the model also makes the code be a bit more ugly and confusing than it feels necessary because it seems to me that its unavoidable to have a line of code where the session is being used to check if the batch norm flag is on or not. The code I am trying to avoid writing is the logic:<\/p>\n\n<pre><\/pre>\n\n<p>it just feels totally unnecessary. It feels the during training the model should know if it should be updating the variables or not. <\/p>\n\n<p>As quick (really ugly) solution to the last problem with the if condition in the session, one can always define  as part of the model (or at least as part of the graph) and accordingly set it equal to true and\/or false when appropriate but when one doesn't actually use the batch norm layer, one actually <strong>does not use<\/strong> the  placeholder in the model even if we set it have a value in the session.run. i.e. the sessions sets it to true or false, but when BN is not being used, it doesn't even matter what one sets it equal to since its not actually being used. Obviously, this makes the code really confusing (since one is defining some variable one doesn't even need), but I can't seem to find a way to hide the  variable. For the moment this is what I am going for because it seems really ugly to have to split (or duplicate) my code between lines that have:<\/p>\n\n<pre><\/pre>\n\n<p>and the ones that don't have it all:<\/p>\n\n<pre><\/pre>\n\n<p>Ideally I want the second solution and have batch norm work regardless if I use the silly  variable.<\/p>\n","answers":[{"AnswerId":"42006922","CreationDate":"2017-02-02T16:05:31.083","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>I don't really have a complete answer to your question, but I have a few observations:<\/p>\n\n<ul>\n<li>The standard practice seems to be to build slightly different graphs for training and for inference, each built with or without <code>is_training<\/code> enabled.<\/li>\n<li>The <code>batch_norm<\/code> layer is designed so that you can use an <code>arg_scope<\/code> to set <code>is_training=True<\/code> for all layers in your model. For example, take a look at how the Inceptionv3 model is defined here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/nets\/inception_v3.py#L571\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/nets\/inception_v3.py#L571<\/a> . This at least makes it much more convenient to set <code>is_training<\/code> once in your Python code that builds a model and to have it apply everywhere.<\/li>\n<li>Tensorflow's underlying infrastructure doesn't distinguish between training and inference time\u2014it's just running graphs of operators. <code>tf.Session<\/code> doesn't really know anything about Neural Networks, training, or inference, so it isn't the right place for this kind of logic.<\/li>\n<li>One could imagine that an <code>Optimizer<\/code> should rewrite the graph to enable <code>is_training<\/code> for those operators that support it. I don't have a strong opinion about this; you might try filing a Tensorflow Github issue making that feature request to see what others think about it. It might seem a bit too \"magical\".<\/li>\n<\/ul>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41927513,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-30T00:25:38.980","AcceptedAnswerId":null,"OwnerUserId":1277928.0,"Title":"saver.restore(sess, ckptfile) vs slim.assign_from_checkpoint_fn","Body":"<p>I have two model files model.ckpt-10       model.ckpt-10.meta \nWhat is the difference between these two options to restore the graph ? \nslim.assign_from_checkpoint_fn() and saver.restore(sess, ckptfile)<\/p>\n\n<p>Both of them lead to similar uninitialized variable  exceptions.<\/p>\n\n<p>Here is the snippet and the exception. <\/p>\n\n<pre><\/pre>\n\n<p>seems to print variables and values causes the following later. \ntensorflow.python.framework.errors.FailedPreconditionError: Attempting to use uninitialized value kaggle_test\/conv1_7_7\/biases_1<\/p>\n","answers":[]}
{"QuestionId":41927602,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><theano><keras>","CreationDate":"2017-01-30T00:38:50.983","AcceptedAnswerId":null,"OwnerUserId":7487684.0,"Title":"Keras backend json is defined to be tensorflow, but Keras still can't find tensorflow","Body":"<p>I have changed the json file for keras to be the following:<\/p>\n\n<pre><\/pre>\n\n<p>But when i run the following simple Keras tutorial for a neural network:<\/p>\n\n<pre><\/pre>\n\n<p>As taken from: <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/getting-started\/sequential-model-guide\/<\/a><\/p>\n\n<p>I still get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm out of ideas on what the issue could be, so some help would be much appreciated. <\/p>\n\n<p>Initially I thought that it may be a python versioning issue. As I'm a bit new to python coding, and Linux, I've been installing\/upgrading all my python versions without really thinking, and I am only hoping that Keras would be using the same python version as my Tensorflow backend (which should be the one used by anaconda3). In hindsight I should really be using virtualenv, but I'm not sure if this is my problem or not (just trying to give as much info as I can). <\/p>\n\n<p>I installed Keras using:<\/p>\n\n<pre><\/pre>\n\n<p>If I run it again I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>Also I am 100% sure my Tensorflow installation works, as I have run (and have coded up) some GPU-Cuda examples for it. <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"41927851","CreationDate":"2017-01-30T01:22:47.503","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>I think you forgot the most obvious thing, TensorFlow is not installed and it is not a Keras dependency. I recommend you to install TensorFlow with:<\/p>\n\n<pre><code>pip install --user tensorflow\n<\/code><\/pre>\n\n<p>This will install TensorFlow in our user folder (~\/.local) and does not require root privileges.<\/p>\n"},{"AnswerId":"45607845","CreationDate":"2017-08-10T08:04:47.653","ParentId":null,"OwnerUserId":"7712022","Title":null,"Body":"<p>You can install all the dependencies for tensorflow along with keras as follows, <\/p>\n\n<p>This setup is for Ubuntu 14.04 server<\/p>\n\n<pre><code># Pick up some TF dependencies\napt-get update &amp;&amp; apt-get install -y --no-install-recommends \\\n    build-essential \\\n    curl \\\n    git \\\n    cmake \\\n    libfreetype6-dev \\\n    libpng12-dev \\\n    libzmq3-dev \\\n    pkg-config \\\n    python \\\n    python-dev \\\n    rsync \\\n    software-properties-common \\\n    unzip \\\n    &amp;&amp; \\\napt-get clean &amp;&amp; \\\nrm -rf \/var\/lib\/apt\/lists\/*\n\ncurl -O https:\/\/bootstrap.pypa.io\/get-pip.py &amp;&amp; \\\npython get-pip.py &amp;&amp; \\\nrm get-pip.py\n\npip --no-cache-dir install --upgrade ipython &amp;&amp; \\\npip --no-cache-dir install \\\n    ipykernel \\\n    jupyter \\\n    matplotlib \\\n    numpy \\\n    scipy \\\n    sklearn \\\n    pandas \\\n    Pillow \\\n    &amp;&amp; \\\npython -m ipykernel.kernelspec\n\n# Install TensorFlow CPU version from central repo\npip --no-cache-dir install https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.1-cp27-none-linux_x86_64.whl\n\n# h5py is optional dependency for keras\napt-get update &amp;&amp; apt-get install -y libhdf5-dev libhdf5-serial-dev\npip install keras h5py\n<\/code><\/pre>\n\n<p>If you're still having environment issues, I would suggest using this <a href=\"https:\/\/github.com\/datmo\/docker-files\/blob\/master\/gpu\/keras\/Dockerfile\" rel=\"nofollow noreferrer\">Dockerfile<\/a>. This allows us to work independently of local python, which I have found very helpful to replicate the environment on any other system. You might also find Datmo conversion useful to facilitate this.<\/p>\n\n<p>DISCLAIMER: I work at this company called <a href=\"https:\/\/beta.datmo.io\/\" rel=\"nofollow noreferrer\">Datmo<\/a>, and we are building a community of developers by simplifying the machine learning workflow<\/p>\n"}]}
{"QuestionId":41927609,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-30T00:40:07.560","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"How is it that the tf.contrib.layers.convolution2d does not require to update the moving_mean and moving_variance explicitly?","Body":"<p>How is it that the  layer does not require the user to explicitly update the moving_statistics?<\/p>\n\n<p>The reason that I am asking is because I recently asked <a href=\"https:\/\/stackoverflow.com\/questions\/41927426\/why-do-we-need-to-explicitly-update-the-moving-mean-and-moving-variance-in-tenso\">Why do we need to explicitly update the moving_mean and moving_variance in TensorFlow&#39;s Batch normalization in tf.contrib.layers.batch_norm?<\/a> trying to figure out how to use  without having to explicitly update the moving_statistics. <\/p>\n\n<p>However, I noticed that there was a different layer <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.layers\/higher_level_ops_for_building_neural_network_layers_#convolution2d\" rel=\"nofollow noreferrer\">tf.contrib.layers.convolution2d<\/a> that allows you to provide a normalization layer (say ) but <strong>does not require to explicitly update the moving statistics<\/strong> (as I detailed on my other question). How is that dark magic happening? i.e. how is the convolution2d able to skip the (what seems) unavoidable requirement? <\/p>\n\n<p>Ideally, I'd like to create a similar API\/interface so that I am not required to update the moving_stats explicitly and work as cleanly as this convolution2d layer.<\/p>\n","answers":[{"AnswerId":"42007051","CreationDate":"2017-02-02T16:11:36.113","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p><code>tf.contrib.layers.convolution2d<\/code> does not know or do anything special about batch normalization or its statistics. If you look at the implementation, all it does is apply a normalizer function of your choice. From <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L914\" rel=\"nofollow noreferrer\"><code>tensorflow\/contrib\/layers\/python\/layers\/layers.py<\/code><\/a>:<\/p>\n\n<pre><code>if normalizer_fn is not None:\n  normalizer_params = normalizer_params or {}\n  outputs = normalizer_fn(outputs, **normalizer_params)\n<\/code><\/pre>\n\n<p>So the convolutional layer isn't really related to your question. It's just the implementation of batch normalization that is relevant. For the remaining part of your question, see my answer at:\n<a href=\"https:\/\/stackoverflow.com\/questions\/41927426\/why-do-we-need-to-explicitly-update-the-moving-mean-and-moving-variance-in-tenso\">Why do we need to explicitly update the moving_mean and moving_variance in TensorFlow&#39;s Batch normalization in tf.contrib.layers.batch_norm?<\/a><\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41927944,"AnswerCount":0,"Tags":"<tensorflow><normalization>","CreationDate":"2017-01-30T01:37:29.767","AcceptedAnswerId":null,"OwnerUserId":720484.0,"Title":"What is the right way to implement the graph of validation process of Batch Normalization in tensorflow?","Body":"<p>When using <strong>Batch Normalization<\/strong> we can use the <strong>batch mean<\/strong> and <strong>batch variance<\/strong> during training.\nHowever this fails to give good enough results in <strong>validation<\/strong> case. We get the best results by providing the <strong>population mean<\/strong> and <strong>population variance<\/strong>.<\/p>\n\n<p>However this means that the two graphs will be almost identical except the part where we need to subtract the mean and divide the standard deviation (we use the population ones)<\/p>\n\n<p>Our current implementation worked ok but it includes too much <strong>boilerplate code<\/strong>.<\/p>\n\n<p><em>Note that in other frameworks you can extract and manipulate layers of a model quite easily for such cases, but tensorflow has an only append policy to its graphs<\/em><\/p>\n\n<p>What is the optimal way to go in such cases where one graph is slightly different from the other?<\/p>\n","answers":[]}
{"QuestionId":41930508,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-30T07:03:44.803","AcceptedAnswerId":41940709.0,"OwnerUserId":2820379.0,"Title":"Is tf.identity() creating a full deep copy?","Body":"<p>I'm often using the following pattern for managing control flow:<\/p>\n\n<pre><\/pre>\n\n<p>However, I'm concerned that tf.identity() might copy the data passed to it which is not what I want. Can somebody confirm that it does or does not create a copy?<\/p>\n","answers":[{"AnswerId":"41940709","CreationDate":"2017-01-30T16:22:37.827","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2607cc67f0e4a6cd313399e1f68cc4b9857fe700\/tensorflow\/core\/kernels\/identity_op.h#L31\" rel=\"nofollow noreferrer\">implementation<\/a> of the <code>tf.identity()<\/code> operation will forward its input to its output without making a deep copy. However, if the <code>tf.identity()<\/code> operation is pinned to a different device from the operation that produces its input, a deep copy <em>will<\/em> occur.<\/p>\n"}]}
{"QuestionId":41930725,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-30T07:22:36.937","AcceptedAnswerId":null,"OwnerUserId":6229482.0,"Title":"tensorflow pass an integer to graph","Body":"<p>I need to loop in graph for different times. I am unable to pass the integer variable through feed_dict. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42007390","CreationDate":"2017-02-02T16:28:10.280","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>When you build a Tensorflow graph, the graph isn't actually evaluated until you call <code>tf.Session.run<\/code>. A consequence of this is that Python constructs like <code>range<\/code> cannot inspect the value of a <code>Tensor<\/code> <em>except<\/em> by calling <code>run()<\/code>. Here, <code>range(st)<\/code> requires that the value of <code>st<\/code> be an integer known to Python, but the value of <code>st<\/code> is not known until the graph actually executes.<\/p>\n\n<p>Put another way: the code that constructs the graph must not depend on the evaluation of that graph.<\/p>\n\n<p>But here, you can just use a Python integer rather than a Tensor and everything works fine:<\/p>\n\n<pre><code>rough1 = tf.Graph()\nwith rough1.as_default():\n  d = tf.Variable(0)   \nfor i in range(3):\n    d = tf.add(d,1)\n\nwith tf.Session(graph = rough1) as sess:\n  sess.run(tf.initialize_all_variables())\n  s = sess.run([d])\n  print s\n<\/code><\/pre>\n\n<p>Hope that helps!<\/p>\n"},{"AnswerId":"44797017","CreationDate":"2017-06-28T08:11:49.960","ParentId":null,"OwnerUserId":"1109028","Title":null,"Body":"<p>Try this and see if it is what you are looking for.<\/p>\n\n<pre><code>st=1\nrough1 = tf.Graph()\nwith rough1.as_default():\n  d = tf.Variable(0)   \nfor i in range(st):\n    d = tf.add(d,1)\n\nwith tf.Session(graph = rough1) as sess:\nsess.run(tf.initialize_all_variables())\n  s = sess.run([d])\n  print s\n<\/code><\/pre>\n"}]}
{"QuestionId":41930964,"AnswerCount":1,"Tags":"<matlab><deep-learning><caffe><matcaffe>","CreationDate":"2017-01-30T07:40:56.520","AcceptedAnswerId":null,"OwnerUserId":7488597.0,"Title":"Error in matcaffe installation","Body":"<p><img src=\"https:\/\/i.stack.imgur.com\/zShHP.png\" alt=\"enter image description here\"><\/p>\n\n<p>The above error showed up when I tried to install matcaffe in ubuntu. I switched to caffe root directory, and ran the following commands:<\/p>\n\n<pre><\/pre>\n\n<p>My matlab version is 2011a.<\/p>\n\n<p>After running , the error above showed up.<\/p>\n\n<p>What should I do to solve this problem?<\/p>\n","answers":[{"AnswerId":"41932680","CreationDate":"2017-01-30T09:38:35.147","ParentId":null,"OwnerUserId":"3187349","Title":null,"Body":"<p>I believe that the functionality for custom unit test caffe(and specifically the package matlab.unittest.TestCase) was introduced in version <code>R2015a<\/code> of <code>MATLAB<\/code>, and that if you're on an older version it won't be available<\/p>\n"}]}
{"QuestionId":41931969,"AnswerCount":1,"Tags":"<ubuntu><tensorflow><tensorflow-serving>","CreationDate":"2017-01-30T08:56:42.927","AcceptedAnswerId":null,"OwnerUserId":6183926.0,"Title":"Error installing tensorflow, serving in ubuntu","Body":"<p>I am installing Tensorflow serving for which I have to install tensorflow in ubuntu. I ran  command in  root directory. \nThis is the output :<\/p>\n\n<pre><\/pre>\n\n<p>There is no folder called . I do have  file in  and  in the folder . I dont know how the configuration file is generating the paths but it cant find cudnn although I have successfully installed caffe whose CMakeLists.txt could easily find the paths to cuda and cudnn installation. How do I fix this problem?<\/p>\n","answers":[{"AnswerId":"45976702","CreationDate":"2017-08-31T08:24:17.757","ParentId":null,"OwnerUserId":"4982729","Title":null,"Body":"<p>Assuming that you really have <code>cudnn<\/code> installed.<br>\nfind the installation location of your cuda using -<br>\n<code>which nvcc<\/code>  <\/p>\n\n<p>in my case it returns - \n<code>\/usr\/local\/cuda-6.5\/bin\/nvcc<\/code><\/p>\n\n<p>so <code>cudnn.h<\/code> is located at <code>\/usr\/local\/cuda-6.5\/include<\/code> (<strong>if<\/strong> cudnn is installed)<\/p>\n\n<p>while configuring tensorflow, you are asked -<br>\n<code>Please specify the location where cuDNN  library is installed. Refer to README.md for more details. [Default is \/usr\/local\/cuda]:<\/code><\/p>\n\n<p>here you have to explicitly specify the location of <code>cudnn<\/code>.<br>\nIn my case it is <code>\/usr\/local\/cuda-6.5\/include\/<\/code>.<\/p>\n"}]}
{"QuestionId":41933830,"AnswerCount":1,"Tags":"<merge><theano><layer><keras><embedding>","CreationDate":"2017-01-30T10:39:22.513","AcceptedAnswerId":41952639.0,"OwnerUserId":6407883.0,"Title":"Keras : How to merge a dense layer and an embedding layer","Body":"<p>I use Keras and I try to concatenate two different layers into a vector (first values of the vector would be values of the first layer, and the other part would be the values of the second layer).\nOne of these layers is a Dense layer and the other layer is a Embedding layer.<\/p>\n\n<p>I know how to merge two embedding layers or two dense layers but I don't know how to merge a embedding layer and a dense layer (dimensional problem).<\/p>\n\n<p>A simple example would be like this:<\/p>\n\n<pre><\/pre>\n\n<p>But this will not work because you can't merge layers with different dimensionalities.<\/p>\n\n<p>PS : Sorry, english is not my native languague and I hope you will understand my problem.<\/p>\n\n<p>Best regards.<\/p>\n","answers":[{"AnswerId":"41952639","CreationDate":"2017-01-31T08:13:38.847","ParentId":null,"OwnerUserId":"5638904","Title":null,"Body":"<p>Use <a href=\"https:\/\/keras.io\/layers\/core\/#flatten\" rel=\"nofollow noreferrer\">Flatten<\/a> layer.<\/p>\n\n<pre><code>L_branch = Sequential()\nL_branch.add(Dense(10, input_shape =  (4,) , activation = 'relu'))\nL_branch.add(BatchNormalization())\n\nR_branch = Sequential()\nR_branch.add(Embedding(1000, 64, input_length=5))\nR_branch.add(Flatten()) # &lt;--\n\nfinal_branch = Sequential() # &lt;--\nfinal_branch.add(Merge([L_branch, R_branch], mode = 'concat'))\n<\/code><\/pre>\n"}]}
{"QuestionId":41933958,"AnswerCount":1,"Tags":"<python><neural-network><keras><recurrent-neural-network><sequence-to-sequence>","CreationDate":"2017-01-30T10:47:00.623","AcceptedAnswerId":41946270.0,"OwnerUserId":996366.0,"Title":"How to code a sequence to sequence RNN in keras?","Body":"<p>I am trying to write a sequence to sequence RNN in keras. I coded this program using what I understood from the web. I first tokenized the text then converted the text into sequence and padded to form feature variable <strong>X<\/strong>. The target variable <strong>Y<\/strong> was obtained first shifting <strong>x<\/strong> to left and then padding it. Lastly I fed my feature and target variable to my LSTM model.<\/p>\n\n<p>This is  my code I written in keras for that purpose.<\/p>\n\n<pre><\/pre>\n\n<p>The problem is its showing an error<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41946270","CreationDate":"2017-01-30T22:08:31.317","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The problem lies in: <\/p>\n\n<pre><code>model.add(Embedding(max_features, 128, input_length=max_len, dropout=0.2))\n<\/code><\/pre>\n\n<p>In the Embedding <a href=\"https:\/\/keras.io\/layers\/embeddings\/\" rel=\"noreferrer\">documentation<\/a> you may see that the first argument provided to it should be set to size of vocabulary + 1. It's because there should be always a place for a <code>null<\/code> word which index is <code>0<\/code>. Because of that you need to change this line to:<\/p>\n\n<pre><code>model.add(Embedding(max_features + 1, 128, input_length=max_len, dropout=0.2))\n<\/code><\/pre>\n"}]}
{"QuestionId":41936988,"AnswerCount":0,"Tags":"<tensorflow><julia>","CreationDate":"2017-01-30T13:22:01.513","AcceptedAnswerId":null,"OwnerUserId":179081.0,"Title":"How to use `sparse_softmax_cross_entropy_with_logits`: without getting Incompatible Shapes Error","Body":"<p>I would like to use the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard4\/tf.nn.sparse_softmax_cross_entropy_with_logits.md\" rel=\"nofollow noreferrer\">sparse_softmax_cross_entropy_with_logits<\/a>\nwith the julia TensorFlow wrapper.<\/p>\n\n<p>The operations is defined in the <a href=\"https:\/\/github.com\/malmaud\/TensorFlow.jl\/blob\/1127fb5864b6513924f8c44cacc46102189fabc5\/src\/ops\/nn.jl#L201\" rel=\"nofollow noreferrer\">code here<\/a>.\nBasically, as I understand it the first argument should be logits, that would normally be fed to softmax to get them to be category probabilities (~1hot output).\nAnd the second should be the correct labels as label ids.<\/p>\n\n<p>I have adjusted the example code from the <a href=\"https:\/\/github.com\/malmaud\/TensorFlow.jl\/tree\/94303d4541785da283cff7fd0039a98a5ad6ff94\" rel=\"nofollow noreferrer\">TensorFlow.jl readme<\/a>\nSee below:<\/p>\n\n<pre><\/pre>\n\n<p>When I run it however, I get an error:<\/p>\n\n<pre><\/pre>\n\n<p>This only happens when I try to train it.\nIf I don't include an optimise function\/output then it works fine.\nSo I am doing something that screws up the gradient math.<\/p>\n","answers":[]}
{"QuestionId":41937617,"AnswerCount":1,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2017-01-30T13:55:06.870","AcceptedAnswerId":null,"OwnerUserId":2324298.0,"Title":"Difficulty in interpreting the output of tensorflow's dynamic_rnn","Body":"<p>I am trying to implement a LSTM cell based RNN in tensorflow. I have a sequence of words in each row which forms the features and a column named labels (which is either 1 or 0) that serves as the target. I am facing difficulties in interpreting the outputs of . The  returns the outputs and the states. \nI have split my data in to batches of 256. I have padded\/sliced off \neach row to have a uniform length of 160. I then embedded them(using glove) so that each word becomes a hundred dimensional vector. After this my each example turns into a 160 by 100 matrix, and each batch to the RNN function becomes a tensor of shape - (256, 160, 100). Now given that the hidden_dimension of LSTM cell is set to 300, the expected dimension of the hidden state vector for each example(of size 160 x 100) is 1 x 300. However, when I check the shape of , I get the following - <\/p>\n\n<pre><\/pre>\n\n<p>I was expecting the dimensions to be 256 x 300. I am not sure how to proceed further. <\/p>\n\n<p>I am adding some code below that I used and their shapes. I have two tensors -   and . I pick 128 samples of each, embed and concatenate them later to form an input batch of 256. Can you tell me where I am going wrong? <\/p>\n\n<pre><\/pre>\n\n<p>I am not able to figure why the dimension is None x 300. I was expecting it to be 256 x 300. Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42010382","CreationDate":"2017-02-02T19:07:14.380","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p><code>get_shape<\/code> is best effort. In general, Tensorflow does not always know the sizes of Tensors at graph construction time. <code>None<\/code> simply means the dimension is unknown at construction time.<\/p>\n\n<p>The only guaranteed way to find the shape of a Tensor from Python is to run the graph (e.g., via <code>session.run()<\/code> or <code>.eval()<\/code>) to evaluate a Tensor to a specific value, whose shape you can inspect.<\/p>\n\n<p>(If you want to consume the shape within the graph, there is also <code>tf.shape<\/code>.)<\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41937719,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2017-01-30T13:59:44.093","AcceptedAnswerId":null,"OwnerUserId":5687439.0,"Title":"Checkpoint deep learning models in Keras","Body":"<p>I need help in implementing the checkpoint function in Keras. I am going to train a large dataset so in order to do that, first I trained a model using the iris flower dataset : <a href=\"http:\/\/machinelearningmastery.com\/multi-class-classification-tutorial-keras-deep-learning-library\/\" rel=\"nofollow noreferrer\">http:\/\/machinelearningmastery.com\/multi-class-classification-tutorial-keras-deep-learning-library\/<\/a><\/p>\n\n<p>since my own dataset is a lot similar to it only the difference is that my dataset is more bigger.<\/p>\n\n<p>For the checkpoint function : <a href=\"http:\/\/machinelearningmastery.com\/check-point-deep-learning-models-keras\/\" rel=\"nofollow noreferrer\">http:\/\/machinelearningmastery.com\/check-point-deep-learning-models-keras\/<\/a><\/p>\n\n<p>I understand the example using pima-indians dataset. \nNow I am trying to implement the same checkpoint function in the iris-flower script. Here is what I tried so far.<\/p>\n\n<pre><\/pre>\n\n<p>This script produced the following error. I don't know how to troubleshoot it or maybe my arrangement in the script is wrong.<\/p>\n\n<pre><\/pre>\n\n<p>I hope somebody can help me with this. Thank you.<\/p>\n","answers":[{"AnswerId":"47670982","CreationDate":"2017-12-06T09:35:46.200","ParentId":null,"OwnerUserId":"4134424","Title":null,"Body":"<p>Instead of KerasClassifier, use the model itself.<\/p>\n\n<pre><code>model = baseline_model()\n#declare your callback methods here\nmodel.fit(x,y, batch_size=32, verbose=0, epochs=10, shuffle=True, validation_split = 0.1, callbacks = &lt;your list of callbacks&gt;)\n<\/code><\/pre>\n"},{"AnswerId":"41937818","CreationDate":"2017-01-30T14:04:15.597","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>I think that the problem is that your <code>baseline_model()<\/code> function is not returning the model it is creating; it should be something like:<\/p>\n\n<pre><code>def baseline_model():\n    model = Sequential()\n    model.add(Dense(4, 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<\/code><\/pre>\n"}]}
{"QuestionId":41937874,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-30T14:06:41.967","AcceptedAnswerId":42010635.0,"OwnerUserId":5621202.0,"Title":"how to make tensorflow placeholder to work properly","Body":"<p>I wrote simple code using tensorflow. I thought the code below should work. The goal of this code is simple. Just calculating simple addition. <\/p>\n\n<pre><\/pre>\n\n<p>But, the interpreter says something like this; <\/p>\n\n<pre><\/pre>\n\n<p>I just copied what is essential to the error. I am spending lots of hours for this simple code, but I don't see any problem. Any tip would be greatly appreciated. <\/p>\n","answers":[{"AnswerId":"42010635","CreationDate":"2017-02-02T19:22:36.863","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>You use <code>x<\/code> as a variable initializer, so you need to feed <code>x<\/code> when executing <code>tf.initialize_all_variables()<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndata = np.array([[35.0, 40.0, 45.0]]).astype(np.float32)\nprint np.shape(data)\nx = tf.placeholder(tf.float32, [1, 3], name='x')\ny = tf.Variable(x, name='y')\n\nmodel = tf.initialize_all_variables()\n\nwith tf.Session() as sess:\n    sess.run(model, feed_dict={x: data})\n    result = sess.run(y)\n    print result\n<\/code><\/pre>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41937915,"AnswerCount":5,"Tags":"<python><macos><tensorflow><pip>","CreationDate":"2017-01-30T14:08:10.807","AcceptedAnswerId":41942396.0,"OwnerUserId":350080.0,"Title":"How to pip install old version of library(tensorflow)?","Body":"<p>I'm trying to install tensorflow r0.11. I tried<\/p>\n\n<pre><\/pre>\n\n<p>But I get this error<\/p>\n\n<pre><\/pre>\n\n<p>I assume pip is no longer support depricated versions, how can I get it?<\/p>\n\n<p>I also tried<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47402682","CreationDate":"2017-11-20T23:13:27.700","ParentId":null,"OwnerUserId":"6334174","Title":null,"Body":"<p>This works for me on Mac OS 10.13.1.<\/p>\n\n<pre><code>pip install --user install tensorflow==1.3.0. \n<\/code><\/pre>\n"},{"AnswerId":"50694243","CreationDate":"2018-06-05T07:16:31.750","ParentId":null,"OwnerUserId":"8900683","Title":null,"Body":"<p>If you want to install tensorflow <strong>r0.11<\/strong>, it means that you want to install the sources from a certain <strong>branch<\/strong> (<strong>r0.11<\/strong>). Therefore, you should follow the official installation instructions from tensorflow: <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources<\/a>:<\/p>\n\n<p>In case you want to install a \"normal\" release. then you can easily use the following command:<\/p>\n\n<pre><code>pip install --upgrade tensorflow==1.3.0.\n<\/code><\/pre>\n"},{"AnswerId":"50505445","CreationDate":"2018-05-24T09:15:16.217","ParentId":null,"OwnerUserId":"8748639","Title":null,"Body":"<p>If you have your own library\/packet on github\/gitlab etc, you have to add tag to commit with concrete version of library e.g. v2.0 then you can install your packet <\/p>\n\n<pre><code>pip install git+ssh:\/\/link\/name\/repo.git@v2.0\n<\/code><\/pre>\n"},{"AnswerId":"41942396","CreationDate":"2017-01-30T17:56:14.707","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>You can install the pip wheel from a URL directly, for example:<\/p>\n\n<pre><code># Ubuntu\/Linux 64-bit, CPU only, Python 2.7\nexport TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.11.0-cp27-none-linux_x86_64.whl\npip install --ignore-installed --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n\n<p>In general, installation instructions for older versions of TensorFlow can be found at <code>https:\/\/www.tensorflow.org\/versions\/r&lt;version&gt;\/get_started\/os_setup<\/code>. For example, the installation instructions for TensorFlow 0.12 can be found at:\n<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup#using_pip\" rel=\"noreferrer\"><code>https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup#using_pip<\/code><\/a>.<\/p>\n"},{"AnswerId":"53623907","CreationDate":"2018-12-05T01:36:13.157","ParentId":null,"OwnerUserId":"7154924","Title":null,"Body":"<p>This worked for me:<\/p>\n\n<p><code>pip3 install 'tensorflow==1.2.0' --force-reinstall<\/code><\/p>\n"}]}
{"QuestionId":41938111,"AnswerCount":1,"Tags":"<docker><tensorflow>","CreationDate":"2017-01-30T14:18:09.023","AcceptedAnswerId":null,"OwnerUserId":6730832.0,"Title":"Restore Tensorflow model on docker","Body":"<p>I am using a face recognition model based on tensorflow. in my local machine - ubuntu 14.04 - everything works.\nwhen I deploy it using docker, I am getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am using python implementation for tensorflow.\nThe model is in the old 11.* format (model.meta &amp; model.ckpt-80000) while the tensorflow python version is 12.* . It shouldn't be a problem, as that's the configuration in my local machine, as well as in the place where I took the model from.\nThe versions of tensorflow, numpy and protobuf are identical in my machine and in the docker machine.<\/p>\n\n<p>Any advice?<\/p>\n\n<hr>\n\n<h2>UPDATE<\/h2>\n\n<p>I created a small script that runs perfectly on my machine. Then, I run the same script on the deployed on virtual machine (AWS instance) BUT NOT on docker. It also failed, with the same error.\nThe deployed machine is ubuntu 16.04.<\/p>\n","answers":[{"AnswerId":"41959507","CreationDate":"2017-01-31T14:02:52.700","ParentId":null,"OwnerUserId":"6730832","Title":null,"Body":"<p>Seems like i was dealing with a corrupted file<\/p>\n"}]}
{"QuestionId":41938327,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><tensorboard>","CreationDate":"2017-01-30T14:28:30.763","AcceptedAnswerId":45697593.0,"OwnerUserId":3747801.0,"Title":"How to use multiple summary writers with `tf.train.Supervisor` for TensorBoard","Body":"<p>I would like to do something similar to the  and  from the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/summaries_and_tensorboard\/\" rel=\"nofollow noreferrer\">TensorBoard tutorial<\/a>. But using . I am however not sure how best to go about this.<\/p>\n\n<p>Pseudo code:<\/p>\n\n<pre><\/pre>\n\n<p>So my question is: <strong>How do I save the  and  do different directories?<\/strong> E.g.  and .<\/p>\n","answers":[{"AnswerId":"45697593","CreationDate":"2017-08-15T16:50:13.960","ParentId":null,"OwnerUserId":"3747801","Title":null,"Body":"<p>You are looking for <code>summary_computed<\/code>. The docstring shows how to create custom summary writers. You cannot get the Supervisor to manage it automatically, but it is quite simple. From <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/supervisor.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/supervisor.py<\/a><\/p>\n\n<pre><code># Create a Supervisor with no automatic summaries.\nsv = Supervisor(logdir='\/tmp\/mydir', is_chief=is_chief, summary_op=None)\n# As summary_op was None, managed_session() does not start the\n# summary thread.\nwith sv.managed_session(FLAGS.master) as sess:\n  for step in xrange(1000000):\n    if sv.should_stop():\n      break\n    if is_chief and step % 100 == 0:\n      # Create the summary every 100 chief steps.\n      sv.summary_computed(sess, sess.run(my_summary_op))\n    else:\n      # Train normally\n      sess.run(my_train_op)\n<\/code><\/pre>\n"}]}
{"QuestionId":41938508,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-01-30T14:37:26.420","AcceptedAnswerId":41940669.0,"OwnerUserId":4479428.0,"Title":"Tensorflow Transfer Learning with Input Pipeline","Body":"<p>I want to use transfer learning with Google's Inception network for an image recognition problem. I am using <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\"><\/a> from the TensorFlow example source for inspiration.<\/p>\n\n<p>In , the Inception graph is loaded and a feed dict is used to feed the new images into the model's input layer. However, I have my data serialized in TFRecord files and have been using an input pipeline to feed in my inputs, as demonstrated <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/#reading_from_files\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>So I have a tensor  which returns my input data in batches when run. But how can I feed these images into Inception? I can't use a feed dict since my inputs are tensors, not NumPy arrays. My two ideas are<\/p>\n\n<p>1) simply call  on each batch to convert it to a NumPy array, and then use a feed dict to pass it to Inception.<\/p>\n\n<p>2) replace the input node in the Inception graph with my own batch input tensor<\/p>\n\n<p>I think (1) would work, but it seems a little inelegant. (2) seems more natural to me, but I can't do exactly that because TensorFlow graphs can only be appended to and not otherwise modified.<\/p>\n\n<p>Is there a better approach?<\/p>\n","answers":[{"AnswerId":"41940669","CreationDate":"2017-01-30T16:20:24.420","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can implement option (2), replacing the input node, but you will need to modify <code>retrain.py<\/code> to do so. The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/framework\/utility_functions#import_graph_def\" rel=\"nofollow noreferrer\"><code>tf.import_graph_def()<\/code><\/a> function supports a limited form of modification to the imported graph, by <em>remapping<\/em> tensors in the imported graph to existing tensors in the target graph.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2607cc67f0e4a6cd313399e1f68cc4b9857fe700\/tensorflow\/examples\/image_retraining\/retrain.py#L254\" rel=\"nofollow noreferrer\">This line in <code>retrain.py<\/code><\/a> calls <code>tf.import_graph_def()<\/code> to import the Inception model, where <code>jpeg_data_tensor<\/code> becomes the tensor that you feed with input data:<\/p>\n\n<pre><code>bottleneck_tensor, jpeg_data_tensor, resized_input_tensor = (\n    tf.import_graph_def(graph_def, name='', return_elements=[\n        BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME,\n        RESIZED_INPUT_TENSOR_NAME]))\n<\/code><\/pre>\n\n<p>Instead of retrieving <code>jpeg_data_tensor<\/code> from the imported graph, you can remap it to an input pipeline that you construct yourself:<\/p>\n\n<pre><code># Output of a training pipeline, returning a `tf.string` tensor containing\n# a JPEG-encoded image.\njpeg_data_tensor = ...\n\nbottleneck_tensor, resized_input_tensor = (\n    tf.import_graph_def(\n        graph_def,\n        input_map={JPEG_DATA_TENSOR_NAME: jpeg_data_tensor},\n        return_elements=[BOTTLENECK_TENSOR_NAME, RESIZED_INPUT_TENSOR_NAME]))\n<\/code><\/pre>\n\n<p>Wherever you previously fed <code>jpeg_data_tensor<\/code>, you no longer need to need it, because the inputs will be read from the input pipeline  you constructed. (Note that you might need to handle <code>resized_input_tensor<\/code> as well... I'm not intimately familiar with <code>retrain.py<\/code>, so some restructuring might be necessary.)<\/p>\n"}]}
{"QuestionId":41939377,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-01-30T15:19:09.030","AcceptedAnswerId":41939583.0,"OwnerUserId":7387749.0,"Title":"How does Keras evaluate loss on test set?","Body":"<p>I am implementing a neural network classifier, to print loss and accuracy  of this NN I'm using: <\/p>\n\n<pre><\/pre>\n\n<p>I would to know how Keras calculate the loss of a model. Specially whether it is evaluated on the first (and only) step of the test set. \nI have search on <a href=\"https:\/\/keras.io\/models\/\" rel=\"noreferrer\">keras.io<\/a>, but I don't have find anything about it.<\/p>\n","answers":[{"AnswerId":"41939583","CreationDate":"2017-01-30T15:28:24.613","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>From the <a href=\"https:\/\/keras.io\/models\/sequential\/#evaluate\" rel=\"noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>evaluate<\/strong><\/p>\n  \n  <p>Computes the loss on some input data, batch by batch.<\/p>\n  \n  <p><strong>Returns<\/strong><\/p>\n  \n  <p>Scalar test loss (if the model has no metrics) or list of scalars (if the model computes other metrics). The attribute <code>model.metrics_names<\/code> will give you the display labels for the scalar outputs.<\/p>\n<\/blockquote>\n\n<p>So it returns you either a single value that represents a loss, or a list of values that correspond to different metrics that were added to your model. These values are calculated based on the whole test set, i. e. all values in <code>x_test<\/code> and <code>y_test<\/code>. <\/p>\n"}]}
{"QuestionId":41939671,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><keras><keras-layer>","CreationDate":"2017-01-30T15:32:20.950","AcceptedAnswerId":41939911.0,"OwnerUserId":1934212.0,"Title":"Is it plausible to get a lower mse with more SpatialDropouts?","Body":"<p>I have made the observation that with a<\/p>\n\n<pre><\/pre>\n\n<p>layer after each of 5 Convolutional2D layers, the training and validation error is much lower during the first few epochs than with the same network without these Dropout layers (all else equal). This seems counter-intuitive, since I would expect the optimization routine to have more trouble finding a minimum if intermediate results are dropped out randomly. <\/p>\n\n<p>So is my observation plausible? And if so, why? <\/p>\n","answers":[{"AnswerId":"41939911","CreationDate":"2017-01-30T15:43:48.260","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>Generally speaking, dropout is one of the techniques that is used to combat overfitting. It is expected to reduce the test error but not the training one. On the contrary, training error might go up when the model stops being overfitted. <\/p>\n\n<p>I recommend to read more about dropout in the <a href=\"http:\/\/www.deeplearningbook.org\/\" rel=\"nofollow noreferrer\">Deep Learning textbook<\/a>, section 7.12.<\/p>\n"}]}
{"QuestionId":41940163,"AnswerCount":1,"Tags":"<numpy><tensorflow><mat-file>","CreationDate":"2017-01-30T15:56:01.437","AcceptedAnswerId":41940507.0,"OwnerUserId":3302759.0,"Title":"How to make tensorflow cifar10 tutorial read from numpy array?","Body":"<p>I am trying to use the CIFAR10 tutorial to create my own training script. My dataset is stored in a MAT file that I convert to a Numpy array using . In the tutorial, they read data using:<\/p>\n\n<pre><\/pre>\n\n<p>whereas, in my case, I use:<\/p>\n\n<pre><\/pre>\n\n<p>The problem is when I try to run the training using a  as they use in CIFAR10 example:<\/p>\n\n<pre><\/pre>\n\n<p>where  is based in MNIST example. Some auxiliary functions:<\/p>\n\n<pre><\/pre>\n\n<p>When I try to run, it obviously return the error  because I didn't create a feed. The point is that I do have the function that creates the feed, but I don't know where should I pass it.<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone help?<\/p>\n\n<p>Thanks,\nDaniel<\/p>\n","answers":[{"AnswerId":"41940507","CreationDate":"2017-01-30T16:12:30.107","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You just need to pass the <code>dict<\/code> created by <code>fill_feed_dict<\/code> every time you call the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session.run\" rel=\"nofollow noreferrer\"><code>run<\/code><\/a> method:<\/p>\n\n<pre><code>mon_sess.run(train_op, feed_dict=fill_feed_dict(data_set, images, labels))\n<\/code><\/pre>\n"}]}
{"QuestionId":41940189,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-30T15:57:34.197","AcceptedAnswerId":41941097.0,"OwnerUserId":1317953.0,"Title":"What is the difference between tf.matmul and tf.batch_matmul in Tensorflow?","Body":"<p>For example if I have the following data:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any difference between the two operations?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41941097","CreationDate":"2017-01-30T16:44:04.883","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p><code>tf.batch_matmul<\/code> is deprecated in favor of <code>tf.matmul<\/code> in version 0.12 and later so no difference in later versions. Earlier versions required rank-2 inputs for <code>matmul<\/code> but allowed larger ranks for <code>batch_matmul<\/code><\/p>\n"}]}
{"QuestionId":41940299,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><tensorboard>","CreationDate":"2017-01-30T16:02:07.147","AcceptedAnswerId":41945173.0,"OwnerUserId":3747801.0,"Title":"TensorFlow: How to use tf.train.SummaryWriter inside supervisor.loop (0.12.0rc1)","Body":"<p>I am using  to manage my session. I am already using the  in the supervisor to write some summaries. I would however, at other intervals, like to write another set of summaries. As fare as I can see the easiest way is to use a . What I was is basically:<\/p>\n\n<p>Pseudo code:<\/p>\n\n<pre><\/pre>\n\n<p>How should I go about this?<\/p>\n","answers":[{"AnswerId":"41945173","CreationDate":"2017-01-30T20:49:19.477","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>You can also provide your own summaries to Supervisor using<\/p>\n\n<pre><code>sv.summary_computed(sess, summary, global_step)\n<\/code><\/pre>\n\n<p>manually. One interesting thing that doesn't seem to be advertised too much is that you can group summaries into collections like so:<\/p>\n\n<pre><code>tf.summary.scalar('learning_rate', p_lr, collections=['train'])\ntf.summary.scalar('loss', t_loss, collections=['train', 'test'])\ns_training = tf.summary.merge_all('train')\n<\/code><\/pre>\n\n<p>and then only write the <code>train<\/code> variables by fetching <code>s_training<\/code> and giving it to the the above function.<\/p>\n"},{"AnswerId":"42000987","CreationDate":"2017-02-02T11:26:18.187","ParentId":null,"OwnerUserId":"3747801","Title":null,"Body":"<p>TensorFlow documents this <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/distributed_execution#Supervisor.__init__\" rel=\"nofollow noreferrer\">here<\/a>, under <strong>Launching additional services<\/strong>:<\/p>\n\n<blockquote>\n  <p>Example: Start a thread to print losses. We want this thread to run every 60 seconds, so we launch it with sv.loop().<\/p>\n\n<pre><code>sv = Supervisor(logdir='\/tmp\/mydir') with sv.managed_session(FLAGS.master) as sess:\n    sv.loop(60, print_loss, (sess)) while not sv.should_stop(): \n        sess.run(my_train_op)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>See answer by @sunside for good tips on how to do this in a smart way.<\/p>\n"}]}
{"QuestionId":41940858,"AnswerCount":0,"Tags":"<parallel-processing><torch>","CreationDate":"2017-01-30T16:30:31.213","AcceptedAnswerId":null,"OwnerUserId":1237313.0,"Title":"Torch7 DataParallelTable is not utilizing any GPU in evaluation mode","Body":"<p>I am using Torch7 DataParallelTable to parallelize my Convolutional Neural Network. I based my implementation on this <a href=\"https:\/\/github.com\/soumith\/imagenet-multiGPU.torch\" rel=\"nofollow noreferrer\">example<\/a><\/p>\n\n<p>The DataParallelTable works great and my training time has reduced significantly. But during the validation phase (after invoking <strong>network:evaluate()<\/strong>), the network ignores GPUs and runs only on GPU (nvidia-smi shows 0% load and htop shows CPU cores burning). I am not sure what is causing this behavior. Because of this, the validation is taking a lot of time and in fact, more time than almost 10 training epochs though the size of validation set is just a quarter of the training set.\nCould someone please help me in understanding this strange behaviour of DataParallelTable and fixing this issue.<\/p>\n\n<p>I am attaching snippets from my code that are relevant to DPT and most of them are just a simple modifications from the example in the link.<\/p>\n\n<p><strong>Creating DPT:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Training part: (This works fine)<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Validation part: (GPU is completely ignored here)<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41941471,"AnswerCount":1,"Tags":"<python><tensorflow><tf-slim>","CreationDate":"2017-01-30T17:05:00.357","AcceptedAnswerId":null,"OwnerUserId":7490963.0,"Title":"Tensorflow slim pre-trained alexnet","Body":"<p>The Tensorflow slim library provides the graph structure for alexnet, however it doesn't seem to provide a pre-trained (on Imagenet) alexnet checkpoint (<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/research\/slim<\/a>). Is there anywhere a pre-trained alexnet checkpoint for TF-slim is provided?<\/p>\n","answers":[{"AnswerId":"47923499","CreationDate":"2017-12-21T11:00:26.787","ParentId":null,"OwnerUserId":"4397349","Title":null,"Body":"<p>Unlike VGG or Inception, TensorFlow doesn\u2019t ship with a pretrained AlexNet.\nCaffe does but it's not a trivial task to convert to tensorflow.<\/p>\n\n<p>Thanks to <a href=\"https:\/\/kratzert.github.io\/aboutme\" rel=\"nofollow noreferrer\">Frederik Kratzert<\/a>, he did that job and share <a href=\"https:\/\/kratzert.github.io\/2017\/02\/24\/finetuning-alexnet-with-tensorflow.html\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>In summary, you need to:<\/p>\n\n<ol>\n<li><p>Download <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_alexnet\" rel=\"nofollow noreferrer\">Caffe model and weights<\/a><\/p><\/li>\n<li><p>Use the <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\" rel=\"nofollow noreferrer\">caffe-to-tensorflow<\/a> to convert the Caffe model to Tensorflow model<\/p><\/li>\n<li><p>Refer the diferences if any between the converted model and the slim model and make the adaptation<\/p><\/li>\n<\/ol>\n\n<p>For more details, please look <a href=\"http:\/\/www.cs.toronto.edu\/~guerzhoy\/tf_alexnet\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
{"QuestionId":41941940,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2017-01-30T17:32:28.530","AcceptedAnswerId":41945333.0,"OwnerUserId":3747801.0,"Title":"TensorFlow: Understanding the `collections` argument in tf.summary.scalar","Body":"<p>I am working with TensorBoard, specifically . In the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/summary.md#tfsummaryscalarname-tensor-collectionsnone-scalar\" rel=\"nofollow noreferrer\">documentation<\/a> it has an arugment , which is described as:<\/p>\n\n<blockquote>\n  <p>: Optional list of graph collections keys. The new summary op is added to these collections. Defaults to .<\/p>\n<\/blockquote>\n\n<p>I don't understand this description, and what  is used for. Can someone please explain this to me, and perhaps point me towards a good example use-case?<\/p>\n","answers":[{"AnswerId":"41945333","CreationDate":"2017-01-30T21:00:47.803","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>It's a hidden gem! You can provide it a list of strings of your choice that label the summary node, e.g.<\/p>\n\n<pre><code>tf.summary.scalar('learning_rate', p_lr, collections=['train'])\ntf.summary.scalar('loss', t_loss, collections=['train', 'test'])\n<\/code><\/pre>\n\n<p>and then fetch the summaries by their label, e.g. like so:<\/p>\n\n<pre><code>s_training = tf.summary.merge_all('train')\ns_test = tf.summary.merge_all('test')\n<\/code><\/pre>\n\n<p>I'm doing it like that because I often want to log extra information during the validation phase; in the above example, I don't have to provide a value for the learning rate placeholder <code>p_lr<\/code> when evaluating (and writing) the accuracy, for example - or anything really that the inference part of the graph relies on.<\/p>\n\n<p>Providing (only) custom categories also has the nice side effect of hiding the node from <code>Supervisor<\/code>, for example. If you really want to have control over when exactly you write a summary (e.g. using <code>sv.summary_computed()<\/code> in case of <code>Supervisor<\/code>), that's an easy way to go.<\/p>\n"}]}
{"QuestionId":41942538,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><deep-learning><except>","CreationDate":"2017-01-30T18:04:13.047","AcceptedAnswerId":57183733.0,"OwnerUserId":7415539.0,"Title":"Tensorflow GPU memory error try-except not catching the error","Body":"<p>I am trying to run a hyperparameter optimization (using spearmint) on a big network with lots of trainable variables. I am worried that when I try a network with the number of hidden units too large, the Tensorflow will throw a GPU memory error. <\/p>\n\n<p>I was wondering if there is a way of catching the GPU memory error thrown by Tensorflow and skip the batch of hyperparameters that causes the memory error.<\/p>\n\n<p>For example, I would like something like<\/p>\n\n<pre><\/pre>\n\n<p>When I try above to test the memory error exception, the code breaks and just prints the GPU memory error and does not progress to the except block.<\/p>\n","answers":[{"AnswerId":"57183733","CreationDate":"2019-07-24T13:08:26.790","ParentId":null,"OwnerUserId":"3816524","Title":null,"Body":"<p>Try this : <\/p>\n\n<pre><code>import tensorflow as tf\n\ntry:\n    with tf.device(\"gpu:0\"):\n        a = tf.Variable(tf.ones((10000, 10000)))\n        sess = tf.Session()\n        sess.run(tf.initialize_all_variables())\nexcept:\n    print(\"Caught error\")\n    import pdb; pdb.set_trace()\n<\/code><\/pre>\n\n<p>source : <a href=\"https:\/\/github.com\/Hak333m\/stuff\/blob\/master\/gpu_oom.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Hak333m\/stuff\/blob\/master\/gpu_oom.py<\/a><\/p>\n"}]}
{"QuestionId":41942885,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><tensorboard>","CreationDate":"2017-01-30T18:24:40.317","AcceptedAnswerId":41943422.0,"OwnerUserId":6182430.0,"Title":"initialize variables in TensorFlow graph","Body":"<p>I'm starting an interactive session in TensorFlow, and after defining all the variables I'm starting to train end evaluate the network.<\/p>\n\n<p>What is the difference between those two commands:<\/p>\n\n<ol>\n<li><\/li>\n<li><\/li>\n<\/ol>\n\n<p>till today I used the second command, but recently I noticed the first command.<\/p>\n\n<p>Thanks :)<\/p>\n","answers":[{"AnswerId":"41943422","CreationDate":"2017-01-30T18:56:19.183","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The two statements are equivalent: both <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/variable_helper_functions#global_variables_initializer\" rel=\"noreferrer\"><code>tf.global_variables_initializer()<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/exporting_and_importing_meta_graphs#initialize_all_variables\" rel=\"noreferrer\"><code>tf.initialize_all_variables()<\/code><\/a> return a <code>tf.Operation<\/code> that, when run, will initialize the global variables in a model. Passing an operation to <code>sess.run()<\/code> or calling <code>operation.run()<\/code> are equivalent when you have created a <code>tf.InteractiveSession<\/code>, or are in a <code>with tf.Session():<\/code> block.<\/p>\n\n<p>The <code>tf.initialize_all_variables()<\/code> function has been deprecated (and will be removed from TensorFlow 1.0) because its name is confusing: it does not initialize <em>all<\/em> variables (i.e. local variables must be initialized separately, using <code>tf.local_variables_initializer()<\/code>), and it doesn't immediately initialize the variables (instead it returns an operation that you have to run yourself).<\/p>\n"}]}
{"QuestionId":41944128,"AnswerCount":1,"Tags":"<tensorflow><mnist>","CreationDate":"2017-01-30T19:41:08.913","AcceptedAnswerId":41945055.0,"OwnerUserId":7302262.0,"Title":"Tensor Flow MNIST Evaluating predictions","Body":"<p>I am working on this <a href=\"https:\/\/pythonprogramming.net\/tensorflow-neural-network-session-machine-learning-tutorial\/?completed=\/tensorflow-deep-neural-network-machine-learning-tutorial\/\" rel=\"nofollow noreferrer\">tutorial<\/a> and I found in this following code: when evaluating predictions he runs accuracy which runs the correct variable which in turn runs prediction which will reintialize again the weights with randoms and reconstruct the NN model. How is this right? What am I missing ?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"41945055","CreationDate":"2017-01-30T20:42:30.570","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>You almost got it right. The <code>accuracy<\/code> tensor indirectly depends on the <code>prediction<\/code> tensor, which is depending on a Tensor <code>x<\/code>. In your code snippet you did not include what <code>x<\/code> actually is; however from the linked tutorial:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>x = tf.placeholder('float', [None, 784])\ny = tf.placeholder('float')\n<\/code><\/pre>\n\n<p>So <code>x<\/code> is a placeholder, i.e. a Tensor that obtains its value directly from the user. It is not entirely clear from the last line of<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_neural_network(x)\n<\/code><\/pre>\n\n<p>that he is not actually calling a transformation function <code>train_neural_network(x)<\/code> that takes an <code>x<\/code> and processes it on the fly, like you would expect from a regular function; rather, the function uses a reference to the previously defined placeholder variables - dummies, really - in order to define a computation graph it then directly executes using a session. \nThe graph, however, is only constructed once using <code>neural_network_model(x)<\/code> and then <em>queried<\/em> for a given number of epochs.<\/p>\n\n<p>What you missed is this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>_, c = sess.run([optimizer, cost], feed_dict={x: epoch_x, y: epoch_y})\n<\/code><\/pre>\n\n<p>This queries the result of the <code>optimizer<\/code> operation and the <code>cost<\/code> tensor given that the input values are <code>epoch_x<\/code> for <code>x<\/code> and <code>epoch_y<\/code> for <code>y<\/code>, pulling data through all defined computation nodes, all the way back \"down\" to <code>x<\/code>. In order to obtain the <code>cost<\/code>, <code>y<\/code> is needed as well. Both are provided by the caller. The <code>AdamOptimizer<\/code> will update all trainable variables as part of its execution, changing the network's weights.<\/p>\n\n<p>After that,<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>accuracy.eval({x: mnist.test.images, y: mnist.test.labels})\n<\/code><\/pre>\n\n<p>or, equivalently<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels})\n<\/code><\/pre>\n\n<p>then issues another evaluation of the <em>same<\/em> graph - without changing it - but this time using the inputs <code>mnist.test.images<\/code> for <code>x<\/code> and <code>mnist.test.labels<\/code> for <code>y<\/code>. \nIt works because <code>prediction<\/code> itself depends on <code>x<\/code>, which is overridden to the user-provided values on each call to <code>sess.run(...)<\/code>.<\/p>\n\n<p>Here is what the graph looks like in TensorBoard. It's hard to tell, but the two placeholder nodes are on the bottom left, next to the orange node called \"Variable\" and in the center right, below the green \"Slice_1\". <\/p>\n\n<p>Here's how the relevant part of the network's graph looks like; I exported this using TensorBoard. It's a bit hard to get since the nodes are not manually labeled (except for a couple I labeled myself), but there are six relevant points here. Placeholders are yellow: On the bottom right you'll find <code>x<\/code> and <code>y<\/code> is on the center left. \nGreen are the intermediate values that make sense to us: On the left is the <code>prediction<\/code> tensor, on the right there's the tensor called <code>correct<\/code>. The blue parts are endpoints of the graph: On the top left there's the <code>cost<\/code> tensor and on the top right you'll find <code>accuracy<\/code>. In essence, data flows from the bottom to the top.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ziEh6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ziEh6.png\" alt=\"Network graph\"><\/a><\/p>\n\n<p>So, whenever you say \"evaluate <code>prediction<\/code> given <code>x<\/code>\", \"evaluate <code>accuracy<\/code> given <code>x<\/code> and <code>y<\/code>\" or \"optimize my network given <code>x<\/code> and <code>y<\/code>\", you really just provide values on the yellow ends and observe the outcome on the green or blue ones.<\/p>\n"}]}
{"QuestionId":41944233,"AnswerCount":1,"Tags":"<python><tensorflow><scikit-learn><deep-learning>","CreationDate":"2017-01-30T19:47:21.420","AcceptedAnswerId":null,"OwnerUserId":6717054.0,"Title":"Tensorflow copy of sklearn MLPRegressor produces other results","Body":"<p>I am trying to reproduce a deep learning regression result in Tensorflow. If I train a neural network with the MLPRegressor class from sklearn I get very nice results of 98% validation.<\/p>\n\n<p>The MLPRegressor:<\/p>\n\n<p><a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.neural_network.MLPRegressor.html#sklearn.neural_network.MLPRegressor\" rel=\"nofollow noreferrer\">http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.neural_network.MLPRegressor.html#sklearn.neural_network.MLPRegressor<\/a><\/p>\n\n<p>I am trying to reproduce the model in Tensorflow. By copying the default values of the MLPRegressor class in a Tensorflow model. However I cannot get the same result. I only get 75% most of the time.<\/p>\n\n<p>My TF model:<\/p>\n\n<pre><\/pre>\n\n<p>My TF session:<\/p>\n\n<pre><\/pre>\n\n<p>Am I missing something in the Tensorflow code? Thanks!<\/p>\n","answers":[{"AnswerId":"41948646","CreationDate":"2017-01-31T02:11:41.630","ParentId":null,"OwnerUserId":"2059708","Title":null,"Body":"<p>Unless you have a very good reason to not use them, regression should have linear output units. I ran into a similar problem a while back and ended up using linear outputs and linear hidden units which seemed to mirror the mlpregressor in my case.<\/p>\n\n<p>There is a great section in Goodfellow's <a href=\"http:\/\/www.deeplearningbook.org\" rel=\"nofollow noreferrer\">Deep Learning Book<\/a> in <a href=\"http:\/\/www.deeplearningbook.org\/contents\/mlp.html\" rel=\"nofollow noreferrer\">chapter 6<\/a>, starting at page 181, that goes over the activation functions.<\/p>\n\n<p>At the very least try this for your output layer<\/p>\n\n<pre><code> layer_2 = tf.matmul(layer_1, W[\"layer_2\"]) + b[\"layer_2\"]\n<\/code><\/pre>\n"}]}
{"QuestionId":41944660,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><keras>","CreationDate":"2017-01-30T20:17:01.027","AcceptedAnswerId":41946137.0,"OwnerUserId":1367204.0,"Title":"Keras: Why isn't observed batch size matching up with the specified batch size?","Body":"<p>I specify a batch size of 500 by doing this in my code: <\/p>\n\n<pre><\/pre>\n\n<p>When I run the code, the first batch size is 500, and the batch sizes after that are like 5000 and larger, why does this happen?<\/p>\n\n<p>The reason I think the batch size are larger is because it seems like the model goes from row 500 to row 6000, which is 5500 rows. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41946137","CreationDate":"2017-01-30T21:58:54.287","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>This is really interesting issue. The part of code which is responsible for displaying a progress bar is a util called <code>progbar<\/code> and it's defined <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/utils\/generic_utils.py\" rel=\"nofollow noreferrer\">here<\/a>. It accepts as a parameter a minimum visual progress update interval which is by default set to <code>0.01<\/code> seconds. This default value is also used during printing a progress bar during <code>fit<\/code> computations and this is probably the reason behind this weird behaviour.<\/p>\n"}]}
{"QuestionId":41945414,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-01-30T21:07:03.427","AcceptedAnswerId":41946033.0,"OwnerUserId":3102085.0,"Title":"How to assign tensor values to a numpy array","Body":"<p>There are ways to initialize tensor with a numpy array. But is there any way to do the opposite. Meaning initialize a numpy array with a tensor in the graph.<\/p>\n","answers":[{"AnswerId":"41946033","CreationDate":"2017-01-30T21:52:10.373","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>You have to evaluate the tensor under a session. Assume you have a tensor <code>t<\/code> defined as<\/p>\n\n<pre><code>x = tf.Variable(...)\ny = tf.Variable(...)\n\nt = tf.add(x, y)\n<\/code><\/pre>\n\n<p>and you want to know its value (given the current <code>x<\/code> and <code>y<\/code>).\nYou then just use a session to fetch it:<\/p>\n\n<pre><code>with tf.Session() as sess:\n    numpy_array = sess.run(t)\n<\/code><\/pre>\n\n<p>or, equivalently,<\/p>\n\n<pre><code>with tf.Session() as sess:\n    numpy_array = t.eval(sess)\n<\/code><\/pre>\n\n<p>That should be all there is to it.<\/p>\n"}]}
{"QuestionId":41945544,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><conv-neural-network><tensorboard>","CreationDate":"2017-01-30T21:16:57.257","AcceptedAnswerId":41992499.0,"OwnerUserId":6182430.0,"Title":"AttributeError while using summary operation","Body":"<p>I built a CNN using TensorFlow. The network worked fine, but I had a problem: I couldn't visualize and plot graphs from the learning process. <\/p>\n\n<p>Therefore I implemented the necessary commands in order to use <strong>TensorBoard<\/strong>, following <a href=\"https:\/\/www.tensorflow.org\/how_tos\/summaries_and_tensorboard\/\" rel=\"nofollow noreferrer\"><strong>this<\/strong><\/a> tutorial.<\/p>\n\n<p>However, when I run the code I get the following <strong>error message:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>Referring to the following commands (specific to the lines with the **):<\/p>\n\n<p>in the <strong><em>main<\/em><\/strong> function:<\/p>\n\n<pre><\/pre>\n\n<p>in <strong><em>variable_summaries<\/em><\/strong> function:<\/p>\n\n<pre><\/pre>\n\n<p>What is this error message? I followed the tutorial step by step and I couldn't find the mistake.<\/p>\n\n<p>Appreciate your help, thanks! :)<\/p>\n\n<p>The whole code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41992499","CreationDate":"2017-02-02T00:47:08.907","ParentId":null,"OwnerUserId":"6182430","Title":null,"Body":"<p>Following <a href=\"https:\/\/stackoverflow.com\/questions\/41945544\/attributeerror-while-using-summary-operation?noredirect=1#comment71073013_41945544\">the comment<\/a> of <a href=\"https:\/\/stackoverflow.com\/users\/195651\/sunside\">sunside<\/a>, I updated my tensorflow version and the problem solved.<\/p>\n\n<p>Apparently, <code>tf.scalar_summary()<\/code> worked at tensorflow version 0.10, but updated to <code>tf.summary.scalar()<\/code> at newer versions (0.12, at least).<\/p>\n\n<p><code>pip install -U tensorflow<\/code> in the terminal solved the problem immediately :)<\/p>\n"}]}
{"QuestionId":41945761,"AnswerCount":1,"Tags":"<neural-network><keras><conv-neural-network><keras-layer>","CreationDate":"2017-01-30T21:32:41.223","AcceptedAnswerId":41945970.0,"OwnerUserId":5051564.0,"Title":"Keras Dense Layer Error: TypeError: 'int' object is not callable","Body":"<p>I'm trying to visualize the output of each convolutional layer in keras, following this link: <a href=\"https:\/\/github.com\/julienr\/ipynb_playground\/blob\/master\/keras\/convmnist\/keras_cnn_mnist.ipynb\" rel=\"nofollow noreferrer\">MNIST Visualisation<\/a>. I have modified some layers to remove errors, but now I'm stuck with the Dense Layer Error.<\/p>\n\n<pre><\/pre>\n\n<p>Any comment or suggestion highly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"41945970","CreationDate":"2017-01-30T21:47:57.010","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>If you check the documentation of a <a href=\"https:\/\/keras.io\/layers\/core\/#dense\" rel=\"nofollow noreferrer\"><code>Dense<\/code><\/a> layer then you'll notice that the first argument it accepts is the shape of output and second is <code>init<\/code> which describes the way how weights of the layer are initiated. In your case you provided the <code>int<\/code> as second positional argument and this caused error. You should change the code to (assuming that you want an output in a form of 128-dimensional vector):<\/p>\n\n<pre><code>model.add(Dense(128))\n<\/code><\/pre>\n"}]}
{"QuestionId":41946591,"AnswerCount":0,"Tags":"<python><caffe><conv-neural-network><pycaffe><resnet>","CreationDate":"2017-01-30T22:31:51.647","AcceptedAnswerId":null,"OwnerUserId":1179925.0,"Title":"Caffe: feature extractor using python","Body":"<p>I'm trying to extract features from <a href=\"https:\/\/github.com\/KaimingHe\/deep-residual-networks\" rel=\"nofollow noreferrer\">pretrained ResNet models<\/a>.<\/p>\n\n<p>Howewer at some point it crash with error:<\/p>\n\n<pre><\/pre>\n\n<p>Howewer I have tested AlexNet model and it seems ok.  shows me that GPU memory is not full(I have gtx 1070 with 8Gb of memory and batch size is 1, so it should not be a problem). <\/p>\n\n<p>What can be the reason of such behaviour?<\/p>\n\n<p>Here is the code:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41946843,"AnswerCount":0,"Tags":"<ubuntu><tensorflow><gpu>","CreationDate":"2017-01-30T22:52:04.057","AcceptedAnswerId":null,"OwnerUserId":7465627.0,"Title":"Unable to use GPU processors for tensorflow LinearRegressor on Ubuntu (Titan-X pascal)","Body":"<p>I have a (CUDA 8) - 3units of Titan-X Pascal on Ubuntu 16.04 with tensorflow v0.12. (Nvidia Driver Version: 370.28)<\/p>\n\n<p>I have trouble running any contrib.learn modules (LinearRegressor) to use the processor (GPU-Util is always showing as 0% in nvidia-smi)<\/p>\n\n<p>I know that my setup is correct because the cifar10 multi-gpu example (here: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py<\/a>) works perfectly fine using all the GPUs.<\/p>\n\n<p>I have also submitted an issue on tensorflow here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6132\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6132<\/a><\/p>\n\n<p>Note: \n- I have changed the run_config to set allow_soft_placement=True\n- I have tried adding specific GPU to model_fn as well as feed_fn separately but no luck. <\/p>\n\n<p>Does anyone else face the same problem for contrib.learn packages? Any luck?<\/p>\n","answers":[]}
{"QuestionId":41946894,"AnswerCount":1,"Tags":"<tensorflow><keras><conv-neural-network>","CreationDate":"2017-01-30T22:56:43.287","AcceptedAnswerId":41952772.0,"OwnerUserId":5656742.0,"Title":"Keras dimension mismatch with ImageDataGenerator","Body":"<p>I am attempting to 'flow' my data into a neural network with Keras.  I am using the .flow_from_directory method and the process is giving me fits.  I am using the basic example from the keras documentation (I am using tensorflow):<\/p>\n\n<pre><\/pre>\n\n<p>Running this i get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have been tinkering around for a long time and found the following--switching the 'model.add' to grayscale input \n    model.add(Convolution2D(4, 4, 4, border_mode='same', input_shape=(64, 64,3), activation='relu'))\ngives me the following error (as expected--but appears to confirm my original input was correct):<\/p>\n\n<pre><\/pre>\n\n<p>So I am passing (in the original) a 4-d array of 32,64,64,3 with the original, but I am getting the error that I THINK means \n   Expected (1,64,64,3) and got (32,64,64,3)<\/p>\n\n<p>As I am sending data in batches of 32.  Curiously enough if I set the batch to zero (to give a 0,64,64,3 input) I get:<\/p>\n\n<pre><\/pre>\n\n<p>Based on the documentation, I cannot figure out the proper way to flow the data into the model--i cannot pass the batch size to the model when using fit_generator, and it appears that the batch_size (num of samples) is the problem.  <\/p>\n\n<p>Any help would be greatly appreciated.<\/p>\n","answers":[{"AnswerId":"41952772","CreationDate":"2017-01-31T08:22:37.483","ParentId":null,"OwnerUserId":"5215538","Title":null,"Body":"<p>There is no problem with your <code>ImageDataGenerator<\/code>. As stated in the error message there is a mismatch between the shape of your model output and the shape of its targets. You use <code>class_mode = 'binary'<\/code>, so expected output of your model is a single value, but instead it yields output of shape <code>(batch_size, 64, 64, 4)<\/code> since you have one convolutional layer and nothing else in your model.<\/p>\n\n<p>Try something like this:<\/p>\n\n<pre><code>model.add(Convolution2D(4, 4, 4, border_mode='same', input_shape=(64, 64,3), activation='relu'))\nmodel.add(Flatten())\nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n<\/code><\/pre>\n"}]}
{"QuestionId":41947039,"AnswerCount":1,"Tags":"<python><neural-network><keras><recurrent-neural-network>","CreationDate":"2017-01-30T23:08:16.310","AcceptedAnswerId":48521460.0,"OwnerUserId":1639886.0,"Title":"Keras RNN with LSTM cells for predicting multiple output time series based on multiple intput time series","Body":"<p>I would like to model RNN with LSTM cells in order to predict multiple output time series based on multiple input time series. To be specific, I have 4 output time series, y1[t], y2[t], y3[t], y4[t], each has a length 3,000 (t=0,...,2999). I also have 3 input time series, x1[t], x2[t], x3[t], and each has a length 3,000 sec (t=0,...,2999). The goal is to predict y1[t],.. y4[t] using all the input time series up to this current time point i.e.:<\/p>\n\n<pre><\/pre>\n\n<p>For a model to have a long term memory, I created a stateful RNN model by following. <a href=\"http:\/\/philipperemy.github.io\/keras-stateful-lstm\/\" rel=\"noreferrer\">keras-stateful-lstme<\/a>. The main difference between my case and <a href=\"http:\/\/philipperemy.github.io\/keras-stateful-lstm\/\" rel=\"noreferrer\">keras-stateful-lstme<\/a> is that I have:<\/p>\n\n<ul>\n<li>more than 1 output time series<\/li>\n<li>more than 1 input time series<\/li>\n<li>the goal is the prediction of continuous time series<\/li>\n<\/ul>\n\n<p>My code is running. However the model's prediction result is bad even with a simple data. So I would like to ask you if I am getting anything wrong.<\/p>\n\n<p>Here is my code with a toy example.<\/p>\n\n<p>In toy example, our input time series are simple cosign and sign waves:<\/p>\n\n<pre><\/pre>\n\n<p>Please notice that y1 at time point t is simply the value of x1 at t - 2.\nPlease also notice that y3 at time point t is simply the value of x1 in the two previous step.<\/p>\n\n<p>Using these functions, I generated 100 sets of time series y1,y2,y3,x1,x2,x3,x4. Half of them go to training data and the remaining half go to testing data.<\/p>\n\n<pre><\/pre>\n\n<p>X, y are both 3 dimensional and each contains:<\/p>\n\n<pre><\/pre>\n\n<p>The example of the time series y1, .. y4 and x1, .., x3 are shown as below:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/u100N.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/u100N.png\" alt=\"enter image description here\"><\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/PtOvt.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/PtOvt.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I standardize these data as:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Create a stateful RNN model with 10 LSTM hidden neurons<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>Now following code is used to train and validate the RNN model:<\/p>\n\n<pre><\/pre>\n\n<p>The output is the following:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, the training loss is NOT decreasing!!<\/p>\n\n<p>As the target time series 1 and 3 have very simple relations with the input time series (y1[t] = x1[t-2] , y3[t] = x4[t-3]), I would expect perfect prediction performance. However, testing R2 at every epoch shows that that is not the case. R2 at the final epoch is just about 0.2 and 0.36. Clearly, the algorithm is not converging. I am very puzzled with this result. Please do let me know what I am missing, and why the algorithm is not converging. <\/p>\n","answers":[{"AnswerId":"48521460","CreationDate":"2018-01-30T12:26:13.997","ParentId":null,"OwnerUserId":"4342311","Title":null,"Body":"<p><strong>Initial note.<\/strong> If time series were short (for example T = 30), we wouldn't need stateful LSTM and classic LSTM would work well.\nIn OP question, time series lengths are T=3000, so learning can be very slow with classic LSTM. Learning will be improved by cutting the time series into pieces and using stateful LSTM.<\/p>\n\n<p><strong>Stateful mode with N=batch_size.<\/strong>\nStateful models are tricky with Keras, because you need to be careful on how you cut time series and select batch size. In OP question, sample size is N=100. Since we can accept to train model with batch of one hundred (it is not a large number), we will select batch_size=100.<\/p>\n\n<p>Selecting batch_size=N simplifies the training because you do not need to reset states inside epochs (so no need to write a callback on_batch_begin).<\/p>\n\n<p>It remains the question of cutting the time series. Cutting is a little technical, so I wrote a wrapper function working in all cases.<\/p>\n\n<pre><code>def stateful_cut(arr, batch_size, T_after_cut):\n    if len(arr.shape) != 3:\n        # N: Independent sample size,\n        # T: Time length,\n        # m: Dimension\n        print(\"ERROR: please format arr as a (N, T, m) array.\")\n\n    N = arr.shape[0]\n    T = arr.shape[1]\n\n    # We need T_after_cut * nb_cuts = T\n    nb_cuts = int(T \/ T_after_cut)\n    if nb_cuts * T_after_cut != T:\n        print(\"ERROR: T_after_cut must divide T\")\n\n    # We need batch_size * nb_reset = N\n    # If nb_reset = 1, we only reset after the whole epoch, so no need to reset\n    nb_reset = int(N \/ batch_size)\n    if nb_reset * batch_size != N:\n        print(\"ERROR: batch_size must divide N\")\n\n    # Cutting (technical)\n    cut1 = np.split(arr, nb_reset, axis=0)\n    cut2 = [np.split(x, nb_cuts, axis=1) for x in cut1]\n    cut3 = [np.concatenate(x) for x in cut2]\n    cut4 = np.concatenate(cut3)\n    return(cut4)\n<\/code><\/pre>\n\n<p>From now, it become easy to train the model. Since the OP example is very simple, we do not need additional preprocessing or regularization. I describe how to proceed step by step <em>(for the impatient, whole self-contained code is available at the very end of this post).<\/em><\/p>\n\n<p>First we load data and reshape it with wrapper function.<\/p>\n\n<pre><code>import numpy as np\nfrom keras.models import Sequential\nfrom keras.layers import Dense, LSTM, TimeDistributed\nimport matplotlib.pyplot as plt\nimport matplotlib.patches as mpatches\n\n##\n# Data\n##\nN = X_train.shape[0] # size of samples\nT = X_train.shape[1] # length of each time series\nbatch_size = N # number of time series considered together: batch_size | N\nT_after_cut = 100 # length of each cut part of the time series: T_after_cut | T\ndim_in = X_train.shape[2] # dimension of input time series\ndim_out = y_train.shape[2] # dimension of output time series\n\ninputs, outputs, inputs_test, outputs_test = \\\n  [stateful_cut(arr, batch_size, T_after_cut) for arr in \\\n  [X_train, y_train, X_test, y_test]]\n<\/code><\/pre>\n\n<p>Then we compile a model with 4 inputs, 3 outputs, and 1 hidden layer containing 10 nodes.<\/p>\n\n<pre><code>##\n# Model\n##\nnb_units = 10\n\nmodel = Sequential()\nmodel.add(LSTM(batch_input_shape=(batch_size, None, dim_in),\n               return_sequences=True, units=nb_units, stateful=True))\nmodel.add(TimeDistributed(Dense(activation='linear', units=dim_out)))\nmodel.compile(loss = 'mse', optimizer = 'rmsprop')\n<\/code><\/pre>\n\n<p>We train the model without resetting states. We can do it only because we have selected batch_size = N. <\/p>\n\n<pre><code>##\n# Training\n##\nepochs = 100\n\nnb_reset = int(N \/ batch_size)\nif nb_reset &gt; 1:\n    print(\"ERROR: We need to reset states when batch_size &lt; N\")\n\n# When nb_reset = 1, we do not need to reinitialize states\nhistory = model.fit(inputs, outputs, epochs = epochs, \n                    batch_size = batch_size, shuffle=False,\n                    validation_data=(inputs_test, outputs_test))\n<\/code><\/pre>\n\n<p>We get evolution of training\/test loss as follows:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Qfvu6.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Qfvu6.png\" alt=\"training and test loss decreasing as a function of epochs as expected\"><\/a><\/p>\n\n<p>Now, we define a 'mime model' which is stateless but containing our stateful weights. [Why like this? Prediction with stateful model through model.predict needs a complete batch in Keras, but we may not have a complete batch to predict...]<\/p>\n\n<pre><code>## Mime model which is stateless but containing stateful weights\nmodel_stateless = Sequential()\nmodel_stateless.add(LSTM(input_shape=(None, dim_in),\n               return_sequences=True, units=nb_units))\nmodel_stateless.add(TimeDistributed(Dense(activation='linear', units=dim_out)))\nmodel_stateless.compile(loss = 'mse', optimizer = 'rmsprop')\nmodel_stateless.set_weights(model.get_weights())\n<\/code><\/pre>\n\n<p>Finally, we can show our incredible predictions on our long time series y1, y2 and y3 (blue for true output ; orange for predicted outputs):<\/p>\n\n<p><em>For y1:<\/em>\n<a href=\"https:\/\/i.stack.imgur.com\/Ar0mn.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Ar0mn.png\" alt=\"Prediction for y1\"><\/a><\/p>\n\n<p><em>For y2:<\/em>\n<a href=\"https:\/\/i.stack.imgur.com\/HkLwt.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HkLwt.png\" alt=\"Prediction for y2\"><\/a><\/p>\n\n<p><em>For y3:<\/em>\n<a href=\"https:\/\/i.stack.imgur.com\/D7PZw.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/D7PZw.png\" alt=\"Prediction for y3\"><\/a><\/p>\n\n<p><strong>Conclusion<\/strong>: It works almost perfectly, unless for the 2-3 first dates where the series is unpredictable by definition. We do not observe any burst when going from one batch for the next batch.<\/p>\n\n<p><strong>Much more<\/strong> When N is large, we would like to select batch_size | N with batch_size &lt; N. I have written full code in <a href=\"https:\/\/github.com\/ahstat\/deep-learning\/blob\/master\/rnn\/4_lagging_and_stateful.py\" rel=\"noreferrer\">https:\/\/github.com\/ahstat\/deep-learning\/blob\/master\/rnn\/4_lagging_and_stateful.py<\/a> (Part C and D). This github path also shows efficiency of classic LSTM for short time series (Part A), and inefficiency for long time series (Part B). I've written a blog post detailing how to use Keras for time series predictions here: <a href=\"https:\/\/ahstat.github.io\/RNN-Keras-time-series\/\" rel=\"noreferrer\">https:\/\/ahstat.github.io\/RNN-Keras-time-series\/<\/a> .<\/p>\n\n<p><strong>Complete self-contained code<\/strong><\/p>\n\n<pre><code>################\n# Code from OP #\n################\nimport numpy as np\ndef random_sample(len_timeseries=3000):\n    Nchoice = 600\n    x1 = np.cos(np.arange(0,len_timeseries)\/float(1.0 + np.random.choice(Nchoice)))\n    x2 = np.cos(np.arange(0,len_timeseries)\/float(1.0 + np.random.choice(Nchoice)))\n    x3 = np.sin(np.arange(0,len_timeseries)\/float(1.0 + np.random.choice(Nchoice)))\n    x4 = np.sin(np.arange(0,len_timeseries)\/float(1.0 + np.random.choice(Nchoice)))\n    y1 = np.random.random(len_timeseries)\n    y2 = np.random.random(len_timeseries)\n    y3 = np.random.random(len_timeseries)\n    for t in range(3,len_timeseries):\n        ## the output time series depend on input as follows: \n        y1[t] = x1[t-2] \n        y2[t] = x2[t-1]*x3[t-2]\n        y3[t] = x4[t-3]\n    y = np.array([y1,y2,y3]).T\n    X = np.array([x1,x2,x3,x4]).T\n    return y, X\ndef generate_data(Nsequence = 1000):\n    X_train = []\n    y_train = []\n    for isequence in range(Nsequence):\n        y, X = random_sample()\n        X_train.append(X)\n        y_train.append(y)\n    return np.array(X_train),np.array(y_train)\n\nNsequence = 100\nprop = 0.5\nNtrain = int(Nsequence*prop)\nX, y = generate_data(Nsequence)\nX_train = X[:Ntrain,:,:]\nX_test  = X[Ntrain:,:,:]\ny_train = y[:Ntrain,:,:]\ny_test  = y[Ntrain:,:,:] \n\n#X.shape = (N sequence, length of time series, N input features)\n#y.shape = (N sequence, length of time series, N targets)\nprint(X.shape, y.shape)\n# (100, 3000, 4) (100, 3000, 3)\n\n####################\n# Cutting function #\n####################\ndef stateful_cut(arr, batch_size, T_after_cut):\n    if len(arr.shape) != 3:\n        # N: Independent sample size,\n        # T: Time length,\n        # m: Dimension\n        print(\"ERROR: please format arr as a (N, T, m) array.\")\n\n    N = arr.shape[0]\n    T = arr.shape[1]\n\n    # We need T_after_cut * nb_cuts = T\n    nb_cuts = int(T \/ T_after_cut)\n    if nb_cuts * T_after_cut != T:\n        print(\"ERROR: T_after_cut must divide T\")\n\n    # We need batch_size * nb_reset = N\n    # If nb_reset = 1, we only reset after the whole epoch, so no need to reset\n    nb_reset = int(N \/ batch_size)\n    if nb_reset * batch_size != N:\n        print(\"ERROR: batch_size must divide N\")\n\n    # Cutting (technical)\n    cut1 = np.split(arr, nb_reset, axis=0)\n    cut2 = [np.split(x, nb_cuts, axis=1) for x in cut1]\n    cut3 = [np.concatenate(x) for x in cut2]\n    cut4 = np.concatenate(cut3)\n    return(cut4)\n\n#############\n# Main code #\n#############\nfrom keras.models import Sequential\nfrom keras.layers import Dense, LSTM, TimeDistributed\nimport matplotlib.pyplot as plt\nimport matplotlib.patches as mpatches\n\n##\n# Data\n##\nN = X_train.shape[0] # size of samples\nT = X_train.shape[1] # length of each time series\nbatch_size = N # number of time series considered together: batch_size | N\nT_after_cut = 100 # length of each cut part of the time series: T_after_cut | T\ndim_in = X_train.shape[2] # dimension of input time series\ndim_out = y_train.shape[2] # dimension of output time series\n\ninputs, outputs, inputs_test, outputs_test = \\\n  [stateful_cut(arr, batch_size, T_after_cut) for arr in \\\n  [X_train, y_train, X_test, y_test]]\n\n##\n# Model\n##\nnb_units = 10\n\nmodel = Sequential()\nmodel.add(LSTM(batch_input_shape=(batch_size, None, dim_in),\n               return_sequences=True, units=nb_units, stateful=True))\nmodel.add(TimeDistributed(Dense(activation='linear', units=dim_out)))\nmodel.compile(loss = 'mse', optimizer = 'rmsprop')\n\n##\n# Training\n##\nepochs = 100\n\nnb_reset = int(N \/ batch_size)\nif nb_reset &gt; 1:\n    print(\"ERROR: We need to reset states when batch_size &lt; N\")\n\n# When nb_reset = 1, we do not need to reinitialize states\nhistory = model.fit(inputs, outputs, epochs = epochs, \n                    batch_size = batch_size, shuffle=False,\n                    validation_data=(inputs_test, outputs_test))\n\ndef plotting(history):\n    plt.plot(history.history['loss'], color = \"red\")\n    plt.plot(history.history['val_loss'], color = \"blue\")\n    red_patch = mpatches.Patch(color='red', label='Training')\n    blue_patch = mpatches.Patch(color='blue', label='Test')\n    plt.legend(handles=[red_patch, blue_patch])\n    plt.xlabel('Epochs')\n    plt.ylabel('MSE loss')\n    plt.show()\n\nplt.figure(figsize=(10,8))\nplotting(history) # Evolution of training\/test loss\n\n##\n# Visual checking for a time series\n##\n## Mime model which is stateless but containing stateful weights\nmodel_stateless = Sequential()\nmodel_stateless.add(LSTM(input_shape=(None, dim_in),\n               return_sequences=True, units=nb_units))\nmodel_stateless.add(TimeDistributed(Dense(activation='linear', units=dim_out)))\nmodel_stateless.compile(loss = 'mse', optimizer = 'rmsprop')\nmodel_stateless.set_weights(model.get_weights())\n\n## Prediction of a new set\ni = 0 # time series selected (between 0 and N-1)\nx = X_train[i]\ny = y_train[i]\ny_hat = model_stateless.predict(np.array([x]))[0]\n\nfor dim in range(3): # dim = 0 for y1 ; dim = 1 for y2 ; dim = 2 for y3.\n    plt.figure(figsize=(10,8))\n    plt.plot(range(T), y[:,dim])\n    plt.plot(range(T), y_hat[:,dim])\n    plt.show()\n\n## Conclusion: works almost perfectly.\n<\/code><\/pre>\n"}]}
{"QuestionId":41947083,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-30T23:12:03.720","AcceptedAnswerId":null,"OwnerUserId":1481013.0,"Title":"How to create a product layer using tensorflow?","Body":"<p>I am trying to implement an SPN (<a href=\"http:\/\/spn.cs.washington.edu\/pubs.shtml\" rel=\"nofollow noreferrer\">http:\/\/spn.cs.washington.edu\/pubs.shtml<\/a>) using tensorflow.\nIn an SPN, we only have sum or product nodes. A value of a sum node is computed by summing the product of its respective edge-weight and corresponding child. A value of a product node is the product of its children.<\/p>\n\n<p>This is a toy example:\n<a href=\"https:\/\/i.stack.imgur.com\/dtwLf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/dtwLf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I could easily compute the first sum layer using this piece of tensorflow code:<\/p>\n\n<pre><\/pre>\n\n<p>Which gives me the correct values for the sum nodes in the first layer:<\/p>\n\n<pre><\/pre>\n\n<p>My question is how can I compute the product layer?<\/p>\n\n<p>The product node should create the following output:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/reduction\" rel=\"nofollow noreferrer\"><strong>tf.reduce_prod<\/strong><\/a>, but it only gives [[1296.0]].<\/p>\n\n<p>Thank you for your thoughts.<\/p>\n","answers":[]}
{"QuestionId":41947157,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-30T23:19:53.313","AcceptedAnswerId":52227168.0,"OwnerUserId":3662820.0,"Title":"Tensorflow - input from DataFrame using batching and sparse\/categorical data","Body":"<p>I have a relatively small dataset that I load into memory using a pandas DataFrame. I'd like to feed this data to a tensorflow model using batching, while maintaining support for sparse (categorical) columns. I'd also like to avoid having to serialize my data to disk in some other format. Although this doesn't seem too complicated I couldn't find a good example in the docs and had a pretty tough time designing a suitable  myself.<\/p>\n\n<p>An toy example dataset would be:<\/p>\n\n<pre><\/pre>\n\n<p>where  is a dense numeric column,  is a categorical column, and  is a binary label column.<\/p>\n\n<p>My solution is below, anything prettier and\/or more efficient would be great.<\/p>\n","answers":[{"AnswerId":"41947158","CreationDate":"2017-01-30T23:19:53.313","ParentId":null,"OwnerUserId":"3662820","Title":null,"Body":"<p>Here's my (OP) solution. The conversion and serialization step is pretty slow (about 3 seconds per 1000 samples). Anything more efficient would be greatly appreciated.<\/p>\n\n<pre><code>import tensorflow as tf\n\n######################################\n# Define Feature conversion functions\n######################################\ndef int64_feature(value):\n    return tf.train.Feature(int64_list=tf.train.Int64List(value=[int(value)]))\n\ndef float_feature(value):\n    return tf.train.Feature(float_list=tf.train.FloatList(value=[float(value)]))\n\ndef bytes_feature(value):\n    return tf.train.Feature(bytes_list=tf.train.BytesList(value=[str(value)]))\n\n####################################################\n# Define tensorflow data feed from pandas DataFrame\n####################################################\ndef input_fn(df, label_col_name, int_col_names, float_col_names, cat_col_names, num_epochs, batch_size, shuffle=False):\n    # Define new column groups\n    feature_col_names = int_col_names + float_col_names + cat_col_names\n    all_col_names = [label_col_name] + feature_col_names\n\n    # Create conversion and parser dicts\n    converters = {}\n    parse_dict = {}\n    for col in all_col_names:\n        if col in cat_col_names:\n            converters[col] = bytes_feature\n            parse_dict[col] = tf.VarLenFeature(tf.string)\n        elif col in float_col_names:\n            converters[col] = float_feature\n            parse_dict[col] = tf.FixedLenFeature([], tf.float32)\n        elif col in int_col_names + [label_col_name]:\n            converters[col] = int64_feature\n            parse_dict[col] = tf.FixedLenFeature([], tf.int64)\n\n    # Convert DataFrame rows to feature Examples, serialize examples to string\n    serialized_examples = []\n    for record in df[all_col_names].to_dict('records'):\n        feat_record = {k: converters[k](v) for k,v in record.iteritems()}\n        example = tf.train.Example(features=tf.train.Features(feature=feat_record))\n        serialized_examples.append(example.SerializeToString())\n\n    # Create input queue\n    example_queue = tf.train.slice_input_producer([serialized_examples], num_epochs=num_epochs, shuffle=shuffle)\n\n    # Create batch\n    example_batch = tf.train.batch(example_queue, batch_size=batch_size, capacity=30, allow_smaller_final_batch=True)\n\n    # Parse batch\n    parsed_example_batch = tf.parse_example(example_batch, parse_dict)\n\n    # Split into features and label\n    feature_batch = {k: parsed_example_batch[k] for k in feature_col_names}\n    label_batch = parsed_example_batch[label_col_name]\n\n    return feature_batch, label_batch\n<\/code><\/pre>\n\n<p>example usage:<\/p>\n\n<pre><code>import functools\nimport numpy as np\nimport pandas as pd\n\n# Create toy dataset\ndf = pd.DataFrame(np.random.randint(1, 4, [7, 3]), columns=['c0', 'c1', 'c2'])\ndf['c1'] = df['c1'].astype(str) + 'g'\ndf['c2'] = (df['c2'] &gt; 2.5).astype(int)\n\n# Specify feature names\ncat_feats = ['c1']\nfloat_feats = []\nint_feats = ['c0']\nlabel_feat = 'c2'\n\n# Create parameterless input function\nepochs = 3\nbatch_size = 2\ninput_fn_train = functools.partial(input_fn, df, label_feat, int_feats, float_feats, cat_feats, epochs, batch_size)\n\n# Define features\ncontinuous_features = [tf.contrib.layers.real_valued_column(feat) for feat in float_feats+int_feats]\ncategorical_features = [tf.contrib.layers.sparse_column_with_hash_bucket(feat, hash_bucket_size=1000) for feat in cat_feats]\nfeatures = continuous_features + categorical_features\n\n# Create and fit model\nmodel = tf.contrib.learn.LinearClassifier(feature_columns=features)\nmodel.fit(input_fn=input_fn_train, steps=1000)\n<\/code><\/pre>\n"},{"AnswerId":"52227168","CreationDate":"2018-09-07T17:29:24.843","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>Firstly, why are you serializing these to tf.Examples, then deserializing them with parse_example? Serializing them, then batching them, then deserializing them, does unnecessary work. The standard way to define input functions in tensorflow is with <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets\" rel=\"nofollow noreferrer\">tf.data<\/a>. For <code>tf.data<\/code> + <code>tf.Estimators<\/code>, <a href=\"https:\/\/www.tensorflow.org\/guide\/datasets_for_estimators\" rel=\"nofollow noreferrer\">this documentation<\/a> might be helpful.  And in this case specifically, this code should work:<\/p>\n\n<pre><code>def input_fn(df, label_feat, num_epochs, batch_size, shuffle=False):\n  # Each element of dataset is one row of the dataframe\n  dataset = tf.data.Dataset.from_tensor_slices(dict(df))\n\n  def map_fn(element, label_feat):\n    # element is a {'c0': int, 'c1': str, 'c2': int} dictionary\n    label = element.pop(label_feat)\n    return (element, label)\n\n  if shuffle:\n    dataset = dataset.shuffle(shuffle_buffer_size)\n\n  # Batch the elements of the dataset\n  dataset = dataset.batch(batch_size)\n  # Repeat the dataset for num_epochs\n  dataset = dataset.repeat(num_epochs)\n\n  # Split it into features, label tuple\n  dataset = dataset.map(lambda elem: map_fn(elem, label_feat)\n\n  # One shot iterator iterates through the (repeated) dataset once, \n  # yielding feature_batch, label_batch\n  iterator = dataset.make_one_shot_iterator()\n  feature_batch, label_batch = iterator.get_next()\n  return feature_batch, label_batch\n<\/code><\/pre>\n\n<p>Additionally, based on your code, it seems like there might be some confusion here about SparseTensors, vs sparse_column. When you used <code>tf.VarLenFeature<\/code>, the feature is parsed as a SparseTensor, and this is only necessary when the feature being parsed has variable shape. In this case, your <code>c1<\/code> features are all scalar string tensors, so <code>FixedLenFeature<\/code> should work for this, there is no need for the feature to be represented as sparse tensors, even if they eventually get represented as a <code>sparse_column<\/code>. <a href=\"https:\/\/www.tensorflow.org\/tutorials\/representation\/linear\" rel=\"nofollow noreferrer\">This documentation<\/a> tells you more about sparse columns.<\/p>\n"}]}
{"QuestionId":41947492,"AnswerCount":1,"Tags":"<tensorflow><neural-network><theano><keras><conv-neural-network>","CreationDate":"2017-01-30T23:53:47.323","AcceptedAnswerId":41948730.0,"OwnerUserId":5051564.0,"Title":"How can I get the output out of a tensor object in keras?","Body":"<p>I'm using Keras' pre-trained VGG16 model, and I want to to visualise the output of each layer. However, layer.output returns a tensor object - how can I convert it something that allows me to get the image outputs?<\/p>\n\n<pre><\/pre>\n\n<p>Also, when I try to access specific node's output, it returns a tensor:<\/p>\n\n<pre><\/pre>\n\n<p>Any help is greatly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"41948730","CreationDate":"2017-01-31T02:25:18.457","ParentId":null,"OwnerUserId":"17328","Title":null,"Body":"<p>You need to evaluate the tensor, which is probably best done by configuring the model to return them when you run predict.<\/p>\n\n<p>e.g. <\/p>\n\n<pre><code>layer_outputs = [layer.output for layer in model.layers]\nviz_model = Model(input=model.input, output=layer_outputs)\n...\nfeatures = viz_model.predict(x)\nfor feature_map in features:\n   ...\n<\/code><\/pre>\n\n<p>Also check out this blog post, which walks through a possibly similar exercise to what you're attempting: <a href=\"https:\/\/blog.keras.io\/how-convolutional-neural-networks-see-the-world.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/how-convolutional-neural-networks-see-the-world.html<\/a><\/p>\n"}]}
{"QuestionId":41948322,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><classification>","CreationDate":"2017-01-31T01:26:21.560","AcceptedAnswerId":null,"OwnerUserId":3447653.0,"Title":"Split data into batches in Tensorflow for classification","Body":"<p>I am training the data using below code. But here I need to split the data into batches (as giving in MNIST tensorflow example:  batch_xs, batch_ys = mnist.train.next_batch(100)).<\/p>\n\n<p>When I looked for next_batch method, it was not available in any library. I am a newbie to Tensor flow. Just wondering whether I can split the data into batches while training.<\/p>\n\n<p>Any help would be really appreciated. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42010989","CreationDate":"2017-02-02T19:43:06.423","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p><code>next_batch<\/code> isn't a general-purpose library method. It is a method specific to the MNIST data set that reads the next minibatch of inputs to feed te Tensorflow.<\/p>\n\n<p>You can see its implementation here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160<\/a><\/p>\n\n<p>You would need to implement something similar for your dataset.\nThis code isn't specific to Tensorflow --- you can use any Python libraries you like to do this.<\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41948367,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-31T01:32:04.700","AcceptedAnswerId":42011097.0,"OwnerUserId":1214630.0,"Title":"How to load a very large into Tensorflow and create minibatchs?","Body":"<p>I have a hdf5 file that has been written into the disk by a python code (I only have the file, not the code). The size of this file is 90GB and the data in this file has the following format: (N, 250,360,3). Just as side note, the data can't fit into memory.<\/p>\n\n<p>Now I want to write a data loader in Tensorflow where each time just loads M samples from this file (M is way smaller than N).<\/p>\n\n<p>What would be the best way to do this? Any pointer to a code would be highly appreciated.<\/p>\n\n<p>Thanks.\nJ<\/p>\n","answers":[{"AnswerId":"42011097","CreationDate":"2017-02-02T19:49:32.937","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>The Tensorflow MNIST tutorial shows how this can be done:\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/<\/a><\/p>\n\n<p>If you look at the implementation on Github, you'll see that it uses a <code>next_batch<\/code> function to read batches of inputs 100 at a time.<\/p>\n\n<p>The implementation of <code>next_batch<\/code> lives here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160<\/a><\/p>\n\n<p>You would need to implement something similar for your data set. I'm not particularly familiar with HDF5, but you can use any Python library to do the loading; it doesn't have to be specific to Tensorflow.<\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41948406,"AnswerCount":4,"Tags":"<machine-learning><tensorflow><nvidia><keras>","CreationDate":"2017-01-31T01:37:53.850","AcceptedAnswerId":null,"OwnerUserId":4926741.0,"Title":"Why is my GPU slower than CPU when training LSTM\/RNN models?","Body":"<p>My machine has the following spec: <\/p>\n\n<p>CPU: Xeon E5-1620 v4<\/p>\n\n<p>GPU: Titan X (Pascal) <\/p>\n\n<p>Ubuntu 16.04<\/p>\n\n<p>Nvidia driver 375.26<\/p>\n\n<p>CUDA tookit 8.0<\/p>\n\n<p>cuDNN 5.1<\/p>\n\n<p>I've benchmarked on the following Keras examples with Tensorflow as the backed <a href=\"https:\/\/github.com\/fchollet\/keras\/tree\/master\/examples\" rel=\"noreferrer\">reference<\/a>: <\/p>\n\n<pre><\/pre>\n\n<p>My gpu is clearly out performing my cpu in non-lstm models. <\/p>\n\n<pre><\/pre>\n\n<p>Has anyone else experienced this? <\/p>\n","answers":[{"AnswerId":"44694528","CreationDate":"2017-06-22T08:47:36.157","ParentId":null,"OwnerUserId":"4893073","Title":null,"Body":"<p>I have got similar issues here:<\/p>\n\n<h1>Test 1<\/h1>\n\n<p>CPU: Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz<\/p>\n\n<p>Ubuntu 14.04<\/p>\n\n<p><code>imdb_bidirectional_lstm.py<\/code>: <strong>155s<\/strong><\/p>\n\n<h1>Test 2<\/h1>\n\n<p>GPU: GTX 860m<\/p>\n\n<p>Nvidia Driver: 369.30<\/p>\n\n<p>CUDA Toolkit: v8.0<\/p>\n\n<p>cuDNN: v6.0<\/p>\n\n<p><code>imdb_bidirectional_lstm.py<\/code>:<strong>450s<\/strong><\/p>\n\n<h1>Analyse<\/h1>\n\n<p>When I observe the GPU load curve, I found one interesting thing: <\/p>\n\n<ul>\n<li>for lstm, GPU load jumps quickly between ~80% and ~10%<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6pnyB.png\" rel=\"nofollow noreferrer\">GPU load<\/a><\/p>\n\n<p>This is mainly due to the sequential computation in LSTM layer. Remember that LSTM requires sequential input to calculate hidden layer weights iteratively, in other words, you must wait for hidden state at time <code>t-1<\/code> to calculate hidden state at time <code>t<\/code>. <\/p>\n\n<p>That's not a good idea for GPU cores, since they are many small cores who like doing computations in parallel, sequential compuatation can't fully utilize their computing powers. That's why we are seeing GPU load around 10% - 20% most of the time.<\/p>\n\n<p>But in the phase of backpropagation, GPU could run derivative computation in parallel, so we can see GPU load peak around 80%.<\/p>\n"},{"AnswerId":"50614523","CreationDate":"2018-05-30T23:13:35.127","ParentId":null,"OwnerUserId":"2050067","Title":null,"Body":"<p>If you use Keras, use <a href=\"https:\/\/keras.io\/layers\/recurrent\/#cudnnlstm\" rel=\"noreferrer\">CuDNNLSTM<\/a> in place of <a href=\"https:\/\/keras.io\/layers\/recurrent\/#lstm\" rel=\"noreferrer\">LSTM<\/a> or <a href=\"https:\/\/keras.io\/layers\/recurrent\/#cudnngru\" rel=\"noreferrer\">CuDNNGRU<\/a> in place of <a href=\"https:\/\/keras.io\/layers\/recurrent\/#gru\" rel=\"noreferrer\">GRU<\/a>. In my case (2 Tesla M60), I am seeing 10x boost of performance. By the way I am using batch size 128 as suggested by @Alexey Golyshev.<\/p>\n"},{"AnswerId":"51470275","CreationDate":"2018-07-23T00:22:25.987","ParentId":null,"OwnerUserId":"8762113","Title":null,"Body":"<p>It's just a tip. <\/p>\n\n<p>Using GPU is powerful when<\/p>\n\n<ol>\n<li><p>your Neural Network is big.<\/p><\/li>\n<li><p>your Batch_Size is big.<\/p><\/li>\n<\/ol>\n\n<p>-- It's what I found from googling. <\/p>\n"},{"AnswerId":"41952863","CreationDate":"2017-01-31T08:28:40.193","ParentId":null,"OwnerUserId":"5638904","Title":null,"Body":"<p>Too small batch size. Try to increase.<\/p>\n\n<p>Results for my GTX1050Ti:<\/p>\n\n<pre>\nimdb_bidirectional_lstm.py\nbatch_size      time\n32 (default)    252\n64              131\n96              87\n128             66\n\nimdb_lstm.py\nbatch_size      time\n32 (default)    108\n64              50\n96              34\n128             25\n<\/pre>\n"}]}
{"QuestionId":41948545,"AnswerCount":0,"Tags":"<theano><keras>","CreationDate":"2017-01-31T01:57:42.633","AcceptedAnswerId":null,"OwnerUserId":4651001.0,"Title":"slicing keras Variable custom objective function","Body":"<p>I've been trying to implement a custom objective function in Keras (the negative log likelihood of the normal distribution)<\/p>\n\n<p>Keras expects one argument for the ground truth tensor, and one for the predictions tensor; for y_pred,I'm passing a tensor that should represent a nx2 matrix where the first column is the mean of the distribution, and the second the precision.<\/p>\n\n<p>My problem is that I haven't been able to get a clear idea how I properly slice y_pred before passing it into the likelihood function without yielding the error<\/p>\n\n<pre><\/pre>\n\n<p>While I understand that I'm feeding l_func arguments of the variable type when it expects an array,I don't seem to be able to grok how to properly split the input y_pred variable into its mean and precision components to plug into the likelihood function.  Here are some attempts; if someone could enlighten me about how to proceed, I would greatly appreciate it.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41948574,"AnswerCount":0,"Tags":"<python><neural-network><keras>","CreationDate":"2017-01-31T02:01:25.383","AcceptedAnswerId":null,"OwnerUserId":7492729.0,"Title":"Real bad accuracy of training test on neural network on keras","Body":"<p>I'm doing a Neural Network for the \"Default of credit card clients\" from <a href=\"http:\/\/archive.ics.uci.edu\/ml\/datasets\/default+of+credit+card+clients\" rel=\"nofollow noreferrer\">http:\/\/archive.ics.uci.edu\/ml\/datasets\/default+of+credit+card+clients<\/a>.<\/p>\n\n<p>But the accuracy of my models is pretty bad, worse than if I predicted all zeros. I already did some research on it and did oversample to correct the classes, and change the optimizer, because adam was not increasing the accuracy. <\/p>\n\n<p>What else could I do?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":41948625,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-31T02:08:16.340","AcceptedAnswerId":null,"OwnerUserId":3834415.0,"Title":"TensorFlow: How to make learning rate in Gradient Descent decrease over time?","Body":"<p>When declaring gradient descent, I apparently have to define the train step prior to initialization: <\/p>\n\n<pre><\/pre>\n\n<p>How do I adjust eta as I increase the number of epochs I learn? <\/p>\n","answers":[]}
{"QuestionId":41949233,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-31T03:35:31.987","AcceptedAnswerId":42011125.0,"OwnerUserId":1212998.0,"Title":"Tensorflow MNIST for Beginners","Body":"<p>Why do I get different results each time I run the code? Granted, they are all around 92% test accuracy, but seeing as there is no randomness involved (weights are initialized to zero), and sampling the training data does not seem to be random, where could the randomness (resulting in different but close accuracy values) be coming from?<\/p>\n\n<p>Or could there be something wrong with my installation? Am using an nvidia 960 gtx gpu, and installed TF from sources (version 0.12.head)<\/p>\n","answers":[{"AnswerId":"42011125","CreationDate":"2017-02-02T19:51:49.713","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>If you look at the implementation of the <code>next_batch<\/code> function, you'll find it randomly shuffles the data set: <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160<\/a><\/p>\n\n<p>Try adding the argument <code>shuffle=False<\/code> to the <code>next_batch<\/code> call.<\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41949633,"AnswerCount":2,"Tags":"<python><scope><tensorflow>","CreationDate":"2017-01-31T04:27:08.250","AcceptedAnswerId":null,"OwnerUserId":4744283.0,"Title":"Tensorflow: how to swap variables between scopes and set variables in scope from another","Body":"<p>I have different scopes and they have variables with same names but with different values. I want to swap values of these variables between scopes.\nExample: <\/p>\n\n<pre><\/pre>\n\n<p>I want to set  to 0,  to 1,  to 2 and  to 3.<\/p>\n\n<p>I was thinking about getting needed variables with  but I can't see how I can change scope of the variable, so probably I need to change values of variables. In this case I need to store one value in temporary variable and then remove that temporary variable. \nI don't sure that it will work and this seems too complicated.\nIs there simple way to do it?<\/p>\n\n<p><strong>UPD1:<\/strong> Also I need to set all variables in one collection from another collection. I think that it's similar issue. \nFor example in code above set  equal to 0 and  to 1.<\/p>\n\n<p><strong>UPD2:<\/strong> This code doesn't work:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"55038509","CreationDate":"2019-03-07T07:44:10.747","ParentId":null,"OwnerUserId":"11163488","Title":null,"Body":"<pre><code>import tensorflow as tf\nimport numpy as np\n\nwith tf.variable_scope('sc1'):\n    a1 = tf.get_variable(name='test_var1', initializer=0.)\n    b1 = tf.Variable(0, name='test_var2')\n\nwith tf.variable_scope('sc2'):\n    a2 = tf.get_variable(name='test_var1', initializer=1.)\n    b2 = tf.Variable(1, name='test_var2')\n\n\ndef swap_tf_scopes(col1, col2):\n    col1_dict = {}\n    for curr_var in col1:\n        curr_var_name = curr_var.name.split('\/')[-1]\n        col1_dict[curr_var_name] = curr_var\n    for curr_var in col2:\n        curr_var_name = curr_var.name.split('\/')[-1]\n        curr_col1_var = col1_dict[curr_var_name]\n        tmp_t =tf.Variable(curr_col1_var.initialized_value()) \n        sess.run(tmp_t.initializer)\n        sess.run(tf.assign(curr_col1_var,curr_var))\n        sess.run(tf.assign(curr_var,tmp_t))\n\ncol1 = tf.get_collection(tf.GraphKeys.VARIABLES, scope='sc1')\ncol2 = tf.get_collection(tf.GraphKeys.VARIABLES, scope='sc2')\nsess = tf.Session()\nsess.run(tf.initialize_all_variables())\nswap_tf_scopes(col1, col2)\nprint(sess.run(col1)) \nprint(sess.run(col2))\n<\/code><\/pre>\n\n<p>Hello\uff01Try this one. I guess it will work.<\/p>\n"},{"AnswerId":"41969243","CreationDate":"2017-01-31T23:15:22.883","ParentId":null,"OwnerUserId":"4744283","Title":null,"Body":"<p>I want to swap relatively small scopes, so it's not a problem have temporary scope for swapping. I made working prototype. It doesn't look cool and actually ugly but works.<\/p>\n\n<pre><code>def swap_tf_collections(col1, col2, tmp_col):\n    col2_dict = {}\n    for i in xrange(len(col1)):\n        curr_var_name = col2[i].name.split('\/')[-1]\n        col2_dict[curr_var_name] = col2[i]\n\n    col1_dict = {}\n    for i in xrange(len(col1)):\n        curr_var_name = col1[i].name.split('\/')[-1]\n        col1_dict[curr_var_name] = col1[i]\n\n    # set values from second scope to tmp_dict \n    tmp_dict = {}\n    assigns_arr = []\n    for i in xrange(len(tmp_col)):\n        curr_var_name = tmp_col[i].name.split('\/')[-1]\n        tmp_dict[curr_var_name] = tmp_col[i]\n        assign0 = tmp_dict[curr_var_name].assign(col2_dict[curr_var_name])\n        assigns_arr.append(assign0)\n\n    for i in xrange(len(col2)):\n        curr_var_name = col2[i].name.split('\/')[-1]\n        curr_col1_var = col1_dict[curr_var_name]\n        tmp_t = tmp_dict[curr_var_name]\n        with tf.control_dependencies(assigns_arr):\n            assign1 = col2[i].assign(curr_col1_var)\n            assigns_arr.append(assign1)\n            with tf.control_dependencies(assigns_arr):\n                assign2 = curr_col1_var.assign(tmp_t)\n                assigns_arr.append(assign2)\n    return assigns_arr\n\n\n\n# first scope\nwith tf.variable_scope('sc1'):\n    a1 = tf.get_variable(name='test_var1', initializer=0.)\n    b1 = tf.Variable(0, name='test_var2')\n# second scope    \nwith tf.variable_scope('sc2'):\n    a2 = tf.get_variable(name='test_var1', initializer=1.)\n    b2 = tf.Variable(1, name='test_var2')\n\n# getting them as collections\ncol1 = tf.get_collection(tf.GraphKeys.VARIABLES, scope='sc1')\ncol2 = tf.get_collection(tf.GraphKeys.VARIABLES, scope='sc2')\n\n# creating temporary scope. It MUST have same variables names as our scopes but it doesn't have to have same data as second scope\nwith tf.variable_scope('tmp_scope_for_scopes_swap'):\n    for i in xrange(len(col2)):\n        col2_var = col2[i]\n        col2_var_name = col2[i].name.split('\/')[-1].split(':')[0]\n        var = tf.Variable(col2_var.initialized_value(), name=col2_var_name)\n        tmp_col.append(var)\n\n# exec\nsess = tf.Session()\nwith sess.as_default():\n    sess.run(tf.initialize_all_variables())\n    tf_ops_t = swap_tf_collections(col1, col2, tmp_col) \n    sess.run(tf_ops_t)  # swap will not work without this line\n\ncol1_dict = {i.name:i for i in col1}\ncol2_dict = {i.name:i for i in col2}\nprint sess.run(col1_dict)\nprint sess.run(col2_dict)\n<\/code><\/pre>\n\n<p>Note that I use dependency control! Without it function results will be undefined.<\/p>\n\n<p>Also despite that function name is <code>swap_tf_collections<\/code> I think that it will not work for arbitrary collections (to be fair I doubt even about scopes).<\/p>\n"}]}
{"QuestionId":41949689,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-31T04:34:21.447","AcceptedAnswerId":null,"OwnerUserId":4561965.0,"Title":"import tensorflow failed in eclipse","Body":"<p>Installed tensorflow for python3.5 on Mac, try the testing code as follows:<\/p>\n\n<pre><\/pre>\n\n<p>It is working when run it from the terminal.\nBut if I run it from eclipse, will get this exception:<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone help here?\nThanks a lot.<\/p>\n","answers":[]}
{"QuestionId":41950583,"AnswerCount":0,"Tags":"<cuda><tensorflow><gpu>","CreationDate":"2017-01-31T05:54:37.697","AcceptedAnswerId":null,"OwnerUserId":2745266.0,"Title":"Can I use different GPUs (GTX 970 and GTX 1060) together for distributed TensorFlow?","Body":"<p>I currently own a 4GB Asus GTX 970 GPU. As it is now unavailable in the market, I am planning to buy 6GB Asus GTX 1060 GPU. I wish to use these GPUs together with distributed TensorFlow framework. Will TensorFlow support GTX 970 (4GB) and GTX 1060 (6GB) together? I have seen Titan and GTX being incompatible (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4013\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4013<\/a>) as 'they can not talk to each other and they are of different generations'. Will above mentioned two cards work together?<\/p>\n","answers":[]}
{"QuestionId":41951433,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-01-31T06:58:16.523","AcceptedAnswerId":null,"OwnerUserId":2324298.0,"Title":"Tensorflow - ValueError: Shape must be rank 1 but is rank 0 for 'ParseExample\/ParseExample'","Body":"<p>I have a  file of the Ubuntu Dialog Corpus. I am trying to read in the whole dataset so that I can split the contexts and utterances into batches. Using  I was able to read in a single example. I tried using  but I get the following error<\/p>\n\n<pre><\/pre>\n\n<p>I am not sure what to make of it. The code I used to get to the error - <\/p>\n\n<pre><\/pre>\n\n<p>Any ideas<\/p>\n","answers":[{"AnswerId":"48712647","CreationDate":"2018-02-09T19:15:36.560","ParentId":null,"OwnerUserId":"2762495","Title":null,"Body":"<p>I had a similar problem just now. Try putting brackets around the serialized_example to turn it into a list:<\/p>\n\n<pre><code>features = tf.parse_example([serialized_example], \nfeatures = {\n\"context\" : tf.FixedLenFeature([160], tf.int64),\n\"context_len\" : tf.FixedLenFeature([1], tf.int64),\n\"utterance\" : tf.FixedLenFeature([80], tf.int64),\n\"utterance_len\" : tf.FixedLenFeature([1], tf.int64),\n\"label\" : tf.FixedLenFeature([1], tf.int64)\n})\n<\/code><\/pre>\n"},{"AnswerId":"44056966","CreationDate":"2017-05-18T20:11:01.310","ParentId":null,"OwnerUserId":"8033317","Title":null,"Body":"<p>to use tf.parse_example you need to first batch the examples:<\/p>\n\n<pre><code>batch = tf.train.batch([serialized_example], num_examples, capacity=num_examples)\nparsed_examples = tf.parse_example(batch, feature_spec)\n<\/code><\/pre>\n"}]}
{"QuestionId":41951657,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-31T07:12:12.310","AcceptedAnswerId":null,"OwnerUserId":7492867.0,"Title":"Using tf.trainable_variables() to show names of trainable variables","Body":"<p>I'm dealing with <em>weights<\/em> in a . <\/p>\n\n<p>I have the following code<\/p>\n\n<pre><\/pre>\n\n<p>In another file called , I am printing the names of the trainable variables just appended to . However, I find that the \"name\" attribute of the trainable variables is not useful because not very descriptive.<\/p>\n\n<p>Do the weights of an  have names?<\/p>\n\n<p>If that's useful, here's another possibly relevant piece of code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43035867","CreationDate":"2017-03-26T23:36:05.563","ParentId":null,"OwnerUserId":"6684307","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>variables_names = [v.name for v in tf.trainable_variables()]\nvalues = sess.run(variables_names)\nfor k, v in zip(variables_names, values):\n    print \"Variable: \", k\n    print \"Shape: \", v.shape\n    print v\n<\/code><\/pre>\n"}]}
{"QuestionId":41952642,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><caffe>","CreationDate":"2017-01-31T08:13:44.683","AcceptedAnswerId":41973977.0,"OwnerUserId":1279459.0,"Title":"Getting and editing gradient parameters in Caffe","Body":"<p>Is it possible to get the gradients with respect to each layer in Caffe in CNNs, edit them and again apply the new gradients in the training process? If possible, using pycaffe interface.<\/p>\n\n<p>For example in TensorFlow, it could be done by means of functions:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41973977","CreationDate":"2017-02-01T07:15:18.180","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I'm not sure what you mean by \"apply the new gradients in the training process\", but you can access the gradients in the pycaffe interface:<\/p>\n\n<pre><code>import caffe\nnet = caffe.Net('\/path\/to\/net.prototxt', '\/path\/to\/weights.caffemodel', caffe.TEST)\n# provide inputs to the net, do a pass so that meaningful data\/gradients propagate to all the layers\nnet.forward_backward_all()\n\n# once data\/gradients are updated, you can access them\nnet.blobs['blob_name'].diff  # access the gradient of blob 'blob_name'\nnet.layers[5].blobs[0].diff  # access the gradient of the first parameter blob of the 6th layer\n<\/code><\/pre>\n\n<p>To map between layer names and layer indices, you can use this code:<\/p>\n\n<pre><code>list(net._layer_names).index('layer_name')\n<\/code><\/pre>\n\n<p>This will return the index of layer <code>'layer_name'<\/code>.<\/p>\n"}]}
{"QuestionId":41952768,"AnswerCount":0,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-01-31T08:22:30.670","AcceptedAnswerId":null,"OwnerUserId":7493814.0,"Title":"what parameter affect the training time in CNN?","Body":"<p>I am doing a project where I need to change the parameter or something in cifar10 by CNN, and then collect the different training time of them. Does anyone know which parameters should I change? \nThanks!<\/p>\n","answers":[]}
{"QuestionId":41953610,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-01-31T09:14:25.657","AcceptedAnswerId":null,"OwnerUserId":4229089.0,"Title":"TensorFlow assign concat-result tensor to tensor of different shape","Body":"<p>I'm using concat to append a tensor t2 of shape [150, 1] onto a tensor t1 which has shape initially of [150, 0]. I want to create a tensor which has shape [150, 1] (concatenate over axis=1), and override t1 with the new tensor. Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>The error I'm getting is:<\/p>\n\n<pre><\/pre>\n\n<p>Oddly enough, I'm getting the same error when I set axis=0. No idea what's going on; this should be a very easy thing to do. I suspect it has something to do with t1 having a Dimension 1 of 0. Any suggestions? Thanks in advance.<\/p>\n\n<p>[EDIT]<\/p>\n\n<p>Turns out, the problem is with \"assign\", as I'm setting t1 (shape (150, 0)) to the concat result (shape (150, 1)), which is making TF scream at me (rightfully so). Any way around this?<\/p>\n\n<p>To provide more context, t1 is a tensor which I'm using to store the logit t2 generated from the unrolling of a seq2seq RNN, in which the output sequence has an unknown length. So the size of t1 will be growing in Dimension 1 with every unrolling.<\/p>\n","answers":[{"AnswerId":"41956835","CreationDate":"2017-01-31T11:49:46.030","ParentId":null,"OwnerUserId":"5861256","Title":null,"Body":"<p>Use <code>reshape<\/code>:  <\/p>\n\n<pre><code> t1 = tf.placeholder(tf.float32,[150,1])  \n t2 = tf.placeholder(tf.float32,[150])      \n tf.concat(1,[t1,tf.reshape(t2,[150,1])])  \n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>&lt;tf.Tensor 'concat_2:0' shape=(150, 2) dtype=float32&gt;\n<\/code><\/pre>\n"},{"AnswerId":"44756381","CreationDate":"2017-06-26T09:13:27.533","ParentId":null,"OwnerUserId":"3921457","Title":null,"Body":"<p>The problem is related with the shape update on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/assign\" rel=\"nofollow noreferrer\">tf.assign<\/a>. You can set <code>validate_shape<\/code> to <code>False<\/code>. This will remove the error on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/assign\" rel=\"nofollow noreferrer\">tf.assign<\/a>, but will generate a bad behavior because of the shape will not be updated.<\/p>\n\n<p>It's a TensorFlow's bug and there is a GitHub's issue related with it waiting for R's implementation here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5492\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5492<\/a><\/p>\n"}]}
{"QuestionId":41953678,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-31T09:18:30.380","AcceptedAnswerId":null,"OwnerUserId":5182224.0,"Title":"Documentation for Inference from saved model in Tensorflow","Body":"<p>As explained at <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/#build_a_multilayer_convolutional_network\" rel=\"nofollow noreferrer\">Multilayer CNN for MNIST Digit Recognition<\/a> I created a CNN for MNIST image recognition dataset. I don't want to test it in the same script so I trained the model, saved the checkpoint files and graph structure and then using <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"nofollow noreferrer\">freeze_graph.py<\/a> script. I froze it into one single .pb file. Now using this .pb file I want to infer for MNIST test dataset and check the accuracy of the model but there is no clear documentation as to how to do it. There is one example written in examples\/label_images directory but that too is in C++ and for inception network. I have already parsed the .pb file and made a tf.Graph object. And also why can't a single inference script be provided since whole graph structure can be read from graph saved in the .pb file.<\/p>\n","answers":[]}
{"QuestionId":41954198,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><gradient>","CreationDate":"2017-01-31T09:44:53.380","AcceptedAnswerId":null,"OwnerUserId":3227411.0,"Title":"optimizer.compute_gradients how the gradients are calculated programatically?","Body":"<p>I'm new to machine learning.I was going through tensorflow and i have a doubt on a particular function.<\/p>\n\n<p>grads_and_vars = optimizer.compute_gradients(loss) can someone explain how the gradients are calculated programatically(i.e what formula does it use to compute the gradients)?<\/p>\n","answers":[{"AnswerId":"42011465","CreationDate":"2017-02-02T20:12:16.180","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>Tensorflow uses an algorithm called reverse-mode automatic differentiation. It's too complex a topic to explain here, but the Wikipedia page is a good starting point:<\/p>\n\n<p><a href=\"https:\/\/en.wikipedia.org\/wiki\/Automatic_differentiation\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Automatic_differentiation<\/a><\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41954308,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-31T09:49:42.953","AcceptedAnswerId":41972119.0,"OwnerUserId":681870.0,"Title":"Loss function works with reduce_mean but not reduce_sum","Body":"<p>I'm new to tensor flow, and have been looking at the examples <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/\" rel=\"noreferrer\">here<\/a>. I wanted to rewrite the multilayer perceptron classification model to be a regression model. However I encountered some strange behaviour when modifying the loss function. It works fine with , but if I try using  it gives nan's in the output. This seems very strange, as the functions are very similar - the only difference is that the mean divides the sum result by the number of elements? So I can't see how nan's could be introduced by this change?<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The problematic variable  is commented out. It should be used in place of .<\/p>\n\n<p>With  the output looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>and with  it ends up like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41972119","CreationDate":"2017-02-01T04:50:00.873","ParentId":null,"OwnerUserId":"728291","Title":null,"Body":"<p>The loss by summing across the batch is 1000 times larger (from skimming the code I think your training batch size is 1000) so your gradients and parameter updates are also 1000 times larger.  The larger updates apparently lead to <code>nan<\/code>s.<\/p>\n\n<p>Generally learning rates are expressed per example so the loss to find the gradients for updates should be per example also.  If the loss is per batch then the learning rate needs to be reduced by the batch size to get comparable training results. <\/p>\n"}]}
{"QuestionId":41954366,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><face-recognition>","CreationDate":"2017-01-31T09:51:38.480","AcceptedAnswerId":null,"OwnerUserId":6191357.0,"Title":"I have built a CNN for detecting human faces . From the very first epoch I am getting higher accuracy. What might be the reason for it?","Body":"<p>The training data set consists of face images taken from youtube faces database labelled as one and non face images are taken from 256Object categories 25k images are chosen for both positive and negative data. so totally 50k for training and another 10k images is taken from yooutube faces and 256 object categories which are not repeated.<\/p>\n\n<p>The problem is I am getting 99% accuracy after just 12k iterations in first epoch itself and I am printing cost value also it is also starts from very high value like 596014.000 like that. When it is tested against the other face images it performs very badly.\n<a href=\"https:\/\/i.stack.imgur.com\/dekes.png\" rel=\"nofollow noreferrer\">cost vs epoch graph<\/a> <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41966228","CreationDate":"2017-01-31T19:45:43.607","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>The line<\/p>\n\n<pre><code>correct_pred = tf.equal(tf.greater(sigmoid_output, 0.5),\n                        tf.greater(output_data, 0.5))\n<\/code><\/pre>\n\n<p>is likely to be wrong. Your code seems not to have <code>sigmoid_output<\/code>, but only <code>pred<\/code>.<\/p>\n\n<p>If that is not the problem, I would look into <code>train_data<\/code>. How many \"True\" labels > 0.5 do you have? How many labels which evaluate to false do you have?<\/p>\n"}]}
{"QuestionId":41954416,"AnswerCount":1,"Tags":"<python><optimization><machine-learning><keras><mean-square-error>","CreationDate":"2017-01-31T09:54:03.470","AcceptedAnswerId":null,"OwnerUserId":4658980.0,"Title":"Add constraint to mean square error cost function in multi output regression","Body":"<p>I am using mean square error to compute the loss function of a multi output regressor. I used a recurrent neural network model with the one to many architecture. My output vector is of size 6 (1*6) and the values are monotonic (non decreasing).<\/p>\n\n<p>example: \ny_i = [1,3,6,13,30,57,201]<\/p>\n\n<p>I would like to force the model to learn this dependency. Hence adding a constraint to the cost function. I am getting an error equal to 300 on the validation set. I believe after editing the mean square error loss function i will be able to get a better performance.<\/p>\n\n<p>I am using keras for the implementation. Here is the core model.<\/p>\n\n<pre><\/pre>\n\n<p>Any other tips concerning learning rate, decay, etc.. are appreciated.<\/p>\n","answers":[{"AnswerId":"43990181","CreationDate":"2017-05-15T23:17:06.617","ParentId":null,"OwnerUserId":"8016701","Title":null,"Body":"<p>Keep the mean squared error as just a metric. Use Smooth L1 loss instead. Here's my implementation.<\/p>\n\n<pre><code>#Define Smooth L1 Loss\ndef l1_smooth_loss(y_true, y_pred):\n    abs_loss = tf.abs(y_true - y_pred)\n    sq_loss = 0.5 * (y_true - y_pred)**2\n    l1_loss = tf.where(tf.less(abs_loss, 1.0), sq_loss, abs_loss - 0.5)\n    return tf.reduce_sum(l1_loss, -1)    \n#And\n\nModel.compile(loss='l1_smooth_loss', optimizer='rmsprop')\n<\/code><\/pre>\n"}]}
{"QuestionId":41954468,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-01-31T09:56:16.993","AcceptedAnswerId":null,"OwnerUserId":4325508.0,"Title":"Variable sharing tensorflow","Body":"<p><strong>Background<\/strong><\/p>\n\n<p>While going through the documentation of tensorflow,I came across the convolutional neural net example on the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/variable_scope\/\" rel=\"nofollow noreferrer\">Variable Sharing<\/a>.The examples motivates the need of variable sharing by describing a problem in which two images are exposed to image filter which would want to reuse the variables. <\/p>\n\n<p><strong>Problem<\/strong><\/p>\n\n<p>I am not able to get my head around this approach.Instead of introducing concept of variable sharing can we not place all the images in a matrix(in form of pixel values with each row denoting a new image) and perform necessary operations like filters on the whole matrix by using single set of variables.The approach I am suggesting is similar to what we use in vanilla neural networks.We dont use the concept of reusing variables for every training example rather we stack them up in form of matrix and have a common weight matrix for all the training examples to perform the necessary operations.<\/p>\n\n<p>Can someone point where am I wrong in setting up the congruency between the two approaches?<\/p>\n","answers":[{"AnswerId":"41955635","CreationDate":"2017-01-31T10:48:51.750","ParentId":null,"OwnerUserId":"15416","Title":null,"Body":"<p>In the forward pass, you can't distinguish between a shared variable and two identical variables. But in training, the backward pass is different. If you had two identical variables, you'll have two adjustments, and those will in general not be identical so the two variables will diverge. If you have a single shared variable, you have only a single adjustment.<\/p>\n"}]}
{"QuestionId":41954906,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><sha>","CreationDate":"2017-01-31T10:16:12.567","AcceptedAnswerId":null,"OwnerUserId":6842947.0,"Title":"SHA Hashing for training\/validation\/testing set split","Body":"<p>Following is a small snippet from the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"noreferrer\">full code<\/a><\/p>\n\n<p>I am trying to understand the logical process of this methodology of split.<\/p>\n\n<ul>\n<li>SHA1 encoding is 40 characters in hexadecimal. What kind of probability has been computed in the expression ? <\/li>\n<li>What is the reason for (MAX_NUM_IMAGES_PER_CLASS + 1) ? Why add 1 ?<\/li>\n<li>Does setting different values to MAX_NUM_IMAGES_PER_CLASS have an effect on the split quality ?<\/li>\n<li><p>How good a quality of split would we get out of this ? Is this is a recommended way of splitting datasets ?<\/p>\n\n<pre><\/pre><\/li>\n<\/ul>\n","answers":[{"AnswerId":"55985262","CreationDate":"2019-05-04T17:43:37.393","ParentId":null,"OwnerUserId":"8586227","Title":null,"Body":"<p>This code is simply distributing file names \u201crandomly\u201d (but reproducibly) over a number of bins and then grouping the bins into just the three categories.  The number of bits in the hash is irrelevant (so long as it\u2019s \u201cenough\u201d, which is probably about 35 for this sort of work).<\/p>\n\n<p>Reducing modulo <em>n<\/em>+1 produces a value on [0,<em>n<\/em>], and multiplying that by 100\/<em>n<\/em> obviously produces a value on [0,100], which is being interpreted as a percentage.  <em>n<\/em> being <code>MAX_NUM_IMAGES_PER_CLASS<\/code> is meant to control the rounding error in the interpretation to be no more than \u201cone image\u201d.<\/p>\n\n<p>This strategy is reasonable, but looks a bit more sophisticated than it is (since there is still rounding going on, and the remainder introduces a bias\u2014although with numbers this large it is utterly unobservable).  You could make it simpler and more accurate by simply precalculating ranges over the whole space of 2^160 hashes for each class and just checking the hash against the two boundaries.  That still notionally involves rounding, but with 160 bits it\u2019s only that intrinsic to representing decimals like 31% in floating point.<\/p>\n"}]}
{"QuestionId":41955089,"AnswerCount":0,"Tags":"<classification><keras><conv-neural-network>","CreationDate":"2017-01-31T10:24:37.077","AcceptedAnswerId":null,"OwnerUserId":7493893.0,"Title":"Keras model classifying well but predicted probabilities are always 1.0 or 0.0","Body":"<p>I am using Keras to build a multi-class (3 classes) image classifier.\nI trained the following model with a dataset of approximately 2000 images (1500 training\/ 500 validation).<\/p>\n\n<pre><\/pre>\n\n<p>These are the figures for the training\/validation <a href=\"https:\/\/i.stack.imgur.com\/se3ey.png\" rel=\"noreferrer\">accuracy<\/a> and <a href=\"https:\/\/i.stack.imgur.com\/cxbMV.png\" rel=\"noreferrer\">loss<\/a>.<\/p>\n\n<p>I achieve a 95% training accuracy, 92% validation accuracy and 94% on another separate chunk of images (aside from the 2000 of the dataset) that I have. <\/p>\n\n<p>Therefore, the model seems to classify reasonably well. However, my problem is that the predicted probabilities for an input image (obtained with the function <em>predict_proba()<\/em>) are always either 1.0 or 0.0. \nLikewise, if I give as an input an image that doesn't belong to any of the 3 classes I would expect some low probabilities (perhaps higher in the most similar class) but I still get 1.0 in one of the classes and 0.0 in the others.<\/p>\n\n<p>What could be causing that? It seems to me that there is no over fitting. Is there any issue with the model? \nCould it be that the images of each class are quite similar between them so somehow the model is quickly too confident on its decision?<\/p>\n","answers":[]}
{"QuestionId":41955404,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-01-31T10:36:49.033","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"Tensorflow: How to ignore specific labels during semantic segmentation?","Body":"<p>I'm using tensorflow for semantic segmentation. How can I tell tensorflow to ignore a specific label when computing the pixelwise loss?  <\/p>\n\n<p>I've read <a href=\"https:\/\/stackoverflow.com\/questions\/39091432\/tensorflow-softmax-ignore-negative-labels-just-like-caffe\/39113607\">in this post<\/a> that for image classification one can set the label to  and it will be ignored. If that is true, given the label-tensor, how can I modify my labels such that certain values are changed to ?<\/p>\n\n<p>In Matlab it would be something like:  <\/p>\n\n<pre><\/pre>\n\n<p>But I don't know how to do this in TF?<\/p>\n\n<p>Some background info:<br>\nThis is how the labels are loaded:  <\/p>\n\n<pre><\/pre>\n\n<p>This is how the loss is currently computed:  <\/p>\n\n<pre><\/pre>\n\n<p>with <\/p>\n\n<pre><\/pre>\n\n<p>I'm using the <a href=\"https:\/\/github.com\/DrSleep\/tensorflow-deeplab-resnet\" rel=\"nofollow noreferrer\">tensorflow-deeplab-resnet model<\/a> which transfers the Resnet model implemented in Caffe to tensorflow using <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\" rel=\"nofollow noreferrer\">caffe-tensorflow<\/a>. <\/p>\n","answers":[{"AnswerId":"41956383","CreationDate":"2017-01-31T11:26:36.290","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>According to the documentation, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.softmax_cross_entropy_with_logits<\/code><\/a> must be called with valid probability distributions on <code>labels<\/code>, or otherwise the computation will be incorrect, and using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/classification#sparse_softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>tf.nn.sparse_softmax_cross_entropy_with_logits<\/code><\/a> (which might be more convenient in your case) with negative labels will either cause an error or return NaN values. I wouldn't rely on it to have some labels ignored.<\/p>\n\n<p>What I would do is to replace the logits for the ignored class with infinity in those pixels where the correct class is the ignored one, so they will contribute nothing to the loss:<\/p>\n\n<pre><code>ignore_label = ...\n# Make zeros everywhere except for the ignored label\ninput_batch_ignored = tf.concat(input_batch.ndims - 1,\n    [tf.zeros_like(input_batch[:, :, :, :ignore_label]),\n     tf.expand_dims(input_batch[:, :, :, ignore_label], -1),\n     tf.zeros_like(input_batch[:, :, :, ignore_label + 1:])])\n# Make corresponding logits \"infinity\" (a big enough number)\npredictions_fix = tf.select(input_batch_ignored &gt; 0,\n    1e30 * tf.ones_like(predictions), predictions)\n# Compute loss with fixed logits\nloss = tf.nn.softmax_cross_entropy_with_logits(prediction, gt)\n<\/code><\/pre>\n\n<p>The only problem with this is that you are considering that pixels of the ignored class are always predicted correctly, which means that the loss for images containing a lot of those pixels will be artificially smaller. Depending on the case this may or may not be significant, but if you want to be really accurate, you would have to weight the loss of each image according to the number of not ignored pixels, instead of just taking the mean.<\/p>\n\n<pre><code># Count relevant pixels on each image\ninput_batch_relevant = 1 - input_batch_ignored\ninput_batch_weight = tf.reduce_sum(input_batch_relevant, [1, 2, 3])\n# Compute relative weights\ninput_batch_weight = input_batch_weight \/ tf.reduce_sum(input_batch_weight)\n# Compute reduced loss according to weights\nreduced_loss = tf.reduce_sum(loss * input_batch_weight)\n<\/code><\/pre>\n"},{"AnswerId":"49482629","CreationDate":"2018-03-26T01:07:13.387","ParentId":null,"OwnerUserId":"9550068","Title":null,"Body":"<p>Sorry, I'm new to this but I believe in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/g3doc\/faq.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/research\/deeplab\/g3doc\/faq.md<\/a>, it is mentioned that new datasets need to be added here. In \"segmentation_dataset.py\", at each dataset, you get to specify ignore_label. For example,<\/p>\n\n<pre><code>_PASCAL_VOC_SEG_INFORMATION = DatasetDescriptor(\n    splits_to_sizes={\n        'train': 1464,\n        'trainval': 2913,\n        'val': 1449,\n    },\n    num_classes=21,\n    ignore_label=255,\n)\n<\/code><\/pre>\n"}]}
{"QuestionId":41955567,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-31T10:44:49.850","AcceptedAnswerId":41961529.0,"OwnerUserId":2492058.0,"Title":"Incrementing variable as a copy in Tensorflow","Body":"<p>I currently have the following code which I would like to use to provide a \"stream\" of incrementing integers.<\/p>\n\n<pre><\/pre>\n\n<p>But the issue is that  is just a reference to , when I just want to enqueue a copy of  after incrementing it.<\/p>\n\n<p>Is this the correct way to go about this, or am I missing something?<\/p>\n","answers":[{"AnswerId":"41961529","CreationDate":"2017-01-31T15:40:29.217","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The current TensorFlow variables have unfortunate semantics when interacting with other stateful constructs (such as queues). The problem stems from \"reference types\" (note that <code>my_variable.dtype<\/code> is <code>tf.int64_ref<\/code>, which means that it is a <em>mutable<\/em> tensor reference), which most operations&mdash;including queues&mdash;implicitly \"dereference\" by creating a \"constant\" tensor that aliases the mutable buffer. We are in the process of fixing this bug in TensorFlow's memory model for variables, but the change is not in the public API yet.<\/p>\n\n<p>In the meantime, your best option is to force a copy when inserting the variable into the queue. This easiest solution relies on undocumented behavior, but <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/queues#QueueBase.enqueue_many\" rel=\"nofollow noreferrer\"><code>tf.QueueBase.enqueue_many()<\/code><\/a> will always copy its values into the queue, even when you enqueue a single element. When using it via <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#batch\" rel=\"nofollow noreferrer\"><code>tf.train.batch()<\/code><\/a>, you just need to reshape the variable (e.g. using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/shapes_and_shaping#expand_dims\" rel=\"nofollow noreferrer\"><code>tf.expand_dims()<\/code><\/a>) and pass <code>enqueue_many=True<\/code>. For example:<\/p>\n\n<pre><code>my_variable = tf.Variable(0, dtype=dtypes.int64)\n# ...\nqueued_increment = tf.train.batch((tf.expand_dims(my_variable, 1),), 1,\n                                  enqueue_many=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":41955726,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-01-31T10:53:20.183","AcceptedAnswerId":41975104.0,"OwnerUserId":3260576.0,"Title":"Tensorflow: Unable to visualize jpeg images","Body":"<p>I used the a script similar to the one <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/data\/build_imagenet_data.py\" rel=\"nofollow noreferrer\">here<\/a> to convert my dataset to sharded tfrecords. But when I attempt to read it using script below tensorflow freezes and I have to kill the process using kill. (Note: Right now I am working in CPU mode)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41975104","CreationDate":"2017-02-01T08:32:03.310","ParentId":null,"OwnerUserId":"3260576","Title":null,"Body":"<p>Was able to resolve this. I thought TFRecord is sort of dictionary and you need to give only the keys required, but upon giving the entire feature map along with small changes to how image was processed later, it worked.<\/p>\n\n<p>Another mistake I made was that queue_runner should be started after calling tf.train.shuffle_batch(). I dont know if it's a bug or a gap in my understanding <\/p>\n\n<p>Here's the working code for reading the data<\/p>\n\n<pre><code>def getImage(filename):\n    # convert filenames to a queue for an input pipeline.\n    filenameQ = tf.train.string_input_producer([filename],num_epochs=None)\n\n    # object to read records\n    recordReader = tf.TFRecordReader()\n\n    # read the full set of features for a single example \n    key, fullExample = recordReader.read(filenameQ)\n\n    # parse the full example into its' component features.\n    features = tf.parse_single_example(\n        fullExample,\n        features={\n            'image\/height': tf.FixedLenFeature([], tf.int64),\n            'image\/width': tf.FixedLenFeature([], tf.int64),\n            'image\/colorspace': tf.FixedLenFeature([], dtype=tf.string,default_value=''),\n            'image\/channels':  tf.FixedLenFeature([], tf.int64),            \n            'image\/class\/label': tf.FixedLenFeature([],tf.int64),\n            'image\/class\/text': tf.FixedLenFeature([], dtype=tf.string,default_value=''),\n            'image\/format': tf.FixedLenFeature([], dtype=tf.string,default_value=''),\n            'image\/filename': tf.FixedLenFeature([], dtype=tf.string,default_value=''),\n            'image\/encoded': tf.FixedLenFeature([], dtype=tf.string, default_value='')\n        })\n\n\n    # now we are going to manipulate the label and image features\n\n    label = features['image\/class\/label']\n    image_buffer = features['image\/encoded']\n\n    # Decode the jpeg\n    with tf.name_scope('decode_jpeg',[image_buffer], None):\n        # decode\n        image = tf.image.decode_jpeg(image_buffer, channels=3)\n\n        # and convert to single precision data type\n        image = tf.image.convert_image_dtype(image, dtype=tf.float32)\n\n\n    # cast image into a single array, where each element corresponds to the greyscale\n    # value of a single pixel. \n    # the \"1-..\" part inverts the image, so that the background is black.\n    # re-define label as a \"one-hot\" vector \n    # it will be [0,1] or [1,0] here. \n    # This approach can easily be extended to more classes.\n    image=tf.reshape(image,[height,width,3])\n    label=tf.pack(tf.one_hot(label-1, nClass))\n\n    return label, image\n\n\nlabel, image = getImage(\"train-00000-of-00001\")\nimageBatch, labelBatch = tf.train.shuffle_batch(\n    [image, label], batch_size=2,\n    capacity=20,\n    min_after_dequeue=10)\n\nsess = tf.InteractiveSession()\nsess.run(tf.initialize_all_variables())\n# start the threads used for reading files\ncoord = tf.train.Coordinator()\nthreads = tf.train.start_queue_runners(sess=sess,coord=coord)\nbatch_xs, batch_ys = sess.run([imageBatch, labelBatch])\nprint batch_xs\n\ncoord.request_stop()\ncoord.join(threads)\n<\/code><\/pre>\n\n<p>Note: I wasnt clear about sharded records so I used just one shard.<\/p>\n\n<p>Credits to <a href=\"https:\/\/agray3.github.io\/2016\/11\/29\/Demystifying-Data-Input-to-TensorFlow-for-Deep-Learning.html\" rel=\"nofollow noreferrer\">https:\/\/agray3.github.io\/2016\/11\/29\/Demystifying-Data-Input-to-TensorFlow-for-Deep-Learning.html<\/a><\/p>\n"}]}
{"QuestionId":41957574,"AnswerCount":5,"Tags":"<python><numpy><tensorflow><theano><numba>","CreationDate":"2017-01-31T12:26:24.843","AcceptedAnswerId":42079546.0,"OwnerUserId":2480410.0,"Title":"Python: rewrite a looping numpy math function to run on GPU","Body":"<p>Can someone help me rewrite this one function <em>(the  function)<\/em> to do the calculations on the GPU? I used a few good days now trying to get my head around it but to no result. I wonder maybe somebody can help me rewrite this function in whatever way you may seem fit as log as I gives the same result at the end. I tried to use  from  but for some reason it is actually much slower than running the code as usual. With a huge sample size, the goal is to decrease the execution time considerably so naturally I believe the GPU is the fastest way to do it. <\/p>\n\n<p>I'll explain a little what is actually happening. The real data, which looks almost identical as the sample data created in the code below is divided into sample sizes of approx 5.000.000 rows each sample or around 150MB per file. In total there are around 600.000.000 rows or 20GB of data. I must loop through this data, sample by sample and then row by row in each sample, take the last 2000 (or another) rows as of each line and run the  function which returns a result. That result is then saved back to the hardrive where I can do some other things with it with another program. As you can see below, I do not need all of the results of all the rows, only those bigger than a specific amount. If I run my function as it is right now in python I get about 62seconds per 1.000.000 rows. This is a very long time considering all the data and how fast it should be done with.<\/p>\n\n<p>I must mention that I upload the real data file by file to the RAM with the help of  so uploading the data is not the problem as it takes only about 0.29 seconds per file. Once uploaded I run the entire code below. What takes the longest time is the  function. I am willing to give all of my 500 reputation points I have on stackoverflow as a reward for somebody willing to help me rewrite this simple code to run on the GPU. My interest is specifically in the GPU, I really want to see how it is done on this problem at hand.<\/p>\n\n<p><strong>EDIT\/UPDATE 1:<\/strong>\nHere is a link to a small sample of the real data: <a href=\"https:\/\/mab.to\/Nx8GvwTjQ\" rel=\"noreferrer\">data_csv.zip<\/a> About 102000 rows of real data1 and 2000 rows for real data2a and data2b. Use  on the real sample data <\/p>\n\n<p><strong>EDIT\/UPDATE 2:<\/strong>\nFor those following this post here is a short summary of the answers below. Up until now we have 4 answers to the original solution. The one offered by @Divakar are just tweaks to the original code. Of the two tweaks only the first one is actually applicable to this problem, the second one is a good tweak but does not apply here. Out of the other three answers, two of them are CPU based solutions and one tensorflow-GPU try. The Tensorflow-GPU by Paul Panzer seems to be promising but when i actually run it on the GPU it is slower than the original, so the code still needs improvement.<\/p>\n\n<p>The other two CPU based solutions are submitted by @PaulPanzer (a pure numpy solution) and @MSeifert (a numba solution). Both solutions give very good results and both process data extremely fast compared to the original code. Of the two the one submitted by Paul Panzer is faster. It processes about 1.000.000 rows in about 3 seconds. The only problem is with smaller batchSizes, this can be overcome by either switching to the numba solution offered by MSeifert, or even the original code after all the tweaks that have been discussed below.<\/p>\n\n<p>I am very happy and thankful to @PaulPanzer and @MSeifert for the work they did on their answers. Still, since this is a question about a GPU based solution, i am waiting to see if anybody is willing to give it a try on a GPU version and see how much faster the data can be processed on the GPU when compared to the current CPU solutions. If there will be no other answers outperforming @PaulPanzer's pure numpy solution then i'll accept his answer as the right one and gets the bounty :) <\/p>\n\n<p><strong>EDIT\/UPDATE 3:<\/strong>\n@Divakar has posted a new answer with a solution for the GPU. After my testings on real data, the speed is not even comparable to the CPU counterpart solutions. The GPU processes about 5.000.000 in about 1,5 seconds. This is incredible :) I am very excited about the GPU solution and i thank @Divakar for posting it. As well as i thank @PaulPanzer and @MSeifert for their CPU solutions :) Now my research continues with an incredible speed due to the GPU :) <\/p>\n\n<pre><\/pre>\n\n<p>The PC specs I am working on:<\/p>\n\n<pre><\/pre>\n\n<p>As a side question, would a second video card in SLI help on this problem?<\/p>\n","answers":[{"AnswerId":"42031876","CreationDate":"2017-02-03T19:43:17.750","ParentId":null,"OwnerUserId":"5393381","Title":null,"Body":"<p>Before you start tweaking the target (GPU) or using anything else (i.e. parallel executions ), you might want to consider how to improve the already existing code. You used the <a href=\"\/questions\/tagged\/numba\" class=\"post-tag\" title=\"show questions tagged &#39;numba&#39;\" rel=\"tag\">numba<\/a>-tag so I'll use it to improve the code: First we operate on arrays not on matrices:<\/p>\n\n<pre><code>data1 = np.array(np.random.uniform(1, 100, (sampleSize + batchSize, 4)))\ndata2a = np.array(np.random.uniform(0, 1, batchSize)) #upper limit\ndata2b = np.array(np.random.uniform(0, 1, batchSize)) #lower limit\n<\/code><\/pre>\n\n<p>Each time you call <code>doTheMath<\/code> you expect an integer back, however you use a lot of arrays and create a lot of intermediate arrays:<\/p>\n\n<pre><code>abcd = ((((A  - Cmin) \/ dif) + ((B  - Cmin) \/ dif) + ((C   - Cmin) \/ dif) + ((D - Cmin) \/ dif)) \/ 4)\nreturn np.where(((abcd &lt;= data2a) &amp; (abcd &gt;= data2b)), 1, 0).sum()\n<\/code><\/pre>\n\n<p>This creates an intermediate array each step:<\/p>\n\n<ul>\n<li><code>tmp1 = A-Cmin<\/code>, <\/li>\n<li><code>tmp2 = tmp1 \/ dif<\/code>, <\/li>\n<li><code>tmp3 = B - Cmin<\/code>, <\/li>\n<li><code>tmp4 = tmp3 \/ dif<\/code><\/li>\n<li>... you get the gist.<\/li>\n<\/ul>\n\n<p>However this is a reduce function (array -> integer) so having a lot of intermediate arrays is unnecessary weight, just calculate the value of the \"fly\".<\/p>\n\n<pre><code>import numba as nb\n\n@nb.njit\ndef doTheMathNumba(tmpData, data2a, data2b):\n    Bmax = np.max(tmpData[:, 1])\n    Cmin = np.min(tmpData[:, 2])\n    diff = (Bmax - Cmin)\n    idiff = 1 \/ diff\n    sum_ = 0\n    for i in range(tmpData.shape[0]):\n        val = (tmpData[i, 0] + tmpData[i, 1] + tmpData[i, 2] + tmpData[i, 3]) \/ 4 * idiff - Cmin * idiff\n        if val &lt;= data2a[i] and val &gt;= data2b[i]:\n            sum_ += 1\n    return sum_\n<\/code><\/pre>\n\n<p>I did something else here to avoid multiple operations:<\/p>\n\n<pre><code>(((A - Cmin) \/ dif) + ((B - Cmin) \/ dif) + ((C - Cmin) \/ dif) + ((D - Cmin) \/ dif)) \/ 4\n= ((A - Cmin + B - Cmin + C - Cmin + D - Cmin) \/ dif) \/ 4\n= (A + B + C + D - 4 * Cmin) \/ (4 * dif)\n= (A + B + C + D) \/ (4 * dif) - (Cmin \/ dif)\n<\/code><\/pre>\n\n<p>This actually cuts down the execution time by almost a factor of 10 on my computer:<\/p>\n\n<pre><code>%timeit doTheMath(tmp_df, data2a, data2b)       # 1000 loops, best of 3: 446 \u00b5s per loop\n%timeit doTheMathNumba(tmp_df, data2a, data2b)  # 10000 loops, best of 3: 59 \u00b5s per loop\n<\/code><\/pre>\n\n<p>There are certainly also other improvements, for example using a rolling min\/max to calculate <code>Bmax<\/code> and <code>Cmin<\/code>, that would make at least part of the calculation run in <code>O(sampleSize)<\/code> instead of <code>O(samplesize * batchsize)<\/code>. This would also make it possible to reuse some of the <code>(A + B + C + D) \/ (4 * dif) - (Cmin \/ dif)<\/code> calculations because if <code>Cmin<\/code> and <code>Bmax<\/code> don't change for the next sample these values don't differ. It's a bit complicated to do because the comparisons differ. But definitely possible! See here:<\/p>\n\n<pre><code>import time\nimport numpy as np\nimport numba as nb\n\n@nb.njit\ndef doTheMathNumba(abcd, data2a, data2b, Bmax, Cmin):\n    diff = (Bmax - Cmin)\n    idiff = 1 \/ diff\n    quarter_idiff = 0.25 * idiff\n    sum_ = 0\n    for i in range(abcd.shape[0]):\n        val = abcd[i] * quarter_idiff - Cmin * idiff\n        if val &lt;= data2a[i] and val &gt;= data2b[i]:\n            sum_ += 1\n    return sum_\n\n@nb.njit\ndef doloop(data1, data2a, data2b, abcd, Bmaxs, Cmins, batchSize, sampleSize, minimumLimit, resultArray):\n    found = 0\n    for rowNr in range(data1.shape[0]):\n        if(abcd[rowNr:rowNr + batchSize].shape[0] == batchSize):\n            result = doTheMathNumba(abcd[rowNr:rowNr + batchSize], \n                                    data2a, data2b, Bmaxs[rowNr], Cmins[rowNr])\n            if (result &gt;= minimumLimit):\n                resultArray[found, 0] = rowNr\n                resultArray[found, 1] = result\n                found += 1\n    return resultArray[:found]\n\n#Declare variables\nbatchSize = 2000\nsampleSize = 50000\nresultArray = []\nminimumLimit = 490 #use 400 on the real sample data \n\ndata1 = np.array(np.random.uniform(1, 100, (sampleSize + batchSize, 4)))\ndata2a = np.array(np.random.uniform(0, 1, batchSize)) #upper limit\ndata2b = np.array(np.random.uniform(0, 1, batchSize)) #lower limit\n\nfrom scipy import ndimage\nt0 = time.time()\nabcd = np.sum(data1, axis=1)\nBmaxs = ndimage.maximum_filter1d(data1[:, 1], \n                                 size=batchSize, \n                                 origin=-((batchSize-1)\/\/2-1))  # correction for even shapes\nCmins = ndimage.minimum_filter1d(data1[:, 2], \n                                 size=batchSize, \n                                 origin=-((batchSize-1)\/\/2-1))\n\nresult = np.zeros((sampleSize, 2), dtype=np.int64)\ndoloop(data1, data2a, data2b, abcd, Bmaxs, Cmins, batchSize, sampleSize, minimumLimit, result)\nprint('Runtime:', time.time() - t0)\n<\/code><\/pre>\n\n<p>This gives me a <code>Runtime: 0.759593152999878<\/code> (after numba compiled the functions!), while your original took had <code>Runtime: 24.68975639343262<\/code>. Now we're 30 times faster!<\/p>\n\n<p>With your sample size it still takes <code>Runtime: 60.187848806381226<\/code> but that's not too bad, right?<\/p>\n\n<p>And even if I haven't done this myself, <a href=\"\/questions\/tagged\/numba\" class=\"post-tag\" title=\"show questions tagged &#39;numba&#39;\" rel=\"tag\">numba<\/a> says that it's possible to write <a href=\"http:\/\/numba.pydata.org\/numba-doc\/dev\/cuda\/index.html#numba-for-cuda-gpus\" rel=\"nofollow noreferrer\">\"Numba for CUDA GPUs\"<\/a> and it doesn't seem to complicated.<\/p>\n"},{"AnswerId":"42032640","CreationDate":"2017-02-03T20:37:38.457","ParentId":null,"OwnerUserId":"7207392","Title":null,"Body":"<p>Here is some code to demonstrate what is possible by just tweaking the algorithm. It's pure numpy but on the sample data you posted gives a roughly 35x speedup over the original version (~1,000,000 samples in ~2.5sec on my rather modest machine):<\/p>\n\n<pre><code>&gt;&gt;&gt; result_dict = master('run')\n[('load', 0.82578349113464355), ('precomp', 0.028138399124145508), ('max\/min', 0.24333405494689941), ('ABCD', 0.015314102172851562), ('main', 1.3356468677520752)]\nTOTAL 2.44821691513\n<\/code><\/pre>\n\n<p>Tweaks used:<\/p>\n\n<ul>\n<li>A+B+C+D, see my other answer<\/li>\n<li>running min\/max, including avoiding to compute (A+B+C+D - 4Cmin)\/(4dif) multiple times with the same Cmin\/dif.<\/li>\n<\/ul>\n\n<p>These are more or less routine. That leaves the comparison with data2a\/b which is expensive O(NK) where N is the number of samples and K is the size of the window.\nHere one can take advantage of the relatively well-behaved data. Using the running min\/max one can create variants of data2a\/b that can be used to test a range of window offsets at a time, if the test fails all these offsets can be ruled out immediately, otherwise the range is bisected.<\/p>\n\n<pre><code>import numpy as np\nimport time\n\n# global variables; they will hold the precomputed pre-screening filters\npreA, preB = {}, {}\nCHUNK_SIZES = None\n\ndef sliding_argmax(data, K=2000):\n    \"\"\"compute the argmax of data over a sliding window of width K\n\n    returns:\n        indices  -- indices into data\n        switches -- window offsets at which the maximum changes\n                    (strictly speaking: where the index of the maximum changes)\n                    excludes 0 but includes maximum offset (len(data)-K+1)\n\n    see last line of compute_pre_screening_filter for a recipe to convert\n    this representation to the vector of maxima\n    \"\"\"\n    N = len(data)\n    last = np.argmax(data[:K])\n    indices = [last]\n    while indices[-1] &lt;= N - 1:\n        ge = np.where(data[last + 1 : last + K + 1] &gt; data[last])[0]\n        if len(ge) == 0:\n            if last + K &gt;= N:\n                break\n            last += 1 + np.argmax(data[last + 1 : last + K + 1])\n            indices.append(last)\n        else:\n            last += 1 + ge[0]\n            indices.append(last)\n    indices = np.array(indices)\n    switches = np.where(data[indices[1:]] &gt; data[indices[:-1]],\n                        indices[1:] + (1-K), indices[:-1] + 1)\n    return indices, np.r_[switches, [len(data)-K+1]]\n\n\ndef compute_pre_screening_filter(bound, n_offs):\n    \"\"\"compute pre-screening filter for point-wise upper bound\n\n    given a K-vector of upper bounds B and K+n_offs-1-vector data\n    compute K+n_offs-1-vector filter such that for each index j\n    if for any offset 0 &lt;= o &lt; n_offs and index 0 &lt;= i &lt; K such that\n    o + i = j, the inequality B_i &gt;= data_j holds then filter_j &gt;= data_j\n\n    therefore the number of data points below filter is an upper bound for\n    the maximum number of points below bound in any K-window in data\n    \"\"\"\n    pad_l, pad_r = np.min(bound[:n_offs-1]), np.min(bound[1-n_offs:])\n    padded = np.r_[pad_l+np.zeros(n_offs-1,), bound, pad_r+np.zeros(n_offs-1,)]\n    indices, switches = sliding_argmax(padded, n_offs)\n    return padded[indices].repeat(np.diff(np.r_[[0], switches]))\n\n\ndef compute_all_pre_screening_filters(upper, lower, min_chnk=5, dyads=6):\n    \"\"\"compute upper and lower pre-screening filters for data blocks of\n    sizes K+n_offs-1 where\n    n_offs = min_chnk, 2min_chnk, ..., 2^(dyads-1)min_chnk\n\n    the result is stored in global variables preA and preB\n    \"\"\"\n    global CHUNK_SIZES\n\n    CHUNK_SIZES = min_chnk * 2**np.arange(dyads)\n    preA[1] = upper\n    preB[1] = lower\n    for n in CHUNK_SIZES:\n        preA[n] = compute_pre_screening_filter(upper, n)\n        preB[n] = -compute_pre_screening_filter(-lower, n)\n\n\ndef test_bounds(block, counts, threshold=400):\n    \"\"\"test whether the windows fitting in the data block 'block' fall\n    within the bounds using pre-screening for efficient bulk rejection\n\n    array 'counts' will be overwritten with the counts of compliant samples\n    note that accurate counts will only be returned for above threshold\n    windows, because the analysis of bulk rejected windows is short-circuited\n\n    also note that bulk rejection only works for 'well behaved' data and\n    for example not on random numbers\n    \"\"\"\n    N = len(counts)\n    K = len(preA[1])\n    r = N % CHUNK_SIZES[0]\n    # chop up N into as large as possible chunks with matching pre computed\n    # filters\n    # start with small and work upwards\n    counts[:r] = [np.count_nonzero((block[l:l+K] &lt;= preA[1]) &amp;\n                                   (block[l:l+K] &gt;= preB[1]))\n                  for l in range(r)]\n\n    def bisect(block, counts):\n        M = len(counts)\n        cnts = np.count_nonzero((block &lt;= preA[M]) &amp; (block &gt;= preB[M]))\n        if cnts &lt; threshold:\n            counts[:] = cnts\n            return\n        elif M == CHUNK_SIZES[0]:\n            counts[:] = [np.count_nonzero((block[l:l+K] &lt;= preA[1]) &amp;\n                                          (block[l:l+K] &gt;= preB[1]))\n                         for l in range(M)]\n        else:\n            M \/\/= 2\n            bisect(block[:-M], counts[:M])\n            bisect(block[M:], counts[M:])\n\n    N = N \/\/ CHUNK_SIZES[0]\n    for M in CHUNK_SIZES:\n        if N % 2:\n            bisect(block[r:r+M+K-1], counts[r:r+M])\n            r += M\n        elif N == 0:\n            return\n        N \/\/= 2\n    else:\n        for j in range(2*N):\n            bisect(block[r:r+M+K-1], counts[r:r+M])\n            r += M\n\n\ndef analyse(data, use_pre_screening=True, min_chnk=5, dyads=6,\n            threshold=400):\n    samples, upper, lower = data\n    N, K = samples.shape[0], upper.shape[0]\n    times = [time.time()]\n    if use_pre_screening:\n        compute_all_pre_screening_filters(upper, lower, min_chnk, dyads)\n    times.append(time.time())\n    # compute switching points of max and min for running normalisation\n    upper_inds, upper_swp = sliding_argmax(samples[:, 1], K)\n    lower_inds, lower_swp = sliding_argmax(-samples[:, 2], K)\n    times.append(time.time())\n    # sum columns\n    ABCD = samples.sum(axis=-1)\n    times.append(time.time())\n    counts = np.empty((N-K+1,), dtype=int)\n    # main loop\n    # loop variables:\n    offs = 0\n    u_ind, u_scale, u_swp = 0, samples[upper_inds[0], 1], upper_swp[0]\n    l_ind, l_scale, l_swp = 0, samples[lower_inds[0], 2], lower_swp[0]\n    while True:\n        # check which is switching next, min(C) or max(B)\n        if u_swp &gt; l_swp:\n            # greedily take the largest block possible such that dif and Cmin\n            # do not change\n            block = (ABCD[offs:l_swp+K-1] - 4*l_scale) \\\n                    * (0.25 \/ (u_scale-l_scale))\n            if use_pre_screening:\n                test_bounds(block, counts[offs:l_swp], threshold=threshold)\n            else:\n                counts[offs:l_swp] = [\n                    np.count_nonzero((block[l:l+K] &lt;= upper) &amp;\n                                     (block[l:l+K] &gt;= lower))\n                    for l in range(l_swp - offs)]\n            # book keeping\n            l_ind += 1\n            offs = l_swp\n            l_swp = lower_swp[l_ind]\n            l_scale = samples[lower_inds[l_ind], 2]\n        else:\n            block = (ABCD[offs:u_swp+K-1] - 4*l_scale) \\\n                    * (0.25 \/ (u_scale-l_scale))\n            if use_pre_screening:\n                test_bounds(block, counts[offs:u_swp], threshold=threshold)\n            else:\n                counts[offs:u_swp] = [\n                    np.count_nonzero((block[l:l+K] &lt;= upper) &amp;\n                                     (block[l:l+K] &gt;= lower))\n                    for l in range(u_swp - offs)]\n            u_ind += 1\n            if u_ind == len(upper_inds):\n                assert u_swp == N-K+1\n                break\n            offs = u_swp\n            u_swp = upper_swp[u_ind]\n            u_scale = samples[upper_inds[u_ind], 1]\n    times.append(time.time())\n    return {'counts': counts, 'valid': np.where(counts &gt;= 400)[0],\n            'timings': np.diff(times)}\n\n\ndef master(mode='calibrate', data='fake', use_pre_screening=True, nrep=3,\n           min_chnk=None, dyads=None):\n    t = time.time()\n    if data in ('fake', 'load'):\n        data1 = np.loadtxt('data1.csv', delimiter=';', skiprows=1,\n                           usecols=[1,2,3,4])\n        data2a = np.loadtxt('data2a.csv', delimiter=';', skiprows=1,\n                            usecols=[1])\n        data2b = np.loadtxt('data2b.csv', delimiter=';', skiprows=1,\n                            usecols=[1])\n        if data == 'fake':\n            data1 = np.tile(data1, (10, 1))\n        threshold = 400\n    elif data == 'random':\n        data1 = np.random.random((102000, 4))\n        data2b = np.random.random(2000)\n        data2a = np.random.random(2000)\n        threshold = 490\n        if use_pre_screening or mode == 'calibrate':\n            print('WARNING: pre-screening not efficient on artificial data')\n    else:\n        raise ValueError(\"data mode {} not recognised\".format(data))\n    data = data1, data2a, data2b\n    t_load = time.time() - t\n    if mode == 'calibrate':\n        min_chnk = (2, 3, 4, 5, 6) if min_chnk is None else min_chnk\n        dyads = (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10) if dyads is None else dyads\n        timings = np.zeros((len(min_chnk), len(dyads)))\n        print('max bisect  ' + ' '.join([\n            '   n.a.' if dy == 0 else '{:7d}'.format(dy) for dy in dyads]),\n              end='')\n        for i, mc in enumerate(min_chnk):\n            print('\\nmin chunk {}'.format(mc), end=' ')\n            for j, dy in enumerate(dyads):\n                for k in range(nrep):\n                    if dy == 0: # no pre-screening\n                        timings[i, j] += analyse(\n                            data, False, mc, dy, threshold)['timings'][3]\n                    else:\n                        timings[i, j] += analyse(\n                            data, True, mc, dy, threshold)['timings'][3]\n                timings[i, j] \/= nrep\n                print('{:7.3f}'.format(timings[i, j]), end=' ', flush=True)\n        best_mc, best_dy = np.unravel_index(np.argmin(timings.ravel()),\n                                            timings.shape)\n        print('\\nbest', min_chnk[best_mc], dyads[best_dy])\n        return timings, min_chnk[best_mc], dyads[best_dy]\n    if mode == 'run':\n        min_chnk = 2 if min_chnk is None else min_chnk\n        dyads = 5 if dyads is None else dyads\n        res = analyse(data, use_pre_screening, min_chnk, dyads, threshold)\n        times = np.r_[[t_load], res['timings']]\n        print(list(zip(('load', 'precomp', 'max\/min', 'ABCD', 'main'), times)))\n        print('TOTAL', times.sum())\n        return res\n<\/code><\/pre>\n"},{"AnswerId":"42079546","CreationDate":"2017-02-07T00:08:41.523","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p><strong>Introduction and solution code<\/strong><\/p>\n\n<p>Well, you asked for it! So, listed in this post is an implementation with <a href=\"https:\/\/documen.tician.de\/pycuda\/\" rel=\"noreferrer\"><strong><code>PyCUDA<\/code><\/strong><\/a> that uses lightweight wrappers extending most of CUDA's capabilities within Python environment. We will its <code>SourceModule<\/code> functionality that lets us write and compile CUDA kernels staying in Python environment.<\/p>\n\n<p>Getting to the problem at hand, among the computations involved, we have sliding maximum and minimum, few differences and divisions and comparisons. For the maximum and minimum parts, that involves block max finding (for each sliding window), we will use reduction-technique  as discussed in some detail <a href=\"http:\/\/developer.download.nvidia.com\/compute\/cuda\/1.1-Beta\/x86_website\/projects\/reduction\/doc\/reduction.pdf\" rel=\"noreferrer\"><code>here<\/code><\/a>. This would be done at block level. For the upper level iterations across sliding windows, we would use the grid level indexing into CUDA resources. For more info on this block and grid format, please refer to <a href=\"http:\/\/gpgpu.org\/wp\/wp-content\/uploads\/2009\/06\/02-CUDA_basic.pdf\" rel=\"noreferrer\"><code>page-18<\/code><\/a>. PyCUDA also supports builtins for computing reductions like max and min, but we lose control, specifically we intend to use specialized memory like shared and constant memory for leveraging GPU at its near to optimum level.<\/p>\n\n<p>Listing out the PyCUDA-NumPy solution code -<\/p>\n\n<p>1] PyCUDA part -<\/p>\n\n<pre><code>import pycuda.autoinit\nimport pycuda.driver as drv\nimport numpy as np\nfrom pycuda.compiler import SourceModule\n\nmod = SourceModule(\"\"\"\n#define TBP 1024 \/\/ THREADS_PER_BLOCK\n\n__device__ void get_Bmax_Cmin(float* out, float *d1, float *d2, int L, int offset)\n{\n    int tid = threadIdx.x;\n    int inv = TBP;\n    __shared__ float dS[TBP][2];\n\n    dS[tid][0] = d1[tid+offset];  \n    dS[tid][1] = d2[tid+offset];         \n    __syncthreads();\n\n    if(tid&lt;L-TBP)  \n    {\n        dS[tid][0] = fmaxf(dS[tid][0] , d1[tid+inv+offset]);\n        dS[tid][1] = fminf(dS[tid][1] , d2[tid+inv+offset]);\n    }\n    __syncthreads();\n    inv = inv\/2;\n\n    while(inv!=0)   \n    {\n        if(tid&lt;inv)\n        {\n            dS[tid][0] = fmaxf(dS[tid][0] , dS[tid+inv][0]);\n            dS[tid][1] = fminf(dS[tid][1] , dS[tid+inv][1]);\n        }\n        __syncthreads();\n        inv = inv\/2;\n    }\n    __syncthreads();\n\n    if(tid==0)\n    {\n        out[0] = dS[0][0];\n        out[1] = dS[0][1];\n    }   \n    __syncthreads();\n}\n\n__global__ void main1(float* out, float *d0, float *d1, float *d2, float *d3, float *lowL, float *highL, int *BLOCKLEN)\n{\n    int L = BLOCKLEN[0];\n    int tid = threadIdx.x;\n    int iterID = blockIdx.x;\n    float Bmax_Cmin[2];\n    int inv;\n    float Cmin, dif;   \n    __shared__ float dS[TBP*2];   \n\n    get_Bmax_Cmin(Bmax_Cmin, d1, d2, L, iterID);  \n    Cmin = Bmax_Cmin[1];\n    dif = (Bmax_Cmin[0] - Cmin);\n\n    inv = TBP;\n\n    dS[tid] = (d0[tid+iterID] + d1[tid+iterID] + d2[tid+iterID] + d3[tid+iterID] - 4.0*Cmin) \/ (4.0*dif);\n    __syncthreads();\n\n    if(tid&lt;L-TBP)  \n        dS[tid+inv] = (d0[tid+inv+iterID] + d1[tid+inv+iterID] + d2[tid+inv+iterID] + d3[tid+inv+iterID] - 4.0*Cmin) \/ (4.0*dif);                   \n\n     dS[tid] = ((dS[tid] &gt;= lowL[tid]) &amp; (dS[tid] &lt;= highL[tid])) ? 1 : 0;\n     __syncthreads();\n\n     if(tid&lt;L-TBP)\n         dS[tid] += ((dS[tid+inv] &gt;= lowL[tid+inv]) &amp; (dS[tid+inv] &lt;= highL[tid+inv])) ? 1 : 0;\n     __syncthreads();\n\n    inv = inv\/2;\n    while(inv!=0)   \n    {\n        if(tid&lt;inv)\n            dS[tid] += dS[tid+inv];\n        __syncthreads();\n        inv = inv\/2;\n    }\n\n    if(tid==0)\n        out[iterID] = dS[0];\n    __syncthreads();\n\n}\n\"\"\")\n<\/code><\/pre>\n\n<p>Please note that <code>THREADS_PER_BLOCK, TBP<\/code> is to be set based on the <code>batchSize<\/code>. The rule of thumb here is to assign power of 2 value to <code>TBP<\/code> that is just lesser than <code>batchSize<\/code>. Thus, for <code>batchSize = 2000<\/code>, we needed <code>TBP<\/code> as <code>1024<\/code>.<\/p>\n\n<p>2] NumPy part -<\/p>\n\n<pre><code>def gpu_app_v1(A, B, C, D, batchSize, minimumLimit):\n    func1 = mod.get_function(\"main1\")\n    outlen = len(A)-batchSize+1\n\n    # Set block and grid sizes\n    BSZ = (1024,1,1)\n    GSZ = (outlen,1)\n\n    dest = np.zeros(outlen).astype(np.float32)\n    N = np.int32(batchSize)\n    func1(drv.Out(dest), drv.In(A), drv.In(B), drv.In(C), drv.In(D), \\\n                     drv.In(data2b), drv.In(data2a),\\\n                     drv.In(N), block=BSZ, grid=GSZ)\n    idx = np.flatnonzero(dest &gt;= minimumLimit)\n    return idx, dest[idx]\n<\/code><\/pre>\n\n<p><strong>Benchmarking<\/strong><\/p>\n\n<p>I have tested on GTX 960M. Please note that PyCUDA expects arrays to be of contiguous order. So, we need to slice the columns and make copies. I am expecting\/assuming that the data could be read from the files such that the data is spread along rows instead of being as columns. Thus, keeping those out of the benchmarking function for now.<\/p>\n\n<p>Original approach -<\/p>\n\n<pre><code>def org_app(data1, batchSize, minimumLimit):\n    resultArray = []\n    for rowNr in  range(data1.shape[0]-batchSize+1):\n        tmp_df = data1[rowNr:rowNr + batchSize] #rolling window\n        result = doTheMath(tmp_df, data2a, data2b)\n        if (result &gt;= minimumLimit):\n            resultArray.append([rowNr , result]) \n    return resultArray\n<\/code><\/pre>\n\n<p>Timings and verification -<\/p>\n\n<pre><code>In [2]: #Declare variables\n   ...: batchSize = 2000\n   ...: sampleSize = 50000\n   ...: resultArray = []\n   ...: minimumLimit = 490 #use 400 on the real sample data\n   ...: \n   ...: #Create Random Sample Data\n   ...: data1 = np.random.uniform(1, 100000, (sampleSize + batchSize, 4)).astype(np.float32)\n   ...: data2b = np.random.uniform(0, 1, (batchSize)).astype(np.float32)\n   ...: data2a = data2b + np.random.uniform(0, 1, (batchSize)).astype(np.float32)\n   ...: \n   ...: # Make column copies\n   ...: A = data1[:,0].copy()\n   ...: B = data1[:,1].copy()\n   ...: C = data1[:,2].copy()\n   ...: D = data1[:,3].copy()\n   ...: \n   ...: gpu_out1,gpu_out2 = gpu_app_v1(A, B, C, D, batchSize, minimumLimit)\n   ...: cpu_out1,cpu_out2 = np.array(org_app(data1, batchSize, minimumLimit)).T\n   ...: print(np.allclose(gpu_out1, cpu_out1))\n   ...: print(np.allclose(gpu_out2, cpu_out2))\n   ...: \nTrue\nFalse\n<\/code><\/pre>\n\n<p>So, there's some differences between CPU and GPU countings. Let's investigate them -<\/p>\n\n<pre><code>In [7]: idx = np.flatnonzero(~np.isclose(gpu_out2, cpu_out2))\n\nIn [8]: idx\nOut[8]: array([12776, 15208, 17620, 18326])\n\nIn [9]: gpu_out2[idx] - cpu_out2[idx]\nOut[9]: array([-1., -1.,  1.,  1.])\n<\/code><\/pre>\n\n<p>There are four instances of non-matching counts. These are off at max by <code>1<\/code>.  Upon research, I came across some information on this. Basically, since we are using math intrinsics for max and min computations and those I think are causing the last binary bit in the floating pt representation to be diferent than the CPU counterpart. This is termed as ULP error and has been discused in detail <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/784309\/why-accuracy-cpu-and-gpu-not-equal-\/\" rel=\"noreferrer\"><code>here<\/code><\/a> and <a href=\"http:\/\/developer.download.nvidia.com\/assets\/cuda\/files\/NVIDIA-CUDA-Floating-Point.pdf\" rel=\"noreferrer\"><code>here<\/code><\/a>.<\/p>\n\n<p>Finally, puting the issue aside, let's get to the most important bit, the performance -<\/p>\n\n<pre><code>In [10]: %timeit org_app(data1, batchSize, minimumLimit)\n1 loops, best of 3: 2.18 s per loop\n\nIn [11]: %timeit gpu_app_v1(A, B, C, D, batchSize, minimumLimit)\n10 loops, best of 3: 82.5 ms per loop\n\nIn [12]: 2180.0\/82.5\nOut[12]: 26.424242424242426\n<\/code><\/pre>\n\n<p>Let's try with bigger datasets. With <code>sampleSize = 500000<\/code>, we get -<\/p>\n\n<pre><code>In [14]: %timeit org_app(data1, batchSize, minimumLimit)\n1 loops, best of 3: 23.2 s per loop\n\nIn [15]: %timeit gpu_app_v1(A, B, C, D, batchSize, minimumLimit)\n1 loops, best of 3: 821 ms per loop\n\nIn [16]: 23200.0\/821\nOut[16]: 28.25822168087698\n<\/code><\/pre>\n\n<p>So, the speedup stays constant at around <strong><code>27<\/code><\/strong>.<\/p>\n\n<p><strong>Limitations :<\/strong><\/p>\n\n<p>1) We are using <code>float32<\/code> numbers, as GPUs work best with those. Double precision specially on non-server GPUs aren't popular when it comes to performance and since you are working with such a GPU, I tested with float32.<\/p>\n\n<p><strong>Further improvement :<\/strong><\/p>\n\n<p>1) We could use faster <code>constant memory<\/code> to feed in <code>data2a<\/code> and <code>data2b<\/code>, rather than use <code>global memory<\/code>.<\/p>\n"},{"AnswerId":"41959759","CreationDate":"2017-01-31T14:13:26.943","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p><strong>Tweak #1<\/strong><\/p>\n\n<p>Its usually advised to vectorize things when working with NumPy arrays. But with very large arrays, I think you are out of options there. So, to boost performance, a minor tweak is possible to optimize on the last step of summing.<\/p>\n\n<p>We could replace the step that makes an array of <code>1s<\/code> and <code>0s<\/code> and does summing :<\/p>\n\n<pre><code>np.where(((abcd &lt;= data2a) &amp; (abcd &gt;= data2b)), 1, 0).sum()\n<\/code><\/pre>\n\n<p>with <code>np.count_nonzero<\/code> that works efficiently to count <code>True<\/code> values in a boolean array, instead of converting to <code>1s<\/code> and <code>0s<\/code> -<\/p>\n\n<pre><code>np.count_nonzero((abcd &lt;= data2a) &amp; (abcd &gt;= data2b))\n<\/code><\/pre>\n\n<p>Runtime test -<\/p>\n\n<pre><code>In [45]: abcd = np.random.randint(11,99,(10000))\n\nIn [46]: data2a = np.random.randint(11,99,(10000))\n\nIn [47]: data2b = np.random.randint(11,99,(10000))\n\nIn [48]: %timeit np.where(((abcd &lt;= data2a) &amp; (abcd &gt;= data2b)), 1, 0).sum()\n10000 loops, best of 3: 81.8 \u00b5s per loop\n\nIn [49]: %timeit np.count_nonzero((abcd &lt;= data2a) &amp; (abcd &gt;= data2b))\n10000 loops, best of 3: 28.8 \u00b5s per loop\n<\/code><\/pre>\n\n<p><strong>Tweak #2<\/strong><\/p>\n\n<p>Use a pre-computed reciprocal when dealing with cases that undergo implicit broadcasting. Some more info <a href=\"https:\/\/stackoverflow.com\/a\/41945198\/3293881\"><code>here<\/code><\/a>. Thus, store reciprocal of <code>dif<\/code> and use that instead at the step :<\/p>\n\n<pre><code>((((A  - Cmin) \/ dif) + ((B  - Cmin) \/ dif) + ...\n<\/code><\/pre>\n\n<p>Sample test -<\/p>\n\n<pre><code>In [52]: A = np.random.rand(10000)\n\nIn [53]: dif = 0.5\n\nIn [54]: %timeit A\/dif\n10000 loops, best of 3: 25.8 \u00b5s per loop\n\nIn [55]: %timeit A*(1.0\/dif)\n100000 loops, best of 3: 7.94 \u00b5s per loop\n<\/code><\/pre>\n\n<p>You have four places using division by <code>dif<\/code>. So, hopefully this would bring out noticeable boost there too!<\/p>\n"},{"AnswerId":"42016258","CreationDate":"2017-02-03T03:27:18.120","ParentId":null,"OwnerUserId":"7207392","Title":null,"Body":"<p><del>This is technically off-topic (not GPU) but I'm sure you'll be interested.<\/del><\/p>\n\n<p>There is one obvious and rather large saving:<\/p>\n\n<p>Precompute <code>A + B + C + D<\/code> (not in the loop, on the whole data: <code>data1.sum(axis=-1)<\/code>), because <code>abcd = ((A+B+C+D) - 4Cmin) \/ (4dif)<\/code>. This will save quite a few ops.<\/p>\n\n<p>Surprised nobody spotted that one before ;-)<\/p>\n\n<p>Edit:<\/p>\n\n<p>There is another thing, though I suspect that's only in your example, not in your real data:<\/p>\n\n<p>As it stands roughly half of <code>data2a<\/code> will be smaller than <code>data2b<\/code>. In these places your conditions on <code>abcd<\/code> cannot be both True, so you needn't even compute <code>abcd<\/code> there.<\/p>\n\n<p>Edit:<\/p>\n\n<p>One more tweak I used below but forgot to mention: If you compute the max (or min) over a moving window. When you move one point to the right, say, how likely is the max to change? There are only two things that can change it: the new point on the right is larger (happens roughly once in windowlength times, and even if it happens, you immediately know the new max) or the old max falls off the window\n on the left (also happens roughly once in windowlength times). Only in this last case you have to search the entire window for the new max.<\/p>\n\n<p>Edit:<\/p>\n\n<p>Couldn't resist giving it a try in tensorflow. I don't have a GPU, so you yourself have to test it for speed. Put \"gpu\" for \"cpu\" on the marked line.<\/p>\n\n<p>On cpu it is about half as fast as your original implementation (i.e. without Divakar's tweaks). Note that I've taken the liberty of changing the inputs from matrix to plain array. Currently tensorflow is a bit of a moving target, so make sure you have the right version. I used Python3.6 and tf 0.12.1 If you do a pip3 install tensorflow-gpu today it <del>should<\/del> might work.<\/p>\n\n<pre><code>import numpy as np\nimport time\nimport tensorflow as tf\n\n# currently the max\/min code is sequential\n# thus\nparallel_iterations = 1\n# but you can put this in a separate loop, precompute and then try and run\n# the remainder of doTheMathTF with a larger parallel_iterations\n\n# tensorflow is quite capricious about its data types\nddf = tf.float64\nddi = tf.int32\n\ndef worker(data1, data2a, data2b):\n    ###################################\n    # CHANGE cpu to gpu in next line! #\n    ###################################\n    with tf.device('\/cpu:0'):\n        g = tf.Graph ()\n        with g.as_default():\n            ABCD = tf.constant(data1.sum(axis=-1), dtype=ddf)\n            B = tf.constant(data1[:, 1], dtype=ddf)\n            C = tf.constant(data1[:, 2], dtype=ddf)\n            window = tf.constant(len(data2a))\n            N = tf.constant(data1.shape[0] - len(data2a) + 1, dtype=ddi)\n            data2a = tf.constant(data2a, dtype=ddf)\n            data2b = tf.constant(data2b, dtype=ddf)\n            def doTheMathTF(i, Bmax, Bmaxind, Cmin, Cminind, out):\n                # most of the time we can keep the old max\/min\n                Bmaxind = tf.cond(Bmaxind&lt;i,\n                                  lambda: i + tf.to_int32(\n                                      tf.argmax(B[i:i+window], axis=0)),\n                                  lambda: tf.cond(Bmax&gt;B[i+window-1], \n                                                  lambda: Bmaxind, \n                                                  lambda: i+window-1))\n                Cminind = tf.cond(Cminind&lt;i,\n                                  lambda: i + tf.to_int32(\n                                      tf.argmin(C[i:i+window], axis=0)),\n                                  lambda: tf.cond(Cmin&lt;C[i+window-1],\n                                                  lambda: Cminind,\n                                                  lambda: i+window-1))\n                Bmax = B[Bmaxind]\n                Cmin = C[Cminind]\n                abcd = (ABCD[i:i+window] - 4 * Cmin) * (1 \/ (4 * (Bmax-Cmin)))\n                out = out.write(i, tf.to_int32(\n                    tf.count_nonzero(tf.logical_and(abcd &lt;= data2a,\n                                                    abcd &gt;= data2b))))\n                return i + 1, Bmax, Bmaxind, Cmin, Cminind, out\n            with tf.Session(graph=g) as sess:\n                i, Bmaxind, Bmax, Cminind, Cmin, out = tf.while_loop(\n                    lambda i, _1, _2, _3, _4, _5: i&lt;N, doTheMathTF,\n                    (tf.Variable(0, dtype=ddi), tf.Variable(0.0, dtype=ddf),\n                     tf.Variable(-1, dtype=ddi),\n                     tf.Variable(0.0, dtype=ddf), tf.Variable(-1, dtype=ddi),\n                     tf.TensorArray(ddi, size=N)),\n                    shape_invariants=None,\n                    parallel_iterations=parallel_iterations,\n                    back_prop=False)\n                out = out.pack()\n                sess.run(tf.initialize_all_variables())\n                out, = sess.run((out,))\n    return out\n\n#Declare variables\nbatchSize = 2000\nsampleSize = 50000#00\nresultArray = []\n\n#Create Sample Data\ndata1 = np.random.uniform(1, 100, (sampleSize + batchSize, 4))\ndata2a = np.random.uniform(0, 1, (batchSize,))\ndata2b = np.random.uniform(0, 1, (batchSize,))\n\nt0 = time.time()\nout = worker(data1, data2a, data2b)\nprint('Runtime (tensorflow):', time.time() - t0)\n\n\ngood_indices, = np.where(out &gt;= 490)\nres_tf = np.c_[good_indices, out[good_indices]]\n\ndef doTheMath(tmpData1, data2a, data2b):\n    A = tmpData1[:, 0]\n    B  = tmpData1[:,1]\n    C   = tmpData1[:,2]\n    D = tmpData1[:,3]\n    Bmax = B.max()\n    Cmin  = C.min()\n    dif = (Bmax - Cmin)\n    abcd = ((((A  - Cmin) \/ dif) + ((B  - Cmin) \/ dif) + ((C   - Cmin) \/ dif) + ((D - Cmin) \/ dif)) \/ 4)\n    return np.where(((abcd &lt;= data2a) &amp; (abcd &gt;= data2b)), 1, 0).sum()\n\n#Loop through the data\nt0 = time.time()\nfor rowNr in  range(sampleSize+1):\n    tmp_df = data1[rowNr:rowNr + batchSize] #rolling window\n    result = doTheMath(tmp_df, data2a, data2b)\n    if (result &gt;= 490):\n        resultArray.append([rowNr , result])\nprint('Runtime (original):', time.time() - t0)\nprint(np.alltrue(np.array(resultArray)==res_tf))\n<\/code><\/pre>\n"}]}
{"QuestionId":41958096,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-01-31T12:51:26.493","AcceptedAnswerId":null,"OwnerUserId":775948.0,"Title":"Cannot load checkpoint file","Body":"<p>I am running a python script on two tensorflow models. The script just read the checkpoint file and prints its content. I'm using this script to validate that the checkpoint file is valid.\nThe script:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/inspect_checkpoint.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/inspect_checkpoint.py<\/a><\/p>\n\n<p>Model 1 - facenet: <a href=\"https:\/\/drive.google.com\/file\/d\/0B5MzpY9kBtDVSTgxX25ZQzNTMGc\/view\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/0B5MzpY9kBtDVSTgxX25ZQzNTMGc\/view<\/a><\/p>\n\n<p>model 2:  <a href=\"http:\/\/download.tensorflow.org\/models\/inception_v1_2016_08_28.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/inception_v1_2016_08_28.tar.gz<\/a><\/p>\n\n<p>on desktop Ubuntu 14.4, docker on Ubuntu 14.4, docker on OSX - it works perfect.\non ubuntu 16.4 on AWS and on docker on ubuntu 16.4 on AWS - it doesn't.<\/p>\n\n<p>assuming the docker eliminates the os version, I afraid the error related to the fact that it runs on AWS.<\/p>\n\n<p>the error:<\/p>\n\n<pre><\/pre>\n\n<p>I wonder if there is any work-around for that.\nI'm using tensorflow 0.12.1<\/p>\n","answers":[]}
{"QuestionId":41958115,"AnswerCount":1,"Tags":"<machine-learning><neural-network><nlp><keras><conv-neural-network>","CreationDate":"2017-01-31T12:52:19.257","AcceptedAnswerId":41961452.0,"OwnerUserId":5368083.0,"Title":"Max over time pooling in Keras","Body":"<p>I'm using CNNs in  for an NLP task and instead of max pooling, I'm trying to achieve max over time pooling.<\/p>\n\n<p>Any ideas\/hacks on how to achieve this?<\/p>\n\n<p>What I mean by max over time pooling is to pool the highest value, no matter where they are in the vector<\/p>\n","answers":[{"AnswerId":"41961452","CreationDate":"2017-01-31T15:35:53.227","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Assuming that your data shape is <code>(batch_size, seq_len, features)<\/code> you may apply:<\/p>\n\n<pre><code>seq_model = Reshape((seq_len * features, 1))(seq_model)\nseq_model = GlobalMaxPooling1D()(seq_model)\n<\/code><\/pre>\n"}]}
{"QuestionId":41958566,"AnswerCount":4,"Tags":"<python-3.x><neural-network><keras><pruning>","CreationDate":"2017-01-31T13:14:12.917","AcceptedAnswerId":41970980.0,"OwnerUserId":1901114.0,"Title":"Pruning in Keras","Body":"<p>I'm trying to design a neural network using Keras with priority on prediction performance, and I cannot get sufficiently high accuracy by further reducing the number of layers and nodes per layer. I have noticed that very large portion of my weights are effectively zero (>95%). Is there a way to prune dense layers in hope of reducing prediction time?<\/p>\n","answers":[{"AnswerId":"56453628","CreationDate":"2019-06-05T02:40:37.170","ParentId":null,"OwnerUserId":"9867252","Title":null,"Body":"<p>See this dedicated tooling for Keras. <a href=\"https:\/\/www.tensorflow.org\/model_optimization\/guide\/pruning\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/model_optimization\/guide\/pruning<\/a><\/p>\n\n<p>As the overview suggests, support for latency improvements is a work in progress<\/p>\n"},{"AnswerId":"41970980","CreationDate":"2017-02-01T02:31:51.707","ParentId":null,"OwnerUserId":"3991470","Title":null,"Body":"<p>Not a dedicated way :(<\/p>\n\n<p>There's currently no easy (dedicated) way of doing this with Keras.<\/p>\n\n<p>A discussion is ongoing at <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/oEecCWayJrM\" rel=\"noreferrer\">https:\/\/groups.google.com\/forum\/#!topic\/keras-users\/oEecCWayJrM<\/a>.<\/p>\n\n<p>You may also be interested in this paper: <a href=\"https:\/\/arxiv.org\/pdf\/1608.04493v1.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1608.04493v1.pdf<\/a>.<\/p>\n"},{"AnswerId":"51681226","CreationDate":"2018-08-03T23:39:47.873","ParentId":null,"OwnerUserId":"3880948","Title":null,"Body":"<p>Take a look at Keras Surgeon:\n<a href=\"https:\/\/github.com\/BenWhetton\/keras-surgeon\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BenWhetton\/keras-surgeon<\/a><\/p>\n\n<p>I have not tried it myself, but the documentation claims that it has functions to remove or insert nodes.<\/p>\n\n<p>Also, after looking at some papers on pruning, it seems that many researchers create a new model with less channels (or less layers), and then copy the weights from the original model to the new model.<\/p>\n"},{"AnswerId":"56069261","CreationDate":"2019-05-10T00:39:23.710","ParentId":null,"OwnerUserId":"623735","Title":null,"Body":"<p>If you set an individual weight to zero won't that prevent it from being updated  during back propagation? Shouldn't thatv weight remain zero from one epoch to the next? That's why you set the initial weights to nonzero values before training. If you want to \"remove\" an entire node, just set all of the weights on that node's output to zero and that will prevent that nodes from having any affect on the output throughout training.<\/p>\n"}]}
{"QuestionId":41958921,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-31T13:32:32.850","AcceptedAnswerId":41964549.0,"OwnerUserId":7460675.0,"Title":"Tensorflow graph editor reroute complex network","Body":"<p>I try to wrap operation with customized operation.<br>\nI solved input of target operation (A in picture) but fail with wrapping output.<\/p>\n\n<p>Init network operations looks like it.<\/p>\n\n<pre><\/pre>\n\n<p>and assume every operation has 1 output tensor. I want to add operation 'E'<\/p>\n\n<pre><\/pre>\n\n<p>I'm tried with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.graph_editor\/\" rel=\"nofollow noreferrer\">graph editor(API r0.12)<\/a>,<br>\nmy strategy was that connect A-E and disconnect (C,D)-A and connect (C,D)-E <\/p>\n\n<ol>\n<li><strong>ge.connect([E], [C,D])<\/strong> - fail because op C has 2 input<\/li>\n<li>use <strong>ge.swap_inputs<\/strong> 2 times - fail, I can attach E and D but still problem with disconnect C-A and connect C-E<\/li>\n<\/ol>\n\n<p>How can I change?<br>\nThanks<\/p>\n","answers":[{"AnswerId":"41964549","CreationDate":"2017-01-31T18:08:41.937","ParentId":null,"OwnerUserId":"1252925","Title":null,"Body":"<p>You need to do some subgraph <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.graph_editor\/module_subgraph#SubGraphView.remap_outputs\" rel=\"noreferrer\">remapping<\/a> to make sure the signatures of the two subgraphs match. To do so it's helpful to print the subgraph.<\/p>\n\n<pre><code>tf.reset_default_graph()\n\na = tf.placeholder(dtype=tf.float32, name=\"a\")\nb = tf.placeholder(dtype=tf.float32, name=\"b\")\nc = tf.add(a, b, name=\"c\")\nd = tf.identity(a, name=\"d\")\n\ne = tf.identity(a, name=\"e\")\nprint(ge.sgv(e.op))\nprint(ge.sgv(c.op, d.op).remap_inputs([0]))\nge.connect(ge.sgv(e.op), ge.sgv(c.op, d.op).remap_inputs([0]))\n<\/code><\/pre>\n\n<p>Outputs:<\/p>\n\n<pre><code>SubGraphView (graphid=241109520):\n** ops[1]:\n  e\n** inputs[1]:\n  a:0\n** outputs[1]:\n  e:0\n\nSubGraphView (graphid=241109520):\n** ops[2]:\n  c\n  d\n** inputs[1]:\n  a:0\n** outputs[2]:\n  c:0\n  d:0\n<\/code><\/pre>\n"}]}
{"QuestionId":41959318,"AnswerCount":4,"Tags":"<tensorflow><google-cloud-platform><keras><google-cloud-ml>","CreationDate":"2017-01-31T13:52:44.537","AcceptedAnswerId":null,"OwnerUserId":7495302.0,"Title":"Deploying Keras Models via Google Cloud ML","Body":"<p>I am looking to use Google Cloud ML to host my Keras models so that I can call the API and make some predictions. I am running into some issues from the Keras side of things. <\/p>\n\n<p>So far I have been able to build a model using TensorFlow and deploy it on CloudML. In order for this to work I had to make some changes to my basic TF code. The changes are documented here: <a href=\"https:\/\/cloud.google.com\/ml\/docs\/how-tos\/preparing-models#code_changes\" rel=\"noreferrer\">https:\/\/cloud.google.com\/ml\/docs\/how-tos\/preparing-models#code_changes<\/a><\/p>\n\n<p>I have also been able to train a similar model using Keras. I can even save the model in the same export and export.meta format as I would get with TF.<\/p>\n\n<pre><\/pre>\n\n<p>The part I am missing is how do I add the placeholders for input and output into the graph I build on Keras?<\/p>\n","answers":[{"AnswerId":"44232441","CreationDate":"2017-05-28T22:01:32.150","ParentId":null,"OwnerUserId":"5722895","Title":null,"Body":"<p>After training your model on Google Cloud ML Engine (check out <a href=\"http:\/\/liufuyang.github.io\/2017\/04\/02\/just-another-tensorflow-beginner-guide-4.html\" rel=\"noreferrer\">this awesome tutorial<\/a>  ), I named the input and output of my graph with <\/p>\n\n<pre><code>signature = predict_signature_def(inputs={'NAME_YOUR_INPUT': new_Model.input},\n                                  outputs={'NAME_YOUR_OUTPUT': new_Model.output})\n<\/code><\/pre>\n\n<p>You can see the full exporting example for an already trained keras model 'model.h5' below. <\/p>\n\n<pre><code>import keras.backend as K\nimport tensorflow as tf\nfrom keras.models import load_model, Sequential\nfrom tensorflow.python.saved_model import builder as saved_model_builder\nfrom tensorflow.python.saved_model import tag_constants, signature_constants\nfrom tensorflow.python.saved_model.signature_def_utils_impl import predict_signature_def\n\n# reset session\nK.clear_session()\nsess = tf.Session()\nK.set_session(sess)\n\n# disable loading of learning nodes\nK.set_learning_phase(0)\n\n# load model\nmodel = load_model('model.h5')\nconfig = model.get_config()\nweights = model.get_weights()\nnew_Model = Sequential.from_config(config)\nnew_Model.set_weights(weights)\n\n# export saved model\nexport_path = 'YOUR_EXPORT_PATH' + '\/export'\nbuilder = saved_model_builder.SavedModelBuilder(export_path)\n\nsignature = predict_signature_def(inputs={'NAME_YOUR_INPUT': new_Model.input},\n                                  outputs={'NAME_YOUR_OUTPUT': new_Model.output})\n\nwith K.get_session() as sess:\n    builder.add_meta_graph_and_variables(sess=sess,\n                                         tags=[tag_constants.SERVING],\n                                         signature_def_map={\n                                             signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature})\n    builder.save()\n<\/code><\/pre>\n\n<p>You can also see my <a href=\"https:\/\/github.com\/Lausbert\/Exermote\/blob\/master\/ExermotePreprocessingAndTraining\/trainer\/exermote.py\" rel=\"noreferrer\">full implementation<\/a>. <\/p>\n\n<p>edit: And if my answer solved your problem, just leave me an uptick here :)<\/p>\n"},{"AnswerId":"42034086","CreationDate":"2017-02-03T22:34:39.603","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>I don't know much about Keras. I consulted with some experts, and the following should work:<\/p>\n\n<pre><code>from keras import backend as k\n\n# Build the model first\nmodel = ...    \n\n# Declare the inputs and outputs for CloudML\ninputs = dict(zip((layer.name for layer in model.input_layers),\n                  (t.name for t in model.inputs)))\ntf.add_to_collection('inputs', json.dumps(inputs))\n\noutputs = dict(zip((layer.name for layer in model.output_layers),\n                   (t.name for t in model.outputs)))\ntf.add_to_collection('outputs', json.dumps(outputs))\n\n# Fit\/train the model\nmodel.fit(...)\n\n# Export the model\nsaver = tf.train.Saver()\nsession = K.get_session()\nsaver.save(session, 'export')\n<\/code><\/pre>\n\n<p>Some important points:<\/p>\n\n<ul>\n<li>You have to call tf.add_to_collection after you create the model\nbut before you ever call  K.get_session(), fit etc., <\/li>\n<li>You should be sure set the name of input and output layers when\nyou add them to the graph because you'll need to refer to them\nwhen you send prediction requests.<\/li>\n<\/ul>\n"},{"AnswerId":"43243108","CreationDate":"2017-04-05T23:13:17.337","ParentId":null,"OwnerUserId":"5368759","Title":null,"Body":"<p>I found out that in order to use keras on google cloud one has to install it with a setup.py script and put it on the same place folder where you run the gcloud command:<\/p>\n\n<pre><code>\u251c\u2500\u2500 setup.py\n\u2514\u2500\u2500 trainer\n    \u251c\u2500\u2500 __init__.py\n    \u251c\u2500\u2500 cloudml-gpu.yaml\n    \u251c\u2500\u2500 example5-keras.py\n<\/code><\/pre>\n\n<p>And in the setup.py you put content such as:<\/p>\n\n<pre><code>from setuptools import setup, find_packages\n\nsetup(name='example5',\n  version='0.1',\n  packages=find_packages(),\n  description='example to run keras on gcloud ml-engine',\n  author='Fuyang Liu',\n  author_email='fuyang.liu@example.com',\n  license='MIT',\n  install_requires=[\n      'keras',\n      'h5py'\n  ],\n  zip_safe=False)\n<\/code><\/pre>\n\n<p>Then you can start your job running on gcloud such as:<\/p>\n\n<pre><code>export BUCKET_NAME=tf-learn-simple-sentiment\nexport JOB_NAME=\"example_5_train_$(date +%Y%m%d_%H%M%S)\"\nexport JOB_DIR=gs:\/\/$BUCKET_NAME\/$JOB_NAME\nexport REGION=europe-west1\n\ngcloud ml-engine jobs submit training $JOB_NAME \\\n  --job-dir gs:\/\/$BUCKET_NAME\/$JOB_NAME \\\n  --runtime-version 1.0 \\\n  --module-name trainer.example5-keras \\\n  --package-path .\/trainer \\\n  --region $REGION \\\n  --config=trainer\/cloudml-gpu.yaml \\\n  -- \\\n  --train-file gs:\/\/tf-learn-simple-sentiment\/sentiment_set.pickle\n<\/code><\/pre>\n\n<p>To use GPU then add a file such as <code>cloudml-gpu.yaml<\/code> in your module with the following content:<\/p>\n\n<pre><code>trainingInput:\n  scaleTier: CUSTOM\n  # standard_gpu provides 1 GPU. Change to complex_model_m_gpu for 4 \nGPUs\n  masterType: standard_gpu\n  runtimeVersion: \"1.0\"\n<\/code><\/pre>\n"},{"AnswerId":"51700445","CreationDate":"2018-08-06T04:09:35.267","ParentId":null,"OwnerUserId":"630752","Title":null,"Body":"<p>Here's another answer that may help. Assuming you already have a keras model you should be able to append this to the end of your script and get an ML Engine compatible version of the model (protocol buffer). Note that you need to upload the <code>saved_model.pb<\/code> file and the sibling directory with variables to ML Engine for it to work. Note also that the .pb file must be named <code>saved_model.pb<\/code> or <code>saved_model.pbtxt<\/code>.<\/p>\n\n<p>Assuming your model is name <code>model<\/code><\/p>\n\n<pre><code>from tensorflow import saved_model\n\nmodel_builder = saved_model.builder.SavedModelBuilder(\"exported_model\")                                                     \ninputs = {                                                                          \n    'input': saved_model.utils.build_tensor_info(model.input)                    \n}                                                                                   \noutputs = {                                                                         \n    'earnings': saved_model.utils.build_tensor_info(model.output)                \n}                                                                                                                                                \nsignature_def = saved_model.signature_def_utils.build_signature_def(             \n    inputs=inputs,                                                                  \n    outputs=outputs,                                                                \n    method_name=saved_model.signature_constants.PREDICT_METHOD_NAME              \n)                                                                            \nmodel_builder.add_meta_graph_and_variables(                                         \n    K.get_session(),                                                                \n    tags=[saved_model.tag_constants.SERVING],                                    \n    signature_def_map={saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY: signature_def\n    })                                                                                                                                                                       \nmodel_builder.save()   \n<\/code><\/pre>\n\n<p>will export the model to directory <code>\/exported_model<\/code>.<\/p>\n"}]}
{"QuestionId":41960319,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-31T14:41:31.627","AcceptedAnswerId":41982937.0,"OwnerUserId":2324298.0,"Title":"Tensorflow taking too long to read a tfrecord file","Body":"<p>I have tfrecord file that has five columns - context, context_len, utterance, uuterance_len and label. The <a href=\"https:\/\/drive.google.com\/open?id=0B_bZck-ksdkpVEtVc1R6Y01HMWM\" rel=\"nofollow noreferrer\">source of the tfrecord file<\/a> is the Ubuntu Dialog Corpus. Problem is that its taking way too long to read in the file<\/p>\n\n<pre><\/pre>\n\n<p>When the code took too long to read the file, I thought it must be due to the file size. So I subsetted the original csv file down to ten records and converted it into a tfrecords file but no difference. I'm still facing the same issue. So now I'm convinced it has something to do with my code. My objective is to parse a single example and use it later to train a RNN.<\/p>\n","answers":[{"AnswerId":"41982937","CreationDate":"2017-02-01T15:01:42.253","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>you need to start queue runners before the first run call<\/p>\n"}]}
{"QuestionId":41961117,"AnswerCount":0,"Tags":"<tensorflow><gpu>","CreationDate":"2017-01-31T15:18:28.313","AcceptedAnswerId":null,"OwnerUserId":4157666.0,"Title":"Confirm of using GPU in tensorflow","Body":"<p>I am to confirm that I am using GPU with tensorflow but I am kinda confused. So, I am using the following code to check if GPU is used:<\/p>\n\n<pre><\/pre>\n\n<p>which outputs the following:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HG1qd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/HG1qd.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>From my understanding it seems from the output that GPU is used for calculations. But when I monitor it with:<\/p>\n\n<pre><\/pre>\n\n<p>I have the following output, which constantly gives the following result:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GdDtn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GdDtn.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>From my undestanding in the last pic it doesnt use GPU. Is it something that I missunderstound? Any tip?<\/p>\n\n<p><strong>Edit<\/strong> (after Yaroslav Bulatov comment, which helped a lot)<\/p>\n\n<p>I can confirm now that I am using GPU. For a  I got the following results from the ctf file:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/tZtLW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tZtLW.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However there are some functions from tensorflow that seems they are using CPU. For example :\n<a href=\"https:\/\/i.stack.imgur.com\/Gb5OS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Gb5OS.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Interesting..<\/p>\n","answers":[]}
{"QuestionId":41961341,"AnswerCount":1,"Tags":"<python><tensorflow><supervised-learning><unsupervised-learning><autoencoder>","CreationDate":"2017-01-31T15:30:16.380","AcceptedAnswerId":42011904.0,"OwnerUserId":7495037.0,"Title":"Using weights from Autoencoder to initialize neural network in tensorflow","Body":"<p>I built an Autoencoder using Python and Tensorflow. To build the Autoencoder I used the Tensorflow tutorial on how to build an Autoencoder to read the MNIST Data set on handwritten digits. I used it to find features of CGRA compositions.<\/p>\n\n<p>So far I restructured the Code for the Autoencoder in a way that I can use it on my own data. I found features and already managed to reconstruct the Input,\nup to a certain error. NOW, I am trying to use the Autoencoders weights, to initialize a Neural Network with parameters similiar to the encoder part of my Autoencoder. Then, add one extra Layer with a single neuron and a linear activation function to perform Regression Analysis(or basically supervised learning). <\/p>\n\n<p>So my question is: How do I initialize a neural network with specific weights (not random) using tensorflow?<\/p>\n\n<p>I'd be grateful for any kind of help. Links to Tutorials or other links to other Threads.<\/p>\n\n<p>Tahnks in Advance!<\/p>\n","answers":[{"AnswerId":"42011904","CreationDate":"2017-02-02T20:39:08.003","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>When you build <code>tf.Variable<\/code>, the first argument is an <code>initial_value<\/code>.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/variables#Variable.__init__\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/variables#Variable.<strong>init<\/strong><\/a><\/p>\n\n<p>You can provide any Tensor you like to initialize the variables, not just, say, random initialization.<\/p>\n\n<p>Another option is you can assign values to the variables after construction, if you find that easier.<\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41961649,"AnswerCount":1,"Tags":"<tensorflow><gradient><distribution>","CreationDate":"2017-01-31T15:46:52.843","AcceptedAnswerId":54076221.0,"OwnerUserId":1938262.0,"Title":"Gradients of Bernoulli Samples","Body":"<p>I'm trying to calculate the gradients of the samples from a Bernoulli distribution w.r.t. the probabilities  (of a sample being ).<\/p>\n\n<p>I tried using both the implementation of the Bernoulli distribution provided in  and my own simple implementation based on this <a href=\"https:\/\/stackoverflow.com\/questions\/35487598\/sampling-bernoulli-random-variables-in-tensorflow\">discussion<\/a>. However both methods fail when I try to calculate the gradients.<\/p>\n\n<p>Using the  implementation:<\/p>\n\n<pre><\/pre>\n\n<p>The above code gives me the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Using my implementation:<\/p>\n\n<pre><\/pre>\n\n<p>Same error:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to calculate the gradients of Bernoulli samples?<\/p>\n\n<p>(My TensorFlow version is 0.12).<\/p>\n","answers":[{"AnswerId":"54076221","CreationDate":"2019-01-07T14:26:35.830","ParentId":null,"OwnerUserId":"914693","Title":null,"Body":"<p>You cannot backprop through a discrete stochastic node for obvious reasons. As gradients are not defined.\nHowever if you approximate the Bernoulli with a continuos distribution controlled by a temperature parameter, yes you can.<\/p>\n\n<p>This idea is called reparametrization trick and is implemented in the RelaxedBernoulli in Tensorflow Probability (or also in TF.contrib library)<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/distributions\/RelaxedBernoulli\" rel=\"nofollow noreferrer\">Relaxed bernoulli<\/a><\/p>\n\n<p>You can specify the probability <code>p<\/code> of your Bernoulli, which is your random variable, et voil\u00e0.<\/p>\n"}]}
{"QuestionId":41962131,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-01-31T16:07:36.667","AcceptedAnswerId":41962132.0,"OwnerUserId":3642234.0,"Title":"WARNING:tensorflow - initialize_all_variables (from tensorflow.python.ops.variables) is deprecated","Body":"<p>While initializaing tensorflow <\/p>\n\n<pre><\/pre>\n\n<p>get below warning<\/p>\n\n<pre><\/pre>\n\n<p>How to remove the warning and use updated code?<\/p>\n","answers":[{"AnswerId":"41962132","CreationDate":"2017-01-31T16:07:36.667","ParentId":null,"OwnerUserId":"3642234","Title":null,"Body":"<p>Solution is to use the updated method <code>global_variables_initializer()<\/code> instead of <code>initialize_all_variables<\/code><\/p>\n\n<pre><code>init = tf.global_variables_initializer()\n<\/code><\/pre>\n"}]}
{"QuestionId":41962854,"AnswerCount":1,"Tags":"<neural-network><theano><keras><conv-neural-network><keras-layer>","CreationDate":"2017-01-31T16:43:29.607","AcceptedAnswerId":null,"OwnerUserId":991077.0,"Title":"Keras convolution along single feature row","Body":"<p>I have a multiclass classification problem. Say I have a feature matrix:<\/p>\n\n<pre><\/pre>\n\n<p>And labels:<\/p>\n\n<pre><\/pre>\n\n<p>I want try to apply convolution kernels along each single feature row with Keras. Say nb_filter=2 and batch_size=3. So I expect input shape for convolution layer to be (3, 4) and output shape to be (3, 3) (as it is applied for AB, BC, CD).<\/p>\n\n<p>Here what I tried with Keras (v1.2.1, Theano backend):<\/p>\n\n<pre><\/pre>\n\n<p>Shapes:<\/p>\n\n<pre><\/pre>\n\n<p>Output:<\/p>\n\n<pre><\/pre>\n\n<p>When I try to run this code I get and exception:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to reshape y_train:<\/p>\n\n<pre><\/pre>\n\n<p>This gives me exception:<\/p>\n\n<pre><\/pre>\n\n<ol>\n<li>Is this approach with Convolution1D correct to achieve my goal?<\/li>\n<li>If #1 is yes, how can I fix my code?<\/li>\n<\/ol>\n\n<p>I've already read numerous github issues and some questions (<a href=\"https:\/\/stackoverflow.com\/questions\/38364415\/keras-building-a-network-for-9-dimensional-feature-vector\">1<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/38656566\/input-dimensions-to-a-one-dimensional-convolutional-network-in-keras?rq=1\">2<\/a>) here, but it didn't really help.<\/p>\n\n<p>Thanks.<\/p>\n\n<p><strong>UPDATE1:<\/strong>\nAccording to Matias Valdenegro comment.\nHere are shapes after reshaping 'X' and after onehot encoding for 'y':<\/p>\n\n<pre><\/pre>\n\n<p>Output:<\/p>\n\n<pre><\/pre>\n\n<p><strong>UPDATE2:<\/strong> Thanks again to Matias Valdenegro. X reshaping is done after creating model for sure it was a copy paste issue. The code should look like:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41970239","CreationDate":"2017-02-01T01:01:58.377","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The input to a 1D convolution should have dimensions (num_samples, channels, width). So that means you need to reshape X_train and X_test, not y_train:<\/p>\n\n<pre><code>X_train = X_train.reshape(X_train.shape[0], 1, X_train.shape[1])\nX_test = X_test.reshape(X_test.shape[0], 1, X_test.shape[1])\n<\/code><\/pre>\n"}]}
{"QuestionId":41963839,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-01-31T17:29:19.230","AcceptedAnswerId":41963961.0,"OwnerUserId":6405712.0,"Title":"TensorFlow: Evaluating functions across a tensor","Body":"<p>Sci-Py lets me do the following with the norm.pdf function:<\/p>\n\n<pre><\/pre>\n\n<p>This will basically give me an array, where each element corresponds to the probability density of x given the corresponding mean and variance.<\/p>\n\n<p>I'm trying to do this, with the difference being that x, mu, and var are now tensors (x is rank-0, mu and var are rank 1 tensors). I want norm.pdf to be a rank 1 tensor as well. I'm trying to do this on TensorFlow but not sure how to.<\/p>\n\n<p>Suggestions?<\/p>\n","answers":[{"AnswerId":"41963961","CreationDate":"2017-01-31T17:35:25.830","ParentId":null,"OwnerUserId":"1611389","Title":null,"Body":"<p>Using TensorFlow, you can do it in 2 steps with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.distributions\/univariate__scalar__distributions#Normal\" rel=\"nofollow noreferrer\"><code>tf.contrib.distributions.Normal<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.distributions\/base_classes#Distribution.pdf\" rel=\"nofollow noreferrer\"><code>tf.contrib.distributions.Distribution.pdf<\/code><\/a><\/p>\n\n<pre><code># Define a batch of 4 scalar valued Normals (mu=means, sigma=standard deviations)\ndist = tf.contrib.distributions.Normal(mu=[1,2,3,4], sigma=[1,1,1,1])\n\n# Evaluate the pdf of the first distribution on 1\n# returning a length two tensor.\ndist.pdf(1)\n<\/code><\/pre>\n"}]}
{"QuestionId":41964212,"AnswerCount":4,"Tags":"<c++><opencv><tensorflow><protocol-buffers>","CreationDate":"2017-01-31T17:49:47.540","AcceptedAnswerId":41975050.0,"OwnerUserId":7496463.0,"Title":"Conflict Protobuf version when using Opencv and Tensorflow c++","Body":"<p>I am currently trying to use Tensorflow's shared library in a non-bazel project, so I creat a .so file from tensorflow using bazel.<\/p>\n\n<p>but when I launch a c++ program that uses both Opencv and Tensorflow, it makes me the following error :<\/p>\n\n<blockquote>\n  <p>[libprotobuf FATAL external\/protobuf\/src\/google\/protobuf\/stubs\/common.cc:78] This program was compiled against version 2.6.1 of the Protocol Buffer runtime library, which is not compatible with the installed version (3.1.0).  Contact the program author for an update.  If you compiled the program yourself, make sure that your headers are from the same version of Protocol Buffers as your link-time library.  (Version verification failed in \"\/build\/mir-pkdHET\/mir-0.21.0+16.04.20160330\/obj-x86_64-linux-gnu\/src\/protobuf\/mir_protobuf.pb.cc\".)\n  terminate called after throwing an instance of 'google::protobuf::FatalException'<\/p>\n  \n  <p>what():  This program was compiled against version 2.6.1 of the Protocol Buffer runtime library, which is not compatible with the installed version (3.1.0).  Contact the program author for an update.  If you compiled the program yourself, make sure that your headers are from the same version of Protocol Buffers as your link-time library.  (Version verification failed in \"\/build\/mir-pkdHET\/mir-0.21.0+16.04.20160330\/obj-x86_64-linux-gnu\/src\/protobuf\/mir_protobuf.pb.cc\".)\n  Abandon (core dumped)<\/p>\n<\/blockquote>\n\n<p>Can you help me?<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"41975050","CreationDate":"2017-02-01T08:28:43.537","ParentId":null,"OwnerUserId":"2686899","Title":null,"Body":"<p>The error indicates that the program was complied using headers (<code>.h<\/code> files) from protobuf 2.6.1. These headers are typically found in <code>\/usr\/include\/google\/protobuf<\/code> or <code>\/usr\/local\/include\/google\/protobuf<\/code>, though they could be in other places depending on your OS and how the program is being built. You need to update these headers to version 3.1.0 and recompile the program.<\/p>\n"},{"AnswerId":"59000058","CreationDate":"2019-11-22T18:46:39.937","ParentId":null,"OwnerUserId":"7994111","Title":null,"Body":"<p>The error complains about the Protocol Buffer runtime library, which is not compatible with the installed version. This error is coming from the GTK3 library. GTK3 use Protocol Buffer 2.6.1.  If you use GTK3 to support Opencv, you get this error. The easiest way to fix this, you can use QT instead of GTK3. <\/p>\n\n<p>If you use Cmake GUI to install Opencv, just select QT support instead of using GTK3. You can install QT using the following command.<\/p>\n\n<pre><code>sudo apt install qtbase5-dev \n<\/code><\/pre>\n"},{"AnswerId":"42218534","CreationDate":"2017-02-14T05:28:11.350","ParentId":null,"OwnerUserId":"221802","Title":null,"Body":"<p>This is indeed a pretty serious problem! I get the below error similar to you: <\/p>\n\n<pre><code>$.\/ceres_single_test \n[libprotobuf FATAL google\/protobuf\/stubs\/common.cc:78] This program was compiled against version 2.6.1 of the Protocol Buffer runtime library, which is not compatible with the installed version (3.1.0).  Contact the program author for an update.  If you compiled the program yourself, make sure that your headers are from the same version of Protocol Buffers as your link-time library.  (Version verification failed in \"\/build\/mir-pkdHET\/mir-0.21.0+16.04.20160330\/obj-x86_64-linux-gnu\/src\/protobuf\/mir_protobuf.pb.cc\".)\nterminate called after throwing an instance of 'google::protobuf::FatalException'\nAborted\n<\/code><\/pre>\n\n<p>My workaround:<\/p>\n\n<pre><code>cd \/usr\/lib\/x86_64-linux-gnu\nsudo mkdir BACKUP\nsudo mv libmirprotobuf.so* .\/BACKUP\/\n<\/code><\/pre>\n\n<p>Now, the executable under test works, cool. What is not cool, however, is that things like gedit no longer work without running from a shell that has the BACKUP path added to LD_LIBRARY_PATH :-(<br>\n  Hopefully there's a better fix out there? <\/p>\n"},{"AnswerId":"46659714","CreationDate":"2017-10-10T06:12:48.463","ParentId":null,"OwnerUserId":"2805617","Title":null,"Body":"<p>You should rebuild TensorFlow with a linker script to avoid making third party symbols global in the shared library that Bazel creates. This is how the Android Java\/JNI library for TensorFlow is able to coexist with the pre-installed protobuf library on the device (look at the build rules in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/android\" rel=\"nofollow noreferrer\">tensorflow\/contrib\/android<\/a> for a working example)<\/p>\n\n<p>Here's a BUILD file that I adapted from the Android library to do this:<\/p>\n\n<pre><code>package(default_visibility = [\"\/\/visibility:public\"])\n\nlicenses([\"notice\"])  # Apache 2.0\n\nexports_files([\"LICENSE\"])\n\nload(\n    \"\/\/tensorflow:tensorflow.bzl\",\n    \"tf_copts\",\n    \"if_android\",\n)\n\nexports_files([\n    \"version_script.lds\",\n])\n\n# Build the native .so.\n# bazel build \/\/tensorflow\/contrib\/android_ndk:libtensorflow_cc_inference.so \\\n#   --crosstool_top=\/\/external:android\/crosstool \\\n#   --host_crosstool_top=@bazel_tools\/\/tools\/cpp:toolchain \\\n#   --cpu=armeabi-v7a\nLINKER_SCRIPT = \"\/\/tensorflow\/contrib\/android:version_script.lds\"\n\ncc_binary(\n    name = \"libtensorflow_cc_inference.so\",\n    srcs = [],\n    copts = tf_copts() + [\n        \"-ffunction-sections\",\n        \"-fdata-sections\",\n    ],\n    linkopts = if_android([\n        \"-landroid\",\n        \"-latomic\",\n        \"-ldl\",\n        \"-llog\",\n        \"-lm\",\n        \"-z defs\",\n        \"-s\",\n        \"-Wl,--gc-sections\",\n        \"-Wl,--version-script\",  # This line must be directly followed by LINKER_SCRIPT.\n        LINKER_SCRIPT,\n    ]),\n    linkshared = 1,\n    linkstatic = 1,\n    tags = [\n        \"manual\",\n        \"notap\",\n    ],\n    deps = [\n        \"\/\/tensorflow\/core:android_tensorflow_lib\",\n        LINKER_SCRIPT,\n    ],\n)\n<\/code><\/pre>\n\n<p>And the contents of <code>version_script.lds<\/code>:<\/p>\n\n<pre><code>{\n  global:\n    extern \"C++\" {\n        tensorflow::*;\n    };\n  local:\n    *;\n};\n<\/code><\/pre>\n\n<p>This will make everything in the <code>tensorflow<\/code> namespace global and available through the library, while hiding the reset and preventing it from conflicting with protobuf.<\/p>\n\n<p>(wasted a ton of time on this so I hope it helps!)<\/p>\n"}]}
{"QuestionId":41964248,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-01-31T17:51:13.220","AcceptedAnswerId":null,"OwnerUserId":2820379.0,"Title":"In dynamic_rnn() is it valid to include variables in the state?","Body":"<p>I'm implementing a RNN cell around BasicLSTMCell where I want to be able to look back on past hidden states (across batch boundaries).  I'm using  and the basic pattern I use is:<\/p>\n\n<pre><\/pre>\n\n<p> and  are variables.  is a function that uses  to write the new hidden states to the buffer and returns the result.<\/p>\n\n<p>I rely on the assumption that accesses to scatter updates return value guarantee that the new variable value is used (similar to <a href=\"https:\/\/stackoverflow.com\/questions\/35148121\/assign-op-in-tensorflow-what-is-the-return-value\">this<\/a>) so that caching of variables does not mess things up.<\/p>\n\n<p>From debug prints it <em>seems<\/em> to work but it would be nice to get some confirmation or suggestions on alternatives. <\/p>\n","answers":[]}
{"QuestionId":41965187,"AnswerCount":2,"Tags":"<cuda><tensorflow>","CreationDate":"2017-01-31T18:44:40.960","AcceptedAnswerId":41975926.0,"OwnerUserId":3094625.0,"Title":"Nvidia device error in tensorflow","Body":"<p>To test my tensorflow installation I am using the mnist example provided in tensorflow repository, but when I execute the convolutional.py script I have this output:<\/p>\n\n<pre><\/pre>\n\n<p>My first idea was that maybe I had problems in cuda installation but I tested using one of the examples provided for nvidia. In this case I used this example: <\/p>\n\n<blockquote>\n  <p>NVIDIA_CUDA-8.0_Samples\/6_Advanced\/c++11_cuda<\/p>\n<\/blockquote>\n\n<p>And the output is this:<\/p>\n\n<pre><\/pre>\n\n<p>Then my conclusion is the cuda is installed correctly. But I don not have any idea what is happening here. If someone can help me I will appreciated.<\/p>\n\n<p>For more information this is my gpu configuration:<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>It is normal the two nvidia cards have the same physical id? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41975926","CreationDate":"2017-02-01T09:15:58.147","ParentId":null,"OwnerUserId":"681865","Title":null,"Body":"<p>The important points in the output you have shown is this:<\/p>\n\n<pre><code>I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:885] Found device 0 with properties: \nname: GeForce GTX 980 Ti\nmajor: 5 minor: 2 memoryClockRate (GHz) 1.2405\npciBusID 0000:03:00.0\nTotal memory: 5.93GiB\nFree memory: 5.83GiB\n<\/code><\/pre>\n\n<p>i.e. the compute device you want is enumerated as device 0 and<\/p>\n\n<pre><code>E tensorflow\/core\/common_runtime\/direct_session.cc:137] Internal: failed initializing StreamExecutor for CUDA device ordinal 1: Internal: failed call to cuDevicePrimaryCtxRetain: CUDA_ERROR_INVALID_DEVICE\n<\/code><\/pre>\n\n<p>i.e. the compute device generating the error is enumerated as device 1. Device 1 is your display GPU, which can't be used for computation in Tensorflow. If you either mark that device as compute prohibited with <code>nvidia-smi<\/code>, or use the <code>CUDA_VISIBLE_DEVICES<\/code> environment variable to only make your compute device visible to CUDA, the error should probably disappear.<\/p>\n"},{"AnswerId":"45511194","CreationDate":"2017-08-04T16:03:32.983","ParentId":null,"OwnerUserId":"570918","Title":null,"Body":"<p>I encountered a similar error when I attempted to run the <code>classify_image.py<\/code> script that is part of <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_recognition\" rel=\"nofollow noreferrer\">the image recognition tutorial<\/a>.  Since I already had a running Python session (elpy) in which I had run some TensorFlow code, the GPUs were allocated there and thus were not available for the script I was attempting to run from shell.<\/p>\n\n<p>Quitting the existing Python session resolved the error.<\/p>\n"}]}
{"QuestionId":41965202,"AnswerCount":1,"Tags":"<variables><tensorflow>","CreationDate":"2017-01-31T18:45:42.700","AcceptedAnswerId":41965889.0,"OwnerUserId":4744283.0,"Title":"Tensorflow: how to assign variables properly","Body":"<p>It's not duplicate of <a href=\"https:\/\/stackoverflow.com\/questions\/34220532\/how-to-assign-value-to-a-tensorflow-variable\">How to assign value to a tensorflow variable?<\/a><\/p>\n\n<p>I was trying to do simpliest thing: just swap variables <a href=\"https:\/\/stackoverflow.com\/questions\/41949633\/tensorflow-how-to-swap-variables-between-scopes-and-set-variables-in-scope-from\">Tensorflow: how to swap variables between scopes and set variables in scope from another<\/a>,\n and I still can't do it.<\/p>\n\n<p>BUT now I know that  changes even copy of tensor which I get with . I don't want this. I need copy of variable for swapping. <\/p>\n\n<pre><\/pre>\n\n<p>How can I assign value to  without changing ?<\/p>\n","answers":[{"AnswerId":"41965889","CreationDate":"2017-01-31T19:26:38.367","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#identity\" rel=\"nofollow noreferrer\"><code>tf.identity()<\/code><\/a> operation is stateless. When you have a <code>tf.Variable<\/code> called <code>a<\/code>, the value of <code>tf.identity(a)<\/code> will always be the same as the value of <code>a<\/code>. If you want <code>b<\/code> to remember a previous value of <code>a<\/code>, you should create <code>b<\/code> as a <code>tf.Variable<\/code> as well:<\/p>\n\n<pre><code>a = tf.Variable(1)\nb = tf.Variable(a.initialized_value())\n\nsess.run(tf.global_variables_initializer())\n\n# Initially, variables `a` and `b` have the same value.\nprint(sess.run([a, b])) ==&gt; [1, 1]\n\n# Update the value of `a` to 2.\nassign_op = a.assign(2)\nsess.run(assign_op)\n\n# Now, `a` and `b` have different values.\nprint(sess.run([a, b])) ==&gt; [2, 1]\n<\/code><\/pre>\n"}]}
{"QuestionId":41966657,"AnswerCount":1,"Tags":"<python><theano><keras><gpuimage>","CreationDate":"2017-01-31T20:12:27.740","AcceptedAnswerId":null,"OwnerUserId":7371868.0,"Title":"ValueError: GpuCorrMM images and kernel must have the same stack size","Body":"<p>My input data shape is [n, 3, 64, 64]<\/p>\n\n<p>I got this after run the code on Stampede.<\/p>\n\n<pre><\/pre>\n\n<p>what's going on with the code and how to solve this problem? Thanks<\/p>\n\n<p>My code:<\/p>\n\n<pre><\/pre>\n\n<p>Additionally, Does someone know what is Inputs shapes: [(32, 8, 16, 1024), (256, 512, 5, 5)] stands for? How could I fix this problem?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"41967420","CreationDate":"2017-01-31T21:00:54.177","ParentId":null,"OwnerUserId":"7371868","Title":null,"Body":"<p>When I run this code on cpu, the gan_model.summary() result is :\n0.0 1.0\nX_train shape ---  (29404, 3, 64, 64)\n29404 train samples<\/p>\n\n<hr>\n\n<pre><code>Layer (type)                     Output Shape          Param #     Connected to                     \n====================================================================================================\ndense_1 (Dense)                  (None, 16384)         1654784     dense_input_1[0][0]              \n____________________________________________________________________________________________________\nbatchnormalization_1 (BatchNorma (None, 16384)         65536       dense_1[0][0]                    \n____________________________________________________________________________________________________\nactivation_1 (Activation)        (None, 16384)         0           batchnormalization_1[0][0]       \n____________________________________________________________________________________________________\nreshape_1 (Reshape)              (None, 1024, 4, 4)    0           activation_1[0][0]               \n____________________________________________________________________________________________________\nupsampling2d_1 (UpSampling2D)    (None, 1024, 8, 8)    0           reshape_1[0][0]                  \n____________________________________________________________________________________________________\nconvolution2d_1 (Convolution2D)  (None, 512, 8, 8)     13107712    upsampling2d_1[0][0]             \n____________________________________________________________________________________________________\nbatchnormalization_2 (BatchNorma (None, 512, 8, 8)     32          convolution2d_1[0][0]            \n____________________________________________________________________________________________________\nactivation_2 (Activation)        (None, 512, 8, 8)     0           batchnormalization_2[0][0]       \n____________________________________________________________________________________________________\nupsampling2d_2 (UpSampling2D)    (None, 512, 16, 16)   0           activation_2[0][0]               \n____________________________________________________________________________________________________\nconvolution2d_2 (Convolution2D)  (None, 256, 16, 16)   3277056     upsampling2d_2[0][0]             \n____________________________________________________________________________________________________\nbatchnormalization_3 (BatchNorma (None, 256, 16, 16)   64          convolution2d_2[0][0]            \n____________________________________________________________________________________________________\nactivation_3 (Activation)        (None, 256, 16, 16)   0           batchnormalization_3[0][0]       \n____________________________________________________________________________________________________\nupsampling2d_3 (UpSampling2D)    (None, 256, 32, 32)   0           activation_3[0][0]               \n____________________________________________________________________________________________________\nconvolution2d_3 (Convolution2D)  (None, 128, 32, 32)   819328      upsampling2d_3[0][0]             \n____________________________________________________________________________________________________\nbatchnormalization_4 (BatchNorma (None, 128, 32, 32)   128         convolution2d_3[0][0]            \n____________________________________________________________________________________________________\nactivation_4 (Activation)        (None, 128, 32, 32)   0           batchnormalization_4[0][0]       \n____________________________________________________________________________________________________\nupsampling2d_4 (UpSampling2D)    (None, 128, 64, 64)   0           activation_4[0][0]               \n____________________________________________________________________________________________________\nconvolution2d_4 (Convolution2D)  (None, 64, 64, 64)    204864      upsampling2d_4[0][0]             \n____________________________________________________________________________________________________\nbatchnormalization_5 (BatchNorma (None, 64, 64, 64)    256         convolution2d_4[0][0]            \n____________________________________________________________________________________________________\nactivation_5 (Activation)        (None, 64, 64, 64)    0           batchnormalization_5[0][0]       \n____________________________________________________________________________________________________\nconvolution2d_5 (Convolution2D)  (None, 3, 64, 64)     4803        activation_5[0][0]               \n____________________________________________________________________________________________________\nactivation_6 (Activation)        (None, 3, 64, 64)     0           convolution2d_5[0][0]            \n====================================================================================================\nTotal params: 19,134,563\nTrainable params: 19,101,555\nNon-trainable params: 33,008\n____________________________________________________________________________________________________\n____________________________________________________________________________________________________\nLayer (type)                     Output Shape          Param #     Connected to                     \n====================================================================================================\nconvolution2d_6 (Convolution2D)  (None, 64, 32, 32)    4864        convolution2d_input_1[0][0]      \n____________________________________________________________________________________________________\nleakyrelu_1 (LeakyReLU)          (None, 64, 32, 32)    0           convolution2d_6[0][0]            \n____________________________________________________________________________________________________\nconvolution2d_7 (Convolution2D)  (None, 128, 16, 16)   204928      leakyrelu_1[0][0]                \n____________________________________________________________________________________________________\nleakyrelu_2 (LeakyReLU)          (None, 128, 16, 16)   0           convolution2d_7[0][0]            \n____________________________________________________________________________________________________\nconvolution2d_8 (Convolution2D)  (None, 256, 8, 8)     819456      leakyrelu_2[0][0]                \n____________________________________________________________________________________________________\nleakyrelu_3 (LeakyReLU)          (None, 256, 8, 8)     0           convolution2d_8[0][0]            \n____________________________________________________________________________________________________\nconvolution2d_9 (Convolution2D)  (None, 512, 4, 4)     3277312     leakyrelu_3[0][0]                \n____________________________________________________________________________________________________\nleakyrelu_4 (LeakyReLU)          (None, 512, 4, 4)     0           convolution2d_9[0][0]            \n____________________________________________________________________________________________________\nflatten_1 (Flatten)              (None, 8192)          0           leakyrelu_4[0][0]                \n____________________________________________________________________________________________________\ndense_2 (Dense)                  (None, 1024)          8389632     flatten_1[0][0]                  \n____________________________________________________________________________________________________\nleakyrelu_5 (LeakyReLU)          (None, 1024)          0           dense_2[0][0]                    \n____________________________________________________________________________________________________\ndropout_1 (Dropout)              (None, 1024)          0           leakyrelu_5[0][0]                \n____________________________________________________________________________________________________\ndense_3 (Dense)                  (None, 2)             2050        dropout_1[0][0]                  \n====================================================================================================\nTotal params: 12,698,242\nTrainable params: 12,698,242\nNon-trainable params: 0\n____________________________________________________________________________________________________\n____________________________________________________________________________________________________\nLayer (type)                     Output Shape          Param #     Connected to                     \n====================================================================================================\ninput_2 (InputLayer)             (None, 100)           0                                            \n____________________________________________________________________________________________________\nsequential_1 (Sequential)        (None, 3, 64, 64)     19134563    input_2[0][0]                    \n____________________________________________________________________________________________________\nsequential_2 (Sequential)        (None, 2)             12698242    sequential_1[1][0]               \n====================================================================================================\nTotal params: 31,832,805\nTrainable params: 31,799,797\nNon-trainable params: 33,008\n____________________________________________________________________________________________________\nPre-training generator...\n<\/code><\/pre>\n"}]}
{"QuestionId":41966701,"AnswerCount":2,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-01-31T20:15:13.603","AcceptedAnswerId":41969162.0,"OwnerUserId":7302262.0,"Title":"Tensorflow MNIST using convolution parameters","Body":"<p>I dont understand why in <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/pros\/\" rel=\"nofollow noreferrer\">official documentation<\/a> he use  the  bias_variable of size 32, as i know the bias num is equal to num of neurons in the layer and in this case the number of neurons in the first layer is equal to 28 because the image pixels = 28 and he use the , why it is equal 32 not 28?<\/p>\n","answers":[{"AnswerId":"41969162","CreationDate":"2017-01-31T23:07:49.880","ParentId":null,"OwnerUserId":"6471140","Title":null,"Body":"<p>Remember that mnist is using convolutional networks, not conventional neural networks and hence, you are dealing with convolutions(not neurons) and in this example , and in convolutions you commonly use a bias for every output channel, and this example uses 32 output channels in the first convolution layer and that gives you 32 biases.<\/p>\n"},{"AnswerId":"44095588","CreationDate":"2017-05-21T10:09:45.653","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>They use bias of size 32 to be compatible with the weights:<\/p>\n\n<pre><code>W_conv1 = weight_variable([5, 5, 1, 32])\nb_conv1 = bias_variable([32])\n<\/code><\/pre>\n\n<p>They use weights in conv2d function <code>tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')<\/code>.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\">tf.nn.conv2d()<\/a> tells that the second parameter represent your filter and consists of <code>[filter_height, filter_width, in_channels, out_channels]<\/code>. So <code>[5, 5, 1, 32]<\/code> means that your <code>in_channels<\/code> is 1: you have a greyscale image, so no surprises here. <\/p>\n\n<p>32 means that during our learning phase, the network will try to learn 32 different kernels which will be used during the prediction. You can change this number to any other number as it is a hyperparameter that you can tune.<\/p>\n"}]}
{"QuestionId":41967234,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning>","CreationDate":"2017-01-31T20:49:53.443","AcceptedAnswerId":null,"OwnerUserId":2802468.0,"Title":"Modifying variables names after restoring from checkpoint in TensorFlow","Body":"<p>let's say I have two models with different configurations. I train and checkpoint them.<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to load these models in the same session so I can use them at the same time. The variable names are mostly the same for the above two models, so to avoid name mangling, I add a name scope for each model.<\/p>\n\n<pre><\/pre>\n\n<p>To restore from the checkpoint for , I collect the variables and variable names but want to rename with the proper scope.<\/p>\n\n<pre><\/pre>\n\n<p>But I get the following error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"53374646","CreationDate":"2018-11-19T12:27:38.000","ParentId":null,"OwnerUserId":"1676547","Title":null,"Body":"<p>The <code>vars<\/code> dict in <code>saver = tf.train.Saver(var_list=vars)<\/code> must be a dict whose value is the reference to the <strong>tf.Variable<\/strong> of current graph in current session.<br>\nBut in your case <code>var = tf.contrib.framework.load_variable(path, name)<\/code>, the problem is that <code>var<\/code> is <strong>numpy.ndarray<\/strong><\/p>\n"}]}
{"QuestionId":41967963,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><keras><kaggle>","CreationDate":"2017-01-31T21:38:40.933","AcceptedAnswerId":null,"OwnerUserId":3715132.0,"Title":"Poor results with Keras inbuilt VGG16 model on cats vs. dogs dataset","Body":"<p>I am trying to apply the inbuilt VGG16 Keras model to the <a href=\"https:\/\/www.kaggle.com\/c\/dogs-vs-cats-redux-kernels-edition\" rel=\"nofollow noreferrer\">Kaggle Cats vs Dogs<\/a> dataset. <\/p>\n\n<p>However, I get 52% accuracy which is barely better than complete hazard.<\/p>\n\n<p>Any idea why this would be the case? <a href=\"https:\/\/www.dropbox.com\/s\/h8gbliyt76bsurr\/vgg_own.ipynb?dl=0\" rel=\"nofollow noreferrer\">Download notebook<\/a><\/p>\n\n<p>Vlad<\/p>\n","answers":[{"AnswerId":"41990520","CreationDate":"2017-02-01T21:55:25.693","ParentId":null,"OwnerUserId":"3715132","Title":null,"Body":"<p>The problem is that I was using a very high learning rate (0.01).<\/p>\n\n<p>I have since amended the learning rate to <code>lr=0.0001<\/code> and the accuracy after one epoch is closer to 95%, which is the expected result.<\/p>\n"}]}
{"QuestionId":41968605,"AnswerCount":1,"Tags":"<python><python-2.7><machine-learning><tensorflow><neural-network>","CreationDate":"2017-01-31T22:25:54.450","AcceptedAnswerId":null,"OwnerUserId":5368816.0,"Title":"Attempting to use uninitialized value Variable in Tensorflow ( sess.run(tf.global_variables_initializer()) is used!)","Body":"<p>I try to divide my neural network model and restore() function with setting random weights to zero.\nHere's the model code: <a href=\"http:\/\/pastebin.com\/TqN6kkeb\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/TqN6kkeb<\/a>\n(it works properly).<\/p>\n\n<p>And here's the function:<\/p>\n\n<pre><\/pre>\n\n<p>The problems are:\n1) is that it keeps writing me \nFailedPreconditionError (see above for traceback): Attempting to use uninitialized value Variable in this line: <\/p>\n\n<pre><\/pre>\n\n<p>no matter what. I have read these topics: <a href=\"https:\/\/stackoverflow.com\/questions\/40975371\/prettytensor-attempting-to-use-uninitialized-value\">Prettytensor: Attempting to use uninitialized value<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/34935464\/update-a-subset-of-weights-in-tensorflow\">Update a subset of weights in TensorFlow<\/a> (and many others), but advices from there didn't help to fix my bug.\nAnd I don't understand, what's the problem with the initialization as long as I run tf.global_variables_initializer();<\/p>\n\n<p>2) all of the weights seem to be setting to zero instead of the half, and I can't understand why.<\/p>\n\n<p>Please, help, I really stuck.<\/p>\n","answers":[{"AnswerId":"43661639","CreationDate":"2017-04-27T15:11:47.893","ParentId":null,"OwnerUserId":"1226095","Title":null,"Body":"<p>Just for the record (and others finding this post), the method name has changed, as per the page here: <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/how_tos\/variables\/#initialization\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.10\/how_tos\/variables\/#initialization<\/a><\/p>\n\n<p>you should run the <code>initialize_all_variables()<\/code> method like this:<\/p>\n\n<pre><code>import tensorflow as tf\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n<\/code><\/pre>\n"}]}
{"QuestionId":41968751,"AnswerCount":1,"Tags":"<tensorflow><keras><convolution>","CreationDate":"2017-01-31T22:35:22.990","AcceptedAnswerId":41968877.0,"OwnerUserId":288609.0,"Title":"Regarding the way to setup Activation and BatchNormalization in Keras","Body":"<p>I once saw the convolutional layers are defined as follows in <\/p>\n\n<pre><\/pre>\n\n<p>It seems to me that  can be also setup in  as well. What are the reasons to setup  in an independent layer as the above? Moreover, between  and , there is , can we move  after ? Will there be any difference then?<\/p>\n","answers":[{"AnswerId":"41968877","CreationDate":"2017-01-31T22:44:53.487","ParentId":null,"OwnerUserId":"7287271","Title":null,"Body":"<p>It's not a settled matter. Most people have batch normalization before the activation function. However, in practice some have found that having batch normalization after the activation function also works, and as far as I know there is no consensus on what's better. I suggest you try both on your particular problem and data. Don't expect any major differences however.<\/p>\n"}]}
{"QuestionId":41969081,"AnswerCount":1,"Tags":"<python><numpy><neural-network><keras><conv-neural-network>","CreationDate":"2017-01-31T23:00:36.710","AcceptedAnswerId":41969286.0,"OwnerUserId":5051564.0,"Title":"How to change 1D numpy array from keras layer output to image (3D numpy array)","Body":"<p>I have the output or feature maps of a keras layer, but how can I convert that into images (3D numpy array) that I can display.<\/p>\n\n<pre><\/pre>\n\n<p>Any comment or suggestion is greatly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"41969286","CreationDate":"2017-01-31T23:20:42.100","ParentId":null,"OwnerUserId":"2605604","Title":null,"Body":"<p>You could add each feature map as a subplot while iterating over each one:<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nfrom pylab import cm\n\nm = np.random.rand(1,224,224,64)\n\nfig = plt.figure()\nfig.suptitle(\"Feature Maps\")\n\nfor j in range(m.shape[3]):\n    ax = fig.add_subplot(8, 8, j+1)\n    ax.matshow(m[0,:,:,j], cmap=cm.gray)\n    plt.xticks(np.array([]))\n    plt.yticks(np.array([]))\n\nplt.show()\n<\/code><\/pre>\n\n<p>Which will give you something that looks like this (just noise in my case):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/v22Zo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/v22Zo.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":41969388,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><computer-vision><deep-learning><caffe>","CreationDate":"2017-01-31T23:30:07.210","AcceptedAnswerId":null,"OwnerUserId":288609.0,"Title":"Regarding two convolutional layers placed following each other in SegNet architecture","Body":"<p>In <a href=\"https:\/\/arxiv.org\/pdf\/1511.00561.pdf\" rel=\"nofollow noreferrer\">SegNet<\/a>, the architecture proposed by authors is shown as follows. <a href=\"https:\/\/i.stack.imgur.com\/DhAjw.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DhAjw.jpg\" alt=\"enter image description here\"><\/a>.<\/p>\n\n<p>What confuses me is that there are two convolutional layers following with each other in each building block, just as what shown in the figure as 1 and 2. What are the major motivations to place convolution layers this way instead of aggregating them into a single convolutional layer?<\/p>\n","answers":[{"AnswerId":"42434106","CreationDate":"2017-02-24T08:38:01.237","ParentId":null,"OwnerUserId":"804643","Title":null,"Body":"<p>SegNet uses the 13 convolutional layers from VGG. (2+2+3+3+3)<\/p>\n\n<p>Check <a href=\"http:\/\/ethereon.github.io\/netscope\/#\/preset\/vgg-16\" rel=\"nofollow noreferrer\">this visualization<\/a> and <a href=\"https:\/\/arxiv.org\/abs\/1409.1556\" rel=\"nofollow noreferrer\">the paper<\/a> for more information.<\/p>\n\n<p>From the paper:<\/p>\n\n<blockquote>\n  <p>It is easy to see that a stack of two 3\u00d73 conv. layers (without spatial pooling in between) has an effective receptive field of 5\u00d75 such layers have a 7 \u00d7 7 effective receptive field. So what have we gained by using, for instance, a stack of three 3\u00d73 conv. layers instead of a single 7\u00d77 layer? First, we incorporate three non-linear rectification layers instead of a single one, which makes the decision function more discriminative. Second, we decrease the number of parameters: assuming that both the input and the output of a three-layer 3 \u00d7 3 convolution stack has C channels, the stack is parametrised by <a href=\"https:\/\/i.stack.imgur.com\/1ikBj.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1ikBj.gif\" alt=\"enter image description here\"><\/a> weights; at the same time, a single 7 \u00d7 7 conv. layer would require <a href=\"https:\/\/i.stack.imgur.com\/5TImF.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/5TImF.gif\" alt=\"enter image description here\"><\/a> parameters, i.e. 81% more. This can be seen as imposing a regularisation on the 7 \u00d7 7 conv. filters, forcing them to have a decomposition through the 3 \u00d7 3 filters (with non-linearity injected in between).<\/p>\n<\/blockquote>\n"},{"AnswerId":"41973502","CreationDate":"2017-02-01T06:45:09.627","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you look at the legend at the bottom of the figure you attached, you'll see that in this <code>SegNet<\/code> illustration the blue layer stands for \"Conv + BatchNormalization + ReLU\": That is there is a non-linear activation, <code>\"ReLU\"<\/code> between the two linear convolutions.<\/p>\n\n<p>Regardless of this example, one might want to place two linear units one on top of the other <strong>without<\/strong> any non-linearity in order to explicitly control\/regularize the rank of the linear operation. See, for example, <a href=\"https:\/\/stackoverflow.com\/a\/40481789\/1714410\">how to reduce dimensionality of a fully connected layer using SVD trick<\/a>.<\/p>\n"}]}
{"QuestionId":41969810,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-02-01T00:08:25.970","AcceptedAnswerId":41973731.0,"OwnerUserId":4868654.0,"Title":"How to make a non-static Caffe network architecture?","Body":"<p>I would like to implement a neural network architecture in Caffe which will perform differently based on some iterable variable. For example: the full network might use 10 layers for 4 out of 5 training or testing iterations, but for all other iterations it will truncate the network and only use the last 5 layers. This would require that the input to the first layer and the input to the 5th layer have the same dimensionality of course, but my primary question is how to implement this switching between the two architectures during training\/testing.<\/p>\n","answers":[{"AnswerId":"41973731","CreationDate":"2017-02-01T06:59:23.320","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I guess you can do that using pycaffe and <code>caffe.NetSpec()<\/code>, but this is not going to be a very nice code...<br>\nOn the other hand, why don't you train for <code>i<\/code> iterations the full net, save a snapshot, and then \"warm start\" the reduced model with the snapshot you saved?<br>\nThat is: have <code>'full_trainval.prototxt'<\/code> with <code>'full_solver.prototxt'<\/code> configured to train the full net for <code>i<\/code> iterations, and <code>'top_trainval.prototxt'<\/code> with <code>'top_solver.prototxt'<\/code> configured to train only the top layers of the net. Then<\/p>\n\n<pre><code>~$ $CAFFE_ROOT\/build\/tools\/caffe train -solver full_solver.prototxt\n<\/code><\/pre>\n\n<p>When this stage is through, make sure you have the final sanpshot saved, and then  <\/p>\n\n<pre><code>~$ $CAFFE_ROOT\/build\/tools\/caffe train -solver top_solver.prototxt -snapshot full_train_last_snapshot.solverstate\n<\/code><\/pre>\n\n<p>Finally, you could use <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow noreferrer\"><code>net_surgery<\/code><\/a> to merge the weights of the two phases into a single full net.<\/p>\n"}]}
{"QuestionId":41970675,"AnswerCount":2,"Tags":"<tensorflow><bit-manipulation>","CreationDate":"2017-02-01T01:51:44.383","AcceptedAnswerId":41972138.0,"OwnerUserId":1452257.0,"Title":"Tensorflow bit-wise NOT operation for floats","Body":"<p>I have a Tensorflow boolean vector that I need to bit-wise invert such that all the values that are 0 (False) becomes 1 (True) and vice-versa. <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/api_docs\/python\/control_flow_ops\/logical_operators\" rel=\"nofollow noreferrer\">tf.logical_not<\/a> exists, but it only takes Boolean vectors. Does there exist a Tensorflow that does this on a vector of ints or floats so I don't need to recast my tensor?<\/p>\n","answers":[{"AnswerId":"41972138","CreationDate":"2017-02-01T04:52:25.863","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow doesn't include a specific operator for performing bitwise NOT, but you could simulate it with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/casting#bitcast\" rel=\"nofollow noreferrer\"><code>tf.bitcast()<\/code><\/a> and arithmetic operations on integers. For example, the following will invert the bits in a single floating-point value:<\/p>\n\n<pre><code>input_value = tf.constant(37.0)\n\nbitcast_to_int32 = tf.bitcast(input_value, tf.int32)\n\ninvert_bits = tf.constant(-1) - bitcast_to_int32\n\nbitcast_to_float = tf.bitcast(invert_bits, tf.float32)\n<\/code><\/pre>\n\n<p>You can confirm that the input and output are bitwise inverses using <a href=\"https:\/\/stackoverflow.com\/a\/16444778\/3574081\">this answer<\/a> to convert the floating-point values to binary.<\/p>\n"},{"AnswerId":"41970945","CreationDate":"2017-02-01T02:28:05.143","ParentId":null,"OwnerUserId":"1611389","Title":null,"Body":"<p>The only way I see is to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/casting#cast\" rel=\"nofollow noreferrer\"><code>tf.cast<\/code><\/a> the tensor to <code>tf.bool<\/code> then invoke <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/logical_operators#logical_not\" rel=\"nofollow noreferrer\"><code>tf.logical_not<\/code><\/a> then cast back.<\/p>\n\n<pre><code>dtype = tensor.dtype\nbtensor = tf.cast(tensor, tf.bool)\nnot_tensor = tf.logical_not(btensor)\ntensor = tf.cast(not_tensor, dtype)\n<\/code><\/pre>\n"}]}
{"QuestionId":41971587,"AnswerCount":4,"Tags":"<python><keras><keras-layer><sequence-to-sequence>","CreationDate":"2017-02-01T03:51:33.073","AcceptedAnswerId":null,"OwnerUserId":996366.0,"Title":"How to convert predicted sequence back to text in keras?","Body":"<p>I have a sequence to sequence learning model which works fine and able to predict some outputs. The problem is I have no idea how to convert the output back to text sequence.<\/p>\n\n<p>This is my code.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43927939","CreationDate":"2017-05-12T01:28:22.013","ParentId":null,"OwnerUserId":"232371","Title":null,"Body":"<p>Here is a solution I found:<\/p>\n\n<pre><code>reverse_word_map = dict(map(reversed, tokenizer.word_index.items()))\n<\/code><\/pre>\n"},{"AnswerId":"53391718","CreationDate":"2018-11-20T11:10:53.627","ParentId":null,"OwnerUserId":"1652219","Title":null,"Body":"<p>I had to resolve the same problem, so here is how I ended up doing it (inspired by @Ben Usemans reversed dictionary).<\/p>\n\n<pre><code># Importing library\nfrom keras.preprocessing.text import Tokenizer\n\n# My texts\ntexts = ['These are two crazy sentences', 'that I want to convert back and forth']\n\n# Creating a tokenizer\ntokenizer = Tokenizer(lower=True)\n\n# Building word indices\ntokenizer.fit_on_texts(texts)\n\n# Tokenizing sentences\nsentences = tokenizer.texts_to_sequences(texts)\n\n&gt;sentences\n&gt;[[1, 2, 3, 4, 5], [6, 7, 8, 9, 10, 11, 12, 13]]\n\n# Creating a reverse dictionary\nreverse_word_map = dict(map(reversed, tokenizer.word_index.items()))\n\n# Function takes a tokenized sentence and returns the words\ndef sequence_to_text(list_of_indices):\n    # Looking up words in dictionary\n    words = [reverse_word_map.get(letter) for letter in list_of_indices]\n    return(words)\n\n# Creating texts \nmy_texts = list(map(sequence_to_text, sentences))\n\n&gt;my_texts\n&gt;[['these', 'are', 'two', 'crazy', 'sentences'], ['that', 'i', 'want', 'to', 'convert', 'back', 'and', 'forth']]\n<\/code><\/pre>\n"},{"AnswerId":"53451482","CreationDate":"2018-11-23T18:30:05.777","ParentId":null,"OwnerUserId":"10226602","Title":null,"Body":"<p>You can use directly the inverse <code>tokenizer.sequences_to_texts<\/code> function. <\/p>\n\n<p><code>text = tokenizer.sequences_to_texts(&lt;list of the integer equivalent encodings&gt;)<\/code><\/p>\n\n<p>I have tested the above and it works as expected. <\/p>\n\n<p>PS.: Take extra care to make the argument be the list of the integer encodings and not the One Hot ones.<\/p>\n"},{"AnswerId":"43928501","CreationDate":"2017-05-12T02:47:42.973","ParentId":null,"OwnerUserId":"3626961","Title":null,"Body":"<p>You can make the dictionary that map index back to character.<\/p>\n\n<pre><code>index_word = {v: k for k, v in tk.word_index.items()} # map back\nseqs = tk.texts_to_sequences(txt1)\nwords = []\nfor seq in seqs:\n    if len(seq):\n        words.append(index_word.get(seq[0]))\n    else:\n        words.append(' ')\nprint(''.join(words)) # output\n\n&gt;&gt;&gt; 'what makes this problem difficult is that the sequences can vary in length  \n&gt;&gt;&gt; be comprised of a very large vocabulary of input symbols and may require the model  \n&gt;&gt;&gt; to learn the long term context or dependencies between symbols in the input sequence '\n<\/code><\/pre>\n\n<p>However, in the question, you're trying to use sequence of characters to predict output of 10 classes which is not the sequence to sequence model. In this case, you cannot just turn prediction (or <code>pred.argmax(axis=1)<\/code>) back to sequence of characters.<\/p>\n"}]}
{"QuestionId":41971793,"AnswerCount":1,"Tags":"<image-processing><computer-vision><deep-learning><caffe>","CreationDate":"2017-02-01T04:15:48.377","AcceptedAnswerId":41987381.0,"OwnerUserId":3755376.0,"Title":"Regarding to backward of convolution layer in Deep learning","Body":"<p>I understood the way to compute the forward part in Deep learning. Now, I want to understand the backward part. Let's take  as an example. The backward at the position  can compute as the figure bellow<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Jjn7L.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Jjn7L.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My question is that where is  (such as ,...) in the formula? How to compute it at the first iteration?<\/p>\n","answers":[{"AnswerId":"41987381","CreationDate":"2017-02-01T18:48:24.143","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p><strong>SHORT ANSWER<\/strong><\/p>\n\n<p>Those terms are in the final expansion at the bottom of the slide; they contribute to the summation for dE\/dX(2,2).  In your first back-propagation, you start at the end and work backwards (hence the name) -- and the <strong>Y<\/strong> values are the ground-truth labels.  So much for computing them.  :-)<\/p>\n\n<p><strong>LONG ANSWER<\/strong><\/p>\n\n<p>I'll keep this in more abstract, natural-language terms.  I'm hopeful that the alternate explanation will help you see the large picture as well as sorting out the math.<\/p>\n\n<p>You start the training with assigned weights that may or may not be at all related to the ground truth (labels).  You move blindly forward, making predictions at each layer based on naive faith in those weights.  The Y(i,j) values are the resulting meta-pixels from that faith.<\/p>\n\n<p>Then you hit the labels at the end.  You work backward, adjusting each weight.  Note that, at the last layer, the <strong>Y<\/strong> values are the ground-truth labels.<\/p>\n\n<p>At each layer, you mathematically deal with two factors:<\/p>\n\n<ol>\n<li>How far off was this prediction?<\/li>\n<li>How heavily did this parameter contribute to that prediction?<\/li>\n<\/ol>\n\n<p>You adjust the X-to-Y weight by \"off * weight * learning_rate\".\nWhen you complete that for layer N, you back up to layer N-1 and repeat.<\/p>\n\n<p><strong>PROGRESSION<\/strong><\/p>\n\n<p>Whether you initialize your weights with fixed or random values (I generally recommend the latter), you'll notice that there's really not much progress in the early iterations.  Since this is slow adjustment from guess-work weights, it takes several iterations to get a glimmer of useful learning into the last layers.  The first layers are still cluelessly thrashing at this point.  The loss function will bounce around close to its initial values for a while.  For instance, with GoogLeNet's image recognition, this flailing lasts for about 30 epochs.<\/p>\n\n<p>Then, finally, you get some valid learning in the latter layers, the patterns stabilize enough that some consistency percolates back to the early layers.  At this point, you'll see the loss function drop to a \"directed experimentation\" level.  From there, the progression depends a lot on the paradigm and texture of the problem: some have a sharp drop, then a gradual convergence; others have a more gradual drop, almost an exponential decay to convergence; more complex topologies have additional sharp drops as middle or early phases \"get their footing\".<\/p>\n"}]}
{"QuestionId":41971887,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-01T04:26:23.053","AcceptedAnswerId":null,"OwnerUserId":7498429.0,"Title":"Saving the Tensorflow cifar10 graph to a protobuf file","Body":"<p>I've been going through the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">Tensorflow CNN tutorial<\/a>, which uses the cifar10 dataset. Now, I've been trying to modify the given cifar10 code (which is a little out of date) to output a protobuf file that I'll be able to use in a small c++ image recognition program - similar to the format of the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_recognition\/\" rel=\"nofollow noreferrer\">Image Recognition<\/a> tutorial. The overall aim is that I'll be able to create and save a graph from my own dataset.<\/p>\n\n<p>However, I've had no success. Most of my attempts have been based around the tensorflow freeze_graph.py, which is the only way I've been able to find that is meant to save the graph. However, I've had no success implementing this and have been met with a myriad of errors, such as:<\/p>\n\n<pre><\/pre>\n\n<p>(when calling freeze_graph.py with the cifar10_train.py generated graph.pbtxt as the input graph, which I am not sure is even valid as an input)<\/p>\n\n<p>and also:<\/p>\n\n<pre><\/pre>\n\n<p>(when attempting to save my own checkpoints and input graph to act as inputs for freeze_graph.py).<\/p>\n\n<p>Any idea what I could possible be doing wrong?<\/p>\n\n<p>Am I even on the right track, or is there a simpler way to do this?<\/p>\n\n<p>UPDATE:<\/p>\n\n<p>I've had some success with the following in writing the protobuf file (read: it doesn't crash\/give errors):<\/p>\n\n<pre><\/pre>\n\n<p>However the pb file produced doesn't produce any tensors when used in my recognition code. Should this be producing a valid protobuf file?<\/p>\n","answers":[]}
{"QuestionId":41972015,"AnswerCount":1,"Tags":"<performance><theano><keras><lstm>","CreationDate":"2017-02-01T04:39:59.697","AcceptedAnswerId":49408944.0,"OwnerUserId":5345784.0,"Title":"Keras LSTM on CPU faster than GPU?","Body":"<p>I am testing LSTM networks on Keras and I am getting much faster training on CPU (5 seconds\/epoch on i2600k 16GB) than on GPU (35secs on Nvidia 1060 6GB). GPU utilisation runs at around 15%, and I never see it over 30% when trying other LSTM networks including the Keras examples. When I run other types of networks MLP and CNN the GPU is much faster. I am using the latest theano 0.9.0dev4 and keras 1.2.0 <\/p>\n\n<p>The sequence has 50,000 timesteps with 3 inputs (ints).<\/p> <p> If the inputs are descending (3,2,1) the output is 0, and 1 if ascending, except if the last two were also ascending, then the output is 0 instead of 1. <\/p><\/p>\n\n<p> After 250 epochs I get 99.97% accuracy, but why is the GPU so much slower? am I doing something wrong in the model? I tried various batch settings and still had the same issue.<\/p>\n\n<pre><\/pre>\n\n<p>Any thoughts? Thank you!<\/p>\n","answers":[{"AnswerId":"49408944","CreationDate":"2018-03-21T14:26:43.000","ParentId":null,"OwnerUserId":"5339857","Title":null,"Body":"<p>Use Keras' <code>CuDNNLSTM<\/code> cells for accelerated compute on Nvidia GPUs: <a href=\"https:\/\/keras.io\/layers\/recurrent\/#cudnnlstm\" rel=\"noreferrer\">https:\/\/keras.io\/layers\/recurrent\/#cudnnlstm<\/a><\/p>\n\n<p>It's simply changing the <code>LSTM<\/code> line to:<\/p>\n\n<pre><code>model.add(CuDNNLSTM(20, input_shape=(50000,3), return_sequences=True))\n<\/code><\/pre>\n"}]}
{"QuestionId":41972573,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><convolution><lstm><recurrent-neural-network>","CreationDate":"2017-02-01T05:33:17.857","AcceptedAnswerId":null,"OwnerUserId":5543916.0,"Title":"Tensorflow: jointly training CNN + LSTM","Body":"<p>There are quite a few examples on how to use LSTMs alone in TF, but I couldn't find any good examples on how to train CNN + LSTM jointly. \nFrom what I see, it is not quite straightforward how to do such training, and I can think of a few options here:<\/p>\n\n<ul>\n<li>First, I believe the simplest solution (or the most primitive one) would be to train CNN independently to learn features and then to train LSTM on CNN features without updating the CNN part, since one would probably have to extract and save these features in numpy and then feed them to LSTM in TF. But in that scenario, one would probably have to use a differently labeled dataset for pretraining of CNN, which eliminates the advantage of end to end training, i.e. learning of features for final objective targeted by LSTM (besides the fact that one has to have these additional labels in the first place). <\/li>\n<li>Second option would be to concatenate all time slices in the batch\ndimension (4-d Tensor), feed it to CNN then somehow repack those\nfeatures to 5-d Tensor again needed for training LSTM and then apply a cost function. My main concern, is if it is possible to do such thing. Also, handling variable length sequences becomes a little bit tricky. For example, in prediction scenario you would only feed single frame at the time. Thus, I would be really happy to see some examples if that is the right way of doing joint training. Besides that, this solution looks more like a hack, thus, if there is a better way to do so, it would be great if someone could share it.<\/li>\n<\/ul>\n\n<p>Thank you in advance !<\/p>\n","answers":[{"AnswerId":"42841913","CreationDate":"2017-03-16T18:16:47.003","ParentId":null,"OwnerUserId":"7643996","Title":null,"Body":"<p>For joint training, you can consider using tf.map_fn as described in the documentation <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/map_fn\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/map_fn<\/a>.<\/p>\n\n<p>Lets assume that the CNN is built along similar lines as described here <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10.py<\/a>.<\/p>\n\n<pre><code>def joint_inference(sequence):\n    inference_fn = lambda image: inference(image)\n    logit_sequence = tf.map_fn(inference_fn, sequence, dtype=tf.float32, swap_memory=True)\n    lstm_cell = tf.contrib.rnn.LSTMCell(128)\n    output_state, intermediate_state = tf.nn.dynamic_rnn(cell=lstm_cell, inputs=logit_sequence)\n    projection_function = lambda state: tf.contrib.layers.linear(state, num_outputs=num_classes, activation_fn=tf.nn.sigmoid)\n    projection_logits = tf.map_fn(projection_function, output_state)\n    return projection_logits\n<\/code><\/pre>\n\n<p>Warning: You might have to look into device placement as described here <a href=\"https:\/\/www.tensorflow.org\/tutorials\/using_gpu\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/using_gpu<\/a> if your model is larger than the memory gpu can allocate.<\/p>\n\n<p>An Alternative would be to flatten the video batch to create an image batch, do a forward pass from CNN and reshape the features for LSTM.<\/p>\n"}]}
{"QuestionId":41974081,"AnswerCount":2,"Tags":"<python><tensorflow><scikit-learn><gpu>","CreationDate":"2017-02-01T07:22:20.760","AcceptedAnswerId":null,"OwnerUserId":3886579.0,"Title":"Using GPU to train model","Body":"<p>I have a little knowledge of using GPU to train model. \nI am using K-means from scikit-learn to train my model.\nSince my data is very large, is it possible to train this model using GPU to reduce computation time?\nor could you please suggest any methods to use GPU power?<\/p>\n\n<p>The other question is if I use TensorFlow to build the K-means as shown in this blog.<\/p>\n\n<p><a href=\"https:\/\/blog.altoros.com\/using-k-means-clustering-in-tensorflow.html\" rel=\"nofollow noreferrer\">https:\/\/blog.altoros.com\/using-k-means-clustering-in-tensorflow.html<\/a><\/p>\n\n<p>It will use GPU or not?<\/p>\n\n<p>Thank you in advance.<\/p>\n","answers":[{"AnswerId":"43086821","CreationDate":"2017-03-29T07:17:03.423","ParentId":null,"OwnerUserId":"4352606","Title":null,"Body":"<p>To check if your GPU supports CUDA: <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-gpus<\/a><\/p>\n\n<p>Scikit-learn hasn't supported CUDA so far. You may want to use TensorFlow: <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_linux<\/a><\/p>\n\n<p>I hope this helps.<\/p>\n"},{"AnswerId":"44824673","CreationDate":"2017-06-29T12:05:28.713","ParentId":null,"OwnerUserId":"3323084","Title":null,"Body":"<p>If you have CUDA enabled GPU with Compute Capability 3.0 or higher and install GPU supported version of Tensorflow, then it will definitely use GPU for training.<\/p>\n\n<p>For additions information on NVIDIA requirements to run TensorFlow with GPU support check the following link: <\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#nvidia_requirements_to_run_tensorflow_with_gpu_support\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_linux#nvidia_requirements_to_run_tensorflow_with_gpu_support<\/a><\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41974579,"AnswerCount":1,"Tags":"<grid><tensorflow><lstm><chatbot><bidirectional>","CreationDate":"2017-02-01T07:57:10.963","AcceptedAnswerId":null,"OwnerUserId":7497449.0,"Title":"Implement BidirectionalGridLSTM","Body":"<p>I\u2019m implementing a chatbot using Tensorflow\u2019s seq2seq model[1], feeding it with data from the Ubuntu Dialogue Corpus. I want to compare an RNN using standard LSTM cells with Grid LSTM cells described in Kalchbrenner et al [2]. <\/p>\n\n<p>I\u2019m trying to implement the Grid LSTM cell in the translation model described in section 4.4 [2], but I\u2019m struggling with the bidirectional part. <\/p>\n\n<p>I have tried using BidirectionalGridLSTMCell, but I\u2019m not sure what they mean by num_frequency_block. They do not mention that in the paper. Does anyone know what they mean by num_frequency_block? In the api docs it says:\n: [required] A list of frequency blocks needed to cover the whole input feature splitting defined by  and .<\/p>\n\n<p>Further, I have tried to create my own cell. First I do the forward processing with the inputs, then I reverse the inputs, and do the backward processing. But when I concatenate these results, the shape changes. E.g. when I try to run the network with a batch size of 32, then i get this error:\nValueError: Dimensions must be equal, but are 64 and 32\nHow can I concatenate the results without changing the shape? Is that even possible?<\/p>\n\n<p>Does anyone have any other tips, on how I can implement Bidirectional Grid LSTM?<\/p>\n\n<p>[1] <a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/<\/a>\n[2] <a href=\"https:\/\/arxiv.org\/abs\/1507.01526\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1507.01526<\/a><\/p>\n","answers":[{"AnswerId":"42709493","CreationDate":"2017-03-10T03:09:41.467","ParentId":null,"OwnerUserId":"2233648","Title":null,"Body":"<p>-tensorflow has bidirectional LSTMs built-in: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3_NeuralNetworks\/bidirectional_rnn.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/notebooks\/3_NeuralNetworks\/bidirectional_rnn.ipynb<\/a><\/p>\n\n<ul>\n<li>here's a tutorial for using bidirectional LSTMs for intent matching: <a href=\"https:\/\/blog.themusio.com\/2016\/07\/18\/musios-intent-classifier-2\/\" rel=\"nofollow noreferrer\">https:\/\/blog.themusio.com\/2016\/07\/18\/musios-intent-classifier-2\/<\/a><\/li>\n<\/ul>\n\n<p>you're missing your second [2] reference link.<\/p>\n\n<p>is this a helpful baseline, even if they don't provide grids? \nmay i ask what you are using it for?<\/p>\n"}]}
{"QuestionId":41975060,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-02-01T08:29:30.220","AcceptedAnswerId":null,"OwnerUserId":7499265.0,"Title":"Why my tensorflow example code training result is increasing?","Body":"<p>Hi I am learning tensorflow.\nThis is my code, a simple multivariable tensorflow example.\nRunning environment is Python3.5.3, Tensorflow 0.12.1, Windows7.<\/p>\n\n<pre><\/pre>\n\n<p>I expected the result will decrease as training loops.<\/p>\n\n<p>but it increases infinitely.<\/p>\n\n<p>Same code on one variable worked well and it decreased.<\/p>\n\n<p>I don't know why 2-variable increases...<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41978272","CreationDate":"2017-02-01T11:09:36.203","ParentId":null,"OwnerUserId":"3883940","Title":null,"Body":"<p>The first solution I've found was to decrease the learning rate to 0.01. It seems like the steps are changing your parameters too radically. Probably if you'd use some kind of a regularization technique (Like L2) this wouldn't happen.<\/p>\n\n<p>Second, your code need some refinement. Use tensorflow matrix operations and initialize the bias with zero. Strangely when using TF functions for operations even the 0.1 learning rate is okay.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\n# Input data &amp; output data\nx1_data = [1.0, 0.0, 3.0, 0.0, 5.0]\nx2_data = [0.0, 2.0, 0.0, 4.0, 5.0]\ny_data =  [1.0, 2.0, 3.0, 4.0, 5.0]\n\ninput_X = tf.Variable(np.row_stack((x1_data, x2_data)).astype(np.float32))\nW = tf.Variable(tf.random_uniform([1,2], -1.0, 1.0))\nb = tf.Variable(tf.zeros([1,1]))\n\n# Our hypothesis\nhypothesis = tf.add(tf.matmul(W,input_X),b)\n# Simplified cost function\ncost = tf.reduce_mean(tf.square(hypothesis - y_data))\n\n# Minimize\na = tf.Variable(0.1) # Learning Rate\noptimizer = tf.train.GradientDescentOptimizer(a)\ntrain = optimizer.minimize(cost)\n\n# Initialise\ninit = tf.global_variables_initializer()\n\n# Launch\nsess = tf.Session()\nsess.run(init)\n\n# Train loop\nfor step in range(10):\n    sess.run(train)\n    print(step, sess.run(cost), sess.run(W), sess.run(b))\n<\/code><\/pre>\n"}]}
{"QuestionId":41975983,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning><keras><openai-gym>","CreationDate":"2017-02-01T09:19:46.583","AcceptedAnswerId":41979384.0,"OwnerUserId":3139545.0,"Title":"Multiple objectives regression with bounded output in Keras","Body":"<p>Im trying to use Keras to solve the following OpenAi gym <a href=\"https:\/\/gym.openai.com\/envs\/Humanoid-v1\" rel=\"nofollow noreferrer\">environment<\/a>.\nIt uses  for input and then it uses  with the . All the examples I have found online uses much simpler output layers with a single objective and no bounds.<\/p>\n\n<p>My questions are:<\/p>\n\n<ol>\n<li>Do I need any special functions since the outputs are bounded?<\/li>\n<li>Are there any example on how to constuct such a output layer with Keras?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"41979384","CreationDate":"2017-02-01T12:05:18.463","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>I would use mean squared error for such a task. <\/p>\n\n<p>However you can define your own metric to be optimized against. An example definition can be found here:: (<a href=\"https:\/\/github.com\/EdwardTyantov\/ultrasound-nerve-segmentation\/blob\/master\/metric.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/EdwardTyantov\/ultrasound-nerve-segmentation\/blob\/master\/metric.py<\/a>).<\/p>\n"}]}
{"QuestionId":41976484,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><deep-learning><reinforcement-learning><q-learning>","CreationDate":"2017-02-01T09:45:05.190","AcceptedAnswerId":null,"OwnerUserId":7499546.0,"Title":"Debugging Deep Q-Learning CNN","Body":"<p>I'm quite new to TF. I have written a Deep Q-Learning CNN to control a simple driving simulator. I've managed to plot the weights, biases and outputs of my fully connected layers on TensorBoard however, I'm not sure what I'm looking for. I've watched a youtube video that mentions a normally distributed set of weights is good. I was wondering if anyone has any other debugging tips. \nAlso, I'm using Relu for my activation functions but many of my 'rewards' and hence my 'qualities' output by the CNN are negative should I be using something like a Sigmoid instead? I'll grab some screenshots of my TensorBoard and post them too.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0rK4c.png\" rel=\"nofollow noreferrer\">The Sim<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/zZx6S.png\" rel=\"nofollow noreferrer\">What does it all mean Basil?<\/a><\/p>\n","answers":[]}
{"QuestionId":41977498,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><recurrent-neural-network>","CreationDate":"2017-02-01T10:34:01.577","AcceptedAnswerId":41978561.0,"OwnerUserId":1105579.0,"Title":"RNN is not overfitting on simple data","Body":"<p>I am trying to predict class for each number in input vector. There are 3 classes. Class 1, if input value changed from 0 to 1. Class 2, if it changed from 1 to 0. Class 0 otherwise.<\/p>\n\n<p>After second epoch and onward accuracy is stuck at 0.8824. Higher number of training epoch does not change anything. I have tried switching  to  or , this changes nothing. I also tried to generate longer input vectors and multiple batches, same without success.<\/p>\n\n<p>Only thing that helped is  increasing size of LSTM layers to 128, adding three  layers and  after each layer including LSTM. But it looks like overkill for such a simple problem and not giving perfect results anyway.<\/p>\n\n<p>I've spend more than a day trying to make it work. What could be a problem? \nThanks!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41978561","CreationDate":"2017-02-01T11:23:08.093","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You could tell that something was wrong because the loss value with your code was NaN after the first epoch. The problem is a common catch that has bitten me more than once in the past, which is that the <code>fit<\/code> method takes a <code>batch_size<\/code> parameter that defaults to 32. Since you only have one example you cannot even fill one batch, a situation that apparently is not detected by Keras but will cause wrong loss calculation. So you just need to change it to:<\/p>\n\n<pre><code>model.fit(X_r, Y_cat, nb_epoch=10, batch_size=1)\n<\/code><\/pre>\n\n<p>Still, it might to take some effort to fit that data. It can be hard for a neural net to work with a single feature (maybe it would be easier if the input was one-hot encoded) and the data is tiny. Also, the bigger\/deeper the net the more it will take to fit the weights (even if it is supposed to be more powerful). But at least now you will be able to see how the loss decreases. I have been able to reach 100% accuracy with your example removing the second and third LSTM layer and training for around 250 epochs (of course if you had more and longer examples the number would surely be smaller).<\/p>\n"}]}
{"QuestionId":41978177,"AnswerCount":1,"Tags":"<tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-02-01T11:05:12.290","AcceptedAnswerId":null,"OwnerUserId":1736294.0,"Title":"Making sense of Inception V3 layers","Body":"<p>The last layers of the Inception V3 network include a 8x8x2048 \"mixed10\" layer followed by a 1x1x2048 \"avg_pool\" layer. What is the real difference between these two layers ie. does the \"mixed10\" layer capture all the features of an image for example or is that only accomplished in the \"avg_pool\" layer? <\/p>\n","answers":[{"AnswerId":"41978930","CreationDate":"2017-02-01T11:42:06.037","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>As the name suggests, the avg_pool layer is obtained by, for each of the 2048 feature maps in mixed10, taking the global average (over the 64 values in the 8x8 feature map). Depending on your application, it can make sense to use either - if you want to use the network for classification and use the network as a feature extractor, using the average-pooled version and training a classifier on top makes sense.<\/p>\n"}]}
{"QuestionId":41978221,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-01T11:07:32.760","AcceptedAnswerId":null,"OwnerUserId":2324298.0,"Title":"Tensorflow - Next batch of data from tf.train.shuffle_batch","Body":"<p>I have a  file from which I am looking to create batches of data. I am using  to create a single batch. In my training I want to call the batches and pass them. And this is where I am stuck. I read that the  the poistion of the  gets saved in the state of the graph and the next example is read from the subsequent position. The trouble is I'm not able to figure how do I load the next batch. I'm using the below code to create the batches.<\/p>\n\n<pre><\/pre>\n\n<p>This gives me a single batch of data. I want to use the  paradigm to pass the batches for the training wherein on each iteration, a new batch gets passed in. How do I load these batches? Will calling the  along with  again call the next batch?<\/p>\n","answers":[{"AnswerId":"41982923","CreationDate":"2017-02-01T15:01:16.733","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>The <code>read_and_decode_single_example()<\/code> function creates a (sub-)graph for the network that is used to load data; you only call that once. It might be more appropriately called <code>build_read_and_decode_single_example_graph()<\/code>, but that's a bit long.<\/p>\n\n<p>The \"magic\" lies in evaluating (i.e. using) the <code>_batch<\/code> tensors multiple times, e.g.<\/p>\n\n<pre><code>batch_size = 100\n# ...\n\nwith tf.Session() as sess:\n    # get the first batch of 100 values\n    first_batch = sess.run([contexts_batch, context_lens_batch,\n                            utterances_batch, utterance_lens_batch,\n                            labels_batch])\n\n    # second batch of different 100 values\n    second_batch = sess.run([contexts_batch, context_lens_batch,\n                            utterances_batch, utterance_lens_batch,\n                            labels_batch])\n    # etc.\n<\/code><\/pre>\n\n<p>Of course, rather than fetching these values from a session manually, you would feed them into some other part of the network instead. The mechanism is the same: Whenever one of these tensors is fetched directly or indirectly, the batching mechanism will take care of providing you a new batch (of different values) each time.<\/p>\n"}]}
{"QuestionId":41979003,"AnswerCount":1,"Tags":"<scikit-learn><keras>","CreationDate":"2017-02-01T11:45:23.633","AcceptedAnswerId":null,"OwnerUserId":1768201.0,"Title":"How to access weights of regressor within Scikit-learn pipeline","Body":"<p>I have used Keras regressor to fit a regression to data. I used Scikit-learn wrapper and Pipeline to first standardize the data and then fit it on Keras regressor. Sort of like this:<\/p>\n\n<pre><\/pre>\n\n<p>I then Grid Search for best fit through GridSearchCv and get a best fit in a variable:<\/p>\n\n<pre><\/pre>\n\n<p>The clf variable has 2 process as defined in pipeline. \nMy question is how do I extract weights and biases of keras regressor for best fit (clf), through get_params function?:<\/p>\n\n<pre><\/pre>\n\n<p>I couldn't find good documentation for this.<\/p>\n","answers":[{"AnswerId":"54835949","CreationDate":"2019-02-22T22:16:26.057","ParentId":null,"OwnerUserId":"10041917","Title":null,"Body":"<p>weights = KerasRegressor.model.layers[0].get_weights()[0]\nbiases = KerasRegressor.model.layers[0].get_weights()[1]<\/p>\n"}]}
{"QuestionId":41981929,"AnswerCount":2,"Tags":"<deep-learning><caffe><conv-neural-network>","CreationDate":"2017-02-01T14:15:08.157","AcceptedAnswerId":null,"OwnerUserId":7500694.0,"Title":"caffe could not open or find file","Body":"<p>I'm new to caffe and after successfully running an example I'm trying to use my own data. However, when trying to either write my data into the lmdb data format or directly trying to use the solver, in both cases I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>The path is right, but it's weird that the label  is part of this error message. That implies that it's reading it as part of the path instead of as a label. The text file looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Is it because the labels are too high? Or maybe because the labels don't start with 0? I've searched for this error and all I found were examples with relatively few labels, about ~1-5, but I have about 4096 classes of which I don't always actually have examples in the training data. Maybe this is a problem, too (certainly for learning, at least, but I didn't expect for it to give me an actual error message). Usually, the label does not seem to be part of this error message.\nFor the creation of the lmdb  file, I use the  from the caffe examples. For solving, I use:<\/p>\n\n<pre><\/pre>\n\n<p>I tried different image data types, too: PNG, JPEG and BMP. So this isn't the culprit, either.\nIf it is really because of my choice of labels, what would be a viable workaround for this problem?<\/p>\n\n<p>Thanks a lot for your help!<\/p>\n","answers":[{"AnswerId":"43891424","CreationDate":"2017-05-10T11:39:01.523","ParentId":null,"OwnerUserId":"2194226","Title":null,"Body":"<p>I had the same issue. Check that lines in your text file don't have spaces in the end.<\/p>\n"},{"AnswerId":"43896084","CreationDate":"2017-05-10T14:55:08.463","ParentId":null,"OwnerUserId":"4115398","Title":null,"Body":"<p>I was facing a similar problem with <em>convert_imageset<\/em>. I have solved just removing the trailing spaces in the text file which contains the labels.<\/p>\n"}]}
{"QuestionId":41983214,"AnswerCount":3,"Tags":"<cmake><deep-learning><caffe><pycaffe>","CreationDate":"2017-02-01T15:15:15.247","AcceptedAnswerId":null,"OwnerUserId":2225050.0,"Title":"Compiling caffe","Body":"<p>I'm trying to compile caffe using cmake on a Linux machine.  I get this complaint when running cmake:<\/p>\n\n<pre><\/pre>\n\n<p>I later get this error<\/p>\n\n<pre><\/pre>\n\n<p>What am I doing wrong?  I tried setting an environment variable for PYTHON_LIBRARIES but that didn't help.  I'm using my locally installed version of python (2.7.13), so I think that might be causing the problem, but I don't know how to get past this.<\/p>\n","answers":[{"AnswerId":"41986953","CreationDate":"2017-02-01T18:23:57.527","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>Check carefully through the installation instructions.  There's a variety of dependencies to install (python-devel, glog-devel, ...) and some environment variables to set.  Using a local Python installation isn't a problem, as long as you have that properly in your paths.  Remember to check your PATH and LD_LIBRARY_PATH variables as well.<\/p>\n\n<p>You might check with<\/p>\n\n<pre><code>env | egrep -i \"py\"\n<\/code><\/pre>\n\n<p>to see what variables <em>do<\/em> have Python references.  If you suspect a version problem with an executable, use<\/p>\n\n<pre><code>ldd &lt;executable&gt;\n<\/code><\/pre>\n\n<p>To see what objects and versions get loaded.  Among other things, it's vaguely possible that something got compiled against an older Python in the system directories.<\/p>\n"},{"AnswerId":"42069247","CreationDate":"2017-02-06T13:41:20.060","ParentId":null,"OwnerUserId":"2225050","Title":null,"Body":"<p>I was able to \"fix it\" by editing the file build\/CMakeCache.txt and manually setting these<\/p>\n\n<pre><code>PYTHON_INCLUDE_DIR\nPYTHON_LIBRARY\n<\/code><\/pre>\n\n<p>I don't know if that's the correct way to resolve this. If not, what is the right way to resolve this?<\/p>\n"},{"AnswerId":"49365788","CreationDate":"2018-03-19T14:49:44.443","ParentId":null,"OwnerUserId":"344159","Title":null,"Body":"<p>I had the same error while installing caffe on Ubuntu 14.04; I solved this by installing boost-python:<\/p>\n\n<p>sudo apt-get install libboost-python-dev<\/p>\n"}]}
{"QuestionId":41983943,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-ml>","CreationDate":"2017-02-01T15:47:00.673","AcceptedAnswerId":null,"OwnerUserId":7397552.0,"Title":"tf.parse_example.features for Multi-label photo classification","Body":"<p>I am trying to apply the tutorial code from cloudml-samples\/flowers\/ on a set of photos with multi-label. Environment is Google Cloud Shell. \"preprocess\"ed all training and evaluation set. Running into error when started the training task. <\/p>\n\n<p>I called the trainer.task through python and returning the below error messages. Please let me know if the log info (only some generic information) through running gcloud beta ml command, or anything else will be helpful. <\/p>\n\n<p>Thanks in advance for taking the look.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42059753","CreationDate":"2017-02-06T02:56:07.567","ParentId":null,"OwnerUserId":"3818158","Title":null,"Body":"<p>The flowers sample does not directly support multi-label classification.  The exact problem in your code I believe is related to the <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/blob\/master\/flowers\/trainer\/model.py#L226\" rel=\"nofollow noreferrer\">shape=[1] in model.py<\/a>.<\/p>\n\n<p>Alas plugging in multi-label examples will not work out of the box with this code, which was designed to be a good example of how to get started with a complex workflow (ie restoring from checkpoints, transfer learning, embedding in Dataflow, etc), more so than an image modeling Swiss army knife.<\/p>\n\n<p>There should be only a few lines you would have to make changes to model.py to get your multi-label example working though.  Furthermore, I'm sure other members of the community would greatly appreciate any progress you make so please consider creating a <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/pulls\" rel=\"nofollow noreferrer\">pull request<\/a>, should you get a sample that does both multi and single-label problems!<\/p>\n"}]}
{"QuestionId":41984876,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-02-01T16:33:16.510","AcceptedAnswerId":null,"OwnerUserId":5776278.0,"Title":"Restore and evaluate a Tensorflow model","Body":"<p>I am already saving this model during the training, but I am having a hard time to load it and evaluate it.<\/p>\n\n<p>I tried some different approaches but I wasn't able to load the saved model and evaluate it to get its prediction over some test sample which is an image file.<\/p>\n\n<p>Can anyone help with that? as I saw it not seems to be so hard but I am miss anything to do it right. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42015398","CreationDate":"2017-02-03T01:39:12.360","ParentId":null,"OwnerUserId":"5776278","Title":null,"Body":"<p>Now I am loading it correctly now<\/p>\n\n<pre><code>saver = tf.train.import_meta_graph('model\/model.ckpt.meta')\n\ninit = tf.group(tf.initialize_all_variables(),\n                tf.initialize_local_variables())\nsess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True))\n\nsaver.restore(sess, 'model\/model.ckpt')\n<\/code><\/pre>\n\n<p>Now I am trying to get the prediction I think that is something like that, but I don't know how I will get the variables from the model which I created before in order to get its prediction:<\/p>\n\n<pre><code>prediction=tf.argmax(y_conv,1)\nprediction.eval(feed_dict={x: [imvalue],keep_prob: 1.0}, session=sess)\n<\/code><\/pre>\n"},{"AnswerId":"41985741","CreationDate":"2017-02-01T17:16:55.537","ParentId":null,"OwnerUserId":"6471140","Title":null,"Body":"<p>How are you saving it? Have you tried:\n(for saving)<\/p>\n\n<pre><code>saver = tf.train.Saver()\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nsaver.save(sess, 'my-model')\n<\/code><\/pre>\n\n<p>(for loading)<\/p>\n\n<pre><code>sess = tf.Session()\nnew_saver = tf.train.import_meta_graph('my-model.meta')\nnew_saver.restore(sess, tf.train.latest_checkpoint('.\/'))\n<\/code><\/pre>\n\n<p>The official reference for this: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops\/exporting_and_importing_meta_graphs\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops\/exporting_and_importing_meta_graphs<\/a> (or substitute a release number such as <code>r0.12<\/code> for <code>master<\/code> in the URL).<\/p>\n"}]}
{"QuestionId":41984881,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-01T16:33:26.530","AcceptedAnswerId":41985832.0,"OwnerUserId":6718080.0,"Title":"Unexpected behavior after updating a placeholder","Body":"<p>The output of the following snippet<\/p>\n\n<pre><\/pre>\n\n<p>is , and not  as I would expected.<\/p>\n\n<p>I've been answered that the  line overwrote the original placeholder so I'm not feeding what I think I'm feeding.<\/p>\n\n<p>However, I didn't fully understand this answer. If  was a variable then the output was  of course. So what am I feeding? What does  hold after the update op?<\/p>\n","answers":[{"AnswerId":"41985832","CreationDate":"2017-02-01T17:22:20.217","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>I think the confusion stems from the difference between TensorFlow variables and Python variables.  <\/p>\n\n<p>In TensorFlow, a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/placeholders#placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a> op holds no state, and there is no supported notion of \"updating\" it: think of it as the formal parameter to a function in other languages (e.g. <code>x<\/code> in <code>def f(x): ...<\/code> in Python). However, you can assign a <code>tf.placeholder()<\/code> to a <em>Python<\/em> variable so that you can use it elsewhere in your program.<\/p>\n\n<p>When you write <code>ph += 1000<\/code>, you are creating a new tensor whose value is 1000 greater than <code>ph<\/code> (it is shorthand for <code>ph = ph + 1000<\/code>), and then assigning it to the <em>Python<\/em> variable <code>ph<\/code>. An equivalent way to write your program is:<\/p>\n\n<pre><code>ph = tf.placeholder(tf.int32)\nnew_ph = ph + 1000\na = new_ph + 0\nprint(sess.run([a], {new_ph: 5}))\n<\/code><\/pre>\n\n<p>Note that you are actually feeding a value for the tensor <code>new_ph<\/code>&mdash;i.e. the <em>output<\/em> of the addition&mdash;so the operations that compute that addition are pruned from the graph and don't execute.<\/p>\n"}]}
{"QuestionId":41985060,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-02-01T16:42:18.410","AcceptedAnswerId":41985538.0,"OwnerUserId":223963.0,"Title":"How to Argsort in Tensorflow?","Body":"<p>How can I argsort a 25 x 5 x 5 matrix (tensor) along the 2nd axis? Essentially, I am looking for tensorflow's equivalent (function or methodology) to numpy's argsort, e.g. .<\/p>\n","answers":[{"AnswerId":"57081015","CreationDate":"2019-07-17T17:07:20.297","ParentId":null,"OwnerUserId":"3222797","Title":null,"Body":"<p>For reference, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/argsort\" rel=\"nofollow noreferrer\">tf.argsort<\/a> is now supported in Tensorflow.<\/p>\n\n<p><strong>Example<\/strong>:<\/p>\n\n<pre><code>import tensorflow as tf\n\n\ntensor = tf.constant(\n    [\n        [8, 7, 11],\n        [5, 3, 4],\n        [17, 33, 23],\n    ]\n)\n\narg_sort_op = tf.argsort(tensor, axis=-1)\n\nwith tf.Session() as sess:\n    out = sess.run(arg_sort_op)\n\nprint(out)\n\n#  [[1 0 2]\n#   [1 2 0]\n#   [0 2 1]]\n<\/code><\/pre>\n"},{"AnswerId":"41985538","CreationDate":"2017-02-01T17:06:11.443","ParentId":null,"OwnerUserId":"5903959","Title":null,"Body":"<p>In your case you could probably use <code>top_k<\/code> which returns the highest <code>k<\/code> values. <code>k<\/code> can be a 1D vector defining how many values to 'top' per dimensions. In your case, if you want the second axis set <code>k=[0, 5, 0]<\/code> might do it.<\/p>\n\n<pre><code>tf.nn.top_k(matrix, k=[0,5,0], sorted=True)\n<\/code><\/pre>\n\n<p>I didn't run it tho.\nHope this helps<\/p>\n"}]}
{"QuestionId":41985207,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-01T16:49:11.313","AcceptedAnswerId":null,"OwnerUserId":7251162.0,"Title":"Visualize influence of an input in TensorFlow","Body":"<p>Is there a way to visualize the influence of an Input image on the output of a net?\nI have a VGG net for example, and I use a dog image and get the correct label at the output. I want to be able to see what parts of the input image had the influence that gave correct output that correspond with what the network look for in a dog image.Is there a way to create a influence heat-map?\nI tried to work with  with no success. \nThanks<\/p>\n","answers":[]}
{"QuestionId":41985509,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><tiff>","CreationDate":"2017-02-01T17:05:02.017","AcceptedAnswerId":null,"OwnerUserId":2470337.0,"Title":"Tensorflow Machine Learning: No Decoder for TIFF Images?","Body":"<p>I have noticed the   package provides standard procedures for decoding ,  and  images after reading files. For instance for :<\/p>\n\n<pre><\/pre>\n\n<p>However, the  format decoder seems to be missing. <\/p>\n\n<p>So what solutions exist for  files? Surely, I could convert my input images to , but this doesn't seem to be a very smart solution.<\/p>\n","answers":[{"AnswerId":"41986552","CreationDate":"2017-02-01T17:59:56.670","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Correct, there's no decoder for TIFF images. Look in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/27711108b5fce2e1692f9440631a183b3808fa01\/tensorflow\/core\/kernels\" rel=\"noreferrer\">tensorflow\/core\/kernels<\/a> and you see<\/p>\n\n<pre><code>decode_csv_op.cc\ndecode_gif_op.cc\ndecode_jpeg_op.cc\ndecode_png_op.cc\ndecode_raw_op.cc\n<\/code><\/pre>\n\n<p>No <code>decode_tiff_op.cc<\/code>. This could be a good target for community contribution.<\/p>\n"}]}
{"QuestionId":41986542,"AnswerCount":1,"Tags":"<tensorflow><sparse-matrix><matrix-multiplication><numpy-einsum>","CreationDate":"2017-02-01T17:59:33.723","AcceptedAnswerId":41991924.0,"OwnerUserId":5511389.0,"Title":"What is the efficient way of multiplying chain of tensors in tensorflow","Body":"<p>I have 3 sparse tensors of dimensions A = P*N, B = Q*N and C = R*N. What is the efficient way to compute the product matrix A*B*C such that dimension of the product matrix is P*Q*R in tensorflow.? <\/p>\n\n<p>I have tried with tf.matmul and followed by tf.reshape but it won't give the product matrix with the dimension as specified above.<\/p>\n\n<p>Thanks. <\/p>\n","answers":[{"AnswerId":"41991924","CreationDate":"2017-02-01T23:43:19.367","ParentId":null,"OwnerUserId":"3128177","Title":null,"Body":"<p><code>tf.einsum<\/code> Should help you. Though I'm not really sure what you mean by <code>A*B*C<\/code> since dimensions are incompatible for matrix multiplication. Probably something like this:<\/p>\n\n<pre><code>R = tf.einsum('il,jl,kl-&gt;ijk', A, B, C)\n<\/code><\/pre>\n"}]}
{"QuestionId":41986583,"AnswerCount":1,"Tags":"<python><tensorflow><multi-gpu>","CreationDate":"2017-02-01T18:01:48.480","AcceptedAnswerId":42007153.0,"OwnerUserId":5895928.0,"Title":"Tensorflow multi gpu example error: Variable conv1\/weights\/ExponentialMovingAverage\/ does not exist","Body":"<p>I am running the code mentioned in this tutorial: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/<\/a> <\/p>\n\n<p>I downloaded the code from here: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/cifar10\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/image\/cifar10\/<\/a> <\/p>\n\n<p>I am running the code on a g2.4xlarge machine in AWS on ubuntu 14.04. The single gpu example runs fine without any errors.<\/p>\n\n<p>Can someone help how to fix this? I am running 0.12 version.<\/p>\n\n<hr>\n\n<p>ubuntu@ip-xxx-xx-xx-xx:~\/pythonworkspace\/tensorflowdev\/models-master\/tutorials\/image\/cifar10$ python -c 'import tensorflow as tf; print(tf.<strong>version<\/strong>)'<\/p>\n\n<h2>0.12.head<\/h2>\n\n<p>ubuntu@ip-xxx-xx-xx-xx:~\/pythonworkspace\/tensorflowdev\/models-master\/tutorials\/image\/cifar10$ python cifar10_multi_gpu_train.py --num_gpus=2<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42007153","CreationDate":"2017-02-02T16:17:10.380","ParentId":null,"OwnerUserId":"4550763","Title":null,"Body":"<p>you can find the answer to your problem here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6220\" rel=\"nofollow noreferrer\">Issue 6220<\/a><\/p>\n\n<p>You need to put:<br>\n<code>with tf.variable_scope(tf.get_variable_scope())<\/code><br>\nin front of the loop which runs over your devices ...  <\/p>\n\n<p>so, do that:  <\/p>\n\n<pre><code>with tf.variable_scope(tf.get_variable_scope()):\n    for i in xrange(FLAGS.num_gpus):\n        with tf.device('\/gpu:%d' % i):\n<\/code><\/pre>\n\n<p>The explanation is given in the link...<br>\nHere the quote:<\/p>\n\n<blockquote>\n  <p>When you do tf.get_variable_scope().reuse_variables() you set the\n  current scope to reuse variables. If you call the optimizer in such\n  scope, it's trying to reuse slot variables, which it cannot find, so\n  it throws an error. If you put a scope around, the\n  tf.get_variable_scope().reuse_variables() only affects that scope, so\n  when you exit it, you're back in the non-reusing mode, the one you\n  want.<\/p>\n  \n  <p>Hope that helps, let me know if I should clarify more.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":41986822,"AnswerCount":1,"Tags":"<python><generator><keras>","CreationDate":"2017-02-01T18:16:52.927","AcceptedAnswerId":41988349.0,"OwnerUserId":1694564.0,"Title":"How do I reference the epoch number within a Keras generator?","Body":"<p>I'm building a Python generator to be consumed by the Keras fit_generator function. I want to use the current epoch count in my generator to adjust a value. Is there any way to reference this number?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41988349","CreationDate":"2017-02-01T19:45:01.757","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>How about:<\/p>\n\n<pre><code>def generate_arrays_from_file(path, samples_per_epoch):\n    samples_produced_in_current_epoch = 0\n    epoch_number = 1\n    while 1:\n    f = open(path)\n    for line in f:\n        x, y = process_line(line)\n        x = x + epoch_number\n        samples_produced += x.shape[0]\n        if samples_produced_in_current_epoch &gt; samples_per_epoch:\n            epoch_number += 1\n            samples_produced_in_current_epoch = 0      \n        yield (x, y)\n\n    f.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":41986934,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2017-02-01T18:23:10.607","AcceptedAnswerId":42100764.0,"OwnerUserId":2251132.0,"Title":"Is it possible to get the hessian through a gamma function in theano?","Body":"<pre><\/pre>\n\n<p>gives an unnamed . Since the equivalent code works fine with e.g.  instead of  I guess theano just doesn't know the second derivation of ? <\/p>\n","answers":[{"AnswerId":"42100764","CreationDate":"2017-02-07T21:55:04.203","ParentId":null,"OwnerUserId":"949321","Title":null,"Body":"<p>The grad of gammaln use psi. But we didn't implement the grad of psi.<\/p>\n\n<p>You can implement the method grad() of the class Psi in the file theano\/scalar\/basic_scipy.py. This would make this case work.<\/p>\n\n<p>This page of the documentation explain how to add a new Op in Theano and explain how to implement the grad() method: <a href=\"http:\/\/deeplearning.net\/software\/theano\/extending\/extending_theano.html\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/extending\/extending_theano.html<\/a><\/p>\n\n<p>If you need help implementing this, using theano-dev mailing list is better suited then stackoverflow for ongoing discussion.<\/p>\n\n<p>If you do it, it would be a great contribution to Theano.<\/p>\n"}]}
{"QuestionId":41987105,"AnswerCount":0,"Tags":"<python><neural-network><deep-learning><theano>","CreationDate":"2017-02-01T18:33:38.160","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Theano Neural network implementation for a chess engine","Body":"<p><strong>Intro<\/strong><\/p>\n\n<p>Please bear with me as this will be a long post.<\/p>\n\n<p>Basically I am trying to implement a Chess engine using NN. My idea is to have a single layer NN trained based on millions of positions (I split the file, and each file consist of few thousand positions). The input of my NN is 64 nodes (1 for each square), and the value of the inputs are between , 13 pieces including no piece.<\/p>\n\n<p>I am trying to implement a single hidden layer NN using Theano. I read over the tutorial on Theano, and I have a rough understanding of NN (I took the ML course taught by Andrew Ng). I think it should be enough to accomplish what I want to do.<\/p>\n\n<p><strong>My input files<\/strong><\/p>\n\n<p>Each FEN file consist of few thousand positions. I used stockfish to analyse each position, and obtained a static evaluation score Y. After, I converted each FEN position into a  X vector like so: <\/p>\n\n<pre><\/pre>\n\n<p>I have 5 million of these.<\/p>\n\n<p><strong>My problem<\/strong> <\/p>\n\n<p>I am having a hard time actually implementing the training section of NN in Theano. Here is a snippet of my NN code (mostly its taken from the tutorial, if i use  then it means I have the same code as the tutorial <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html#logreg\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/tutorial\/logreg.html#logreg<\/a>).<\/p>\n\n<pre><\/pre>\n\n<p>I understand I am asking for a lot of help, and I thank you for being patient reading my post. But I will really appreciate it if someone can show me how this is done, or even point me towards the right direction. This project is not work related or school related, but it is still very important to me because I really want to see it through.<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":41987690,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-02-01T19:08:15.503","AcceptedAnswerId":42000594.0,"OwnerUserId":58634.0,"Title":"How to adjust labels with ImageDataGenerator in Keras?","Body":"<p>I want to augment my dataset with Keras's <a href=\"https:\/\/keras.io\/preprocessing\/image\/#imagedatagenerator\" rel=\"nofollow noreferrer\">ImageDataGenerator<\/a> for use with <a href=\"https:\/\/keras.io\/models\/sequential\/#sequential-model-methods\" rel=\"nofollow noreferrer\">model.fit_generator()<\/a>.  I see I can randomly flip images.  For flipped images, I need to modify the corresponding label.  How can I do that?<\/p>\n\n<p>EDIT: I'm doing regression, not classification, so if an image is flipped I need to adjust the label.  The actual images are from a self-driving car simulator, and the labels are the steering angles.  If I horizontally flip an image, I need to negate the steering angle.<\/p>\n","answers":[{"AnswerId":"42000594","CreationDate":"2017-02-02T11:06:25.840","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>You might do something like:<\/p>\n\n<pre><code>import numpy\n\ndef fliping_gen(image_generator, flip_p=0.5):\n    for x, y in image_generator:\n        flip_selector = numpy.random.binomial(1, flip_p, size=x.shape[0]) == 1\n        x[flip_selector,:,:,:] = x[flip_selector,:,::-1,:]\n        y[flip_selector] = (-1) * y[flip_selector]\n        yield x, y\n<\/code><\/pre>\n"}]}
{"QuestionId":41988036,"AnswerCount":1,"Tags":"<python><macos><installation><tensorflow>","CreationDate":"2017-02-01T19:26:51.020","AcceptedAnswerId":null,"OwnerUserId":1752609.0,"Title":"tensorflow ImportError \"@rpath\/libcudart.8.0.dylib\" Image Not Found","Body":"<p>Trying Tensorflow in Mac Sierra with cuda_gpu, python 2.7 from system default, no virtual environment.<\/p>\n\n<p>Encounter ImportError when doing basic test \"import tensorflow\"<\/p>\n\n<pre><\/pre>\n\n<p>I've tried the steps of \"local_config_cuda\" suggested by @norman_h from <a href=\"https:\/\/stackoverflow.com\/questions\/38710339\/library-not-loaded-rpath-libcudart-7-5-dylib-tensorflow-error-on-mac\">&#39;Library not loaded: @rpath\/libcudart.7.5.dylib&#39; TensorFlow Error on Mac<\/a>, but still not working.<\/p>\n\n<p>CUDA samples work, so it looks an issue of lib reference... can one shed me \nsome light? thanks!<\/p>\n","answers":[{"AnswerId":"42005879","CreationDate":"2017-02-02T15:15:42.553","ParentId":null,"OwnerUserId":"1752609","Title":null,"Body":"<p>Per above comment from @Yaroslav, disabling SIP resolves the issue immediately.<\/p>\n"}]}
{"QuestionId":41988267,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-01T19:40:32.500","AcceptedAnswerId":41988547.0,"OwnerUserId":4351454.0,"Title":"Simple Tensorflow example not working in Jupyter Notebook","Body":"<pre><\/pre>\n\n<p>I am running this Tensorflow example in Jupyter Notebook in a cell.\nWhen I run the code the first time it works fine. But, when I re-run the cell, it gives the following error, I have tried a lot but can't seem to figure out the error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41988547","CreationDate":"2017-02-01T19:57:42.307","ParentId":null,"OwnerUserId":"2299382","Title":null,"Body":"<p>The result from an optimization is 'None' so you can't print that. If you want to print the loss after an optimization step you can do<\/p>\n\n<pre><code>loss,_ = sess.run([loss,grads_and_vars]);\nprint(loss)\n<\/code><\/pre>\n"}]}
{"QuestionId":41989585,"AnswerCount":0,"Tags":"<python><parsing><tensorflow>","CreationDate":"2017-02-01T20:54:52.077","AcceptedAnswerId":null,"OwnerUserId":4040030.0,"Title":"Output Tensor Wrong Shape When Parsing TFRecord","Body":"<p>When I attempt to parse a TFRecord file I am getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>As far as I can tell the line error implies that the output tensor is the wrong shape. I have been following the information <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow noreferrer\">here<\/a> for the most part. Under the heading <strong>Standard TensorFlow format<\/strong> it lists two github links that both encode and parse the mnist dataset. In the file that performs the parsing there is no reference to setting the output tensors shape as far as I can tell. I'm not sure how to correct this and I have tried everything I can think of. I have included my code below. <\/p>\n\n<pre><\/pre>\n\n<p>--<\/p>\n\n<p>EDIT: So I figured out that I can get rid of the error and have things work by changing the line <\/p>\n\n<pre><\/pre>\n\n<p>to <\/p>\n\n<pre><\/pre>\n\n<p>This forces the shape of the output tensor. I'm still unclear why I need to specify this for the case I have presented while the example in the tensorflow directions does not. If someone could clarify I'd greatly appreciate it.<\/p>\n","answers":[]}
{"QuestionId":41990014,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2017-02-01T21:23:49.120","AcceptedAnswerId":41991989.0,"OwnerUserId":6174438.0,"Title":"load multiple models in Tensorflow","Body":"<p>I have written the following convolutional neural network (CNN) class in Tensorflow [<em>I have tried to omit some lines of code for clarity.<\/em>]<\/p>\n\n<pre><\/pre>\n\n<p>I have trained two CNNs for two separate tasks independently. Each took around 1 day. Say, model1 and model2  are saved as '' and '' (with their corresponding meta files) respectively. I can test each model and compute its performance separately.<\/p>\n\n<p>Now, I want to load these two models in a single script. I would naturally try to do as below:<\/p>\n\n<pre><\/pre>\n\n<p>I encounter the error [<em>The error message is long. I just copied\/pasted a snippet of it.<\/em>]<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to load from these two files two separate CNNs? Any suggestion\/comment\/feedback is welcome.<\/p>\n\n<p>Thank you,<\/p>\n","answers":[{"AnswerId":"44800867","CreationDate":"2017-06-28T11:13:16.793","ParentId":null,"OwnerUserId":"7996483","Title":null,"Body":"<p>I encountered the same problem and could not solve the problem (without retraining) with any solution i found on the internet. So what I did is load each model in two separate threads which communicate with the main thread. It is simple enough to write the code, you just have to be careful when you synchronize the threads. \nIn my case each thread received the input for its problem and returned to the main thread the output. It works without any observable overhead.<\/p>\n"},{"AnswerId":"41991989","CreationDate":"2017-02-01T23:48:44.633","ParentId":null,"OwnerUserId":"3128177","Title":null,"Body":"<p>Yes there is. Use separate graphs.<\/p>\n\n<pre><code>g1 = tf.Graph()\ng2 = tf.Graph()\n\nwith g1.as_default():\n    cnn1 = CNN(..., restore_file='snapshot-model1-10000',..........) \nwith g2.as_default():\n    cnn2 = CNN(..., restore_file='snapshot-model2-10000',..........)\n<\/code><\/pre>\n\n<p>EDIT:<\/p>\n\n<p>If you want them into same graph. You'll have to rename some variables. One idea is have each CNN in separate scope and let saver handle variables in that scope e.g.:<\/p>\n\n<pre><code>saver = tf.train.Saver(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES), scope='model1')\n<\/code><\/pre>\n\n<p>and in cnn wrap all your construction in scope:<\/p>\n\n<pre><code>with tf.variable_scope('model1'):\n    ...\n<\/code><\/pre>\n\n<p>EDIT2:<\/p>\n\n<p>Other idea is renaming variables which saver manages (since I assume you want to use your saved checkpoints without retraining everything. Saving allows different variable names in graph and in checkpoint, have a look at documentation for initialization.<\/p>\n"},{"AnswerId":"55022410","CreationDate":"2019-03-06T11:47:58.903","ParentId":null,"OwnerUserId":"6502806","Title":null,"Body":"<p>One way is to clear your session if you want to train or load multiple models in succession. You can easily do this using <\/p>\n\n<pre><code>from keras import backend as K \n\n# load and use model 1\n\nK.clear_session()\n\n# load and use  model 2\n\nK.clear_session()`\n<\/code><\/pre>\n\n<p>K.clear_session() destroys the current TF graph and creates a new one.\nUseful to avoid clutter from old models \/ layers.<\/p>\n"},{"AnswerId":"51290092","CreationDate":"2018-07-11T16:15:11.770","ParentId":null,"OwnerUserId":"5161420","Title":null,"Body":"<p>This should be a comment to the most up-voted answer. But I do not have enough reputation to do that.<\/p>\n\n<p>Anyway.\nIf you(anyone searched and got to this point) still having trouble with the solution provided by <a href=\"https:\/\/stackoverflow.com\/a\/41991989\/5161420\">lpp<\/a> AND you are using Keras, check following quote from <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/8538#issuecomment-358720718\" rel=\"noreferrer\">github<\/a>.<\/p>\n\n<blockquote>\n  <p>This is because the keras share a global session if no default tf session provided<\/p>\n  \n  <p>When the model1 created, it is on graph1\n  When the model1 loads weight, the weight is on a keras global session which is associated with graph1<\/p>\n  \n  <p>When the model2 created, it is on graph2\n  When the model2 loads weight, the global session does not know the graph2<\/p>\n  \n  <p>A solution below may help,<\/p>\n<\/blockquote>\n\n<pre><code>graph1 = Graph()\nwith graph1.as_default():\n    session1 = Session()\n    with session1.as_default():\n        with open('model1_arch.json') as arch_file:\n            model1 = model_from_json(arch_file.read())\n        model1.load_weights('model1_weights.h5')\n        # K.get_session() is session1\n\n# do the same for graph2, session2, model2\n<\/code><\/pre>\n"}]}
{"QuestionId":41990665,"AnswerCount":2,"Tags":"<python><neural-network><deep-learning><theano>","CreationDate":"2017-02-01T22:04:32.793","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Theano ValueError: y_i value out of bound","Body":"<p>I am implementing a Neural network using Theano. Where my input size is 64 nodes, hidden layer is 500 node, and output is 1 node. <\/p>\n\n<p>My input is a (1,000,000 * 64) matrix, and output is (1,000,000*1) matrix.<\/p>\n\n<p>I am implementing my Neural network using the following tutorial. <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html#mlp\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/tutorial\/mlp.html#mlp<\/a><\/p>\n\n<p>Please help!<\/p>\n\n<p>i am getting out of bound error on the following line its causing me error<\/p>\n\n<pre><\/pre>\n\n<p>ERROR:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44102173","CreationDate":"2017-05-21T21:54:14.983","ParentId":null,"OwnerUserId":"2552811","Title":null,"Body":"<p>In case anyone still got the same error, I ran into this issue today and the reason is because I labeled the classes from 1 instead of 0.<\/p>\n"},{"AnswerId":"52582907","CreationDate":"2018-09-30T23:15:51.693","ParentId":null,"OwnerUserId":"2847782","Title":null,"Body":"<p>When you add a logistic regression layer to your MLP, the number of outputs must be the number of classes you have. Supposing you are dealing with binary classification, your <code>n_out<\/code> must be 2.<\/p>\n"}]}
{"QuestionId":41990876,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-01T22:17:57.450","AcceptedAnswerId":null,"OwnerUserId":6112394.0,"Title":"TFRecords QueueRunner Error","Body":"<p>I am trying to load many csv files into a single TFRecord file then be able to feed that TFRecord to my model. I below is all my code and I have tried to break it down as to what I think I am doing. <\/p>\n\n<p><strong>Generate data.. the target variable will be the last column.<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Functions for making TFRecord File<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>The function to load csvs into Pandas. Then take the last column and make it my target variable, y. The Pandas dataframes get converted to numpy arrays and written to the TFRecords file.<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>The function to read the TFRecords file.<\/strong> <\/p>\n\n<pre><\/pre>\n\n<p><strong>Created the batches in Tensorflow<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Make the TFRecord file from the data that was created.<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Attempt to load the TFRecord file into a queueRunner.<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>ERROR<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT:<\/strong>\nThe below code is what seems to be the root cause of the problem. I think I am not parsing the TFRecord file properly (duh*). I think maybe I am not reading it in as the correct data type. Almost the exact same code will read pictures into a TFRecord and back out.. Only difference is that I am trying to send float32 values through it all. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"41991079","CreationDate":"2017-02-01T22:33:16.953","ParentId":null,"OwnerUserId":"2405736","Title":null,"Body":"<p>There's a lot to follow there so I'm not sure, but the quickest thing to check is whether your \"num_epochs\" is set properly. Those OutOfRangeErrors are thrown when the epoch limit has been reached.<\/p>\n"}]}
{"QuestionId":41991101,"AnswerCount":8,"Tags":"<tensorflow>","CreationDate":"2017-02-01T22:35:02.367","AcceptedAnswerId":null,"OwnerUserId":7490399.0,"Title":"ImportError: libcudnn when running a TensorFlow program","Body":"<p>I encountered the following error when trying to run a TensorFlow program:<\/p>\n\n<pre>\nImportError: libcudnn.<i>Version<\/i>: cannot open shared object file: No such file or director\n<\/pre>\n","answers":[{"AnswerId":"43784285","CreationDate":"2017-05-04T13:29:43.693","ParentId":null,"OwnerUserId":"837451","Title":null,"Body":"<p>At this time Tensorflow does not support cuDNN 6.  If your error is that it cannot find <code>libcudnn.so.5<\/code> and you have only installed cuDNN 6 which provides <code>libcudnn.so.6<\/code>, you'll have to install cuDNN 5 until cuDNN 6 support is introduced.  There's an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8828\" rel=\"noreferrer\">open bug report<\/a> that you can follow on the Tensorflow Github page to find out when they start supporting cuDNN 6.<\/p>\n"},{"AnswerId":"43901888","CreationDate":"2017-05-10T20:11:17.493","ParentId":null,"OwnerUserId":"4240868","Title":null,"Body":"<p>Make sure the <code>$LD_LIBRARY_PATH<\/code> environment variable is set to the correct path.<\/p>\n\n<p>From the <a href=\"https:\/\/developer.nvidia.com\/compute\/machine-learning\/cudnn\/secure\/v5.1\/prod\/doc\/cudnn_install-txt\" rel=\"nofollow noreferrer\">cuDNN Install Guide<\/a>: <\/p>\n\n<p>ALL PLATFORMS<\/p>\n\n<pre><code>Extract the cuDNN archive to a directory of your choice, referred to below as &lt;installpath&gt;.\nThen follow the platform-specific instructions as follows.\n<\/code><\/pre>\n\n<p>LINUX<\/p>\n\n<pre><code>cd &lt;installpath&gt;\nexport LD_LIBRARY_PATH=`pwd`:$LD_LIBRARY_PATH\n\nAdd &lt;installpath&gt; to your build and link process by adding -I&lt;installpath&gt; to your compile\nline and -L&lt;installpath&gt; -lcudnn to your link line.\n<\/code><\/pre>\n\n<p>OS X<\/p>\n\n<pre><code>cd &lt;installpath&gt;\nexport DYLD_LIBRARY_PATH=`pwd`:$DYLD_LIBRARY_PATH\n\nAdd &lt;installpath&gt; to your build and link process by adding -I&lt;installpath&gt; to your compile\nline and -L&lt;installpath&gt; -lcudnn to your link line.\n<\/code><\/pre>\n\n<p>WINDOWS<\/p>\n\n<pre><code>Add &lt;installpath&gt; to the PATH environment variable.\n\nIn your Visual Studio project properties, add &lt;installpath&gt; to the Include Directories \nand Library Directories lists and add cudnn.lib to Linker-&gt;Input-&gt;Additional Dependencies.\n<\/code><\/pre>\n"},{"AnswerId":"44170463","CreationDate":"2017-05-25T00:28:43.520","ParentId":null,"OwnerUserId":"4772009","Title":null,"Body":"<p>I faced this issue. In my case my eclipse gave this error but tensorflow was running smoothly in terminal.\nSo after cuDNN installation or pasting the files related to cudNN in <code>\/usr\/local\/cuda\/lib64<\/code> and <code>\/usr\/local\/cuda\/include<\/code>, I figured that the command <code>sudo ldconfig -v<\/code> also has to be executed in the terminal.<\/p>\n\n<p>After that it should work.<\/p>\n\n<p>This is assuming that one has already set the following environment variables as per the specifications in CUDA installation procedure by NVIDIA: <code>LD_LIBRARY_PATH<\/code>, <code>PATH<\/code>, <code>CUDA_HOME<\/code><\/p>\n"},{"AnswerId":"44180695","CreationDate":"2017-05-25T12:33:32.583","ParentId":null,"OwnerUserId":"3987993","Title":null,"Body":"<p>Just <a href=\"https:\/\/developer.nvidia.com\/compute\/machine-learning\/cudnn\/secure\/v5.1\/prod_20161129\/8.0\/cudnn-8.0-linux-x64-v5.1-tgz\" rel=\"noreferrer\">download cuDNN 5.1<\/a> and follow the steps (Tested on Ubuntu 16.04, CUDA toolkit 8.0 )<\/p>\n\n<pre><code>$ tar xvzf cudnn-8.0-linux-x64-v5.1-ga.tgz\n$ sudo cp -P cuda\/include\/cudnn.h \/usr\/local\/cuda\/include\n$ sudo cp -P cuda\/lib64\/libcudnn* \/usr\/local\/cuda\/lib64\n$ sudo chmod a+r \/usr\/local\/cuda\/include\/cudnn.h \/usr\/local\/cuda\/lib64\/libcudnn*\n<\/code><\/pre>\n\n<p>Now set Path variables<\/p>\n\n<pre><code>$ vim ~\/.bashrc\n\nexport LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH:\/usr\/local\/cuda\/lib64:\/usr\/local\/cuda\/extras\/CUPTI\/lib64\"\nexport CUDA_HOME=\/usr\/local\/cuda\n<\/code><\/pre>\n\n<p>and done<\/p>\n\n<p>For more details, you can check this <a href=\"https:\/\/medium.com\/@ikekramer\/installing-cuda-8-0-and-cudnn-5-1-on-ubuntu-16-04-6b9f284f6e77\" rel=\"noreferrer\">site<\/a><\/p>\n"},{"AnswerId":"55665390","CreationDate":"2019-04-13T12:23:53.550","ParentId":null,"OwnerUserId":"11355682","Title":null,"Body":"<p>Download and install cuDNN from <a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-download\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/rdp\/cudnn-download<\/a><\/p>\n\n<p>for ubuntu 18.04 and cuda 10.1: <\/p>\n\n<pre><code>sudo dpkg -i libcudnn7_7.5.0.56-1+cuda10.1_amd64.deb\n<\/code><\/pre>\n"},{"AnswerId":"41991123","CreationDate":"2017-02-01T22:36:51.890","ParentId":null,"OwnerUserId":"7490399","Title":null,"Body":"<p>You may not have installed the correct version of cuDNN. <\/p>\n\n<p>To determine the correct version of cuDNN, see the NVIDIA requirements to run TensorFlow with GPU support. <\/p>\n"},{"AnswerId":"45725222","CreationDate":"2017-08-17T01:18:36.583","ParentId":null,"OwnerUserId":"111776","Title":null,"Body":"<p>It appears TensorFlow now supports cudNN 6.0. I was getting errors with cudNN 5.1.<\/p>\n\n<p>Using TensorFlow version 1.3.0.<\/p>\n\n<p>If you're getting <\/p>\n\n<pre><code>ImportError: libcudnn.so.6: cannot open shared object file: No such file or directory\n<\/code><\/pre>\n\n<p>Try using 6.0<\/p>\n"},{"AnswerId":"44417083","CreationDate":"2017-06-07T15:35:18.887","ParentId":null,"OwnerUserId":"8126598","Title":null,"Body":"<p>I was able to solve problem by changing my cudNN installation from cudNN 6.0 to cudNN 5.1. cudNN 6.0 has the file libcudnn.so.6 while cudNN 5.1 has the file libcudnn.so.5.<\/p>\n"}]}
{"QuestionId":41991397,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-01T22:57:27.970","AcceptedAnswerId":41991764.0,"OwnerUserId":4474386.0,"Title":"How to replace a dimension of a tensor in tensorflow?","Body":"<p>Say I have a tensor with shape (?, 5, 1, 20)<\/p>\n\n<p>For each occurence of the last dimension I do some computation (getting the k max values) on the last dimension that produces a smaller tensor b. What do I do if want to replace the last dimension of my original tensor with b?<\/p>\n\n<p>What (preferably pure tensorflow) path should I take?<\/p>\n","answers":[{"AnswerId":"41991764","CreationDate":"2017-02-01T23:28:04.960","ParentId":null,"OwnerUserId":"3128177","Title":null,"Body":"<p>You're doing some computation on last dimension...That is you want to go from (?, 5, 1, 20) -> (?, 5, 1, b) if I understood correctly? <\/p>\n\n<p>What kind of computation?<\/p>\n\n<p>You could reshape your tensor, do the computation (such as matrix multiplication) and reshape back. <\/p>\n\n<pre><code>a = tf.reshape(X, [-1, 20])\na = tf.matmul(a, X)\na = tf.reshape(a, [-1, b])\n<\/code><\/pre>\n\n<p>Or you could use <code>tf.einsum()<\/code> to achieve similar feat. For non-linear computation depends what you want to do. <\/p>\n\n<p>EDIT:\nAlso you could hack it with Conv2D and using filter of size [1,1, 20, b]. Does the same thing and more efficiently<\/p>\n"}]}
{"QuestionId":41991756,"AnswerCount":2,"Tags":"<python-2.7><neural-network><keras><conv-neural-network><keras-layer>","CreationDate":"2017-02-01T23:26:56.987","AcceptedAnswerId":null,"OwnerUserId":5051564.0,"Title":"ValueError: Tensor:(...) is not an element of this graph","Body":"<p>I am using keras' pre-trained model and the error came up when trying to get predictions. I have the following code in flask server:<\/p>\n\n<pre><\/pre>\n\n<p>This is what I have in my NeuralNetwork.py file:<\/p>\n\n<pre><\/pre>\n\n<p>However, running this code below, does not create any error:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the full error:\n<a href=\"https:\/\/i.stack.imgur.com\/mDAgW.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mDAgW.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Any comment or suggestion is greatly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"51528998","CreationDate":"2018-07-25T23:26:36.307","ParentId":null,"OwnerUserId":"7993099","Title":null,"Body":"<p>Considering the backend is set to tensorflow. You should set the Keras session to the graph of tensorflow<\/p>\n\n<pre><code>from tensorflow import Graph, Session\nfrom keras import backend \nmodel = 'model path'\ngraph1 = Graph()\nwith graph1.as_default():\n    session1 = Session(graph=graph1)\n    with session1.as_default():\n        model_1.load_model(model) # depends on your model type\n\nmodel2 = 'model path2'\ngraph2 = Graph()\nwith graph2.as_default():\n    session2 = Session(graph=graph2)\n    with session2.as_default():\n        model_2.load_model(model2) # depends on your model type\n<\/code><\/pre>\n\n<p>and for predicting<\/p>\n\n<pre><code>K.set_session(session#)\nwith graph#.as_default():\n    prediction = model_#.predict(img_data)\n<\/code><\/pre>\n"},{"AnswerId":"42111079","CreationDate":"2017-02-08T10:50:39.233","ParentId":null,"OwnerUserId":"7533224","Title":null,"Body":"<p><strong>EDIT:<\/strong> What I wrote below <strong>doesn't seem to work<\/strong> when deploying the application (I was just testing locally until now). The model in app.config is loaded too often it seems. (at each request?)<\/p>\n\n<p>Coincidentally I had the same problem yesterday. It seems there are some problems between the interaction of TensorFlow and Flask. Unfortunately I don't know enough about the internals of either one to really understand the problem but I can provide a hack which helped me get it to work. (Note: I was using Python3 but I don't think this makes a difference here.)<\/p>\n\n<p>The problem seems to occur when initializing a model in the global namespace of your flask application. Therefore I loaded the model directly into app.config:<\/p>\n\n<pre><code>app.config.update({\"MODEL\":VGG16(weights='imagenet', include_top=True)})\n# ...\napp.config[\"MODEL\"].predict(x)\n<\/code><\/pre>\n\n<p>Maybe you can load the model in your server.py instead of your NeuralNetwork.py and pass it to <code>getVGG16Prediction<\/code> together with <code>img_path<\/code>?<\/p>\n"}]}
{"QuestionId":41992132,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-02-02T00:04:16.367","AcceptedAnswerId":42008906.0,"OwnerUserId":3296050.0,"Title":"Tensorflow error when restoring trained model for eval, invalid shapes issue?","Body":"<p>I'm getting an error when trying to restore a trained model for evaluation, but only when evaluating on the test set. The error is:<\/p>\n\n<pre><\/pre>\n\n<p>Note that lhs shape = [2325,11] and rhs shape = [4891,11] corresponds to 2325 images in the test set and 4891 images in the training set; and the 11 is one-hot encoding for 11 classes - so these likely correspond to the labels. When I run evaluation on the training set the dimensions match and no error results. Help would be appreciated!<\/p>\n\n<p>Full stack trace below:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>I just looked into the tensor shapes from the checkpoint file and it looks like the saver was saving even the inputs to the model. I'll need to reconfigure my training code or otherwise figure out how to exclude the model inputs (labels and images) from the checkpointing:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42008906","CreationDate":"2017-02-02T17:43:16.160","ParentId":null,"OwnerUserId":"5235369","Title":null,"Body":"<p>It would be useful to see the code defining your model; but from what I can understand it looks like you may have defined you input as a <code>tf.Variable<\/code>. A variable is a value that the optimiser is allowed to change in order to minimise the loss function. Variables are the learning weights of your model which is why Tensorflow is saving them so they can be restored later.<\/p>\n\n<p>You should feed your input data to the graph using a <code>tf.Placeholder<\/code>.<\/p>\n"}]}
{"QuestionId":41992190,"AnswerCount":0,"Tags":"<tensorflow><beam-search>","CreationDate":"2017-02-02T00:10:54.373","AcceptedAnswerId":null,"OwnerUserId":3052875.0,"Title":"How to implement a custom beam search in TensorFlow?","Body":"<p>In the original <a href=\"https:\/\/arxiv.org\/abs\/1606.04870\" rel=\"nofollow noreferrer\">Google's SmartReply paper<\/a>:<\/p>\n\n<blockquote>\n  <p>Our search is conducted as follows. First, the elements of R (the set of all possible responses) are organized into a trie. Then, we conduct a left-to-right beam search, but only retain hypotheses that appear in the trie. This search process has complexity O(bl) for beam size b and maximum response length l. Both b and l are typically in the range of 10-30, so this method dramatically reduces the time to find the top responses and is a critical element of making this system deployable.<\/p>\n<\/blockquote>\n\n<p>I want to implement something similar. Some questions:<\/p>\n\n<ol>\n<li>How to represent the set of all possible responses (R) in a trie data structure in TF?<\/li>\n<li>How to modify the current beam search to retain only hypothesis that are available in R?<\/li>\n<\/ol>\n","answers":[]}
{"QuestionId":41992346,"AnswerCount":2,"Tags":"<c++><tensorflow><artificial-intelligence><deep-learning><tensorflow-serving>","CreationDate":"2017-02-02T00:29:29.563","AcceptedAnswerId":42103559.0,"OwnerUserId":2168245.0,"Title":"TensorFlow 0.12 Model Files","Body":"<p>I train a model and save it using:<\/p>\n\n<pre><\/pre>\n\n<p>Besides the <em>checkpoint<\/em> file, which simply contains pointers to the most recent checkpoints of the model, this creates the following 3 files in the current path:<\/p>\n\n<ol>\n<li><em>my_model_name.meta<\/em><\/li>\n<li><em>my_model_name.index<\/em><\/li>\n<li><em>my_model_name.data-00000-of-00001<\/em><\/li>\n<\/ol>\n\n<p>I wonder what each of these files contains. <\/p>\n\n<p>I'd like to load this model in C++ and run the inference. The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/label_image\" rel=\"nofollow noreferrer\">label_image<\/a> example loads the model from a single <strong>.bp<\/strong> file using . I wonder how I can load it from these 3 files. What is the C++ equivalent of the following?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43639040","CreationDate":"2017-04-26T15:51:27.317","ParentId":null,"OwnerUserId":"1076564","Title":null,"Body":"<p>What your saver creates is called \"Checkpoint V2\" and was introduced in TF 0.12.<\/p>\n\n<p>I got it working quite nicely (though the docs on the C++ part are horrible, so it took me a day to solve). Some people suggest <a href=\"https:\/\/stackoverflow.com\/a\/35663177\/1076564\">converting all variables to constants<\/a> or <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/615#issuecomment-184784850\" rel=\"noreferrer\">freezing the graph<\/a>, but none of these is actually needed.<\/p>\n\n<p><strong>Python part (saving)<\/strong><\/p>\n\n<pre><code>with tf.Session() as sess:\n    tf.train.Saver(tf.trainable_variables()).save(sess, 'models\/my-model')\n<\/code><\/pre>\n\n<p>If you create the <code>Saver<\/code> with <code>tf.trainable_variables()<\/code>, you can save yourself some headache and storage space. But maybe some more complicated models need all data to be saved, then remove this argument to <code>Saver<\/code>, just make sure you're creating the <code>Saver<\/code> <strong>after<\/strong> your graph is created. It is also very wise to give all variables\/layers unique names, otherwise you can run in different problems.<\/p>\n\n<p><strong>C++ part (inference)<\/strong><\/p>\n\n<p>Note that <code>checkpointPath<\/code> isn't a path to any of the existing files, just their common prefix. If you mistakenly put there path to the <code>.index<\/code> file, TF won't tell you that was wrong, but it will die during inference due to uninitialized variables.<\/p>\n\n<pre><code>#include &lt;tensorflow\/core\/public\/session.h&gt;\n#include &lt;tensorflow\/core\/protobuf\/meta_graph.pb.h&gt;\n\nusing namespace std;\nusing namespace tensorflow;\n\n...\n\/\/ set up your input paths\nconst string pathToGraph = \"models\/my-model.meta\"\nconst string checkpointPath = \"models\/my-model\";\n...\n\nauto session = NewSession(SessionOptions());\nif (session == nullptr) {\n    throw runtime_error(\"Could not create Tensorflow session.\");\n}\n\nStatus status;\n\n\/\/ Read in the protobuf graph we exported\nMetaGraphDef graph_def;\nstatus = ReadBinaryProto(Env::Default(), pathToGraph, &amp;graph_def);\nif (!status.ok()) {\n    throw runtime_error(\"Error reading graph definition from \" + pathToGraph + \": \" + status.ToString());\n}\n\n\/\/ Add the graph to the session\nstatus = session-&gt;Create(graph_def.graph_def());\nif (!status.ok()) {\n    throw runtime_error(\"Error creating graph: \" + status.ToString());\n}\n\n\/\/ Read weights from the saved checkpoint\nTensor checkpointPathTensor(DT_STRING, TensorShape());\ncheckpointPathTensor.scalar&lt;std::string&gt;()() = checkpointPath;\nstatus = session-&gt;Run(\n        {{ graph_def.saver_def().filename_tensor_name(), checkpointPathTensor },},\n        {},\n        {graph_def.saver_def().restore_op_name()},\n        nullptr);\nif (!status.ok()) {\n    throw runtime_error(\"Error loading checkpoint from \" + checkpointPath + \": \" + status.ToString());\n}\n\n\/\/ and run the inference to your liking\nauto feedDict = ...\nauto outputOps = ...\nstd::vector&lt;tensorflow::Tensor&gt; outputTensors;\nstatus = session-&gt;Run(feedDict, outputOps, {}, &amp;outputTensors);\n<\/code><\/pre>\n\n<p>For completeness, here's the Python equivalent:<\/p>\n\n<p><strong>Inference in Python<\/strong><\/p>\n\n<pre><code>with tf.Session() as sess:\n    saver = tf.train.import_meta_graph('models\/my-model.meta')\n    saver.restore(sess, tf.train.latest_checkpoint('models\/'))\n    outputTensors = sess.run(outputOps, feed_dict=feedDict)\n<\/code><\/pre>\n"},{"AnswerId":"42103559","CreationDate":"2017-02-08T02:19:40.503","ParentId":null,"OwnerUserId":"5166390","Title":null,"Body":"<p>I'm currently struggling with this myself, I've found it's not very straightforward to do currently. The two most commonly cited tutorials on the subject are: \n<a href=\"https:\/\/medium.com\/jim-fleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f#.goxwm1e5j\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/jim-fleming\/loading-a-tensorflow-graph-with-the-c-api-4caaff88463f#.goxwm1e5j<\/a>\nand\n<a href=\"https:\/\/medium.com\/@hamedmp\/exporting-trained-tensorflow-models-to-c-the-right-way-cf24b609d183#.g1gak956i\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@hamedmp\/exporting-trained-tensorflow-models-to-c-the-right-way-cf24b609d183#.g1gak956i<\/a><\/p>\n\n<p>The equivalent of<\/p>\n\n<pre><code>new_saver = tf.train.import_meta_graph('.\/my_model_name.meta')\nnew_saver.restore(session, '.\/my_model_name')\n<\/code><\/pre>\n\n<p>Is just <\/p>\n\n<pre><code>Status load_graph_status = LoadGraph(graph_path, &amp;session);\n<\/code><\/pre>\n\n<p>Assuming you've \"frozen the graph\" (Used a script with combines the graph file with the checkpoint values).\nAlso, see the discussion here: <a href=\"https:\/\/stackoverflow.com\/questions\/35508866\/tensorflow-different-ways-to-export-and-run-graph-in-c\">Tensorflow Different ways to Export and Run graph in C++<\/a><\/p>\n"}]}
{"QuestionId":41993311,"AnswerCount":3,"Tags":"<machine-learning><tensorflow><neural-network><deep-learning><mnist>","CreationDate":"2017-02-02T02:32:12.133","AcceptedAnswerId":41999270.0,"OwnerUserId":1182159.0,"Title":"Adding more layers to tensorflow MNIST tutorial makes accuracy drop","Body":"<p>Novice to deep learning. \nWith MNIST_SOFTMAX.py tutorial from gogoel tensorflow (   <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_softmax.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_softmax.py<\/a>), I added two new layers just to see what happen. <\/p>\n\n<pre><\/pre>\n\n<p>Changed code above into <\/p>\n\n<pre><\/pre>\n\n<p>it drops accuracy from 0.9188 to 0.1028. can I get some idea of why it drops.<\/p>\n","answers":[{"AnswerId":"41996911","CreationDate":"2017-02-02T07:59:17.890","ParentId":null,"OwnerUserId":"1397946","Title":null,"Body":"<p>You suffer from the same problem as answered in <a href=\"https:\/\/stackoverflow.com\/questions\/41872353\/how-to-test-if-my-implementation-of-back-propagation-neural-network-is-correct\/41899530#41899530\">this post<\/a>. Essentially, your first hidden layers learn much slower than last. Normally your network should learn the correct weights. Here, however, most likely the weights in the first layer change very little and the error propagates to the next layers. It's so large that the subsequent layers can't possibly correct for it. Check the weights.<\/p>\n"},{"AnswerId":"41993395","CreationDate":"2017-02-02T02:44:50.313","ParentId":null,"OwnerUserId":"4561314","Title":null,"Body":"<p>You need to add a non-linear activation function between your layers. Try ReLU.<\/p>\n"},{"AnswerId":"41999270","CreationDate":"2017-02-02T10:06:28.780","ParentId":null,"OwnerUserId":"728291","Title":null,"Body":"<p>I think you need both <a href=\"https:\/\/www.quora.com\/Why-do-we-apply-symmetry-breaking-to-the-weight-matrices-in-Neural-Networks-but-not-to-the-Biases-during-initilization\" rel=\"nofollow noreferrer\">symmetry breaking in the weights<\/a> and a non-linear activation between layers:<\/p>\n\n<pre><code>W = tf.Variable(tf.random_normal([784, 256], stddev=0.1)) \nW1 = tf.Variable(tf.random_normal([256, 256], stddev=0.1))\nW2 = tf.Variable(tf.random_normal([256, 10], stddev=0.1))\nb = tf.Variable(tf.zeros([256]))\nb1 = tf.Variable(tf.zeros([256]))\nb2 = tf.Variable(tf.zeros([10]))\n\ny = tf.matmul(x, W) + b\ny = tf.nn.relu(y)\ny = tf.matmul(y, W1) + b1\ny = tf.nn.relu(y)\ny = tf.matmul(y, W2) + b2\n<\/code><\/pre>\n\n<p>That gets accuracy of 0.9653.<\/p>\n"}]}
{"QuestionId":41993634,"AnswerCount":2,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-02-02T03:18:00.613","AcceptedAnswerId":41999400.0,"OwnerUserId":2839624.0,"Title":"tensorflow image retraining with serving","Body":"<p>I am trying to serve retrained inception graph using tensorflow serving. For retraining, I am using this <a href=\"https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/\" rel=\"nofollow noreferrer\">example<\/a>. However I need to make changes to this graph to get it working with <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/master\/tensorflow_serving\/example\/inception_saved_model.py\" rel=\"nofollow noreferrer\">serving export code<\/a>.<\/p>\n\n<p>Since in tensorflow serving, you will receive serialized images as input, graph input should start with this:<\/p>\n\n<pre><\/pre>\n\n<p>This images tensor should be input to retrained inception graph. However I don't know if its possible to prepend one graph to another in tensorflow like you can append easily using placeholder_with_input (which has been done in retraining code).<\/p>\n\n<pre><\/pre>\n\n<p>Ideally, in image retraining code, I receive a placeholder tensor . I need to append tensor  to this placeholder tensor  and export it as single graph using exporter so that it can be served using tensorflow serving. However I don't any tensorflow instruction that does it. Are there any other alternatives apart from this method?<\/p>\n","answers":[{"AnswerId":"41999400","CreationDate":"2017-02-02T10:13:12.163","ParentId":null,"OwnerUserId":"6933420","Title":null,"Body":"<p>One way of going about it is:<\/p>\n\n<pre><code>model_path = 'trained\/export.pb' \nwith tf.Graph().as_default():\n    with tf.Session() as sess:       \n        with gfile.FastGFile(model_path, 'rb') as f:\n            graph_def = tf.GraphDef()\n            graph_def.ParseFromString(f.read())\n            # Your prepending ops here\n            images_placeholder = tf.placeholder(tf.string, name='tf_example')\n            ...\n            images = tf.map_fn(preprocess_image, jpegs, dtype=tf.float32)\n            tf.import_graph_def(graph_def, name='inception', input_map={'ResizeBilinear:0': images})\n<\/code><\/pre>\n\n<p>Notice especially the <code>input_map<\/code> argument. <code>ResizeBilinear:0<\/code> is likely not the correct name of the operation you need - you can list the ops by:<\/p>\n\n<pre><code>[n.name for n in tf.get_default_graph().as_graph_def().node]  \n<\/code><\/pre>\n\n<p>I realize this is not a full answer and perhaps not the most efficient but hopefully it can get you started. Just a heads-up, there is also <a href=\"https:\/\/cloud.google.com\/blog\/big-data\/2016\/12\/how-to-train-and-classify-images-using-google-cloud-machine-learning-and-cloud-dataflow\" rel=\"nofollow noreferrer\">this blogpost<\/a>. <\/p>\n"},{"AnswerId":"45207408","CreationDate":"2017-07-20T06:57:00.630","ParentId":null,"OwnerUserId":"3910238","Title":null,"Body":"<p>So since you have already retrained the model, I'm assuming that the model is a Protobuf, but you can just load that up into a Python object and serve off of that object using a custom function that processes either a batch or an atomic operation.<\/p>\n\n<p>And to your graph question, as far as I know, when you load a tf.Graph() object you are only working with that object and can't work with other graphs... that being said if you had another graph that is an extension of the existing Inception-V3 Graph then you can easily add that to the computation graph for your custom graph quite easily.<\/p>\n"}]}
{"QuestionId":41993837,"AnswerCount":1,"Tags":"<python><machine-learning><keras><conv-neural-network><lstm>","CreationDate":"2017-02-02T03:44:05.753","AcceptedAnswerId":42002116.0,"OwnerUserId":3466910.0,"Title":"Convolutional1D in Keras convolves on time steps instead of features?","Body":"<p>So, my input is of shape , or, more generally, . This corresponds to 1 sample of 893 time steps, each with 463 features. The output shape is , i.e. .<\/p>\n\n<p>My network structure looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Compiled like this:<\/p>\n\n<pre><\/pre>\n\n<p>Problem is, when I do , I get the following error: , tracing back to the  line. <\/p>\n\n<p>I logged the shapes of the outputs of each of the layers using <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">this<\/a> technique, coming up with this:<\/p>\n\n<pre><\/pre>\n\n<p>I suspect this occurs when the model tries to calculate accuracy, and finds that for 893 correct outputs, it only has 9 predictions. For some reason, the second  layer starts convolving on the time steps, not on the features, as the first one did. <\/p>\n\n<p>Why is this, and how do I fix this?<\/p>\n\n<h1>EDIT:<\/h1>\n\n<p>Model summary:<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to make a CNN-LSTM classifier which, given time series data, will give an output for each time step.<\/p>\n\n<p>Full error message:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42002116","CreationDate":"2017-02-02T12:19:12.993","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>Right now all the convolution\/pooling layers are acting over time. If you want to apply them in the features space you would need to make them <code>TimeDistributed<\/code> and add an extra dimension at the end of your input. Then you would need to remove the extra dimension before passing the data to the <code>GRU<\/code> layer. However, you would only be able to do that if you have one channel output per convolution:<\/p>\n\n<pre><code>import keras.backend as K\n\nmodel = Sequential()\nmodel.add(Lambda(lambda x: K.expand_dims(x, -1)))\nmodel.add(TimeDistributed(Convolution1D(1, 5, input_dim = one_input_length, border_mode = \"same\", W_regularizer = l2(0.01))))\nmodel.add(TimeDistributed(MaxPooling1D(10, border_mode = \"same\")))\nmodel.add(TimeDistributed(Convolution1D(1, 5, border_mode = \"same\", W_regularizer = l2(0.01))))\nmodel.add(TimeDistributed(MaxPooling1D(10, border_mode = \"same\")))\nmodel.add(Lambda(lambda x: K.squeeze(x, -1)))\nmodel.add(GRU(300, return_sequences = True, W_regularizer = l2(0.01), U_regularizer = l2(0.01)))\nmodel.add(TimeDistributed(Dense(2, activation='sigmoid')))\n<\/code><\/pre>\n\n<p>If you want to use multiple output channel in your convolution then you would need to create some kind of \"matrix\" of GRU units.<\/p>\n"}]}
{"QuestionId":41993924,"AnswerCount":1,"Tags":"<python><theano><theano.scan>","CreationDate":"2017-02-02T03:53:19.957","AcceptedAnswerId":null,"OwnerUserId":7503833.0,"Title":"How to iterate over the rows of a theano matrix using the scan function in theano?","Body":"<p>I am writing a simple code to compute one-hot encoding for a list of indices. \nEg: [1,2,3] => [[0,1,0,0],[0,0,1,0],[0,0,0,1]]<\/p>\n\n<p>I write a function to do the same for a single vector:<\/p>\n\n<pre><\/pre>\n\n<p>To repeat the same function over the rows of a matrix, I do the following,<\/p>\n\n<pre><\/pre>\n\n<p>I am expecting a 3D tensor with each slice corresponding to the one-hot encoding of the rows of the matrix.<\/p>\n\n<p>I am getting the following error while compiling the function,<\/p>\n\n<pre><\/pre>\n\n<p>Can someone please help me understand where I am going wrong and how I can modify the code to obtain what I need? <\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42661941","CreationDate":"2017-03-08T02:34:01.260","ParentId":null,"OwnerUserId":"5250620","Title":null,"Body":"<p>Here is the code that work <\/p>\n\n<pre><code># input a matrix, expect scan to work with each row of matrix\nmy_matrix = np.asarray([[1,2,3],[1,3,2],[1,1,1]])\n\nx = T.imatrix()\n\ndef encoding(idx):\n    z = theano.tensor.zeros((idx.shape[0], 4))\n    one_hot = theano.tensor.set_subtensor(z[theano.tensor.arange(idx.shape[0]), idx], 1)\n    return one_hot\n\nm, update = theano.scan(fn=encoding,\n                        sequences=x)\n\n\nf = theano.function([x], m)\n\n##########3\nresult = f(my_matrix)\nprint (result)\n<\/code><\/pre>\n"}]}
{"QuestionId":41994847,"AnswerCount":1,"Tags":"<logging><tensorflow><tracking>","CreationDate":"2017-02-02T05:32:46.033","AcceptedAnswerId":42013113.0,"OwnerUserId":7474000.0,"Title":"Tracking every tensor variable's changing while training","Body":"<p>Is it possible to track a variable's value while training and test?\nI'm going to log every step's tensor value and make some program see\nit.<\/p>\n\n<p>But before doing this: is it possible using tensorflow's summary\nfunction?  If possible, using only tensorboard and web site, then it\nwouldn't give me detail info and I couldn't get intuition.<\/p>\n","answers":[{"AnswerId":"42013113","CreationDate":"2017-02-02T22:00:05.673","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>You should be able to do this with Tensorboard and summary ops. Take a look at the examples here: <a href=\"https:\/\/www.tensorflow.org\/how_tos\/summaries_and_tensorboard\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/summaries_and_tensorboard\/<\/a><\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":41995052,"AnswerCount":2,"Tags":"<neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-02-02T05:49:26.913","AcceptedAnswerId":41999781.0,"OwnerUserId":5009875.0,"Title":"How good is training in pycaffe compared to caffe?","Body":"<p>I want to try using pycaffe interface to train caffe models.\nParticularly I want to use a python layer for input data.<\/p>\n\n<p>Will this training be slow or will pycaffe just act as an interface to actual caffe?What are the effects?<\/p>\n\n<p>Any attempts to enlighten the novice are thanked.<\/p>\n","answers":[{"AnswerId":"41999781","CreationDate":"2017-02-02T10:30:35.907","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>pycaffe is a Python interface to caffe, so running from pycaffe just calls caffe methods under the hood.<\/p>\n\n<p>On the other hand, having a Python layer in your model does not compel you to use pycaffe. You can run this model like any other from command line.<\/p>\n"},{"AnswerId":"42278259","CreationDate":"2017-02-16T15:33:26.433","ParentId":null,"OwnerUserId":"6786718","Title":null,"Body":"<p>This can be asked as a general question, not only to caffe. What pycaffe does is simply calling the C++ methods that are implemented. Of coruse you will have an overhead to to the python interface and the calling of the function. This blog evaluates the <a href=\"https:\/\/ilovesymposia.com\/2015\/12\/10\/the-cost-of-a-python-function-call\/\" rel=\"nofollow noreferrer\">The cost of a Python function call<\/a>. <\/p>\n\n<p>Furthermore, if you use a Python IDE like e.g Spyder you give some resources to the IDE. If you run your script in a python-console it may take less resources. <\/p>\n\n<p>All in all it can be said that if you have a average machine pycaffe will work just as good as the C++ implementation and you will not directly feel any different. <\/p>\n\n<p>Pycaffe will be easier to handle than the C++ impelementation thanks to abstraction. <\/p>\n"}]}
{"QuestionId":41995376,"AnswerCount":0,"Tags":"<ubuntu><tensorflow><classification>","CreationDate":"2017-02-02T06:15:23.803","AcceptedAnswerId":null,"OwnerUserId":5838783.0,"Title":"TypeError: argument of type 'NoneType' is not iterable Tensorflow estimator.py DNNClassifier","Body":"<p>Currently I am working on a DNNClassifier to predict stock trends for next day.<\/p>\n\n<p>Below is the code for DNNClassifier created using tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>When executed the code throws below error:<\/p>\n\n<pre><\/pre>\n\n<p>Could you please help me in solving this. If this is an existing issue, a link to that would be of great help.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
{"QuestionId":41995678,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-02-02T06:36:08.720","AcceptedAnswerId":41995977.0,"OwnerUserId":2736559.0,"Title":"How can I get layer type in pycaffe?","Body":"<p>Is it possible at all to get each layer's type (e.g: Convolution, Data, etc) in pycaffe? \nI searched the examples provided, but I couldn't find anything. currently I'm using layers name to do my job which is extremely bad and limiting . <\/p>\n","answers":[{"AnswerId":"41995977","CreationDate":"2017-02-02T06:59:03.340","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It's easy!<\/p>\n\n<pre><code>import caffe\nnet = caffe.Net('\/path\/to\/net.prototxt', '\/path\/to\/weights.caffemodel', caffe.TEST)\n\n# get type of 5-th layer\nprint \"type of 5-th layer is \", net.layers[5].type\n<\/code><\/pre>\n\n<p>To map between layer names and indices you can use this simple trick:<\/p>\n\n<pre><code>idx = list(net._layer_names).index('my_layer')\nprint \"The index of \\'my_layer\\' is \", idx, \" and the type is \", net.layers[idx].type\n<\/code><\/pre>\n"}]}
{"QuestionId":41996006,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-02T07:00:58.400","AcceptedAnswerId":null,"OwnerUserId":2849934.0,"Title":"Should I enable SIMD extensions together with CUDA support when building Tensorflow?","Body":"<p>My CPU has support for SSE4.2, AVX2 and FMA.<\/p>\n\n<ol>\n<li>Is there any gain in passing  along with ?<\/li>\n<li><strike>If all of them are enabled, will Tensorflow choose the fastest one among them for completing a certain task? Is there a precedence order?<\/strike> GPU is always preferred.<\/li>\n<li><strike>What does  do?<\/strike> Allocate more registers (including SSE) for floating point operations? What difference does it make when SSE4.2 is already enabled?<\/li>\n<\/ol>\n\n<p><strong>UPDATE<\/strong><\/p>\n\n<p>I tested <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/translate\" rel=\"nofollow noreferrer\">\"translate\" example<\/a> with 4 different configurations in 2 different machines:<\/p>\n\n<ol>\n<li><p>In an i5-6200U with 930MX machine:<\/p>\n\n<ul>\n<li><p><\/p>\n\n<p>The \"step-time\" ranged from 0.65 to 0.81<\/p><\/li>\n<li><p><\/p>\n\n<p>The \"step-time\" ranged from 0.66 to 0.89 (I'd consider this noise)<\/p><\/li>\n<\/ul><\/li>\n<li><p>In an i7-4770 machine (no GPU):<\/p>\n\n<ul>\n<li><p><\/p>\n\n<p>The \"step-time\" ranged from 1.75 - 1.93<\/p><\/li>\n<li><p><\/p>\n\n<p>The \"step-time\" ranged from 1.26 - 1.40 (definitely an improvement)<\/p><\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p><em>UPDATE2:<\/em> I meant to pass , but incorrectly passed  instead! From this, I will (lazily) assume that passing  will indeed enable all  features supported in the current platform.<\/p>\n","answers":[{"AnswerId":"42028085","CreationDate":"2017-02-03T15:51:33.270","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>To summarize, the following will enable all hardware-specific optimizations in (in TF 1.0rc1 or later):<\/p>\n\n<p><code>bazel build --config=opt<\/code><\/p>\n\n<p>For GPU support, use<\/p>\n\n<p><code>bazel build --config=opt --config=cuda<\/code><\/p>\n"}]}
{"QuestionId":41996593,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2017-02-02T07:39:23.497","AcceptedAnswerId":null,"OwnerUserId":2736559.0,"Title":"Is it possible in Caffe to calculate number of Operations taking place in an architecture?","Body":"<p>I have seen in <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"noreferrer\">tensorflow tutorial<\/a>s, they provide some interesting statistics about different architectures, such as number of operations taking place, etc . <\/p>\n\n<blockquote>\n  <p>This model achieves a peak performance of about 86% accuracy within a\n  few hours of training time on a GPU. Please see below and the code for\n  details. It consists of 1,068,298 learnable parameters and requires\n  about 19.5M multiply-add operations to compute inference on a single\n  image.<\/p>\n<\/blockquote>\n\n<p>How can I achieve something similar using Caffe?<br>\nHow can compute or compile such interesting statistics myself?<br>\n(by the way I'm not talking about tensorboard visualizations, Just statistics such as number of operations and stuff like that)<\/p>\n","answers":[{"AnswerId":"41997609","CreationDate":"2017-02-02T08:41:08.133","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>There is an open <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/2507\" rel=\"nofollow noreferrer\">issue #2507<\/a> in BVLC\/caffe github requesting this feature.<br>\nYou are more than welcome to make a contribution to caffe implementing this feature!<\/p>\n"}]}
{"QuestionId":41996660,"AnswerCount":1,"Tags":"<machine-learning><scikit-learn><caffe>","CreationDate":"2017-02-02T07:43:38.303","AcceptedAnswerId":null,"OwnerUserId":4928920.0,"Title":"Generating test set in an augmented dataset","Body":"<p>I have an image retrieval task at hand. I have a dataset with not a lot of images per class, so I ended up augmenting the dataset(3 copies per image in the class) with some inspiration from the Keras <a href=\"https:\/\/blog.keras.io\/building-powerful-image-classification-models-using-very-little-data.html\" rel=\"nofollow noreferrer\">blog<\/a>. I used the  method in  to split the dataset generated from the augmented dataset. Now because this is a random split, when testing there are chances that the image that is being used to test is being tested against its own augmented counterpart image, for example a flipped and slightly skewed image tested against its original. In such a case there are chances that results can be skewed? Is there a way to overcome this problem apart from just getting new data solely for testing.  <\/p>\n","answers":[{"AnswerId":"41999054","CreationDate":"2017-02-02T09:56:56.350","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Yes, that is called \"leaking\", and unless the data augmentation transformations are so radical that the generated images might be similar to a real test set, then its better to split the train\/test sets BEFORE doing augmentation, and only do augmentation in the train set.<\/p>\n"}]}
{"QuestionId":41997719,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-02T08:47:36.730","AcceptedAnswerId":null,"OwnerUserId":3705573.0,"Title":"Tensorflow Compilation Speeding up CPU","Body":"<p>As I use the following command after importing tensorflow in python 2.7:\n<\/p>\n\n<p><strong>Warnings\/errors:<\/strong><\/p>\n\n<blockquote>\n  <p>tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use SSE4.2 instructions, but these are\n  available on your machine and could speed up CPU computations.<\/p>\n  \n  <p>2017-02-02 00:41:48.616602: W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use AVX instructions, but these are\n  available on your machine and could speed up CPU computations.<\/p>\n  \n  <p>2017-02-02 00:41:48.616614: W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use AVX2 instructions, but these are\n  available on your machine and could speed up CPU computations.<\/p>\n  \n  <p>2017-02-02 00:41:48.616624: W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use FMA instructions, but these are\n  available on your machine and could speed up CPU computations.<\/p>\n<\/blockquote>\n\n<p>Please help me fix this so I may use my machine at its optimal power.<\/p>\n","answers":[{"AnswerId":"45975748","CreationDate":"2017-08-31T07:32:13.980","ParentId":null,"OwnerUserId":"2536294","Title":null,"Body":"<p>These warnings you see, are telling you that the compiled code does not use these instructions which you have, but not all CPUs out there. When maintainers compile codes for repositories, they need to compile it such that it supports majority of CPUs out there, which means they tell the compiler to use architecture specific instructions.<\/p>\n\n<p>If you want the package to use all the instructions you have, you need to compile it yourself, or as it's called <em>install from source<\/em>. You can find documentation about how to do that <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"nofollow noreferrer\">here<\/a>, and once you're comfortable compiling <code>tensorflow<\/code> from source, then you should go and read the <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide\" rel=\"nofollow noreferrer\">performance specific instructions<\/a>.<\/p>\n\n<p>However, at the end of the day, for realworld applications you might really need a GPU. It is true that these CPU instructions give you a bit of performance boost, but that's not comparable to using a GPU.<\/p>\n"},{"AnswerId":"42458606","CreationDate":"2017-02-25T16:31:22.927","ParentId":null,"OwnerUserId":"4847576","Title":null,"Body":"<p>Those warnings are just saying if you build TensorFlow from source it can run faster on your machine. There is no fix as it's not an issue but intended behavior to provide this information to users.<\/p>\n\n<p>Those CPU instructions were not enabled by default to provide a broader compatibility with most machines.<\/p>\n\n<p>As the docs says:<\/p>\n\n<blockquote>\n  <p>TensorFlow checks on startup whether it has been compiled with the optimizations available on the CPU. If the optimizations are not included, TensorFlow will emit warnings, e.g. AVX, AVX2, and FMA instructions not included.<\/p>\n<\/blockquote>\n\n<p>For all details on that see the <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide\" rel=\"nofollow noreferrer\">Performance Guide<\/a>.<\/p>\n"}]}
{"QuestionId":41998358,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-02T09:21:32.693","AcceptedAnswerId":null,"OwnerUserId":2736559.0,"Title":"How does TensorFlow calculate FLOPS?","Body":"<p>I wanted to implement something similar in Caffe for myself but I'm clueless how this is done.<br>\nHow exactly is  calculating such Statistics ? <\/p>\n","answers":[]}
{"QuestionId":41998735,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-02-02T09:41:51.763","AcceptedAnswerId":null,"OwnerUserId":7505030.0,"Title":"weighted average of tensor","Body":"<p>I am trying to implement a custom objective function in keras frame.\nRespectively a weighted average function that takes the two arguments tensors y_true and y_pred ; the weights information is derived from y_true tensor.<\/p>\n\n<p>Is there a weighted average function in tensorflow ?\nOr any other suggestions on how to implement this kind of loss function ?<\/p>\n\n<p>My function would look something like this:<\/p>\n\n<p>function(y_true,y_pred)\n  A=(y_true-y_pred)**2\n  w - derivable from y_true, tensor of same shape as y_true\n  return average(A, weights=w) &lt;-- a scalar<\/p>\n\n<p>y_true and y_pred are 3D tensors.<\/p>\n","answers":[{"AnswerId":"41999129","CreationDate":"2017-02-02T10:00:15.400","ParentId":null,"OwnerUserId":"7438048","Title":null,"Body":"<p>you can use one of the existing objectives (also called loss) on <code>keras<\/code> from <a href=\"https:\/\/keras.io\/objectives\/#usage-of-objectives\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>you may also implement your own custom function loss:<\/p>\n\n<pre><code>from keras import backend as K\n\ndef my_loss(y_true, y_pred):\n    return K.mean(K.square(y_pred - y_true), axis=-1)\n\n# Let's train the model using RMSprop\nmodel.compile(loss=my_loss, optimizer='SGD', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>notice the <code>K<\/code> module, its the keras backend you should use to fully utilize keras performance, dont do something like this unless you dont care from performance issues:<\/p>\n\n<pre><code>def my_bad_and_slow_loss(y_true, y_pred):\n    return sum((y_pred - y_true) ** 2, axis=-1)\n<\/code><\/pre>\n\n<p>for your specific case, please write your desired objective function if you need help to write it.<\/p>\n\n<h1>Update<\/h1>\n\n<p>you can try this to provide weights - <code>W<\/code> as loss function:<\/p>\n\n<pre><code>def my_loss(y_true, y_pred):\n    W = np.arange(9) \/ 9.  # some example W\n    return K.mean(K.pow(y_true - y_pred, 2) * W)\n<\/code><\/pre>\n"}]}
{"QuestionId":41999686,"AnswerCount":1,"Tags":"<python><optimization><neural-network><keras><python-3.5>","CreationDate":"2017-02-02T10:26:18.890","AcceptedAnswerId":42009206.0,"OwnerUserId":3338246.0,"Title":"Keras classifier accuracy steadily increases during training then drops to 0.25 (local minimum?)","Body":"<p>I have the following neural network, written in Keras using Tensorflow as the backend, which I'm running on Python 3.5 (Anaconda) on Windows 10:<\/p>\n\n<pre><\/pre>\n\n<p>I'm training on my GPU. During training (10000 epochs), the accuracy of the naive network steadily increases from 0.25 to somewhere between 0.7 and 0.9, before suddenly dropping and sticking at 0.25:<\/p>\n\n<pre><\/pre>\n\n<p>I'm guessing that this is due to the optimiser falling into a local minimum where it assigns all data to one category. How can I inhibit it from doing this?<\/p>\n\n<p>Things I've tried (but didn't seem to stop this from happening):<\/p>\n\n<ol>\n<li>Using a different optimiser (adam)<\/li>\n<li>Ensuring that the training data included an equal number of examples from each category<\/li>\n<li>Increasing the volume of training data (currently at 6000)<\/li>\n<li>Varying the number of categories between 2 to 5<\/li>\n<li>Increasing the number of hidden layers in the network from 1 to 5<\/li>\n<li>Changing the width of the layers (from 50 to 500)<\/li>\n<\/ol>\n\n<p>None of these helped. Any other ideas why this is happening and\/or how to inhibit it? Could it be a bug in Keras? Many thanks in advance for any suggestions.<\/p>\n\n<p><strong>Edit:<\/strong>\nThe problem <em>appears<\/em> to have been solved by changing the final activation to softmax (from sigmoid) and adding maxnorm(3) regularization to the final two hidden layers:<\/p>\n\n<pre><\/pre>\n\n<p>Many thanks for the suggestions.<\/p>\n","answers":[{"AnswerId":"42009206","CreationDate":"2017-02-02T17:59:20.523","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The problem lied in <code>sigmoid<\/code> function as an activation in a last layer. In this case the output of your final layer cannot be interpreted as a probability distribution of an example given belonging to a single class. The output from this layer usually doesn't even sum up to 1. In this case the optimization may lead to unexpected behaviour. In my opinion adding a <code>maxnorm<\/code> constrain is not necessary but I strongly advise you to use a <code>categorical_crossentropy<\/code> instead of <code>mse<\/code> loss as it's proven that this function works better for this optimization case.<\/p>\n"}]}
{"QuestionId":42000375,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><fsm><towers-of-hanoi>","CreationDate":"2017-02-02T10:56:56.960","AcceptedAnswerId":null,"OwnerUserId":4697424.0,"Title":"Recommended Transition from FSM to ML","Body":"<p>I have been doing a bit of machine learning stuff with tensorflow and tflearn, and decided to revisit an old problem. the specific problem is writing a program that can effectively solve the towers of hanoi from any state to any other state. I have previously done this using a FSM but would like to take a crack at it using ML, and was wondering if there is there any particular type of ANN I should try first?<\/p>\n","answers":[]}
{"QuestionId":42000831,"AnswerCount":0,"Tags":"<python-2.7><tensorflow><multiprocessing><distributed>","CreationDate":"2017-02-02T11:18:21.927","AcceptedAnswerId":null,"OwnerUserId":3399275.0,"Title":"Creating session hangs on distributed tensorflow","Body":"<p>I am working my way through tensorflow distributed tutorial:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/distributed\/<\/a><\/p>\n\n<p>I am trying to use python multiprocessing Processes instead of launching scripts in separate command shells. Unfortunately, code hangs on the stage of opening a session. <\/p>\n\n<p>Any ideas are welcome. I have prepared the very minimal example of code, which basically just launches few parallel processes:<\/p>\n\n<pre><\/pre>\n\n<p>Python 2.7. Tensorflow 0.12.1 (CPU), Mint 17 (Ubuntu x64)<\/p>\n\n<p>EDIT:<\/p>\n\n<p>Problem didn't reproduce on Tensorflow CUDA version.<\/p>\n","answers":[]}
{"QuestionId":42001507,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><deep-learning><caffe><reduction>","CreationDate":"2017-02-02T11:50:52.703","AcceptedAnswerId":42003153.0,"OwnerUserId":5498819.0,"Title":"Caffe - How to use reduction layer?","Body":"<p>I have a question regarding the reduction layer in caffe. I haven't found any examples on how to use this layer in my .prototxt file. So I would appreciate it, if anybody could give me a short example of how to use this layer.<\/p>\n\n<p>This is the documentation: <a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1ReductionLayer.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1ReductionLayer.html<\/a> , but there is no example given :-\/<\/p>\n\n<p>I want to use this layer to reduce an output matrix of 1x18x18 to one scalar value. So i want the absolute sum of this matrix.<\/p>\n","answers":[{"AnswerId":"42003153","CreationDate":"2017-02-02T13:11:59.660","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Try:<\/p>\n\n<pre><code>layer {\n  name: \"reduction\"\n  type: \"Reduction\"\n  bottom: \"in\"\n  top: \"out\"\n  reduction_param {\n    axis: 0 # reduce all dims after first\n    operation: ASUM  # use absolute sum\n  }\n}\n<\/code><\/pre>\n\n<p>For more information, see <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L956-L983\" rel=\"nofollow noreferrer\"><code>caffe.proto<\/code><\/a> and the new <a href=\"http:\/\/caffe.help\/manual\/layers\/reduction.html\" rel=\"nofollow noreferrer\">caffe.help<\/a>.<\/p>\n"}]}
{"QuestionId":42001566,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tensorboard>","CreationDate":"2017-02-02T11:53:58.023","AcceptedAnswerId":42005052.0,"OwnerUserId":3747801.0,"Title":"TensorFlow: How to merge multiple 'collections'?","Body":"<p>I have some  that I would like to track with TensorBoard using a supervisor. In the Supervisor initializer I would like something to the effect         <\/p>\n\n<pre><\/pre>\n\n<p>But I get the error , because the  must be a string (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/summary\/generation_of_summaries_#merge_all\" rel=\"nofollow noreferrer\">see documentation<\/a>).<\/p>\n\n<p><br><\/p>\n\n<hr>\n\n<p><strong>Edit<\/strong>:<\/p>\n\n<p>This doesn't work either:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42005052","CreationDate":"2017-02-02T14:39:58.900","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>Try <code>tf.summary.merge()<\/code>, e.g. like so:<\/p>\n\n<pre><code>summary_op = tf.summary.merge([\n        tf.summary.merge_all('test'),\n        tf.summary.merge_all('train')],\n    collections='merged')\n<\/code><\/pre>\n\n<p>This would merge all summaries from the <code>test<\/code> and <code>train<\/code> collections and add them to a new <code>merged<\/code> collection. Keep in mind that this will result in strange effects if the <em>same<\/em> summary is used multiple times during the <em>same<\/em> time step:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YkhOB.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YkhOB.jpg\" alt=\"Same summaries, same time step\"><\/a><\/p>\n\n<p>Here I was accidentally (manually!) storing validation summaries during training runs and then <em>again<\/em> in a separate validation run.<\/p>\n\n<p>Also I'm not sure if this is the most efficient way to go about it, but it certainly works.<\/p>\n"}]}
{"QuestionId":42002083,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><recurrent-neural-network>","CreationDate":"2017-02-02T12:17:54.950","AcceptedAnswerId":42005426.0,"OwnerUserId":4154826.0,"Title":"Tensorflow Session.Run() Tensor object is not callable","Body":"<p>I have a rnn model trained for the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\/\" rel=\"nofollow noreferrer\">ptb example<\/a> with tensorflow's <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/ptb_word_lm.py\" rel=\"nofollow noreferrer\">ptb_word<\/a>.\nBellow I have a code where I'm trying to print a few examples to test the model trained. I'm getting a error  when running this code on the line I make <\/p>\n\n<p>What exactly causes this error?<\/p>\n\n<p>here is the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42005426","CreationDate":"2017-02-02T14:57:20.993","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>Looking at the referenced code at GitHub I cannot find <code>output_props<\/code>, so maybe the versions differ. However, since <code>mtest.initial_state<\/code> is a <code>@property<\/code>, I assume that <code>mtest.output_props<\/code> is one as well. That is, try<\/p>\n\n<pre><code>probs, state = sess.run([mtest.output_probs, mtest.final_state], feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>instead, i.e. without using the parentheses.<\/p>\n\n<p>Also <code>mtest._final_state<\/code> is an internal variable and shouldn't be used directly. You probably want to use <code>mtest.final_state<\/code> instead.<\/p>\n"}]}
{"QuestionId":42002569,"AnswerCount":1,"Tags":"<lua><tensorflow>","CreationDate":"2017-02-02T12:41:41.127","AcceptedAnswerId":null,"OwnerUserId":5891602.0,"Title":"Function similar to Lua's nn.Identity()() in tensorflow","Body":"<p>Hello can some one please suggest a similar function to Lua's nn.Identity()() that can be used in Tensorflow.<\/p>\n","answers":[{"AnswerId":"42005526","CreationDate":"2017-02-02T15:01:03.517","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>There is <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/control_flow_ops\/control_flow_operations#identity\" rel=\"nofollow noreferrer\"><code>tf.identity(x)<\/code><\/a> which creates a new (named) node of the same content as <code>x<\/code>.<\/p>\n"}]}
{"QuestionId":42002717,"AnswerCount":2,"Tags":"<python><neural-network><keras><lstm><sequence-to-sequence>","CreationDate":"2017-02-02T12:49:06.960","AcceptedAnswerId":42084491.0,"OwnerUserId":996366.0,"Title":"How should we pad text sequence in keras using pad_sequences?","Body":"<p>I have coded a <a href=\"https:\/\/stackoverflow.com\/questions\/41933958\/how-to-code-a-sequence-to-sequence-rnn-in-keras\">sequence to sequence<\/a> learning LSTM in keras myself using the knowledge gained from the web tutorials and my own intuitions. I converted my sample text to sequences and then padded using  function in keras. <\/p>\n\n<pre><\/pre>\n\n<p>After a carefully inspection I found my padded sequence looks like this <\/p>\n\n<pre><\/pre>\n\n<p>Is the padded sequence suppose to look like this? Except the last column in the array the rest are all zeros. I think I made some mistake in padding the text to sequence and if so can you tell me where I made the error?<\/p>\n","answers":[{"AnswerId":"42009085","CreationDate":"2017-02-02T17:52:53.640","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The problem is in this line:<\/p>\n\n<pre><code>tk = Tokenizer(nb_words=2000, filters=base_filter(), lower=True, split=\" \")\n<\/code><\/pre>\n\n<p>When you set such split (by <code>\" \"<\/code>), due to nature of your data, you'll get each sequence consisting of a single word. That's why your padded sequences have only one non-zero element. To change that try:<\/p>\n\n<pre><code>txt=\"a b c d e f g h i j k l m n \"*100\n<\/code><\/pre>\n"},{"AnswerId":"42084491","CreationDate":"2017-02-07T07:55:01.293","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>If you want to tokenize by char, you can do it manually, it's not too complex: <\/p>\n\n<p>First build a vocabulary for your characters:<\/p>\n\n<pre><code>txt=\"abcdefghijklmn\"*100\nvocab_char = {k: (v+1) for k, v in zip(set(txt), range(len(set(txt))))}\nvocab_char['&lt;PAD&gt;'] = 0\n<\/code><\/pre>\n\n<p>This will associate a distinct number for every character in your txt. The character with index 0 should be preserved for the padding. <\/p>\n\n<p>Having the reverse vocabulary will be usefull to decode the output.<\/p>\n\n<pre><code>rvocab = {v: k for k, v in vocab.items()}\n<\/code><\/pre>\n\n<p>Once you have this, you can first split your text into sequences, say you want to have sequences of length  <code>seq_len = 13<\/code> :<\/p>\n\n<pre><code>[[vocab_char[char] for char in txt[i:(i+seq_len)]] for i in range(0,len(txt),seq_len)]\n<\/code><\/pre>\n\n<p>your output will look like :<\/p>\n\n<pre><code>[[9, 12, 6, 10, 8, 7, 2, 1, 5, 13, 11, 4, 3], \n [14, 9, 12, 6, 10, 8, 7, 2, 1, 5, 13, 11, 4],\n ...,\n [2, 1, 5, 13, 11, 4, 3, 14, 9, 12, 6, 10, 8], \n [7, 2, 1, 5, 13, 11, 4, 3, 14]]\n<\/code><\/pre>\n\n<p>Note that the last sequence doesn't have the same length, you can discard it or pad your sequence to max_len = 13, it will add 0's to it. <\/p>\n\n<p>You can build your targets Y the same way, by shifting everything by 1. :-)<\/p>\n\n<p>I hope this helps.<\/p>\n"}]}
{"QuestionId":42002743,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-02T12:50:30.307","AcceptedAnswerId":42011617.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Will more memory be consumed if I handle secondary computations within a graph?","Body":"<p>If I have a trained Imagnet model from Google (inception-resnet-v2) and the model actually gives me two outputs back: the , and a list called end_points where I can extract the final predicted layer that has been performed with softmax activation, i.e. a variable called . However, this does not definitively give me the class label which I would require for predictions. To do so, I would either have to do  <strong>after<\/strong> I defined the train_op in the graph, so that I won't affect the original computation. <\/p>\n\n<p>Alternatively, I can use  which is computed out of the graph.<\/p>\n\n<p>My question is if I choose to do the first approach, would it consume more memory and affect my computation (in terms of the batch_size I can use)? Is it safer and better to just compute the necessary labels out of the graph?<\/p>\n","answers":[{"AnswerId":"42011617","CreationDate":"2017-02-02T20:21:43.947","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>When you issue multiple <code>.run<\/code> calls, the graph definition is cached. If you modify Graph, it'll need to re-encode it and send it again. So there may be a bit of extra memory used by <code>graph_def.SerializeToString<\/code> the first time you run modified graph, but it shouldn't affect <code>.run<\/code> steps after that.<\/p>\n\n<p>Relevant logic is in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/b00fc538638f87ac45be9105057b9865f0f9418b\/tensorflow\/python\/client\/session.py#L1037\" rel=\"nofollow noreferrer\">session.py<\/a>, note the line which checks <code>self._graph.version &gt; self._current_version<\/code><\/p>\n\n<pre><code> def _extend_graph(self):\n    # Ensure any changes to the graph are reflected in the runtime.\n    with self._extend_lock:\n      if self._graph.version &gt; self._current_version:\n        # pylint: disable=protected-access\n        graph_def, self._current_version = self._graph._as_graph_def(\n            from_version=self._current_version,\n            add_shapes=self._add_shapes)\n        # pylint: enable=protected-access\n\n        with errors.raise_exception_on_not_ok_status() as status:\n          tf_session.TF_ExtendGraph(\n              self._session, graph_def.SerializeToString(), status)\n        self._opened = True\n<\/code><\/pre>\n"}]}
{"QuestionId":42003846,"AnswerCount":2,"Tags":"<android><python><tensorflow>","CreationDate":"2017-02-02T13:45:52.373","AcceptedAnswerId":44973456.0,"OwnerUserId":1965273.0,"Title":"Retraining Inception5h model from TensorFlow Android Camera Demo","Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\/\" rel=\"nofollow noreferrer\">TensorFlow Android Camera Demo<\/a> uses <a href=\"https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\" rel=\"nofollow noreferrer\">Inception5h model<\/a> for live image recognition which delivers exceptional performance. Since I haven't had success retraining Inception5h I've gone with <a href=\"http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz\" rel=\"nofollow noreferrer\">InceptionV3 model<\/a> but it's not quite as snappy at image recognition. So I'm back at the beginning trying to retrain (or transfer learn) Inception5h model. I've tried modifying <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">retrain.py<\/a> but it's clearly written just for the v3 model. 5h model doesn't contain \"pool_3\/_reshape:0\", \"DecodeJpeg\/contents:0\" or \"ResizeBilinear:0\" tensors to begin with. There are other differences as well. <\/p>\n\n<p>I'm a bit of a newbie at machine learning and TensorFlow so I'd greatly appreciate clear steps as to what I have to do.<\/p>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"42022624","CreationDate":"2017-02-03T10:56:51.687","ParentId":null,"OwnerUserId":"6631259","Title":null,"Body":"<p>At the end of retrain.py script you can notice these lines:<\/p>\n\n<pre><code>output_graph_def = graph_util.convert_variables_to_constants(\n  sess, graph.as_graph_def(), [FLAGS.final_tensor_name])\nwith gfile.FastGFile(FLAGS.output_graph, 'wb') as f:\n   f.write(output_graph_def.SerializeToString())\n<\/code><\/pre>\n\n<p>Here all the variables are saved as constants in a protocol buffer (pb) file which is binary ('wb'). You should also save in a text file the names of the model's classes. Then as the android documentation mentions, you should save these 2 files in a folder named \"assets\" in the android path of tensorflow. Then there are some modifications that should be done to load the inception-v3 model which you can see here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1269\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1269<\/a>\nI hope this will help! <\/p>\n"},{"AnswerId":"44973456","CreationDate":"2017-07-07T14:35:00.970","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>It looks like the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\"><code>retrain.py<\/code><\/a> script and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/docs_src\/tutorials\/image_retraining.md\" rel=\"nofollow noreferrer\">tutorial<\/a> was just updated to work with the <a href=\"https:\/\/research.googleblog.com\/2017\/06\/mobilenets-open-source-models-for.html\" rel=\"nofollow noreferrer\">mobilenet<\/a> architecture.<\/p>\n\n<p>So that solves the first part of your problem, it's not actually inception5h, but it runs well on mobile with much better accuracy than inception5h.<\/p>\n\n<p>To actually get it to run in the android example you'll still need to update <a href=\"https:\/\/github.com\/googlecodelabs\/tensorflow-for-poets-2\/blob\/master\/android\/src\/org\/tensorflow\/demo\/ClassifierActivity.java#L61\" rel=\"nofollow noreferrer\">these settings<\/a>.<\/p>\n\n<p>I think you should be able to just copy <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py#L925\" rel=\"nofollow noreferrer\">the settings determined for the mobilenet you choose, from the retrain script<\/a> and you might be okay.<\/p>\n\n<hr>\n\n<p>If you wanted to use a different network, that didn't have the settings in <code>retrain.py<\/code> then the easiest way I can think of to determine them would be to explore the graph with TensorBoard.<\/p>\n\n<p>So if you <strong>really<\/strong> wanted to use inception 5h, you could download and unzip it:<\/p>\n\n<pre><code>curl -O https:\/\/storage.googleapis.com\/download.tensorflow.org\/models\/inception5h.zip\nunzip -d inception5h inception5h.zip\n<\/code><\/pre>\n\n<p>Then grab this simple script, from the <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets-2\/#0\" rel=\"nofollow noreferrer\">Tensorflow for Poets: 2 codelab<\/a> <a href=\"https:\/\/github.com\/googlecodelabs\/tensorflow-for-poets-2\/\" rel=\"nofollow noreferrer\">repo<\/a>, to convert the graph <code>.pb<\/code> file to something tensorboard can use:<\/p>\n\n<pre><code>curl -O https:\/\/raw.githubusercontent.com\/googlecodelabs\/tensorflow-for-poets-2\/master\/scripts\/graph_pb2tb.py\n<\/code><\/pre>\n\n<p>And run it on your graph.pb:<\/p>\n\n<pre><code>mkdir tb_graph\npython graph_pb2tb.py tb\/inception5h inception5h\/tensorflow_inception_graph.pb \n<\/code><\/pre>\n\n<p>And <a href=\"http:\/\/0.0.0.0:6006#graphs\" rel=\"nofollow noreferrer\">open it in tensorboard<\/a>:<\/p>\n\n<pre><code>tensorboard --logdir tb_graph\n<\/code><\/pre>\n\n<p>Then it might be relatively simple to poke around in the graph and find the names of the nodes you need to fill up your own <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py#L925\" rel=\"nofollow noreferrer\"><code>model_info<\/code><\/a> dict.<\/p>\n\n<p>I think this is the node you'd want to set as your <code>bottleneck_tensor<\/code>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LZQMT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LZQMT.png\" alt=\"TensorBoard screenshot of inception 5h with avgpool0\/reshape highlighted\"><\/a><\/p>\n"}]}
{"QuestionId":42004584,"AnswerCount":0,"Tags":"<python><numpy><deep-learning><theano><lasagne>","CreationDate":"2017-02-02T14:19:41.230","AcceptedAnswerId":null,"OwnerUserId":4326513.0,"Title":"Multi-label classification with Theano and Lasagne?","Body":"<p>I am trying to make the  tutorial into multi-label classification using CNN. I am very new on this field and my goal is to understand the architecture, the input-output data format in order to be more familiar and run my own multi-label classification problems.<\/p>\n\n<p>From what I have read until now the modifications I have done are:<\/p>\n\n<ol>\n<li><p>Change the training and test labels to 2d numpy arrays where the first dimension (rows) corresponds to instances and the second dimension (columns) to the specific label(s) corresponding to each instance. For mnist example if I want to convert it to a 10 multi-label classification the y_train &amp; y_test should be like<\/p>\n\n<p><\/p><\/li>\n<li><p>On the network code we change the non-linearity in the final layer from a softmax to a sigmoid. <\/p>\n\n<p><\/p>\n\n<p>The  remain 10 right? Again it is a 10-class classification problem or not?<\/p><\/li>\n<li><p>On theano variables we change the  from . Then according to <a href=\"https:\/\/stackoverflow.com\/questions\/35356608\/input-dimension-mismatch-binary-crossentropy-lasagne-and-theano\" title=\"this post\">this post<\/a> I change the loss functions from categorical cross-entropy to binary cross-entropy<\/p>\n\n<p><\/p>\n\n<p> <\/p><\/li>\n<li><p>I also change the test_acc variable because I was getting error with mismatch input size. From the previous post this solution seems to work<\/p>\n\n<p><\/p><\/li>\n<\/ol>\n\n<p>Well, the I get no errors now but I am not pretty sure that the code is correct. When you are doing muli-label classifaction you have to define other metrics than the typical accuracy, right? Any suggestions??<\/p>\n","answers":[]}
{"QuestionId":42004948,"AnswerCount":2,"Tags":"<keras>","CreationDate":"2017-02-02T14:34:41.507","AcceptedAnswerId":42014657.0,"OwnerUserId":4897119.0,"Title":"Accuracy from callback and progress bar in Keras doesnt match","Body":"<p>I'm trying to learn Keras and are using LSTM for a classification problem. I want to be able to plot the\naccuracy and loss and update the plot during training. For that I'm using the <a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">callback function<\/a>.\nFor some reason the accuracy and loss I receive in the callbacks does not match with\nthe accuracy and loss printed by the fit function. <\/p>\n\n<p>Here are the relevant lines of my code:<\/p>\n\n<pre><\/pre>\n\n<p>When running the program I get this output (first row of each batch is printed by callback, second is printed by model.fit):<\/p>\n\n<pre><\/pre>\n\n<p>I have tried to print logs.get('acc'), as well as saved the accuracies to a list in the PlotCallbacks object and print the list, but the problem remains. <\/p>\n\n<p>Does anyone have a clue what the problem may be?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42014657","CreationDate":"2017-02-03T00:09:52.267","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p><code>on_batch_end()<\/code> type callback function gets the accuracy of the batch that just got trained. Whereas the logs printed by keras is the average over all the batches that it has seen in the current epoch. You can easily observe that in your logs.. say in first 2 batches one accuracy was <code>0.0<\/code> and <code>1.0<\/code>, which made the overall accuracy over 2 batches seen as <code>0.5000<\/code>. <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/utils\/generic_utils.py#L267-L276\" rel=\"nofollow noreferrer\">here<\/a> is exactly where the average is calculated. <\/p>\n\n<p>Also accuracy as a metric is usually reported from epoch to epoch so you can change the callback to <code>on_epoch_end()<\/code>. <\/p>\n"},{"AnswerId":"51935113","CreationDate":"2018-08-20T16:36:57.530","ParentId":null,"OwnerUserId":"2835597","Title":null,"Body":"<p>Here is a simple callback that records the accuracy in the way defined by Keras' progbar (average over all the batches seen so far in the current epoch):<\/p>\n\n<pre><code>class AccHistory(Callback):\n    def on_train_begin(self, logs={}):\n        self.accs = []\n        self.acc_avg = []\n\n    def on_batch_end(self, batch, logs={}):\n        self.accs.append(round(1e2*float(logs.get('acc')),4))\n        self.acc_avg.append(round(np.mean(self.accs,dtype=np.float64),4))\n\n    def on_epoch_end(self, batch, logs={}):\n        self.accs = []\n<\/code><\/pre>\n"}]}
{"QuestionId":42004995,"AnswerCount":2,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2017-02-02T14:36:41.900","AcceptedAnswerId":44783729.0,"OwnerUserId":3380374.0,"Title":"Slice multiple slices at once with tensorflow","Body":"<p>I am preparing the input tensor for the tensorflow RNN.<br>\nCurrently I am doing the following<\/p>\n\n<pre><\/pre>\n\n<p>Would it be possible to do this at once, without loop, with some tensorflow function?<\/p>\n","answers":[{"AnswerId":"42013306","CreationDate":"2017-02-02T22:13:44.607","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>Try <code>tf.split<\/code> or <code>tf.split_v<\/code>. See here:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/split\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/split<\/a><\/p>\n\n<p>Does that help?<\/p>\n"},{"AnswerId":"44783729","CreationDate":"2017-06-27T15:04:09.840","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p>As suggested by Peter Hawkins, you can use <code>gather_nd<\/code> with the appropriate indices to get there.<\/p>\n\n<p>Your uniform cropping on the inner dimension can simply be done before the call to <code>gather_nd<\/code>.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nsess = tf.InteractiveSession()\n\n# integer image simply because it is more readable to me\nim0 = np.random.randint(10, size=(20,20))\nim = tf.constant(im0)\n\nmax_steps = 3\nbatch_size = 10\n\n# create the appropriate indices here\nindices = (np.arange(max_steps) +\n    np.arange(batch_size)[:,np.newaxis])[...,np.newaxis]\n# crop then call gather_nd\nres = tf.gather_nd(im[:,:10], indices).eval()\n\n# check that the resulting tensors are equal to what you had previously\nfor each in range(batch_size):\n  assert(np.all(tf.slice(im, [each,0],[max_steps,10]).eval() == res[each]))\n<\/code><\/pre>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>If your slices indices are in a tensor, you simply replace numpy's operations with tensorflow's operations when creating <code>indices<\/code>:<\/p>\n\n<pre><code># indices stored in a 1D array\nmy_indices = tf.constant([1, 8, 3, 0, 0])\nindices = (np.arange(max_steps) +\n    my_indices[:,tf.newaxis])[...,tf.newaxis]\n<\/code><\/pre>\n\n<p>Further remarks:<\/p>\n\n<ul>\n<li><code>indices<\/code> is created by taking advantage of <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">broadcasting<\/a> during the addition: arrays are virtually tiled so that their dimensions match. Broadcasting is supported by numpy and by tensorflow in a similar fashion.<\/li>\n<li>Ellipsis <code>...<\/code> is part of the standard <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html\" rel=\"nofollow noreferrer\">numpy slicing notation<\/a>, it basically fills all remaining dimensions left by the other slicing indices. So <code>[..., newaxis]<\/code> is basically equivalent to <code>expand_dims(\u00b7, -1)<\/code>.<\/li>\n<\/ul>\n"}]}
{"QuestionId":42006089,"AnswerCount":1,"Tags":"<variables><tensorflow>","CreationDate":"2017-02-02T15:26:29.907","AcceptedAnswerId":42007040.0,"OwnerUserId":7474000.0,"Title":"Reason why setting tensorflow's variable with small stddev","Body":"<p>I have a question about a reason why setting TensorFlow's variable with small .\nI guess many people do test MNIST test code from TensorFlow beginner's guide.\nAs following it, the first layer's weights are initiated by using  with  0.1.\nAnd I guessed if setting it with more bigger value, then it would be the same result, which is exactly accurate. \nBut although increasing epoch count, it doesn't work.\nIs there anybody know this reason?<\/p>\n\n<p>original :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>modified :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42007040","CreationDate":"2017-02-02T16:11:02.443","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>The reason is because you want to keep all the layer's variances (or standard deviations) approximately the same, and sane. It has to do with the error backpropagation step of the learning process and the activation functions used.<\/p>\n\n<p>In order to learn the network's weights, the backpropagation step requires knowledge of the network's <em>gradient<\/em>, a measure of how strong each weight influences the input to reach the final output; layer's weight variance directly influences the propagation of gradients.<\/p>\n\n<p>Say, for example, that the activation function is sigmoidal (e.g. <code>tf.nn.sigmoid<\/code> or <code>tf.nn.tanh<\/code>); this implies that all input values are squashed into a fixed output value range. For the sigmoid, it is the range <code>0..1<\/code>, where essentially all values <code>z<\/code> greater or smaller than +\/- 4 are very close to one (for <code>z &gt; 4<\/code>) or zero (for <code>z &lt; -4<\/code>) and only values within that range tend to have some meaningful \"change\".<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/auTc9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/auTc9.png\" alt=\"Sigmoid\"><\/a><\/p>\n\n<p>Now the difference between the values <code>sigmoid(5)<\/code> and <code>sigmoid(1000)<\/code> is barely noticeable. Because of that, all very large or very small values will optimize very slowly, since their influence on the result <code>y = sigmoid(W*x+b)<\/code> is extremely small. Now the pre-activation value <code>z = W*x+b<\/code> (where <code>x<\/code> is the input) depends on the actual input <code>x<\/code> and the current weights <code>W<\/code>. If either of them is large, e.g. by initializing the weights with a high variance (i.e. standard deviation), the result will necessarily be (relatively) large, leading to said problem. This is also the reason why <code>truncated_normal<\/code> is used rather than a correct normal distribution: The latter only guarantees that <em>most of<\/em> the values are very close to the mean, with some less than 5% chance that this is not the case, while <code>truncated_normal<\/code> simply clips away every value that is too big or too small, guaranteeing that all weights are in the same range, while still being normally distributed.<\/p>\n\n<p>To make matters worse, in a typical neural network - especially in deep learning - each network layer is followed by one or many others. If in each layer the output value range is big, the gradients will get bigger and bigger as well; this is known as the <em>exploding gradients<\/em> problem (a variation of the <em>vanishing gradients<\/em>, where gradients are getting smaller).\nThe reason that this is a problem is because learning starts at the very last layer and each weight is adjusted depending on how much it contributed to the error. If the gradients are indeed getting very big towards the end, the very last layer is the first one to pay a high toll for this: Its weights get adjusted very strongly - likely overcorrecting the actual problem - and then only the \"remaining\" error gets propagated further back, or up, the network. Here, since the last layer was already \"fixed a lot\" regarding the measured error, only smaller adjustments will be made. This may lead to the problem that the first layers are corrected only by a tiny bit or not at all, effectively preventing all learning there.  The same basically happens if the learning rate is too big.<\/p>\n\n<p>Finding the best weight initialization is a topic by itself and there are somewhat more sophisticated methods such as <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/contrib.layers\/initializers#xavier_initializer\" rel=\"nofollow noreferrer\">Xavier initialization<\/a> or Layer-sequential unit variance, however small normally distributed values are usually simply a good guess.<\/p>\n"}]}
{"QuestionId":42006320,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2017-02-02T15:38:00.270","AcceptedAnswerId":null,"OwnerUserId":7490399.0,"Title":"TensorFlow pip installation issue: cannot import name 'descriptor'","Body":"<p>I'm seeing the following error when installing TensorFlow:<\/p>\n\n<pre>\nImportError: Traceback (most recent call last):\nFile \"...\/graph_pb2.py\", line 6, in \nfrom google.protobuf import descriptor as _descriptor\nImportError: cannot import name 'descriptor'\n<\/pre>\n","answers":[{"AnswerId":"42006351","CreationDate":"2017-02-02T15:39:56.823","ParentId":null,"OwnerUserId":"7490399","Title":null,"Body":"<p>This error signals a mismatch between protobuf and TensorFlow versions.<\/p>\n\n<p>Take the following steps to fix this error:<\/p>\n\n<ol>\n<li>Uninstall TensorFlow.<\/li>\n<li>Uninstall protobuf (if protobuf is installed).<\/li>\n<li>Reinstall TensorFlow, which will also install the correct protobuf dependency.<\/li>\n<\/ol>\n"},{"AnswerId":"49096146","CreationDate":"2018-03-04T13:50:45.817","ParentId":null,"OwnerUserId":"9441725","Title":null,"Body":"<p>I faced the similar issue , after trial and error, i used the below logic to run the program.\n$pip install --upgrade --no-deps --force-reinstall tensorflow<\/p>\n\n<p>this will make sure to uninstall and reinstall the program from fresh.it works!!!<\/p>\n"},{"AnswerId":"54955702","CreationDate":"2019-03-02T06:02:37.743","ParentId":null,"OwnerUserId":"10331178","Title":null,"Body":"<p>I would be extra careful before uninstalling\/reinstalling other packages such as protobuf. What I think would most likely be the issue is difference in versions. As of writing this, the most recent release of python is 3.7 while tensorflow is only compatible up to 3.6. <\/p>\n\n<p>If you're using a 3rd party distribution like Anaconda, this can get hidden from you. In this case I would recommend creating a new environment in Anaconda, with python 3.6 and then installing tensorflow: <a href=\"https:\/\/conda.io\/projects\/conda\/en\/latest\/user-guide\/getting-started.html#managing-python\" rel=\"nofollow noreferrer\">https:\/\/conda.io\/projects\/conda\/en\/latest\/user-guide\/getting-started.html#managing-python<\/a><\/p>\n"},{"AnswerId":"42007152","CreationDate":"2017-02-02T16:17:09.140","ParentId":null,"OwnerUserId":"7173909","Title":null,"Body":"<p>Try this:<\/p>\n\n<ol>\n<li><p>pip uninstall protobuf<\/p><\/li>\n<li><p>brew install protobuf<\/p><\/li>\n<li><p>mkdir -p\n\/Users\/alexeibendebury\/Library\/Python\/2.7\/lib\/python\/site-packages<\/p><\/li>\n<li><p>echo 'import site;\nsite.addsitedir(\"\/usr\/local\/lib\/python2.7\/site-packages\")' >>\n\/Users\/alexeibendebury\/Library\/Python\/2.7\/lib\/python\/site-packages\/homebrew.pth<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42006965,"AnswerCount":1,"Tags":"<tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-02-02T16:07:31.653","AcceptedAnswerId":null,"OwnerUserId":1036017.0,"Title":"How to create a specialized machine vision OCR solution?","Body":"<p>We need to read some text from photos of sales receipts taken by iPad camera. Here is a sample similar to what we need to read from:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/duEX6.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/duEX6.png\" alt=\"Receipts\"><\/a><\/p>\n\n<p>There are a few constraints to this problem:<\/p>\n\n<ol>\n<li>We need to read the total amount which always appears after a text marker (such as <strong>Grand Total<\/strong> in this example).<\/li>\n<li>The font is always the same.<\/li>\n<li>The app must work offline without network connectivity.<\/li>\n<\/ol>\n\n<p>This is what we have tried so far:<\/p>\n\n<ul>\n<li>Google Mobile Vision text extraction worked like magic. But text extraction is available in Android only. And we need to build the solution in iOS.<\/li>\n<li>Google and Microsoft have cloud based machine vision solutions which also work very accurately. But our app needs to work offline.<\/li>\n<li>Use tesseract OCR. It performed very poorly. No doubt because we have a photo instead of scanned black &amp; white image.<\/li>\n<\/ul>\n\n<p>We are now thinking of creating a custom solution using convolutional NN. The question I have is how can we build a model that takes advantage of these two constraints to create a simpler and yet very accurate solution?<\/p>\n\n<ol>\n<li>The total amount always appears after a text marker. We can safely ignore the rest of the text.<\/li>\n<li>The text is always in English and in the same font.<\/li>\n<\/ol>\n\n<p>This is the general pipeline we have come up with so far.<\/p>\n\n<ol>\n<li>Straighten the image and scale it to a standard size.<\/li>\n<li>Doing conv net to locate the text marker (<strong>Grad Total<\/strong>) should be fairly easy. We can completely skip the top half of the image.<\/li>\n<\/ol>\n\n<p>We are not sure what else to do at this point. Any tips, advice and help will be great.<\/p>\n\n<p>PS. I realize this is a question about design methodology and not a specific programming question. I apologize if this violates SO guidelines.<\/p>\n","answers":[{"AnswerId":"46757905","CreationDate":"2017-10-15T17:27:10.477","ParentId":null,"OwnerUserId":"677824","Title":null,"Body":"<p>I propose for you to consider deeplearning4j.org solution. You can train their network on powerful machine and then save state of network and use it at android. <a href=\"https:\/\/deeplearning4j.org\/android\" rel=\"nofollow noreferrer\">Here<\/a> they explained how to use their network at android app with help of java. <\/p>\n"}]}
{"QuestionId":42006967,"AnswerCount":0,"Tags":"<tensorflow><distributed>","CreationDate":"2017-02-02T16:07:35.880","AcceptedAnswerId":null,"OwnerUserId":7450158.0,"Title":"Scalability issues related to distributed tensorflow","Body":"<p>I was running some experiments to test the scalability of distributed tensorflow. My goal is to check how is the speedup of using 4gpus on 2nodes (in total 8gpus) compared with using single gpu.<\/p>\n\n<p>I chose ResNet as the network structure and modified the ResNet example code from the official model zoos (github.com\/tensorflow\/models\/tree\/master\/resnet) to be the one can do distributed training.\nIn the experiments, ResNet 56 layers is used for training on CIFAR10 data set and the settings follow the original paper.<\/p>\n\n<p>First thing I would like to ensure is by using Synchronous SGD (SyncReplicasOptimizer), 8 gpus can get a similar curve as 1gpu in terms of <strong>epoch-to-accuracy<\/strong> figure. I used the same effective batch size (i.e., ) on the two experiments.<\/p>\n\n<p>Please check the experimental figures<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pnNTc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pnNTc.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<blockquote>\n  <p>tf-0: single gpu version<\/p>\n  \n  <p>tf-1: distributed version (2ps, 8workers)<\/p>\n  \n  <p>x-axis: #epoch (per one means going through the whole training data once)<\/p>\n  \n  <p>y-axis: accuracy of CIFAR-10 <strong>testing data<\/strong><\/p>\n<\/blockquote>\n\n<p>The single gpu (<strong>tf-0<\/strong>) result is similar to the original paper and the ones reproduced by many other people (e.g., <a href=\"http:\/\/torch.ch\/blog\/2016\/02\/04\/resnets.html\" rel=\"nofollow noreferrer\">http:\/\/torch.ch\/blog\/2016\/02\/04\/resnets.html<\/a>, <a href=\"https:\/\/github.com\/ppwwyyxx\/tensorpack\/tree\/master\/examples\/ResNet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ppwwyyxx\/tensorpack\/tree\/master\/examples\/ResNet<\/a>, or more examples here <a href=\"https:\/\/github.com\/KaimingHe\/deep-residual-networks#third-party-re-implementations\" rel=\"nofollow noreferrer\">https:\/\/github.com\/KaimingHe\/deep-residual-networks#third-party-re-implementations<\/a>).<\/p>\n\n<p>However, you can observe that <strong>tf-1<\/strong> (8gpus curve) has a very different pattern, which is not what I expected. My understanding of Synchronous SGD is that optimizer accumulate  gradients and then apply to the variables. So the procedure should be exactly the same as single gpu if the same effective batch size is set.<\/p>\n\n<p>Interestingly, when I set #worker = 2 in the distributed version, the curve (<strong>tf-2<\/strong>) become close to the single gpu version. But 4gpus (<strong>tf-3<\/strong>) got closer to <strong>tf-1<\/strong>.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/D5Vup.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/D5Vup.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<blockquote>\n  <p>tf-2: distributed version (2ps, 2workers)<\/p>\n  \n  <p>tf-3: distributed version (1ps, 4workers)<\/p>\n<\/blockquote>\n\n<p>Here are some detail settings (basically follows the original paper):<\/p>\n\n<blockquote>\n  <p>batch size=128, num_residual_units=9, relu_leakiness=0<\/p>\n  \n  <p>learning rate scheduler: 0.1 for 1-80 epoch, 0.01 for 81-120 epoch, 0.001 for 120-160 epoch <\/p>\n<\/blockquote>\n\n<p>Environments:<\/p>\n\n<blockquote>\n  <p>Tensorflow version: 0.12.1<\/p>\n  \n  <p>Installed version of CUDA and cuDNN: CUDA 7.5, cuDNN 5.1<\/p>\n<\/blockquote>\n\n<p>The single gpu version code can be found <a href=\"http:\/\/github.com\/tensorflow\/models\/blob\/22036b6f44efeb648fe2b2b4551f2b982edf0f5f\/resnet\/resnet_main.py\" rel=\"nofollow noreferrer\">here<\/a> <\/p>\n\n<p>The code for distributed version can be found <a href=\"http:\/\/gist.github.com\/infwinston\/d01721d9609eef4a38b8e5e57719b194\" rel=\"nofollow noreferrer\">here<\/a> .<\/p>\n\n<p>Does anyone has ideas why it got gradually worse when #workers increase? Thanks.<\/p>\n\n<hr>\n\n<p><strong>Updated figure<\/strong><\/p>\n\n<p>According to commit ecdee38a (I think v0.12.1 includes this commit), tensorflow seems to use <strong>averaging<\/strong> instead of <strong>summing up<\/strong> gradients calculated by each worker. so I tried the one with 4x learning rate in 4gpus case. But got a even worse result.<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/Ogwtp.png\" alt=\"experiments of modified learning rate\"><\/p>\n\n<blockquote>\n  <p>1xlr: 1x learning rate (0.1\/0.01\/0.001)<\/p>\n  \n  <p>4xlr: 4x learning rate (0.4\/0.04\/0.004)<\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":42007193,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform><google-cloud-ml>","CreationDate":"2017-02-02T16:19:30.140","AcceptedAnswerId":42007789.0,"OwnerUserId":3627572.0,"Title":"Does Google Cloud ML support GPU?","Body":"<p>I'm testing Google Cloud ML for speeding up my ML model using Tensorflow.<\/p>\n\n<p>Unfortunately, it seems like Google Cloud ML is extremely slow. My Mainstream-Level PC is at least 10x faster than Google Cloud ML.<\/p>\n\n<p>I doubt it uses GPU, so I did a test. I modified a sample <a href=\"https:\/\/github.com\/GoogleCloudPlatform\/cloudml-samples\/blob\/38fb21771da3fe6fae661c343f2ece8eaba1c8c5\/mnist\/trainable\/trainer\/task.py\" rel=\"nofollow noreferrer\">code<\/a> to force using GPU.<\/p>\n\n<pre><\/pre>\n\n<p>This training code works at my PC () but not in cloud. It gives errors like this:<\/p>\n\n<pre><\/pre>\n\n<p>Does Google Cloud ML support GPU?<\/p>\n","answers":[{"AnswerId":"42007789","CreationDate":"2017-02-02T16:47:13.673","ParentId":null,"OwnerUserId":"4392784","Title":null,"Body":"<p>GPUs are now in Beta and all Cloud ML customers have access.<\/p>\n\n<p>Here are the <a href=\"https:\/\/cloud.google.com\/ml\/docs\/how-tos\/using-gpus\" rel=\"nofollow noreferrer\">docs<\/a> for using GPUs with Cloud ML.<\/p>\n"}]}
{"QuestionId":42007591,"AnswerCount":1,"Tags":"<python><keras><object-detection><training-data>","CreationDate":"2017-02-02T16:38:49.780","AcceptedAnswerId":42181934.0,"OwnerUserId":4508333.0,"Title":"Single object detection keras","Body":"<p>I want to make a system that recognizes a single object using keras. In my case I will be detecting car wheels. How do I train my system just for 1 object? I did classification task before using cats and dogs, but now its a completely different task. Do I still \"classify\", with class 0= wheels, class = 1 non wheels(just random images of anything)? How do I do following steps in this problem?\n1) Train system for 1 object\n2) Detect object(sliding window or heatmap)<\/p>\n","answers":[{"AnswerId":"42181934","CreationDate":"2017-02-11T22:24:20.897","ParentId":null,"OwnerUserId":"7544784","Title":null,"Body":"<p>Your task is a so-called binary classification. \nMake sure, that your final layer has got only one neuron (e.g. for Sequential model model.add(Dense(1, ... other parameters ... ))) and use the binary_crossentropy as loss function.<\/p>\n\n<p>Hope this helps. <\/p>\n"}]}
{"QuestionId":42007638,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-02-02T16:40:33.603","AcceptedAnswerId":null,"OwnerUserId":7506591.0,"Title":"Can I change Inv operation into Reciprocal in an existing graph in Tensorflow?","Body":"<p>I am working on an image classification problem with tensorflow. I have 2 different CNNs trained separately (in fact 3 in total but I will deal with the third later), for different tasks and on a AWS (Amazon) machine. One tells if there is text in the image and the other one tells if the image is safe for work or not. Now I want to use them in a single script on my computer, so that I can put an image as input and get the results of both networks as output.<\/p>\n\n<p>I load the two graphs in a single tensorflow Session, using the  API and the import_scope argument and putting each subgraph in a separate scope. Then I just use the  method of the created saver, giving it the common Session as argument.<\/p>\n\n<p>Then, in order to run inference, I retrieve the placeholders and final output with  and  before using it in  (I think I could just have put 'name' in  instead of fetching the output tensor and putting it in a variable, but this is not my problem).<\/p>\n\n<p><strong>My problem is<\/strong> the text CNN works perfectly fine, but the nsfw detector always gives me the same output, no matter the input (even with ). I have tried both separately and same story: text works but not nsfw. So I don't think the problem comes from using two networks simultaneaously.<\/p>\n\n<p>I also tried on the original AWS machine I trained it on, and this time the nsfw CNN worked perfectly.<\/p>\n\n<p>Both networks are very similar. I checked on <a href=\"https:\/\/i.stack.imgur.com\/Bwv6S.png\" rel=\"nofollow noreferrer\">Tensorboard<\/a> if everything was fine and I think it is ok. The differences are in the number of hidden units and the fact that I use batch normalization in the nsfw model and not in the text one. Now <strong>why this title<\/strong> ? I observed that I had a warning when running the nsfw model that I didn't have when using only the text model:<\/p>\n\n<blockquote>\n  <p>W tensorflow\/core\/framework\/op_def_util.cc:332] Op Inv is deprecated. It will cease to work in GraphDef version 17. Use Reciprocal.<\/p>\n<\/blockquote>\n\n<p>So I thougt maybe this was the reason, everything else being equal. I checked my GraphDef version, which seems to be 11, so Inv should still work in theory. By the way the AWS machine use tensroflow version 0.10 and I use version 0.12.<\/p>\n\n<p>I noticed that the text network only had one Inv operation (via a filtering on the names of the operations given by ), and that the nsfw model had the same operation plus multiple Inv operations due to the batch normalization layers. As precised in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\" rel=\"nofollow noreferrer\">release notes<\/a>,  has simply been renamed to , so I tried to change the names of the operations to Reciprocal with , as proposed <a href=\"https:\/\/stackoverflow.com\/questions\/37949588\/is-there-an-no-op-pass-through-operation-in-tensorflow\">here<\/a>, but it didn't work. I have seen that using  and changing the name could also work, but from what I understand, tensorflow graphs are an append-only structure, so we can't really modify its operations (which seems to be immutable anyway).<\/p>\n\n<p><strong>The thing is<\/strong>:<\/p>\n\n<ol>\n<li>as I said, the Inv operation should still work in my GraphDef version;<\/li>\n<li>this is only a warning;<\/li>\n<li>the Inv operations only appear under name scopes that begin with 'gradients' so, from my understanding, this shouldn't be used for inference;<\/li>\n<li>the text model also have an Inv operation.<\/li>\n<\/ol>\n\n<p>For these reasons, I have a big doubt on my diagnosis. So <strong>my final questions are<\/strong>:<\/p>\n\n<ul>\n<li>do you have another diagnosis?<\/li>\n<li>if mine is correct, is it possible to replace Inv operations with Reciprocal operations, or do you have any other solution?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"42021969","CreationDate":"2017-02-03T10:23:18.947","ParentId":null,"OwnerUserId":"7506591","Title":null,"Body":"<p>After a thorough examination of the output of relevant nodes, with the help of Tensorboard, I am now pretty certain that the renaming of Inv to Reciprocal has nothing to do with my problem.<\/p>\n\n<p>It appears that the last batch normalization layer eliminates almost any variance of its output when the inputs varies. I will ask why elsewhere.<\/p>\n"}]}
{"QuestionId":42009190,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-02T17:58:25.260","AcceptedAnswerId":null,"OwnerUserId":7490399.0,"Title":"TensorFlow installation error: Operation not permitted ...\/_markerlib","Body":"<p>Despite using , the  command sometimes generates an error similar to the following:<\/p>\n\n<pre>\n...\nInstalling collected packages: setuptools, protobuf, wheel, numpy, tensorflow\nFound existing installation: setuptools 1.1.6\nUninstalling setuptools-1.1.6:\nException:\n\u2026\n[Errno 1] Operation not permitted: \n'\/tmp\/pip-a1DXRT-uninstall\/System\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/Extras\/lib\/python\/_markerlib'\n<\/pre>\n","answers":[{"AnswerId":"45916589","CreationDate":"2017-08-28T10:10:30.430","ParentId":null,"OwnerUserId":"785212","Title":null,"Body":"<p>Try to install using <code>--ignore-installed<\/code> option:<\/p>\n\n<pre><code>sudo pip install --ignore-installed tensorflow\n<\/code><\/pre>\n\n<p>It should work.<\/p>\n"}]}
{"QuestionId":42009335,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-02T18:06:53.613","AcceptedAnswerId":null,"OwnerUserId":7507448.0,"Title":"Tensorflow transfer learning error - invalid literal for float(): 0&324597","Body":"<p>I am running a tensorflow transfer learning retrain session in Ubuntun 16.10 with the following code:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>The retraining session has encountered an 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\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>Can't figure out why. Any pointers to solving this issues will be of great help!<\/p>\n\n<p>Cheers,<\/p>\n\n<p>Rahul<\/p>\n","answers":[]}
{"QuestionId":42010114,"AnswerCount":2,"Tags":"<tensorflow><translate>","CreationDate":"2017-02-02T18:52:16.643","AcceptedAnswerId":42257055.0,"OwnerUserId":470184.0,"Title":"Training custom dataset with translate model","Body":"<p>Running the model out of the box generates these files in the data dir : <\/p>\n\n<pre><\/pre>\n\n<p>Reading the src of translate.py : <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/translate\/translate.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/translate\/translate.py<\/a><\/p>\n\n<pre><\/pre>\n\n<p>To utilize my own training data I created dirs my-from-train-data &amp; to-from-train-data and add my own training data to each of these dirs, training data is contained in the files mydata.from &amp; mydata.to<\/p>\n\n<pre><\/pre>\n\n<p>I could not find documentation as to using own training data or what format it should take so I inferred this from the translate.py src and contents of data dir created when executing translate model out of the box.<\/p>\n\n<p>Contents of mydata.from : <\/p>\n\n<pre><\/pre>\n\n<p>Contents of mydata.to : <\/p>\n\n<pre><\/pre>\n\n<p>I then attempt to train the model using : <\/p>\n\n<pre><\/pre>\n\n<p>This returns with an error : <\/p>\n\n<pre><\/pre>\n\n<p>Appears I need to create file my-from-train-data.ids40000 , what should it's contents be ? Is there an example of how to train this model using custom data ?<\/p>\n","answers":[{"AnswerId":"42258406","CreationDate":"2017-02-15T19:28:23.987","ParentId":null,"OwnerUserId":"539075","Title":null,"Body":"<p>Quick answer to :<\/p>\n\n<blockquote>\n  <p>Appears I need to create file my-from-train-data.ids40000 , what should it's contents be ? Is there an example of how to train this model using custom data ?<\/p>\n<\/blockquote>\n\n<p>Yes, that's the vocab\/ word-id file missing, which is generated when preparing to create the data.<\/p>\n\n<p>Here is a tutorial from the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/\" rel=\"nofollow noreferrer\">Tesnorflow documentation<\/a>.<\/p>\n\n<p>quick over-view of the files and why you might be confused by the files outputted vs what to use:<\/p>\n\n<ul>\n<li><code>python\/ops\/seq2seq.py<\/code>: >>   Library for building sequence-to-sequence models.<\/li>\n<li><code>models\/rnn\/translate\/seq2seq_model.py<\/code>: >>   Neural translation sequence-to-sequence model.<\/li>\n<li><code>models\/rnn\/translate\/data_utils.py<\/code>: >>  Helper functions for preparing translation data.<\/li>\n<li><code>models\/rnn\/translate\/translate.py<\/code>: >>   Binary that trains and runs the translation model.<\/li>\n<\/ul>\n\n<p>The Tensorflow <code>translate.py<\/code> file requires several files to be generated when using your own corpus to translate.<\/p>\n\n<ol>\n<li><p>It needs to be aligned, meaning: <code>language line 1 in file 1.<\/code> &lt;> <code>language line 1 file 2.<\/code> This\nallows the model to do encoding and decoding.<\/p><\/li>\n<li><p>You want to make sure the Vocabulary have been generated from the dataset using this file:\nCheck these steps:<\/p><\/li>\n<\/ol>\n\n<p><code>python translate.py\n  --data_dir [your_data_directory] --train_dir [checkpoints_directory]\n  --en_vocab_size=40000 --fr_vocab_size=40000<\/code> <\/p>\n\n<p>Note! If the Vocab-size is lower, then change that value.<\/p>\n\n<p>There is a longer discussion here <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/600\" rel=\"nofollow noreferrer\">tensorflow\/issues\/600 <\/a><\/p>\n\n<p>If all else fails, check out this <a href=\"https:\/\/github.com\/buriburisuri\/ByteNet\" rel=\"nofollow noreferrer\">ByteNet<\/a> implementation in Tensorflow which does translation task as well.<\/p>\n"},{"AnswerId":"42257055","CreationDate":"2017-02-15T18:11:38.957","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>blue-sky<\/p>\n\n<p>Great question, training a model on your own data is way more fun than using the standard data. An example of what you could put in the terminal is: <\/p>\n\n<p><code>python translate.py --from_train_data mydatadir\/to_translate.in --to_train_data mydatadir\/to_translate.out --from_dev_data mydatadir\/test_to_translate.in --to_dev_data mydatadir\/test_to_translate.out --train_dir train_dir_model --data_dir mydatadir<\/code><\/p>\n\n<p>What goes wrong in your example is that you are not pointing to a file, but to a folder. from_train_data should always point to a plaintext file, whose rows should be aligned with those in the to_train_data file. <\/p>\n\n<p>Also: as soon as you run this script with sensible data (more than one line ;) ), translate.py will generate your ids (40.000 if from_vocab_size and to_vocab_size are not set). Important to know is that this file is created in the folder specified by data_dir... if you do not specify one this means they are generated in \/tmp (I prefer them at the same place as my data). <\/p>\n\n<p>Hope this helps! <\/p>\n"}]}
{"QuestionId":42010699,"AnswerCount":2,"Tags":"<lua><computer-vision><deep-learning><torch>","CreationDate":"2017-02-02T19:25:42.510","AcceptedAnswerId":42035109.0,"OwnerUserId":1749629.0,"Title":"How to interpret the output of deepmask getTopScores function","Body":"<p>I am using facebookresearch's <a href=\"https:\/\/github.com\/facebookresearch\/deepmask\" rel=\"nofollow noreferrer\">deepmask<\/a> to segment an image. I am trying to modify the <a href=\"https:\/\/github.com\/facebookresearch\/deepmask\/blob\/master\/computeProposals.lua\" rel=\"nofollow noreferrer\">computeProposals.lua<\/a> module to segment my own images. The key function is getTopProps. In the code block below it is being used to find the masks associated with the top proposals.<\/p>\n\n<pre><\/pre>\n\n<p>The infer class is from the <a href=\"https:\/\/github.com\/facebookresearch\/deepmask\/blob\/master\/InferDeepMask.lua\" rel=\"nofollow noreferrer\">InferDeepMask.lua<\/a> module. getTopProps returns two things: the masks and the scores. The scores come from the Infer:getTopScores() function.<\/p>\n\n<p><strong>Question: How do I interpret the output of the getTopScores function?<\/strong><\/p>\n\n<p>The code comment:<\/p>\n\n<pre><\/pre>\n\n<p>Example output (converted into a numpy array):<\/p>\n\n<pre><\/pre>\n\n<p>The first column is clearly a score out of one. What is scaleNb, and what are the positions of M(:)? They are not the pixel locations on the mask.<\/p>\n\n<p>NB Each mask is of shape (336, 448)<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42032703","CreationDate":"2017-02-03T20:42:36.533","ParentId":null,"OwnerUserId":"1749629","Title":null,"Body":"<p>I believe (but have not proven) that the output is thus:<\/p>\n\n<ol>\n<li>the score value - a value between 0 and 1 describing how good a mask this is.<\/li>\n<li>the scaleNb - a value that indicates the rough size of the bounding box used to generate the mask.<\/li>\n<li>&amp; 4. position(of M(:)) - some way of relating the position of the bounding box to a smaller version of the original image.<\/li>\n<\/ol>\n"},{"AnswerId":"42035109","CreationDate":"2017-02-04T00:24:29.823","ParentId":null,"OwnerUserId":"893697","Title":null,"Body":"<p>By looking at the code:<\/p>\n\n<ul>\n<li><p>the score is fairly self-explanatory<\/p><\/li>\n<li><p>the scaleNb is the index of the adopted scale, i.e. the index <code>k<\/code> in the loop <code>for k = 1,nScales do<\/code> where the table containing the values of the different scales is created by the line <code>for scale = -3,2,.25 do table.insert(self.scales,scale)<\/code><\/p><\/li>\n<li><p>M(:) is the tricky part. If I understand the code correctly, <code>pos<\/code> is a tensor containing <code>nScales<\/code> ones, and the <code>k-th<\/code> element of the tensor gets incremented every time an object at scale <code>k<\/code> is selected (they are ordered by score). <code>temp<\/code> (which is the variable that is then used to compute the <code>x<\/code> and <code>y<\/code> values forming what they call <code>M(:)<\/code> is given by the line <code>local temp=sortedIds[pos[scale]][scale]<\/code>.\nWhat does that contain? It appears to be containing the position of the mask at the scale it has been detected. See the following piece of code:<\/p>\n\n<p><code>local sc=sc:view(h*w)\nlocal sS,sIds=torch.sort(sc,true)\nlocal sz = sS:size(1)\nsortedScores:narrow(2,s,1):narrow(1,1,sz):copy(sS)\nsortedIds:narrow(2,s,1):narrow(1,1,sz):copy(sIds)<\/code><\/p><\/li>\n<\/ul>\n\n<p><code>sortedIds<\/code> appears to be containing the <code>x<\/code>,<code>y<\/code> position of the score that is currently being evaluated.<\/p>\n"}]}
{"QuestionId":42010966,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><lstm>","CreationDate":"2017-02-02T19:41:42.317","AcceptedAnswerId":42011123.0,"OwnerUserId":223963.0,"Title":"LSTM Batches vs Timesteps","Body":"<p>I've followed the TensorFlow RNN tutorial to create a LSTM model. However, in the process, I've grown confused as to the difference, if any, between 'batches' and 'timesteps', and I'd appreciate help in clarifying this matter.<\/p>\n\n<p>The tutorial code (see following) essentially creates 'batches' based on a designated number of steps:<\/p>\n\n<pre><\/pre>\n\n<p>However, the following appears to do the same:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42011123","CreationDate":"2017-02-02T19:51:35.607","ParentId":null,"OwnerUserId":"2878004","Title":null,"Body":"<p>Assuming you are working with text, BATCH_SIZE would be the number of sentences that you are processing in parallel and num_steps would be the maximum number of words in any sentence. These are different dimensions of your input to the LSTM.<\/p>\n"}]}
{"QuestionId":42011070,"AnswerCount":23,"Tags":"<tensorflow>","CreationDate":"2017-02-02T19:48:04.830","AcceptedAnswerId":null,"OwnerUserId":7490399.0,"Title":"On Windows, running \"import tensorflow\" generates No module named \"_pywrap_tensorflow\" error","Body":"<p>On Windows, TensorFlow reports either or both of the following errors after executing an  statement:<\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<\/ul>\n","answers":[{"AnswerId":"43877887","CreationDate":"2017-05-09T18:59:57.393","ParentId":null,"OwnerUserId":"67878","Title":null,"Body":"<p>In my case the \"cudnn64_6.dll\" file in the \/bin folder had to be renamed to \"cudnn64_5.dll\" for the error to go away. I easily spent two hours to figure this out, and I followed the official install guide to the letter. This is true for installation via pip (officially supported) and conda (community supported).<\/p>\n"},{"AnswerId":"46762326","CreationDate":"2017-10-16T03:00:05.867","ParentId":null,"OwnerUserId":"3796113","Title":null,"Body":"<p>I will try to give the solution that worked for me. It seems that different set of problems can lead to this situation.<\/p>\n\n<p>32 bit software works in 64 bit OS. I installed anaconda-3 (32 bit) in my 64 bit OS. It was working perfectly fine. I decided to install tensorflow in my machine and it wouldn't install at first. I was using conda environment to install tensorflow and got this error.<\/p>\n\n<p>Solution is <strong>if you are running 64 bit OS, install 64 bit anaconda and if 32 bit OS then 32 bit anaconda<\/strong>. Then follow the standard procedure mentioned in tensorflow website for windows (anaconda installation). This made it possible to install tensorflow without any problem.<\/p>\n"},{"AnswerId":"57855072","CreationDate":"2019-09-09T13:30:43.383","ParentId":null,"OwnerUserId":"4788546","Title":null,"Body":"<p>Ran into the same problem (<strong>in <em>2019\/09\/09<\/em><\/strong>) while investigating <a href=\"https:\/\/stackoverflow.com\/questions\/57588589\/error-while-traning-using-the-estimator-api-in-tensorflow\">[SO]: Error while training using the estimator API in tensorflow<\/a>.<\/p>\n\n<p>Setup:<\/p>\n\n<ul>\n<li><em>Win 10 x64<\/em><\/li>\n<li><em>Python 3.7.3<\/em> (<em>x64<\/em>)<\/li>\n<li><em>TensorFlow-<strong>GPU<\/strong> 1.13.1<\/em> (<a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">[TensorFlow]: Install TensorFlow with pip<\/a>)<\/li>\n<\/ul>\n\n<p><strong>Error<\/strong>:<\/p>\n\n\n\n<blockquote>\n<pre class=\"lang-bat prettyprint-override\"><code>[cfati@CFATI-5510-0:e:\\Work\\Dev\\StackOverflow\\q057588589]&gt; \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\Scripts\\python.exe\" -c \"import tensorflow as tf\"\nTraceback (most recent call last):\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\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 \"&lt;string&gt;\", line 1, in &lt;module&gt;\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\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 \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\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 \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 58, in &lt;module&gt;\n    from tensorflow.python.pywrap_tensorflow_internal import *\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 28, in &lt;module&gt;\n    _pywrap_tensorflow_internal = swig_import_helper()\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow_internal.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow_internal', fp, pathname, description)\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\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<\/blockquote>\n\n<p>Looking at the \"faulty\" module (thanks to <a href=\"http:\/\/www.dependencywalker.com\" rel=\"nofollow noreferrer\">Dependency Walker<\/a>), it turns out that <strong>it's not itself that's missing, but some of its dependencies<\/strong> (the <em>cu*_<strong>100<\/strong>.dll<\/em> files).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vqAg8.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vqAg8.png\" alt=\"Img0\"><\/a><\/p>\n\n<p>Check <a href=\"https:\/\/stackoverflow.com\/questions\/57187566\/python-ctypes-multidimensional-arrays-loading-dll-throws-oserror-winerror-1\/57297745#57297745\">[SO]: Python Ctypes - Multidimensional arrays: loading dll throws OSError: [WinError 193] %1 is not a valid Win32 application (@CristiFati's answer)<\/a> (the <strong>Conslusions<\/strong> section at the end) for more details on this kind of errors. <br>I had an older <em>CUDA Toolkit<\/em> version (<strong><em>8<\/em><\/strong>), and as a consequence, the <em>cu*_<strong>80<\/strong>.dll<\/em> files.<\/p>\n\n<p>Upgrading to <em>TensorFlow-GPU 1.<strong>14<\/strong>.0<\/em> (<code>\"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\Scripts\\python.exe\" -m pip install --upgrade tensorflow-gpu<\/code>), made the error a bit clearer (and also shorter):<\/p>\n\n<blockquote>\n<pre class=\"lang-bat prettyprint-override\"><code>[cfati@CFATI-5510-0:e:\\Work\\Dev\\StackOverflow\\q057588589]&gt; \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\Scripts\\python.exe\" -c \"import tensorflow as tf\"\nTraceback (most recent call last):\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\platform\\self_check.py\", line 75, in preload_check\n    ctypes.WinDLL(build_info.cudart_dll_name)\n  File \"c:\\install\\x64\\python\\python\\03.07.03\\Lib\\ctypes\\__init__.py\", line 356, in __init__\n    self._handle = _dlopen(self._name, mode)\nOSError: [WinError 126] The specified module could not be found\n\nDuring handling of the above exception, another exception occurred:\n\nTraceback (most recent call last):\n  File \"&lt;string&gt;\", line 1, in &lt;module&gt;\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\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 \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\__init__.py\", line 49, in &lt;module&gt;\n    from tensorflow.python import pywrap_tensorflow\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\lib\\site-packages\\tensorflow\\python\\pywrap_tensorflow.py\", line 30, in &lt;module&gt;\n    self_check.preload_check()\n  File \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\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))\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<\/blockquote>\n\n<p><strong>Steps<\/strong>:<\/p>\n\n<ul>\n<li>Uninstall any <em>CUDA Toolkit<\/em> version (optional)<\/li>\n<li>Install <a href=\"https:\/\/developer.nvidia.com\/cuda-10.0-download-archive\" rel=\"nofollow noreferrer\">[nVidia.Developer]: CUDA Toolkit 10.0 Archive<\/a>\n\n<ul>\n<li><strong>Make sure to install <em>v10.0<\/em><\/strong> (that <strong>this <em>TensorFlow-GPU<\/em> version was built against<\/strong>). I installed <em>v10.<strong>1<\/em><\/strong> (which was the latest, and also the recommended version at the answer time), and the <em>.dll<\/em> names didn't match (<em>cu*_10<strong>1<\/strong>.dll<\/em>). Since I didn't want to install <em>v10.0<\/em>, I created some <em>symlink<\/em>s (with the \"correct\" names) to the existing files, and it worked. But <strong>bear in mind that this is unsupported!!! You may experience funny behavior (including crashes)<\/strong>. This is a (lame) workaround (<em>gainarie<\/em>)<\/li>\n<li>Additionally, a <strong>compatible<\/strong> (meaning that it's for a specific <em>CUDA Toolkit<\/em> version) <em>cuDNN<\/em> version (<a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-archive\" rel=\"nofollow noreferrer\">[nVidia.Developer]: cuDNN Archive<\/a>) is required. In order to access the download <em>URL<\/em>, <em>nVidia<\/em> membership is required<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>After the above steps, and also setting the correct paths, it worked:<\/p>\n\n<blockquote>\n<pre class=\"lang-bat prettyprint-override\"><code>[cfati@CFATI-5510-0:e:\\Work\\Dev\\StackOverflow\\q057588589]&gt; set PATH=%PATH%;%CUDA_PATH%\\bin;f:\\Install\\x64\\NVidia\\GPU Computing Toolkit\\cuDNN\\7.6\\bin\n\n[cfati@CFATI-5510-0:e:\\Work\\Dev\\StackOverflow\\q057588589]&gt; \"e:\\Work\\Dev\\VEnvs\\py_064_03.07.03_test0\\Scripts\\python.exe\" -c \"import tensorflow;print(\\\"Success!!!\\\")\"\nSuccess!!!\n<\/code><\/pre>\n<\/blockquote>\n"},{"AnswerId":"54115581","CreationDate":"2019-01-09T17:36:25.160","ParentId":null,"OwnerUserId":"6851744","Title":null,"Body":"<p>For the people finding this post in 2019, this error could also occur because the Python version 3.7 does not have support for TensorFlow (see <a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip<\/a>). So, check the Python version:<\/p>\n\n<pre><code>python --version\n<\/code><\/pre>\n\n<p>In case it is larger than 3.6, it should be downgraded to 3.6. For Anaconda:<\/p>\n\n<pre><code>conda install python=3.6\n<\/code><\/pre>\n\n<p>Then, install TensorFlow.<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>Btw, I did not have the GPU version, so there were no CUDA related issues in my case.<\/p>\n"},{"AnswerId":"43935739","CreationDate":"2017-05-12T10:40:12.713","ParentId":null,"OwnerUserId":"4233146","Title":null,"Body":"<p>The problem for me was the cuDNN library which didn't match the requirements of the graphics card. I downloaded the 6.0 version but for my GTX980ti but the recommended compute capability on the nvidia website was 5.1 ( <a href=\"http:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">http:\/\/developer.nvidia.com\/cuda-gpus<\/a> ) so I downloaded 5.1 and replaced the 6.0 version and as soon as I've done that it started working.<\/p>\n"},{"AnswerId":"43566341","CreationDate":"2017-04-23T01:06:23.467","ParentId":null,"OwnerUserId":"1656124","Title":null,"Body":"<p>The problem was the cuDNN Library for me - for whatever reason cudnn-8.0-windows10-x64-v6.0 was NOT working - I used cudnn-8.0-windows10-x64-v5.1 - ALL GOOD!<\/p>\n\n<p>My setup working with Win10 64 and the Nvidia GTX780M:<\/p>\n\n<ul>\n<li>Be sure you have the lib MSVCP140.DLL by checking your system\/path - if not get it <a href=\"https:\/\/www.microsoft.com\/en-us\/download\/details.aspx?id=48145\" rel=\"noreferrer\">here<\/a><\/li>\n<li>Run the windows installer for python 3.5.3-amd64 from <a href=\"https:\/\/www.python.org\/downloads\/release\/python-352\/\" rel=\"noreferrer\">here<\/a> - DO NOT try newer versions as they probably won't work <\/li>\n<li>Get the cuDNN v5.1 for CUDA 8.0 from <a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-download\" rel=\"noreferrer\">here<\/a> - put it under your users folder or in another known location (you will need this in your path)<\/li>\n<li>Get CUDA 8.0 x86_64 from <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"noreferrer\">here<\/a><\/li>\n<li>Set PATH vars as expected to point at the cuDNN libs and python (the python path should be added during the python install)<\/li>\n<li>Make sure that \".DLL\" is included in your PATHEXT variable<\/li>\n<li>If you are using tensorflow 1.3 then you want to use cudnn64_6.dll <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7705\" rel=\"noreferrer\">github.com\/tensorflow\/tensorflow\/issues\/7705<\/a><\/li>\n<\/ul>\n\n<p>If you run Windows 32 be sure to get the 32 bit versions of the files mentioned above.<\/p>\n"},{"AnswerId":"43690973","CreationDate":"2017-04-29T01:38:16.610","ParentId":null,"OwnerUserId":"5595995","Title":null,"Body":"<p>The problem was the cuDNN Library for me. I was able to run <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">the test code<\/a> after adding the directory (possibly bin folder) of the cuDNN DLL (not LIB file) in the Windows PATH.  <\/p>\n\n<p>For the reference, I installed TensorFlow from the source using PIP and my OS: Windows 7 and IDE: Visual Studio 2015.<\/p>\n"},{"AnswerId":"43632183","CreationDate":"2017-04-26T10:56:01.447","ParentId":null,"OwnerUserId":"4679530","Title":null,"Body":"<p>cuDNN causes my problem. PATH variable doesn't work for me. I have to copy the files in my cuDNN folders into respectful CUDA 8.0 folder structure. <\/p>\n"},{"AnswerId":"51606627","CreationDate":"2018-07-31T06:22:01.853","ParentId":null,"OwnerUserId":"10157759","Title":null,"Body":"<p>I posted a general approach for troubleshooting the \"DLL load failed\" problem in <a href=\"https:\/\/stackoverflow.com\/a\/51606222\/10157759\">this post<\/a> on Windows systems.<\/p>\n\n<ol>\n<li><p>Use the DLL dependency analyzer <a href=\"https:\/\/stackoverflow.com\/a\/51606222\/10157759\">Dependencies<\/a> to analyze <code>&lt;Your Python Dir&gt;\\Lib\\site-packages\\tensorflow\\python\\_pywrap_tensorflow_internal.pyd<\/code> and determine the exact missing DLL (indicated by a <code>?<\/code> beside the DLL). The path of the .pyd file is based on the TensorFlow 1.9 GPU version that I installed. I am  not sure if the name and path is the same in other TensorFlow versions.<\/p><\/li>\n<li><p>Look for information of the missing DLL and install the appropriate package to resolve the problem.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"47008832","CreationDate":"2017-10-30T05:52:48.733","ParentId":null,"OwnerUserId":"7212273","Title":null,"Body":"<p>tensorflow 1.3 does not support <strong>cuda 9.0<\/strong> yet.\nI degrade to <strong>cuda 8.0<\/strong>, then it works.<\/p>\n"},{"AnswerId":"46940714","CreationDate":"2017-10-25T19:33:13.340","ParentId":null,"OwnerUserId":"5990044","Title":null,"Body":"<p>my answer is for windows 10 users only  as I have tried the following on windows 10.\nExtending some of the answers above I suggest this : \nIf you are using anaconda then you can avoid everything and simply install <strong>anaconda-navigator<\/strong> using the command<\/p>\n\n<pre><code>conda install -c anaconda anaconda-navigator\n<\/code><\/pre>\n\n<p>Then you can launch the navigator from command prompt  using the command<\/p>\n\n<pre><code>anaconda-navigator\n<\/code><\/pre>\n\n<p>On running this command you get a simple gui where you can create an virtual environment, create the environment with python=3.5.2 and install module tensorflow-gpu or tensorflow by searching the module in the search box using gui, it will also take care of installing correct cuda files for you. Using anaconda navigator is the simplest solution.<\/p>\n\n<p>If you are not using anaconda then take care about the following<\/p>\n\n<p>tensorflow-gpu 1.3 requires python 3.5.2, cuda development kit 8.0 and cudaDNN 6.0, hence when installing make sure you run the command <\/p>\n\n<pre><code>pip install tensorflow-gpu==1.3\n<\/code><\/pre>\n\n<p>tensorflow-gpu 1.2.1 or less requires python 3.5.2, cuda development kit 8.0 and cudaDNN 5.1 hence when installing make sure you run the command <\/p>\n\n<pre><code>pip install tensorflow-gpu==1.2.1\n<\/code><\/pre>\n\n<p>Below are the steps you need to follow for both of the above processes\nSetting up you path variables\nYou must have the following system variables<\/p>\n\n<pre><code>CUDA_HOME = \"C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\"\nCUDA_PATH = \"C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\"\nCUDA_PATH_V8.0 = \"C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\"\n<\/code><\/pre>\n\n<p>You PATHTEXT must include \".DLL\" along with other extensions <\/p>\n\n<pre><code>\".COM;.EXE;.BAT;.CMD;.VBS;.VBE;.JS;.JSE;.WSF;.WSH;.MSC;.PY;.DLL\"\n<\/code><\/pre>\n\n<p>Also Add the following to you path<\/p>\n\n<pre><code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\lib\\x64\nC:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\extras\\CUPTI\\libx64;\nC:\\Windows\\SysWOW64;\nC:\\Windows\\System32        \n<\/code><\/pre>\n\n<p>If you are getting errors you can download the run the below code by mrry, this code will check your setup and tell you if something is wrong\n<a href=\"https:\/\/gist.github.com\/mrry\/ee5dbcfdd045fa48a27d56664411d41c\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/mrry\/ee5dbcfdd045fa48a27d56664411d41c<\/a><\/p>\n\n<p>References : \n<a href=\"http:\/\/blog.nitishmutha.com\/tensorflow\/2017\/01\/22\/TensorFlow-with-gpu-for-windows.html\" rel=\"nofollow noreferrer\">http:\/\/blog.nitishmutha.com\/tensorflow\/2017\/01\/22\/TensorFlow-with-gpu-for-windows.html<\/a><\/p>\n\n<p>The above reference is very useful.\nPlease comment for improvements to this answer.\nHope this helps, Thanks.<\/p>\n"},{"AnswerId":"42011114","CreationDate":"2017-02-02T19:50:30.567","ParentId":null,"OwnerUserId":"7490399","Title":null,"Body":"<p>Either error indicates that your system has not installed <code>MSVCP140.DLL<\/code>,\nwhich TensorFlow requires.<\/p>\n\n<p>To fix this error:<\/p>\n\n<ol>\n<li>Determine whether <code>MSVCP140.DLL<\/code> is in your <code>%PATH%<\/code> variable.<\/li>\n<li>If <code>MSVCP140.DLL<\/code> is not in your <code>%PATH%<\/code>, install the \n <a href=\"https:\/\/www.microsoft.com\/en-us\/download\/details.aspx?id=5555\" rel=\"noreferrer\">Visual C++ 2015 redistributable<\/a> (x64 version), which contains this DLL.<\/li>\n<\/ol>\n"},{"AnswerId":"45247795","CreationDate":"2017-07-21T22:48:44.223","ParentId":null,"OwnerUserId":"3127650","Title":null,"Body":"<p>In case you are trying to install tensorflow GPU in Windows, you can find <a href=\"https:\/\/nitishmutha.github.io\/tensorflow\/2017\/01\/22\/TensorFlow-with-gpu-for-windows.html\" rel=\"nofollow noreferrer\">this<\/a> easy interesting tutorial.<\/p>\n\n<p>Note: If you are using PyCharm for example, you have to change the interpreter to the created conda environment.<\/p>\n"},{"AnswerId":"42579688","CreationDate":"2017-03-03T13:06:34.973","ParentId":null,"OwnerUserId":"7653614","Title":null,"Body":"<p>Dll not found. Install Visual C++ 2015 redistributable to fix.<\/p>\n"},{"AnswerId":"42468780","CreationDate":"2017-02-26T12:54:07.063","ParentId":null,"OwnerUserId":"6017315","Title":null,"Body":"<p><code>TensorFlow<\/code> requires <code>MSVCP140.DLL<\/code>, which may not be installed on your system.\nTo solve it open the terminal en type or paste this link: <\/p>\n\n<pre><code>C:\\&gt; pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_amd64.whl \n<\/code><\/pre>\n\n<p><strong>Note this is to install the CPU-only version of TensorFlow.<\/strong> <\/p>\n"},{"AnswerId":"44466083","CreationDate":"2017-06-09T20:17:14.877","ParentId":null,"OwnerUserId":"4749857","Title":null,"Body":"<p>I have Win7 Pro 64-bit on AMD cpu, no gpu. I was following the instructions under \"Installing with native pip\" at <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a>. The installation step went ok but the attempt to import tensorflow produced the infamous:<\/p>\n\n<p>ImportError: No module named '_pywrap_tensorflow_internal'<\/p>\n\n<p>This seems to be one of those situations where a lot of unrelated things can go wrong, depending on configuration, which all cascade through to the same error.<\/p>\n\n<p>In my case, installing MSVCP140.DLL was the answer.<\/p>\n\n<p>You have MSVCP140.DLL already if (a) you have a file C:\\Windows\\System43\\MSVCP140.DLL, AND (b) if you have a 64 bit system, then you additionally have C:\\Windows\\SysWOW64\\MSVCP140.DLL. <\/p>\n\n<p>I installed it manually, which was unnecessary (the redistributable is not the whole Visual C++ development mess and isn't large). Use the link posted earlier in this thread to install it: <a href=\"https:\/\/www.microsoft.com\/en-us\/download\/details.aspx?id=5555http:\/\/\" rel=\"noreferrer\">Visual C++ 2015 redistributable<\/a>.<\/p>\n\n<p>Also, I recommend that you override the default install directory for Python and put it anywhere not under C:\\Program Files, because Windows tries to write-protect files there, which causes problems later.<\/p>\n"},{"AnswerId":"45784000","CreationDate":"2017-08-20T15:48:26.840","ParentId":null,"OwnerUserId":"1831452","Title":null,"Body":"<p>For each Tensorflow's version, it requires different version of CuDnn. On www.tensorflow.org, they did not mentioned about that in installation guide!<\/p>\n\n<p>My case use tensorflow version 1.3 which uses cuDNN 6. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/releases<\/a>.<\/p>\n\n<p>Please check your tensorfow version and cuDNN version if they are match together.<\/p>\n\n<p>And please set path environment for cuDNN, if it still does not work, please check the comment from @Chris Han.<\/p>\n"},{"AnswerId":"45541189","CreationDate":"2017-08-07T07:09:22.470","ParentId":null,"OwnerUserId":"605027","Title":null,"Body":"<p>After much trial and error, and making sure <em>VC++ 2015 Redistributable<\/em>, <em>cuDNN<\/em> DLL and all other dependencies are accessible from PATH, it looks like <em>Tensorflow GPU<\/em> works only with Python <code>3.5.2<\/code> (as of this writing)<\/p>\n\n<p>So if you're using <code>Anaconda<\/code><\/p>\n\n<ul>\n<li><code>conda create -n tensorflow-gpu python=3.5.2<\/code><\/li>\n<li><code>activate tensorflow-gpu<\/code><\/li>\n<li><code>pip install tensorflow-gpu<\/code><\/li>\n<\/ul>\n\n<p>Then open the python interpreter and verify<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n<\/code><\/pre>\n\n<blockquote>\n  <p>Found device 0 with properties:<br>\n  name: GeForce 940M<br>\n  major: 5 minor: 0<br>\n  memoryClockRate (GHz) 1.176<br>\n  pciBusID 0000:06:00.0<br>\n  Total memory: 2.00GiB<br>\n  Free memory: 1.66GiB  <\/p>\n<\/blockquote>\n\n<p>Credits: <a href=\"https:\/\/nitishmutha.github.io\/tensorflow\/2017\/01\/22\/TensorFlow-with-gpu-for-windows.html\" rel=\"nofollow noreferrer\">this neat guide<\/a><\/p>\n"},{"AnswerId":"44757622","CreationDate":"2017-06-26T10:24:27.367","ParentId":null,"OwnerUserId":"3558618","Title":null,"Body":"<p>One may be tempted to keep the Powershell\/cmd open on Windows. I've spent reasonable time till I decided to close and reopen my Powershell only to realize that I've done everything right. <\/p>\n"},{"AnswerId":"46189616","CreationDate":"2017-09-13T05:38:24.300","ParentId":null,"OwnerUserId":"2411048","Title":null,"Body":"<p>My two cents:<\/p>\n\n<p>I had a ton of problems trying to get my CUDA 8.0 installed properly on Windows 7. I had a previous version installed and I wanted to upgrade so I uninstalled it and tried to install CUDA 8.0 (for tensorflow 1.3). The installation failed every single time, I tried to downgrade to CUDA 7.5 and was able to install it but had a ton of problems with tensorflow (similar to the PATH problem described here). Long story short: what worked for me was:<\/p>\n\n<p>1) Uninstall EVERY NVIDIA component (except the display graphics driver)<\/p>\n\n<p>2) Download CUDA toolkit 8.0 (and the patch) <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cuda-downloads<\/a><\/p>\n\n<p>3) Check the CheckSum MD5 (I used MS <a href=\"https:\/\/www.microsoft.com\/en-ca\/download\/confirmation.aspx?id=11533\" rel=\"nofollow noreferrer\">https:\/\/www.microsoft.com\/en-ca\/download\/confirmation.aspx?id=11533<\/a> but any would do) to make sure they were OK (it happened several times that the installer was not dowloaded properly because my WiFi router apparently).<\/p>\n\n<p>4) Run the CUDA toolkit installer as root<\/p>\n\n<p>5) download the cudnn 8.0 v6 and add its location to the PATH variable <a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-download\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/rdp\/cudnn-download<\/a><\/p>\n\n<p>Hope that helps and saves some headaches... <\/p>\n\n<p>NOTE: This script helped me a lot to debug the problem! (Thank you mrry)\n<a href=\"https:\/\/gist.github.com\/mrry\/ee5dbcfdd045fa48a27d56664411d41c\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/mrry\/ee5dbcfdd045fa48a27d56664411d41c<\/a> <\/p>\n"},{"AnswerId":"46080926","CreationDate":"2017-09-06T17:12:03.313","ParentId":null,"OwnerUserId":"5204436","Title":null,"Body":"<p>With TensorFlow release 1.3.0, you need to use Cudnn 6.0 instead of Cudnn 5.0 as Cudnn 5.0 is giving this error. Don't forget to add path variable to Cudnn 6.0 .With cudnn64_6.dll your Tensorflow will work fine. Read the link below.\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md#release-130\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md#release-130<\/a><\/p>\n"},{"AnswerId":"49730981","CreationDate":"2018-04-09T10:37:37.033","ParentId":null,"OwnerUserId":"3131433","Title":null,"Body":"<p><strong>For tensorflow with CPU only:<\/strong><\/p>\n\n<hr>\n\n<p>I had installed tensorflow using command:<\/p>\n\n<pre><code>pip3 install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>This installed <code>tensorflow 1.7<\/code> \n<br>But could not import the tensorflow from withing <code>python 3.6.5 amd64<\/code> using:<\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<p>So, i downgraded the tensorflow version from <code>1.7<\/code> to <code>1.5<\/code> using following command:<\/p>\n\n<pre><code>pip3 install tensorflow==1.5\n<\/code><\/pre>\n\n<p>This uninstalled the previous version and installed <code>1.5<\/code>. Now it works.<\/p>\n\n<p>Seems that, my <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17807\" rel=\"noreferrer\">CPU does not support AVX instruction<\/a> set that is needed in <code>tensorflow 1.7<\/code><\/p>\n\n<p>I had <code>MSVCP140.DLL<\/code> in the system folders and .DLL in the PATHEXT variable in Environment Variable.<\/p>\n"},{"AnswerId":"49353124","CreationDate":"2018-03-18T21:43:07.987","ParentId":null,"OwnerUserId":"7009114","Title":null,"Body":"<p>For Those Running on Older Hardware:<\/p>\n\n<p>You may get this same error due to having an older CPU using tensorflow-gpu 1.6.<\/p>\n\n<p>If your cpu was made before 2011, then your max tensorflow-gpu version is 1.5.<\/p>\n\n<p>Tensorflow 1.6 requires AVX instructions on your cpu.  Verified here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md\" rel=\"noreferrer\">Tensorflow Github docs<\/a><\/p>\n\n<p>AVX enabled CPUs:  <a href=\"https:\/\/en.wikipedia.org\/wiki\/Advanced_Vector_Extensions#CPUs_with_AVX\" rel=\"noreferrer\">Wiki AVX CPUs<\/a><\/p>\n\n<p>What I did in my conda environment for tensorflow:<\/p>\n\n<pre><code>pip install --ignore-installed --upgrade tensorflow-gpu==1.5\n<\/code><\/pre>\n"}]}
{"QuestionId":42011419,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2017-02-02T20:09:30.757","AcceptedAnswerId":null,"OwnerUserId":7071746.0,"Title":"Is it possible to call tensorboard smooth function manually?","Body":"<p>I have two arrays X and Y. <\/p>\n\n<p>Is there a function I can call in tensorboard to do the smooth? <\/p>\n\n<p>Right now I can do an alternative way in python like:\n\nBut I am not sure what's way tensorboard do smooth. Is there a function I can call? <\/p>\n\n<p>Thanks. <\/p>\n","answers":[{"AnswerId":"56456661","CreationDate":"2019-06-05T08:19:44.707","ParentId":null,"OwnerUserId":"4792548","Title":null,"Body":"<p>So far I haven't found a way to call it manually, but you can construct a similar function,<\/p>\n\n<p>based on this <a href=\"https:\/\/stackoverflow.com\/questions\/42281844\/what-is-the-mathematics-behind-the-smoothing-parameter-in-tensorboards-scalar\">answer<\/a>, the function will be something like<\/p>\n\n<pre><code>def smooth(scalars, weight):  # Weight between 0 and 1\n    last = scalars[0]  # First value in the plot (first timestep)\n    smoothed = list()\n    for point in scalars:\n        smoothed_val = last * weight + (1 - weight) * point  # Calculate smoothed value\n        smoothed.append(smoothed_val)                        # Save it\n        last = smoothed_val                                  # Anchor the last smoothed value\n\n    return smoothed\n<\/code><\/pre>\n"}]}
{"QuestionId":42011471,"AnswerCount":0,"Tags":"<python><theano><joblib><theano-cuda>","CreationDate":"2017-02-02T20:12:26.670","AcceptedAnswerId":null,"OwnerUserId":1462738.0,"Title":"parallelizing with joblib.Parallel() on CPU fails after running a parallelized CUDA-based code on GPU","Body":"<p>I have a Python-based machine learning code that runs three algorithms on my data: Random Forest (implementation in scikit-learn), Gradient Boosting (implementation in XGBoost), and Recurrent Neural Network (implementation in Theano\/Keras). The first two are running on the CPU and are parallelized using  and the latter runs on the GPU using CUDA. I pass the name of the algorithm I want to use (, , ) in the variable  to a function, and it will proceed with running a parallel implementation of each algorithm.<\/p>\n\n<pre><\/pre>\n\n<p> is a function that performs nested cross-validation to select the best performing hyper-parameters and to estimate the performance of the selected model on new data, and  is the number of different hyper-parameter configurations to be sampled from a grid (I use, say, 30 different configurations).<\/p>\n\n<p>If the run order is (1) , (2) , (3) , everything runs fine; the first two are parallelized on a 4-core CPU and the latter runs on a Tesla K80 GPU; at the end, I get three predictions which I can later merge. However, after running the Theano-based, CUDA-parallelized neural network, if I rerun either  or , I will get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>It seems to me that once the CUDA-based code (recurrent neural network) is executed, any call to  will get rerouted to the GPU instead of the CPU, since the error I get is  even though I had called a scikit-learn Random Forest classifier which should run on the CPU. In fact, if my first run is ordered like (1) , (2) , (3) , I get the same error after finishing the first pass.<\/p>\n\n<p>I should mention that I am running these in an IPython session, so I am typing the commands by hand in the terminal (although I am not sure if that matters or not).<\/p>\n\n<p>Any ideas on why this happens and how I can route any execution of  or  to the CPU regardless of whether a  method was previously executed on the GPU or not? I'd appreciate your help.<\/p>\n","answers":[]}
{"QuestionId":42011491,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-02-02T20:13:38.393","AcceptedAnswerId":null,"OwnerUserId":4738644.0,"Title":"Tensorflow: Using a value of a placeholder into another placeholder","Body":"<p>I am using different sets of datasets in the same neural network, each one with a different number of outputs (9,12,15,21,65 each one) In order to do this, I tried to use an integer placeholder and use this parameter as parameter for another placeholder (in order to maintain modifiable size of the outputs) like this:<\/p>\n\n<pre><\/pre>\n\n<p>Once I do this, I receive a long exception telling me that I cannot put a tensor value where a should put a number or a string. Ok, To try to solve this then I made a condition based on the placeholder which I use<\/p>\n\n<pre><\/pre>\n\n<p>But then it creates a 65 placeholder and when I try to add the dataset of 9 possible labels, I receive an exception telling me I'm trying to insert an array of size 9 into a tensor of size 65<\/p>\n\n<p>I kept making tricks into this and then I decided to play the whole thing so I try to use the conditions as depending on the numb_labels as I could like this<\/p>\n\n<pre><\/pre>\n\n<p>... and also (this is an example of 5 cases which are same as this one)<\/p>\n\n<pre><\/pre>\n\n<p>However when I'm going to train, I receive an exception telling me that the train_step is not defined... my tricks are finished. Can someone please tell me how can I put a placeholder value in another placeholder without causing so much trouble?? Is that possible or should I keep each architecture separated as till now. <\/p>\n\n<p>I also thought of using<\/p>\n\n<pre><\/pre>\n\n<p>But with my output in the network I do need the fixed number of labels (or at least what TF says)<\/p>\n\n<p>Is there a way of using a placeholder value as input for another placeholder or should I just keep 5 different scripts for 5 different datasets?<\/p>\n\n<p>Thanks a lot.<\/p>\n","answers":[]}
{"QuestionId":42012272,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2017-02-02T21:05:09.433","AcceptedAnswerId":null,"OwnerUserId":1214630.0,"Title":"Changing recurrent connections and equations of RNN\/LSTM in Tensorflow","Body":"<p>According to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/rnn_cell\/rnn_cells_for_use_with_tensorflow_s_core_rnn_methods#BasicLSTMCell\" rel=\"nofollow noreferrer\">tensorflow documantaion<\/a>, the RNN in Tensorflow supports the following equation:<\/p>\n\n<pre><\/pre>\n\n<p>Now, what would be the best practice to implemnet RNN as follow:<\/p>\n\n<pre><\/pre>\n\n<p>By the best practice, I mean the fast and efficient way to implement RNN\/LSTM with slightly different equations: edit tf.nn.rnn_cell such that it can support this or? [For this example, assume we can't concat input_1 and input_2 since they have different dimensions] <\/p>\n\n<p>In general, if we want to change a bit the equations of LSTM\/RNN (i.e. add new inputs, new recurrent connections, and etc), where we should start in TF? any suggestions? Which files should be changed? A pointer to any tutorial or code will be appreciated.<\/p>\n","answers":[{"AnswerId":"42070933","CreationDate":"2017-02-06T15:05:02.473","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>This should be easy to do using Tensorflow's RNN cell library. Each RNN cell type is implemented a subclass of RNNCell; you can easily add your own new subclass. Take a look at the code that defines some of the common RNN cells:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py<\/a><\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":42012337,"AnswerCount":3,"Tags":"<tensorflow><metal><metal-performance-shaders>","CreationDate":"2017-02-02T21:09:14.617","AcceptedAnswerId":45702172.0,"OwnerUserId":3815547.0,"Title":"Deconvolution with Metal Performance Shaders","Body":"<p>Turns out there is no such operation as  in . The closest analogue in  is . <\/p>\n\n<p>Is it possible to sort of plug-in custom operations between  default operations?<\/p>\n","answers":[{"AnswerId":"42353988","CreationDate":"2017-02-20T20:51:24.023","ParentId":null,"OwnerUserId":"7501629","Title":null,"Body":"<p>You can write your own Metal compute kernels and execute those in between the MPS operations.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code>let commandBuffer = commandQueue.makeCommandBuffer()\n\n. . .\n\n\/\/ Do something with an MPSCNN layer:\nlayer1.encode(commandBuffer: commandBuffer, sourceImage: img1, destinationImage: img2)\n\n\/\/ Perform your own compute kernel:\nlet encoder = commandBuffer.makeComputeCommandEncoder()\nencoder.setComputePipelineState(yourOwnComputePipeline)\nencoder.setTexture(img2.texture, at: 0)\nencoder.setTexture(img3.texture, at: 1)\nlet threadGroupSize = MTLSizeMake(. . .)\nlet threadGroups = MTLSizeMake(img2.texture.width \/ threadGroupSize.width,\n                               img2.texture.height \/ threadGroupSize.height, 1)\nencoder.dispatchThreadgroups(threadGroups, threadsPerThreadgroup: threadGroupSize)\nencoder.endEncoding()\n\n\/\/ Do something with another MPSCNN layer:\nlayer2.encode(commandBuffer: commandBuffer, sourceImage: img3, destinationImage: img4)\n\n. . .\n\ncommandBuffer.commit()\n<\/code><\/pre>\n\n<p>You have to write your own compute kernel in the Metal Shading Language and load this into the <code>yourOwnComputePipeline<\/code> object. Then you can encode it into the current command buffer whenever you want.<\/p>\n"},{"AnswerId":"45702172","CreationDate":"2017-08-15T21:57:32.307","ParentId":null,"OwnerUserId":"3166255","Title":null,"Body":"<p>MPS now provides MPSCNNConvolutionTranspose in macOS X.13 and tvOS\/iOS 11. <\/p>\n"},{"AnswerId":"42984784","CreationDate":"2017-03-23T19:00:58.143","ParentId":null,"OwnerUserId":"7501629","Title":null,"Body":"<p>[I am adding this as a new answer because it's a different solution.]<\/p>\n\n<p>Note that deconvolution in deep learning is also known as \"transposed convolution\", which means that it's the same as doing a regular convolution but with the kernels horizontally and vertically flipped.<\/p>\n\n<p>So you should be able to use a regular <code>MPSCNNConvolution<\/code> layer that takes the <code>MPSImage<\/code> that you wish to deconvolve as input, and that uses the same kernels as the \"forward\" convolution step but flipped horizontally and vertically.<\/p>\n\n<p>The advantage of doing this over writing your own compute kernel is that you can use the very fast kernels from MPS.<\/p>\n\n<p><strong>Edit:<\/strong> An example. Let's say your conv kernel weights look like this:<\/p>\n\n<pre><code>1, 2, 3\n4, 5, 6\n7, 8, 9\n<\/code><\/pre>\n\n<p>Then after flipping the kernel, the weights look like this:<\/p>\n\n<pre><code>9, 8, 7\n6, 5, 4\n3, 2, 1\n<\/code><\/pre>\n\n<p>In other words, you need to make a copy of your weights array and reverse it. In memory the first weights look like this:<\/p>\n\n<pre><code>1, 2, 3, 4, 5, 6, 7, 8, 9\n<\/code><\/pre>\n\n<p>The flipped kernel looks like this in memory, so it's simply the original kernel but in reverse order:<\/p>\n\n<pre><code>9, 8, 7, 6, 5, 4, 3, 2, 1\n<\/code><\/pre>\n\n<p>Then you make a new convolution layer using that reversed array. This is now your deconv layer.<\/p>\n\n<p>I don't have Metal sample code to show you, but it's really no different from making a regular <code>MPSCNNConvolution<\/code> layer. You just have to reverse the weights for the layer.<\/p>\n"}]}
{"QuestionId":42012715,"AnswerCount":1,"Tags":"<deep-learning><caffe><loss>","CreationDate":"2017-02-02T21:36:04.953","AcceptedAnswerId":null,"OwnerUserId":7508264.0,"Title":"Caffe loss layer - backward function","Body":"<p>I have a question about backward function in Caffe's loss layer. \nI have seen implementation of a Euclidean loss layer at : \n<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/pycaffe\/layers\/pyloss.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/pycaffe\/layers\/pyloss.py<\/a><\/p>\n\n<p>At the end, the <strong>backward<\/strong> function is defined as follow : <\/p>\n\n<pre><\/pre>\n\n<p>I know that <strong>bottom[0]<\/strong> indicates the predicted value and <strong>bottom[1]<\/strong> is target values (ground truth).<br>\nCan you help me understand why <strong>sign<\/strong> is +1 for prediction and -1 for targets? \nI thought that we didn't need to assign any values to <strong>bottom[1].diff<\/strong> for targets. <\/p>\n\n<p>If I have multi-label problem, how can I update the code ? <\/p>\n","answers":[{"AnswerId":"42013330","CreationDate":"2017-02-02T22:15:59.843","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>When <code>bottom[1]<\/code> is ground truth you should expect <code>propagate_down[1]<\/code> to be <code>False<\/code> thus no gradient is propagated to the labels.<br>\nHowever, if you use this loss layer to compare two <strong>trainable<\/strong> blobs you would like to propagate gradients to both <code>bottom<\/code>s.<\/p>\n\n<p>The sign is only due to the derivative of this particular loss, write the derivative and you'll see where it comes from.<\/p>\n\n<p>The one thing I am missing here is <code>top[0].diff<\/code>. If I understand correctly it should hold the <code>loss_weight<\/code> defined for this layer...<\/p>\n"}]}
{"QuestionId":42012906,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-02T21:48:35.350","AcceptedAnswerId":48876774.0,"OwnerUserId":4880003.0,"Title":"Create a custom Tensorflow histogram summary","Body":"<p>There are a couple of SO answers on creating a custom scalar summary in TF (<a href=\"https:\/\/stackoverflow.com\/questions\/37902705\/how-to-manually-create-a-tf-summary\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/37530228\/how-do-i-add-an-arbitrary-value-to-a-tensorflow-summary\">here<\/a>), but I can't find anything on creating a custom <em>histogram<\/em> summary. The documentation seems to be very lacking for custom summaries. I have a numpy array of that I'd like to make a summary of - any ideas on how?<\/p>\n\n<p>(tf.Summary.Value has a histo field that I tried using, but it requires a tensorflow::HistogramProto; there's no documentation on that class either, so I'm at a loss on how to make it. I've tried creating a minimal failing example below).<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"48876774","CreationDate":"2018-02-20T01:54:54.840","ParentId":null,"OwnerUserId":"7242633","Title":null,"Body":"<p>This piece of code works:<\/p>\n\n<pre><code>import tensorflow as tf\n\nimport numpy as np\n\ndef log_histogram(writer, tag, values, step, bins=1000):\n    # Convert to a numpy array\n    values = np.array(values)\n\n    # Create histogram using numpy\n    counts, bin_edges = np.histogram(values, bins=bins)\n\n    # Fill fields of histogram proto\n    hist = tf.HistogramProto()\n    hist.min = float(np.min(values))\n    hist.max = float(np.max(values))\n    hist.num = int(np.prod(values.shape))\n    hist.sum = float(np.sum(values))\n    hist.sum_squares = float(np.sum(values**2))\n\n    # Requires equal number as bins, where the first goes from -DBL_MAX to bin_edges[1]\n    # See https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/summary.proto#L30\n    # Thus, we drop the start of the first bin\n    bin_edges = bin_edges[1:]\n\n    # Add bin edges and counts\n    for edge in bin_edges:\n        hist.bucket_limit.append(edge)\n    for c in counts:\n        hist.bucket.append(c)\n\n    # Create and write Summary\n    summary = tf.Summary(value=[tf.Summary.Value(tag=tag, histo=hist)])\n    writer.add_summary(summary, step)\n    writer.flush()\n\nsess = tf.Session()\nplaceholder = tf.placeholder(tf.float32)\n\ntf.summary.histogram('N(0,1)', placeholder)\nsummaries = tf.summary.merge_all()\nwriter = tf.summary.FileWriter('.\/summaries')\n\nmu, sigma = 0, 0.1 # mean and standard deviation\ns = np.random.normal(mu, 1, 10000)\nlog_histogram(writer, 'N(0,1)', s, 1, bins=100)\n<\/code><\/pre>\n\n<p>The <a href=\"https:\/\/gist.github.com\/gyglim\/1f8dfb1b5c82627ae3efcfbbadb9f514\" rel=\"noreferrer\">source<\/a>.<\/p>\n"}]}
{"QuestionId":42013138,"AnswerCount":1,"Tags":"<tensorflow><neural-network><keras><conv-neural-network><keras-layer>","CreationDate":"2017-02-02T22:02:05.640","AcceptedAnswerId":null,"OwnerUserId":5051564.0,"Title":"ValueError: Tensor Tensor(...) is not an element of this graph. When using global variable keras model","Body":"<p>I'm running a web server using flask and the error comes up when I try to use vgg16, which is the global variable for keras' pre-trained VGG16 model. I have no idea why this error rises or whether it has anything to do with the Tensorflow backend.\nHere is my code:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the full error:\n<a href=\"https:\/\/i.stack.imgur.com\/jLzba.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jLzba.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Any comment or suggestion is greatly appreciated. Thank you.<\/p>\n","answers":[{"AnswerId":"50917853","CreationDate":"2018-06-18T21:40:32.450","ParentId":null,"OwnerUserId":"1193738","Title":null,"Body":"<p>Take a look at <code>avital<\/code>'s answer on <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/2397\" rel=\"nofollow noreferrer\">this github issue<\/a>. Quoting the relevant part here:<\/p>\n\n<blockquote>\n  <p>Right after loading or constructing your model, save the TensorFlow graph:<\/p>\n\n<pre><code>graph = tf.get_default_graph()\n<\/code><\/pre>\n  \n  <p>In the other thread (or perhaps in an asynchronous event handler), do:<\/p>\n\n<pre><code>global graph\nwith graph.as_default():\n    (... do inference here ...)\n<\/code><\/pre>\n<\/blockquote>\n\n<p>I modified this a bit and stored the graph in my app's config object instead of making it a global. <\/p>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_default_graph\" rel=\"nofollow noreferrer\">TensorFlow documentation<\/a> for <code>get_default_graph<\/code> explains why this is necessary:<\/p>\n\n<blockquote>\n  <p>NOTE: The default graph is a property of the current thread. If you create a new thread, and wish to use the default graph in that thread, you must explicitly add a with g.as_default(): in that thread's function.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42013316,"AnswerCount":12,"Tags":"<tensorflow>","CreationDate":"2017-02-02T22:14:54.007","AcceptedAnswerId":null,"OwnerUserId":7490399.0,"Title":"After building TensorFlow from source, seeing libcudart.so and libcudnn errors","Body":"<p>I'm building TensorFlow from source code.  The build appears to succeed; however, when my TensorFlow program invokes , one or both of the following errors appear:<\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<\/ul>\n","answers":[{"AnswerId":"45631282","CreationDate":"2017-08-11T09:12:46.253","ParentId":null,"OwnerUserId":"7045929","Title":null,"Body":"<p>Mysteriously, my <code>libcudnn.so.5<\/code> was installed at <code>~\/cuda\/lib64<\/code>. For people like me, you need to change<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH:~\/cuda\/lib64\"\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH:\/home\/yourusername\/cuda\/lib64\"\n<\/code><\/pre>\n"},{"AnswerId":"42013350","CreationDate":"2017-02-02T22:18:03.447","ParentId":null,"OwnerUserId":"7490399","Title":null,"Body":"<p>The preceding errors are typically caused by not specifying a version number for the Cuda SDK or cuDNN when you run the <code>configure<\/code> script. In other words, when running the <code>configure<\/code> script, always specify a version number in response to the following two questions:<\/p>\n\n<ul>\n<li><code>Please specify the Cuda SDK version you want to use, e.g. 7.0.<\/code><\/li>\n<li><code>Please specify the cuDNN version you want to use.<\/code><\/li>\n<\/ul>\n\n<p>Don't accept the system defaults.<\/p>\n"},{"AnswerId":"43143994","CreationDate":"2017-03-31T15:07:03.480","ParentId":null,"OwnerUserId":"7798212","Title":null,"Body":"<p><strong>First<\/strong>, for the following error:<\/p>\n\n<blockquote>\n  <p>ImportError: libcudart.so.8.0: cannot open shared object file: No such file or directory<\/p>\n<\/blockquote>\n\n<p>make sure your <code>LD_LIBRARY_PATH<\/code> includes your <code>lib64<\/code> directory in whichever path you installed your cuda package in. You can do this by adding an <code>export<\/code> line in your <code>.bashrc<\/code>. For Omar, it looked like the following:<\/p>\n\n<blockquote>\n  <p>I fixed this just adding the cuda path to my .bashrc<\/p>\n  \n  <blockquote>\n    <p><code>export LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64\/<\/code><\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<hr>\n\n<p><strong>For me, I had to do Omar's line and also:<\/strong>\n <code>export LD_LIBRARY_PATH=\/usr\/local\/cuda-8.0\/lib64\/<\/code>\nbecause I have two directories involving cuda (probably not the best).<\/p>\n\n<hr>\n\n<p><strong>Second<\/strong>, are you sure you installed cuDNN? Note that this is different from the regular cuda package. You will need to register, then download and install the package from the following page:\n<a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"noreferrer\">https:\/\/developer.nvidia.com\/cudnn<\/a><\/p>\n\n<hr>\n\n<p><strong>Third<\/strong>, I had this same problem:<\/p>\n\n<blockquote>\n  <p>ImportError: libcudnn.5: cannot open shared object file: No such file or directory<\/p>\n<\/blockquote>\n\n<p>It turns out there is no <code>libcudnn.5<\/code> in my <code>\/usr\/local\/cuda\/lib64<\/code> or <code>\/usr\/local\/cuda-8.0\/lib64<\/code> directories. However, I do have a <code>libcudnn.so.6.*<\/code> file. To solve the problem, I created a soft link:<\/p>\n\n<pre><code>ln -s libcudnn.so.6.* libcudnn.so.5\n<\/code><\/pre>\n\n<p>in my <code>\/usr\/local\/cuda\/lib64<\/code> directory. Now everything works for me. Your directory might be different if you already had <code>cuDNN<\/code>, and your <code>libcudnn.so.6.*<\/code> might be a different version, so check that.<\/p>\n"},{"AnswerId":"43387484","CreationDate":"2017-04-13T08:44:23.013","ParentId":null,"OwnerUserId":"4561409","Title":null,"Body":"<p>On MacOS, this issue is often caused by bazel running in a sandbox environment, thus not respecting the LD_LIBRARY_PATH set in your local shell. I wouldn't bother going into the merit of deep integration of sandboxing in a build tool.<\/p>\n\n<p>The simple workaround is to symlink the libraries into \/usr\/local\/lib.<\/p>\n\n<p><code>cd \/usr\/local\/lib &amp;&amp; ln -s ..\/cuda\/lib\/libcudart.8.0.dylib<\/code><\/p>\n"},{"AnswerId":"51496265","CreationDate":"2018-07-24T10:26:31.243","ParentId":null,"OwnerUserId":"2764243","Title":null,"Body":"<p>Common workaround general problem related to GPU, CUDA, and Docker:<\/p>\n\n<p><strong>A.<\/strong> If you are dealing with machine learning\/ deep learning related deployment, use nvidia-docker and not native-docker. To install nvidia-docker follow these simple steps.<\/p>\n\n<blockquote>\n  <p>Docker containers are platform-agnostic, but also hardware-agnostic. This presents a problem when using specialized hardware such as NVIDIA GPUs which require kernel modules and user-level libraries to operate.  As a result, Docker does not natively support NVIDIA GPUs within containers.<\/p>\n<\/blockquote>\n\n<p><strong>B.<\/strong> If you want to access GPU with docker, never build a container from scratch, you will be fried with a number of errors.  Instead just use any container from Nvidia-Docker hub. Choose any specific image, copy its Dockerfile and run  <code>sudo nvidia-docker build -t happyapp .<\/code>.  [happyapp is your new app name]. In 5 min you will get your container ready (depends on network speed :p). <\/p>\n\n<p><strong>C.<\/strong> Never download a nvidia-docker having any cuda\/cudnn version If you want to install and run Tensorflow on it. If you do so you will get <strong>libcudnn.so.6<\/strong> or <strong>libcudnn.so.9<\/strong> or <strong>libcusolver.so.8.0<\/strong> related errors and you will hardly get around this errors. \nInstead just use pre-built Tensorflow Docker image: <code>sudo nvidia-docker run -it tensorflow\/tensorflow:latest-gpu \/bin\/bash<\/code><\/p>\n"},{"AnswerId":"44147506","CreationDate":"2017-05-24T01:28:26.703","ParentId":null,"OwnerUserId":"5046896","Title":null,"Body":"<p>I came across the same issue<\/p>\n\n<pre><code>In [1]: import tensorflow\n---------------------------------------------------------------------------\nImportError                               Traceback (most recent call last)\n\/usr\/local\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py in &lt;module&gt;()\n     40     sys.setdlopenflags(_default_dlopen_flags | ctypes.RTLD_GLOBAL)\n---&gt; 41   from tensorflow.python.pywrap_tensorflow_internal import *\n     42   from tensorflow.python.pywrap_tensorflow_internal import __version__\n\n\/usr\/local\/lib\/python3.5\/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\/usr\/local\/lib\/python3.5\/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\/usr\/local\/lib\/python3.5\/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\/usr\/local\/lib\/python3.5\/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: libcudnn.so.5: cannot open shared object file: No such file or directory\n\nDuring handling of the above exception, another exception occurred:\n\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-1-a649b509054f&gt; in &lt;module&gt;()\n----&gt; 1 import tensorflow\n\n\/usr\/local\/lib\/python3.5\/site-packages\/tensorflow\/__init__.py in &lt;module&gt;()\n     22\n     23 # pylint: disable=wildcard-import\n---&gt; 24 from tensorflow.python import *\n     25 # pylint: enable=wildcard-import\n     26\n\n\/usr\/local\/lib\/python3.5\/site-packages\/tensorflow\/python\/__init__.py in &lt;module&gt;()\n     49 import numpy as np\n     50\n---&gt; 51 from tensorflow.python import pywrap_tensorflow\n     52\n     53 # Protocol buffers\n\n\/usr\/local\/lib\/python3.5\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py in &lt;module&gt;()\n     50 for some common reasons and solutions.  Include the entire stack trace\n     51 above this error message when asking for help.\"\"\" % traceback.format_exc()\n---&gt; 52   raise ImportError(msg)\n     53\n     54 # pylint: enable=wildcard-import,g-import-not-at-top,unused-import,line-too-long\n\nImportError: Traceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.5\/site-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\/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 \"\/usr\/local\/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 \"\/usr\/local\/lib\/python3.5\/imp.py\", line 242, in load_module\n    return load_dynamic(name, filename, file)\n  File \"\/usr\/local\/lib\/python3.5\/imp.py\", line 342, in load_dynamic\n    return _load(spec)\nImportError: libcudnn.so.5: 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><strong>I have installed cudnn 6.0 while it needs <code>libcudnn.so.5<\/code>, apparently it couldn't find <code>libcudnn.so.5<\/code>. It seems that your tensorflow needs cudnn 5.x, so install cudnn 5.x<\/strong><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pJOfP.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pJOfP.jpg\" alt=\"CUDNN INSTALLATION\"><\/a><\/p>\n\n<p><strong>Make sure<\/strong> you have already installed cuda 8.0 and exported the <code>PATH<\/code> and <code>LD_LIBRARY_PATH<\/code><\/p>\n\n<p>To install cudnn 5.x, try the following commands<\/p>\n\n<h3>Extract tgz files<\/h3>\n\n<pre><code>$ tar -zxvf cudnn-8.0-linux-x64-v5.1.tgz\n<\/code><\/pre>\n\n<h3>Check the files<\/h3>\n\n<pre><code>$ cd cuda\/lib64\/\n$ ls -l\ntotal 150908\nlrwxrwxrwx 1 doom doom       13 Nov  7  2016 libcudnn.so -&gt; libcudnn.so.5\nlrwxrwxrwx 1 doom doom       18 Nov  7  2016 libcudnn.so.5 -&gt; libcudnn.so.5.1.10\n-rwxr-xr-x 1 doom doom 84163560 Nov  7  2016 libcudnn.so.5.1.10\n-rw-r--r-- 1 doom doom 70364814 Nov  7  2016 libcudnn_static.a\n<\/code><\/pre>\n\n<p>Here you will see 2 symbolic link files, and just copy <code>libcudnn.so.5.1.10<\/code> and <code>libcudnn_static.a<\/code> to <code>\/usr\/local\/cuda\/lib64<\/code><\/p>\n\n<h3>Make symbolic link files<\/h3>\n\n<pre><code>$ cd \/usr\/local\/cuda\/lib64\/\n$ sudo ln -s libcudnn.so.5.1.10 libcudnn.so.5\n$ sudo ln -s libcudnn.so.5 libcudnn.so\n$ ls -l libcudnn*\nlrwxrwxrwx 1 root root       13 May 24 09:24 libcudnn.so -&gt; libcudnn.so.5\nlrwxrwxrwx 1 root root       18 May 24 09:24 libcudnn.so.5 -&gt; libcudnn.so.5.1.10\n-rwxr-xr-x 1 root root 84163560 May 24 09:23 libcudnn.so.5.1.10\n-rw-r--r-- 1 root root 70364814 May 24 09:23 libcudnn_static.a\n<\/code><\/pre>\n\n<h3>Copy <code>cudnn.h<\/code> in <code>include<\/code> directory  to <code>\/usr\/local\/cuda\/include<\/code><\/h3>\n\n<pre><code>$ sudo cp cudnn.h \/usr\/local\/cuda\/include\/\n<\/code><\/pre>\n\n<p>Hope it will help you!<\/p>\n"},{"AnswerId":"49455109","CreationDate":"2018-03-23T17:10:55.120","ParentId":null,"OwnerUserId":"4568212","Title":null,"Body":"<p>As of now tensor-flow supports cuda-9.0<\/p>\n\n<p>Do following things. Hope it helps :<\/p>\n\n<pre><code>$ sudo apt-get install cuda-9.0\n$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:\/usr\/local\/cuda-9.0\/lib64\n<\/code><\/pre>\n\n<p>Download cuDNN for 9.0 (You need to register before downloading)\n<a href=\"https:\/\/developer.nvidia.com\/rdp\/form\/cudnn-download-survey\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/rdp\/form\/cudnn-download-survey<\/a><\/p>\n\n<pre><code>$ sudo dpkg -i libcudnn7_7.1.2.21-1+cuda9.0_amd64.deb\n<\/code><\/pre>\n\n<p>Close all terminal and open new<\/p>\n\n<pre><code>$ source activate tensorflow\n$ python\n&gt;&gt; import tensorflow as tf\n<\/code><\/pre>\n\n<p>You should not get any error after this. <\/p>\n"},{"AnswerId":"45787225","CreationDate":"2017-08-20T22:11:16.020","ParentId":null,"OwnerUserId":"534708","Title":null,"Body":"<p>I have seen a similar error (bottom of this post), but complaining about <strong>libcudnn.so.6<\/strong> instead of libcudart.so.8.0 (see a note below).<\/p>\n\n<p><strong>Solution:<\/strong><\/p>\n\n<ol>\n<li>Download '<strong>cuDNN v6.0 Library for Linux<\/strong>':\n\n<ul>\n<li>Go to <a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-download\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/rdp\/cudnn-download<\/a><\/li>\n<li>Click on '<strong>Download cuDNN v6.0 (April 27, 2017), for CUDA 8.0<\/strong>'<\/li>\n<li>Then click on '<strong>cuDNN v6.0 Library for Linux<\/strong>'. A file (named '<strong>cudnn-8.0-linux-x64-v6.0.tgz<\/strong>') will be downloaded.<\/li>\n<\/ul><\/li>\n<li>Follow the instructions of <strong>Alexander Yau<\/strong> <a href=\"https:\/\/stackoverflow.com\/questions\/42013316\/after-building-tensorflow-from-source-seeing-libcudart-so-and-libcudnn-errors\/44147506#44147506\"><strong>above<\/strong><\/a> to install the cuDNN v6.0 library.<\/li>\n<\/ol>\n\n<p><br>\n<strong>Note:<\/strong> <\/p>\n\n<p>the Tensorflow installation instructions (as of 20\/Aug\/2017) <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#nvidia_requirements_to_run_tensorflow_with_gpu_support\" rel=\"nofollow noreferrer\">require installing cuDNN v5.1<\/a>, but my Tensorflow installation (following the instructions for <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#InstallingVirtualenv\" rel=\"nofollow noreferrer\">installing in a virtualenv<\/a>) required cuDNN v6.x (as indicated by the error). I don't know if it is a mistake on my side or a Tensorflow documentation one. Nevertheless, above solution worked for me.<\/p>\n\n<p><br>\n<strong>Encountered error:<\/strong><\/p>\n\n<pre class=\"lang-html prettyprint-override\"><code>In [1]: import tensorflow as tf\n---------------------------------------------------------------------------\nImportError                               Traceback (most recent call last)\n&lt;ipython-input-1-41389fad42b5&gt; in &lt;module&gt;()\n----&gt; 1 import tensorflow as tf\n\n\/home\/haseeb\/.virtualenvs\/attention_transformer\/local\/lib\/python2.7\/site-packages\/tensorflow\/__init__.py in &lt;module&gt;()\n     22 \n     23 # pylint: disable=wildcard-import\n---&gt; 24 from tensorflow.python import *\n     25 # pylint: enable=wildcard-import\n     26 \n\n\/home\/haseeb\/.virtualenvs\/attention_transformer\/local\/lib\/python2.7\/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\/home\/haseeb\/.virtualenvs\/attention_transformer\/local\/lib\/python2.7\/site-packages\/tensorflow\/python\/pywrap_tensorflow.py in &lt;module&gt;()\n     50 for some common reasons and solutions.  Include the entire stack trace\n     51 above this error message when asking for help.\"\"\" % traceback.format_exc()\n---&gt; 52   raise ImportError(msg)\n     53 \n     54 # pylint: enable=wildcard-import,g-import-not-at-top,unused-import,line-too-long\n\nImportError: Traceback (most recent call last):\n  File \"\/home\/haseeb\/.virtualenvs\/attention_transformer\/local\/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 \"\/home\/haseeb\/.virtualenvs\/attention_transformer\/local\/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 \"\/home\/haseeb\/.virtualenvs\/attention_transformer\/local\/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: libcudnn.so.6: 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\n"},{"AnswerId":"45794113","CreationDate":"2017-08-21T09:50:35.073","ParentId":null,"OwnerUserId":"3127650","Title":null,"Body":"<p>First, Install CUDA library (version 7.5) from <a href=\"https:\/\/developer.nvidia.com\/cuda-75-downloads-archive\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>Installation Instructions:\n1- <code>sudo dpkg -i cuda-repo-ubuntu1404-7-5-local_7.5-18_amd64.deb<\/code>\n2- <code>sudo apt-get update<\/code>\n3- <code>sudo apt-get install cuda<\/code><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0aH9e.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0aH9e.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Second, install the cuDNN from <a href=\"https:\/\/developer.nvidia.com\/rdp\/cudnn-download\" rel=\"nofollow noreferrer\">here<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/0Je4K.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0Je4K.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Third, export cuDNN path:<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64\/\n<\/code><\/pre>\n\n<p>In case you have an error like \"<strong>The package libcudnnX needs to be reinstalled<\/strong>\", follow those steps <a href=\"http:\/\/www.ihaveapc.com\/2011\/10\/fix-annoying-the-package-needs-to-be-reinstalled-but-i-cant-find-an-archive-for-it-error-in-linux-mint-ubuntu\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"48332903","CreationDate":"2018-01-19T01:35:47.350","ParentId":null,"OwnerUserId":"2452792","Title":null,"Body":"<p>Check the <strong>NVIDIA requirements to run TensorFlow with GPU support<\/strong> (<a href=\"https:\/\/www.tensorflow.org\/install\/install_linux\" rel=\"nofollow noreferrer\">link<\/a>):<\/p>\n\n<ul>\n<li><p>CUDA\u00ae Toolkit 8.0<\/p><\/li>\n<li><p>The NVIDIA drivers associated with CUDA Toolkit 8.0<\/p><\/li>\n<li><p>cuDNN v6.0<\/p><\/li>\n<li><p>GPU card with CUDA Compute <strong>Capability 3.0<\/strong> or higher<\/p><\/li>\n<li><p>The libcupti-dev library, which is the NVIDIA CUDA Profile Tools Interface<\/p><\/li>\n<\/ul>\n\n<p>I installed the cuda v5.1 and the message below still remains:<\/p>\n\n<pre><code>ImportError: libcudart.so.8.0: cannot open shared object file:\n  No such file or directory\n<\/code><\/pre>\n\n<p>I so I got pissed off because everything looks fine, so I decide to check my GPU with the command (on Linux):<\/p>\n\n<pre><code>glxinfo | grep GeForce\n<\/code><\/pre>\n\n<p>And I noticed that my NVIDIA GPU is not supported:<\/p>\n\n<pre><code>OpenGL renderer string: **GeForce GTX 560M**\/PCIe\/SSE2\n<\/code><\/pre>\n\n<p>In this <a href=\"https:\/\/developer.nvidia.com\/cuda-gpus\" rel=\"nofollow noreferrer\">link<\/a> you can find a list, like that:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lWdZd.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lWdZd.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>So my solution was use tensor flow without GPU support. So I do:<\/p>\n\n<pre><code>pip uninstall tensorflow-gpu\n<\/code><\/pre>\n\n<p>I install whithout support:<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"42779596","CreationDate":"2017-03-14T07:01:44.380","ParentId":null,"OwnerUserId":"1653443","Title":null,"Body":"<p>I fixed this just adding the cuda path to my .bashrc<\/p>\n\n<pre><code>export LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64\/\n<\/code><\/pre>\n\n<p>Just have in mind that first you need to go to nvidia Deep Learning page, register and download cuDNN, extract and copy the files from include and lib64 folders into your cuda installation.<\/p>\n"},{"AnswerId":"45414825","CreationDate":"2017-07-31T11:33:46.010","ParentId":null,"OwnerUserId":"329263","Title":null,"Body":"<p>TensorFlow 1.2.1 is compatible with cuDNN 5.1, but not yet with 6.0. So just install cuDNN 5.1. Besides that you seem to be missing CUDA 8.0.<\/p>\n"}]}
{"QuestionId":42013977,"AnswerCount":1,"Tags":"<machine-learning><cygwin><anaconda><theano><keras>","CreationDate":"2017-02-02T23:02:57.470","AcceptedAnswerId":null,"OwnerUserId":7469788.0,"Title":"How should I configure Theano to allow Keras handle an ANN using the GPU (& her CUDA cores) instead of the CPU?","Body":"<p>Trying to change  in '' instead of '', in <strong>Python 3.5<\/strong> under <strong>Anaconda3<\/strong> (mounted on a <em>Windows 10 Workstation<\/em>).<\/p>\n\n<p>I get as response the message :<\/p>\n\n<pre><\/pre>\n\n<p>On the other hand, if I run '' from bash <strong>Cygwin64<\/strong>, going back in <strong>Anaconda<\/strong>, the value remains fixed as ''.<\/p>\n\n<p><strong>Attention:<\/strong> Somewhere I read \"\". Ok, would you try doing it under Windows?<\/p>\n","answers":[{"AnswerId":"42018172","CreationDate":"2017-02-03T06:35:59.927","ParentId":null,"OwnerUserId":"5638904","Title":null,"Body":"<p>My installation steps:<\/p>\n\n<ol>\n<li>install Anaconda Python<\/li>\n<li>in command prompt:<\/li>\n<\/ol>\n\n<pre>\nconda install git libpython mingw\n\npip install git+git:\/\/github.com\/Theano\/Theano.git\n\npip install git+git:\/\/github.com\/fchollet\/keras.git\n<\/pre>\n\n<ol start=\"3\">\n<li>in command prompt:<\/li>\n<\/ol>\n\n<pre>\npython\n\nimport keras\n<\/pre>\n\n<p>At the first run you will see an error that you don't have Tensorflow backend.<\/p>\n\n<p>Win+R to open the Run dialog<\/p>\n\n<p>enter <code>%HOMEPATH%<\/code><\/p>\n\n<p>go in <code>.keras<\/code> folder and edit <code>keras.json<\/code> in notepad: replace 'tensorflow' with 'theano'<\/p>\n\n<ol start=\"4\">\n<li>close and open command prompt again:<\/li>\n<\/ol>\n\n<pre>\npython\n\nimport keras\n<\/pre>\n\n<p>Now it should work.<\/p>\n\n<p><strong>Additional steps for GPU support:<\/strong><\/p>\n\n<ol start=\"5\">\n<li><p>install Visual Studio 2015 (see step 1 in <a href=\"https:\/\/github.com\/philferriere\/dlwin\" rel=\"nofollow noreferrer\">Phil Ferriere's guide<\/a>)<\/p><\/li>\n<li><p>install CUDA 8.0<\/p><\/li>\n<li><p>extract cuDNN 5.1 files in <code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0<\/code><\/p><\/li>\n<li><p>configure Theano:<\/p><\/li>\n<\/ol>\n\n<p>Win+R to open the Run dialog<\/p>\n\n<p>enter <code>%HOMEPATH%<\/code><\/p>\n\n<p>create <code>.theanorc.txt<\/code> with config in it:<\/p>\n\n<pre>\n[global]\n\nfloatX = float32\n\ndevice = gpu0\n\nmode = FAST_RUN\n<\/pre>\n"}]}
{"QuestionId":42014191,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><neural-network><deep-learning>","CreationDate":"2017-02-02T23:23:05.963","AcceptedAnswerId":null,"OwnerUserId":7508594.0,"Title":"Why am I getting \"IndexError: tuple index out of range\"?","Body":"<p>When I run this code I receive an error at line 53 \"IndexError: tuple index out of range.\" When I try changing the Out.shape functions from 1 and 2 to 0 and 1, this error goes away but the LSTM model doesn't work since it requires a 3-D array, and with 0 and 1, I am inputting a 2-D array? How can I fix this and why am I having this issue? Thank you!<\/p>\n\n<pre><\/pre>\n\n<p>Line below is where the error occurs --->>>><\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42014552,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-02T23:59:27.757","AcceptedAnswerId":null,"OwnerUserId":7375883.0,"Title":"Run AlexNet: cuda_error_out_of_memory error","Body":"<p>Gpu memory:11.9G<\/p>\n\n<p>Init alexnet: need 34G, so out of memory.<\/p>\n\n<p>use the one of the two ways of tensorflow official web as below, but still the old question.<\/p>\n\n<pre><\/pre>\n\n<p>why?? any good ideas, thank you very much!!!<\/p>\n","answers":[{"AnswerId":"42029915","CreationDate":"2017-02-03T17:34:45.830","ParentId":null,"OwnerUserId":"7375883","Title":null,"Body":"<p>The real reason is that in shuffle the min_after_dequeue is too big<\/p>\n"}]}
{"QuestionId":42015116,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><autoencoder>","CreationDate":"2017-02-03T01:05:18.790","AcceptedAnswerId":null,"OwnerUserId":5896621.0,"Title":"Decoder's weights of Autoencoder with tied weights in Keras","Body":"<p>I have implemented a tied weights Auto-encoder in Keras and have successfully trained it. <\/p>\n\n<p>My goal is to use only the decoder part of the Auto-encoder as the last layer of another network, to fine tune both the network and the decoder.<\/p>\n\n<p>Thing is, as you can see below from the summary, the decoder has no parameters with my tied weights implementation, so there is nothing to be fine tuned. ( returns )<\/p>\n\n<p>My question is: Should I change the implementation of the tied weights, so that the tied layer can still hold weights, that is the transposed weights of the encoder? If yes, how? <\/p>\n\n<p>Or am I just way off?<\/p>\n\n<p>Below is the summary of the autoencoder model as well as the class of the tied Dense layer (slightly modified from <a href=\"https:\/\/github.com\/nanopony\/keras-convautoencoder\/blob\/master\/autoencoder_layers.py\" rel=\"noreferrer\">https:\/\/github.com\/nanopony\/keras-convautoencoder\/blob\/master\/autoencoder_layers.py.<\/a>)<\/p>\n\n<hr>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"55981431","CreationDate":"2019-05-04T10:03:36.037","ParentId":null,"OwnerUserId":"5140223","Title":null,"Body":"<p>It's been more than 2 years since this question was asked, but this answer might still be relevant for some.<\/p>\n\n<p>The function <code>Layer.get_weights()<\/code> retrieves from <code>self.trainable_weights<\/code> and <code>self.non_trainable_weights<\/code> (see <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/base_layer.py#L1014\" rel=\"nofollow noreferrer\">keras.engine.base_layer.Layer.weights<\/a>). In your custom layer, your weights <code>self.W<\/code> and <code>self.b<\/code> are not being added to any of these collections and that's why the layer has 0 parameters.<\/p>\n\n<p>You could tweak your implementation as follows:<\/p>\n\n<pre><code>class TiedtDense(Dense):\n    def __init__(self, output_dim, master_layer, **kwargs):\n        self.master_layer = master_layer\n        super(TiedtDense, self).__init__(output_dim, **kwargs)\n\n    def build(self, input_shape):\n        assert len(input_shape) &gt;= 2\n        input_dim = input_shape[-1]\n        self.input_dim = input_dim\n\n        self.kernel = tf.transpose(self.master_layer.kernel)\n        self.bias = K.zeros((self.units,))\n        self.trainable_weights.append(self.kernel)\n        self.trainable_weights.append(self.bias)\n<\/code><\/pre>\n\n<p><strong>NOTE:<\/strong> I am excluding the regularizers and constraints for simplicity. If you want those, please refer to <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/base_layer.py#L215\" rel=\"nofollow noreferrer\">keras.engine.base_layer.Layer.add_weight<\/a>.<\/p>\n"}]}
{"QuestionId":42015156,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2017-02-03T01:10:54.307","AcceptedAnswerId":null,"OwnerUserId":5029595.0,"Title":"The order of pooling and normalization layer in convnet","Body":"<p>I'm looking at <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10.py\" rel=\"nofollow noreferrer\">TensorFlow implementation of ORC on CIFAR-10<\/a>, and I noticed that after the first convnet layer, they do pooling, then normalization, but after the second layer, they do normalization, then pooling.<\/p>\n\n<p>I'm just wondering what would be the rationale behind this, and any tips on when\/why we should choose to do norm before pool would be greatly appreciated. Thanks!<\/p>\n","answers":[{"AnswerId":"48659250","CreationDate":"2018-02-07T08:37:50.267","ParentId":null,"OwnerUserId":"3908644","Title":null,"Body":"<p>Normalization is just normalization. After normalizaion, max value is still max value among all.<\/p>\n\n<p>So Normalization->Pooling or Pooling->Normaliazaion results same.<\/p>\n"}]}
{"QuestionId":42015170,"AnswerCount":1,"Tags":"<python><macos><python-3.x><tensorflow><pip>","CreationDate":"2017-02-03T01:12:54.697","AcceptedAnswerId":null,"OwnerUserId":4446167.0,"Title":"installing tensorflow on mac os with pip3","Body":"<p>So I'm having some trouble getting tensorflow working on my mac.\nI checked my current version of python and ran the install.\nFrom the result I can clearly see it's already there:<\/p>\n\n<pre><\/pre>\n\n<p>But when I run python and try importing the module it can't be found. Why is this the case?<\/p>\n\n<pre><\/pre>\n\n<p>Edit: I tried doing what wpercy suggested and got an error<\/p>\n\n<pre><\/pre>\n\n<p>So then I tried<\/p>\n\n<pre><\/pre>\n\n<p>This one didn't give me an error, but I did receive this warning:<\/p>\n\n<pre><\/pre>\n\n<p>It installed successfully, but again, when I try importing it in python the module is not recognized. Does this warning have something to do with it?<\/p>\n\n<p>Alright so I finally got it working, and it required me to do the installation in a conda environment (from the instructions in the tensorflow docs):<\/p>\n\n<pre><\/pre>\n\n<p>but now if I want to use tensorflow I am required to activate this conda environment... This is OK... but it still did not resolve my original problem. Does anyone know how I can go about using tensorflow outside of a conda environment?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42015222","CreationDate":"2017-02-03T01:19:18.337","ParentId":null,"OwnerUserId":"2399775","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#anaconda_installation\" rel=\"nofollow noreferrer\">tensorflow docs<\/a> suggest a special approach when you're using Anaconda.<\/p>\n\n<pre><code># Mac OS X, CPU only, Python 3.4 or 3.5:\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.1-py3-none-any.whl\n\n# Mac OS X, GPU enabled, Python 3.4 or 3.5:\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/mac\/gpu\/tensorflow_gpu-0.12.1-py3-none-any.whl\n\n# Install with Python 3\n$ pip3 install --ignore-installed --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n"}]}
{"QuestionId":42015608,"AnswerCount":1,"Tags":"<python><tensorflow><gpu>","CreationDate":"2017-02-03T02:08:38.350","AcceptedAnswerId":42020112.0,"OwnerUserId":7509001.0,"Title":"Using 2 GPU at same time for different traning's TensorFlow\\Python","Body":"<ol>\n<li>Python 2.7 <\/li>\n<li>TensorFlow<\/li>\n<li>Ubuntu<\/li>\n<\/ol>\n\n<p>Hello my dear friends, have question. Is it possible to use multi GPU for training different models at same time? It's not a problem to use them for one training with simply code manipulation, but what if I want try to do 2 of them at the same time with different parameters at same time?\nI am sorry, if this question is easy, new in programming and TensorFlow. At guide \n<a href=\"https:\/\/www.tensorflow.org\/how_tos\/using_gpu\/\" rel=\"nofollow noreferrer\">Using GPU's<\/a> I tried to find a question about it, but here is for one process only, as I can understand.<\/p>\n","answers":[{"AnswerId":"42020112","CreationDate":"2017-02-03T08:44:27.113","ParentId":null,"OwnerUserId":"5370721","Title":null,"Body":"<p>It looks possible. For your models, you use \"with tf.device..\" and specify different gpus for different models. And then run both the program and see what happens. I tried with a simple program on different gps, it ran. let us know the output you find.<\/p>\n"}]}
{"QuestionId":42015797,"AnswerCount":1,"Tags":"<python><tensorflow><pycharm><keras>","CreationDate":"2017-02-03T02:33:24.007","AcceptedAnswerId":null,"OwnerUserId":5051564.0,"Title":"Is Tensorflow loading twice in PyCharm?","Body":"<p>I'm using flask as the web server running keras with tensor flow backend. Every time I restart it, it shows up like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lUZqn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lUZqn.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>Does this mean that Tensorflow is loading twice? Or is it because I'm running it in PyCharm and the debugger is active - how can I disable it? The 3 lines below don't come up when running on IDLE. Tensorflow takes quite a while to load, so it would be great if I can only have it load once.<\/p>\n\n<p>Thank you for any comment and suggestions.<\/p>\n","answers":[{"AnswerId":"45104702","CreationDate":"2017-07-14T13:50:10.820","ParentId":null,"OwnerUserId":"712203","Title":null,"Body":"<p>I ran into the same error. In my case it was because <code>predict<\/code> was called in a different thread than the one in which <code>load_model<\/code> was called. Moved them to the same thread the the problem was gone.<\/p>\n"}]}
{"QuestionId":42016559,"AnswerCount":1,"Tags":"<neural-network><torch>","CreationDate":"2017-02-03T04:02:28.770","AcceptedAnswerId":null,"OwnerUserId":1159685.0,"Title":"Torch: back-propagation from loss computed over a subset of the output","Body":"<p>I have a simple convolutional neural network, whose output is a single channel 4x4 feature map. During training, the (regression) loss needs to be computed only on a single value among the 16 outputs. <em>The location of this value will be decided after the forward pass.<\/em> How do I compute the loss from just this one output, while making sure all irrelevant gradients are zero'ed out during back-prop.<\/p>\n\n<p>Let's say I have the following simple model in torch:<\/p>\n\n<pre><\/pre>\n\n<p>I have looked at nn.SelectTable and it seems like if I convert the output into tabular form I would be able to implement what I want?<\/p>\n","answers":[{"AnswerId":"42034588","CreationDate":"2017-02-03T23:23:00.057","ParentId":null,"OwnerUserId":"1159685","Title":null,"Body":"<p>This is my current solution. It works by splitting the output into a table, and then using nn.SelectTable():backward() to get the full gradient:<\/p>\n\n<pre><code>require 'nn'\n\n-- the input\nlocal batch_sz = 2\nlocal x = torch.Tensor(batch_sz, 3, 100, 100):uniform(-1,1)\n\n-- the model\nlocal net = nn.Sequential()\nnet:add(nn.SpatialConvolution(3, 128, 9, 9, 9, 9, 1, 1))\nnet:add(nn.SpatialConvolution(128, 1, 3, 3, 3, 3, 1, 1))\nnet:add(nn.Squeeze(1, 3))\n\n-- convert output into a table format\nnet:add(nn.View(1, -1))         -- vectorize\nnet:add(nn.SplitTable(1, 1))    -- split all outputs into table elements\n\nprint(net)\n\n-- the loss\nlocal loss = nn.SmoothL1Criterion()\n\n-- forward'ing x through the network would result in a (2)x4x4 output\ny = net:forward(x)\n\nprint(y)\n\n-- returns the output table's index belonging to specific location\nfunction get_sample_idx(feat_h, feat_w, smpl_idx, feat_r, feat_c)\n    local idx = (smpl_idx - 1) * feat_h * feat_w\n    return idx + feat_c + ((feat_r - 1) * feat_w)\nend\n\n-- I want to back-propagate the loss of this sample at this feature location\nlocal smpl_idx = 2\nlocal feat_r = 3\nlocal feat_c = 4\n-- get the actual index location in the output table (for a 4x4 output feature map)\nlocal out_idx = get_sample_idx(4, 4, smpl_idx, feat_r, feat_c)\n\n-- the (fake) ground-truth\nlocal gt = torch.rand(1)\n\n-- compute loss on the selected feature map location for the selected sample\nlocal err = loss:forward(y[out_idx], gt)\n-- compute loss gradient, as if there was only this one location\nlocal dE_dy = loss:backward(y[out_idx], gt)\n-- now convert into full loss gradient (zero'ing out irrelevant losses)\nlocal full_dE_dy = nn.SelectTable(out_idx):backward(y, dE_dy)\n-- do back-prop through who network\nnet:backward(x, full_dE_dy)\n\nprint(\"The full dE\/dy\")\nprint(table.unpack(full_dE_dy))\n<\/code><\/pre>\n\n<p>I would really appreciate it somebody points out a simpler OR more efficient method.<\/p>\n"}]}
{"QuestionId":42017411,"AnswerCount":1,"Tags":"<tensorflow><linear-regression>","CreationDate":"2017-02-03T05:30:42.287","AcceptedAnswerId":null,"OwnerUserId":243755.0,"Title":"Why I have to shuffle input data for linear regression in tensorflow","Body":"<p>I am using tensorflow to build linear regression model, the following is my code. But from what I experiment, I have to shuffle the train data, otherwise weight and bias would be estimated as na. Could anyone explain to me why I have to shuffle the data ? Thanks<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"58582463","CreationDate":"2019-10-27T19:05:13.123","ParentId":null,"OwnerUserId":"4891729","Title":null,"Body":"<p>sometimes data is ordered by some columns\nand when you split you data to ratio of 75% vs 25%\nyou are blind for some values that exists in the last 25% split.\nso you learn everything except the values that exists in the test (last 25% rows) se.\nthats why the best is to shuffle, to make sure you break some ordering in the data and that you learn all the variety of possible values that exist<\/p>\n"}]}
{"QuestionId":42017833,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-03T06:09:10.253","AcceptedAnswerId":null,"OwnerUserId":1452257.0,"Title":"Tensorflow load graph into specific scope","Body":"<p>I'm loading a pretrained network into Tensorflow using the methods below that are within a  class (hence the calls to self.xyz). First,  is called, then I do initialization of other variables and optimizers, then  is called.<\/p>\n\n<p>However, despite using  the variables from the graph are loaded into the generic space of variables. This is problematic as I have two instances of this class that each load the same network and I want to separate the out into different scopes.<\/p>\n\n<p>How can I load the variables into a specific scope?<\/p>\n\n<p>P.S. Feel free to correct me on any errors in my code!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42021548","CreationDate":"2017-02-03T10:00:34.067","ParentId":null,"OwnerUserId":"6933420","Title":null,"Body":"<p>I think your problem can be solved by adding an argument into  <\/p>\n\n<pre><code>self.saver = tf.train.import_meta_graph(self.model_path + '.meta', 'import_scope'=self.name)\n<\/code><\/pre>\n\n<p>Here's the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/exporting_and_importing_meta_graphs#import_meta_graph\" rel=\"nofollow noreferrer\">reference<\/a><\/p>\n"}]}
{"QuestionId":42018010,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-03T06:23:09.170","AcceptedAnswerId":null,"OwnerUserId":3425082.0,"Title":"Copy tensorflow contrib layers","Body":"<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I change the value of w1 using some optimizer, the value of w2 is also changed. It seems that the copy function only returns the reference of w1 instead of a new instance of it. I also tried tf.identity, which gave the same result.<\/p>\n\n<p>How can I create a new instance of w1 whose its content doesn't change when I change w1?<\/p>\n","answers":[]}
{"QuestionId":42019818,"AnswerCount":0,"Tags":"<tensorflow><importerror>","CreationDate":"2017-02-03T08:27:28.253","AcceptedAnswerId":null,"OwnerUserId":6201311.0,"Title":"Tensorflow: load_op_library load incorrectly","Body":"<p>I've added a new Op (for example, it's name actv). I have a GPU core for it, so I have source files: actv.cc and actv_kernel.cu.cc. Then I compiled it (as in Adding new Op manual), so now I have actv.so. Then<\/p>\n\n<pre><\/pre>\n\n<p>In the output I see most important for me -- 'actv'. So, everything is fine.\nLet's try do it in another directory (same sources, same compiler flags, same two lines of code):<\/p>\n\n<pre><\/pre>\n\n<p>And there isn't 'actv' in it. Even more, if I just copy test\/actv.so to test2\/test3\/actv.so, problem still arises. So, two questions: \n1) What's a king of black magic? \n2) How can I make such loading by relative path?<\/p>\n","answers":[]}
{"QuestionId":42020571,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-03T09:11:47.717","AcceptedAnswerId":42020971.0,"OwnerUserId":5046896.0,"Title":"Are the ckpt data under Python 2.7 and Python 3.4 compatible?","Body":"<p>My model is written with Python, and I got ckpt data under Python 2.7 and Python 3.4, Are they(ckpt data) compatible? Since Python 2.7 is not compatible with Python 3.x<\/p>\n\n<p>It will cost a lot time to train the model to get the ckpt data, so didn't try with both ckpt data under Python 2.7 and Python 3.4 to test whether they(ckpt data) are compatible.<\/p>\n","answers":[{"AnswerId":"42020971","CreationDate":"2017-02-03T09:30:43.470","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>The checkpoints are stored using <a href=\"https:\/\/developers.google.com\/protocol-buffers\/\" rel=\"nofollow noreferrer\">Protocol Buffers<\/a> (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/saver.py\" rel=\"nofollow noreferrer\"><code>Saver<\/code> source<\/a>), and therefore its format is independent of the Python version. In fact, in theory you could even load checkpoint files in C++, for example.<\/p>\n\n<p>Checkpoint compatibility could break if the Protocol Buffers specs for TensorFlow models change after a release, but that seems rather unlikely.<\/p>\n"}]}
{"QuestionId":42020606,"AnswerCount":1,"Tags":"<tensorflow><crf>","CreationDate":"2017-02-03T09:13:25.510","AcceptedAnswerId":null,"OwnerUserId":6227444.0,"Title":"How to correctly pass initial value of transition_params in tensorflow linear chain CRF","Body":"<p>I'm trying to use the linear chain CRF in my work. I took the help of the example usage code provided in -- <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r1.0\/tensorflow\/contrib\/crf\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r1.0\/tensorflow\/contrib\/crf<\/a><\/p>\n\n<p>My question is how to supply some initial value of \"transition_params\" in \"crf_log_likelihood()\". For concreteness of the example, say, I want to initialize it with standard random normal distribution. In the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/contrib\/crf\/python\/ops\/crf.py#L128\" rel=\"nofollow noreferrer\">api doc<\/a>, I saw that \"transition_params\" can, in fact, be passed as an input argument. Inside the method I see that if no \"transition_params\" is passed, it is obtained by doing a \"vs.get_variable()\" with name = \"transitions\".<\/p>\n\n<p>So should I do something similar to this, before creating the 'crf_log_likelihood' op? Something like -- transition_params = vs.get_variable(\"transitions\", [num_tags, num_tags], initializer=tf.random_normal_initializer()) -- and then change the call of \"crf_log_likelihood()\" to \"log_likelihood, transition_params = tf.contrib.crf.crf_log_likelihood(unary_scores, y_t, sequence_lengths_t, transition_params)\"?<\/p>\n","answers":[{"AnswerId":"42071521","CreationDate":"2017-02-06T15:34:18.943","ParentId":null,"OwnerUserId":"6392807","Title":null,"Body":"<p>The <code>get_variable()<\/code> inside the definition of <code>crf_log_likelihood()<\/code> will create a fresh, randomly-initialized variable to represent the transition parameters, if you don't provide one yourself. You only need to provide an explicit <code>transition_params<\/code> if you don't want the default behavior.<\/p>\n\n<p>To understand the behavior of <code>get_variable()<\/code>, see here:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sharing_variables#get_variable\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sharing_variables#get_variable<\/a><\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":42022950,"AnswerCount":2,"Tags":"<python><tensorflow><random-seed>","CreationDate":"2017-02-03T11:16:19.043","AcceptedAnswerId":58409022.0,"OwnerUserId":1934212.0,"Title":"Which seeds have to be set where to realize 100% reproducibility of training results in tensorflow?","Body":"<p>In a general tensorflow setup like<\/p>\n\n<pre><\/pre>\n\n<p>Where  contains the model definition including random initialization of weights () and  executes the training of the model -<\/p>\n\n<p>Which seeds do I have to set where to ensure 100% reproducibility between repeated runs of the code snippet above? <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/random\/set_random_seed\" rel=\"nofollow noreferrer\">The documentation<\/a> for  may be concise, but left me a bit confused. I tried:<\/p>\n\n<pre><\/pre>\n\n<p>But got different results each time. <\/p>\n","answers":[{"AnswerId":"42023225","CreationDate":"2017-02-03T11:29:54.147","ParentId":null,"OwnerUserId":"3531215","Title":null,"Body":"<p>One possible reason is that when constructing the model, there are some code using numpy.random module. So maybe you can try to set the seed for numpy, too. <\/p>\n"},{"AnswerId":"58409022","CreationDate":"2019-10-16T08:33:14.387","ParentId":null,"OwnerUserId":"9828197","Title":null,"Body":"<p>The best solution which works as of today with GPU is to install tensorflow-determinism with the following:<\/p>\n\n<pre><code>pip install tensorflow-determinism\n<\/code><\/pre>\n\n<p>Then include the following code to your code<\/p>\n\n<pre><code>import tensorflow as tf\nimport os\nos.environ['TF_DETERMINISTIC_OPS'] = '1'\n\n<\/code><\/pre>\n\n<p>source: <a href=\"https:\/\/github.com\/NVIDIA\/tensorflow-determinism\" rel=\"nofollow noreferrer\">https:\/\/github.com\/NVIDIA\/tensorflow-determinism<\/a><\/p>\n"}]}
{"QuestionId":42024795,"AnswerCount":1,"Tags":"<tensorflow><dask>","CreationDate":"2017-02-03T12:52:32.133","AcceptedAnswerId":42025125.0,"OwnerUserId":3318517.0,"Title":"Dask compute() blocks and does not do anything","Body":"<p>I have a TensorFlow model that only uses 2 cores during training. I have 8 cores and want to train a bunch of these models on different random samples to compare. I thought I would save time by training 4 models in parallel.<\/p>\n\n<pre><\/pre>\n\n<p> just blocks indefinitely and there is no CPU use. What did I mess up?<\/p>\n\n<p>I looked at <a href=\"http:\/\/dask.pydata.org\/en\/latest\/inspect.html\" rel=\"nofollow noreferrer\">http:\/\/dask.pydata.org\/en\/latest\/inspect.html<\/a> and tried printing . I think it looks fine:<\/p>\n\n<pre><\/pre>\n\n<p>What else can I check to understand what's going on? It's an Ubuntu 16.04 system.<\/p>\n","answers":[{"AnswerId":"42025125","CreationDate":"2017-02-03T13:10:29.433","ParentId":null,"OwnerUserId":"616616","Title":null,"Body":"<h3>Try using Threads<\/h3>\n\n<p>By default dask.bag parallelizes using separate processes on your computer.  This is ideal for pure Python code (because of the GIL) but may not be ideal for numeric code like Tensorflow, especially if the Tensorflow library doesn't play well with forked processes (which may be the case?)<\/p>\n\n<p>You can do this globally by setting the following<\/p>\n\n<pre><code>import dask\ndask.set_options(get=dask.threaded.get)\n<\/code><\/pre>\n\n<p>or setting <code>get=<\/code> in the compute call<\/p>\n\n<pre><code>accuracies.compute(get=dask.threaded.get)\n<\/code><\/pre>\n\n<h3>Consider using dask.delayed<\/h3>\n\n<p>Dask.bag provides a pretty simple interface, somewhat similar to the Spark RDD.  For more complex algorithms in the future, you might also try <a href=\"http:\/\/dask.pydata.org\/en\/latest\/delayed.html\" rel=\"nofollow noreferrer\">dask.delayed<\/a>.<\/p>\n"}]}
{"QuestionId":42025077,"AnswerCount":2,"Tags":"<python><numpy><lambda><tensorflow><matrix-multiplication>","CreationDate":"2017-02-03T13:07:48.023","AcceptedAnswerId":null,"OwnerUserId":2294865.0,"Title":"How to apply calculations to elements of a multidimensional matrix in Tensorflow?","Body":"<p>I'm relatively new to Python and even more so to Tensorflow so I've been working through some tutorials such as this <a href=\"http:\/\/learningtensorflow.com\/lesson4\/\" rel=\"nofollow noreferrer\">tutorial<\/a>. A challenge given was to make an image greyscale. One approach taken <a href=\"https:\/\/stackoverflow.com\/questions\/37860136\/converting-rgb-to-grayscale-manually-tensorflow\">here<\/a> is to just take one colour channel value and duplicate it across all channels. Another is to take an average which can be achieved using tf.reduce_mean as done <a href=\"https:\/\/stackoverflow.com\/questions\/40235030\/grayscale-convertion-using-tf-reduce-mean-tf-concat\">here<\/a>. However there are many ways to make an image monochromatic as anyone who has played with GIMP or Photoshop will know. One standard method defined adjusts for the way humans perceive colour and requires that the three colour channels are individually adjusted this way:<\/p>\n\n<pre><\/pre>\n\n<p>Anyway I've achieved it by doing this:<\/p>\n\n<pre><\/pre>\n\n<p>This to me seems hugely inelegant having to cut up the data and apply calculations and then recombine them. A matrix multiplication on individual RGB tuples would be efficient or barring that a function that takes an individual RGB tuple and returns a greyscaled tuple. I've looked at tf.map_fn but can't seem to make it work for this.<\/p>\n\n<p>Any suggestions or improvements?<\/p>\n","answers":[{"AnswerId":"42030757","CreationDate":"2017-02-03T18:29:27.720","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>How about this?  <\/p>\n\n<pre><code>img = tf.ones([100, 100, 3])\nr, g, b = tf.unstack(img, axis=2)\ngrey = r * 0.2126 + g * 0.7152 + b * 0.0722\nout = tf.stack([grey, grey, grey], axis=2)\nout = tf.cast(out, tf.uint8)\n<\/code><\/pre>\n\n<p>sample of <code>map_fn<\/code><br>\nshape of x is <code>(2, 4)<\/code>, so shape of <code>elms_fn<\/code> is <code>(4,)<\/code><br>\nif shape of x is <code>(100, 100, 3)<\/code>, shape of <code>elms_fn<\/code> will be <code>(100, 3)<\/code><\/p>\n\n<pre><code>x = tf.constant([[1, 2, 3, 4],\n                 [5, 6, 7, 8]], dtype=tf.float32)\n\ndef avg_fc(elms_fn):\n    # shape of elms_fn is (4,)\n    # compute average for each row and return it\n    avg = tf.reduce_mean(elms_fn)\n    return avg\n\n# map_fn will stack avg at axis 0\nres = tf.map_fn(avg_fc, x)\n\nwith tf.Session() as sess:\n    a = sess.run(res) #[2.5, 6.5]\n<\/code><\/pre>\n"},{"AnswerId":"42058641","CreationDate":"2017-02-05T23:45:49.003","ParentId":null,"OwnerUserId":"2294865","Title":null,"Body":"<p>So having really looked in to this topic, in the current release of tensorflow (r0.12) there doesn't appear to be a simple way to apply custom functions to tuples of values, especially if the result does not effect a reduce. As my initial effort and that of the <a href=\"https:\/\/stackoverflow.com\/a\/42030757\/2294865\">answer from @xxi<\/a> you pretty much have to dis-aggregate the tuples before applying a function to them collectively.<\/p>\n\n<p>I figured out another way to get the result that I wanted without slicing or unstacking but instead reshaping and matrix multiplication which is:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.image as mpimg\nimport matplotlib.pyplot as plt\n\nfilename = \"MarshOrchid.jpg\"\nraw_image_data = mpimg.imread(filename)\n\nimage = tf.placeholder(\"float\", [None, None, 3])\n\nout = tf.reshape(image, [-1,3])\nout = tf.matmul(out,[[0.2126, 0, 0], [0, 0.7152, 0], [0, 0, 0.0722]])\nout = tf.reduce_sum(out, 1, keep_dims=True)\n\nout = tf.concat(1, [out, out, out])\nout = tf.reshape(out, tf.shape(image))\nout = tf.cast(out, tf.uint8)\n\n\nwith tf.Session() as session:\n\n    result = session.run(out, feed_dict={image: raw_image_data})\n\n    plt.imshow(result)\n    plt.show()\n<\/code><\/pre>\n\n<p>This worked for the narrow purpose of greyscaling an image but doesn't really give a design pattern to apply for dealing with more generic calculations.<\/p>\n\n<p>Out of curiosity I profiled these three methods in terms of execution time and memory usage. So which was better?<\/p>\n\n<ul>\n<li>Method 1 - Slicing:    1.6 seconds &amp; 1.0 GiB memory usage<\/li>\n<li>Method 2 - Unstacking: 1.6 seconds &amp; 1.1 GiB memory usage<\/li>\n<li>Method 3 - Reshape:    1.4 seconds &amp; 1.2 GiB memory usage<\/li>\n<\/ul>\n\n<p>So no major differences in performance but interesting nonetheless.<\/p>\n\n<p>In case you were wondering why the process is so slow, the image used is 5528 x 3685 pixels. But yeah still pretty slow compared to Gimp and others.<\/p>\n"}]}
{"QuestionId":42025093,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><conv-neural-network><tensorboard>","CreationDate":"2017-02-03T13:08:39.010","AcceptedAnswerId":null,"OwnerUserId":3442956.0,"Title":"How to plot grid of images in tensorboard?","Body":"<p>I am able to plot images using  and this plots 'max_images' number of images but I want to plot images as a grid, something like 3x3 grid of images.<\/p>\n\n<p>How to do this?<\/p>\n","answers":[{"AnswerId":"55247716","CreationDate":"2019-03-19T18:24:44.287","ParentId":null,"OwnerUserId":"11199580","Title":null,"Body":"<p>Try to use the <strong>family<\/strong> parameter of <code>tf.summary.image(name, family='...')<\/code> to group several images into one part.<\/p>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/image\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/image<\/a> for more details.<\/p>\n"}]}
{"QuestionId":42025364,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-03T13:23:52.470","AcceptedAnswerId":null,"OwnerUserId":5107084.0,"Title":"TensorFlow: Checking whether a pre-trained model is successfully restored?","Body":"<p>I am currently running a  using a supervisor to test the performance of the official inception-resnet-v2 model checkpoint file (offered by Google) on a few dog images (Border Collie images that I took from Google images). However I am seeing a very strange behavior: The losses for a managed session run by a supervisor with an  that takes a saver that restores all the variables from the checkpoint file and another managed session run by a supervisor without an  at all gives almost an entirely similar loss (around 6). What is funny is that the supervisor with no  even has a lesser loss of under 6 (around 5.9 to be specifc), while the one with  has a loss slightly over 6. I understand that even if the loss may be random at times, how could the performance of both be so similar?<\/p>\n\n<p>I am wary that my model might be reusing the same weights from a previous session (just a guess), and so I used  at the very start of my code. Also, I also made sure to clear the log directory where the supervisor might look for checkpoints from earlier sessions. Yet, the performance is so similar. Is it just that the model performing too well with or without a checkpoint file? All the images have been successfully predicted with a label of 232 (border collie). <\/p>\n\n<p>Just to be sure, I did an initial prediction before any of my training started. The non-restored model has an initial prediction of class 411 (apron), while the restored model has the correct prediction of 232. But I am not sure if this is indicative that my restored model is working - how do I definitively check whether the model is successfully restored?<\/p>\n\n<p>Also, what is considered as an acceptable loss for a softmax_cross_entropy loss? I think a value of around 6 may be too high, but I can't find the official loss obtained by the Google paper on the inception-resnet-v2 model. <\/p>\n\n<p>Here are some details about my model:\n1. Due the my GPU memory limitations, I could only run a batch size of 6. I included only 8 images to test.<\/p>\n\n<ol start=\"2\">\n<li>The non-restored model has an initial loss of around 8, while the restored one has an initial model of around 6.<\/li>\n<\/ol>\n\n<p>The main difference in my code for both supervisors is this:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42025409,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2017-02-03T13:26:36.450","AcceptedAnswerId":null,"OwnerUserId":482601.0,"Title":"Remove small values from matrix","Body":"<p>I have a  tensor  where I want to zero out all values smaller than some threshold . I.e.,<\/p>\n\n<pre><\/pre>\n\n<p>I was wondering, is there a more efficient way to do this? Because  in my setup is huge and as I understand it, the  constructs an other tensor that needs as much memory as .<\/p>\n","answers":[{"AnswerId":"42038575","CreationDate":"2017-02-04T09:16:04.423","ParentId":null,"OwnerUserId":"7514887","Title":null,"Body":"<pre><code>foo = tf.constant([1., 2., 3., 4., 5., 6., 7., 8., 9., 10.])\nthreshold_map = tf.greater(foo, tf.constant(5.))\nthreshold_map_index = tf.reshape(tf.where(threshold_map), [-1])\nfoo_threshold = tf.gather(foo, threshold_map_index)\n# foo_threshold = [6., 7., 8., 9., 10.]\n<\/code><\/pre>\n\n<p>( this <strong>won't<\/strong> work with more then one-dimension )<\/p>\n"},{"AnswerId":"42030441","CreationDate":"2017-02-03T18:08:39.760","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>I am not sure if this will more efficient  <\/p>\n\n<pre><code>x = tf.constant([1, 2, 3, 4, 5, 6, 7])\ny = tf.where(tf.greater(x, tf.constant(5)),\n             x, # if ture\n             tf.zeros_like(x)) # if false\n\nwith tf.Session() as sess:\n    a = sess.run(y)\n    # a is [0, 0, 0, 0, 0, 6, 7]\n<\/code><\/pre>\n"},{"AnswerId":"42025523","CreationDate":"2017-02-03T13:33:08.507","ParentId":null,"OwnerUserId":"3416356","Title":null,"Body":"<p>What is wrong with the good old<\/p>\n\n<pre><code>for i in range(n):\n   for j in range(m):\n      if X[n][m] &lt; t: X[n][m] = 0\n<\/code><\/pre>\n"},{"AnswerId":"42025526","CreationDate":"2017-02-03T13:33:13.570","ParentId":null,"OwnerUserId":"3510686","Title":null,"Body":"<p>If X is your matrix (a numpy array I assume) you can try:<\/p>\n\n<pre><code>x[x&lt;small_value]=0\n<\/code><\/pre>\n\n<p>if creating the boolean array takes too much memory you can try doing that through a loop by individual columns.<\/p>\n"}]}
{"QuestionId":42025534,"AnswerCount":0,"Tags":"<python-3.x><keras><theano-cuda><keras-layer>","CreationDate":"2017-02-03T13:33:51.747","AcceptedAnswerId":null,"OwnerUserId":4730816.0,"Title":"Computation in fixed point or int","Body":"<p>I am using fixed point numbers within my network based on keras framework. My concern is when there are multiplication operations in the network on theano variables, the result is float32 ( even if the numbers supplied are in fixed point). Is there any intrinsic way to get the result in fixed point format, or even int.<\/p>\n\n<p>If not, what can be alternative approaches?<\/p>\n","answers":[]}
{"QuestionId":42025546,"AnswerCount":2,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-02-03T13:34:40.357","AcceptedAnswerId":42038250.0,"OwnerUserId":5155425.0,"Title":"Keras: network doesn't train with fit_generator()","Body":"<p>I'm using Keras on the large dataset (Music autotagging with MagnaTagATune dataset). So I've tried to use fit_generator() fuction with a custom data generator. But the value of loss function and metrics doesn't change during the training process. It looks like my network doesen't train at all.<\/p>\n\n<p>When I use fit() function instead of fit_generator() everything is OK, but I can't keep the whole dataset in memory.<\/p>\n\n<p>I've tried with both Theano and TensorFlow backends<\/p>\n\n<p><strong>Main code:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>class generator_v2:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>To prepare arrays for fit() function I use this code:<\/p>\n\n<pre><\/pre>\n\n<p>Sorry, if the style of my code is not good. And thank you!<\/p>\n\n<p><strong>UPD 1:<\/strong>\nI've tried to use  instead of  but nothing changes.<\/p>\n\n<p><strong>Part of my new generator class:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>fit_generator call:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Logs:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42026763","CreationDate":"2017-02-03T14:41:36.860","ParentId":null,"OwnerUserId":"6833523","Title":null,"Body":"<p>In method 'generate', there is a while statement.<\/p>\n\n<pre><code>def generate(self, begin, end):\n    while(1): # this\n        for count in range(begin, end):\n            try:\n                # something\n                yield(...)\n\n            except FileNotFoundError:\n                continue\n<\/code><\/pre>\n\n<p>I think this statement is not needed, so <\/p>\n\n<pre><code>def generate(self, begin, end):\n    for count in range(begin, end):\n        try:\n            # something\n            yield(...)\n\n        except FileNotFoundError:\n            continue\n<\/code><\/pre>\n"},{"AnswerId":"42038250","CreationDate":"2017-02-04T08:34:30.153","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>I had the same problem as you with the yield method. So i just stored the current index and returned one batch per call with the return statement. <\/p>\n\n<p>So I just used <code>return (X, y)<\/code> instead of <code>yield (X,y)<\/code> and it worked. I am not sure why this is. It would be cool if someone could shed a light on this.<\/p>\n\n<p>Edit: \nYou need to pass in the generator to the function not only call the function. Something like this:  <\/p>\n\n<pre><code>model.fit_generator(gen, samples_per_epoch=750,\n                                  nb_epoch=80,\n                                  validation_data=gen,\n                                  nb_val_samples=250)\n<\/code><\/pre>\n\n<p>Keras will call your __next__ function, while training on the data.<\/p>\n"}]}
{"QuestionId":42025660,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform><keras><keras-layer>","CreationDate":"2017-02-03T13:41:10.580","AcceptedAnswerId":null,"OwnerUserId":7495302.0,"Title":"Keras set_learning_phase for Dropout when saving TensorFlow Session","Body":"<p>I have been following <a href=\"https:\/\/blog.keras.io\/keras-as-a-simplified-interface-to-tensorflow-tutorial.html\" rel=\"nofollow noreferrer\">this post<\/a>, in particular part II, to use Keras as an interface to TensorFlow. <\/p>\n\n<p>As an example I have been training a CNN using the MNIST data set. My aim is to train and evaluate a model in a TF session then save the session using  so that I can deploy the model on CloudML. <\/p>\n\n<p>I am able to do this for a model which does not use Dropout, however, when I include Dropout layers in Keras you need to specify the learning_phase (training = 1, testing = 0), this is done through the feed_dict (see code below). <\/p>\n\n<p>Locally I am able to control this by doing something like<\/p>\n\n<pre><\/pre>\n\n<p>However when I upload my model to CloudML and try to test I get the following error<\/p>\n\n<pre><\/pre>\n\n<p>I know it is because of the line in the feed_dict but I have no idea how to get around it. In the blog post section IV it talks about this problem in the context of TensorFlow serving where the model is loaded and re-saved. I couldn't get this to work for my approach because I need to export the session export and export.meta, not the Keras model. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42793050","CreationDate":"2017-03-14T17:45:29.243","ParentId":null,"OwnerUserId":"4743732","Title":null,"Body":"<p>As this is a very Keras oriented programming issue, it would be best to post this question directly to their GitHub <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\" rel=\"nofollow noreferrer\">Issue Tracker<\/a>. <\/p>\n\n<p>You can also find that <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2310\" rel=\"nofollow noreferrer\">this same issue<\/a> was reported &amp; solved <a href=\"https:\/\/github.com\/fchollet\/keras\/issues?utf8=%E2%9C%93&amp;q=keras_learning_phase\" rel=\"nofollow noreferrer\">many times<\/a> in their Issue Tracker already. The solution to your issue might also be covered in the <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-obtain-the-output-of-an-intermediate-layer\" rel=\"nofollow noreferrer\">Keras documentation<\/a>. <\/p>\n"}]}
{"QuestionId":42026835,"AnswerCount":2,"Tags":"<python><numpy><machine-learning><deep-learning><keras>","CreationDate":"2017-02-03T14:44:58.150","AcceptedAnswerId":null,"OwnerUserId":4140027.0,"Title":"Problems while trying to understand matrix shapes in Keras?","Body":"<p>I am learning keras and this is the first time I am using it in a toy example, so I tried the following linear regression, however I got a :<\/p>\n\n<p>In:<\/p>\n\n<pre><\/pre>\n\n<p>Out:<\/p>\n\n<pre><\/pre>\n\n<p>Data:<\/p>\n\n<pre><\/pre>\n\n<p>labels:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42029039","CreationDate":"2017-02-03T16:45:28.683","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p><code>input_shape<\/code> parameter of <code>Dense()<\/code> should be an integer. So, try the following:<\/p>\n\n<pre><code>model.add(Dense(nb_classes, input_dim=dims))\n<\/code><\/pre>\n\n<p>From <a href=\"https:\/\/keras.io\/layers\/core\/#dense\" rel=\"nofollow noreferrer\">official documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><strong>keras.layers.core.Dense<\/strong>(output_dim, init='glorot_uniform', activation=None, weights=None, W_regularizer=None, b_regularizer=None, activity_regularizer=None, W_constraint=None, b_constraint=None, bias=True, input_dim=None)<\/p>\n<\/blockquote>\n\n<p>where<\/p>\n\n<blockquote>\n  <p><strong>input_dim<\/strong>: dimensionality of the input (integer). This argument (or alternatively, the keyword argument input_shape) is required when using this layer as the first layer in a model.<\/p>\n<\/blockquote>\n"},{"AnswerId":"42029692","CreationDate":"2017-02-03T17:21:55.053","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The first parameter of a Dense layer is the number of outputs of that layer. So in your case, you have an input with shape (1000, 20), and labels with shape (1000,). This means you have 1000 training samples, with 20 features and each label is one dimensional (and you have 1000 of them). Then you need to modify the model to have a single output:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(1, input_dim=dims))\nmodel.compile(optimizer='sgd', loss='mean_squared_error')\n<\/code><\/pre>\n\n<p>Here I used the fact that you want to implement linear regression to remove the softmax (which is for classification), and use the mean square error loss instead of cross entropy.<\/p>\n"}]}
{"QuestionId":42028066,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-03T15:50:26.130","AcceptedAnswerId":42079737.0,"OwnerUserId":7241513.0,"Title":"TensorFlow CSV import: adding features and labels to Summary for TensorBoard reads double the lines","Body":"<p>I have a very basic TensorFlow app to test loading the data from CSV line-by-line and adding various summaries and visualizations to TensorBoard. My input CSV file has 18 rows and a bunch of columns -- first XX columns are 'features' and subsequent YY columns are 0s and 1s representing the label.<\/p>\n\n<p>I noticed that when I create Summaries for the variables that hold Features and Labels, TensorFlow reads twice as many lines from the CSV, so instead of looping 18 times given my 18 rows, I only loop 9. As soon as I remove the code that adds Features and Labels to Summary, the reading goes back to normal. Adding other variables to Summary, such as Cost\/etc. that are the output of functions, rather than the CSV, does not have this issue.<\/p>\n\n<p>I don't yet know whether I will actually need that information in TensorBoard, so I can live without it, but I'd prefer to put in as much info as possible into TensorBoard first, and then decide what I need to keep. <\/p>\n\n<p>Is this expected behavior or am I doing something wrong? <\/p>\n\n<pre><\/pre>\n\n<p>Thanks for the input!<\/p>\n","answers":[{"AnswerId":"42079737","CreationDate":"2017-02-07T00:31:16.410","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>The issue is that each session.run call pulls from the queue (the first one does it explicitly, the second because the summary ops rely on queue data). Rather than using feed_dict to feed previously pulled data, if you instead have the summaries and the actual use of the queue data in the same session.run call, there will be no data thrown away. Something like:<\/p>\n\n<pre><code>examples, labels = input_pipeline(sampling_size, try_epochs)\n\nx, y_ = examples, labels\n\nW = tf.Variable(tf.random_normal([TS, TL], stddev=1), name='weights')\nb = tf.Variable(tf.random_normal([TL], stddev=1), name='biaes')\ny = tf.matmul(x, W) + b\n\n\n# this one causes the issue\nwith tf.name_scope('Features'):\n  features = examples\n# this one also causes the issue\nwith tf.name_scope('Labels'):\n  labDisp = labels\nwith tf.name_scope('Model'):\n  myModel = tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y)\nwith tf.name_scope('Loss'):\n  lossFn = tf.reduce_mean(myModel)\nwith tf.name_scope('Optimizer'):\n  train_step = tf.train.AdamOptimizer(.05).minimize(lossFn)\n\na1 = tf.summary.histogram(\"Features\", features)\na2 = tf.summary.histogram(\"Labels\", labDisp)\na3 = tf.summary.histogram(\"Model\", myModel)\na4 = tf.summary.scalar(\"Loss\", lossFn)\n\nmerged_summary_op = tf.summary.merge([a1, a2, a3, a4])\n\nwith tf.Session() as sess:\n  gInit = tf.global_variables_initializer().run()\n  lInit = tf.local_variables_initializer().run()\n\n  summary_writer = tf.summary.FileWriter(logs_path, graph=tf.get_default_graph())\n\n  coord = tf.train.Coordinator()\n  threads = tf.train.start_queue_runners(coord=coord)\n\n  try:\n    while not coord.should_stop():\n      act = tf.argmax(labels, 1)\n      fit = tf.argmax(y, 1)\n      _, pAct, pFit, l, summary = sess.run([train_step, act, fit, lossFn,\n                                            merged_summary_op])\n      summary_writer.add_summary(summary, i)\n      print(pAct)\n      print(pFit)\n\n  except tf.errors.OutOfRangeError:\n    print('Finished')\n  finally:\n    coord.request_stop()\n  coord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":42028191,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><bazel>","CreationDate":"2017-02-03T15:58:19.337","AcceptedAnswerId":null,"OwnerUserId":7510512.0,"Title":"No Checkpoint file found in none","Body":"<p>I'm using <a href=\"https:\/\/github.com\/tensorflow\/models\/im2txt\" rel=\"nofollow noreferrer\">im2txt<\/a> from TensorFlow.<\/p>\n\n<p>When I run:<\/p>\n\n<pre><\/pre>\n\n<p>I get an error:<\/p>\n\n<blockquote>\n  <p>No Checkpoint file found in :None <\/p>\n<\/blockquote>\n\n<p>Checking  returns None but checkpoint is stored inside .<\/p>\n","answers":[]}
{"QuestionId":42029046,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-03T16:45:45.437","AcceptedAnswerId":null,"OwnerUserId":3296050.0,"Title":"Tensorflow batch loss spikes when restoring model for training from saved checkpoint?","Body":"<p>I'm encountering a strange issue that I've been trying to debug, without much luck. My model starts training properly with batch loss decreasing consistently (from ~6000 initially to ~120 after 20 epochs). However, when I pause training and resume training later by restoring the model from the checkpoint, the batch loss seems to spike unexpectedly from the previous batch loss (before pausing), and resumes decreasing from that higher loss point. My worry is that when I restore the model for evaluation, I may not be using the trained model that I think I am.<\/p>\n\n<p>I have combed over my code several times, comparing to the Tensorflow tutorials. I tried to ensure that I was saving and restoring using the tutorial-suggested methods. Here is the code snapshot: <a href=\"https:\/\/github.com\/KaranKash\/DigitSpeak\/tree\/b7dad3128c88061ee374ae127579ec25cc7f5286\" rel=\"nofollow noreferrer\">https:\/\/github.com\/KaranKash\/DigitSpeak\/tree\/b7dad3128c88061ee374ae127579ec25cc7f5286<\/a> - the train.py file contains the saving and restoring steps, the graph setup and training process; while model.py creates the network layers and computes loss.<\/p>\n\n<p>Here is an example from my print statements - notice batch loss rises sharply when resuming training from epoch 7's checkpoint:<\/p>\n\n<pre><\/pre>\n\n<p>I've printed the results of the inspect_checkpoint.py script. I've also experimented with other loss functions (Adam and GradientDescentOptimizer) and noticed the same behavior with respect to spiked loss after resuming training.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"53562826","CreationDate":"2018-11-30T18:16:07.833","ParentId":null,"OwnerUserId":"10729015","Title":null,"Body":"<p>I ran into this issue and found it was the fact that I was initializing the graph variables when restoring the graph -- throwing away all learned parameters, to be replaced with whatever initialization values were originally specified for each respective tensor in the original graph definition.<\/p>\n\n<p>For example, if you used <strong>tf.global_variable_initializer()<\/strong> to initialize variables as part of your model program, whatever your control logic to indicate that a saved graph will be restored, make sure the graph restore flow omits:  <strong>sess.run(tf.global_variable_initializer())<\/strong><\/p>\n\n<p>This was a simple, but costly mistake for me, so I hope someone else is saved a few grey hairs (or hairs, in general).<\/p>\n"}]}
{"QuestionId":42030350,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-03T18:03:03.283","AcceptedAnswerId":null,"OwnerUserId":7512656.0,"Title":"ImportError: No module named 'tensorflow' error occur","Body":"<p>I am running my first python Tensorflow programm in spyder IDE.  I have also change python interpreter preference but still I get <\/p>\n\n<blockquote>\n  <p>ImportError: No module named 'tensorflow' error occur. <\/p>\n<\/blockquote>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/K3yBs.png\" alt=\"ImportError: No module named &#39;tensorflow&#39;\"><\/p>\n","answers":[{"AnswerId":"42030425","CreationDate":"2017-02-03T18:07:55.550","ParentId":null,"OwnerUserId":"6668266","Title":null,"Body":"<p>Check module is present in current working directory or in site-package folder or if you have downloaded at some other location include that path in path variable<\/p>\n"}]}
{"QuestionId":42031158,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-03T18:53:25.807","AcceptedAnswerId":null,"OwnerUserId":7512933.0,"Title":"OutOfRangeError on FIFOQueues of different experiments","Body":"<p>I am running some experiments with Tensorflow on a 4-GPU machine using CIFAR10 with the FIFOQueue implementation for input data pipeline.<\/p>\n\n<p>I want to simultaneously run one different instance of my experiment on each of the four GPUs,\nbut because the data processing queues are always places on the CPU I get an OutOfRange error and\ncannot run more than one instance per machine.<\/p>\n\n<p>Is there a way to overcome this? (other than placing the data processing on GPUs)<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":42031173,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><cudnn>","CreationDate":"2017-02-03T18:54:37.063","AcceptedAnswerId":null,"OwnerUserId":7511763.0,"Title":"Tensorflow conv2d result discrepancy on cpu vs gpu","Body":"<p>We've been working on a problem here that has yielded some discrepancies between results based on what we've observed to be caused by a couple of factors which should theoretically have no impact.  <\/p>\n\n<p>In our problem, we're dealing with a data set that is varying in size.  During training, the input is fixed to a specific size by either padding or chopping.  However, in the interest of performance and speed, we were hoping to take advantage of the fact that we should be able to pre-calculate the results of feeding an input consisting of only padding through the layers of our convolutional neural net.  Thus for inputs significantly smaller than the fixed size set in training, we would determine the bare minimum amount of padding required to have the activations generated by the final conv layer match those of the same input if it were padded out to that large fixed size, in addition to as much of the pre-calculated all-pad activations that would need to be appended to the result.<\/p>\n\n<p>The conv-net portion of our model that we were experimenting with this on consists of 9 layers, the first eight performing a conv2d followed by a maxpool, and the final one being a single conv2d.  The training for the model was performed on gpu's, however, for testing we only have the cpu available.  Testing on a large set of data using the variable padding approach yielded for the most part identical results to what we would expect to see compared to the fixed size input.  However, there were a few cases where we observed discrepancies between the values obtained using the two approaches, sometimes on the order of e-4.<\/p>\n\n<p>This issue looked to manifest itself only in input which were on the smaller end of the scale.  Digging down into these cases revealed that the activations from each layer in the model were always identical, and it wasn't until you reach the final convolution layer where the discrepancies appeared.  Surprisingly though, we discovered that adding enough padding (still far below the fixed size) could affect the value of the first few activations at the final layer, even though those cells don't depend on that extra padding at all.<\/p>\n\n<p>We were able to replicate the problem using canned data, and even demonstrate it on a single conv2d operation.  In the code below, A simple conv2d with a single length:5 stride:1 filter is applied to a tensor of length 7 to produce 3 activations.  However, when you append 1 or 2 extra values on to the input, the values of the 2nd and 3rd activations change even though they have no dependence on these new input values at all.  Replicating the results by hand using tf.mul and reduce_sum ops, produces the same results as the input with the extra values, indicating that it's the correct result.  This is only observed when the ops are executed on the cpu, however, assigning them to the gpu instead always produces the same values regardless of whether the input is of length 7, 8, or 9.  Is this expected when performing conv on the cpu?  Is there a way to explain the variance in the conv output, eg. does it happen due to the algorithm used for small inputs?  All of the above refers to conv2d with VALID padding used.<\/p>\n\n<p>example:<\/p>\n\n<pre><\/pre>\n\n<p>sample output:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42031974,"AnswerCount":2,"Tags":"<numpy><tensorflow>","CreationDate":"2017-02-03T19:49:27.807","AcceptedAnswerId":null,"OwnerUserId":3678911.0,"Title":"Weighted sum of matrices in Tensorflow","Body":"<p>I have a 3-dimensional tensor A of size (M,N,N) and an 1 dimensional tensor p of size M. I want to compute the weighted sum of matrices:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9v5FU.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9v5FU.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>In NumPy I am implementing the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I want to do the same in TensorFlow but I don't seem to find any inbuilt operations to carry out the same. I tried  and  but they don't seem to be giving the desired result. Can someone suggest me a right way to do this in TensorFlow?<\/p>\n","answers":[{"AnswerId":"53994124","CreationDate":"2019-01-01T08:44:02.783","ParentId":null,"OwnerUserId":"7389608","Title":null,"Body":"<p>If you want to compute a tensor B of size <code>(K,N,N)<\/code> when you have a P matrix of size <code>(K,M)<\/code> and a tensor A of size <code>(M,N,N)<\/code>, you can follow it. <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nK = 2\nM = 3\nN = 2\n\nnp.random.seed(0)\nA = tf.constant(np.random.randint(1,5,(M,N,N)),dtype=tf.float64)\n\n# when K.shape=(K,M)\nP = tf.constant(np.random.randint(1,5,(K,M)),dtype=tf.float64)\n# when K.shape=(M,)\n# P = tf.constant(np.random.randint(1,5,(M,)),dtype=tf.float64)\n\nP_new = tf.expand_dims(tf.expand_dims(P,-1),-1)\n\n# if K.shape=(K,M) set axis=1,if K.shape=(M,) set axis=0,\nB = tf.reduce_sum(tf.multiply(P_new , A),axis=1)\n\nwith tf.Session()as sess:\n    print(sess.run(P))\n    print(sess.run(A))\n    print(sess.run(B))\n\n[[1. 4. 3.]\n [1. 1. 1.]]\n[[[1. 4.]\n  [2. 1.]]\n\n [[4. 4.]\n  [4. 4.]]\n\n [[2. 4.]\n  [2. 3.]]]\n[[[23. 32.]\n  [24. 26.]]\n\n [[ 7. 12.]\n  [ 8.  8.]]]\n<\/code><\/pre>\n\n<p>The above code can be modified to include the solution of problem in your question.<\/p>\n"},{"AnswerId":"42034514","CreationDate":"2017-02-03T23:14:53.210","ParentId":null,"OwnerUserId":"2283175","Title":null,"Body":"<p>well it seems it's quite easy to do with explicit broadcasting. But I'm not sure how efficient is this!<\/p>\n\n<pre><code>import numpy as np\n\nA = np.array([[[1, 7],\n              [4, 4]],\n             [[2, 6],\n              [5, 3]],\n             [[3, 5],\n              [6, 2]]])\n\np = np.array([4,3,2])\n\nM = 3\nN = 2\n\n#numpy version\ntemp=np.array([p[m]*A[m] for m in range(M)])\nB=sum(temp);\n\n#tensorflow version\nimport tensorflow as tf\nA_tf = tf.constant(A,dtype=tf.float64)\np_tf = tf.constant(p,dtype=tf.float64)\np_tf_broadcasted = tf.tile(tf.reshape(p_tf,[M,1,1]), tf.pack([1,N,N]))\nB_tf = tf.reduce_sum(tf.mul(A_tf,p_tf_broadcasted), axis=0)\n\nsess=tf.Session()\nB_tf_ = sess.run(B_tf)\n\n\n#make sure they're equal\nB_tf_ == B\n\n\n#array([[ True,  True],\n#       [ True,  True]], dtype=bool)\n<\/code><\/pre>\n"}]}
{"QuestionId":42032250,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-02-03T20:09:11.067","AcceptedAnswerId":42032773.0,"OwnerUserId":5896419.0,"Title":"Keras 3D Convolution: Error when checking model input: expected covolution3d_input_1 to have 5 dimensions, but got array shape (1, 90, 100, 100)","Body":"<p>The goal of the system is to classify video input by which word was pronounced. Each sample is a set of 90, 100x100, grayscale (1-colour channel frame, with a dimension of . Previously, training data was loaded directly onto memory and trained, which worked, but was not efficient, and would have been impossible later on with more training samples. To work around this, the system was modified to preprocess and save training data into an  file, then fit the training data into the model with a generator for on demand loading. However, the following error is now generated as a result of this modification:<\/p>\n\n<blockquote>\n  <p>Exception: Error when checking model input: expected\n  convolution3d_input_1 to have 5 dimensions, but got array with shape\n  (1, 90, 100, 100)<\/p>\n<\/blockquote>\n\n<p>Here is the code for the system:<\/p>\n\n<pre><\/pre>\n\n<p>What confuses me is that the reported input dimensions are the expected input dimensions, but it is complaining about a missing 5th dimension. Should the generator yield a batch of samples instead of a single sample for every iteration to generate a 5-dimensional output?<\/p>\n","answers":[{"AnswerId":"42032773","CreationDate":"2017-02-03T20:47:34.513","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>In case when you are returning a simple example you need to make sure that your output is 5-dimensional with shapes : <code>(batch_size, channels, frames, height, width)<\/code>. This is simply because the dimensionality of every layer shoud be fixed. The simplest way to make this working is:<\/p>\n\n<pre><code>X = training_save_file[\"X\"][[idx]]\n<\/code><\/pre>\n\n<p>With this fix your output should match the expected shape.<\/p>\n"}]}
{"QuestionId":42032429,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><tflearn>","CreationDate":"2017-02-03T20:21:41.457","AcceptedAnswerId":42034849.0,"OwnerUserId":7091463.0,"Title":"python tflearn - ValueError: Cannot feed value of shape (10, 250, 250, 3) for Tensor 'TargetsData\/Y:0', which has shape '(?, 2)'","Body":"<p>I am trying to create a model to recognize faces. But I keep running into this error and none of the other answers to similar questions have solved this particular problem. The code is the following:<\/p>\n\n<pre><\/pre>\n\n<p>I keep getting <\/p>\n\n<blockquote>\n  <p>ValueError: Cannot feed value of shape (10, 250, 250, 3) for Tensor 'TargetsData\/Y:0', which has shape '(?, 2)'. <\/p>\n<\/blockquote>\n\n<p>I have tried everything at this point, and can't exactly understand how to go about solving the problem.<\/p>\n","answers":[{"AnswerId":"42034849","CreationDate":"2017-02-03T23:53:20.880","ParentId":null,"OwnerUserId":"3928385","Title":null,"Body":"<p>Your input is of shape <code>(?, 250, 250, 3)<\/code> (based on the comment at the beginning and on the fact that you use Convolution layers early on), your output is of shape <code>(?, 2)<\/code> (based on the fast that your last layer is a fully connected layer with 2 output neurons). Yet, you feed the same dataset to both:<\/p>\n\n<pre><code>X = pickle.load(open('dataset.pkl', 'rb')).astype('float32')\nY = pickle.load(open('dataset.pkl', 'rb')).astype('float32')\n<\/code><\/pre>\n\n<p>^^ Note that you load the same file for both <code>X<\/code> and <code>Y<\/code>.<\/p>\n\n<p>Since I don't know what you try to achieve, there are two possible solutions:<\/p>\n\n<ol>\n<li><p>If you are trying to build an auto encoder of some sort (in which case feeding the same data set to both input and output would make sense), you need to change the architecture of your network, the convolution layers should feed into deconvolution layers. How to do it is beyond the scope of a single Stack Overflow answer<\/p><\/li>\n<li><p>If you are trying to build some sort of a classifier, then you are not reading the right file for <code>Y<\/code>. <code>Y<\/code> should contain the labels you are trying to predict, not the images.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42032517,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-03T20:29:00.103","AcceptedAnswerId":42079312.0,"OwnerUserId":4474386.0,"Title":"How to omit zeros in a 4-D tensor in tensorflow?","Body":"<p>Say I have a tensor:<\/p>\n\n<pre><\/pre>\n\n<p>I want to omit zeros and be left with a tensor of shape: (1, 2, 2, 4), with 4 being the number of non zero elements in my tensor like<\/p>\n\n<pre><\/pre>\n\n<p>I've used boolean mask to to do this on a 1-D tensor. How can I omit the zeros in a 4-D tensor. Can it be generalized for higher ranks?<\/p>\n","answers":[{"AnswerId":"42079312","CreationDate":"2017-02-06T23:44:38.400","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>Using TensorFlow 1.12:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef batch_of_vectors_nonzero_entries(batch_of_vectors):\n  \"\"\"Removes non-zero entries from batched vectors.\n\n  Requires that each vector have the same number of non-zero entries.\n\n  Args:\n    batch_of_vectors: A Tensor with length-N vectors, having shape [..., N].\n  Returns:\n    A Tensor with shape [..., M] where M is the number of non-zero entries in\n    each vector.\n  \"\"\"\n  nonzero_indices = tf.where(tf.not_equal(\n      batch_of_vectors, tf.zeros_like(batch_of_vectors)))\n  # gather_nd gives us a vector containing the non-zero entries of the\n  # original Tensor\n  nonzero_values = tf.gather_nd(batch_of_vectors, nonzero_indices)\n  # Next, reshape so that all but the last dimension is the same as the input\n  # Tensor. Note that this will fail unless each vector has the same number of\n  # non-zero values.\n  reshaped_nonzero_values = tf.reshape(\n      nonzero_values,\n      tf.concat([tf.shape(batch_of_vectors)[:-1], [-1]], axis=0))\n  return reshaped_nonzero_values\n\nt = tf.Variable(\n    [[[[0., 235., 0., 0., 1006., 0., 0., 23., 42.],\n       [77., 0., 0., 12., 0., 0., 33., 55., 0.]],\n      [[0., 132., 0., 0., 234., 0., 1., 24., 0.],\n       [43., 0., 0., 124., 0., 0., 0., 52., 645]]]])\nnonzero_t = batch_of_vectors_nonzero_entries(t)\n\nwith tf.Session():\n    tf.global_variables_initializer().run()\n    result_evaled = nonzero_t.eval()\n    print(result_evaled.shape, result_evaled)\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>(1, 2, 2, 4) [[[[  2.35000000e+02   1.00600000e+03   2.30000000e+01   4.20000000e+01]\n   [  7.70000000e+01   1.20000000e+01   3.30000000e+01   5.50000000e+01]]\n\n  [[  1.32000000e+02   2.34000000e+02   1.00000000e+00   2.40000000e+01]\n   [  4.30000000e+01   1.24000000e+02   5.20000000e+01   6.45000000e+02]]]]\n<\/code><\/pre>\n\n<p>It may be useful to look into <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/sparse_ops\/\" rel=\"nofollow noreferrer\">SparseTensor<\/a>s if the result ever ends up being ragged.<\/p>\n"}]}
{"QuestionId":42032595,"AnswerCount":2,"Tags":"<python><tensorflow><sparse-matrix><backpropagation>","CreationDate":"2017-02-03T20:34:18.560","AcceptedAnswerId":null,"OwnerUserId":821806.0,"Title":"Back-propagating gradients through a sparse tensor?","Body":"<p>I have a normal feed-forward network that produces a vector v. The elements of v are then used as the non-zero entries of a sparse matrix M (assume the coordinates are predefined). The sparse matrix is then multiplied by a dense vector and a loss is defined on the resulting scalar. I want to back-propagate the loss w.r.t. the weights of the network, which entails going through the sparse matrix. <\/p>\n\n<p>This seems like a perfectly reasonable use-case for a sparse matrix, but it appears that such functionality is not supported. Indeed, even calling tf.gradients(M,[v]) produces an error:<\/p>\n\n<blockquote>\n  <p>AttributeError: 'SparseTensor' object has no attribute 'value_index'<\/p>\n<\/blockquote>\n\n<p>Am I doing something wrong or am I correct in presuming that this functionality doesn't (yet?) exist? If the latter, then is there a work-around for this particular use-case short of rewriting all of the sparse tensor operations with gradients defined? <\/p>\n","answers":[{"AnswerId":"42034637","CreationDate":"2017-02-03T23:29:32.813","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>I'm fishing around in the dark here, working from code and documentation, not experience. <\/p>\n\n<p>The <code>Tensor<\/code> class creator is:<\/p>\n\n<pre><code>def __init__(self, op, value_index, dtype):\n    #  value_index: An `int`. Index of the operation's endpoint that produces this tensor.\n<\/code><\/pre>\n\n<p>The <code>value_index<\/code> is used to generate the <code>Tensor<\/code> name.<\/p>\n\n<p>The <code>SparseTensor<\/code> one is<\/p>\n\n<pre><code>def __init__(self, indices, values, dense_shape):\n<\/code><\/pre>\n\n<p>Nowhere in it's definition file <code>tensorflow\/tensorflow\/python\/framework\/sparse_tensor.py<\/code> is <code>value_index<\/code> referenced.<\/p>\n\n<p>Its arguments are Tensors, presumably each with its own <code>value_index<\/code>.  <\/p>\n\n<p>Else where it appears that a <code>SparseTensor<\/code> is an alternative to a <code>IndexedSlices<\/code>, which also contains tensors.<\/p>\n\n<p>The inputs to <code>tf.gradients<\/code> are all<\/p>\n\n<pre><code>A `Tensor` or list of tensors \n<\/code><\/pre>\n\n<p>The <code>gradients<\/code> definition file has a <code>_IndexedSlicesToTensor<\/code> method, but nothing equivalent for <code>SparseTensor<\/code>.  So there seems to be some sort of automatic conversion to dense in the case of <code>IndexedSlices<\/code> (with a warning if the result is too big), but not for <code>SparseTensors<\/code>.  I don't know if that'a case of incomplete development, or an incompatibility that makes it impossible.<\/p>\n"},{"AnswerId":"42102604","CreationDate":"2017-02-08T00:34:21.970","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>A slight variation on this does work, taking the gradient of the <code>values<\/code> of a <code>SparseTensor<\/code> directly:<\/p>\n\n<pre><code>import tensorflow as tf\nsparse_values = tf.identity(tf.Variable(tf.constant([1., 2., 3.])))\nsparse_indices = tf.constant([[0, 0], [1, 1], [2, 2]], dtype=tf.int64)\nsparse_matrix = tf.SparseTensor(sparse_indices, sparse_values, [3, 3])\nmultiplied = tf.sparse_tensor_dense_matmul(sparse_matrix, tf.eye(3))\nloss = tf.reduce_sum(multiplied)\ngradients = tf.gradients(loss, [sparse_values])\nwith tf.Session() as session:\n    tf.global_variables_initializer().run()\n    print(session.run(gradients))\n<\/code><\/pre>\n\n<p>Prints (on TensorFlow 0.12.1):<\/p>\n\n<pre><code>[array([ 1.,  1.,  1.], dtype=float32)]\n<\/code><\/pre>\n\n<p>Why the <code>tf.identity<\/code> op is necessary for the gradient to be defined I haven't quite figured out (probably something to do with ref dtypes). <\/p>\n"}]}
{"QuestionId":42035150,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2017-02-04T00:31:08.990","AcceptedAnswerId":null,"OwnerUserId":5194256.0,"Title":"image classification of vehicles using caffe","Body":"<p>Hi guys I am new in deep learning and caffe as well but i am trying to build an architecture for vehicle classification into Cars,Van,Bus and Truck using Caffe.\n        I used an already trained models but the loss value was constant as well as the accuracy.I therefore decided to start from scratch starting with 2 convolutional layers 2 pooling layers and 2 fully connected layers but I still have the same problem of constant loss and constant accuracy.\n        I have fine-tuned the model changing most values but to no avail. My data set in the text file are shuffled as well hence I am confused where my mistakes are.I am therefore asking if anyone can be of help\n        Below is my my_train_val.prototxt and solver file<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42035400,"AnswerCount":0,"Tags":"<session><parallel-processing><tensorflow><distributed><reinforcement-learning>","CreationDate":"2017-02-04T01:07:14.293","AcceptedAnswerId":null,"OwnerUserId":3399275.0,"Title":"Parallel processes in distributed tensorflow","Body":"<p>I have neural network in tensorflow with trained parameters, it is \"policy\" for agent. Network is being updated in training loop in main tensorflow session in core program. <\/p>\n\n<p>In the end of each training cycle I need to pass this network to few parallel processes (\"workers\"), which will use it for collecting samples from interactions of agent's policy with environment. <\/p>\n\n<p>I need to do it in parallel, because simulating environment takes most of the time and runs only single-core. So, few parallel sampling processes are needed. I am struggling how to structure this in distributed tensorflow. What I considered so far:<\/p>\n\n<ol>\n<li>Create main session in core program, where global network is updated.\nSpawn processes with python multiprocessing and pass them network's global parameters (can I just pass network as an argument?). Then in each process separate Session is created, where network runs.<\/li>\n<\/ol>\n\n<p>Minimal example for this is here (also inserted code below): \n<a href=\"https:\/\/gist.github.com\/dd210\/e1808efcc4362cab949ad0337ba600a9\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/dd210\/e1808efcc4362cab949ad0337ba600a9<\/a><\/p>\n\n<p>Problem with this example that it sometimes hangs on sess.run in 2nd process. And sometimes it runs smoothly (!). So, there got to be some fundamental issue with this approach..\nIn my real code 2nd agent just always hangs on sess.run.<\/p>\n\n<ol start=\"2\">\n<li>Somehow use between-graph replication and Supervisor (for managing sessions) in order to create one master session (in core program) and use replicas of global network on the workers.\nSeems more accurate, but here I don't know how to structure it. Example of code from official tutorial is quite different. <\/li>\n<\/ol>\n\n<p>Would be grateful for advice.<\/p>\n\n<p>Code to p.1: <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42035506,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-04T01:23:53.587","AcceptedAnswerId":null,"OwnerUserId":223963.0,"Title":"How To Use Loops In TensorFlow Scopes and Sessions?","Body":"<p>How can the following error be resolved?<\/p>\n\n<blockquote>\n  <p>ValueError: Variable\n  RNN\/MultiRNNCell\/Cell0\/BasicLSTMCell\/Linear\/Matrix does not exist, or\n  was not created with tf.get_variable(). Did you mean to set reuse=None\n  in VarScope?<\/p>\n<\/blockquote>\n\n<p>This error is thrown when I loop over \"training\" data to provide batched data to a MultiRNNCell based model (see <strong>Main<\/strong>). The  (MultiRNNCell) and its constituent parts, i.e. BasicLSTMCell(s) in the code (see <strong>Model<\/strong>) below are not created via  and thus repeated calls to the model cause the error.<\/p>\n\n<p><strong>Main<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Model<\/strong>:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42036187,"AnswerCount":0,"Tags":"<theano><keras><theano-cuda><cudnn>","CreationDate":"2017-02-04T03:35:40.303","AcceptedAnswerId":null,"OwnerUserId":6343425.0,"Title":"nvcc report compile error when train model using Keras","Body":"<p>1.Problem: I designed a simple model using Keras ( backend: Theano) to train on stock data set. When configure Theano to use cpu, no problem appears. However, when using gpu, there will appear the problem as shown below.        <\/p>\n\n<pre><\/pre>\n\n<p>2.Hardware: Titan X Maxwell, Driver Version 367.48, cuda V8.0.44. cuDNN 5. I noticed that I was warned my cuDNN version is too recent than Theano when installing it. <\/p>\n\n<p>3.This is the source code:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42037432,"AnswerCount":0,"Tags":"<java><android><tensorflow>","CreationDate":"2017-02-04T06:46:35.613","AcceptedAnswerId":null,"OwnerUserId":2190503.0,"Title":"Modify tensorflow android's DetectorActivity.java file to use retrain model, but the app crash","Body":"<p>I follow the doc of <a href=\"https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/image_retraining\/<\/a> retrain the model, then put the output_graph.pb and output_labels.txt files to android assets directory. <\/p>\n\n<p>I modify this ClassifierActivity.java and it works well. But I modify this DetectorActivity.java <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/19d932f4cb49182c111d646267ac6c0bbc4f2f00\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/DetectorActivity.java#L55-L62\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/19d932f4cb49182c111d646267ac6c0bbc4f2f00\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/DetectorActivity.java#L55-L62<\/a> , it will crash on starting  app.<\/p>\n\n<p>The file diff is this:<\/p>\n\n<pre class=\"lang-java prettyprint-override\"><\/pre>\n\n<p>this is crash logs:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42039116,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-04T10:12:49.937","AcceptedAnswerId":null,"OwnerUserId":6102784.0,"Title":"Is FIFOQueue supported in TensorFlow for iOS?","Body":"<p>Is FIFOQueue included with TensorFlow for iOS?<\/p>\n\n<p>I trained a model, froze it, and optimized for inference, yet I'm getting an error message about FIFOQueue on iOS<\/p>\n\n<p>\"Could not create TensorFlow Graph: Invalid argument: No OpKernel was registered to support Op 'FIFOQueue' with these attrs. Registered devices: [CPU], Registered kernels: \n[[Node: fifo_queue = FIFOQueuecapacity=30, component_types=[DT_FLOAT, DT_FLOAT, DT_FLOAT], container=\"\", shapes=[[640,480,3], [300,1,2], [300,1,4]], shared_name=\"\"]]\"<\/p>\n\n<p>I even tried setting is_training=False in tensorflow\/contrib\/slim\/python\/slim\/nets\/inception_v1.py before training and getting my input_graph.pb . (I tried the checkpoint created with is_training=True)<\/p>\n\n<p>Is FIFOQueue not supported in TensorFlow for iOS?<\/p>\n","answers":[{"AnswerId":"42189330","CreationDate":"2017-02-12T15:19:31.697","ParentId":null,"OwnerUserId":"6102784","Title":null,"Body":"<p>I added \"tensorflow\/core\/kernels\/fifo_queue_op.cc\" to tensorflow\/contrib\/makefile\/tf_op_files.txt then ran compile_ios_tensorflow.sh, and the errors went away.  <\/p>\n"}]}
{"QuestionId":42039548,"AnswerCount":1,"Tags":"<neural-network><deep-learning><keras><keras-layer>","CreationDate":"2017-02-04T10:58:24.083","AcceptedAnswerId":null,"OwnerUserId":7452543.0,"Title":"How to check the weights after every epoc in Keras model","Body":"<p>I am using the sequential model in Keras. I would like to check the weight of the model after every epoch. Could you please guide me on how to do so. <\/p>\n\n<pre><\/pre>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42076588","CreationDate":"2017-02-06T20:21:28.200","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>What you are looking for is a <code>CallBack<\/code> function. A callback is a Keras function which is called repetitively during the training at key points. It can be after a batch, an epoch or the whole training. See <a href=\"https:\/\/keras.io\/callbacks\/#usage-of-callbacks\" rel=\"noreferrer\">here<\/a> for doc and the list of callbacks existing.<\/p>\n\n<p>What you want is a custom CallBack that can be created with a LambdaCallBack object.<\/p>\n\n<pre><code>from keras.callbacks import LambdaCallback\n\nmodel = Sequential()\nmodel.add(Embedding(max_features, 128, dropout=0.2))\nmodel.add(LSTM(128, dropout_W=0.2, dropout_U=0.2))  \nmodel.add(Dense(1))\nmodel.add(Activation('sigmoid'))\n\nprint_weights = LambdaCallback(on_epoch_end=lambda batch, logs: print(model.layers[0].get_weights()))\n\nmodel.compile(loss='binary_crossentropy',optimizer='adam',metrics['accuracy'])\nmodel.fit(X_train, \n          y_train, \n          batch_size=batch_size, \n          nb_epoch=5 validation_data=(X_test, y_test), \n          callbacks = [print_weights])\n<\/code><\/pre>\n\n<p>the code above should print your embedding weights <code>model.layers[0].get_weights()<\/code> at the end of every epoch. Up to you to print it where you want to make it readable, to dump it into a pickle file,...<\/p>\n\n<p>Hope this helps<\/p>\n"}]}
{"QuestionId":42039583,"AnswerCount":1,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2017-02-04T11:03:16.653","AcceptedAnswerId":null,"OwnerUserId":7501011.0,"Title":"Getting way worse results after switching from bidirectional_rnn to bidirectional_dynamic_rnn","Body":"<p>basically tried to substitute bidirectional_rnn with bidirectional_dynamic_rnn (reshaped inputs), and got way worse results on a classification task.\nAm I doing something wrong? Reshaping?<\/p>\n\n<p>bidirectional_rnn version (code excerpt): <\/p>\n\n<pre><\/pre>\n\n<p>Classification Accuracy: 95%<\/p>\n\n<p>bidirectional_dynamic_rnn version (code excerpt): <\/p>\n\n<pre><\/pre>\n\n<p>Classification Accuracy: 70%<\/p>\n","answers":[{"AnswerId":"42040292","CreationDate":"2017-02-04T12:21:26.463","ParentId":null,"OwnerUserId":"7501011","Title":null,"Body":"<p>Okay, so I found that tf.reshape is not suitable for this task, I should have used tf.stack and tf.transpose instead.\nSo basically it was using messed up inputs, and was no longer able to learn.<\/p>\n\n<p>Wrong:<\/p>\n\n<pre><code>birnn_inputs = tf.reshape(encoder_embedded_inputs, [-1,enc_size,emb_size])\n<\/code><\/pre>\n\n<p>Right:<\/p>\n\n<pre><code>birnn_inputs = tf.stack(encoder_embedded_inputs)\nbirnn_inputs = tf.transpose(birnn_inputs, [1,0,2])\n<\/code><\/pre>\n\n<p>So now it works well.<\/p>\n"}]}
{"QuestionId":42040402,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-04T12:32:51.947","AcceptedAnswerId":null,"OwnerUserId":5688175.0,"Title":"LSTM example with MNIST","Body":"<p>I am currently trying to understand the meaning of  and  of the  function in :<\/p>\n\n<pre><\/pre>\n\n<p>through the <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/recurrent_network.py\" rel=\"nofollow noreferrer\">LSTM MNIST tutorial<\/a>.<\/p>\n\n<p>Indeed, with respect to the following tutorial, <a href=\"http:\/\/colah.github.io\/posts\/2015-08-Understanding-LSTMs\/\" rel=\"nofollow noreferrer\">Understanding LSTM<\/a>, I am wondering what correspond to these variables.<\/p>\n\n<p>In my opinion,  correspond to the hidden state (denoted as  in the previous link) but I am not sure.<\/p>\n\n<p>Thus, I understand that  is a list of  tensors of shape . But why does  is a a list of  tensors of shape . Is it just the cell state for the last time step?<\/p>\n","answers":[{"AnswerId":"44119978","CreationDate":"2017-05-22T18:42:23.470","ParentId":null,"OwnerUserId":"4950320","Title":null,"Body":"<p>This seems like an old question.  Sorry if you figured it out already.<\/p>\n\n<p>The <code>outputs<\/code> are the values coming out of the \"top\" of your LSTM units in the diagram.  That is the output of your LSTM layer in total.<\/p>\n\n<p>As you suggest in your question, the hidden states are actually the temporal hidden states that move from one time period to the next.  They are the <code>C_t<\/code> and <code>h_t<\/code> values that move along the time axis as output\/input to the unrolled LSTM layer. That is why you get <code>2<\/code> tensors each with size <code>(batch_size,n_hidden)<\/code><\/p>\n"}]}
{"QuestionId":42042338,"AnswerCount":1,"Tags":"<python><padding><keras>","CreationDate":"2017-02-04T15:52:56.410","AcceptedAnswerId":null,"OwnerUserId":996366.0,"Title":"What keras pad_sequence do?","Body":"<p>I am trying to code a text  in keras for that purpose I have to first convert text to sequence and then pad the sequence. But I am having a lot of trouble in implementing this step itself. I believe its because of my skewed or lack of understanding of this function () itself.  I tried to google it and didnt find any good tutorial and their is not much explained in the keras <a href=\"https:\/\/keras.io\/preprocessing\/sequence\/\" rel=\"nofollow noreferrer\">docs<\/a> also.<\/p>\n\n<p>Can any one tell me how and what is  ? Why we should pad the sequence (here character level) before feeding to.<\/p>\n\n<p>Please consider this text as an example?<\/p>\n\n<blockquote>\n  <p>Take the 50-year-old man diagnosed with prostate cancer in my clinic\n  at Brigham and Women's Hospital in Boston. He received a novel\n  procedure to remove his prostate, and later received focused radiation\n  to try to eradicate any remaining cancer. Unfortunately, his disease\n  returned a year later. But after two new therapies, his cancer now\n  appears in check. And if his cancer does spread, a host of other\n  treatments \u2014 including many not even on the market yet \u2014 may put his\n  cancer back in remission.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42072582","CreationDate":"2017-02-06T16:26:18.747","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>The way we train RNN's is to feed them a serie of sequences.<\/p>\n\n<p>The RNN's have well known issues with backpropagation of the gradient (see <a href=\"http:\/\/www.dsi.unifi.it\/~paolo\/ps\/tnn-94-gradient.pdf\" rel=\"noreferrer\">Bengio &amp; al<\/a>). This is the reason why we usually feed limited sequences to the RNN to train it. \nSo in your example, you should cut the text into smaller pieces (sentences?) in order to build your training set.<\/p>\n\n<p>For the sake of simplicity of implementation, keras only accepts sequences of the same length in a batch (<a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/40\" rel=\"noreferrer\">Recurrent Models with sequences of mixed length<\/a>). So if your sequences don't have the same length, this is where the <code>pad_sequence<\/code> is useful. <\/p>\n\n<p><code>pad_sequence<\/code> takes a LIST of sequences as an input (list of list) and will return a list of padded sequences. <\/p>\n\n<p>To get your example to work, you will have to somehow cut the text into sequences of chars. \nTo do that you can pick the separator of your choice ('.' ? ) and then pad all the sentences to the same length. Or, smarter in my opinion, consider the text as a sequence of char (even spaces and \\n), cut every n char and then feed this list of sequences as training data.\nThis will avoid you to use padding except for the last sequence (if the number of char in your data is not a multiple of your sequence length n).<\/p>\n\n<p>Of course, don't forget to tokenize your characters and embed them in a vector space before feeding them into the RNN. The RNN won't work on categorical data.<\/p>\n"}]}
{"QuestionId":42042360,"AnswerCount":1,"Tags":"<python><theano><theano-cuda>","CreationDate":"2017-02-04T15:55:31.837","AcceptedAnswerId":null,"OwnerUserId":4603642.0,"Title":"ImportError: dynamic module does not define module export function (PyInit_m3d1cf20adb1014f04986e6a344a55bde)","Body":"<p>This is an error message I get when I execute :<\/p>\n\n<blockquote>\n  <p>ImportError: dynamic module does not define module export function (PyInit_m3d1cf20adb1014f04986e6a344a55bde)<\/p>\n<\/blockquote>\n\n<p>I'm using Python 3.5.2 on Windows 10.<\/p>\n\n<p>I have also asked this of the authors of Theano on GitHub. I'm asking here hoping somebody will be able to help me even if he\/she doesn't know Theano's internals. I suspect this has more to do with Python 2 versus 3.<\/p>\n\n<p>Here's the full error:<\/p>\n\n<pre><\/pre>\n\n<p>As far as I can tell, Theano compiles a cpp file. Here are the files after the compilation:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the content of mod.cpp:<\/p>\n\n<pre><\/pre>\n\n<p>Function  seems to be present, so what's the problem?<\/p>\n\n<h1>edit 1<\/h1>\n\n<p>I copied the files to d:\\test and managed to recompile the .pyd file using the same exact command used by Theano (thanks to procmon):<\/p>\n\n<pre><\/pre>\n\n<p>I read that pyd files are just DLLs (at least on Windows). The odd thing is that the compiled pyd file doesn't seem to export any function. Maybe I don't understand how Python finds exported functions:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42047098","CreationDate":"2017-02-05T00:00:26.177","ParentId":null,"OwnerUserId":"4603642","Title":null,"Body":"<p>I've finally solved this puzzle!<\/p>\n\n<p>Let's have a look at the g++ command again:<\/p>\n\n<pre><code>\"C:\\Users\\Kiuhnm\\Anaconda3\\Library\\mingw-w64\\bin\\g++.exe\"  -shared -g -O3 -fno-math-errno\n -Wno-unused-label -Wno-unused-variable -Wno-write-strings -Wl,-rpath, -march=haswell -mmmx\n -mno-3dnow -msse -msse2 -msse3 -mssse3 -mno-sse4a -mcx16 -msahf -mmovbe -maes -mno-sha -mpclmul\n -mpopcnt -mabm -mno-lwp -mfma -mno-fma4 -mno-xop -mbmi -mbmi2 -mno-tbm -mavx -mavx2 -msse4.2\n -msse4.1 -mlzcnt -mno-rtm -mno-hle -mrdrnd -mf16c -mfsgsbase -mno-rdseed -mno-prfchw -mno-adx\n -mfxsr -mxsave -mxsaveopt -mno-avx512f -mno-avx512er -mno-avx512cd -mno-avx512pf -mno-prefetchwt1\n -mno-clflushopt -mno-xsavec -mno-xsaves -mno-avx512dq -mno-avx512bw -mno-avx512vl -mno-avx512ifma\n -mno-avx512vbmi -mno-clwb -mno-pcommit -mno-mwaitx --param l1-cache-size=32 --param\n l1-cache-line-size=64 --param l2-cache-size=8192 -mtune=haswell\n -DNPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION -m64 -DMS_WIN64\n -I\"C:\\Users\\Kiuhnm\\Anaconda3\\lib\\site-packages\\numpy\\core\\include\"\n -I\"C:\\Users\\Kiuhnm\\Anaconda3\\include\"\n -I\"C:\\Users\\Kiuhnm\\Anaconda3\\lib\\site-packages\\theano\\theano\\gof\"\n -L\"C:\\Users\\Kiuhnm\\Anaconda3\\libs\" -L\"C:\\Users\\Kiuhnm\\Anaconda3\" -o\n d:\\test\\m3d1cf20adb1014f04986e6a344a55bde.pyd d:\\test\\mod.cpp -lcudnn -lpython35\n<\/code><\/pre>\n\n<p>Do you see that <code>-Wl,-rpath,<\/code> on the second row? Well, it turns out that, at least in Windows, that swallows up everything which follows it!<\/p>\n\n<p><code>-Wl,a,b,c,d<\/code> is used to pass \"a b c d\" to the linker, but that trailing comma followed by a space is problematic in Windows. Maybe it has to do with the fact that Windows allow spaces in filenames.<\/p>\n\n<p>After removing the offending part, the output file becomes 40 KB bigger suggesting that it now includes the object code for mod.cpp. Importing the module succeeds, finally, but now I'll have to delve into Theano code to see what's wrong, since the command is dynamically generated.<\/p>\n\n<h1>EDIT: It wasn't difficult. All one has to do is add<\/h1>\n\n<pre><code>[dnn]\nlibrary_path=d:\\whatever\n<\/code><\/pre>\n\n<p>to <code>.theanorc<\/code>. I use a dummy directory because cuDNN's files are already reachable on my system (I must have updated <code>PATH<\/code>).<\/p>\n\n<p>If you don't have cuDNN installed, then you should never experience this issue.<\/p>\n\n<p>Now <code>-Wl,-rpath,<\/code> should be <code>-Wl,-rpath,d:\\whatever<\/code>, but I didn't check.<\/p>\n\n<p>Just for completeness, the relevant code is in <code>[...]Lib\\site-packages\\Theano\\theano\\gpuarray\\dnn.py<\/code> at line 267:<\/p>\n\n<pre><code>def c_compile_args(self):\n    return ['-Wl,-rpath,' + config.dnn.library_path]\n<\/code><\/pre>\n\n<p><code>config.dnn.library_path<\/code> is now \"d:\\whatever\" and all is good.<\/p>\n"}]}
{"QuestionId":42042567,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><gradient-descent>","CreationDate":"2017-02-04T16:16:46.240","AcceptedAnswerId":null,"OwnerUserId":4892874.0,"Title":"Implementing Feedback Alignment in Tensorflow","Body":"<p>I want to implement Direct Feedback Alignemnt in Tensorflow. <\/p>\n\n<p>Reference paper: <a href=\"https:\/\/arxiv.org\/pdf\/1609.01596v5.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1609.01596v5.pdf<\/a>, N\u00f8kland (2016)<\/p>\n\n<p>I implemented a simple network that does DFA in pure Python, having explicitly the backprop, I just switched the backward pass of BP with the one of DFA. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/BRanT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/BRanT.png\" alt=\"Backward passes\"><\/a><\/p>\n\n<p>where B1 and B2 are two random matrices with appropriate dimensions.<\/p>\n\n<p>Tensorflow training obviously does backprop by default. Where and what should I modify in Tensorflow code in order to get this training behavior?<\/p>\n","answers":[]}
{"QuestionId":42042695,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-04T16:28:38.557","AcceptedAnswerId":null,"OwnerUserId":3678911.0,"Title":"Efficient implementation of a tensor dot product in tensorflow","Body":"<p>I have three tensors  of size  respectively. I want to compute a matrix  of size  so that:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9X358.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9X358.gif\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>For now I am using the following implementation in tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>I want to know if there are any other effecient procedures to implement this. I tried using inbuilt matrix operations such as  and  but I am unable to find a solution using them. Any useful suggestions would be helpful. Thanks!<\/p>\n","answers":[]}
{"QuestionId":42043079,"AnswerCount":1,"Tags":"<statistics><tensorflow><keras><cross-validation><hypothesis-test>","CreationDate":"2017-02-04T17:04:08.960","AcceptedAnswerId":null,"OwnerUserId":1367204.0,"Title":"Keras: how to figure out the Null hypothesis?","Body":"<p>I am training a deep neural net using keras.  One of the scores is called val_acc.  I get like a 70% val_acc.  How do I know if this is good or bad? The neural net is a binary classifier,  so I am trying to predict a 1 or a 0. The data itself is about 65% 0's and 35% 1's. Is my 70% val_acc any good? <\/p>\n","answers":[{"AnswerId":"42043341","CreationDate":"2017-02-04T17:29:56.067","ParentId":null,"OwnerUserId":"4706171","Title":null,"Body":"<p>Accuracy is not always the right metric for the evaluation of a classifier. For example, it could be more important for you to classify the 1s more correctly than 0s (for example fraud detection) or the other way. So you may be interested to have a classifier with higher precision (specificity) or recall (sensitivity).  In other words, false positives may be more expensive for you than false negatives. If you have some idea about the costs of misclassifications (e.g. for FPs &amp; FNs) then you can precisely compute the specific threshold that will be optimal (instead of default 0.5) for 0-1 classification. You can use ROC curves and AUC to find performance of your classifier as well (the higher AUC the better). Finally you may want to consider kappa statistics to find how useful \/ effective your classifier is.<\/p>\n"}]}
{"QuestionId":42043488,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-04T17:45:13.213","AcceptedAnswerId":null,"OwnerUserId":4171672.0,"Title":"Replace nan values in tensorflow tensor","Body":"<p>I'm working on a convolutional neural network in tensorflow and I have a problem. The problem is the input image I read through tfrecords contains a certain number of nan values. The cause of this is the image represents a depthmap which has some infinite values in it, and in the process of encoding it in the tfrecord and then decoding to feed it to the net these infinite values become nan values.<\/p>\n\n<p>Now, since in my situation replacing the infinite values in the original image before encoding it in the tfrecors is not an option, there is any way I can replace the nan values in my image tensor as an operation to do before I feed it to the net?<\/p>\n","answers":[{"AnswerId":"42044859","CreationDate":"2017-02-04T19:48:51.550","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>A combination of tf.where and tf.is_nan should work:<\/p>\n\n<pre><code>import tensorflow as tf\nwith tf.Session():\n    has_nans = tf.constant([float('NaN'), 1.])\n    print(tf.where(tf.is_nan(has_nans), tf.zeros_like(has_nans), has_nans).eval())\n<\/code><\/pre>\n\n<p>Prints (using TensorFlow 0.12.1):<\/p>\n\n<pre><code>[ 0.  1.]\n<\/code><\/pre>\n"},{"AnswerId":"58791048","CreationDate":"2019-11-10T17:13:45.570","ParentId":null,"OwnerUserId":"12066144","Title":null,"Body":"<p>If someone is looking for the solution in Tensorflow 2.0, the adapted code of <a href=\"https:\/\/stackoverflow.com\/a\/42044859\/12066144\">Allen Lavoie<\/a> is :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nwith tf.compat.v1.Session():\n    has_nans = tf.constant([float('NaN'), 1.])\n    print(tf.where(tf.math.is_nan(has_nans), tf.zeros_like(has_nans), has_nans).eval())\n<\/code><\/pre>\n"}]}
{"QuestionId":42043728,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-04T18:06:16.583","AcceptedAnswerId":null,"OwnerUserId":7508550.0,"Title":"Tensorflow [Errno 2] - not creating \"\/svg\/summary-icon.svg\" so cannot visualize in TensorBoard","Body":"<p>I get \"WARNING:root:Error [Errno 2] No such file or directory: 'home\/...\/svg\/summary-icon.svg' (The svg part of the path is missing in my file system) when trying to generate a TensorFlow graph to then look at in TensorBoard. I suspect this is a known bug and wonder of rebuilding TensorFlow with Bazel is the way to go? I am not a software engineer so hope there is an easier way to get the graph to show inside TensorBoard. Any help will be gratefully welcomed.<\/p>\n","answers":[]}
{"QuestionId":42044608,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><lstm><reinforcement-learning>","CreationDate":"2017-02-04T19:23:08.227","AcceptedAnswerId":42052804.0,"OwnerUserId":5365672.0,"Title":"Does dropout improve models even with access to infinite data?","Body":"<p>Are there well understood guidelines of when to use dropout versus simply get more data? I had previously understood that, given sufficient data, one wouldn't want to use dropout. However, I recently had a model (4 layer LSTM processing audio input) that empirically converged to a certain loss no matter how much data I provided - and was then improved significantly when I added dropout.<\/p>\n\n<p>Is this phenomena well understood? Should one always use dropout then, even when there is more (potentially infinite) data available?<\/p>\n\n<p>Follow-up: if so, I haven't seen much mention of dropout in RL papers. I'd assumed this is because there is infinite generable data. Are there other nuances to consider here about the fraction of the state space explored, or the heterogeneity of the training data available etc?<\/p>\n","answers":[{"AnswerId":"42045070","CreationDate":"2017-02-04T20:10:55.307","ParentId":null,"OwnerUserId":"3507061","Title":null,"Body":"<p>There is no precise prescription for dropout - as <a href=\"https:\/\/www.cs.toronto.edu\/~hinton\/absps\/JMLRdropout.pdf\" rel=\"nofollow noreferrer\">Hinton's group 2014 paper<\/a> seems to confirm.  However they write there that:<\/p>\n\n<blockquote>\n  <p>Dropout introduces a significant amount of noise in the gradients\n  compared to standard stochastic gradient descent. Therefore, a lot of\n  gradients tend to cancel each other. In order to make up for this, a\n  dropout net should typically use 10-100 times the learning rate that\n  was optimal for a standard neural net.<\/p>\n<\/blockquote>\n\n<p>So my naive guess is a factor of 10-100 of additional training data is necessary.  Note also the using dropout can practically triple the number of training regimes needed - an additional factor to bear in mind.<\/p>\n"},{"AnswerId":"42052804","CreationDate":"2017-02-05T13:46:43.697","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Dropout reduces your train accuracy in exchange for test accuracy. However your performance on the data is bounded by the training accuracy. In a world with infinite independent distributed data it would be possible to obtain performance on the problem equal to the training set, because in essence its the same data. <\/p>\n\n<p>However in practice you will never have such a good training set. Even if you obtain infinite data it will not represent your distribution and will definitely not be independent.<\/p>\n"}]}
{"QuestionId":42045092,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><keras>","CreationDate":"2017-02-04T20:12:34.090","AcceptedAnswerId":42047044.0,"OwnerUserId":7375294.0,"Title":"Does K.function method of Keras with Tensorflow backend work with network layers?","Body":"<p>I recently have started using Keras to build neural networks. I built a simple CNN to classify MNIST dataset. Before learning the model I used  in order to plot a convolutional layer weights. Right now I am trying to visualize convolutional layer output with  method, but I keep getting error.<\/p>\n\n<p>Here is what I want to do for now:<\/p>\n\n<pre><\/pre>\n\n<p>And this is what I get:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"55515158","CreationDate":"2019-04-04T11:42:35.903","ParentId":null,"OwnerUserId":"9481613","Title":null,"Body":"<p>I think you can also use K.function to get gradients.  <\/p>\n\n<pre><code>self.action_gradients = K.gradients(Q_values, actions)\nself.get_action_gradients=K.function[*self.model.input, K.learning_phase()], outputs=action_gradients) \n<\/code><\/pre>\n\n<p>which basically runs the graph to obtain the Q-value to calculate the gradient of the Q-value w.r.t. action vector in DDPG.  Source code here (lines 64 to 70):  <a href=\"https:\/\/github.com\/nyck33\/autonomous_quadcopter\/blob\/master\/criticSolution.py#L65\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nyck33\/autonomous_quadcopter\/blob\/master\/criticSolution.py#L65<\/a><\/p>\n\n<p>In light of the accepted answer and this usage here (originally from project 5 autonomous quadcopter in the Udacity Deep Learning nanodegree), a question remains in my mind, ie. is <code>K.function()<\/code> something that can be used fairly flexibly to run the graph and to designate as outputs of <code>K.function()<\/code> for example outputs of a particular layer, gradients or even weights themselves?<\/p>\n\n<p>Lines 64 to 67 here:  <a href=\"https:\/\/github.com\/nyck33\/autonomous_quadcopter\/blob\/master\/actorSolution.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nyck33\/autonomous_quadcopter\/blob\/master\/actorSolution.py<\/a><\/p>\n\n<p>It is being used as a custom training function for the actor network in DDPG:<\/p>\n\n<pre><code>#caller\nself.actor_local.train_fn([states, action_gradients, 1])\n#called\nself.train_fn = K.function(inputs=[self.model.input, action_gradients, K.learning_phase()], \\\n            outputs=[], updates=updates_op)\n<\/code><\/pre>\n\n<p><code>outputs<\/code> is given a value of an empty list because we merely want to train the actor network with the action_gradients from the critic network.  <\/p>\n"},{"AnswerId":"42047044","CreationDate":"2017-02-04T23:52:49.283","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You should do the following changes:<\/p>\n\n<pre><code>output_fn = K.function([input_layer], [output_layer])\noutput_image = output_fn([input_image])\n<\/code><\/pre>\n\n<p><code>K.function<\/code> takes the input and output tensors as list so that you can create a function from many input to many output. In your case one input to one output.. but you need to pass them as a list none the less. <\/p>\n\n<p>Next <code>K.function<\/code> returns a tensor function and not a model object where you can use <code>predict()<\/code>. The correct way of using is just to call as a function<\/p>\n"}]}
{"QuestionId":42046018,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-04T21:50:23.617","AcceptedAnswerId":null,"OwnerUserId":2002890.0,"Title":"TensorFlow adding duplicate key when using optimizers other than GradientDescentOptimizer","Body":"<p>When I use the simple GradientDescentOptimizer, my model runs fine, however if I change it to any of the other ones (Adam, RMSProp, etc.), I get an InvalidArgumentError that looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>The problem seems to have to do with partitioning the softmax layer. When I remove the partitioner the problem goes away as well (and this is the case for both variable and fixed size partitioners). The optimizer code looks like this:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42046238,"AnswerCount":1,"Tags":"<tensorflow><normalization>","CreationDate":"2017-02-04T22:13:32.877","AcceptedAnswerId":null,"OwnerUserId":7509951.0,"Title":"Batch normaliztion on tensorflow - tf.contrib.layers.batch_norm works good on training but poor testing\/validation results","Body":"<p>I trying to implement CNN with the function tf.contrib.layers.batch_norm on Mnist dataset.<\/p>\n\n<p>When I train and check the model I see that the loss is decreasing (good!) but the accuracy of the test dataset is remain random (~10%)(BAD!!!)<\/p>\n\n<p>If I use the same model without the batch normalization I see that the test accuracy is increasing as expected.<\/p>\n\n<p>you can see in the code below how I use the batch normalization function.\nif I use for the test dataset to set the <strong>is_training=True<\/strong> I get good results so the problem is the <strong>is_training=False<\/strong> mode of the batch normalization function.<\/p>\n\n<p>Please help me with this. Thanks in advance for all.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"45748146","CreationDate":"2017-08-18T03:38:41.867","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>During training, batch-norm uses statistics based on the batch. During evaluation\/testing (whenever <code>is_training<\/code> is <code>False<\/code>), it uses population statistics.<\/p>\n\n<p>Internally, the population statistics are updated via implicitly created update ops which are added to the <code>tf.GraphKeys.UPDATE_OPS<\/code> collection - but you have to force tensorflow to run these operations. A simple way of doing this is to introduce <code>control_dependencies<\/code> on your optimization op.<\/p>\n\n<pre><code>update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)\nwith tf.control_dependencies(update_ops):\n    train_op = optimizer.minimize(loss, step)\n<\/code><\/pre>\n"}]}
{"QuestionId":42046264,"AnswerCount":1,"Tags":"<gpu><theano><theano-cuda>","CreationDate":"2017-02-04T22:16:33.963","AcceptedAnswerId":42115105.0,"OwnerUserId":3733772.0,"Title":"Theano setting in .theanorc file different beetween gpu and cuda","Body":"<p>To configure theano we create a .theanorc file in your home folder and add the following to set up theano to run on GPU.<\/p>\n\n<pre><\/pre>\n\n<p>but sometime i save in configure that smbd put device = cuda, what's the difference?\nAs i understans, if you use cuda, it should work faster because cuda drivers will manage gpu more better,not getting other jobs for gpu at calculation time<\/p>\n","answers":[{"AnswerId":"42115105","CreationDate":"2017-02-08T13:57:32.920","ParentId":null,"OwnerUserId":"1953384","Title":null,"Body":"<p>Setting <code>device=gpu<\/code> in your <code>.theanorc<\/code> file instructs Theano to use the <code>cuda<\/code> backend. On the other hand, setting <code>device=cuda<\/code> instructs theano to use the <code>libgpuarray<\/code> backend. Both lead to the use of the GPU (in contrast to setting <code>device=cpu<\/code> in your <code>.theanorc<\/code> file which would lead to the use of the CPU), and the difference is the low-level API they use to communicate with it. The <code>libgpuarray<\/code> backend is the newer one of the two that is a wrapper that allows Theano to communicate with both <code>cuda<\/code> (for NVIDIA GPUs) and <code>opencl<\/code> (for non-NVIDIA GPUs). <\/p>\n\n<p>Note that the <code>cuda<\/code> backend will be deprecated in the next release of Theano and it is recommended that you always use the <code>libgpuarray<\/code> backend henceforth.<\/p>\n"}]}
{"QuestionId":42046372,"AnswerCount":0,"Tags":"<python-3.x><machine-learning><tensorflow>","CreationDate":"2017-02-04T22:30:23.677","AcceptedAnswerId":null,"OwnerUserId":7038807.0,"Title":"TensorFlow, losses after training the model are different than losses printed during the last Epoch of Stochastic Gradient Descent.","Body":"<p>I'm trying to do binary classification on two spirals.  For testing, I am feeding my neural network the exact spiral data with no noise, and the model seems to work as the losses near 0 during SGD.  However, after using my model to infer the exact same data points after SGD has completed, I get completely different losses than what was printed during the last epoch of SGD.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42046525,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-04T22:48:21.207","AcceptedAnswerId":42055776.0,"OwnerUserId":2845263.0,"Title":"Two staged training for different dependent graphs in tensorflow","Body":"<p>I am trying to perform two tasks (A and B) which have inputs  and  and corresponding outputs , .<\/p>\n\n<p>Task A is to be first achieved by a Graph . After Task A is finished, I wish to use the weights of  into a new graph  which is a bigger graph ( a superset of ). <\/p>\n\n<p>I am unsure how to do this in tensorflow. <\/p>\n","answers":[{"AnswerId":"42055776","CreationDate":"2017-02-05T18:40:33.013","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>I am using this kind of split for training and validation purposes, where I create dedicated input and output pipelines but share the inception part of the graph, although I'm using the same <em>graph<\/em> (as of <code>tf.Graph()<\/code>), but different (unconnected) <em>subgraphs<\/em> within it.<\/p>\n\n<p>Within one <code>tf.Graph()<\/code> the general concept is <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variable_scope\/\" rel=\"nofollow noreferrer\">variable sharing<\/a> which you can achieve by using <code>tf.variable_scope()<\/code> to group your variables by concept and them creating and refetching them by using <code>tf.get_variable()<\/code> (instead of using <code>tf.Variable()<\/code> directly). The first time it's called it will create the variables, the second time it will reuse them - provided the name of the variable stays the same.<\/p>\n\n<p>However I found it much easier to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sharing_variables#make_template\" rel=\"nofollow noreferrer\"><code>tf.make_template()<\/code><\/a> instead, which will wrap a function that creates a subgraph entirely and on every call creates a new instance of the graph while sharing all of its variables.\nThe documentation example for that is<\/p>\n\n<pre><code>def my_op(x, scalar_name):\n  var1 = tf.get_variable(scalar_name,\n                         shape=[],\n                         initializer=tf.constant_initializer(1))\n  return x * var1\n\ncreate_instance = tf.make_template('scale_by_y', my_op, scalar_name='y')\n\nz = create_instance(input1)\nw = create_instance(input2)\n<\/code><\/pre>\n\n<p>Here, each call to <code>create_instance<\/code> will create a new node called <code>scale_by_y<\/code> in the graph that performs the operation defined by <code>my_op()<\/code> while sharing its internal variables. (In the example, the parameter <code>scalar_name<\/code> is statically bound to the value of <code>y<\/code>, resulting in variable <code>scale_by_y\/y<\/code> to be created (and reused) in the graph. I find that to be more confusing than helpful.) \nIt does not care about parent scopes, so<\/p>\n\n<pre><code>with tf.variable_scope('training'):\n    z1 = create_instance(input1)\n\nwith tf.variable_scope('validation'):\n    z2 = create_instance(input2)\n<\/code><\/pre>\n\n<p>works. It also might or might not work across different <code>tf.Graph()<\/code> instances, though I doubt it.<\/p>\n"}]}
{"QuestionId":42046944,"AnswerCount":1,"Tags":"<tensorflow><distributed>","CreationDate":"2017-02-04T23:40:48.930","AcceptedAnswerId":null,"OwnerUserId":3399275.0,"Title":"Between-graph replication in tensorflow: sessions and variables","Body":"<p>Question about between-graph replication in distributed Tensorflow, because I didn't get few moments from tutorials. As I understood the current model:<\/p>\n\n<p>We have parameter server which we just launch in separate process and make server.join().\nWe have workers, each of them builds the similar computational graph which contains parameter nodes linked to parameter server (through tf.train.replica_device_setter) and calculation nodes placed on workers themselves.<\/p>\n\n<p>What I didn't find:<\/p>\n\n<ol>\n<li><p>How sessions are working in this model? Because in examples\/tutorials it is hidden behind tf.train.Supervisor. \nDo we have separate sessions on each worker or just one huge session that accumulate graphs from all the workers and parameter server?<\/p><\/li>\n<li><p>How global variables are initialized on parameter server? I wonder that I can initialize them in one of the worker processes (choose it as a \"master\"), if I linked these parameters on the worker through tf.train.replica_device_setter. Is that correct?<\/p><\/li>\n<li><p>In the following gist:<\/p><\/li>\n<\/ol>\n\n<p><a href=\"https:\/\/gist.github.com\/yaroslavvb\/ea1b1bae0a75c4aae593df7eca72d9ca\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/yaroslavvb\/ea1b1bae0a75c4aae593df7eca72d9ca<\/a><\/p>\n\n<p>global variables initialized just in parameter server process and all the workers consider them initialized. How is that possible, given that they even work in different sessions? I could not replicate it in simpler example.<\/p>\n\n<ol start=\"4\">\n<li>I have main session in core program where I perform training of the model. Part of training loop is collection of data, which in turn requires calculation on the tensorflow cluster. So I need to create this cluster, put on the parameter server current state of the trained model, then collect data from calculation and continue with training loop. How can I: 1) pass current trained model to the cluster? 2) Extract collected data from cluster and pass it to main program?<\/li>\n<\/ol>\n\n<p>Thanks in advance!<\/p>\n\n<p>EDIT:<\/p>\n\n<p>To q.3: \nIt was answered previously (<a href=\"https:\/\/stackoverflow.com\/questions\/36145808\/in-tensorflow-is-variable-value-the-only-context-information-a-session-stores\/36145962#36145962\">In tensorflow, is variable value the only context information a session stores?<\/a>) that in distributed runtime variables are shared between sessions. \nDoes it mean that when I create session with some \"target\", then all the variables will be shared between those sessions that run on the same graph?<\/p>\n","answers":[{"AnswerId":"42047796","CreationDate":"2017-02-05T01:57:07.790","ParentId":null,"OwnerUserId":"3399275","Title":null,"Body":"<p>Guess I can try answering these questions by myself, at least it may be helpful for other newbies trying to harness distributed Tensorflow, because as of now there is lack of concise and clear blog posts on that topic.<\/p>\n\n<p>Hope more knowledgeable people will correct me if needed.<\/p>\n\n<ol>\n<li>We have separate sessions on all the servers, and these sessions share their resources (variables, queues, and readers) <strong>but only in distributed setting<\/strong> (i.e. you pass server.target to tf.Session constructor).<\/li>\n<\/ol>\n\n<p>Ref: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session<\/a><\/p>\n\n<ol start=\"2\">\n<li><p>Parameter variables usually are initialized in one \"master\" process. It can be process where parameter server is launched. <em>But it is not strictly necessary to do it in just one process.<\/em><\/p><\/li>\n<li><p>Because of p.1. Replicated :)<\/p><\/li>\n<li><p>Thanks to ideas from @YaroslavBulatov, I came to the following approach, which appears to be the simplest possible:<\/p>\n\n<ul>\n<li>Cluster: one local \"calculation server\" and N \"workers\".<\/li>\n<li>\"Calculation server\" keeps all the parameters of global network and performs training steps. All training ops are assigned to it.<\/li>\n<li>\"Workers\" collect data in parallel and then put it in Queue; these data are used by \"calculation server\" when doing training steps.<\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p>So, high-level algorithm:<\/p>\n\n<ol>\n<li>launch all the units in cluster <\/li>\n<li>build comp graph and training ops on calculation server <\/li>\n<li>build comp graph on workers (variables are linked to calculation\nserver).<\/li>\n<li>collect data with workers<\/li>\n<li>perform training step on calculation server and update global\nnetwork<\/li>\n<\/ol>\n\n<p>repeat 4-5 till convergence :) <\/p>\n\n<p>As of now I did coordination between calculation server and workers through Queues (when to start collection of data and when to start training step), which is definitely not the most elegant solution. Any feedback is very welcome.<\/p>\n"}]}
{"QuestionId":42047313,"AnswerCount":1,"Tags":"<tensorflow><keras><lstm>","CreationDate":"2017-02-05T00:32:27.267","AcceptedAnswerId":null,"OwnerUserId":3365859.0,"Title":"Issue with LSTMs in Keras with Tensorflow version 1.0.0-rc1","Body":"<p>I use Tensorflow as the backend in my keras. I upgraded both today, my keras version is 1.2.1, Tensorflow version is 1.0.0-rc1<\/p>\n\n<p>I used LSTMs in a simple example<\/p>\n\n<pre><\/pre>\n\n<p>It worked fine in the previous version of Tensorflow 0.9.0<\/p>\n\n<p>Now, I get an error<\/p>\n\n<blockquote>\n  <p>TypeError: Expected int32, got\n   of\n  type 'Variable' instead.<\/p>\n<\/blockquote>\n\n<p>in the LSTM line.<\/p>\n\n<p>Tensorflow released the current version a week ago. Any suggestions about this?<\/p>\n","answers":[{"AnswerId":"42544216","CreationDate":"2017-03-01T23:36:27.543","ParentId":null,"OwnerUserId":"4945455","Title":null,"Body":"<p>It seems keras 1 doesn't work with tensorflow 1.0.0<\/p>\n\n<p>See: <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/5443\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/5443<\/a><\/p>\n"}]}
{"QuestionId":42047497,"AnswerCount":2,"Tags":"<python><memory-management><out-of-memory><keras><grid-search>","CreationDate":"2017-02-05T01:05:59.833","AcceptedAnswerId":42047606.0,"OwnerUserId":7517192.0,"Title":"Keras: Out of memory when doing hyper parameter grid search","Body":"<p>I'm running multiple nested loops to do hyper parameter grid search. Each nested loop runs through a list of hyper parameter values and inside the innermost loop, a Keras sequential model is built and evaluated each time using a generator. (I'm not doing any training, I'm just randomly initializing and then evaluating the model multiple times and then retrieving the average loss).<\/p>\n\n<p>My problem is that during this process, Keras seems to be filling up my GPU memory, so that I eventually get an OOM error.<\/p>\n\n<p>Does anybody know how to solve this and free up the GPU memory each time after a model is evaluated?<\/p>\n\n<p>I do not need the model anymore at all after it has been evaluated, I can throw it away entirely every time before building a new one in the next pass of the inner loop.<\/p>\n\n<p>I'm using the Tensorflow backend.<\/p>\n\n<p>Here is the code, although much of it isn't relevant to the general problem. The model is built inside the fourth loop,<\/p>\n\n<pre><\/pre>\n\n<p>I guess the details of how the model is built don't matter much, but here is all of it anyway:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"53521666","CreationDate":"2018-11-28T14:25:36.013","ParentId":null,"OwnerUserId":"3806238","Title":null,"Body":"<p>Using the tip given by indraforyou, I added the code to clear the TensorFlow session inside the function I pass to GridSearchCV, like this:<\/p>\n\n<pre><code>def create_model():\n    # cleanup\n    K.clear_session()\n\n    inputs = Input(shape=(4096,))\n    x = Dense(2048, activation='relu')(inputs)\n    p = Dense(2, activation='sigmoid')(x)\n    model = Model(input=inputs, outputs=p)\n    model.compile(optimizer='SGD',\n              loss='mse',\n              metrics=['accuracy'])\n    return model\n<\/code><\/pre>\n\n<p>And then I can invoke the grid search: <\/p>\n\n<pre><code>model = KerasClassifier(build_fn=create_model)\ngrid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=1)\n<\/code><\/pre>\n\n<p>It should work.<\/p>\n\n<p>Cheers!<\/p>\n"},{"AnswerId":"42047606","CreationDate":"2017-02-05T01:25:01.267","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>As indicated, the backend being used is Tensorflow. With the Tensorflow backend the current model is not destroyed, so you need to clear the session.<\/p>\n\n<p>After the usage of the model just put:<\/p>\n\n<pre><code>if K.backend() == 'tensorflow':\n    K.clear_session()\n<\/code><\/pre>\n\n<p>Include the backend:<\/p>\n\n<pre><code>from keras import backend as K\n<\/code><\/pre>\n\n<p>Also you can use sklearn wrapper to do grid search. Check this example: <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/mnist_sklearn_wrapper.py\" rel=\"nofollow noreferrer\">here<\/a>. Also for more advanced hyperparameter search you can use <a href=\"https:\/\/github.com\/maxpumperla\/hyperas\" rel=\"nofollow noreferrer\">hyperas<\/a>.<\/p>\n"}]}
{"QuestionId":42048635,"AnswerCount":1,"Tags":"<python><theano><theano-cuda>","CreationDate":"2017-02-05T04:46:00.377","AcceptedAnswerId":null,"OwnerUserId":6082378.0,"Title":"Using CUDA8 in theano","Body":"<p>I have working installtion of CUDA8 and have installed theano, while importing the theano it searches for CUDA7.5 instead of CUDA8, How can tell  to use CUDA8 instead of CUDA7.5?<\/p>\n\n<p>My sytem only have CUDA8, and it doesn't contain mixed environment cuda(i.e. having both CUDA7.5 and CUDA8).<\/p>\n\n<p>Here is a output of  <\/p>\n\n<pre><\/pre>\n\n<p>Here is a output of <\/p>\n\n<pre><\/pre>\n\n<p>While importing theano in ipython it fails to run in gpu mode with error that, it can't find the <\/p>\n\n<pre><\/pre>\n\n<p>Here is a content of my <\/p>\n\n<pre><\/pre>\n\n<p>I tried to build  from source, after uninstalling the previous installtion of it, that too is not working. I did cleared the theano-cache with \/ and my manually deleting the content under  directory, which too couldn't helped.<\/p>\n\n<p>With more debugging I get error here\n<a href=\"https:\/\/github.com\/Theano\/Theano\/blob\/8b9f73365e4932f1c005a0a37b907d28985fbc5f\/theano\/gof\/cmodule.py#L302\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Theano\/Theano\/blob\/8b9f73365e4932f1c005a0a37b907d28985fbc5f\/theano\/gof\/cmodule.py#L302<\/a><\/p>\n\n<p>when  tries to load the  from  in theano cache<\/p>\n\n<p>comiplation phase for mod.cu runs without error.<\/p>\n\n<p>In this case linker is pointing to wrong <\/p>\n\n<pre><\/pre>\n\n<p>I assume  is properly caching the cuda libraries<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42152251","CreationDate":"2017-02-10T05:42:06.563","ParentId":null,"OwnerUserId":"6082378","Title":null,"Body":"<p>After more digging about the my issue, I refactored my original question and posted here <a href=\"https:\/\/stackoverflow.com\/questions\/42127432\/nvcc-is-picking-wrong-libcudart-library\">nvcc is picking wrong libcudart library<\/a> which resolved my problem.<\/p>\n"}]}
{"QuestionId":42048739,"AnswerCount":1,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2017-02-05T05:05:58.990","AcceptedAnswerId":null,"OwnerUserId":3204416.0,"Title":"Tflearn Custom Objective function","Body":"<p>I am creating a custom objective function for my tflearn model.  The objective function is complex and requires me to iterate through the predicted and correct outputs and add certain parts not based on index.  I cannot find a way to make it work with the tensor datatype.<\/p>\n\n<p>I have coded a version using standard lists below.<\/p>\n\n<pre><\/pre>\n\n<p>There does not seem to be a way to loop through the individual values of a tensor.  Should I be creating a new session in the objective function and evaluating the tensor?<\/p>\n\n<p>Thanks for the help in advance<\/p>\n","answers":[{"AnswerId":"44180260","CreationDate":"2017-05-25T12:10:23.520","ParentId":null,"OwnerUserId":"6493994","Title":null,"Body":"<p>You can add any new loss function to tflearn\/objectives.py file (<a href=\"https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/tflearn\/objectives.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/tflearn\/objectives.py<\/a>). To use it, you just need to call its name in regression layer.<\/p>\n\n<pre><code>def my_own_Loss(y_pred, y_true):\n    for value, index in enumerate(y_true):\n        if y_true[index] == 0:\n            errorBuild += y_pred[index]\n        else:\n            errorBuild += y_pred[index] - y_true[index]\n            errorCheck += math.abs(errorBuild)\n\n    return errorCheck\n<\/code><\/pre>\n\n<p>and then call it in:<\/p>\n\n<pre><code>net = tflearn.regression(net, optimizer='momentum',\n                         loss='my_own_Loss',\n                         learning_rate=0.1)\n<\/code><\/pre>\n"}]}
{"QuestionId":42049256,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-05T06:38:39.020","AcceptedAnswerId":42075126.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Does queuing examples in parallel speed up a batch creation if tf.train.batch is already dequeuing examples in parallel?","Body":"<p>In the TensorFlow-slim documentation, there is a ParallelReader object that can read TFRecords data in parallel through having multiple readers to take in example strings into queue. However, if I am not mistaken tf.train.batch dequeues examples from a queue and is able to do so in parallel with the argument . If that is the case, is it necessary that both the batch creation and reading of data must have the same speed otherwise one will cause a slower creation of a batch?<\/p>\n\n<p>I am actually not very sure whether the dequeue operation happens in tf.train.batch or when a queue_runner is manually created to dequeue the examples since I believe tf.train.batch can effectively replace the queue_runner operation. Is this correct?<\/p>\n","answers":[{"AnswerId":"42075126","CreationDate":"2017-02-06T18:53:06.563","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>As far as your first paragraph goes: the answer is that each step in your input pipeline should have enough bandwidth to feed the model (and probably not too much more). You can easily run into bottlenecks if e.g. moving from one queue to another is fast but not enough data is being fed to the first queue (because of I\/O issues, for example). You could also imagine having huge numbers of parallel readers and plenty of I\/O bandwidth but only a single thread responsible for collecting all of those examples being the bottleneck. Some back-of-the-envelope computation may help, but ultimately you should be <a href=\"https:\/\/stackoverflow.com\/questions\/40191367\/tensorflow-get-amount-of-samples-in-queue\">monitoring the number of examples in each queue<\/a>.<\/p>\n\n<p>As for the second question, <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#batch\" rel=\"nofollow noreferrer\">tf.train.batch does start its own queue runner<\/a>. The actual enqueuing\/dequeuing happens asynchronously, which is why it's important to monitor queue sizes to make sure that training isn't consuming data faster than the input pipeline is producing it.<\/p>\n"}]}
{"QuestionId":42049365,"AnswerCount":0,"Tags":"<python><arrays><numpy><tensorflow><keras>","CreationDate":"2017-02-05T06:55:27.747","AcceptedAnswerId":null,"OwnerUserId":4644648.0,"Title":"Tensorflow ValueError: Cannot feed value of shape when the shape of the numpy array is the same as input","Body":"<p>I have a numpy array with the shape (155508, 50).\nI feed it into my TF model which is made with keras, and I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>I am absolutely certain the numpy array I am feeding into my model is the correct shape of (155508, 50)!<\/p>\n\n<p>I use Keras to define my model like so:<\/p>\n\n<pre><\/pre>\n\n<p>I then train it with:<\/p>\n\n<pre><\/pre>\n\n<p>and that runs fine.<\/p>\n\n<p>The issue occurs when I run<\/p>\n\n<pre><\/pre>\n\n<p>I am 100% certain that my numpy array has the shape (155508, 50), so I am confused as to why the error says \"shape (155508, 32)\"\nI confirm it in my terminal like so:<\/p>\n\n<pre><\/pre>\n\n<p>I'm starting to tear my hair out as to what could possibly make this happen.<\/p>\n","answers":[]}
{"QuestionId":42049828,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-05T08:06:54.337","AcceptedAnswerId":42051906.0,"OwnerUserId":7045779.0,"Title":"How to concatenates tensors by both two dims in tensorflow","Body":"<p>I would like to concatenates tensors by both two dims in tensorflow.<\/p>\n\n<p>For example, there is four tensors with 4 dimensions. All tensors is something like images in tensorflow, so each dimension means the followings : [batch_size, image_width_size, image_height_size, image_channel_size].<\/p>\n\n<pre><\/pre>\n\n<p>Without consideration about batch size and channel size(I mean to say, only with consideration about image width and image height), I would like to work out the following :<\/p>\n\n<pre><\/pre>\n\n<p>Thus, 's shape should be .<\/p>\n\n<p>How should I deal with this operation?<\/p>\n","answers":[{"AnswerId":"42051906","CreationDate":"2017-02-05T12:13:09.717","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can use <code>tf.concat<\/code> to concat the tensors along the axis needed.<\/p>\n\n<p>Here:<\/p>\n\n<pre><code>import tensorflow as tf\n\nimage_tensor_1 = 1*tf.ones([60, 2, 2, 3])\nimage_tensor_2 = 2*tf.ones([60, 2, 2, 3])\nimage_tensor_3 = 3*tf.ones([60, 2, 2, 3])\nimage_tensor_4 = 4*tf.ones([60, 2, 2, 3])\n\ntry:\n    temp_1 = tf.concat_v2([image_tensor_1, image_tensor_2], 2)\n    temp_2 = tf.concat_v2([image_tensor_3, image_tensor_4], 2)\n    result = tf.concat_v2([temp_1, temp_2], 1)\nexcept AttributeError:\n    temp_1 = tf.concat(2, [image_tensor_1, image_tensor_2])\n    temp_2 = tf.concat(2, [image_tensor_3, image_tensor_4])\n    result = tf.concat(1, [temp_1, temp_2])\n\n\nsess = tf.Session()\nprint sess.run([result[0,:,:,0]])\n<\/code><\/pre>\n"},{"AnswerId":"42051919","CreationDate":"2017-02-05T12:14:11.683","ParentId":null,"OwnerUserId":"6791223","Title":null,"Body":"<p>Really have no idea how to do this in one line, so I've come up with the following:<\/p>\n\n<pre><code>import tensorflow as tf\n\nimage_tensor_1 = 1 * tf.ones([60, 2, 2, 3])\nimage_tensor_2 = 2 * tf.ones([60, 2, 2, 3])\nimage_tensor_3 = 3 * tf.ones([60, 2, 2, 3])\nimage_tensor_4 = 4 * tf.ones([60, 2, 2, 3])\n\n# make two tensors with shapes of [60, 2, 4, 3]\nconcat1 = tf.concat(2, [image_tensor_1, image_tensor_2])\nconcat2 = tf.concat(2, [image_tensor_3, image_tensor_4])\n# stack two tensors together to obtain desired result with shape [60, 4, 4, 3]\nresult = tf.concat(1, [concat1, concat2])\n<\/code><\/pre>\n\n<p>Following code:<\/p>\n\n<pre><code>sess = tf.Session()\nprint(sess.run(result[0, :, :, 0]))\n<\/code><\/pre>\n\n<p>results with<\/p>\n\n<pre><code>[[ 1.  1.  2.  2.]\n [ 1.  1.  2.  2.]\n [ 3.  3.  4.  4.]\n [ 3.  3.  4.  4.]]\n<\/code><\/pre>\n\n<p>as desired.<\/p>\n\n<p>Too late, haha :)<\/p>\n"}]}
{"QuestionId":42051371,"AnswerCount":2,"Tags":"<keras><image-segmentation>","CreationDate":"2017-02-05T11:13:41.703","AcceptedAnswerId":null,"OwnerUserId":1447257.0,"Title":"Deep Net with keras for image segmentation","Body":"<p>I am pretty new to deep learning; I want to train a network on image patches of size (256, 256, 3) to predict three labels of pixel-wise segmentation. As a start I want to provide one convolutional layer:<\/p>\n\n<pre><\/pre>\n\n<p>The model output so far is an image with 32 channels. Now, I want to add a dense layer which merges all of these 32 channels into three channels, each one predicting the probability of a class for one pixel.<\/p>\n\n<p>How can I do that?<\/p>\n","answers":[{"AnswerId":"42054278","CreationDate":"2017-02-05T16:11:53.280","ParentId":null,"OwnerUserId":"4955956","Title":null,"Body":"<p>The simplest method to merge your 32 channels back to 3 would be to add another convolution, this time with three filters (I arbitrarily set the filter sizes to be 1x1):<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Convolution2d(32, 16, 16, input_shape=(3, 256, 256))\nmodel.add(Convolution2d(3, 1, 1))\n<\/code><\/pre>\n\n<p>And then finally add an activation function for segmentation<\/p>\n\n<pre><code>model.add(Activation(\"tanh\"))\n<\/code><\/pre>\n\n<p>Or you could add it all at once if you want to with <strong>activation<\/strong> parameter (arbitrarily chosen to be tanh)<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Convolution2d(32, 16, 16, input_shape=(3, 256, 256))\nmodel.add(Convolution2d(3, 1, 1,activation=\"tanh\"))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/keras.io\/layers\/convolutional\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/convolutional\/<\/a><\/p>\n"},{"AnswerId":"42058982","CreationDate":"2017-02-06T00:41:20.827","ParentId":null,"OwnerUserId":"5256563","Title":null,"Body":"<p>You have to use flatten between the convolution layers and the dense layer:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Convolution2d(32, 16, 16, input_shape=(3, 256, 256))\n# Do not forget to add an activation layer after your convolution layer, so here.\n\nmodel.add(Flatten())\n\nmodel.add(Dense(3))\nmodel.add(Activation(\"sigmoid\")) # whatever activation you want.\n<\/code><\/pre>\n"}]}
{"QuestionId":42052750,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-05T13:40:37.890","AcceptedAnswerId":null,"OwnerUserId":6690269.0,"Title":"Why slim.nets.vgg use conv2d instead of fully_connected layers?","Body":"<p>In tensorflow\/models\/slim\/nets, here is the link of relative snippets of <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/nets\/vgg.py#L164-L174\" rel=\"noreferrer\">vgg<\/a>. I'am curious about why slim.nets.vgg use conv2d instead of fully_connected layers, although it works out the same way actually? Is it for the purpose of speed?\nI appreciate some explanations. Thanks!<\/p>\n","answers":[{"AnswerId":"43013975","CreationDate":"2017-03-25T08:01:28.153","ParentId":null,"OwnerUserId":"6690269","Title":null,"Body":"<p>After a while, I think there is at least one reason that it can avoid weights converting mistakes.<\/p>\n\n<p>Tensorflow\/slim as well as other high-level libraries allows tensor formats either <code>BHWC<\/code>(batch_size, height, width, channel. Same below) by default or <code>BCHW<\/code>(for better performance).<\/p>\n\n<p>When converting weights between these two formats, the weights <code>[in_channel, out_channel]<\/code> of first fc (fully connected layer, after conv layer) have to been reshaped to <code>[last_conv_channel, height, width, out_channel]<\/code> for example, then transposed to <code>[height, width, last_conv_channel, out_channel]<\/code> and reshaped again to <code>[in_channel, out_channel]<\/code>.<\/p>\n\n<p>And if you use conv weights rather than fully connected weights, such transformation will explicitly applied on fc layer (conv weights actually). Surely will avoid converting mistakes.<\/p>\n"}]}
{"QuestionId":42053170,"AnswerCount":1,"Tags":"<machine-learning><keras>","CreationDate":"2017-02-05T14:24:44.187","AcceptedAnswerId":42053507.0,"OwnerUserId":7387749.0,"Title":"Keras - How can I get biases from a trained model?","Body":"<p>I have built a simple neural network,<\/p>\n\n<pre><\/pre>\n\n<p>and I would get its weights by:<\/p>\n\n<pre><\/pre>\n\n<p>but, in this way, I only get the weights matrices (5x20 , 1x20) without the biases. How can I get the biases values?<\/p>\n","answers":[{"AnswerId":"42053507","CreationDate":"2017-02-05T14:59:02.430","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Quite simple, its just the second element in the array returned by get_weights() (For Dense layers):<\/p>\n\n<pre><code>B_Input_Hidden = model.layers[0].get_weights()[1]\nB_Output_Hidden = model.layers[1].get_weights()[1]\n<\/code><\/pre>\n"}]}
{"QuestionId":42053348,"AnswerCount":1,"Tags":"<python><tensorflow><lstm>","CreationDate":"2017-02-05T14:43:41.410","AcceptedAnswerId":null,"OwnerUserId":5891602.0,"Title":"Dimension error in LSTM Tensorflow, how to proceed?","Body":"<p>I have a sentence in the form of :<\/p>\n\n<pre><\/pre>\n\n<p>Where: <\/p>\n\n<ul>\n<li>d is the dimensionality of word embedding.  <\/li>\n<li>P is no. of tokens in sentence.<\/li>\n<\/ul>\n\n<p>I am trying to get the hidden representation of P using LSTM<\/p>\n\n<pre><\/pre>\n\n<p>Where:<\/p>\n\n<ul>\n<li>l is the dimensionality of the hidden vectors<\/li>\n<\/ul>\n\n<p>I tried the following <br><\/p>\n\n<pre><\/pre>\n\n<p>Embedding tensor of dimension [A, V ] and BasicLSTMCell(l)  l is hidden units.<\/p>\n\n<p>I am getting a dimension error: <\/p>\n\n<pre><\/pre>\n\n<p>Any suggestion please.<\/p>\n","answers":[{"AnswerId":"42056907","CreationDate":"2017-02-05T20:25:53.763","ParentId":null,"OwnerUserId":"2470337","Title":null,"Body":"<p>Your compiler might be complaining about the visibility scopes of your variables. As you see, \"para_\" is added to the variable names from the error report you quote. It is likely this is not what you intended. <\/p>\n\n<p>I could probably tell you more if you provided the full error message and a minimal (not) working example of your code.<\/p>\n"}]}
{"QuestionId":42053446,"AnswerCount":0,"Tags":"<linux><docker><caffe>","CreationDate":"2017-02-05T14:53:30.667","AcceptedAnswerId":null,"OwnerUserId":6786718.0,"Title":"Copying a once compiled folder from shared drive into own workspace in a docker image","Body":"<p>I am a user of a docker iamge on a server. On this server the caffe framework is installed on the \"shared\" side so that every docker image is able to access this framework without needing it to install themselves. <\/p>\n\n<p>However, I want to make changes in one python script (e.g the draw_net.py) which is located in a python folder. I now want to copy the files to my own workspace and continue working with them, and only change the draw_net.py but make this file use the compiled data (otherwise I would have to compile it my own, which is a little painful). However since I do not own these files I am not able to change them, therefore I wanted to copy one file into my container and change it, but the dependencies must be still kept up, this is where I run into errors. I do not want to recompile anything on my image, I just want to change one python script a little. <\/p>\n\n<p>The structure looks like the following:<\/p>\n\n<pre><\/pre>\n\n<p>Now I want the files from  , especially the  in my own docker image are, so in . <\/p>\n\n<p>A simple  command resulted in the following errors<\/p>\n\n<blockquote>\n  <p>cp: cannot create symbolic link 'cpcaffe\/build' Operation not\n  supported <\/p>\n  \n  <p>cp: cannot create symbolic link 'cpcaffe\/python\/caffe\/_caffe.so': Operation not supported<\/p>\n<\/blockquote>\n\n<p>Executing gives me the expected reseult that it cannot located the moduel  , which is located in \/opt\/caffe\/build\/lib\/_caffe.so<\/p>\n\n<p>I tried to add the line <\/p>\n\n<pre><\/pre>\n\n<p>at the beginning of the script but this did not change anything.<\/p>\n\n<p>I guess that I somehow to make a symbolic link by myself to the original version, but I am a little newbish around this topic and could need some pointers.<\/p>\n\n<p>I have root access to all folders under root\/ , all other folders are shared folders where I only have read rights.<\/p>\n","answers":[]}
{"QuestionId":42053709,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-05T15:21:42.180","AcceptedAnswerId":null,"OwnerUserId":5241351.0,"Title":"Unsuccessful TensorSliceReader constructor: Failed to find any matching files for","Body":"<p>I tried saving my model out and then tried to restore it, but seems tensorflow is unable to find the location of the matching files :-<\/p>\n\n<p>Code to save model output :-<\/p>\n\n<pre><\/pre>\n\n<p>Code to restore model<\/p>\n\n<pre><\/pre>\n\n<p>I am getting below errors :-<\/p>\n\n<p>W tensorflow\/core\/framework\/op_kernel.cc:975] Not found: Unsuccessful TensorSliceReader constructor: Failed to find any matching files for model.ckpt<\/p>\n\n<p>W tensorflow\/core\/framework\/op_kernel.cc:975] Not found: Unsuccessful TensorSliceReader constructor: Failed to find any matching files for model.ckpt<\/p>\n","answers":[{"AnswerId":"42059546","CreationDate":"2017-02-06T02:21:51.963","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>saver.restore()<\/code> method will fail unless you pass a path\u2014and not just a filename\u2014as the second argument. To work around this problem, you can call <code>saver.restore(sess, '.\/model.ckpt')<\/code> if you are running the script from the directory containing the checkpoint.<\/p>\n"}]}
{"QuestionId":42053727,"AnswerCount":0,"Tags":"<tensorflow><recurrent-neural-network><tflearn>","CreationDate":"2017-02-05T15:23:53.960","AcceptedAnswerId":null,"OwnerUserId":5762613.0,"Title":"How to use pre-built word embedding in tfLearn?","Body":"<p>I would like to build recurrent neural net binary classifier in tfLearn. But I want to use my pre-built word embedding, which I have saved in the pickle. Following is the tflearn code I am using in the model.<\/p>\n\n<pre><\/pre>\n\n<p>How can I use my pre-built word embedding in this model (line 2)?<\/p>\n","answers":[]}
{"QuestionId":42054451,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-05T16:28:36.267","AcceptedAnswerId":46008803.0,"OwnerUserId":7287271.0,"Title":"How do I do global average pooling in TensorFlow?","Body":"<p>How do I do global average pooling in TensorFlow? If I have a tensor of shape , is it enough to just use  as long as channels = classes?<\/p>\n","answers":[{"AnswerId":"42056850","CreationDate":"2017-02-05T20:20:23.137","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>You don't need to pass a stride. With padding='valid' (the default), if the spatial extent of the pooling filter is the same as the spatial extent of the image, you will get a 1x1 image out. Other than that, that is how you do it, yes.<\/p>\n"},{"AnswerId":"46008803","CreationDate":"2017-09-01T22:41:53.520","ParentId":null,"OwnerUserId":"3038791","Title":null,"Body":"<p>You could also do <strong>tf.reduce_mean(x, axis=[1,2])<\/strong>, specially if your height and width are not defined.<\/p>\n\n<p>Typically, in a CNN, tensors have a shape of <code>b, h, w, c<\/code> where <code>b<\/code> is the batch size, <code>w<\/code> and <code>h<\/code> correspond to the width and height dimensions, and <code>c<\/code> is the number of channels\/filters.<\/p>\n\n<p>When you reduce along the axis <code>[1,2]<\/code>, you reduce on the first and second dimensions of the tensor (keeping the batch size, and the number of channels\/filters)<\/p>\n"}]}
{"QuestionId":42055750,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-05T18:38:25.220","AcceptedAnswerId":42056846.0,"OwnerUserId":350080.0,"Title":"Continue training after keyboard interrupt?","Body":"<p>I was training tensorflow and then i mashed the keyboard for shits and giggles:<\/p>\n\n<pre><\/pre>\n\n<p>When I'm trying to start training again, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>What causes this and what can I do about it? Is there a proper way to pause\/cancel a training session? (Tensorflow seems to pick up where it left of if you startup with training 50 steps and then set the steps to 100)<\/p>\n","answers":[{"AnswerId":"42056846","CreationDate":"2017-02-05T20:20:11.690","ParentId":null,"OwnerUserId":"2470337","Title":null,"Body":"<p>It seems like your problem is caused by that <code>Tensorflow<\/code> is trying to load a session, which wasn't properly saved when you interrupted your code. Your solutions now would either be not loading the last session when you restart the code (by commenting the loading line), or deleting the saved session files (then it should automatically restart from scratch). It is hard to give a more specific example, since you didn't share your code...<\/p>\n"}]}
{"QuestionId":42056396,"AnswerCount":1,"Tags":"<python><json><tensorflow><neural-network><record>","CreationDate":"2017-02-05T19:38:35.110","AcceptedAnswerId":null,"OwnerUserId":2618346.0,"Title":"Tensorflow feed forward network session doesn't stop","Body":"<p>I am trying to build a simple feed forward neural network using TensorFlow and its tfr formatting. I have been using TensorFlow's tutorials and examples as a reference: \n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/how_tos\/reading_data\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/how_tos\/reading_data<\/a><\/p>\n\n<p>Given \"food\" float values, I want to predict the \"happiness\" float value that it produces.<\/p>\n\n<p>food_test.json is a a JSON file that contains the \"food\" value and its associated \"happiness\" value. This is the format the data is stored in.<\/p>\n\n<p>food_to_record.py is based off of tensorflow's convert_to_records.py. It reads in the food_test.json and converts it to a food_record.tfr file.<\/p>\n\n<p>food_reader.py is based off of tensorflow's fully_connected_reader.py. It reads in the food_record.tfr file and runs the data through a neural network.<\/p>\n\n<p>I run the program in this order:\n1. food_to_record.py\n2. food_reader.py<\/p>\n\n<p>When food_reader.py is run, it starts a TensorFlow session, but the session never terminates, does anyone know the reason for this?<\/p>\n\n<p>food_test.json:<\/p>\n\n<pre><\/pre>\n\n<p>food_to_record.py:<\/p>\n\n<pre><\/pre>\n\n<p>food_reader.py:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42057491","CreationDate":"2017-02-05T21:25:43.413","ParentId":null,"OwnerUserId":"7045779","Title":null,"Body":"<p>You must call <code>tf.train.start_queue_runners<\/code> to populate the queue before you call <code>run<\/code> or <code>eval<\/code> to execute reading files. Otherwise, reading will block while it waits for filenames from the queue. Pleases check <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/fully_connected_reader.py#L120\" rel=\"nofollow noreferrer\">run_training<\/a> method from the original example, or tensorflow's documentation about <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow noreferrer\">how_tos\/reading_data<\/a>.<\/p>\n"}]}
{"QuestionId":42056989,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-05T20:33:41.110","AcceptedAnswerId":42192685.0,"OwnerUserId":7466373.0,"Title":"Tensorflow sequence to sequence models","Body":"<p>I am getting the following errors when I try to execute \"translate.py\" in the sequence to sequence models tutorial <a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/\" rel=\"nofollow noreferrer\">here<\/a>. Can anyone please help me.<\/p>\n\n<p>Errors:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42192685","CreationDate":"2017-02-12T20:12:09.960","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>It appears that the code you are using simply misses the <code>prepare_wmt_data<\/code> function in <code>data_utils.py<\/code>; it's there in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/efa4a6cfaa9d69a74fba73546ee320cf740bbc25\/tutorials\/rnn\/translate\/data_utils.py#L247\" rel=\"nofollow noreferrer\">this version<\/a> of the example. You can also have a look at the complete (and latest) version of the sequence-to-sequence example code at <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/translate\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/translate<\/a><\/p>\n\n<p>They should work with TensorFlow ~0.12.<\/p>\n"}]}
{"QuestionId":42057313,"AnswerCount":1,"Tags":"<python><numpy><docker><theano>","CreationDate":"2017-02-05T21:08:52.447","AcceptedAnswerId":null,"OwnerUserId":6252456.0,"Title":"numpy.distutils missing from numpy installation, theano tries to use it","Body":"<p>I am trying to build a Docker image with Theano and Lasagne. These are the main steps:<\/p>\n\n<pre><\/pre>\n\n<p>When the image is ready I start it up and try to test theano by simply importing theano in ipython. I get the following error (end of it):<\/p>\n\n<pre><\/pre>\n\n<p>The installed versions according to pip are: numpy (1.12.0) and Theano (0.9.0b1). I have another image that works for some reason.  is not present there either but Theano does not try to call it on  command.<\/p>\n\n<ol>\n<li>Why isn't  installed with ?<\/li>\n<li>Why does Theano try to use it?<\/li>\n<li>Most importantly: how can I fix this?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"42058334","CreationDate":"2017-02-05T23:02:58.143","ParentId":null,"OwnerUserId":"4237316","Title":null,"Body":"<p>Installing <code>https:\/\/github.com\/Theano\/Theano\/archive\/master.zip<\/code> gives you the beta release of the library, and it doesn't seem to work properly for every system.<\/p>\n\n<p>Unless you absolutely need some feature from <code>Theano 0.9<\/code>, I would suggest that you install from pip's own repositories instead. Probably do the same for <code>Lasagne<\/code>.<\/p>\n\n<pre><code>RUN pip install theano lasagne\n<\/code><\/pre>\n"}]}
{"QuestionId":42057529,"AnswerCount":1,"Tags":"<lua><tensorflow>","CreationDate":"2017-02-05T21:29:10.113","AcceptedAnswerId":null,"OwnerUserId":5891602.0,"Title":"function similar to Lua nn.Linear() in tensorflow","Body":"<p>I am trying to implement a code from lua to tensorflow. What is the similar function for nn.Linear() in tensorflow. Any replacement or similar function to get the same result.<\/p>\n","answers":[{"AnswerId":"42057824","CreationDate":"2017-02-05T21:59:05.493","ParentId":null,"OwnerUserId":"7045779","Title":null,"Body":"<p>Just use <code>slim.fully_connected<\/code> in <code>tf.contrib.slim<\/code>.<\/p>\n\n<p>In Torch, <\/p>\n\n<pre><code>module= nn.Linear(input_dims, output_dims)\n<\/code><\/pre>\n\n<p>is equivalent with<\/p>\n\n<pre><code>input = tf.placeholder(tf.float32, [batch_size] + input_dims, name='real_images')\noutput = slim.fully_connected(input, output_dims, activation_fn=None)\n<\/code><\/pre>\n\n<p>in tensorflow without considering about data's batch opertion.<\/p>\n\n<p>Please check <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/slim\" rel=\"nofollow noreferrer\">tf.contrib.slim documentation<\/a>.<\/p>\n"}]}
{"QuestionId":42057667,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-05T21:42:42.327","AcceptedAnswerId":null,"OwnerUserId":7520522.0,"Title":"How can we use MNIST dataset one class as an input using tensorflow?","Body":"<p>I want to find the accuracy of one class in MMNIST dataset .So how can i split it on the basis of classes?<\/p>\n","answers":[{"AnswerId":"42057766","CreationDate":"2017-02-05T21:52:42.480","ParentId":null,"OwnerUserId":"7106579","Title":null,"Body":"<p>Not sure exactly what you are asking. I will answer about what I understood. In case you want to predict only one class for example digit 5 and rest of the digits. Then first you need to label your vectors in such a way that maybe label all those vectors as 'one' which has ground truth 5 and 'zero' to those vectors whose ground truth is not 5. \nThen design your network with only two nodes in output, where first node will show the probability that the input vector belongs to class 'one' (or digit 5) and second node will show the probability of belonging to class 'zero'. Then just train your network. \nTo find accuracy, you can simple techniques like just count how many times you predict right i.e. if probability is higher than 0.5 for right class classify it as that class. <\/p>\n\n<p>I hope that helps, if not maybe it would be better if you could explain your question more precisely. <\/p>\n"}]}
{"QuestionId":42058690,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-05T23:53:33.347","AcceptedAnswerId":null,"OwnerUserId":395573.0,"Title":"Separate gradients in tf.gradients","Body":"<p>Based on this <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/gradient_computation#gradients\" rel=\"nofollow noreferrer\">reference<\/a>,  returns the partial derivatives of <strong>sum<\/strong> of  w.r.t.  in .<\/p>\n\n<p>Is it possible to get separate values for each  in ?<\/p>\n\n<p><strong>Example<\/strong>. Please look at the following code:<\/p>\n\n<pre><\/pre>\n\n<p>It simply calculates  which is . But  returns the sum of :<\/p>\n\n<pre><\/pre>\n\n<p>What I need is a separate value for each row in .<\/p>\n","answers":[]}
{"QuestionId":42059103,"AnswerCount":1,"Tags":"<python><tensorflow><one-hot-encoding>","CreationDate":"2017-02-06T01:03:20.003","AcceptedAnswerId":null,"OwnerUserId":7520522.0,"Title":"how to predict only one class in tensorflow","Body":"<p>In case you want to predict only one class. Then first you need to label your vectors in such a way that maybe label all those vectors as 'one' which has ground truth 5 and 'zero' to those vectors whose ground truth is not 5. \nhow can I implement this in tensorflow using puthon<\/p>\n","answers":[{"AnswerId":"59105698","CreationDate":"2019-11-29T13:26:05.760","ParentId":null,"OwnerUserId":"7941307","Title":null,"Body":"<p>while preparing the data you can use numpy to set all the data points in class 5 as 1 and the others will be set to as 0 using .<\/p>\n\n<p><code>arr = np.where(arr!=5,arr,0)<\/code><\/p>\n\n<p><code>arr = np.where(arr=5,arr,1)<\/code> <\/p>\n\n<p>and then you can create a binary classifier using Tensorflow to classifiy them while using a <code>binary_crossentropy<\/code> loss to optimize the classifier<\/p>\n"}]}
{"QuestionId":42059230,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-02-06T01:28:50.977","AcceptedAnswerId":null,"OwnerUserId":4420961.0,"Title":"Tensorflow 'nan' Loss and '-inf' weights, Even with 0 Learning Rate","Body":"<p>I am training Deep Convolutional Neural Network on AWS GPU Machine.\nDataset -> Google SVHN\nTraining Size -> 200,000+<\/p>\n\n<p>I get Loss = 'nan' and W = '-inf'<\/p>\n\n<h3>Even with 0 Learning Rate<\/h3>\n\n<pre><\/pre>\n\n<p>I have tried following techniques:<\/p>\n\n<ol>\n<li>Used Several Different Optimisers (Adam, SGD, etc)<\/li>\n<li>Used different Activation functions on last layer (ReLU, Sigmoid, tanH)<\/li>\n<li>Initialized Weights and biases in different ways<\/li>\n<li>Tried Different Learning rates and rate-decays (from 0.001 to 0.0001)<\/li>\n<li>I thought there might be an error in my dataset, so removed first 10000 entries. Didn't work<\/li>\n<\/ol>\n\n<p>None of these things seemed to work for me.\nI am still getting 'nan' loss after 1500 steps.<\/p>\n\n<p>My Code :<\/p>\n\n<h3>Weight Initalization<\/h3>\n\n<pre><\/pre>\n\n<h3>Loss<\/h3>\n\n<pre><\/pre>\n\n<h3>Tensorflow Session<\/h3>\n\n<pre><\/pre>\n\n<h3>Since No learning takes place if Learning Rate is 0, How can loss and weights change and ho to nan and -inf.<\/h3>\n\n<p>Any Help is appreciated. <\/p>\n","answers":[{"AnswerId":"42138879","CreationDate":"2017-02-09T14:09:22.520","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>I have seen this happening when one label is out of range. Can you check that your labels are all in the range of (0 - (num_labels-1) ) ?<\/p>\n"},{"AnswerId":"44442947","CreationDate":"2017-06-08T18:19:40.753","ParentId":null,"OwnerUserId":"4385912","Title":null,"Body":"<p>While I cannot say for certain what problem your code has, there are two general ways to debug NaN's and Inf's that I can recommmend.<\/p>\n\n<p>The first is simply going over your code and look for any operation that might not be defined over its input. The operations that come to my mind immediately (since they are very common) are division (by 0) and logs (of negative values). This includes part of your code where this is not obvious, since you apply a more complex function that entails these operations. In your code, this include s<code>tf.reduce_mean<\/code> (which includes a problematic division if your input set sums to 0 - which might happen if it has length 0, too).<\/p>\n\n<p>The second is one of tensorflow's most useful op's: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/add_check_numerics_ops\" rel=\"nofollow noreferrer\">tf.add_check_numerics_ops<\/a> that creates an op (i.e. an operation that you need to call with session.run) that will tell you which computation node is either <code>inf<\/code> of <code>nan<\/code> ...<\/p>\n"},{"AnswerId":"56149380","CreationDate":"2019-05-15T12:28:37.700","ParentId":null,"OwnerUserId":"4958717","Title":null,"Body":"<p>I have a similar issue with a CNN implementation. And after implementing all the checks from other answers (assert input and label values are in range, no division by zero, etc.) the problem still persisted: after one training step the CNN would produce nan-s at output (logits).<\/p>\n\n<p>Interestingly I have stumbled upon a \"workaround\": if I replace <code>tf.train.AdamOptimizer<\/code> with <code>tf.train.RMSPropOptimizer<\/code> the nan values do not appear any more. This is very confusing.<\/p>\n\n<p>Hope this helps someone. I will update the answer when I find out more.<\/p>\n"}]}
{"QuestionId":42059375,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-06T01:52:35.247","AcceptedAnswerId":null,"OwnerUserId":1356280.0,"Title":"tensorflow, how does it work in parallel","Body":"<p>I'm studying machine learning and tensorflow.\nbut i have a question\nI saw documents that Tensorflow by default uses multiple threads.\nSo I tried to check this out by using log.<\/p>\n\n<pre><\/pre>\n\n<p>then I can get result below.<\/p>\n\n<blockquote>\n  <p>Const: (Const): \/job:localhost\/replica:0\/task:0\/cpu:0\n  I tensorflow\/core\/common_runtime\/simple_placer.cc:827] Const: (Const)\/job:localhost\/replica:0\/task:0\/cpu:0\n  add: (Add): \/job:localhost\/replica:0\/task:0\/cpu:0\n  I tensorflow\/core\/common_runtime\/simple_placer.cc:827] add: (Add)\/job:localhost\/replica:0\/task:0\/cpu:0\n  sub: (Sub): \/job:localhost\/replica:0\/task:0\/cpu:0\n  I tensorflow\/core\/common_runtime\/simple_placer.cc:827] sub: (Sub)\/job:localhost\/replica:0\/task:0\/cpu:0\n  add_1: (Add): \/job:localhost\/replica:0\/task:0\/cpu:0\n  I tensorflow\/core\/common_runtime\/simple_placer.cc:827] add_1: (Add)\/job:localhost\/replica:0\/task:0\/cpu:0\n  add\/y: (Const): \/job:localhost\/replica:0\/task:0\/cpu:0\n  I tensorflow\/core\/common_runtime\/simple_placer.cc:827] add\/y: (Const)\/job:localhost\/replica:0\/task:0\/cpu:0\n  sub\/y: (Const): \/job:localhost\/replica:0\/task:0\/cpu:0\n  I tensorflow\/core\/common_runtime\/simple_placer.cc:827] sub\/y: (Const)\/job:localhost\/replica:0\/task:0\/cpu:0\n  238<\/p>\n<\/blockquote>\n\n<p>It seems like work in parallel and synchronous, is that right?\nIs it default?<\/p>\n","answers":[{"AnswerId":"54005347","CreationDate":"2019-01-02T11:16:49.533","ParentId":null,"OwnerUserId":"989329","Title":null,"Body":"<p>Check the implementation in direct_session.c and executor.cc, where ExecutorState is created for each partition of the Graph, and a threadpool is used to execute each node ready to be processed.<\/p>\n\n<p>In below function, \"ready\" is a vector initialised with the input nodes of the graph, so each will run in a separate thread, and continue executing the nodes in that branch which has its dependencies satisfied. As the processing across threads continues, the dependencies for different nodes will be satisfied and will be executed. <\/p>\n\n<pre><code>void ExecutorState::RunAsync(Executor::DoneCallback done)\n    \/\/ Schedule to run all the ready ops in thread pool.\n    ScheduleReady(ready, nullptr);\n}\n<\/code><\/pre>\n\n<p>ExecutorState::ScheduleReady is the function which triggers the parallelism and is called from code where nodes are ready to be processed, which includes ExecutorState::RunAsync and ExecutorState::NodeDone<\/p>\n"}]}
{"QuestionId":42059487,"AnswerCount":1,"Tags":"<windows><tensorflow><install><anaconda>","CreationDate":"2017-02-06T02:14:06.213","AcceptedAnswerId":42083248.0,"OwnerUserId":7509159.0,"Title":"Tensorflow install on Windows with anaconda and no internet connection","Body":"<p>I can not install Tensorflow on Windows because there in no internet connection based on commpany's security policy. <\/p>\n\n<p>I just installed anaconda and python by transmiting files with intranet. <\/p>\n\n<p>Please let me know how to install that with no internet connection. <\/p>\n\n<p>==========================================================================<\/p>\n\n<p>In addition, when I used the below command after installing tensorflow, I found\nother problems..<\/p>\n\n<pre><\/pre>\n\n<p>I don't know how to solve this one.<\/p>\n","answers":[{"AnswerId":"42083248","CreationDate":"2017-02-07T06:36:05.540","ParentId":null,"OwnerUserId":"732003","Title":null,"Body":"<p>If you can download the whl file and transfer it to your workstation, then you can run:<\/p>\n\n<p><code>pip.exe install --upgrade --no-deps &lt;tensorflow whl file name&gt;\n<\/code><\/p>\n\n<p>This should avoid trying to connect to download tensorflow dependencies, as anaconda already has most of these.<\/p>\n"}]}
{"QuestionId":42060176,"AnswerCount":1,"Tags":"<tensorflow><distance><automatic-differentiation>","CreationDate":"2017-02-06T04:06:38.927","AcceptedAnswerId":null,"OwnerUserId":3068233.0,"Title":"Calculate distance between n data points and k clusters in TensorFlow","Body":"<p> is a matrix of data points,  by  in shape.  is a matrix of cluster points,  by  in shape. The smallest distance between a datapoint, , and each cluster can be calculated as follows:<\/p>\n\n<pre><\/pre>\n\n<p>How can the distance between every data point and each cluster be calculated in a tensorflow graph method? A for loop comes to mind but no such functionality exists in tensorflow as far as I understand.<\/p>\n","answers":[{"AnswerId":"42060997","CreationDate":"2017-02-06T05:38:39.493","ParentId":null,"OwnerUserId":"3068233","Title":null,"Body":"<p>Most efficient method I have found so far is the following:<\/p>\n\n<pre><code>################\n# Reshape Tensors for Calculation\n################\nsamples = X;\ncentroids = W;\nexpanded_vectors = tf.expand_dims(samples, 0)\nexpanded_centroids = tf.expand_dims(centroids, 1)\n\n\n################\n# Define objective of model\n################\ndistances = tf.reduce_sum( tf.square(tf.sub(expanded_vectors, expanded_centroids)), 2)\nmin_distances = tf.reduce_min(distances, 0);   \n<\/code><\/pre>\n"}]}
{"QuestionId":42060241,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-02-06T04:17:50.587","AcceptedAnswerId":null,"OwnerUserId":7521228.0,"Title":"How to create shuffled batch list from numpy arrays to feed tensorflow dictionnary","Body":"<p>I am working on a classifier with Tensorflow. My input and output are numpy arrays with examples as rows and parameter as columns. My code is working correctly until now, and I could train my network by directly feeding 100% of my input and output arrays.\nNow, I would like to create a list of shuffled batches to feed tensorflow, but this step is not so clear for me.<\/p>\n\n<p>Does the following creates lists of tensors (my batches)?<\/p>\n\n<pre><\/pre>\n\n<p>Then I am trying to feed this batches to tensorflow :<\/p>\n\n<pre><\/pre>\n\n<p>But nothing happens after this line.<\/p>\n\n<p>How can I correctly prepare batches from numpy arrays and feed it to tensorflow dictionnary?<\/p>\n","answers":[{"AnswerId":"42074540","CreationDate":"2017-02-06T18:15:17.240","ParentId":null,"OwnerUserId":"7521228","Title":null,"Body":"<p>I finally found how to create different batches randomly for each epoch. I post my solution as it may be useful for someone else. So here is the trick.<\/p>\n\n<pre><code>for epoch in range(0,2000):\n    permutation=np.random.permutation(input_size)\n    permutation=permutation[0:batch_size]\n    batch=[train_set[permutation],train_label[permutation]]\n    sess.run(train_step,feed_dict={X:batch[0],Yreal:batch[1]})\n<\/code><\/pre>\n\n<p>It may not be the sexiest way, but it's working. We have a random list (permutation) created at each epoch from which we extract our batches. Each batch is then fed to tensorflow.<\/p>\n"}]}
{"QuestionId":42061142,"AnswerCount":0,"Tags":"<python><tensorflow><scikit-learn><standardized>","CreationDate":"2017-02-06T05:53:37.950","AcceptedAnswerId":null,"OwnerUserId":1616677.0,"Title":"standardization of a tensor in tensorflow","Body":"<p>I have the following functions to make standardization\/inversion of the input\/output to a list of values. Now I want to inverse-apply it to a tensor, so that I can write loss value in \"real\" units. Obviously, this can't be applied to a tensor directly. Surely this can be done by manual computation, but I wonder if there are framework means of doing this? I've found  but not sure this what I need, and how to invert it.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42061224,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-06T06:01:19.143","AcceptedAnswerId":42076464.0,"OwnerUserId":3425082.0,"Title":"Duplicate a network in tensorflow","Body":"\n\n<p>I've asked a similar question <a href=\"https:\/\/stackoverflow.com\/questions\/42018010\/copy-tensorflow-contrib-layers?noredirect=1#comment71210552_42018010\">here<\/a>. <\/p>\n\n<p>However, the answer didn't satisfy my need, and no one has replied to my comment, so I have to repost the question and make it clearer.<\/p>\n\n<p>I have 2 networks, called Target &amp; Source. For simple, the networks are defined as follow:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Now, after some training steps, I want to copy the content (parameters) of layer  to layer  so that  doesn't change as  changed due to the gradient descent &amp;  still gets input from .<\/p>\n\n<p>I've tried the solution that Yaroslav Bulatov suggested, however it didn't work.<\/p>\n\n<p>If I define my network using simply , I can easily copy the variable by , but now I want to use  which is much more simpler &amp; flexible enough to define my own network.<\/p>\n\n<p>If anyone doesn't understand my question, please notify me so that I can fix it.<\/p>\n","answers":[{"AnswerId":"42076464","CreationDate":"2017-02-06T20:13:47.917","ParentId":null,"OwnerUserId":"6938362","Title":null,"Body":"<p>Hi you can use assign to create a copy operation, e.g.<\/p>\n\n<pre><code>s_output = tf.contrib.layers.fully_connected(input = s_input, num_outputs=1, weights_initializer=tf.contrib.layers.xavier_initializer())\nt_output = tf.contrib.layers.fully_connected(input = s_input, num_outputs=1)\n<\/code><\/pre>\n\n<p>Now you can access the trainable variables<\/p>\n\n<pre><code>vars = tf.trainable_variables()\n<\/code><\/pre>\n\n<p>and copy them (the first half are the variables from s_output and the second half from t_output):<\/p>\n\n<pre><code>copy_ops = [vars[ix+len(vars)\/\/2].assign(var.value()) for ix, var in enumerate(vars[0:len(vars)\/\/2])]\n<\/code><\/pre>\n\n<p>Now you can copy the data using:<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\nsess = tf.Session() \nsess.run(init)\nmap(lambda x: sess.run(x), copy_ops)\nprint(sess.run(vars[2]))\n<\/code><\/pre>\n\n<p>Hope this what you're looking for.<\/p>\n"},{"AnswerId":"55886683","CreationDate":"2019-04-28T03:06:53.310","ParentId":null,"OwnerUserId":"9056709","Title":null,"Body":"<p>The simplest way is to do the mathematical operation with tf. First, get the weight of each layer. To get weight from any layer:\nMy fc1 layer looks like this:-<\/p>\n\n<pre><code>with tf.variable_scope(\"fc1\"):\n    fc1 = tf.contrib.layers.fully_connected(inputs = input_,\n                                            num_outputs = 10,\n                                            activation_fn=tf.nn.relu,\n                                            weights_initializer=tf.contrib.layers.xavier_initializer())\n<\/code><\/pre>\n\n<p>To get the weight and biases of the layer the following operation can be done:<code>var_fc1 = tf.trainable_variables('fc1') \nfc1_w_  = np.array(sess.run(var_fc1)[0])  #get weight\nfc1_b_  = np.array(sess.run(var_fc1)[1])  # get biasess<\/code>\nMimic what fc1 is doing :<\/p>\n\n<pre><code>fc1_old = tf.add(tf.matmul(input_ , fc1_w) , fc1_b)\nfc1_old = tf.nn.relu(fc1_old)\n<\/code><\/pre>\n"}]}
{"QuestionId":42061854,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-06T06:47:59.557","AcceptedAnswerId":null,"OwnerUserId":5588279.0,"Title":"Variable length input in Tensorflow","Body":"<p>The shape of input is . And I want to use a certain cnn at each step to make it become . At first glance, I think I should use .<\/p>\n\n<p>But it is difficult to build this graph, the \"maxstep\" and \"batch_size\" are both variable, so how to iterate this tensor?<\/p>\n\n<p>This is my wrong code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42103054","CreationDate":"2017-02-08T01:23:07.153","ParentId":null,"OwnerUserId":"5588279","Title":null,"Body":"<p>With <a href=\"https:\/\/stackoverflow.com\/users\/6824418\/allen-lavoie\">AllenLavoie<\/a>'s help, I finally solved this problem.<\/p>\n\n<p>First I pad the input's shape <code>[batch_size, variable_step_len, 50, 50]<\/code> into <code>[batch_size, max_step_len, 50, 50]<\/code>, and then reshape it into <code>[batch_size*max_step_len, 50, 50]<\/code>.<\/p>\n\n<p>Second I op cnn on it and reshape it back.<\/p>\n"}]}
{"QuestionId":42061855,"AnswerCount":2,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-02-06T06:48:12.203","AcceptedAnswerId":null,"OwnerUserId":350080.0,"Title":"What is loss exactly?","Body":"<p>I'm training the im2txt sample of tensorflow and it outputs the loss:<\/p>\n\n<pre><\/pre>\n\n<p>What is loss? How does it relate to the AI:s probability to perform correctly (is there a formula)? What is ususally an acceptable loss?<\/p>\n","answers":[{"AnswerId":"42065164","CreationDate":"2017-02-06T10:12:08.463","ParentId":null,"OwnerUserId":"241013","Title":null,"Body":"<p>From: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/im2txt\/im2txt\/show_and_tell_model.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/im2txt\/im2txt\/show_and_tell_model.py<\/a><\/p>\n\n<pre><code>  losses = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=targets,\n                                                          logits=logits)\n  batch_loss = tf.div(tf.reduce_sum(tf.multiply(losses, weights)),\n                      tf.reduce_sum(weights),\n                      name=\"batch_loss\")\n<\/code><\/pre>\n\n<p><code>softmax<\/code> is basically a fancy max function that is derivable (you can lookup the exact definition in the docs). It's going to have high values for the largest activations. It can have multiple high activation and this is going to get penalized for all the wrong ones.<\/p>\n\n<p>Loss is the the thing that you want the model to optimize down.<\/p>\n\n<p>This doesn't usually mean much unless you've trained lots of similar models with the exact same loss. Usually you look at the loss graph to see when the model stopped making progress so that you can stop training. Also keep in mind that for other models you will want to add other things to the loss that you might want to optimize (say some input weights that you want to average to 1.0) that will bump up the loss but that doesn't mean the model is worse.<\/p>\n\n<p>If you want to figure out if your model is good or bad add metrics for the things you care about. The obvious things are precision\/recall\/accuracy. There's predefined metrics already that you can use (<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/contrib.metrics\/metric_ops#streaming_accuracy\" rel=\"nofollow noreferrer\">streaming_accuracy<\/a>). Alternatively you can compute the metric and add it as a summary but that's not going to be available from the eval dataset.<\/p>\n\n<p>Another option is to setup a model that's obviously bad (constant or random) and compare the loss of that model with what you are getting.<\/p>\n"},{"AnswerId":"42076606","CreationDate":"2017-02-06T20:22:27.367","ParentId":null,"OwnerUserId":"3935325","Title":null,"Body":"<p>Loss is the target function that the optimization algorithm will try to minimize.<\/p>\n\n<p>In general, you want your loss function to be a measure of how bad your model is. But because the optimization algorithms require a few mathematical properties to work nicely, you can't pick the usual stuff like precision and recall (you want continuous functions that are differentiable in relation to the model parameters).<\/p>\n\n<p>With classification tasks, <code>softmax<\/code> is a common choice. It's a smooth and well-behaved version of <code>argmax<\/code>, which is used to pick the class with highest network activation. With regression, the usual <code>mean squared error<\/code> serves fine.<\/p>\n"}]}
{"QuestionId":42063327,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><jupyter-notebook><keras>","CreationDate":"2017-02-06T08:30:51.723","AcceptedAnswerId":null,"OwnerUserId":3952760.0,"Title":"Keras NoteBook GPU Timeout","Body":"<p>I am trying to run keras with tensorflow on a windows 10 machine with my GTX 980 gpu on a jupyter notebook. If I run tensorflow alone with my gpu, its works perfectly fine without any issues. But problems arise with the keras interface for high number of epochs. <\/p>\n\n<p>The keras model uses the GPU and gives an output if my number of epochs is low like the following<\/p>\n\n<pre><\/pre>\n\n<p>Following is the output<\/p>\n\n<blockquote>\n  <p>Train on 31367 samples, validate on 7842 samples\n  Epoch 1\/2\n  31367\/31367 [==============================] - 3s - loss: 1.7640 - acc: 0.5438 - val_loss: 1.2872 - val_acc: 0.6486 - ETA: 0s - loss: 1.8827 - acc: 0.5145 - ETA: 0s - loss: 1.7732 - acc: 0.5416<\/p>\n  \n  <p>Epoch 2\/2\n  31367\/31367 [==============================] - 2s - loss: 0.8539 - acc: 0.7765 - val_loss: 0.7958 - val_acc: 0.7615<\/p>\n<\/blockquote>\n\n<p>If the number of epochs is high then it will timeout with the following error and the webpage says busy<\/p>\n\n<blockquote>\n  <p>WebSocket ping timeout after 119999 ms.<\/p>\n<\/blockquote>\n\n<p>How do i fix this error?<\/p>\n","answers":[{"AnswerId":"45540623","CreationDate":"2017-08-07T06:30:44.323","ParentId":null,"OwnerUserId":"1314209","Title":null,"Body":"<p>I guess this issue is related to TDR(Timeout Detection and Recovery) on Windows. \nBasically, the OS thought the GPU hang and do not response any more, so OS will reboot the graphics card. You can try to disable the TDR or extend the up limit of TdRDelay. More details can be found <a href=\"https:\/\/docs.microsoft.com\/en-us\/windows-hardware\/drivers\/display\/tdr-registry-keys\" rel=\"nofollow noreferrer\">https:\/\/docs.microsoft.com\/en-us\/windows-hardware\/drivers\/display\/tdr-registry-keys<\/a>.<\/p>\n"}]}
{"QuestionId":42063829,"AnswerCount":0,"Tags":"<tensorflow><google-cloud-platform><google-cloud-ml>","CreationDate":"2017-02-06T09:01:20.967","AcceptedAnswerId":null,"OwnerUserId":2839624.0,"Title":"Google cloud ml cannot deploy graph for prediction api","Body":"<p>Has anybody tried exporting tensorflow model built on local machine and use google cloud ml prediction API uploading this model?<\/p>\n\n<p>I am trying to build a graph using slightly modified version of retraining example given in <a href=\"https:\/\/cloud.google.com\/blog\/big-data\/2016\/12\/how-to-train-and-classify-images-using-google-cloud-machine-learning-and-cloud-dataflow\" rel=\"nofollow noreferrer\">this<\/a> blog on my local machine and upload exported model checkpoints directly to cloud ml for prediction. However it throws an error saying:<\/p>\n\n<pre><\/pre>\n\n<p>When for debugging, I try to predict locally with same graph and same example using command , it gives error:<\/p>\n\n<pre><\/pre>\n\n<p>It seems that it cannot import all the graph elements from model. I have exported model using this command:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42064690,"AnswerCount":1,"Tags":"<machine-learning><neural-network><keras><lstm><word2vec>","CreationDate":"2017-02-06T09:47:22.213","AcceptedAnswerId":null,"OwnerUserId":9655419.0,"Title":"Using pre-trained word2vec with LSTM for word generation","Body":"<p>LSTM\/RNN can be used for text generation.\n<a href=\"https:\/\/blog.keras.io\/using-pre-trained-word-embeddings-in-a-keras-model.html\" rel=\"noreferrer\">This<\/a> shows way to use pre-trained GloVe word embeddings for Keras model.<\/p>\n\n<ol>\n<li>How to use pre-trained Word2Vec word embeddings with Keras LSTM\nmodel? <a href=\"http:\/\/ben.bolte.cc\/blog\/2016\/keras-gensim-embeddings.html\" rel=\"noreferrer\">This<\/a> post did help.<\/li>\n<li>How to predict \/ generate next <em>word<\/em> when the model is provided with the sequence of words as its input?<\/li>\n<\/ol>\n\n<p>Sample approach tried:<\/p>\n\n<pre><\/pre>\n\n<p>Sample code \/ psuedocode to train LSTM and predict will be appreciated. <\/p>\n","answers":[{"AnswerId":"48230654","CreationDate":"2018-01-12T16:51:22.010","ParentId":null,"OwnerUserId":"712995","Title":null,"Body":"\n\n<p>I've created a <a href=\"https:\/\/gist.github.com\/maxim5\/c35ef2238ae708ccb0e55624e9e0252b\" rel=\"noreferrer\">gist<\/a> with a simple generator that builds on top of your initial idea: it's an LSTM network wired to the pre-trained word2vec embeddings, trained to predict the next word in a sentence. The data is the <a href=\"https:\/\/raw.githubusercontent.com\/maxim5\/stanford-tensorflow-tutorials\/master\/data\/arxiv_abstracts.txt\" rel=\"noreferrer\">list of abstracts from arXiv website<\/a>.<\/p>\n\n<p>I'll highlight the most important parts here.<\/p>\n\n<h2>Gensim Word2Vec<\/h2>\n\n<p>Your code is fine, except for the number of iterations to train it. The default <code>iter=5<\/code> seems rather low. Besides, it's definitely not the bottleneck -- LSTM training takes much longer. <code>iter=100<\/code> looks better.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>word_model = gensim.models.Word2Vec(sentences, size=100, min_count=1, \n                                    window=5, iter=100)\npretrained_weights = word_model.wv.syn0\nvocab_size, emdedding_size = pretrained_weights.shape\nprint('Result embedding shape:', pretrained_weights.shape)\nprint('Checking similar words:')\nfor word in ['model', 'network', 'train', 'learn']:\n  most_similar = ', '.join('%s (%.2f)' % (similar, dist) \n                           for similar, dist in word_model.most_similar(word)[:8])\n  print('  %s -&gt; %s' % (word, most_similar))\n\ndef word2idx(word):\n  return word_model.wv.vocab[word].index\ndef idx2word(idx):\n  return word_model.wv.index2word[idx]\n<\/code><\/pre>\n\n<p>The result embedding matrix is saved into <code>pretrained_weights<\/code> array which has a shape <code>(vocab_size, emdedding_size)<\/code>.<\/p>\n\n<h2>Keras model<\/h2>\n\n<p>Your code is almost correct, except for the loss function. Since the model predicts the next word, it's a classification task, hence the loss should be <code>categorical_crossentropy<\/code> or <code>sparse_categorical_crossentropy<\/code>. I've chosen the latter for efficiency reasons: this way it avoids one-hot encoding, which is pretty expensive for a big vocabulary.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>model = Sequential()\nmodel.add(Embedding(input_dim=vocab_size, output_dim=emdedding_size, \n                    weights=[pretrained_weights]))\nmodel.add(LSTM(units=emdedding_size))\nmodel.add(Dense(units=vocab_size))\nmodel.add(Activation('softmax'))\nmodel.compile(optimizer='adam', loss='sparse_categorical_crossentropy')\n<\/code><\/pre>\n\n<p>Note passing the pre-trained weights to <code>weights<\/code>.<\/p>\n\n<h2>Data preparation<\/h2>\n\n<p>In order to work with <code>sparse_categorical_crossentropy<\/code> loss, both sentences and labels must be word indices. Short sentences must be padded with zeros to the common length.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>train_x = np.zeros([len(sentences), max_sentence_len], dtype=np.int32)\ntrain_y = np.zeros([len(sentences)], dtype=np.int32)\nfor i, sentence in enumerate(sentences):\n  for t, word in enumerate(sentence[:-1]):\n    train_x[i, t] = word2idx(word)\n  train_y[i] = word2idx(sentence[-1])\n<\/code><\/pre>\n\n<h2>Sample generation<\/h2>\n\n<p>This is pretty straight-forward: the model outputs the vector of probabilities, of which the next word is sampled and appended to the input. Note that the generated text would be better and more diverse if the next word is <em>sampled<\/em>, rather than <em>picked<\/em> as <code>argmax<\/code>. The temperature based random sampling I've used is <a href=\"https:\/\/medium.com\/machine-learning-at-petiteprogrammer\/sampling-strategies-for-recurrent-neural-networks-9aea02a6616f\" rel=\"noreferrer\">described here<\/a>.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def sample(preds, temperature=1.0):\n  if temperature &lt;= 0:\n    return np.argmax(preds)\n  preds = np.asarray(preds).astype('float64')\n  preds = np.log(preds) \/ temperature\n  exp_preds = np.exp(preds)\n  preds = exp_preds \/ np.sum(exp_preds)\n  probas = np.random.multinomial(1, preds, 1)\n  return np.argmax(probas)\n\ndef generate_next(text, num_generated=10):\n  word_idxs = [word2idx(word) for word in text.lower().split()]\n  for i in range(num_generated):\n    prediction = model.predict(x=np.array(word_idxs))\n    idx = sample(prediction[-1], temperature=0.7)\n    word_idxs.append(idx)\n  return ' '.join(idx2word(idx) for idx in word_idxs)\n<\/code><\/pre>\n\n<h2>Examples of generated text<\/h2>\n\n<pre class=\"lang-py prettyprint-override\"><code>deep convolutional... -&gt; deep convolutional arithmetic initialization step unbiased effectiveness\nsimple and effective... -&gt; simple and effective family of variables preventing compute automatically\na nonconvex... -&gt; a nonconvex technique compared layer converges so independent onehidden markov\na... -&gt; a function parameterization necessary both both intuitions with technique valpola utilizes\n<\/code><\/pre>\n\n<p>Doesn't make too much sense, but is able to produce sentences that look at least grammatically sound (sometimes).<\/p>\n\n<p>The link to the <a href=\"https:\/\/gist.github.com\/maxim5\/c35ef2238ae708ccb0e55624e9e0252b\" rel=\"noreferrer\">complete runnable script<\/a>.<\/p>\n"}]}
{"QuestionId":42064941,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-06T10:00:44.820","AcceptedAnswerId":42077538.0,"OwnerUserId":7522443.0,"Title":"TensorFlow float16 support is broken","Body":"<p>Recently I tried to train a CNN in TF using float16. To my surprise it is broken in various ways even though TF claims to support it for a while. For example, float16 optimization causes NaN loss already on the second step regardless of the network.<\/p>\n\n<pre><\/pre>\n\n<p>To my understanding it is clearly a bug: I apply zero convolutions on zero input, I should get zero gradients that don't change zero loss. It just can't diverge. If dtype is float32 it works. NaN loss occurs both on CPU and GPU versions.<\/p>\n\n<p>However, I was dismissed in GH issues, a random dude closed this issue saying that it is intended behaviour: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7226\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7226<\/a><\/p>\n\n<p>If you uncomment the line with BN, it will break already on graph construction time because BN assumes moving averages (and beta, gamma) are always float32 and does not cast them properly. This issue was also closed and apparently ignored: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7164\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7164<\/a><\/p>\n\n<p>I feel like I am talking to a first line IT support of an ISP.<\/p>\n\n<p>Can anybody explain how I should train with float16 when such a simple \"network\" fails horribly? And what is the recommended way to report bugs now?<\/p>\n","answers":[{"AnswerId":"42077538","CreationDate":"2017-02-06T21:26:35.673","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>It looks like you need a slightly larger epsilon to avoid numerical instability with zero moments in AdamOptimizer (default is 1e-8). This works for me with float16:<\/p>\n\n<pre><code>opt = tf.train.AdamOptimizer(1e-3, epsilon=1e-4)\n<\/code><\/pre>\n\n<p>It would be reasonable to request that epsilon be set based on dtype (and presumably such a request, or better yet a pull request, would be met with a more positive response on GitHub). Note that GradientDescentOptimizer has no such issue.<\/p>\n"}]}
{"QuestionId":42065222,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-06T10:15:10.740","AcceptedAnswerId":null,"OwnerUserId":1274878.0,"Title":"Multiple GPUs, but the same PCI bus id","Body":"<p>Here is a simple script I wrote to utilize multiple GPUs:<\/p>\n\n<pre><\/pre>\n\n<p>Basically, creating a separate process for each graph-gpu combination.I am doing it like this, because my models are completely independent. When I utilize a single GPU, my process completes in , but with all the four processes, the completion time is . In my TF graph, I am using . Could someone please tell me why this overhead is there or how to find out the cause of the overhead?<\/p>\n\n<p>One observation:<\/p>\n\n<p>For all the 4 GPUs, Tensorflow shows . But, when I do , I get <\/p>\n\n<pre><\/pre>\n\n<p>They all have different PCI bus ids as per the system. So, why is Tensorflow showing the same PCI bus ids for all the GPUs? Could that be the bottleneck?<\/p>\n","answers":[]}
{"QuestionId":42065870,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-06T10:46:05.233","AcceptedAnswerId":null,"OwnerUserId":3425890.0,"Title":"How can i use my own images to train my CNN neural network in tensorFlow","Body":"<p>I am currently working on a program that uses a CNN tensorflow neural network and I want to use my own images to train and test it, please I want some advice because I am new in deep learning <\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"50126477","CreationDate":"2018-05-02T02:51:39.323","ParentId":null,"OwnerUserId":"9727942","Title":null,"Body":"<p>Download the ronnie package from python.[{pythonpath}\\scripts\\pip3 ronnie]\nConstruct the training data structure as [label, image pixel] and execute the below code to create the training data.<\/p>\n\n<p>dataFileLoc = os.path.join(dir,\".\/data\/digitalRec\/train.csv\")\norgDF = collector.initial(dataFileLoc)<\/p>\n\n########################################################################### Return Object from buildFormDF are ImgXs and LabelYs in dict Structure\n\n<h1>{meta.DATA_INPUTXs: imgXs, meta.DATA_LABELYs: labelYs}<\/h1>\n\n##########################################################################    trainData       = collector.buildFromDF(orgDF,batchNum=1, batchSize=30000)\n"}]}
{"QuestionId":42066512,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-06T11:20:07.830","AcceptedAnswerId":42072258.0,"OwnerUserId":7522709.0,"Title":"Why does tf.cond() recognize tf.bool as python bool not as tf.bool?","Body":"<h1>Training NN to learn 'Xor'<\/h1>\n\n<p>I'm trying to use 'batch normalization' and I made a batch normalization layer function 'batch_norm1'. <\/p>\n\n<pre><\/pre>\n\n<p>When I run the Code above, I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I used <\/p>\n\n<pre><\/pre>\n\n<p>,whose type is \"tf.bool\", as an input of function 'batch_norm1'. However, the error said that I used python bool, not tf.bool. Can anyone tell me what's going on in this code??<\/p>\n","answers":[{"AnswerId":"42072258","CreationDate":"2017-02-06T16:09:24.483","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This line looks incorrect to me:<\/p>\n\n<pre><code> mean , var = tf.cond(is_training , mean_var_update , lambda :tf.identity(shadow_mean), tf.identity(shadow_var))\n<\/code><\/pre>\n\n<p>In particular, <code>tf.identity(shadow_var)<\/code> is being treated as a fourth argument to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#cond\" rel=\"nofollow noreferrer\"><code>tf.cond()<\/code><\/a> (i.e. the <code>name<\/code> argument) rather than the second return value from the <code>lambda<\/code> (as I think you intend). Adding parentheses to create a tuple of the two <code>tf.identity()<\/code> tensors should remedy this problem:<\/p>\n\n<pre><code> mean, var = tf.cond(is_training,\n                     mean_var_update,\n                     lambda: (tf.identity(shadow_mean), tf.identity(shadow_var)))\n<\/code><\/pre>\n"},{"AnswerId":"45131254","CreationDate":"2017-07-16T17:06:27.553","ParentId":null,"OwnerUserId":"7886651","Title":null,"Body":"<p>Try <code>tf.cast(is_training, tf.bool)<\/code>. That should work out.\nSo you code should become: <\/p>\n\n<pre><code>mean , var = tf.cond(tf.cast(is_training, tf.bool) , mean_var_update , lambda :tf.identity(shadow_mean), tf.identity(shadow_var))\n<\/code><\/pre>\n\n<p>Please let me if it works. <\/p>\n"}]}
{"QuestionId":42068999,"AnswerCount":4,"Tags":"<python><tensorflow><conv-neural-network><mini-batch>","CreationDate":"2017-02-06T13:28:00.713","AcceptedAnswerId":46286109.0,"OwnerUserId":3990607.0,"Title":"Tensorflow: Convolutions with different filter for each sample in the mini-batch","Body":"<p>I would like to have a 2d convolution with a filter which depends on the sample in the mini-batch in tensorflow. Any ideas how one could do that, especially if the number of sample per mini-batch is not known? <\/p>\n\n<p>Concretely, I have input data  of the form , and I have filters  of the form .<\/p>\n\n<p>It is assumed that <\/p>\n\n<p>.<\/p>\n\n<p>I would like to do , but this is not allowed because  cannot have a mini-batch dimension. Any idea how to solve this problem? <\/p>\n","answers":[{"AnswerId":"49030963","CreationDate":"2018-02-28T13:40:56.330","ParentId":null,"OwnerUserId":"5896264","Title":null,"Body":"<p>You could use <code>tf.map_fn<\/code> as follows:<\/p>\n\n<pre><code>inp = tf.placeholder(tf.float32, [None, h, w, c_in]) \ndef single_conv(tupl):\n    x, kernel = tupl\n    return tf.nn.conv2d(x, kernel, strides=(1, 1, 1, 1), padding='VALID')\n# Assume kernels shape is [tf.shape(inp)[0], fh, fw, c_in, c_out]\nbatch_wise_conv = tf.squeeze(tf.map_fn(\n    single_conv, (tf.expand_dims(inp, 1), kernels), dtype=tf.float32),\n    axis=1\n)\n<\/code><\/pre>\n\n<p>It is important to specify <code>dtype<\/code> for <code>map_fn<\/code>. Basically, this solution defines <code>batch_dim_size<\/code> 2D convolution operations.<\/p>\n"},{"AnswerId":"50213363","CreationDate":"2018-05-07T11:40:01.877","ParentId":null,"OwnerUserId":"7146501","Title":null,"Body":"<p>The accepted answer is slightly wrong in how it treats the dimensions, as they are changed by padding = \"VALID\" (he treats them as if padding = \"SAME\"). Hence in the general case, the code will crash, due to this mismatch. I attach his corrected code, with both scenarios correctly treated.<\/p>\n\n<pre><code>inp = tf.placeholder(tf.float32, [MB, H, W, channels_img])\n\n# F has shape (MB, fh, fw, channels, out_channels)\n# REM: with the notation in the question, we need: channels_img==channels\n\nF = tf.transpose(F, [1, 2, 0, 3, 4])\nF = tf.reshape(F, [fh, fw, channels*MB, out_channels)\n\ninp_r = tf.transpose(inp, [1, 2, 0, 3]) # shape (H, W, MB, channels_img)\ninp_r = tf.reshape(inp_r, [1, H, W, MB*channels_img])\n\npadding = \"VALID\" #or \"SAME\"\nout = tf.nn.depthwise_conv2d(\n          inp_r,\n          filter=F,\n          strides=[1, 1, 1, 1],\n          padding=padding) # here no requirement about padding being 'VALID', use whatever you want. \n# Now out shape is (1, H-fh+1, W-fw+1, MB*channels*out_channels), because we used \"VALID\"\n\nif padding == \"SAME\":\n    out = tf.reshape(out, [H, W, MB, channels, out_channels)\nif padding == \"VALID\":\n    out = tf.reshape(out, [H-fh+1, W-fw+1, MB, channels, out_channels)\nout = tf.transpose(out, [2, 0, 1, 3, 4])\nout = tf.reduce_sum(out, axis=3)\n\n# out shape is now (MB, H-fh+1, W-fw+1, out_channels)\n<\/code><\/pre>\n"},{"AnswerId":"42086729","CreationDate":"2017-02-07T09:56:09.920","ParentId":null,"OwnerUserId":"6933420","Title":null,"Body":"<p>They way to go around it is adding an extra dimension using <\/p>\n\n<pre><code>tf.expand_dims(inp, 0)\n<\/code><\/pre>\n\n<p>to create a 'fake' batch size. Then use the <\/p>\n\n<pre><code>tf.nn.conv3d()\n<\/code><\/pre>\n\n<p>operation where the filter-depth matches the batch size. This will result in each filter convolving with only one sample in each batch. <\/p>\n\n<p>Sadly, you will not solve the variable batch size problem this way, only the convolutions. <\/p>\n"},{"AnswerId":"46286109","CreationDate":"2017-09-18T18:39:58.333","ParentId":null,"OwnerUserId":"6002304","Title":null,"Body":"<p>I think the proposed trick is actually not right. What happens with a <code>tf.conv3d()<\/code> layer is that the input gets convolved on depth (=actual batch) dimension AND then summed along resulting feature maps. With <code>padding='SAME'<\/code> the resulting number of outputs then happens to be the same as batch size so one gets fooled!<\/p>\n\n<p>EDIT: I think a possible way to do a convolution with different filters for the different mini-batch elements involves 'hacking' a depthwise convolution. Assuming batch size <code>MB<\/code> is known: <\/p>\n\n<pre><code>inp = tf.placeholder(tf.float32, [MB, H, W, channels_img])\n\n# F has shape (MB, fh, fw, channels, out_channels)\n# REM: with the notation in the question, we need: channels_img==channels\n\nF = tf.transpose(F, [1, 2, 0, 3, 4])\nF = tf.reshape(F, [fh, fw, channels*MB, out_channels)\n\ninp_r = tf.transpose(inp, [1, 2, 0, 3]) # shape (H, W, MB, channels_img)\ninp_r = tf.reshape(inp, [1, H, W, MB*channels_img])\n\nout = tf.nn.depthwise_conv2d(\n          inp_r,\n          filter=F,\n          strides=[1, 1, 1, 1],\n          padding='VALID') # here no requirement about padding being 'VALID', use whatever you want. \n# Now out shape is (1, H, W, MB*channels*out_channels)\n\nout = tf.reshape(out, [H, W, MB, channels, out_channels) # careful about the order of depthwise conv out_channels!\nout = tf.transpose(out, [2, 0, 1, 3, 4])\nout = tf.reduce_sum(out, axis=3)\n\n# out shape is now (MB, H, W, out_channels)\n<\/code><\/pre>\n\n<p>In case <code>MB<\/code> is unknown, it should be possible to determine it dynamically using <code>tf.shape()<\/code> (I think)<\/p>\n"}]}
{"QuestionId":42069147,"AnswerCount":1,"Tags":"<parallel-processing><tensorflow><distributed>","CreationDate":"2017-02-06T13:35:55.363","AcceptedAnswerId":42081222.0,"OwnerUserId":4060761.0,"Title":"Implementation of model parallelism in tensorflow","Body":"<p>I am a beginner to tensorflow. I'm currently working on a system with 2 GPUs each of 12GB. I want to implement model parallelism across the two GPUs to train large models. I have been looking through all over the internet, SO, tensorflow documentation, etc, i was able to find the explanations of model parallelism and its results but nowhere did i find a small tutorial or small code snippets on how to implement it using tensorflow. I mean we have to exchange activations after every layer right? So how do we do that? Is there a specific or cleaner ways of implementing model parallelism in tensorflow? It would be very helpful if you could suggest me a place where i can learn to implement it or a simple code like mnist training on multiple GPU using 'MODEL PARALLELISM'. <\/p>\n\n<p>Note: I have done data parallelism like in CIFAR10 - multi gpu tutorial but i haven't found any implementation of model parallelism.<\/p>\n","answers":[{"AnswerId":"42081222","CreationDate":"2017-02-07T03:30:19.403","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's an example. The model has some parts on GPU0, some parts on GPU1 and some parts on CPU, so this is 3 way model parallelism.<\/p>\n\n<pre><code>with tf.device(\"\/gpu:0\"):\n    a = tf.Variable(tf.ones(()))\n    a = tf.square(a)\nwith tf.device(\"\/gpu:1\"):\n    b = tf.Variable(tf.ones(()))\n    b = tf.square(b)\nwith tf.device(\"\/cpu:0\"):\n    loss = a+b\nopt = tf.train.GradientDescentOptimizer(learning_rate=0.1)\ntrain_op = opt.minimize(loss)\n\nsess = tf.Session()\nsess.run(tf.global_variables_initializer())\nfor i in range(10):\n    loss0, _ = sess.run([loss, train_op])\n    print(\"loss\", loss0)\n<\/code><\/pre>\n"}]}
{"QuestionId":42069231,"AnswerCount":1,"Tags":"<android><tensorflow><bazel>","CreationDate":"2017-02-06T13:40:32.890","AcceptedAnswerId":42312912.0,"OwnerUserId":1545917.0,"Title":"Are linker flags (linkopts) propagats in Bazel?","Body":"<p>I'm building an Android library based on Tensorflow with Bazel.<\/p>\n\n<p>Here the  file<\/p>\n\n<pre><\/pre>\n\n<p>The linker complains about not finding , while I didn't add this flag to .<\/p>\n\n<p>I've checked the executed command, and in fact there is extra flags on it: <\/p>\n\n<p>Where did they came from ? Is there a workaround for this ?<\/p>\n","answers":[{"AnswerId":"42312912","CreationDate":"2017-02-18T08:29:25.433","ParentId":null,"OwnerUserId":"1545917","Title":null,"Body":"<p>I got the answer from the tensorflow's issue tracker.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7288#issuecomment-277761166\" rel=\"nofollow noreferrer\">@jart<\/a><\/p>\n\n<blockquote>\n  <p>Since copts and linkopts are viral and propagate to dependencies, -lpthread is most likely being inherited from @boringssl\/\/:crypto<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7288#issuecomment-277801316\" rel=\"nofollow noreferrer\">@andrewharp<\/a><\/p>\n\n<blockquote>\n  <p>-lpthread is not necessary or possible on Android, so it sounds like the solution would be to add another condition for the select statement as in the linked commit <a href=\"https:\/\/github.com\/google\/protobuf\/pull\/1386\" rel=\"nofollow noreferrer\">google\/protobuf#1386:<\/a><\/p>\n  \n  <p>...<\/p>\n  \n  <p>The only other workaround I know of that doesn't require editing the other repository is to create a dummy libpthread.so target, but that's pretty hacky.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42069304,"AnswerCount":0,"Tags":"<tensorflow><bmp>","CreationDate":"2017-02-06T13:45:01.820","AcceptedAnswerId":null,"OwnerUserId":806160.0,"Title":"reading BMP file with tensorflow","Body":"<p>I want read a image with tensorflow that hase shape of 100*100*3= 30000.<\/p>\n\n<p>I use this code :<\/p>\n\n<pre><\/pre>\n\n<p>the result is 30054. \nI don know why can solve this problem. or there is other way to read bmp files?<\/p>\n\n<p>My solution:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42070153,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-06T14:28:41.217","AcceptedAnswerId":null,"OwnerUserId":3425890.0,"Title":"Detecting images using tensorflow","Body":"<p>i am actually working on a deep learning program and i try to develop an convolutional neural network for image recognition, my problem is how can we identify in the input pictures the differents patterns that we try to extract?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":42070528,"AnswerCount":4,"Tags":"<image-processing><machine-learning><deep-learning><caffe><conv-neural-network>","CreationDate":"2017-02-06T14:45:51.793","AcceptedAnswerId":42071370.0,"OwnerUserId":6786718.0,"Title":"What does global pooling do?","Body":"<p>I recently found the \"global_pooling\" flag in the Pooling layer in caffe, however was unable to find sth about it in the documentation here (<a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers\/pooling.html\" rel=\"noreferrer\">Layer Catalogue)<\/a>\nnor here (<a href=\"http:\/\/caffe.berkeleyvision.org\/doxygen\/classcaffe_1_1PoolingLayer.html\" rel=\"noreferrer\">Pooling doxygen doc<\/a>) . <\/p>\n\n<p>Is there an easy forward examply explanation to this in comparison to the normal Pool-Layer behaviour?<\/p>\n","answers":[{"AnswerId":"42090174","CreationDate":"2017-02-07T12:40:58.163","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>Convolutions can work on any image input size (which is big enough). However, if you have a fully connected layer at the end, this layer needs a fixed input size. Hence the complete network needs a fixed image input size.<\/p>\n\n<p>However, you can remove the fully connected layer and just work with convolutional layers. You can make a convolutional layer at the end which has the same number of filters as you have classes. But you want one value for each class which indicates the probability of that class. Hence you apply a pooling filter over the complete remaining feature map. This pooling is hence \"global\" as it always is as big as necessary. In contrast, usual pooling layers have a fixed size (e.g. of 2x2 or 3x3).<\/p>\n\n<p>This is a general concept. You can also find global pooling in other libraries, e.g. <a href=\"http:\/\/lasagne.readthedocs.io\/en\/latest\/modules\/layers\/pool.html#lasagne.layers.GlobalPoolLayer\" rel=\"nofollow noreferrer\">Lasagne<\/a>. If you want a good reference in literature, I recommend reading <a href=\"https:\/\/arxiv.org\/pdf\/1312.4400v3.pdf\" rel=\"nofollow noreferrer\">Network In Network<\/a>.<\/p>\n"},{"AnswerId":"42071561","CreationDate":"2017-02-06T15:36:15.160","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you are looking for information regarding flags\/parameters of caffe, it is best look them up in the comments of <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto\" rel=\"nofollow noreferrer\"><code>'$CAFFE_ROOT\/src\/caffe\/proto\/caffe.proto'<\/code><\/a>.<br>\nFor <code>'global_pooling'<\/code> parameter the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L911-L913\" rel=\"nofollow noreferrer\">comment says<\/a>:<\/p>\n\n<blockquote>\n<pre><code>\/\/ If global_pooling then it will pool over the size of the bottom by doing\n\/\/ kernel_h = bottom-&gt;height and kernel_w = bottom-&gt;width\n<\/code><\/pre>\n<\/blockquote>\n\n<p>For more information about caffe layers, see <a href=\"http:\/\/caffe.help\/manual\/layers\/pooling.html\" rel=\"nofollow noreferrer\">this help pages<\/a>.<\/p>\n"},{"AnswerId":"42071370","CreationDate":"2017-02-06T15:25:57.243","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>With Global pooling reduces the dimensionality from 3D to 1D. Therefore Global pooling outputs 1 response for every feature map. This can be the maximum or the average or whatever other pooling operation you use.<\/p>\n\n<p>It is often used at the end of the backend of a convolutional neural network to get a shape that works with dense layers. Therefore no flatten has to be applied.<\/p>\n"},{"AnswerId":"51276995","CreationDate":"2018-07-11T03:52:03.300","ParentId":null,"OwnerUserId":"4466275","Title":null,"Body":"<p>We get only one value from entire feature map when we apply GP layer, in which kernel size is the h\u00d7w of the feature map. GP layers are used to reduce the spatial dimensions of a three-dimensional feature map. However, GP layers perform a more extreme type of dimensionality reduction, where a feature map with dimensions h\u00d7w\u00d7d is reduced in size to have dimensions 1\u00d71\u00d7d. GP layers reduce each h\u00d7w feature map to a single number by simply taking the average of all hw values.<\/p>\n"}]}
{"QuestionId":42071074,"AnswerCount":1,"Tags":"<tensorflow><lstm>","CreationDate":"2017-02-06T15:12:18.863","AcceptedAnswerId":null,"OwnerUserId":7523983.0,"Title":"Multidimensional lstm tensorflow","Body":"<p>Can someone suggest an improvement on my implementation of multi-dimensional lstm?<\/p>\n\n<p>It is very slow and uses a lot of memory.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42508243","CreationDate":"2017-02-28T12:00:24.177","ParentId":null,"OwnerUserId":"7523983","Title":null,"Body":"<pre><code>def ln(tensor, scope = None, epsilon = 1e-5):\n    \"\"\" Layer normalizes a 2D tensor along its second axis \"\"\"\n    assert(len(tensor.get_shape()) == 2)\n    m, v = tf.nn.moments(tensor, [1], keep_dims=True)\n    if not isinstance(scope, str):\n        scope = ''\n    with tf.variable_scope(scope + 'layer_norm'):\n        scale = tf.get_variable('scale',\n                                shape=[tensor.get_shape()[1]],\n                                initializer=tf.constant_initializer(1))\n        shift = tf.get_variable('shift',\n                                shape=[tensor.get_shape()[1]],\n                                initializer=tf.constant_initializer(0))\n    LN_initial = (tensor - m) \/ tf.sqrt(v + epsilon)\n\n    return LN_initial * scale + shift\n\n\nclass MultiDimentionalLSTMCell(tf.nn.rnn_cell.RNNCell):\n    \"\"\"\n    Adapted from TF's BasicLSTMCell to use Layer Normalization.\n    Note that state_is_tuple is always True.\n    \"\"\"\n\n    def __init__(self, num_units, forget_bias=0.0, activation=tf.nn.tanh):\n        self._num_units = num_units\n        self._forget_bias = forget_bias\n        self._activation = activation\n\n    @property\n    def state_size(self):\n        return tf.nn.rnn_cell.LSTMStateTuple(self._num_units, self._num_units)\n\n    @property\n    def output_size(self):\n        return self._num_units\n\n    def __call__(self, inputs, state, scope=None):\n        \"\"\"Long short-term memory cell (LSTM).\n        @param: imputs (batch,n)\n        @param state: the states and hidden unit of the two cells\n        \"\"\"\n        with tf.variable_scope(scope or type(self).__name__):\n            c1,c2,h1,h2 = state\n\n            # change bias argument to False since LN will add bias via shift\n            concat = tf.nn.rnn_cell._linear([inputs, h1, h2], 5 * self._num_units, False)\n\n            i, j, f1, f2, o = tf.split(1, 5, concat)\n\n            # add layer normalization to each gate\n            i =  ln(i, scope = 'i\/')\n            j =  ln(j, scope = 'j\/')\n            f1 = ln(f1, scope = 'f1\/')\n            f2 = ln(f2, scope = 'f2\/')\n            o =  ln(o, scope = 'o\/')\n\n            new_c = (c1 * tf.nn.sigmoid(f1 + self._forget_bias) + \n                     c2 * tf.nn.sigmoid(f2 + self._forget_bias) + tf.nn.sigmoid(i) *\n                   self._activation(j))\n\n            # add layer_normalization in calculation of new hidden state\n            new_h = self._activation(ln(new_c, scope = 'new_h\/')) * tf.nn.sigmoid(o)\n            new_state = tf.nn.rnn_cell.LSTMStateTuple(new_c, new_h)\n\n            return new_h, new_state\n\n\ndef multiDimentionalRNN_whileLoop(rnn_size,input_data,sh,dims=None,scopeN=\"layer1\"):\n        \"\"\"Implements naive multidimentional recurent neural networks\n\n        @param rnn_size: the hidden units\n        @param input_data: the data to process of shape [batch,h,w,chanels]\n        @param sh: [heigth,width] of the windows \n        @param dims: dimentions to reverse the input data,eg.\n            dims=[False,True,True,False] =&gt; true means reverse dimention\n        @param scopeN : the scope\n\n        returns [batch,h\/sh[0],w\/sh[1],chanels*sh[0]*sh[1]] the output of the lstm\n        \"\"\"\n        with tf.variable_scope(\"MultiDimentionalLSTMCell-\"+scopeN):\n            cell = MultiDimentionalLSTMCell(rnn_size)\n\n            shape = input_data.get_shape().as_list()\n\n            if shape[1]%sh[0] != 0:\n                offset = tf.zeros([shape[0], sh[0]-(shape[1]%sh[0]), shape[2], shape[3]])\n                input_data = tf.concat(1,[input_data,offset])\n                shape = input_data.get_shape().as_list()\n            if shape[2]%sh[1] != 0:\n                offset = tf.zeros([shape[0], shape[1], sh[1]-(shape[2]%sh[1]), shape[3]])\n                input_data = tf.concat(2,[input_data,offset])\n                shape = input_data.get_shape().as_list()\n\n            h,w = int(shape[1]\/sh[0]),int(shape[2]\/sh[1])\n            features = sh[1]*sh[0]*shape[3]\n            batch_size = shape[0]\n\n            x =  tf.reshape(input_data, [batch_size,h,w, features])\n            if dims is not None:\n                assert dims[0] == False and dims[3] == False\n                x = tf.reverse(x, dims)\n            x = tf.transpose(x, [1,2,0,3])\n            x =  tf.reshape(x, [-1, features])\n            x = tf.split(0, h*w, x)     \n\n            sequence_length = tf.ones(shape=(batch_size,), dtype=tf.int32)*shape[0]\n            inputs_ta = tf.TensorArray(dtype=tf.float32, size=h*w,name='input_ta')\n            inputs_ta = inputs_ta.unpack(x)\n            states_ta = tf.TensorArray(dtype=tf.float32, size=h*w+1,name='state_ta',clear_after_read=False)\n            outputs_ta = tf.TensorArray(dtype=tf.float32, size=h*w,name='output_ta')\n\n            states_ta = states_ta.write(h*w,  tf.nn.rnn_cell.LSTMStateTuple(tf.zeros([batch_size,rnn_size], tf.float32),\n                                                         tf.zeros([batch_size,rnn_size], tf.float32)))\n            def getindex1(t,w):\n                return tf.cond(tf.less_equal(tf.constant(w),t),\n                               lambda:t-tf.constant(w),\n                               lambda:tf.constant(h*w))\n            def getindex2(t,w):\n                return tf.cond(tf.less(tf.constant(0),tf.mod(t,tf.constant(w))),\n                               lambda:t-tf.constant(1),\n                               lambda:tf.constant(h*w))\n\n            time = tf.constant(0)\n\n            def body(time, outputs_ta, states_ta):\n                constant_val = tf.constant(0)\n                stateUp = tf.cond(tf.less_equal(tf.constant(w),time),\n                                  lambda: states_ta.read(getindex1(time,w)),\n                                  lambda: states_ta.read(h*w))\n                stateLast = tf.cond(tf.less(constant_val,tf.mod(time,tf.constant(w))),\n                                    lambda: states_ta.read(getindex2(time,w)),\n                                    lambda: states_ta.read(h*w)) \n\n                currentState = stateUp[0],stateLast[0],stateUp[1],stateLast[1]\n                out , state = cell(inputs_ta.read(time),currentState)  \n                outputs_ta = outputs_ta.write(time,out)\n                states_ta = states_ta.write(time,state)\n                return time + 1, outputs_ta, states_ta\n\n            def condition(time,outputs_ta,states_ta):\n                return tf.less(time ,  tf.constant(h*w)) \n\n            result , outputs_ta, states_ta = tf.while_loop(condition, body, [time,outputs_ta,states_ta]\n                                                           ,parallel_iterations=1)\n\n\n            outputs = outputs_ta.pack()\n            states  = states_ta.pack()\n\n            y =  tf.reshape(outputs, [h,w,batch_size,rnn_size])\n            y = tf.transpose(y, [2,0,1,3])\n            if dims is not None:\n                y = tf.reverse(y, dims)\n\n            return y,states\n<\/code><\/pre>\n"}]}
{"QuestionId":42071538,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-06T15:35:11.587","AcceptedAnswerId":null,"OwnerUserId":7506591.0,"Title":"Tensorflow import_meta_graph returns 'tensor does not exist' error","Body":"<p>I am trying to import a saved neural network in Tensorflow. I saved it after training with:<\/p>\n\n<pre><\/pre>\n\n<p>and in the script I use for inference, I restore it with:<\/p>\n\n<pre><\/pre>\n\n<p>But during the  line, I get this error:<\/p>\n\n<blockquote>\n  <p>KeyError: \"The name 'dropout1\/cond\/dropout\/Shape\/Switch:1' refers to a Tensor which does not exist. The operation, 'dropout1\/cond\/dropout\/Shape\/Switch', does not exist in the graph.\"<\/p>\n<\/blockquote>\n\n<p>I looked at the names of the tensors and operations in the original notebook in which I trained the model, and the names mentionned in the error message do exist. Moreover, I used the same code for saving and importing other models and it works. The only difference is that I trained these on an AWS machine, with an older version of tensorflow, while I trained the problematic one on my computer.<\/p>\n","answers":[]}
{"QuestionId":42072024,"AnswerCount":1,"Tags":"<python><numpy><anaconda><keras><keras-layer>","CreationDate":"2017-02-06T15:58:59.883","AcceptedAnswerId":null,"OwnerUserId":7103753.0,"Title":"ImportError: No module named numpy.distutils.core (during keras installation)","Body":"<p>I want to install keras module (<a href=\"https:\/\/pypi.python.org\/pypi\/Keras\" rel=\"nofollow noreferrer\">github-link<\/a>). I tried .It installed 1.2.1. It does not have Graph module as per <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/4661\" rel=\"nofollow noreferrer\">this<\/a>. Sadly the documentation is also is not updated. So I tried a previous version 1.1.2.and as per the answer by <a href=\"https:\/\/stackoverflow.com\/users\/4802797\/chang-liu\">Chang Liu<\/a> given <a href=\"https:\/\/stackoverflow.com\/a\/35911254\/7103753\">here<\/a> ran  but now I am getting following error<\/p>\n\n<pre><\/pre>\n\n<p>I have already installed anaconda and during its installation it correctly installed numpy as well.Another thing that I tried was  as according to <a href=\"https:\/\/keras.io\/\" rel=\"nofollow noreferrer\">documentation<\/a>  there is class Graph() present in container.Please let me know what should I do?<\/p>\n","answers":[{"AnswerId":"42074168","CreationDate":"2017-02-06T17:53:18.273","ParentId":null,"OwnerUserId":"539075","Title":null,"Body":"<p>As the author from the Issue mentioned <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/4661\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>Graph has been removed. Use the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">functional API (see docs).<\/a>:<\/p>\n\n<blockquote>\n  <p>The Keras functional API is the way to go for defining complex models, such as multi-output models, directed acyclic graphs, or models with shared layers.<\/p>\n<\/blockquote>\n\n<p>more clarification:<\/p>\n\n<ol>\n<li><p>This error could be due to the dependency installations, make sure <code>numpy<\/code> and <code>scikit-learn<\/code> are not having issues during the process.<\/p><\/li>\n<li><p>Perhaps due a fresh install of the conda if this is still an issue.<\/p><\/li>\n<\/ol>\n\n<p>and for the Graph, the first answer still stands, it has been removed.<\/p>\n"}]}
{"QuestionId":42072234,"AnswerCount":2,"Tags":"<python><variables><tensorflow><neural-network><deep-learning>","CreationDate":"2017-02-06T16:08:19.463","AcceptedAnswerId":42073677.0,"OwnerUserId":6182430.0,"Title":"TensorFlow - import meta graph and use variables from it","Body":"<p>I'm training classification CNN using TensorFlow v0.12, and then want to create labels for new data using the trained model.<\/p>\n\n<p>At the end of the training script, I added those lines of code:<\/p>\n\n<pre><\/pre>\n\n<p>After the training completed, the files appearing in the folder are: 1. <em>checkpoint<\/em> ; 2. <em>model.ckpt.data-00000-of-00001<\/em> ; 3. <em>model.ckpt.index<\/em> ; 4. <em>model.ckpt.meta<\/em><\/p>\n\n<p>Then I tried to restore the model using the <em>.meta<\/em> file. Following <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/how_tos\/meta_graph\/\" rel=\"noreferrer\">this tutorial<\/a>, I added the following line into my classification code:<\/p>\n\n<pre><\/pre>\n\n<p>and then:<\/p>\n\n<pre><\/pre>\n\n<p>Before that change, I needed to build the graph again, and then write (instead of line1):<\/p>\n\n<pre><\/pre>\n\n<p>But, deleting the graph building, and using  in order to restore it, raised an error. The error was that I used a variable from the graph inside my code, and the python didn't recognize it:<\/p>\n\n<pre><\/pre>\n\n<p>The python didn't recognize the  parameter. There is a way to restore the variables using the meta graph? if not, what os this restore helping, if I can't use variables from the original graph?<\/p>\n\n<p>I know this question isn't so clear, but it was hard for me to express the problem in words. Sorry about it...<\/p>\n\n<p>Thanks for answering, appreciate your help! Roi.<\/p>\n","answers":[{"AnswerId":"43569135","CreationDate":"2017-04-23T08:50:11.710","ParentId":null,"OwnerUserId":"651945","Title":null,"Body":"<p>Just to add to Roberts's answer - after obtaining a saver from the meta graph, and using it to restore the variables in the current session, you can also use:<\/p>\n\n<pre><code>y_conv = graph.get_tensor_by_name('y_conv:0')\n<\/code><\/pre>\n\n<p>This'll work if you've created the <code>y_conv<\/code> with explicitly adding the <code>name=\"y_conv\"<\/code> argument (all TF ops have this).<\/p>\n"},{"AnswerId":"42073677","CreationDate":"2017-02-06T17:24:21.363","ParentId":null,"OwnerUserId":"6933420","Title":null,"Body":"<p>it is possible, don't worry. Assuming you don't want to touch the graph anymore, do something like this:<\/p>\n\n<pre><code>saver = tf.train.import_meta_graph('model\/export\/{}.meta'.format(model_name))\nsaver.restore(sess, 'model\/export\/{}'.format(model_name))\ngraph = tf.get_default_graph()       \ny_conv = graph.get_operation_by_name('y_conv').outputs[0]\npredictions = sess.run(y_conv, feed_dict={x: patches,keep_prob: 1.0})\n<\/code><\/pre>\n\n<p>A preferred way would however be adding the ops into collections when you build the graph and then referring to them. So when you define the graph, you would add the line: <\/p>\n\n<pre><code>tf.add_to_collection(\"y_conv\", y_conv)\n<\/code><\/pre>\n\n<p>And then after you import the metagraph and restore it, you would call:<\/p>\n\n<pre><code>y_conv = tf.get_collection(\"y_conv\")[0]\n<\/code><\/pre>\n\n<p>It is actually explained in the documentation - the exact page you linked - but perhaps you missed it. <\/p>\n\n<p>Btw, no need for the <code>.ckpt<\/code> extension, it might create some confusion as that is the old way of saving models. <\/p>\n"}]}
{"QuestionId":42073239,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-06T16:59:29.350","AcceptedAnswerId":null,"OwnerUserId":1817595.0,"Title":"tf.get_collection to extract variables of one scope","Body":"<p>I have  (e.g: n=3) scopes and  (e.g: x=4) no of Variables defined in each scope.\nThe scopes are:<\/p>\n\n<pre><\/pre>\n\n<p>Once I compute the loss, I want to extract and provide all the variables from only one of the scope based on a criteria during run-time. Hence the index of the scope  that I select is an argmin tensor cast into int32<\/p>\n\n<pre><\/pre>\n\n<p>I have already tried:<\/p>\n\n<pre><\/pre>\n\n<p>which obviously did not work.\nIs there any way to get all the  Variables belonging to that particular scope using idx to pass into the optimizer.<\/p>\n\n<p>Thanks in advance!<\/p>\n\n<p>Vignesh Srinivasan<\/p>\n","answers":[{"AnswerId":"42074009","CreationDate":"2017-02-06T17:43:11.770","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You can do something like this in TF 1.0 rc1 or later:<\/p>\n\n<pre><code>v = tf.Variable(tf.ones(()))\nloss = tf.identity(v)\nwith tf.variable_scope('adamoptim') as vs:\n   optim = tf.train.AdamOptimizer(learning_rate=0.1).minimize(loss)\noptim_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=vs.name)\nprint([v.name for v in optim_vars]) #=&gt; prints lists of vars created\n<\/code><\/pre>\n"}]}
{"QuestionId":42073336,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-06T17:04:14.577","AcceptedAnswerId":null,"OwnerUserId":7490399.0,"Title":"\"import tensorflow\" terminated by signal SIGSEGV (Address boundary error)","Body":"<p>An  statement triggers the following error:<\/p>\n\n<pre>\n>>> <b>import tensorflow as tf<\/b>\nI tensorflow\/stream_executor\/dso_loader.cc:108] successfully opened CUDA library libcublas.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:108] successfully opened CUDA library libcudnn.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:108] successfully opened CUDA library libcufft.dylib locally\n\"import tensorflow\" terminated by signal SIGSEGV (Address boundary error)\n<\/pre>\n","answers":[{"AnswerId":"42073376","CreationDate":"2017-02-06T17:06:43.307","ParentId":null,"OwnerUserId":"7490399","Title":null,"Body":"<p>By default, CUDA creates <code>libcuda.dylib<\/code>, but TensorFlow tries to load\n<code>libcuda.1.dylib<\/code>. To fix the problem, create a symbolic link between\n<code>libcuda.dylib<\/code> and <code>libcuda.1.dylib<\/code>. For example, if CUDA is installed\nin <code>\/usr\/local\/cuda<\/code>, issue the following <code>ln<\/code> command:<\/p>\n\n<pre><code>ln -sf \/usr\/local\/cuda\/lib\/libcuda.dylib \/usr\/local\/cuda\/lib\/libcuda.1.dylib\n<\/code><\/pre>\n"}]}
{"QuestionId":42073768,"AnswerCount":2,"Tags":"<mobile><machine-learning><tensorflow><voice-recognition>","CreationDate":"2017-02-06T17:29:01.333","AcceptedAnswerId":null,"OwnerUserId":7524599.0,"Title":"Can TensoFlow learn unstructured data?","Body":"<p>I'm new to TensoFlow and I'd like to know if it is possible to make a tensor flow algorithm do a deep learning in a website or in a database, for instance.\nI wish to make and App for iOS using an API, witch someone can ask something about a course in a university for instance and the software would be able to answer.\nThanks!<\/p>\n","answers":[{"AnswerId":"42075663","CreationDate":"2017-02-06T19:22:21.090","ParentId":null,"OwnerUserId":"6248065","Title":null,"Body":"<p>First of all Tenserflow is not an algorithm , it is  an open source software library for machine learning in various kinds of perceptual and language understanding tasks. \nI think you can refer the algorithms that are used in chatbots developed by the team of Google i.e. 'Google Brain' . You can refer to this link (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/seq2seq\/<\/a>) . I know I did not provide you the complete answer but you can take it as a start for your task .<\/p>\n"},{"AnswerId":"50622142","CreationDate":"2018-05-31T10:28:12.350","ParentId":null,"OwnerUserId":"4112804","Title":null,"Body":"<p>Well, tensorflow is a framework which you can use to develop algorithms as well as to make your computations easier.SO it's in your hand, you can develop anything out of it. And yes if you develop your model with mobile support in mind you can run the trained model in your mobile device.<\/p>\n"}]}
{"QuestionId":42074595,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2017-02-06T18:19:19.830","AcceptedAnswerId":42075122.0,"OwnerUserId":6409135.0,"Title":"Convolution neural network output value in range","Body":"<p>i'm trying to learn about convolution neural network using tensorflow , i'm using the code bellow to create my network .<\/p>\n\n<pre><\/pre>\n\n<p>this network take a [80,80,4] object an output a ACTIONS prediction ;(ACTIONS number of classes)<\/p>\n\n<p>My question is how to create a network that output a value between [a,b]<\/p>\n","answers":[{"AnswerId":"42075122","CreationDate":"2017-02-06T18:53:03.063","ParentId":null,"OwnerUserId":"3260576","Title":null,"Body":"<p>if [a,b] is categorical, i.e. you have discrete classes (say n in total), you need to have n neurons as in readout layer, each corresponding to a given class representing its probability. Standard way to do it is using sigmoid cross entropy on logits (using tf.nn.sigmoid_cross_entropy_with_logits in case of tensorflow).<\/p>\n\n<p>In case [a,b] is continuous (which I believe is probably not what you are looking to do) than it becomes a regression problem. So, given sufficient data and an appropriate loss function (squared errors) along with some regularization should ensure your output lies in acceptable range or you can always map one numeric range to another.<\/p>\n\n<p>If none of this makes sense, you need to study some basic machine learning before you dive into neural nets.<\/p>\n"}]}
{"QuestionId":42075397,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-06T19:08:26.767","AcceptedAnswerId":null,"OwnerUserId":7490399.0,"Title":"pip install command causes Broken toolchain error","Body":"<p>Installing TensorFlow with  generates several messages, ending with a  error.  Here's part of the transcript:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42075590","CreationDate":"2017-02-06T19:19:19.737","ParentId":null,"OwnerUserId":"7490399","Title":null,"Body":"<p>The error indicates that the \n<a href=\"https:\/\/developer.apple.com\/xcode\/\" rel=\"nofollow noreferrer\">Xcode<\/a> build tools are installed on your system, \nbut you haven't yet accepted the license agreements. To resolve this issue, do either of the following:<\/p>\n\n<ol>\n<li>Open <a href=\"https:\/\/developer.apple.com\/xcode\/\" rel=\"nofollow noreferrer\">Xcode<\/a> and\naccept the license agreement.<\/li>\n<li><p>Invoke the following command: <\/p>\n\n<pre><code>xcodebuild -license\n<\/code><\/pre>\n\n<p>See \n<a href=\"https:\/\/developer.apple.com\/legacy\/library\/documentation\/Darwin\/Reference\/ManPages\/man1\/xcodebuild.1.html\" rel=\"nofollow noreferrer\">xcodebuild documentation<\/a> for more details.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42075436,"AnswerCount":1,"Tags":"<machine-learning><nan><caffe>","CreationDate":"2017-02-06T19:10:34.147","AcceptedAnswerId":null,"OwnerUserId":1850692.0,"Title":"ResNet - NaN during classification (not training) on same dataset used for validation","Body":"<p>Some very quick background:<\/p>\n\n<p>Training on Ubuntu 16.04 LTS with latest Caffe branch on GTX 1080<\/p>\n\n<p>Classification on OSX<\/p>\n\n<p>I'm wondering if anyone can shed light on this issue. I'm finetuning ResNet-50 using caffe from <a href=\"https:\/\/github.com\/cvjena\/cnn-models\" rel=\"nofollow noreferrer\">here<\/a>. I'm experimenting with a small dataset. For training, I have approximately 1248 images and for validation, I have 1656. I am going through the entire validation set each time I test during training.<\/p>\n\n<p>Now, during training, I am not seeing any NaNs or anything that would indicate gradient blow up or something like that. No NaNs, infs, or anything strange. When I load the net and try to classify the <strong>same images<\/strong> used for validation during training, NaNs show up. Any idea why this is? Could it be the difference between the GPU architecture vs the CPU architecture in carrying out the calculations in training vs classification?<\/p>\n\n<p>[EDIT: I'm not doing any preprocessing]<\/p>\n\n<p>[EDIT2: It seems stochastic. If I run the same image through again, the network will properly output a probability or it will output NaN again. My solution currently is to run the same image through the network until a non-NaN shows up, but sometimes, this takes awhile]<\/p>\n","answers":[{"AnswerId":"42075997","CreationDate":"2017-02-06T19:44:30.230","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Maybe different type of preprocessing. You get NaN output for NaN input. Therefore if you somehow divide by 0 with your preprocessing (Normalization?) this can occur. <\/p>\n"}]}
{"QuestionId":42077395,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-02-06T21:16:16.047","AcceptedAnswerId":42077694.0,"OwnerUserId":6446734.0,"Title":"Python\/Tensorflow: Saver for network","Body":"<p>According to the documentation and numerous SO posts regarding this API, the saver object must be created using <\/p>\n\n<pre><\/pre>\n\n<p>I wanted to know if there is any way to automatically populate the  without having to explicitly list all variables and ops used in my network. <\/p>\n\n<p>Right now my network is only two layers so it is not a huge hassle, but it feels downright stone-age like to have to list all the variables manually. <\/p>\n","answers":[{"AnswerId":"42077694","CreationDate":"2017-02-06T21:37:43.037","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables#Saver.__init__\" rel=\"nofollow noreferrer\">default initializer for <code>tf.train.Saver<\/code><\/a> will create an instance that saves\/restores all saveable objects in your graph, which typically includes all of your model variables. Therefore you should be able to write:<\/p>\n\n<pre><code>saver = tf.train.Saver()\n<\/code><\/pre>\n\n<p>\u2026and get the desired effect without too much trouble.<\/p>\n"}]}
{"QuestionId":42078356,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-02-06T22:22:17.057","AcceptedAnswerId":42079545.0,"OwnerUserId":7387749.0,"Title":"Keras - How does to get misclassifcation?","Body":"<p>I have implemented a neural network by Keras, and I have used  to get model's loss and accuracy. Now I would to get the misclassification error, but I don't know how to set the <em>loss<\/em> parameter in  to obtain this kind of evaluation.\nI don't have found anything about it on the Keras website <a href=\"https:\/\/keras.io\/\" rel=\"nofollow noreferrer\">keras.io<\/a><\/p>\n","answers":[{"AnswerId":"42079545","CreationDate":"2017-02-07T00:08:41.400","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You need to set the <code>metrics<\/code> parameter in <code>model.compile()<\/code>. A metric is a function that is used to judge the performance of your model, like in your case classification error, but not used for training. Check more <a href=\"https:\/\/keras.io\/metrics\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>In your case you can use: <code>model.compile(..., loss='binary_crossentropy', metrics='binary_accuracy')<\/code><\/p>\n\n<p>If you are doing classification you should use: <code>model.compile(..., loss='categorical_crossentropy', metrics='categorical_accuracy')<\/code><\/p>\n"}]}
{"QuestionId":42079301,"AnswerCount":1,"Tags":"<c++><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-02-06T23:43:04.817","AcceptedAnswerId":null,"OwnerUserId":4868654.0,"Title":"Why does the HDF5 Output Caffe layer write data of seemingly incorrect dimensions?","Body":"<p>I am interested in writing the output of a given Caffe layer to a file. I would like to do this for more than one image, so I have made some modifications to the HDF5 Output layer code to enable the creation of a file for each image which contains the features for each image. Here is the modified SaveBlobs function:<\/p>\n\n<pre><\/pre>\n\n<p>The code almost works great since I am able to successfully create files for each image which do in fact contain data. Unfortunately it seems like the wrong data is being written, as both the dimensions shown in the log and in the resulting output file are incorrect. Here is where I specify my output layer (in the network prototext):<\/p>\n\n<pre><\/pre>\n\n<p>My reason for believing that the incorrect data is being saved is because when I observe the network setup, the dimensions for the Top of conv5_3 are shown as follows:<\/p>\n\n<pre><\/pre>\n\n<p>So great, I expect to have data of dimensions 1 512 14 14. Unfortunately when I run inference on the model I see in my log that the wrong dimensions are showing up:<\/p>\n\n<pre><\/pre>\n\n<p>This shows that not only are the dimensions of the output incorrect, but they also vary between iterations (images)! The dimensions shown in the log match the dimensions of the data written to the h5 files, so the log is accurately describing the behavior of the code. My question is why might this be the case? It would seem that I have set everything up correctly, but there must be something that I'm missing...<\/p>\n","answers":[{"AnswerId":"42117246","CreationDate":"2017-02-08T15:32:41.297","ParentId":null,"OwnerUserId":"4868654","Title":null,"Body":"<p>As @hbaderts has helped me uncover, it turns out that the HDF5 layer is correct and does output data of the right dimensions. My confusion with dimensions was due to the seemingly static definition of input dimensions in the testing version of the network prototext. It turns out that because I am loading in data by using the pycaffe function net.forward(**forward_kwargs) the convolutional layers are scaling themselves to meet the varied input dimensions of my input images. This explains both the larger than expected feature dimensions and the fact that they vary between images.<\/p>\n"}]}
{"QuestionId":42079915,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-07T00:53:51.670","AcceptedAnswerId":null,"OwnerUserId":1044011.0,"Title":"Get sum of a batch of model.evaluate results in TensorFlow","Body":"<p>In a sentence, I have several customized evaluation metrics, a large set of evaluation data that have to run in batches during evaluation, and I want to get the sum of these metrics of all evaluation batches.<\/p>\n\n<p>More details:<\/p>\n\n<p>I build a model using TensorFlow's Estimator class, and it works fine for training and evaluation. The problem is, my evaluation data is too big and I need to split them into batches using a customized .<\/p>\n\n<p>The call looks like<\/p>\n\n<pre><\/pre>\n\n<p>In this case, I'll get the evaluation metric dict (let's call it ) output every step through the terminal output, but  only has the result of the last evaluation batch. <\/p>\n\n<p>Let's assume  actually contains a list of , then what I want is the sum across all the elements in  that has the same key.<\/p>\n\n<p>An example is like<\/p>\n\n<pre><\/pre>\n\n<p>I tried to store the metric ops into a graph collection using , and then use  to get them. But the result of  is always empty. I suppose I need to specify the graph name to it but I'm not sure how to find the graph of an Estimator class.<\/p>\n","answers":[{"AnswerId":"42095734","CreationDate":"2017-02-07T17:01:40.690","ParentId":null,"OwnerUserId":"1044011","Title":null,"Body":"<p>Every time <code>model.evaluate<\/code> get called, it will run the <code>model_fn<\/code> again. So if I create several variables using <code>tf.Variable<\/code> and update the metric score of each evaluation batch to a variable using <code>assign_add<\/code>, then I can successfully get the accumulated metric score for all evaluation batches.<\/p>\n\n<p>And because every time <code>model_fn<\/code> get called, the variables get reset to 0 by the <code>tf.Variable(0., dtype=tf.float32)<\/code> call, previous evaluation will not affect the result of next evaluation.<\/p>\n\n<p>A code sample looks like:<\/p>\n\n<pre><code>hr_mean_rank_var = tf.Variable(0, name=\"hr_mean_rank\", dtype=tf.float32, trainable=False)\nhr_mr_update = hr_mean_rank_var.assign_add(mean_rank_eval_op)\neval_metric_dict = {'hr_mr' : hr_mr_update}\n<\/code><\/pre>\n\n<p>This time when I call <\/p>\n\n<pre><code>eval_res = model.evaluate(input_fn=lambda:eval_input_fn(params['eval_data'], params['batch_size']))\n<\/code><\/pre>\n\n<p>I'll get the accumulated metric scores.<\/p>\n"}]}
{"QuestionId":42080215,"AnswerCount":0,"Tags":"<tensorflow><padding><convolution>","CreationDate":"2017-02-07T01:35:00.403","AcceptedAnswerId":null,"OwnerUserId":1293689.0,"Title":"Optimized Custom Padding in Tensorflow Conv2d","Body":"<p>When I do a conv2d step, how do I pad my images (spatially) symmetrically as done for example in Caffe and cuDNN, as opposed to Tensorflow's implementation:<\/p>\n\n<p>From <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/nn\/convolution\" rel=\"nofollow noreferrer\">Tensorflow Convolution Docs<\/a>:<\/p>\n\n<blockquote>\n  <p>Note that the division by 2 means that there might be cases when the\n  padding on both sides (top vs bottom, right vs left) are off by one.\n  In this case, the bottom and right sides always get the one additional\n  padded pixel. For example, when pad_along_height is 5, we pad 2 pixels\n  at the top and 3 pixels at the bottom. Note that this is different\n  from existing libraries such as cuDNN and Caffe, which explicitly\n  specify the number of padded pixels and always pad the same number of\n  pixels on both sides.<\/p>\n<\/blockquote>\n\n<p><strong>Question<\/strong>: How do I use (or hack) Tensorflow in a way that it conforms to the way cuDNN or Caffe does it? I do not want to introduce individual performance-hitting calls , preferably.<\/p>\n\n<hr>\n\n<p>Note this is <strong>not<\/strong> the same question as posed here: <a href=\"https:\/\/stackoverflow.com\/questions\/37659538\/custom-padding-for-convolutions-in-tensorflow\">Custom padding for convolutions in TensorFlow<\/a><\/p>\n","answers":[]}
{"QuestionId":42080835,"AnswerCount":2,"Tags":"<python><windows><tensorflow><pip><install>","CreationDate":"2017-02-07T02:48:00.423","AcceptedAnswerId":null,"OwnerUserId":7016241.0,"Title":"tensorflow pip install on windows > error \" no perl script found in input\"","Body":"<p>im trying to install tensorflow on my computer. I have windows 10. Im doing CPU only pip install. When i paste this (pip install --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-0.12.1-cp35-cp35m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-0.12.1-cp35-cp35m-win_amd64.whl<\/a>) command into my command prompt i get error message that says \"no perl script found in input\". What is the problem? is it referring to Perl programming language or something else? how is perl related to any of this? Thanks!!!<\/p>\n","answers":[{"AnswerId":"42895895","CreationDate":"2017-03-20T04:31:28.180","ParentId":null,"OwnerUserId":"7718218","Title":null,"Body":"<p>what python version installed in your environment ?<\/p>\n\n<p>if it's 3.6 you can simply replace the -cp35-cp35m- become -cp36-cp36m-<\/p>\n\n<p>I hope it'll help<\/p>\n"},{"AnswerId":"57530803","CreationDate":"2019-08-16T20:34:11.373","ParentId":null,"OwnerUserId":"1218483","Title":null,"Body":"<p>I just ran into the \"No Perl script found in input\" error myself, in a different situation (as is, not trying to use pip).<\/p>\n\n<p>The original poster has supplied a comment that answers the question he asked; he had two \"pip\"s in his path and the wrong one was before the right one. He either had to change the order of his path or supply a full pathname.<\/p>\n\n<p>This helped me solve my problem; I was using the command line lwp-download to download a file and suddenly got this error. Running lwp-download.bat solved my problem.<\/p>\n\n<p>If you experience the \"No Perl script found in input\" error, try running \"where\" and then your command; in my example \"where lwp-download\". If you have more than one file in your path with that name, try to be specific. In my case, add the \".bat\"<\/p>\n"}]}
{"QuestionId":42081223,"AnswerCount":0,"Tags":"<neural-network><theano><multicore><lasagne>","CreationDate":"2017-02-07T03:30:26.577","AcceptedAnswerId":null,"OwnerUserId":956730.0,"Title":"Theano CPU Multicore lower utilisation on a larger dataset","Body":"<p>I have a Multilayer Perceptron code using Theano\/Lasagne that when I run on a small dataset uses multiple cores correctly.<\/p>\n\n<p>But, when I run it over a much larger dataset (with the same code) and I watch CPU utilisation in htop, I don't see it parallelize. It creates the number of processes as defined in <em>.theanorc<\/em> file which looks like:<\/p>\n\n<pre><\/pre>\n\n<p>but most of the time (~ 90%) there is only one of the created processes is working (though some short amount of times utilisation goes up).<\/p>\n\n<p>I imagine there is an operation that is not utilising multi-core while other operations are using it because when I run the code on a small dataset I see that most of the time all cores are utilised. All the operations are matrix multiplication (both sparse and dense) so I don't\nknow why some operations don't use multi-core. <\/p>\n\n<p>This is the section of the code that is being run:<\/p>\n\n<pre><\/pre>\n\n<p>My numpy\/blas information:<\/p>\n\n<pre><\/pre>\n\n<p>Note that the input matrices are sparse and I'm doing some extra sparse multiplications within custom layers using S.dot.<\/p>\n","answers":[]}
{"QuestionId":42081257,"AnswerCount":11,"Tags":"<machine-learning><keras><neural-network><deep-learning><conv-neural-network>","CreationDate":"2017-02-07T03:34:55.963","AcceptedAnswerId":46038271.0,"OwnerUserId":1291820.0,"Title":"Keras binary_crossentropy vs categorical_crossentropy performance?","Body":"<p>I'm trying to train a CNN to categorize text by topic. When I use binary crossentropy I get ~80% accuracy, with categorical crossentropy I get ~50% accuracy.<\/p>\n\n<p>I don't understand why this is. It's a multiclass problem, doesn't that mean that I have to use categorical cross-entropy and that the results with binary cross-entropy are meaningless?<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>Then I compile it either it like this using  as the loss function:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>or <\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42098861","CreationDate":"2017-02-07T19:59:55.780","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>It's really interesting case. Actually in your setup the following statement is true:<\/p>\n\n<pre><code>binary_crossentropy = len(class_id_index) * categorical_crossentropy\n<\/code><\/pre>\n\n<p>This means that up to a constant multiplication factor your losses are equivalent. The weird behaviour that you are observing during a training phase might be an example of a following phenomenon:<\/p>\n\n<ol>\n<li>At the beginning the most frequent class is dominating the loss - so network is learning to predict mostly this class for every example.<\/li>\n<li>After it learnt the most frequent pattern it starts discriminating among less frequent classes. But when you are using <code>adam<\/code> - the learning rate has a much smaller value than it had at the beginning of training (it's because of the nature of this optimizer). It makes training slower and prevents your network from e.g. leaving a poor local minimum less possible.<\/li>\n<\/ol>\n\n<p>That's why this constant factor might help in case of <code>binary_crossentropy<\/code>. After many epochs - the learning rate value is greater than in <code>categorical_crossentropy<\/code> case. I usually restart training (and learning phase) a few times when I notice such behaviour or\/and adjusting a class weights using the following pattern:<\/p>\n\n<pre><code>class_weight = 1 \/ class_frequency\n<\/code><\/pre>\n\n<p>This makes loss from a less frequent classes balancing the influence of a dominant class loss at the beginning of a training and in a further part of an optimization process.<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>Actually - I checked that even though in case of maths:<\/p>\n\n<pre><code>binary_crossentropy = len(class_id_index) * categorical_crossentropy\n<\/code><\/pre>\n\n<p>should hold - in case of <code>keras<\/code> it's not true, because <code>keras<\/code> is automatically normalizing all outputs to sum up to <code>1<\/code>. This is the actual reason behind this weird behaviour as in case of multiclassification such normalization harms a training.<\/p>\n"},{"AnswerId":"55708351","CreationDate":"2019-04-16T12:38:28.080","ParentId":null,"OwnerUserId":"7338066","Title":null,"Body":"<p>You are passing a target array of shape (x-dim, y-dim) 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:<\/p>\n\n<pre><code>from keras.utils import to_categorical\ny_binary = to_categorical(y_int)\n<\/code><\/pre>\n\n<p>Alternatively, you can use the loss function <code>sparse_categorical_crossentropy<\/code> instead, which does expect integer targets.<\/p>\n\n<pre><code>model.compile(loss='sparse_categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n"},{"AnswerId":"44498955","CreationDate":"2017-06-12T12:02:11.277","ParentId":null,"OwnerUserId":"2637126","Title":null,"Body":"<p>After commenting @Marcin answer, I have more carefully checked one of my students code where I found the same weird behavior, even after only 2 epochs ! (So @Marcin's explanation was not very likely in my case).<\/p>\n\n<p>And I found that the answer is actually very simple: the accuracy computed with the Keras method <code>evaluate<\/code> is just plain wrong when using binary_crossentropy with more than 2 labels. You can check that by recomputing the accuracy yourself (first call the Keras method \"predict\" and then compute the number of correct answers returned by predict): you get the true accuracy, which is much lower than the Keras \"evaluate\" one.<\/p>\n"},{"AnswerId":"45435674","CreationDate":"2017-08-01T10:43:45.787","ParentId":null,"OwnerUserId":"7403220","Title":null,"Body":"<p>I came across an \"inverted\" issue \u2014 I was getting good results with categorical_crossentropy (with 2 classes) and poor with binary_crossentropy. It seems that problem was with wrong activation function. The correct settings were:<\/p>\n\n<ul>\n<li>for <code>binary_crossentropy<\/code>: sigmoid activation, scalar target <\/li>\n<li>for <code>categorical_crossentropy<\/code>: softmax activation, one-hot encoded target<\/li>\n<\/ul>\n"},{"AnswerId":"46038271","CreationDate":"2017-09-04T13:34:45.440","ParentId":null,"OwnerUserId":"4685471","Title":null,"Body":"<p>The reason for this apparent performance discrepancy between categorical &amp; binary cross entropy is what @xtof54 has already reported in his answer, i.e.:<\/p>\n\n<blockquote>\n  <p>the accuracy computed with the Keras method <code>evaluate<\/code> is just plain\n  wrong when using binary_crossentropy with more than 2 labels<\/p>\n<\/blockquote>\n\n<p>I would like to elaborate more on this, demonstrate the actual underlying issue, explain it, and offer a remedy.<\/p>\n\n<p>This behavior is not a bug; the underlying reason is a rather subtle &amp; undocumented issue at how Keras actually <em>guesses<\/em> which accuracy to use, depending on the loss function you have selected, when you include simply <code>metrics=['accuracy']<\/code> in your model compilation. In other words, while your first compilation option <\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><code>model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>is valid, your second one:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>will not produce what you expect, but the reason is not the use of binary cross entropy (which, at least in principle, is an absolutely valid loss function).<\/p>\n\n<p>Why is that? If you check the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/metrics.py\" rel=\"noreferrer\">metrics source code<\/a>, Keras does not define a single accuracy metric, but several different ones, among them <code>binary_accuracy<\/code> and <code>categorical_accuracy<\/code>. What happens <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/engine\/training.py#L876\" rel=\"noreferrer\">under the hood<\/a> is that, since you have selected binary cross entropy as your loss function and have not specified a particular accuracy metric, Keras (wrongly...) infers that you are interested in the <code>binary_accuracy<\/code>, and this is what it returns - while in fact you are interested in the <code>categorical_accuracy<\/code>.<\/p>\n\n<p>Let's verify that this is the case, using the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/mnist_cnn.py\" rel=\"noreferrer\">MNIST CNN example<\/a> in Keras, with the following modification:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])  # WRONG way\n\nmodel.fit(x_train, y_train,\n          batch_size=batch_size,\n          epochs=2,  # only 2 epochs, for demonstration purposes\n          verbose=1,\n          validation_data=(x_test, y_test))\n\n# Keras reported accuracy:\nscore = model.evaluate(x_test, y_test, verbose=0) \nscore[1]\n# 0.9975801164627075\n\n# Actual accuracy calculated manually:\nimport numpy as np\ny_pred = model.predict(x_test)\nacc = sum([np.argmax(y_test[i])==np.argmax(y_pred[i]) for i in range(10000)])\/10000\nacc\n# 0.98780000000000001\n\nscore[1]==acc\n# False    \n<\/code><\/pre>\n\n<p>To remedy this, i.e. to use indeed binary cross entropy as your loss function (as I said, nothing wrong with this, at least in principle) while still getting the <strong>categorical<\/strong> accuracy required by the problem at hand, you should ask explicitly for <code>categorical_accuracy<\/code> in the model compilation as follows:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>from keras.metrics import categorical_accuracy\nmodel.compile(loss='binary_crossentropy', optimizer='adam', metrics=[categorical_accuracy])\n<\/code><\/pre>\n\n<p>In the MNIST example, after training, scoring, and predicting the test set as I show above, the two metrics now are the same, as they should be:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code># Keras reported accuracy:\nscore = model.evaluate(x_test, y_test, verbose=0) \nscore[1]\n# 0.98580000000000001\n\n# Actual accuracy calculated manually:\ny_pred = model.predict(x_test)\nacc = sum([np.argmax(y_test[i])==np.argmax(y_pred[i]) for i in range(10000)])\/10000\nacc\n# 0.98580000000000001\n\nscore[1]==acc\n# True    \n<\/code><\/pre>\n\n<p>System setup:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>Python version 3.5.3\nTensorflow version 1.2.1\nKeras version 2.0.4\n<\/code><\/pre>\n\n<p><strong>UPDATE<\/strong>: After my post, I discovered that this issue had already been identified in <a href=\"https:\/\/stackoverflow.com\/questions\/45799474\/keras-model-evaluate-vs-model-predict-accuracy-difference-in-multi-class-nlp-ta\/45834857#45834857\">this answer<\/a>.<\/p>\n"},{"AnswerId":"49175655","CreationDate":"2018-03-08T14:34:48.707","ParentId":null,"OwnerUserId":"4949282","Title":null,"Body":"<p>It all depends on the type of classification problem you are dealing with. There are three main categories;<\/p>\n\n<ul>\n<li><strong>binary<\/strong> classification (two target classes)<\/li>\n<li><strong>multi-class<\/strong> classification (more than two <strong>exclusive<\/strong> targets)<\/li>\n<li><strong>multi-label<\/strong> classification (more than two <strong>non exclusive<\/strong> targets) in which multiple target classes can be on at the same time<\/li>\n<\/ul>\n\n<p>In the first case, binary cross-entropy should be used and targets should be encoded as one-hot vectors.<\/p>\n\n<p>In the second case, categorical cross-entropy should be used and targets should be encoded as one-hot vectors.<\/p>\n\n<p>In the last case, binary cross-entropy should be used and targets should be encoded as one-hot vectors. Each output neuron (or unit) is considered as a separate random binary variable, and the loss for the entire vector of outputs is the product of the loss of single binary variables. Therefore it is the product of binary cross-entropy for each single output unit.<\/p>\n\n<p>binary cross-entropy is defined as such:\n<a href=\"https:\/\/i.stack.imgur.com\/Qb9x9.png\" rel=\"noreferrer\">binary cross-entropy<\/a>\nand categorical cross-entropy is defined as such:\n<a href=\"https:\/\/i.stack.imgur.com\/g0b6Y.png\" rel=\"noreferrer\">categorical cross-entropy<\/a><\/p>\n"},{"AnswerId":"48592159","CreationDate":"2018-02-02T23:09:26.013","ParentId":null,"OwnerUserId":"3752261","Title":null,"Body":"<p>when using the <code>categorical_crossentropy<\/code> loss, your targets should be in categorical format (e.g. if you have 10 classes, the target for each sample should be a 10-dimensional vector that is all-zeros except for a 1 at the index corresponding to the class of the sample).<\/p>\n"},{"AnswerId":"42093183","CreationDate":"2017-02-07T15:04:53.700","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>As it is a multi-class problem, you have to use the categorical_crossentropy, the binary cross entropy will produce bogus results, most likely will only evaluate the first two classes only.<\/p>\n\n<p>50% for a multi-class problem can be quite good, depending on the number of classes. If you have n classes, then 100\/n is the minimum performance you can get by outputting a random class.<\/p>\n"},{"AnswerId":"54812024","CreationDate":"2019-02-21T16:34:30.103","ParentId":null,"OwnerUserId":"8785067","Title":null,"Body":"<p>The binary_crossentropy(y_target, y_predict) doesn't need to apply in binary classification problem. .<\/p>\n\n<p>In the source code of <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/backend\/tensorflow_backend.py#L3584\" rel=\"nofollow noreferrer\">binary_crossentropy()<\/a>, the <code>nn.sigmoid_cross_entropy_with_logits(labels=target, logits=output)<\/code> TensorFlow function was actually used. \nAnd, in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sigmoid_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">documentation<\/a>, it says that:<\/p>\n\n<blockquote>\n  <p>Measures the probability error in discrete classification tasks in which each class is independent and not mutually exclusive. For instance, one could perform multilabel classification where a picture can contain both an elephant and a dog at the same time.<\/p>\n<\/blockquote>\n"},{"AnswerId":"53968243","CreationDate":"2018-12-29T09:13:41.543","ParentId":null,"OwnerUserId":"10845805","Title":null,"Body":"<p>a simple example under a multi-class setting to illustrate<\/p>\n\n<p>suppose you have 4 classes (onehot encoded) and below is just one prediction<\/p>\n\n<p>true_label = [0,1,0,0]\npredicted_label = [0,0,1,0]<\/p>\n\n<p>when using categorical_crossentropy, the accuracy is just 0 , it only cares about if you get the concerned class right.<\/p>\n\n<p>however when using binary_crossentropy, the accuracy is calculated for all classes, it would be 50% for this prediction. and the final result will be the mean of the individual accuracies for both cases.<\/p>\n\n<p>it is recommended to use categorical_crossentropy for multi-class(classes are mutually exclusive) problem but binary_crossentropy for multi-label problem.<\/p>\n"},{"AnswerId":"56031472","CreationDate":"2019-05-07T22:59:41.967","ParentId":null,"OwnerUserId":"11467138","Title":null,"Body":"<p>Take a look at the equation you can find that <a href=\"https:\/\/i.stack.imgur.com\/zcqvg.png\" rel=\"nofollow noreferrer\">binary cross entropy<\/a> not only punish those label = 1, predicted =0, but also label = 0, predicted = 1.<\/p>\n\n<p>However <a href=\"https:\/\/i.stack.imgur.com\/dgHW4.png\" rel=\"nofollow noreferrer\">categorical cross entropy<\/a> only punish those label = 1 but predicted = 1.That's why we make assumption that there is only ONE label positive.<\/p>\n"}]}
{"QuestionId":42081917,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><deep-learning>","CreationDate":"2017-02-07T04:43:56.507","AcceptedAnswerId":null,"OwnerUserId":7526688.0,"Title":"Tensor flow, making predictions using a trained network","Body":"<p>So I am training a network to classify images in tensor flow.  After I trained the network I began work on trying to use it to classify other images. The goal is to import an image, feed it to the classifier and have it print the result. I am having some trouble getting that part off the ground though.  Here is what I have so far. I found that having tf.argmax(y,1) gave an error.  I found that changing it to 0 fixed that error.  However I am not convinced that it is actually working.  I tossed 2 images through the classifier and they both got the same class even though they are vastly different.  Just need some perspective here.  Is this valid?  Or is there something wrong here that will always feed me the same class (in this case I got class 0 for both of the images I tried).  <\/p>\n\n<p>Is this even the right way to approach making predictions in tensor flow?  This is just the culmination of my debugging, not sure if it is what should be done or not.  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42083754","CreationDate":"2017-02-07T07:07:41.567","ParentId":null,"OwnerUserId":"7526688","Title":null,"Body":"<p>So what had to happen here was first clear the kernel and outputs. Somewhere along the way my placeholders got muddled up and clearing the kernel fixed that right up.  Then I had to realize what really had to get done here: I had to call up the <code>softmax<\/code> function on my new data.  <\/p>\n\n<p>Like this:<\/p>\n\n<pre><code>pred = tf.nn.softmax(logits)\nclassification = sess.run(pred, feed_dict={x: image_array}) \n<\/code><\/pre>\n"}]}
{"QuestionId":42081945,"AnswerCount":0,"Tags":"<parallel-processing><tensorflow>","CreationDate":"2017-02-07T04:46:51.607","AcceptedAnswerId":null,"OwnerUserId":5590571.0,"Title":"TensorFlow tf.nn.conv2d parallelization","Body":"<p>Is there a way of parallelizing tf.nn.conv2d operation in TensorFlow using GPU when the batch size equals to 1?<\/p>\n\n<p>I want to give only one image to my CNN and I want the computations to be parallelized on a GPU. As far as I understand, TensorFlow does not parallelize efficiently for a batch size of 1 as the process is faster on CPU.<\/p>\n","answers":[]}
{"QuestionId":42082068,"AnswerCount":1,"Tags":"<python><keras><conv-neural-network><mnist>","CreationDate":"2017-02-07T04:58:17.353","AcceptedAnswerId":42082889.0,"OwnerUserId":6471247.0,"Title":"Keras for multiple digit recognition","Body":"<p>I am trying to understand CNNs through some exercises in Keras (theano backend). I am unable to fit the model below (Error: AttributeError: 'Convolution2D' object has no attribute 'get_shape').  This dataset is images (28*28) from the MNIST data concatenated together for a maximum of five images.  So the input shape should be 1, 28, 140 (grayscale = 1, height =28, width = 28*5)<\/p>\n\n<p>The goal is to predict the sequence of numbers. Thank you!!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42082889","CreationDate":"2017-02-07T06:10:06.490","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>The problem is in the input layer. Do the following change:<\/p>\n\n<pre><code>model_input=Input(shape=(img_channels, img_rows, img_cols))\n<\/code><\/pre>\n\n<p>provided your <code>image_dim_ordering<\/code> is <code>th<\/code>. Import the <code>Input<\/code> layer from <code>keras.layers<\/code>.<\/p>\n\n<p>I also notice that there are multiple outputs. So you need to use Function model instead of Sequential. Just change it to:<\/p>\n\n<pre><code>model = Model(input=model_input, output=lst)\n<\/code><\/pre>\n\n<p>Import <code>Model<\/code> from <code>keras.models<\/code>.<\/p>\n"}]}
{"QuestionId":42082452,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><tf-slim>","CreationDate":"2017-02-07T05:34:01.780","AcceptedAnswerId":42100817.0,"OwnerUserId":5861256.0,"Title":"Why is my transfer learning implementation on tensorflow is throwing me an error after a few iterations?","Body":"<p>I am using inception v1 architecture for transfer learning. I have downloded the checkpoints file, nets, pre-processing file from the below github repository<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim<\/a><\/p>\n\n<p>I have 3700 images and pooling out the last pooling layer filters from the graph for each of my image and appending it to a list\u00a0. With every iteration the ram usage is increasing and finally killing the run at around 2000 images. Can you tell me what mistake I have done\u00a0?<\/p>\n\n<p><a href=\"https:\/\/github.com\/Prakashvanapalli\/TensorFlow\/blob\/master\/Transfer_Learning\/inception_v1_finallayer.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Prakashvanapalli\/TensorFlow\/blob\/master\/Transfer_Learning\/inception_v1_finallayer.py<\/a><\/p>\n\n<p>Even if I remove the list appending and just trying to print the results. this is still happening. I guess the mistake is with the way of calling the graph. When I see my ram usage , with every iteration it is becoming heavy and I don't know why this is happening as I am not saving anything nor there is a difference between 1st iteration <\/p>\n\n<p>From my point, I am just sending one Image and getting the outputs and saving them. So it should work irrespective of how many images I send. <\/p>\n\n<p>I have tried it on both GPU (6GB) and CPU (32GB).<\/p>\n","answers":[{"AnswerId":"42100817","CreationDate":"2017-02-07T22:00:06.850","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>You seem to be storing images in your graph as <code>tf.constant<\/code>s. These will be persistent, and will cause memory issues like you're experiencing. Instead, I would recommend either <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/placeholders#placeholder\" rel=\"nofollow noreferrer\">placeholders<\/a> or <a href=\"https:\/\/www.tensorflow.org\/how_tos\/threading_and_queues\/\" rel=\"nofollow noreferrer\">queues<\/a>. Queues are very flexible, and can be very high performance, but can also get quite complicated. You may want to start with just a placeholder.<\/p>\n\n<p>For a full-complexity example of an image input pipeline, you could look at the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/image_processing.py\" rel=\"nofollow noreferrer\">Inception model<\/a>.<\/p>\n"}]}
{"QuestionId":42082913,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-07T06:11:46.463","AcceptedAnswerId":42086540.0,"OwnerUserId":5588279.0,"Title":"tf.reshape doesn't work as expected","Body":"<p>Actually, I don't know how to describe this question. It's so strange.<\/p>\n\n<pre><\/pre>\n\n<p>These are my whole code, I can't find any problem in it.<\/p>\n\n<p>But a  was raised. <\/p>\n\n<p>The error message is strange, because it seems happen at , but how could  has a shape of ? It should be  <\/p>\n\n<p>And, sometimes another  was raised. <\/p>\n\n<p>It is also difficult to understand, why it happened?<\/p>\n\n<p>Hope anyone could give me a hand.<\/p>\n","answers":[{"AnswerId":"42086540","CreationDate":"2017-02-07T09:48:02.063","ParentId":null,"OwnerUserId":"6933420","Title":null,"Body":"<p>Look what happens when you use <code>feed_dict<\/code> - you reference the variables <code>source<\/code> and <code>target<\/code>. However, the python variable no longer refers to the placeholders but rather the reshape ops - hence the op is 'skipped'. <\/p>\n\n<p>The easiest fix is renaming the placeholders to something unique. Further down in the network it is OK to reuse the same name (you could just call every layer <code>net<\/code>), it doesn't matter as long as you no longer need to reference them. <\/p>\n\n<p>Try giving this a go?<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport pickle\n\ndef weight_and_bias(name ,shape):\n    weight = tf.get_variable(\"W\" + name, shape=shape, initializer=tf.contrib.layers.xavier_initializer())\n    bias = tf.get_variable(\"B\" + name, shape=shape[-1], initializer=tf.contrib.layers.xavier_initializer())\n    return weight, bias\n\ndef conv2d_2x2(x, W):\n    return tf.nn.conv2d(x, W, strides=[1, 5, 5, 1], padding='SAME')\n\ndef max_pool_2x2(x):\n    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\n\nsess = tf.InteractiveSession()\n\nsource_placeholder = tf.placeholder(tf.float32, [None, None, 50, 50])\nsource_len = tf.placeholder(tf.int32, [None])\nsource_max_step = tf.shape(source)[1]\n\ntarget_placeholder = tf.placeholder(tf.float32, [None, None, 50, 50])\ntarget_len = tf.placeholder(tf.int32, [None])\ntarget_max_step = tf.shape(target)[1]\n\nW_conv, B_conv = weight_and_bias('conv1', [5, 5, 1, 32])\n\nsource = tf.reshape(source_placeholder, [-1, 50, 50], \"source_reshape\")\nsource_tmp = tf.reshape(source, [-1, 50, 50 ,1])\nsource_conv = tf.nn.relu(conv2d_2x2(source_tmp, W_conv) + B_conv)\nsource_pool = max_pool_2x2(source_conv)\nsource_flat = tf.reshape(source_pool, [-1, 5 * 5 * 32], \"source_pool_reshape\")\nsource = tf.reshape(source_flat, [-1, source_max_step, 5*5*32], \"source_flat_reshape\")\n\nW_conv, B_conv = weight_and_bias('conv2', [5, 5, 1, 32])\n\ntarget = tf.reshape(target_placeholder, [-1, 50, 50], \"target_reshape\")\ntarget_tmp = tf.reshape(target, [-1, 50, 50 ,1])\ntarget_conv = tf.nn.relu(conv2d_2x2(target_tmp, W_conv) + B_conv)\ntarget_pool = max_pool_2x2(target_conv)\ntarget_flat = tf.reshape(target_pool, [-1, 5 * 5 * 32], \"target_pool_reshape\")\ntarget = tf.reshape(target_flat, [-1, target_max_step, 5*5*32], \"target_flat_reshape\")\n\nsource_cell = tf.nn.rnn_cell.LSTMCell(500, initializer=tf.contrib.layers.xavier_initializer())\ntarget_cell = tf.nn.rnn_cell.LSTMCell(500, initializer=tf.contrib.layers.xavier_initializer())\nsource_rnn_output, _ = tf.nn.dynamic_rnn(source_cell, source, source_len, dtype=tf.float32, scope = \"source\")\ntarget_rnn_output, _ = tf.nn.dynamic_rnn(target_cell, target, target_len, dtype=tf.float32, scope = \"target\")\n\nsource_output = tf.transpose(source_rnn_output, [1, 0, 2])\ntarget_output = tf.transpose(target_rnn_output, [1, 0, 2])\n\nsource_final_output = tf.gather(source_output, -1)\ntarget_final_output = tf.gather(target_output, -1)\n\noutput = tf.concat(1, [source_final_output, target_final_output])\n\nW_sf, B_sf = weight_and_bias('sf', [1000, 2])\npredict = tf.nn.softmax(tf.matmul(output, W_sf) + B_sf)\ny = tf.placeholder(tf.float32, [None, 2])\n\ncross_entropy = -tf.reduce_sum(y * tf.log(predict))\ntrain_step = tf.train.RMSPropOptimizer(1e-4).minimize(cross_entropy)\n\ncorrect_prediction = tf.equal(tf.arg_max(predict, 1), tf.arg_max(y, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n\nwith open('set', 'rb') as f:\n    _set = pickle.load(f)\n    training_set = _set[0]\n    training_len = _set[1]\n    training_label = _set[2]\n\nsess.run(tf.global_variables_initializer())\n\nfor i in range(20000):\n\n    if i % 100 == 0:\n        train_accuacy = accuracy.eval(feed_dict = {source_placeholder: training_set[0], target_placeholder: training_set[1], source_len: training_len[0], target_len: training_len[1], y: training_label})\n        print(\"step %d, training accuracy %g\"%(i, train_accuacy))\n    train_step.run(feed_dict = {source_placeholder: training_set[0], target_placeholder: training_set[1], source_len: training_len[0], target_len: training_len[1], y: training_label})\n<\/code><\/pre>\n"}]}
{"QuestionId":42083278,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2017-02-07T06:37:47.707","AcceptedAnswerId":42091786.0,"OwnerUserId":7205119.0,"Title":"how to install tensorflow on window with anaconda","Body":"<p>OS : window 7 64 bit \nanaconda version is 4.3\nI try to install tensorflow using anaconda. <\/p>\n\n<p>First I installed anaconda.\nThe second,<\/p>\n\n<blockquote>\n  <p>conda creat -n tensorflow<\/p>\n<\/blockquote>\n\n<p>Third, the environment is not activated.\nIf you type activate tensorflow, it says that it is not a cmd command.\nCan not you just type <\/p>\n\n<blockquote>\n  <p>activate tensorflow<\/p>\n<\/blockquote>\n\n<p>into anaconda prompt?<\/p>\n\n<p>Where should I enter 'activate tensorflow'?<\/p>\n","answers":[{"AnswerId":"42091786","CreationDate":"2017-02-07T14:00:55.430","ParentId":null,"OwnerUserId":"7483498","Title":null,"Body":"<p>Instead of <code>creat<\/code> use <code>create<\/code><\/p>\n\n<pre><code> conda create -n tensorflow\n\n activate tensorflow\n<\/code><\/pre>\n"}]}
{"QuestionId":42083890,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-07T07:16:12.513","AcceptedAnswerId":null,"OwnerUserId":6033925.0,"Title":"UnicodeDecodeError on Github code","Body":"<p>I'm studying tensorflow now<\/p>\n\n<p>I downloaded codes of YoonKim's paper on this website<\/p>\n\n<p><a href=\"http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/\" rel=\"nofollow noreferrer\">http:\/\/www.wildml.com\/2015\/12\/implementing-a-cnn-for-text-classification-in-tensorflow\/<\/a><\/p>\n\n<p>I use pycharm on windows 10 and I installed python3.5.3, Numpy and tensorflow<\/p>\n\n<p>I just downloaded this code on the website<\/p>\n\n<p>and I read readme.txt file .<\/p>\n\n<p>I followed the instruction. I didn't touch any codes. <\/p>\n\n<p>Following Readme file, I activated train.py on pycharm.<\/p>\n\n<p>But I have error like this <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/cJ5LH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cJ5LH.png\" alt=\"Error\"><\/a><\/p>\n\n<p>It seems I have to change rt-polarity.neg, rt-polarity.pos to UTF8.<\/p>\n\n<p>But I dont know how cause I'm not good at python.<\/p>\n\n<p>Need some help <\/p>\n","answers":[]}
{"QuestionId":42083918,"AnswerCount":0,"Tags":"<python-3.x><machine-learning><nlp><keras><lstm>","CreationDate":"2017-02-07T07:17:27.727","AcceptedAnswerId":null,"OwnerUserId":7523984.0,"Title":"Word level sentence generation using keras","Body":"<p>I am new to keras. I am trying to build word level sentence generation module using keras.\nI use a vocabulary size of 8000 and one-hot-vector representation of words in my corpus.<\/p>\n\n<p>This is my model<\/p>\n\n<pre><\/pre>\n\n<p>here input dimension is 3 (as keras lstm expects 3d input)  input shape such that (batch_size,1,8000). I made the 2nd dimension as 1 because I want to feed word by word.\nBatch_size is 45 because that is the avarage length of a sentence in the corpus.\nAll sentences are pretended with a \"START_TOKEN\" and appended with a \"END_TOKEN\".<\/p>\n\n<p>len(itw) returns length of vocabulary which in my case is 8000<\/p>\n\n<p>Now after training I wished to loop the generated words back as input until a stop token is encountered to generated a sentence.\nBut it seems that the model only considers the current input.<\/p>\n\n<p>I hoped the model to consider the inputs before also and not just the current one.<\/p>\n\n<p>So how should I change my model<\/p>\n\n<p>Also is Keras unfit for nlp applications involving varying time serise<\/p>\n\n<p>How to change the model such that it will generate the next word given n words.\nWhere n is any natural number<\/p>\n","answers":[]}
{"QuestionId":42084742,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-07T08:11:33.490","AcceptedAnswerId":null,"OwnerUserId":7493814.0,"Title":"I run cifar10 on Tensorflow, but it show me some Error","Body":"<p>everyone, I do a little change on  and it occurs an error. \nthe origin code in  is:<\/p>\n\n<pre><\/pre>\n\n<p>and I change it to:<\/p>\n\n<pre><\/pre>\n\n<p>The reason is that I want to change some codes in these three files, so I import these files by myself instead of importing from tensorflow. In this case, I do not change anything in these three files.<\/p>\n\n<p>When I run the , The error is showed below:<\/p>\n\n<pre><\/pre>\n\n<p>Any answer is appreciated, thanks!<\/p>\n","answers":[]}
{"QuestionId":42086214,"AnswerCount":1,"Tags":"<python><tensorflow><pycharm>","CreationDate":"2017-02-07T09:32:11.320","AcceptedAnswerId":null,"OwnerUserId":6693951.0,"Title":"pycharm cannot run script but can debug it","Body":"<p>When I ran a script in , it exited with:<\/p>\n\n<p><\/p>\n\n<p>But when I debug it in , the program ran without any problem. <\/p>\n\n<p>Also, if I ran the script in ubuntu terminal, no problem occurs, too. Why does this happen, or how can I debug this problem?<\/p>\n","answers":[{"AnswerId":"44600199","CreationDate":"2017-06-17T02:23:50.040","ParentId":null,"OwnerUserId":"8174395","Title":null,"Body":"<p>I have run into a similar error running caffe on pycharm. I think it's because of the version of Python. When I installed Python 2.7.13, it worked!<\/p>\n"}]}
{"QuestionId":42086465,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-02-07T09:43:53.580","AcceptedAnswerId":42087839.0,"OwnerUserId":5009875.0,"Title":"caffe `\"Python\"` layer not found?","Body":"<p>I have installed caffe, uncommenting<\/p>\n\n<pre><\/pre>\n\n<p>in <\/p>\n\n<p>When I use a python data layer in my net.prototxt, it says <\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>To cross check it in python interface,\nI tried<\/p>\n\n<pre><\/pre>\n\n<p>this seems to work ,no error then.\nWhere is the problem?<\/p>\n","answers":[{"AnswerId":"42087839","CreationDate":"2017-02-07T10:48:05.653","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can find out what layer types caffe has in python simply by examining <code>caffe.layer_types_list()<\/code>. For example, if you actually have a <code>\"Python\"<\/code> layer, then<\/p>\n\n<pre><code>list(caffe.layer_type_list()).index('Python')\n<\/code><\/pre>\n\n<p>Should actually return an index for its name in the layer types list. <\/p>\n\n<p>As for <code>L.Python()<\/code> - this <code>caffe.NetSpec()<\/code> interface is used to programatically write a net prototxt, and at the writing stage layer types are not checked. You can actually write whatever layer you want:<\/p>\n\n<pre><code>L.YouDontThinkTheyNameALayerLikeThis()\n<\/code><\/pre>\n\n<p>Is totally cool. Even converting it to prototxt:<\/p>\n\n<pre><code>print \"{}\".format(L.YouDontThinkTheyNameALayerLikeThis().to_proto())\n<\/code><\/pre>\n\n<p>Actually results with this:<\/p>\n\n<blockquote>\n<pre><code>layer {\n  name: \"YouDontThinkTheyNameALayerLikeThis1\"\n  type: \"YouDontThinkTheyNameALayerLikeThis\"\n  top: \"YouDontThinkTheyNameALayerLikeThis1\"\n}\n<\/code><\/pre>\n<\/blockquote>\n\n<p>You'll get an error message once you try to run this \"net\" using caffe...<\/p>\n"}]}
{"QuestionId":42087302,"AnswerCount":1,"Tags":"<python><tensorflow><pyspark>","CreationDate":"2017-02-07T10:22:36.127","AcceptedAnswerId":null,"OwnerUserId":2563522.0,"Title":"Tensorflow flags not recognized","Body":"<p>I have a tensorflow code to be run on pyspark.\nCode<\/p>\n\n<pre><\/pre>\n\n<p>I have a RDD for which I am calling a function . But I am not able to print the <strong>FLAGS<\/strong> value.<\/p>\n\n<pre><\/pre>\n\n<p>However I get a following error.<\/p>\n\n<pre><\/pre>\n\n<p>Moreover I am not able to create a broadcast variable of a tensorflow object.<\/p>\n\n<p>For example, if I write<\/p>\n\n<pre><\/pre>\n\n<p>I get an error.<\/p>\n","answers":[{"AnswerId":"51040273","CreationDate":"2018-06-26T10:10:15.197","ParentId":null,"OwnerUserId":"4137497","Title":null,"Body":"<p>Get the latest TensorFlow (>1.4) with Python3 and use <code>FLAGS(sys.argv)<\/code>, since the <code>FLAGS._parse_flags()<\/code> is not supported any more.<\/p>\n\n<pre><code>import sys\nimport tensorflow as tf\n\nFLAGS = tf.app.flags.FLAGS\nunparsed = FLAGS(sys.argv)\n<\/code><\/pre>\n"}]}
{"QuestionId":42087542,"AnswerCount":1,"Tags":"<python><tensorflow><pip><ubuntu-16.04>","CreationDate":"2017-02-07T10:34:18.473","AcceptedAnswerId":42087612.0,"OwnerUserId":null,"Title":"Pip help install","Body":"<p>I am trying to install tensorflow on ubuntu from: <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#virtualenv_installation\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/os_setup#virtualenv_installation<\/a>\nBut when I get to the step: \nI get this error:<\/p>\n\n<blockquote>\n  <p><\/p>\n<\/blockquote>\n\n<p>Any help please<\/p>\n","answers":[{"AnswerId":"42087612","CreationDate":"2017-02-07T10:37:25.507","ParentId":null,"OwnerUserId":"3633742","Title":null,"Body":"<p>That's because the environment variable <code>$TF_BINARY_URL<\/code> is not set. you must export it first, like described in the docs you provided.<\/p>\n\n<blockquote>\n<pre><code># Ubuntu\/Linux 64-bit, CPU only, Python 2.7\n(tensorflow)$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-0.12.1-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n<\/blockquote>\n\n<p>(or any other url specific to arch\/platform)<\/p>\n"}]}
{"QuestionId":42087742,"AnswerCount":1,"Tags":"<lua><torch>","CreationDate":"2017-02-07T10:43:41.730","AcceptedAnswerId":42094168.0,"OwnerUserId":225885.0,"Title":"Openface batch-represent \"unknown object\"","Body":"<p>I am trying to run the script  of OpenFace to train some images and getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am not familiar with Lua but does anybody know how to debug this error please? I guess it is a missing dependency from my machine as it works on another machine.<\/p>\n","answers":[{"AnswerId":"42094168","CreationDate":"2017-02-07T15:49:50.517","ParentId":null,"OwnerUserId":"4687565","Title":null,"Body":"<p>It doesn't really look like missing dependency, the torch library complains about incorrect format of the learning model. The model name is provided to the torch as a parameter.  You can also find out that in order to give that to torch openface does expects a <code>-model<\/code> parameter passed to it. The error log doesn't specify whether you did provide one, but it seems that there's default one in that case \"openface.nn4.small2.v1.t7\" it should be located somewhere. If anything I did find it on the <a href=\"https:\/\/github.com\/pyannote\/pyannote-data\" rel=\"nofollow noreferrer\">internets<\/a>. <\/p>\n\n<p>But it does manage to load something first, if file absence was the cause I believe it would error sooner.<\/p>\n\n<p>So, I'd say you should check the parameters you're calling the program with. And check model file. <\/p>\n"}]}
{"QuestionId":42087772,"AnswerCount":0,"Tags":"<tensorflow><image-segmentation><softmax><cross-entropy>","CreationDate":"2017-02-07T10:45:22.940","AcceptedAnswerId":null,"OwnerUserId":1710775.0,"Title":"How can I normalize my loss function in relation to the number of elements in a class in Tensorflow?","Body":"<p>In an image segmentation problem, we will usually get a background label that dominates the mask and the object(s) occupying only a small area in the mask. \nAs such, during training, I have observed that my neural network does very well in classifying the background label but really poor at classifying the objects. <\/p>\n\n<p>I am using the Tensorflow's tf.nn.sparse_softmax_cross_entropy_with_logits as my lost function. Maybe in relation to this problem, my loss value is also not decreasing\/converging. <\/p>\n\n<p>I was told that maybe I should consider adding a weighting factor to my loss function so that the dominant class will have a small weight compared to other non-dominant class. Can anyone share some insights? <\/p>\n","answers":[]}
{"QuestionId":42090414,"AnswerCount":1,"Tags":"<python><optimization><scipy><tensorflow>","CreationDate":"2017-02-07T12:52:53.540","AcceptedAnswerId":null,"OwnerUserId":7528640.0,"Title":"Optimization in SciPy with noisy loss function","Body":"<p>I have a nasty optimization problem in TensorFlow which requires a non-linear optimizer to solve, the internal tensorflow optimizers (Gradient Descent, AdaGrad, Adam) seem to do significantly worse than using scipy as an external optimizer (CG, BFGS) of the same graph. <\/p>\n\n<p>This would be fine but for the production run I want to do I need to use minibatches of my training dataset to optimize. I have implemented this by each time the loss\/gradient function is called, a new minibatch of data is used to calculate it. (I am using a modified version of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/opt\/python\/training\/external_optimizer.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/opt\/python\/training\/external_optimizer.py<\/a> )  In practice, this means that the loss function is a noisy function of the input parameters. <\/p>\n\n<p>Scipy seems to be having a problem with this, limiting any calling of scipy.minimize to only a few iterations, like so:<\/p>\n\n<pre><\/pre>\n\n<p>In contrast, if I run this optimization with the full dataset (which is feasible now but not later) it will converge to around 0.1 within one call of scipy.minimize (and do around 1000 iterations without quitting). <\/p>\n\n<p>Has anyone encountered this problem? Is there a fix (easy preferred but hacky also OK) to stop scipy from quitting out of these optimization problems? Something like a min_iter keyword would be perfect but as far as I'm aware that isn't implemented.<\/p>\n\n<p>I hope this made sense. Thanks!<\/p>\n\n<p>EDIT: I was asked for code but the full code is several hundreds of lines long so I'll make a short example:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42094111","CreationDate":"2017-02-07T15:47:31.983","ParentId":null,"OwnerUserId":"39123","Title":null,"Body":"<p>Normalizing the coefficients or variables in a preprocessing step could help. Did you do that?   Overflow, underflow, and failure to converge are commonly a symptom of numerical scaling problems.<\/p>\n"}]}
{"QuestionId":42091999,"AnswerCount":1,"Tags":"<deep-learning><keras>","CreationDate":"2017-02-07T14:10:11.790","AcceptedAnswerId":null,"OwnerUserId":3620916.0,"Title":"Merging feature representations of two different Networks using Keras","Body":"<p>A simple hint is appreciated. \nHow can we fuse the feature representations output from 2 different networks using Keras?<\/p>\n\n<p>More explanation: \nThe model consists of a CNN and Gated RNN. <a href=\"https:\/\/i.stack.imgur.com\/6z5ak.png\" rel=\"nofollow noreferrer\">The feature representations output from the networks are fused<\/a> and fed to a standard softmax regression classifier.<\/p>\n","answers":[{"AnswerId":"42092904","CreationDate":"2017-02-07T14:51:59.643","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>You should use a merge layer (<a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">see doc here<\/a>)<\/p>\n\n<p>The mode to chose is up to you, it might be a sum, a concatenation or a custom merge. \nIf they are different features, I would recommend you to concatenate them and then use a Dense layer with softmax activation for your classification prediction.<\/p>\n\n<p>Take care of using the right merge layer : <code>Merge<\/code> is used in sequential models (merging 2 layers) and <code>merge<\/code> is to merge tensors (using the functionnal API)<\/p>\n\n<p>I hope this helps<\/p>\n"}]}
{"QuestionId":42092443,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-07T14:32:42.073","AcceptedAnswerId":42093345.0,"OwnerUserId":5107084.0,"Title":"Machine Learning: Does computing the accuracy score for binary labels always result in a low accuracy score?","Body":"<p>If I have 2 labels (1 and 0), and after I pass my logits through a softmax activation layer, I get something like:<\/p>\n\n<pre><\/pre>\n\n<p>as a predictions output, and my labels are only 1 or 0, does this always result in a low accuracy? Meaning to say if I have a lot more classes,  will my softmax layer give me something close to either 1 or 0 for each of the classes, which give me a higher accuracy score?<\/p>\n\n<p>Further, if I want to use accuracy as my metric, is there a way to scale my probability to either 0 or 1? Can be this be done by applying a mask in TensorFlow that outputs boolean values whenever a probability hits 0.5 or above?<\/p>\n","answers":[{"AnswerId":"42093345","CreationDate":"2017-02-07T15:13:22.453","ParentId":null,"OwnerUserId":"2300116","Title":null,"Body":"<p>After softmax layer you have probabilities in range 0..1, \nso if you want to check accuracy against your labels being only 0 or 1 you have to convert probablities <\/p>\n\n<p>if pred>0.5 then pred=1\nif pred&lt;=0.5 then pred=0<\/p>\n"}]}
{"QuestionId":42092475,"AnswerCount":0,"Tags":"<python><jenkins><keras>","CreationDate":"2017-02-07T14:33:45.760","AcceptedAnswerId":null,"OwnerUserId":7085724.0,"Title":"Configure Keras backend in Jenkins","Body":"<p>I am using Jenkins to run a python script on Windowas.<\/p>\n\n<p>In the script it uses Keras with tensorflow backend (I configured Keras.json). Then I encountered a problem: the Keras in my PyCharm and normal prompt uses Tensorflow backend but uses Theano backend when I run the script in Jenkins.<\/p>\n\n<p>Normally when running in PyCharm, it uses Tensorflow backend as configured.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xEdX8.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xEdX8.jpg\" alt=\"Using Tensorflow backend in PyCharm\"><\/a><\/p>\n\n<p>as described, when running in Jenkins it used Theano.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/OQ7zo.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/OQ7zo.jpg\" alt=\"But changed into Theano backend in Jenkins\"><\/a><\/p>\n\n<p>This is my json config file for keras.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/humSI.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/humSI.jpg\" alt=\"json config file\"><\/a><\/p>\n\n<p>I think I need to config something like keras.json again somewhere but I couldn't find it elsewhere. Can anyone help me to properly configure so that I can use keras with Tensorflow backend on Jenkins . Thank you<\/p>\n","answers":[]}
{"QuestionId":42092508,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-07T14:35:30.503","AcceptedAnswerId":42099639.0,"OwnerUserId":4520592.0,"Title":"How can I encode labels in TensorFlow?","Body":"<p>I need to convert my string labels into vectors like [0, 0, ... , 1, ... 0]. <br>\nAs far as I could understand this is something that called one hot vector. <br>\nI have 10 classes, so 10 different string labels. <br><\/p>\n\n<p>Could anyone please help with direct and inverse transformation? <br>\nI'm newbie in tensorflow so please be kind.<\/p>\n","answers":[{"AnswerId":"42099639","CreationDate":"2017-02-07T20:44:35.010","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>The forward direction is easy, since there's the <code>tf.one_hot<\/code> op:<\/p>\n\n<pre><code>import tensorflow as tf\n\noriginal_indices = tf.constant([1, 5, 3])\ndepth = tf.constant(10)\none_hot_encoded = tf.one_hot(indices=original_indices, depth=depth)\n\nwith tf.Session():\n  print(one_hot_encoded.eval())\n<\/code><\/pre>\n\n<p>Outputs:<\/p>\n\n<pre><code>[[ 0.  1.  0.  0.  0.  0.  0.  0.  0.  0.]\n [ 0.  0.  0.  0.  0.  1.  0.  0.  0.  0.]\n [ 0.  0.  0.  1.  0.  0.  0.  0.  0.  0.]]\n<\/code><\/pre>\n\n<p>The inverse of this isn't too bad either, with <code>tf.where<\/code> to find the non-zero indices:<\/p>\n\n<pre><code>def decode_one_hot(batch_of_vectors):\n  \"\"\"Computes indices for the non-zero entries in batched one-hot vectors.\n\n  Args:\n    batch_of_vectors: A Tensor with length-N vectors, having shape [..., N].\n  Returns:\n    An integer Tensor with shape [...] indicating the index of the non-zero\n    value in each vector.\n  \"\"\"\n  nonzero_indices = tf.where(tf.not_equal(\n      batch_of_vectors, tf.zeros_like(batch_of_vectors)))\n  reshaped_nonzero_indices = tf.reshape(\n      nonzero_indices[:, -1], tf.shape(batch_of_vectors)[:-1])\n  return reshaped_nonzero_indices\n\nwith tf.Session():\n  print(decode_one_hot(one_hot_encoded).eval())\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>[1 5 3]\n<\/code><\/pre>\n"}]}
{"QuestionId":42092727,"AnswerCount":0,"Tags":"<python><tensorflow><install><spyder>","CreationDate":"2017-02-07T14:44:35.500","AcceptedAnswerId":null,"OwnerUserId":7529216.0,"Title":"Tensorflow not executing correctly","Body":"<p>I am trying to run a sample tensorflow script: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py<\/a><\/p>\n\n<p>When I run it the first time, I get an error at the end of the run:<\/p>\n\n<pre><\/pre>\n\n<p>Any subsequent run would break right after loading the data:<\/p>\n\n<pre><\/pre>\n\n<p>The only change I have done is to reduce the number of loops from 1000 to 100 in line 193 of the code (but the error occurs in any case).\nI use Sypder and Python 2 on a Mac.<\/p>\n\n<p>Any ideas how to fix this? Might there be anything wrong with the installation.<\/p>\n\n<p>Also the code should write summaries for using with tensorboard. Where are these files stored? Can't find them in the location of the module (Desktop).<\/p>\n\n<p>Thanks al lot for the help!<\/p>\n","answers":[]}
{"QuestionId":42092737,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-07T14:45:06.140","AcceptedAnswerId":null,"OwnerUserId":195651.0,"Title":"Freezing\/exporting a part of a TensorFlow graph","Body":"<p>My question is related to <a href=\"https:\/\/stackoverflow.com\/q\/38947658\/195651\">this one here<\/a> about persisting graphs: I wonder if it is possible to only export a <em>certain part<\/em> of a graph, e.g. a subgraph prefixed by a given scope, using TensorFlow 0.12 or newer.<\/p>\n\n<p>I'm currently using a combination of ,  and  to export (\"freeze\") a graph into a protocol buffer file which I then can load back using . During the export I can specify which nodes are considered <em>required outputs<\/em> of the graph, so no upstream nodes are thrown away, while during import I can rewire certain parts of the graph to other operations using . <\/p>\n\n<p>This all works fine, but it is missing the notion of <em>unnecessary inputs<\/em> and the problem is that by doing so, the <em>entire<\/em> upstream of the  is written to the file as well, i.e. everything input and preprocessing related.<\/p>\n\n<p>Currently, exporting looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>While importing looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Is there a way to filter and\/or remove the parts I do not require during or before exporting? To be clear, I known I can ignore them after loading, but I do not to export them to begin with. Would using collections help at some point?<\/p>\n","answers":[{"AnswerId":"42117108","CreationDate":"2017-02-08T15:25:51.193","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>The built in graph saving tools need to be reworked in my opinion.<\/p>\n\n<p>The best solution I've found is to create a list of the variables you want saved then save them during the session.  Afterwards you have the freedom to reload the weights and change the graph as needed.<\/p>\n\n<pre><code>parameters = []\nparameters += [wieghts, biases]\n...\ndef load_weights(self, weight_file, sess):\n    weights = np.load(weight_file)\n    keys = sorted(weights.keys())\n    for i, k in enumerate(keys):\n        print i, k, np.shape(weights[k])\n        sess.run(self.parameters[i].assign(weights[k]))\n<\/code><\/pre>\n\n<p>credit for idea goes to Davi Frossard <a href=\"https:\/\/www.cs.toronto.edu\/~frossard\/post\/vgg16\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
{"QuestionId":42093083,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-07T15:00:00.220","AcceptedAnswerId":42095720.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Does creating a near-replica of a graph as an extension to the original graph to test the validation dataset consume a lot of memory?","Body":"<p>Meaning to say if I have the following operations for training purposes in my graph initially:<\/p>\n\n<pre><\/pre>\n\n<p>Which allows me to train my model just fine, but I would like to run a small validation dataset that evaluates my model every once in a while inside my , would it take too much memory to consume a nearly exact replica within the same graph like:<\/p>\n\n<pre><\/pre>\n\n<p>My concern is that to evaluate my validation dataset, I need to set the  argument to  so that I can disable dropout. But will creating an entire inception-resnet-v2 model from scratch just for validation inside the same graph consume too much memory? Or should I just create an entirely new file that runs the validation on my own?<\/p>\n\n<p>Ideally, I wanted to have 3 kinds of dataset - a training one, a small validation dataset to test during training, and a final evaluation dataset. This small validation dataset will help me see if my model is overfitting to the training data. If my proposed idea consumes too much memory, however, would it be equivalent to just occasionally monitor the training data score? Is there a better idea to test the validation set while training?<\/p>\n","answers":[{"AnswerId":"42095720","CreationDate":"2017-02-07T17:01:10.370","ParentId":null,"OwnerUserId":"5903959","Title":null,"Body":"<p>TensorFlow's devs thought about it and made variable ready to be shared.\nYou can see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sharing_variables\" rel=\"nofollow noreferrer\">here the doc<\/a>.<\/p>\n\n<p>Using scopes the right way make it possible to reuse some variable.\nOne very good example (the context is language model but never mind) is <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/ptb_word_lm.py#L329\" rel=\"nofollow noreferrer\">TensorFlow PTB Word LM<\/a>.<\/p>\n\n<p>The global pseudo-code of this approach is something like:<\/p>\n\n<pre><code>class Model:\n  def __init__(self, train=True, params):\n  \"\"\" Build the model \"\"\"\n\n  tf.placeholder( ... ) \n\n  tf.get_variable( ...) \n\n\n\ndef main(_):\n  with tf.Graph.as_default() as g:\n    with tf.name_scope(\"Train\"):\n      with tf.variable_scope(\"Model\", reuse=None):\n        train = Model(train=True, params ) \n    with tf.name_scope(\"Valid\"):\n      # Now reuse variables = no memory cost\n      with tf.variable_scope(\"Model\", reuse=True):\n        # But you can set different parameters\n        valid = Model(train=False, params)\n\n    session = tf.Session\n    ...\n<\/code><\/pre>\n\n<p>Thus you can share some variable without having the exact same model as the parameters may change the model itself. <\/p>\n\n<p>Hope this helps<br>\n<code>pltrdy<\/code><\/p>\n"}]}
{"QuestionId":42093250,"AnswerCount":3,"Tags":"<opencv><machine-learning><tensorflow>","CreationDate":"2017-02-07T15:08:07.267","AcceptedAnswerId":null,"OwnerUserId":7219743.0,"Title":"Obtain negative results from a machine learning algorithm","Body":"<p>I have a set of images of a particular object. I want to find if some of these has anomalies with a machine learning algorithm. For example if I have many photos of glasses I want to find if one of these is broken or has something anomalous. Something like this:<\/p>\n\n<p><strong>GOOD!!<\/strong>\n<a href=\"https:\/\/i.stack.imgur.com\/IGP04.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IGP04.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><strong>BAD!!<\/strong>\n<a href=\"https:\/\/i.stack.imgur.com\/LgwZb.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LgwZb.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>(Obviously I will use the same kind of glasses...)<\/p>\n\n<p>The problem is that I don't know every negative situation, so, for training, I have only positive images.<\/p>\n\n<p>In other words I want an algorithm that recognize if an image has something different from the dataset. Do you have any suggestion?<\/p>\n\n<p>In particular is there a way to use convolutional neural network? <\/p>\n","answers":[{"AnswerId":"42095105","CreationDate":"2017-02-07T16:30:59.720","ParentId":null,"OwnerUserId":"6885902","Title":null,"Body":"<p>This is a typical <strong>Classification<\/strong> problem. I do not understand why you need CNN for this ...... <\/p>\n\n<ol>\n<li>My suggestion would be to build\/train a classification model\ncomprising of only <strong>GOOD<\/strong> images of glass. Here you would possibly\nhave all kinds of glasses that are intact with a <em>regular<\/em> shape.<\/li>\n<li>If the model encounters anything other than <strong>GOOD<\/strong> images, it will\nclassify those as <strong>BAD<\/strong>  images. This so called <strong>BAD<\/strong> images may\ninclude cracked\/broken glasses having an <em>irregular<\/em> shape.<\/li>\n<\/ol>\n"},{"AnswerId":"42095521","CreationDate":"2017-02-07T16:50:55.160","ParentId":null,"OwnerUserId":"2484687","Title":null,"Body":"<p>What you are looking for is usually called anomaly, outlier, or novelty detection.  You have lots of examples of what your data should look like, and you want to know when something doesn't look like your data. <\/p>\n\n<p>A good approach for this problem, since you are using images, you can get a <a href=\"http:\/\/fastml.com\/classifying-images-with-a-pre-trained-deep-network\/\" rel=\"noreferrer\">feature vectorized version using a pre-trained CNN on image net<\/a>. Then you can use an anomaly detector on that feature set. The <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.ensemble.IsolationForest.html#sklearn.ensemble.IsolationForest\" rel=\"noreferrer\">isolation forest<\/a> should be an easier one to get working. <\/p>\n"},{"AnswerId":"42184068","CreationDate":"2017-02-12T03:58:12.250","ParentId":null,"OwnerUserId":"697538","Title":null,"Body":"<p>Another option that might work is to use an autoencoder.\nAutoencoders are unsupervised neural networks with bottleneck architecture that try to reconstruct its own input.\nWe could train a deep convolutional autoencoder with examples of good glasses so that it gets specialized in reconstructing those type of images. You don't need to train autoencoder with bad glasses.<\/p>\n\n<p>Therefore I would expect the trained autoencoder to produce low error rate for good glasses and high error rate for bad glasses. Error rate could be measured with MSE based on the difference between the reconstructed and original values (pixels).<\/p>\n\n<p>From the trained autoencoder you can plot the MSEs for good vs bad glasses to help you define the right threshold. Or you can also try statistic thresholds such as: avg + 2*std, median + 2*MAD, etc.<\/p>\n\n<ul>\n<li>Autoencoder details:<\/li>\n<\/ul>\n\n<p><a href=\"http:\/\/ufldl.stanford.edu\/tutorial\/unsupervised\/Autoencoders\/\" rel=\"nofollow noreferrer\">http:\/\/ufldl.stanford.edu\/tutorial\/unsupervised\/Autoencoders\/<\/a><\/p>\n\n<ul>\n<li>Deep autoencoder for images:<\/li>\n<\/ul>\n\n<p><a href=\"https:\/\/cds.cern.ch\/record\/2209085\/files\/Outlier%20detection%20using%20autoencoders.%20Olga%20Lyudchick%20(NMS).pdf\" rel=\"nofollow noreferrer\">https:\/\/cds.cern.ch\/record\/2209085\/files\/Outlier%20detection%20using%20autoencoders.%20Olga%20Lyudchick%20(NMS).pdf<\/a><\/p>\n"}]}
{"QuestionId":42093428,"AnswerCount":1,"Tags":"<neural-network><deep-learning><complexity-theory><caffe><convolution>","CreationDate":"2017-02-07T15:18:09.597","AcceptedAnswerId":42108281.0,"OwnerUserId":6786718.0,"Title":"Why is the most time spent in the fully connected layers despite its complexity is less than the conv-layers?","Body":"<p>When doing benchmarks of CNNs I found out that the most time is spent in the fully-connected layers. But when it comes to calculate the computational complexity I found out that:<\/p>\n\n<pre><\/pre>\n\n<p>Where <\/p>\n\n<pre><\/pre>\n\n<p>For an example, I have a , a  without padding , and with the , and the <\/p>\n\n<p>This results in the following calculation for the convolution:<\/p>\n\n<blockquote>\n  <p>O(conv) = 512*(1024*11*11*5*5)\/1 = 1 585 971 200<\/p>\n<\/blockquote>\n\n<p>If the same input is used for a fully connected layer, and the desired output is still 512 then:<\/p>\n\n<blockquote>\n  <p>O(fully_connected) = 512*1024*11*11 = 63 438 848<\/p>\n<\/blockquote>\n\n<p>Is this due to the more advanced methods for parallesing the convolutional layers on a GPU and the conv layer has more operations but less computation time cause of parallism issues? Or is my way of calculating the complexity of each layers simply wrong?<\/p>\n","answers":[{"AnswerId":"42108281","CreationDate":"2017-02-08T08:38:39.553","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>You can check if it is only the implementation by converting the fully-connected connections to equivalent convolutions. For every fully connected layer, there is an equivalent convolutional layer (see <a href=\"https:\/\/datascience.stackexchange.com\/q\/12830\/8820\">my question<\/a> for details and examples).<\/p>\n\n<ol>\n<li>You have <code>c<\/code> channels of size <code>w \u00d7 h<\/code> (hence the shape <code>c \u00d7 w \u00d7 h<\/code>) followed by a fully-connected layer with <code>n<\/code> nodes.<\/li>\n<li>Add a reshape layer after the channels to get <code>(c \u22c5 w \u22c5 h) \u00d7 1 \u00d7 1<\/code>.<\/li>\n<li>Add a convolutional layer with <code>n<\/code> filters of size <code>1 \u00d7 1<\/code>.<\/li>\n<\/ol>\n\n<p>Now check the time. If it is faster than the fully connected layer, then it is simply due to a better implementation of convolution.<\/p>\n"}]}
{"QuestionId":42093632,"AnswerCount":3,"Tags":"<tensorflow><deep-learning><keras>","CreationDate":"2017-02-07T15:27:43.700","AcceptedAnswerId":null,"OwnerUserId":7505030.0,"Title":"Semantic Segmentation with Encoder-Decoder CNNs","Body":"<p>Appologizes for misuse of technical terms.\nI am working on a project of semantic segmentation via CNNs ; trying to implement an architecture of type Encoder-Decoder, therefore output is the same size as the input.<\/p>\n\n<p>How do you design the labels ? \nWhat loss function should one apply ? Especially in the situation of heavy class inbalance (but the ratio between the classes is variable from image to image). <\/p>\n\n<p>The problem deals with two classes (objects of interest and background). I am using Keras with tensorflow backend.<\/p>\n\n<p>So far, I am going with designing expected outputs to be the same dimensions as the input images, applying pixel-wise labeling. Final layer of model has either softmax activation (for 2 classes), or sigmoid activation ( to express probability that the pixels belong to the objects class). I am having trouble with designing a suitable objective function for such a task, of type:<\/p>\n\n<p>function(y_pred,y_true),  <\/p>\n\n<p>in agreement with Keras. <\/p>\n\n<p>Please,try to be specific with the dimensions of tensors involved (input\/output of the model). Any thoughts and suggestions are much appreciated. Thank you !<\/p>\n","answers":[{"AnswerId":"42783368","CreationDate":"2017-03-14T10:24:21.727","ParentId":null,"OwnerUserId":"4485678","Title":null,"Body":"<p>Two ways :<\/p>\n\n<ol>\n<li><p>You could try 'flattening':<\/p>\n\n<pre><code>model.add(Reshape(NUM_CLASSES,HEIGHT*WIDTH))  #shape : HEIGHT x WIDTH x NUM_CLASSES\nmodel.add(Permute(2,1)) # now itll be NUM_CLASSES x HEIGHT x WIDTH\n#Use some activation here- model.activation()\n#You can use Global averaging or Softmax\n<\/code><\/pre><\/li>\n<li><p>One hot encoding every pixel:<\/p>\n\n<p>In this case your final layer should Upsample\/Unpool\/Deconvolve to HEIGHT x WIDTH x CLASSES. So your output is essentially of the shape:  (HEIGHT,WIDTH,NUM_CLASSES). <\/p><\/li>\n<\/ol>\n"},{"AnswerId":"42099071","CreationDate":"2017-02-07T20:12:26.543","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Actually when you use a <code>TensorFlow<\/code> backend you could simply apply a predefined <code>Keras<\/code> objectives in a following manner:<\/p>\n\n<pre><code>output = Convolution2D(number_of_classes, # 1 for binary case\n                       filter_height,\n                       filter_width,\n                       activation = \"softmax\")(input_to_output) # or \"sigmoid\" for binary\n... \nmodel.compile(loss = \"categorical_crossentropy\", ...) # or \"binary_crossentropy\" for binary\n<\/code><\/pre>\n\n<p>And then feed either a one-hot encoded feature map or matrix of shape <code>(image_height, image_width)<\/code> with integer encoded classes (remember than in this case you should use <code>sparse_categorical_crossentropy<\/code> as a loss).<\/p>\n\n<p>To deal with a class inbalance (I guess it's beacuse of a <em>backgroud<\/em> class) I strongly recommend you to read carefully answers to this Stack Overflow <a href=\"https:\/\/stackoverflow.com\/questions\/41898437\/possible-maybe-category-in-deep-learning\">question<\/a>. <\/p>\n"},{"AnswerId":"42111073","CreationDate":"2017-02-08T10:50:31.653","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>I suggest starting with a base architecture used in practice like this one in nerve-segmentation: <a href=\"https:\/\/github.com\/EdwardTyantov\/ultrasound-nerve-segmentation\" rel=\"nofollow noreferrer\">https:\/\/github.com\/EdwardTyantov\/ultrasound-nerve-segmentation<\/a>. Here a dice_loss is used as a loss function. This works very well for a two class problem as has been shown in literature: <a href=\"https:\/\/arxiv.org\/pdf\/1608.04117.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1608.04117.pdf<\/a>.<\/p>\n\n<p>Another loss function that has been widely used is cross entropy for such a problem. For problems like yours most commonly long and short skip connections are deployed to stabilize training as denoted in the paper above.  <\/p>\n"}]}
{"QuestionId":42094131,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2017-02-07T15:48:19.750","AcceptedAnswerId":null,"OwnerUserId":1908842.0,"Title":"Fail to download caffe model","Body":"<p>I tried the following command to download bvlc_alexnet model.<\/p>\n\n<p>python scripts\/download_model_binary.py models\/bvlc_alexnet<\/p>\n\n<p>But, it failed, displaying error message:\nFile \"\/usr\/lib\/python2.7\/socket.py\", line 575, in create_connection\n    raise err\nIOError: [Errno socket error] [Errno 110] Connection timed out<\/p>\n\n<p>The following url is also not working.\n<a href=\"http:\/\/dl.caffe.berkeleyvision.org\/bvlc_alexnet.caffemodel\" rel=\"nofollow noreferrer\">http:\/\/dl.caffe.berkeleyvision.org\/bvlc_alexnet.caffemodel<\/a><\/p>\n\n<p>Appreciate any help.<\/p>\n","answers":[{"AnswerId":"42103165","CreationDate":"2017-02-08T01:35:35.063","ParentId":null,"OwnerUserId":"68758","Title":null,"Body":"<p>It seems like an issue with your connection (which is odd since you posed the question). It could be a DNS issue which would explain why I can download the model you linked above and you can't. <\/p>\n\n<p>Try using the IP: 169.229.222.251 Which is what dl.caffe.berkeleyvision.org resolves to. You can also change the domain to the IP in the models\/bvlc_alexnet\/readme.md file to try the script again.<\/p>\n\n<p>I checked the domain here and it looks like there's an issue (at the time of this post): <a href=\"http:\/\/dnscheck.pingdom.com\/?domain=dl.caffe.berkeleyvision.org\" rel=\"nofollow noreferrer\">http:\/\/dnscheck.pingdom.com\/?domain=dl.caffe.berkeleyvision.org<\/a><\/p>\n"}]}
{"QuestionId":42094377,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-07T15:58:27.767","AcceptedAnswerId":42120970.0,"OwnerUserId":1538856.0,"Title":"TypeError running RNN tutorial from Tensorflow","Body":"<p>Simply trying to run the unmodified <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/\" rel=\"nofollow noreferrer\">RNN Tensorflow<\/a> tutorial on a Centos 7 machine.  I have d the machine.  Installed python3.6, pulled the tutorial from github, downloaded and untarred the sample data.  Other tutorials run.  Then I run:<\/p>\n\n<pre><\/pre>\n\n<p>Which returns:<\/p>\n\n<pre><\/pre>\n\n<p>As a total NOOB, I am completely stuck!  Can anyone help?<\/p>\n","answers":[{"AnswerId":"42120970","CreationDate":"2017-02-08T18:32:30.440","ParentId":null,"OwnerUserId":"1538856","Title":null,"Body":"<p>Per the comment on the question above, I installed <code>TensorFlow 1.0.0rc1<\/code>.   I used <code>pip3.6<\/code>, but as I was on Centos, the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup\" rel=\"nofollow noreferrer\">OS Setup guide<\/a> doesn't address it, so I had to try each download:<\/p>\n\n<pre><code>(tensorflow) [tensorflow]# export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0rc1-cp35-cp35m-linux_x86_64.whl\n(tensorflow) [tensorflow]# pip3.6 install --upgrade $TF_BINARY_URL\ntensorflow-1.0.0rc1-cp35-cp35m-linux_x86_64.whl is not a supported wheel on this platform.\n(tensorflow) [tensorflow]# export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0rc1-cp34-cp34m-linux_x86_64.whl\n(tensorflow) [tensorflow]# pip3.6 install --upgrade $TF_BINARY_URL\ntensorflow-1.0.0rc1-cp34-cp34m-linux_x86_64.whl is not a supported wheel on this platform.\n(tensorflow) [tensorflow]# export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0rc1-cp27-none-linux_x86_64.whl\n(tensorflow) [tensorflow]# pip3.6 install --upgrade $TF_BINARY_URL\ntensorflow-1.0.0rc1-cp27-none-linux_x86_64.whl is not a supported wheel on this platform.\n(tensorflow) [tensorflow]# export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0rc1-cp33-cp33m-linux_x86_64.whl\n(tensorflow) [tensorflow]# pip3.6 install --upgrade $TF_BINARY_URL\ntensorflow-1.0.0rc1-cp33-cp33m-linux_x86_64.whl is not a supported wheel on this platform.\n(tensorflow) [tensorflow]# export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0rc1-cp36-cp36m-linux_x86_64.whl\n(tensorflow) [tensorflow]# pip3.6 install --upgrade $TF_BINARY_URL\n    Collecting tensorflow==1.0.0rc1 from https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0rc1-cp36-cp36m-linux_x86_64.whl\n  Downloading https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0rc1-cp36-cp36m-linux_x86_64.whl (43.5MB)\n    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| 43.5MB 28kB\/s \nRequirement already up-to-date: six&gt;=1.10.0 in \/usr\/lib\/python3.6\/site-packages (from tensorflow==1.0.0rc1)\nRequirement already up-to-date: wheel&gt;=0.26 in \/usr\/lib\/python3.6\/site-packages (from tensorflow==1.0.0rc1)\nRequirement already up-to-date: protobuf&gt;=3.1.0 in \/usr\/lib\/python3.6\/site-packages (from tensorflow==1.0.0rc1)\nRequirement already up-to-date: numpy&gt;=1.11.0 in \/usr\/lib64\/python3.6\/site-packages (from tensorflow==1.0.0rc1)\nRequirement already up-to-date: setuptools in \/usr\/lib\/python3.6\/site-packages (from protobuf&gt;=3.1.0-&gt;tensorflow==1.0.0rc1)\nRequirement already up-to-date: packaging&gt;=16.8 in \/usr\/lib\/python3.6\/site-packages (from setuptools-&gt;protobuf&gt;=3.1.0-&gt;tensorflow==1.0.0rc1)\nRequirement already up-to-date: appdirs&gt;=1.4.0 in \/usr\/lib\/python3.6\/site-packages (from setuptools-&gt;protobuf&gt;=3.1.0-&gt;tensorflow==1.0.0rc1)\nRequirement already up-to-date: pyparsing in \/usr\/lib\/python3.6\/site-packages (from packaging&gt;=16.8-&gt;setuptools-&gt;protobuf&gt;=3.1.0-&gt;tensorflow==1.0.0rc1)\nInstalling collected packages: tensorflow\n  Found existing installation: tensorflow 0.12.1\n    Uninstalling tensorflow-0.12.1:\n      Successfully uninstalled tensorflow-0.12.1\nSuccessfully installed tensorflow-1.0.0rc1\n<\/code><\/pre>\n\n<p>So the correct URL for <code>RHEL\/Centos<\/code> is: <code>tensorflow-1.0.0rc1-cp36-cp36m-linux_x86_64.whl<\/code>.  Verify your upgrade by running:<\/p>\n\n<pre><code>(tensorflow) [tensorflow]# python3.6 -c 'import tensorflow as tf; print(tf.__version__)'\n1.0.0-rc1\n<\/code><\/pre>\n\n<p>Now run the tutorial again, the <code>TypeError<\/code> is gone, but is replaced by an <code>AttributeError<\/code>:<\/p>\n\n<pre><code>(tensorflow) [tensorflow]# python3.6 ptb_word_lm.py --data_path=simple-examples\/data\/\nTraceback (most recent call last):\n  ...\n  File \"ptb_word_lm.py\", line 149, in __init__\n     output = tf.reshape(tf.concat_v2(outputs, 1), [-1, size])\nAttributeError: module 'tensorflow' has no attribute 'concat_v2'\n<\/code><\/pre>\n\n<p>Turns out that TWO tensorflow APIs changed: <\/p>\n\n<ol>\n<li><code>concat_v2<\/code> goes away and is replaced simply by: <code>concat<\/code><\/li>\n<li>Likewise, <code>scalar_summary<\/code> has been renamed to: <code>summary.scalar<\/code>.  There are 3 such references in ptb_word_lm.py<\/li>\n<\/ol>\n\n<p>Edit <code>ptb_word_lm.py<\/code> to make both of the changes above AND the tutorial should now run correctly for you!<\/p>\n"}]}
{"QuestionId":42094542,"AnswerCount":3,"Tags":"<python><numpy><matrix><theano>","CreationDate":"2017-02-07T16:05:27.863","AcceptedAnswerId":null,"OwnerUserId":245549.0,"Title":"How to apply outer product for tensors without unnecessary increase of dimensions?","Body":"<p>I have two vectors  and  and I want to make a matrix  out of them such that:<\/p>\n\n<pre><\/pre>\n\n<p>In other words I want to calculate the <strong>outer product<\/strong> of them. I can do it either by using  or by adding new indexes to  and  and using the  product.<\/p>\n\n<pre><\/pre>\n\n<p>Now, I try to solve a bit more general problem. Instead of two vectors  and  I have two matrices (I call them  and  again) and I would like to calculate an outer product of each row from matrix  with the correspondent row of matrix  (i_th row in the first matrix should be multiplied with the i_th row of the second matrix). So, I would like to do something like that:<\/p>\n\n<pre><\/pre>\n\n<p>In other words,  is the matrix corresponding to outer product of  row from the matrix  and  row of the matrix .<\/p>\n\n<p>I see two problems with the above \"solution\". First, it is not really a solution, since the second line of the code is not a proper theano code. So, my first question is how to do this \"advanced slicing\" by forcing some indexes to be equal. For example . Second, I do not like the fact that I first create a 4D tesnor () from two 2D tensors and then I reduce it back to a 3D tensor. It can be very memory consuming. I guess one can avoid it.<\/p>\n","answers":[{"AnswerId":"42220153","CreationDate":"2017-02-14T07:25:41.237","ParentId":null,"OwnerUserId":"4427777","Title":null,"Body":"<p>I can't believe nobody has tried to use <code>np.einsum<\/code>.<\/p>\n\n<pre><code>w\narray([[1, 8, 9, 2],\n       [1, 2, 9, 0],\n       [5, 8, 7, 3],\n       [2, 9, 8, 2]])\n\nv \narray([[1, 4, 5, 9],\n       [9, 1, 3, 7],\n       [9, 6, 1, 5],\n       [4, 9, 7, 0]])\n\nfor i in range(w.shape[0]):\n     print(np.outer(w[i], v[i]))\n\n[[ 1  4  5  9]\n [ 8 32 40 72]\n [ 9 36 45 81]\n [ 2  8 10 18]]\n[[ 9  1  3  7]\n [18  2  6 14]\n [81  9 27 63]\n [ 0  0  0  0]]\n[[45 30  5 25]\n [72 48  8 40]\n [63 42  7 35]\n [27 18  3 15]]\n[[ 8 18 14  0]\n [36 81 63  0]\n [32 72 56  0]\n [ 8 18 14  0]]\n\nnp.einsum('ij,ik-&gt;ijk', w, v)\n\narray([[[ 1,  4,  5,  9],\n        [ 8, 32, 40, 72],\n        [ 9, 36, 45, 81],\n        [ 2,  8, 10, 18]],\n\n       [[ 9,  1,  3,  7],\n        [18,  2,  6, 14],\n        [81,  9, 27, 63],\n        [ 0,  0,  0,  0]],\n\n       [[45, 30,  5, 25],\n        [72, 48,  8, 40],\n        [63, 42,  7, 35],\n        [27, 18,  3, 15]],\n\n       [[ 8, 18, 14,  0],\n        [36, 81, 63,  0],\n        [32, 72, 56,  0],\n        [ 8, 18, 14,  0]]])\n<\/code><\/pre>\n\n<p>It looks like the equivalent Theano function is <code>theano.tensor.batched_dot<\/code> (which is supposed to be even faster than <code>einsum<\/code>), but I have no experience with Theano.<\/p>\n"},{"AnswerId":"42141982","CreationDate":"2017-02-09T16:30:29.310","ParentId":null,"OwnerUserId":"3220135","Title":null,"Body":"<p>you looking for something like this?<\/p>\n\n<pre><code>&gt;&gt;&gt; a = b = np.arange(8).reshape([2,4])\n&gt;&gt;&gt; a[:,None,:]*b[:,:,None]\narray([[[ 0,  0,  0,  0],\n        [ 0,  1,  2,  3],\n        [ 0,  2,  4,  6],\n        [ 0,  3,  6,  9]],\n\n       [[16, 20, 24, 28],\n        [20, 25, 30, 35],\n        [24, 30, 36, 42],\n        [28, 35, 42, 49]]])\n<\/code><\/pre>\n"},{"AnswerId":"42143688","CreationDate":"2017-02-09T17:55:38.723","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>We need to introduce broadcastable dimensions into the two input matrices with <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.dimshuffle\" rel=\"nofollow noreferrer\"><code>dimshuffle<\/code><\/a> and then let <code>broadcasting<\/code> take care of the elementwise multiplication resulting in outer-product between coresponding rows of them.<\/p>\n\n<p>Thus, with <code>V<\/code> and <code>W<\/code> as the theano matrices, simply do -<\/p>\n\n<pre><code>V.dimshuffle(0, 1, 'x')*W.dimshuffle(0, 'x', 1)\n<\/code><\/pre>\n\n<p>In <code>NumPy<\/code>, we have <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/arrays.indexing.html#numpy.newaxis\" rel=\"nofollow noreferrer\"><code>np.newaxis<\/code><\/a> to extend dimensions and <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.transpose.html\" rel=\"nofollow noreferrer\"><code>np.transpose()<\/code><\/a> for permuting dimensions. With <code>theno<\/code>, there's <code>dimshuffle<\/code> to do both of these tasks with a mix of listing dimension IDs and <code>x<\/code>'s for introducing new <em>broadcast-able<\/em> axes.<\/p>\n\n<p><strong>Sample run<\/strong><\/p>\n\n<p>1) Inputs :<\/p>\n\n<pre><code># Numpy arrays\nIn [121]: v = np.random.randint(11,99,(3,4))\n     ...: w = np.random.randint(11,99,(3,5))\n     ...: \n\n# Perform outer product on corresponding rows in inputs\nIn [122]: for i in range(v.shape[0]):\n     ...:     print(np.outer(v[i],w[i]))\n     ...:     \n[[2726 1972 1740 2117 1972]\n [8178 5916 5220 6351 5916]\n [7520 5440 4800 5840 5440]\n [8648 6256 5520 6716 6256]]\n[[8554 3458 8918 4186 4277]\n [1786  722 1862  874  893]\n [8084 3268 8428 3956 4042]\n [2444  988 2548 1196 1222]]\n[[2945 2232 1209  372  682]\n [2565 1944 1053  324  594]\n [7125 5400 2925  900 1650]\n [6840 5184 2808  864 1584]]\n<\/code><\/pre>\n\n<p>2) Theano part :<\/p>\n\n<pre><code># Get to theano : Get the theano matrix versions \nIn [123]: V = T.matrix('v')\n     ...: W = T.matrix('w')\n     ...: \n\n# Use proposed code\nIn [124]: OUT = V.dimshuffle(0, 1, 'x')*W.dimshuffle(0, 'x', 1)\n\n# Create a function out of it and then use on input NumPy arrays\nIn [125]: f = function([V,W], OUT)\n<\/code><\/pre>\n\n<p>3) Verify results :<\/p>\n\n<pre><code>In [126]: f(v,w)    # Verify results against the earlier loopy results\nOut[126]: \narray([[[ 2726.,  1972.,  1740.,  2117.,  1972.],\n        [ 8178.,  5916.,  5220.,  6351.,  5916.],\n        [ 7520.,  5440.,  4800.,  5840.,  5440.],\n        [ 8648.,  6256.,  5520.,  6716.,  6256.]],\n\n       [[ 8554.,  3458.,  8918.,  4186.,  4277.],\n        [ 1786.,   722.,  1862.,   874.,   893.],\n        [ 8084.,  3268.,  8428.,  3956.,  4042.],\n        [ 2444.,   988.,  2548.,  1196.,  1222.]],\n\n       [[ 2945.,  2232.,  1209.,   372.,   682.],\n        [ 2565.,  1944.,  1053.,   324.,   594.],\n        [ 7125.,  5400.,  2925.,   900.,  1650.],\n        [ 6840.,  5184.,  2808.,   864.,  1584.]]])\n<\/code><\/pre>\n"}]}
{"QuestionId":42095575,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-02-07T16:53:46.013","AcceptedAnswerId":42097355.0,"OwnerUserId":6005072.0,"Title":"How are the past gradients accumulated in Momentum\/Adam when we preprocess the gradients before tf.train.Optimizer.apply_gradients()?","Body":"<p>I want to preprocess the gradients before , and want the past gradients to be accumulated upon <strong>processed<\/strong> gradients when  or  is used. I know we can preprocess the gradients between  and  as shown <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/train\/optimizers\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>My question is, in the above case, is the history gradients accumulated upon the <strong>capped<\/strong> gradients or <strong>non-capped<\/strong> ones?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42097355","CreationDate":"2017-02-07T18:32:56.890","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>All of the state that optimizers keep is updated in apply_gradients. There is a bit of a complicated call chain (best followed in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/optimizer.py\" rel=\"nofollow noreferrer\">optimizer.py<\/a>), but the short of it is that apply_gradients eventually calls one of apply_sparse or apply_dense (ignoring resource variables).<\/p>\n\n<p>Going back to Adam, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/adam.py#L154\" rel=\"nofollow noreferrer\">apply_sparse<\/a> is relatively easy to read since it's an agglomeration of ops rather than a single C++ op. You can see that it updates all of the moments and the variables.<\/p>\n\n<p>So to answer your question, if you cap the gradients before calling apply_gradients, then capped gradients will be accumulated in Adam's moments (and likewise for other optimizers).<\/p>\n\n<p>There is a bit of a gotcha if you're dealing with sparse gradients (IndexedSlices), since these are disaggregated as they pass through the graph. So if you cap them in the disaggregated form, repeated indices may sum up to more than your cap. This will only be an issue if you're doing gather() or using embeddings, but it's worth keeping in mind.<\/p>\n"}]}
{"QuestionId":42095625,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-07T16:56:35.970","AcceptedAnswerId":42095969.0,"OwnerUserId":7528619.0,"Title":"What does the function control_dependencies do?","Body":"<p>I would like to have an example illustrating the use of the function <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/control_dependencies\" rel=\"noreferrer\"><\/a>. For example, I want to create two tensors  and  and if they are equal do or print something. <\/p>\n\n<pre><\/pre>\n\n<p>In the code above,  is clearly not equal to . What is  doing in this case?<\/p>\n","answers":[{"AnswerId":"42095969","CreationDate":"2017-02-07T17:14:09.373","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p><code>control_dependencies<\/code> is not a conditional. It is a mechanism to add dependencies to whatever ops you create in the <code>with<\/code> block. More specifically, what you specify in the argument to <code>control_dependencies<\/code> is ensured to be evaluated before anything you define in the <code>with<\/code> block. <\/p>\n\n<p>In your example, you don't add any (TensorFlow) operations in the <code>with<\/code> block, so the block does nothing.<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/33950177\/524436\">This answer<\/a> has an example of how to use <code>control_dependencies<\/code>, where it is used to make sure the assignments happen before the batchnorm operations are evaluated.<\/p>\n"}]}
{"QuestionId":42096015,"AnswerCount":2,"Tags":"<python><import><caffe>","CreationDate":"2017-02-07T17:16:48.087","AcceptedAnswerId":53809328.0,"OwnerUserId":6129428.0,"Title":"How to add caffe to anaconda on windows?","Body":"<p>I'm trying to use <strong>Caffe<\/strong> in <strong>Anaconda3<\/strong> on windows, but when I try to import caffe in iPython, it shows this error:<\/p>\n\n<blockquote>\n  <p>ImportError: No module named caffe<\/p>\n<\/blockquote>\n\n<p>I tried to add a new system variable named <strong>PYTHONPATH<\/strong> which looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>It still shows the same error. I would appreciate if anyone mentioned what am I missing here?<\/p>\n","answers":[{"AnswerId":"53809328","CreationDate":"2018-12-17T05:17:20.850","ParentId":null,"OwnerUserId":"3367646","Title":null,"Body":"<p>REM Create a new conda environment to install Caffe\nREM You can use python 3.5, 3.6 or 3.7<\/p>\n\n<pre><code>conda create -n caffe python=3.7\n<\/code><\/pre>\n\n<p>REM activate the environment\nREM Depending on your conda version you may have to use activate caffe<\/p>\n\n<pre><code>conda activate caffe\n<\/code><\/pre>\n\n<p>REM add the anaconda channel to resolve all dependencies<\/p>\n\n<pre><code>conda config --add channels anaconda\n<\/code><\/pre>\n\n<p>REM Install caffe with cuda support or install caffe-cpu if you do not want CUDA support<\/p>\n\n<pre><code>conda install caffe -c willyd\n<\/code><\/pre>\n\n<p>Answer taken from here: <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/6569#issuecomment-432491851\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/issues\/6569#issuecomment-432491851<\/a><\/p>\n"},{"AnswerId":"58601574","CreationDate":"2019-10-29T04:52:22.440","ParentId":null,"OwnerUserId":"11266523","Title":null,"Body":"<pre><code>{conda create -n caffe python=x.x}    \n{conda activate caffe}     \n{conda config --add channels anaconda}     \n{conda install caffe -c willyd}   \n<\/code><\/pre>\n\n<p>Change the version of python, it should work then.\nAnswer taken from here: <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/6569#issuecomment-432491851\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/issues\/6569#issuecomment-432491851<\/a> <\/p>\n"}]}
{"QuestionId":42096358,"AnswerCount":1,"Tags":"<json><tensorflow><protocol-buffers>","CreationDate":"2017-02-07T17:36:29.643","AcceptedAnswerId":42100674.0,"OwnerUserId":3002273.0,"Title":"TensorFlow Shape and Type Inference","Body":"<p>I learned that I can program TensorFlow by modifying JSON that come from protobuf. See <a href=\"https:\/\/stackoverflow.com\/questions\/41575442\/tensorflow-export-compute-graph-to-xml-json-etc\/41576075\">here<\/a><\/p>\n\n<p>If I modify this JSON, then I sometimes have a problem where I need to manually edit the JSON to propagate the proper shapes of the various inputs and outputs. Is there a way to make TF do this automatically for me so that I can specify inputs via Placeholders and then the shapes and types are automatically propagated? <\/p>\n","answers":[{"AnswerId":"42100674","CreationDate":"2017-02-07T21:49:14.553","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>If you know which types of modifications you're going to be making, you could just remove that shape information from the placeholders. The uncertainty will propagate automatically. For example:<\/p>\n\n<pre><code>import tensorflow as tf\nplaceholder = tf.placeholder(dtype=tf.float32, shape=[None])\nderived = (placeholder \/ 3)[1:, None]\nprint(placeholder.get_shape(), derived.get_shape())\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>(TensorShape([Dimension(None)]), TensorShape([Dimension(None), Dimension(1)]))\n<\/code><\/pre>\n\n<p>So there will be no static shape information saved for the length of <code>placeholder<\/code>. You can even have Tensors with unknown rank.<\/p>\n\n<p>Re-computing static shapes is a tempting thought, but this is not currently supported because graph construction can depend on static shape information. For example:<\/p>\n\n<pre><code>placeholder = tf.placeholder(dtype=tf.float32, shape=[2])\nif placeholder.get_shape()[0].value % 2 == 0:\n    derived = placeholder\nelse:\n    derived = tf.concat(0, [placeholder, [0]])\n<\/code><\/pre>\n\n<p>This isn't a recommended graph construction technique (much better to use <code>tf.shape<\/code> with <code>cond<\/code>), but it does happen. This kind of static-shape-conditional graph structure is not captured in the meta-graph, unfortunately.<\/p>\n"}]}
{"QuestionId":42096647,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-02-07T17:52:49.127","AcceptedAnswerId":null,"OwnerUserId":7530285.0,"Title":"What is the most optimal approach for training Inception v3 on a multi-label dataset like CIFAR-100?","Body":"<p>I am trying to train Inception v3 on <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow noreferrer\">CIFAR-100 dataset<\/a>, where each image belongs to a class and a superclass as indicated by its \"fine\" label and \"coarse\" labels. My goal is to get the model to infer the image class and with greater accuracy its superclass. For example, predict with 95% certainty that a photo of an apple contains \"fruit and vegetables\" and with 80% certainty that it contains \"apples\".<\/p>\n\n<p>What would be the most optimal approach assuming I already have generated a TFRecords file with the both labels included:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42096921,"AnswerCount":1,"Tags":"<python><scikit-learn><keras>","CreationDate":"2017-02-07T18:09:04.453","AcceptedAnswerId":42097513.0,"OwnerUserId":1870832.0,"Title":"Passing input_dim to KerasClassifier (sklearn wrapper\/interface)","Body":"<p>I originally tried the same approach and ran into the same error as <a href=\"https:\/\/stackoverflow.com\/questions\/40393629\/how-to-pass-a-parameter-to-scikit-learn-keras-model-function\">this SO questioner<\/a>. However, using the accepted (and only) answer there gave me another error: \"input_dim is not a legal parameter.\"<\/p>\n\n<p>I then tried to use the solution (\"add an input_dim keyarg to the KerasClassifier constructor\") on the original question, and got the same error again. Am I doing something wrong, or is there a new way to pass a first layer's input_dim through the sklearn wrapper KerasClassifier now?<\/p>\n\n<p>Minimal code example below:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>ValueError: input_dim is not a legal parameter<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42097513","CreationDate":"2017-02-07T18:42:21.620","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You need to pass the <code>input_dim<\/code> as one of the parameters to <code>create_model()<\/code> <\/p>\n\n<pre><code>def create_model(input_dim):\n    # create model\n    model = Sequential()\n    # model.add(Dense(12, input_dim=4, init='uniform', activation='relu'))\n    model.add(Dense(12, input_dim=input_dim, init='uniform', activation='relu'))\n<\/code><\/pre>\n\n<p>Though you need not use <code>input_dim<\/code> inside <code>create_model<\/code> for the sake of just removing the error.<\/p>\n"}]}
{"QuestionId":42097115,"AnswerCount":1,"Tags":"<performance><tensorflow><gpu><cpu><keras>","CreationDate":"2017-02-07T18:19:27.033","AcceptedAnswerId":null,"OwnerUserId":6101419.0,"Title":"Keras (Tensorflow backend) slower on GPU than on CPU when training certain networks","Body":"<p>I am having some difficulty understanding exactly why the GPU and CPU speeds are similar with networks of small size (CPU is sometimes faster), and GPU is faster with networks of larger size. The code at the bottom of the question runs in 103.7s on an i7-6700k, but when using tensorflow-gpu, the code runs in 29.5 seconds.<\/p>\n\n<p>However, when I train a network that has 100 hidden neurons, instead of 1000 like in the example below, I get ~20 seconds when using the GPU, and ~15 seconds when using the CPU.<\/p>\n\n<p>I read on another stack overflow answer that CPU->GPU transfers take long, I'm assuming this is in reference to loading the data examples on the GPU.<\/p>\n\n<p>Can someone explain why this occurs, and possibly reference some change in the code that I can make to maximize speed?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"46949749","CreationDate":"2017-10-26T08:50:07.690","ParentId":null,"OwnerUserId":"4520330","Title":null,"Body":"<p>In case of tiny networks batch loading may be the culprit here.<\/p>\n\n<p>Keras is loading each minibatch from RAM to GPU at the start of each iteration, thus creating a bottleneck in tiny networks (where forward\/backward computation is very quick).<br>\nYou can try using <code>model.fit_generator<\/code> instead of plain <code>fit<\/code>, so that CPU thread which loads minibatches works in parallel. <\/p>\n\n<p>Unfortunately, there is no way I am aware of to preload the whole dataset on GPU for Keras (see <a href=\"https:\/\/stackoverflow.com\/questions\/46688822\/preload-whole-dataset-on-gpu-for-training-keras-model\">my issue<\/a>)<\/p>\n\n<p>If you're using Tensorflow backend, you can use Google Timeline profiling tool to see what causes the slowdowns. For the reference, see <a href=\"https:\/\/stackoverflow.com\/questions\/44050661\/keras-real-amount-of-gpu-memory-used\">this issue<\/a><\/p>\n"}]}
{"QuestionId":42099701,"AnswerCount":0,"Tags":"<tensorflow><windows-10><jupyter-notebook>","CreationDate":"2017-02-07T20:48:03.457","AcceptedAnswerId":null,"OwnerUserId":2908528.0,"Title":"AttributeError: 'Graph' object has no attribute 'SerializeToString'","Body":"<p>I just wrote a simple program to test tensorflow library in my windows 10.\nIt is running well but only problem is that I am not being able to use tensorboard to visualize graphs.<\/p>\n\n<p><strong>Following is my code<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Error log<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Tensorflow version<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Tensorflow library upgrade<\/strong><\/p>\n\n<p>When I tried to upgrade tensorflow library to latest 0.12 using pip in jupyter notebook it downloaded and installed lot of stuffs and after that I again tested the version but it was same.So I again tried to upgrade but it displayed following message.<\/p>\n\n<p>Requirement already up-to date<\/p>\n\n<p>How to get out of this mess?<\/p>\n","answers":[]}
{"QuestionId":42101041,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-02-07T22:14:41.327","AcceptedAnswerId":null,"OwnerUserId":2202866.0,"Title":"Keras error when finetuning InceptionV3","Body":"<p>I am trying to follow the \"<a href=\"https:\/\/keras.io\/applications\/#inceptionv3\" rel=\"nofollow noreferrer\">Fine-tune InceptionV3 on a new set of classes\"<\/a> sample code to freeze the first 172 layers and re-train the last layers on cats\/dogs dataset. I keep getting an error which I have noted at the bottom. Please help. I am using Ubuntu 16.04, keras 1.2.1, theano 0.9.0beta1.dev, numpy 1.12.0 and python 3.5.<\/p>\n\n<pre><\/pre>\n\n<p>The exception is: padding must be zero for average_exc_pad<\/p>\n\n<p>Here is the full stack-trace:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42102899","CreationDate":"2017-02-08T01:06:15.577","ParentId":null,"OwnerUserId":"4955956","Title":null,"Body":"<p>Fine-tuning in that situation possibly means using the convolutional layers as pre-trained feature extractors. So you don't really want the top layers (densely connected layers) of the Inception network.<\/p>\n\n<p>Changing <\/p>\n\n<pre><code>base_model = InceptionV3(weights='imagenet', include_top=True)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>base_model = InceptionV3(weights='imagenet', include_top=False)\n<\/code><\/pre>\n\n<p>should work.<\/p>\n\n<p>Also, if you have 200 classes you should change <\/p>\n\n<pre><code># and a logistic layer -- let's say we have 200 classes\npredictions = Dense(2, activation='softmax')(x)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>predictions = Dense(200, activation='softmax')(x)\n<\/code><\/pre>\n\n<p>So your last layer will have the desired 200 elements.<\/p>\n"}]}
{"QuestionId":42101285,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-07T22:32:30.190","AcceptedAnswerId":42102008.0,"OwnerUserId":6093883.0,"Title":"What types of operations are permitted inside tf.cond","Body":"<p>It seems that  executes possible dependencies for both branches, so any computation we would like to perform if and only if the conditions hold have to be put into the function fn1 fn2. <\/p>\n\n<p>However I am confused as to what fn actually is. Every variable\/op in tensorflow should be a node of the computation graph, but fn is actually a python function. This leads to many questions. For example, is this function re-evaluated every time  is executed? Can this function return different computation graphs each time? Can placeholders be defined in them, and if not how to avoid supplying values to placeholders we know will not be used when, for example, there is a switch variable that chooses between different inputs? <\/p>\n","answers":[{"AnswerId":"42102008","CreationDate":"2017-02-07T23:34:54.633","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>The functions passed to <code>tf.cond<\/code> are only run when the op is defined, not during graph execution. And <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/control_flow_ops.py#L1760\" rel=\"nofollow noreferrer\">both of them are run<\/a>, exactly once as far as I can see. The functions themselves are just a way to indicate exactly which ops should have the conditional execution behavior: note the <code>context_t.Enter()<\/code>\/<code>context_t.Exit()<\/code> calls surrounding each function call.<\/p>\n\n<p>Hopefully that clarifies things. The functions are a useful way of grouping ops during graph definition. There's no function execution magic going on in the TensorFlow graph.<\/p>\n"}]}
{"QuestionId":42102174,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-07T23:49:50.340","AcceptedAnswerId":42102395.0,"OwnerUserId":3291312.0,"Title":"Tensorflow InvalidArgumentError (see above for traceback): Incompatible shapes: [10000,10] vs. [10000]","Body":"<p>I'm trying my first attempt at tensorflow - I'm attempting to learn a linear regressor by taking an input vector of 10 dimensions, X and outputting a scalar, Y. Specifically I'm trying to use the closed-form gradient based solution.<\/p>\n\n<p>I've received the following error and I'm not sure what I've done wrong. Anything to point me in the right direction would be greatly appreciated!<\/p>\n\n<pre><\/pre>\n\n<p>Here is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42102395","CreationDate":"2017-02-08T00:12:29.823","ParentId":null,"OwnerUserId":"5067311","Title":null,"Body":"<p>I don't know tensorflow and I'm not sure exactly what's going on in your code, so I'm trying to make an educated guess based on numpy's behaviour. I was going to add this as a comment, but it got too long.<\/p>\n\n<p>When you load the training data, <code>train_X<\/code> is of shape <code>(10000,10)<\/code> since it has 10 columns (it's a 2d array), while <code>train_Y<\/code> is of shape <code>(10000,)<\/code>, since it is a single column (it's a 2d array). These two shapes can't be broadcast together, so <code>pred<\/code> and <code>Y<\/code> in <code>pred-Y<\/code> have incompatible shapes. You either need to transpose <code>train_X<\/code> for this, or turn <code>train_Y<\/code> into an array of shape <code>(10000,1)<\/code> for them to be compatible. The former you can do by passing <code>unpack=True<\/code> to <code>np.loadtxt<\/code>; the latter with something like <code>train_Y = train_Y[:,None]<\/code> (at least my suspicion is that <code>unpack=True<\/code> won't help in this case, but it's worth a try anyway).<\/p>\n\n<p>But in case you transpose your arrays, you need to be careful that your training loop still works. Currently your <code>(10000,10)<\/code>-shaped array is equivalent to a length-10000 list of length-ten lists, while your <code>(10000,)<\/code>-shaped array is equivalent to a single length-10000 list. These can nicely be <code>zip<\/code>ped together. If you transpose, say <code>train_X<\/code> to make broadcasting work, then you'll need to modify this loop:<\/p>\n\n<pre><code>    for (x, y) in zip(train_X.T, train_Y):\n        sess.run(optimizer, feed_dict = {X: x, Y: y})\n<\/code><\/pre>\n\n<p>In hindsight, I just probably mean that you need to transpose back <code>train_X<\/code> in order to ensure that the first dimensions of <code>train_X<\/code> and <code>train_Y<\/code> match up for zipping.<\/p>\n\n<p>Transposition might affect later steps too, such as plotting. In case there's any weird output or errors, you need to transpose back. Or even better: only transpose for the tensorflow-specific operations (but I'm unfamiliar with this part, so I don't know if and how this could be done idiomatically).<\/p>\n"}]}
{"QuestionId":42102998,"AnswerCount":0,"Tags":"<c++><database><stream><tensorflow><protocol-buffers>","CreationDate":"2017-02-08T01:17:24.700","AcceptedAnswerId":null,"OwnerUserId":3002273.0,"Title":"TensorFlow Database Streaming Operations","Body":"<p>I would like to write a custom C++ operation in TensorFlow that streams data out in such a way that I can replace the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow noreferrer\">Placeholder\/Feed<\/a> way of streaming data into the compute graph.<\/p>\n\n<pre><\/pre>\n\n<p>I would like to have something that looked like,<\/p>\n\n<pre><\/pre>\n\n<p>Concerns are signaling that there isn't anymore data. I would like to have this compiled into a shared library like in <a href=\"https:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/\" rel=\"nofollow noreferrer\">Adding an Op<\/a> which <a href=\"https:\/\/www.tensorflow.org\/how_tos\/new_data_formats\/\" rel=\"nofollow noreferrer\">Custom Data Readers<\/a> appears to not support due to header dependency. <\/p>\n\n<p>I looked at Queues, but one of my nit picky points is that I need to do this in such a way that I have a graph element that has the database connection information. I need this because I need to do a <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/how_tos\/meta_graph\/\" rel=\"nofollow noreferrer\">MetaGraph dump<\/a> and reload.<\/p>\n","answers":[]}
{"QuestionId":42103182,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-08T01:36:53.210","AcceptedAnswerId":42103279.0,"OwnerUserId":7520802.0,"Title":"How to predict unlabeled images in Tensorflow","Body":"<p>I've been going through several Tensorflow tutorials and have not seen anything on using a model after it's been trained\/tested with. I looked through stackoverflow and found a few solutions that aren't working for me like <a href=\"https:\/\/stackoverflow.com\/questions\/33711556\/making-predictions-with-a-tensorflow-model\">here<\/a><\/p>\n\n<p>So I'm using the code <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/convolutional_network.py\" rel=\"nofollow noreferrer\">here<\/a> with the exception that I altered the code so I could try running predictions afterwards instead of the session closing. For the prediction I'm just using a test sample, but trying to work it out without giving a label. I want to see what class is predicted.<\/p>\n\n<pre><\/pre>\n\n<p>From the stack overflow page I listed above I should be able to do something like this<\/p>\n\n<pre><\/pre>\n\n<p>Though it looks like this has been removed as tensorflow is saying there is not run function. A comment in that same page suggested to do this<\/p>\n\n<pre><\/pre>\n\n<p>but I'm getting this error<\/p>\n\n<pre><\/pre>\n\n<p>With this I found I needed to run what was stated above but I'm getting an issue with the tensor being the incorrect size<\/p>\n\n<pre><\/pre>\n\n<p>So from here it looks like the data is not aligned correctly? I've tried using reshape with no success. If anyone could point me into the right direction so I can figure out how to actual use my model for applications that would be great.<\/p>\n\n<p>EDIT: Here's a simpler program. I'm having the same issue<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting the same issue as above<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42103279","CreationDate":"2017-02-08T01:48:12.783","ParentId":null,"OwnerUserId":"2723734","Title":null,"Body":"<p>The link to your notebook is a local link, so no one else can see it.  Also, you are missing some code.  Where is \"y\" defined?  <\/p>\n\n<p>All you need to do is to evaluate y, e.g.: y.eval(feed_dict={x: NEW_DATA})<\/p>\n"}]}
{"QuestionId":42104176,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-08T03:31:22.103","AcceptedAnswerId":42107308.0,"OwnerUserId":7493814.0,"Title":"When I run cifar10 on tensorflow, it occurs AttributeError: no attribute 'random_crop'","Body":"<p>I download cifar10 code with tensorflow from the internet, and when I run , It occurs an error below:<\/p>\n\n<pre><\/pre>\n\n<p>Is there anyone knows what happened? Thanks so much!<\/p>\n","answers":[{"AnswerId":"42107308","CreationDate":"2017-02-08T07:39:44.757","ParentId":null,"OwnerUserId":"1844390","Title":null,"Body":"<p>This code is maybe from an older version of tensorflow than the one you installed.<\/p>\n\n<p>Try changing the line <\/p>\n\n<pre><code>tf.image.random_crop(reshaped_image, [height, width])\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>tf.random_crop(reshaped_image, [height, width])\n<\/code><\/pre>\n"}]}
{"QuestionId":42104395,"AnswerCount":1,"Tags":"<php><python><anaconda><theano>","CreationDate":"2017-02-08T03:54:42.550","AcceptedAnswerId":null,"OwnerUserId":5574453.0,"Title":"Working Python script throw out error when PHP calls it with exec","Body":"<p>I have a simple php code:<\/p>\n\n<pre><\/pre>\n\n<p>And  simply import a library called <\/p>\n\n<pre><\/pre>\n\n<p>It works under ssh but throw out:<\/p>\n\n<pre><\/pre>\n\n<p>when looking at the php in browser.<\/p>\n\n<p>What's happeneing here? Is there any way that I can see full trace of error msg?<\/p>\n","answers":[{"AnswerId":"42104725","CreationDate":"2017-02-08T04:27:30.533","ParentId":null,"OwnerUserId":"5347875","Title":null,"Body":"<p>You can try this for the PHP side of things, but I think KeyError is a Python thing:<\/p>\n\n<pre><code>&lt;?php \n\n$output = array();\n\nexec('\/opt\/anaconda2\/bin\/python test2.py 2&gt;&amp;1', $output, $returnCode); \n\necho 'Output is: ' . PHP_EOL;\nvar_dump($output);\necho 'Return code is: ' . PHP_EOL;\nvar_dump($returnCode);\n\n?&gt;  \n<\/code><\/pre>\n"}]}
{"QuestionId":42104577,"AnswerCount":1,"Tags":"<python><tensorflow><convolution><softmax>","CreationDate":"2017-02-08T04:13:16.930","AcceptedAnswerId":null,"OwnerUserId":7532349.0,"Title":"Error with training Faster R-CNN","Body":"<p>I'm trying to run Faster R-CNN on Ubuntu 16.04 with my computer with some peers.\nBut I'm stuck.\nAnd I can't find same error like me on google.<\/p>\n\n<pre><\/pre>\n\n<p>I have no idea what this is error is about and how to solve this error.\nPlease help me<\/p>\n","answers":[{"AnswerId":"42107194","CreationDate":"2017-02-08T07:33:26.680","ParentId":null,"OwnerUserId":"1844390","Title":null,"Body":"<p>Instead of this:<\/p>\n\n<pre><code>rpn_cross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(rpn_cls_score, rpn_label))\n<\/code><\/pre>\n\n<p>try writing:<\/p>\n\n<pre><code>tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(rpn_label, rpn_cls_score))\n<\/code><\/pre>\n\n<p>if it doesn't work try:<\/p>\n\n<pre><code>tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=rpn_label, logits=rpn_cls_score))\n<\/code><\/pre>\n"}]}
{"QuestionId":42106032,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><word2vec>","CreationDate":"2017-02-08T06:20:47.203","AcceptedAnswerId":42108270.0,"OwnerUserId":6033925.0,"Title":"tensorflow Word2Vec error","Body":"<p>I downloaded source code of word2vec in github below.\n\nI am using tensorflow on pycharm.\nI'm using windows 10.\nI installed tensorflow, python, numpy which are needed to use tensorflow on windows. \nIn word2vec.py source code, I set the , , and .\nI downloaded the training text file from \nwhich the source code recommended.\nBut when I ran the code I get the error below:<\/p>\n\n<blockquote>\n  <p>C:\\Users\\Sungjin\\AppData\\Local\\Programs\\Python\\Python35\\python.exe\n  C:\/Users\/Sungjin\/PycharmProjects\/untitled\/ImpW2V.py Traceback (most\n  recent call last):   File\n  \"C:\/Users\/Sungjin\/PycharmProjects\/untitled\/ImpW2V.py\", line 43, in\n  \n      word2vec = tf.load_op_library(os.path.join(os.path.dirname(os.path.realpath(<strong>file<\/strong>)),\n  'word2vec_ops.so'))   File\n  \"C:\\Users\\Sungjin\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\framework\\load_library.py\",\n  line 64, in load_op_library\n      None, None, error_msg, error_code) tensorflow.python.framework.errors_impl.NotFoundError:\n  C:\\Users\\Sungjin\\PycharmProjects\\untitled\\word2vec_ops.so not found<\/p>\n  \n  <p>Process finished with exit code 1<\/p>\n<\/blockquote>\n\n<p>I checked Readme.md file on github but there is nothing mentioned about the file.\nIt seems I have to make .so file but I don't know how.<\/p>\n","answers":[{"AnswerId":"42108270","CreationDate":"2017-02-08T08:38:21.537","ParentId":null,"OwnerUserId":"1485311","Title":null,"Body":"<p>You're using windows and .so files are (almost) the equivalent of DLL files but under unix systems. So, even if you could find the .so file, that wouldn't work under windows.<\/p>\n\n<p>If you're willing to install linux and do everything from there, in the same place where you found word2vec.py you can also find how to build the .so file.\n( <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/embedding\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/embedding<\/a> , it says <code>You will need to compile the ops as follows<\/code> ). If you don't know how linux and g++ \/ gcc work, you should probably switch to another project that is windows specific.<\/p>\n"}]}
{"QuestionId":42106461,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-02-08T06:48:41.937","AcceptedAnswerId":null,"OwnerUserId":6038015.0,"Title":"How to get started with TensorFlow?","Body":"<p>How to get start TensorFlow, is there any book to read?\nI'm good at java programming, and have some background of machine learning. I'd like to find a book or a course about TensorFlow. Is there any recommendation?<\/p>\n","answers":[{"AnswerId":"42111918","CreationDate":"2017-02-08T11:30:16.160","ParentId":null,"OwnerUserId":"6248065","Title":null,"Body":"<p>Though this page its not the right place to ask such question , A programmer should get answer wherever possible<\/p>\n\n<p>You can get some of the tutorials that could be taken as initiation for your study at the very its own website <a href=\"https:\/\/www.tensorflow.org\/tutorials\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/<\/a> .Also you can refer the youtube <a href=\"https:\/\/www.youtube.com\/playlist?list=PLxePmhCQRxb9I6tlcB1yJpOq5_4ryCy7n\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/playlist?list=PLxePmhCQRxb9I6tlcB1yJpOq5_4ryCy7n<\/a> .<\/p>\n\n<p>provided below is the github repository link where you can find <\/p>\n\n<p><a href=\"https:\/\/github.com\/jtoy\/awesome-tensorflow#video\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jtoy\/awesome-tensorflow#video<\/a><\/p>\n\n<ul>\n<li>Books<\/li>\n<li>Models\/Projects<\/li>\n<li><p>Libraries<\/p><\/li>\n<li><p>Videos<\/p><\/li>\n<li><p>Papers <\/p><\/li>\n<li><p>Blog posts<\/p><\/li>\n<li><p>Tutorials <\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":42108006,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2017-02-08T08:24:07.000","AcceptedAnswerId":null,"OwnerUserId":245549.0,"Title":"How to put constraints on indexes involving different tensors in Theano?","Body":"<p>I want to multiply two big matrices and then to take only diagonal elements of the resulting matrix:<\/p>\n\n<pre><\/pre>\n\n<p>I can do it this way. However, doing it this way involves a lot of unnecessary operations. A better way to do it would be:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to \"force\" to do it the second way?<\/p>\n","answers":[{"AnswerId":"42110276","CreationDate":"2017-02-08T10:15:28.943","ParentId":null,"OwnerUserId":"245549","Title":null,"Body":"<p>The solution:<\/p>\n\n<pre><code>T.sum(a * b.dimshuffle(1,0), axis = 1)\n<\/code><\/pre>\n\n<p>The explanation:<\/p>\n\n<p>To \"pair\" indices of the first tensor with the indices of the second tensor we can use a pairwise multiplication. In this case all the dimensions of the first tensor will be paired with all the corresponding dimensions of the second tensor. However, to be able to do it we might need to reorder the dimensions since the i_th dimensions of the first tensor is always paired with i_th dimension of the second tensor. So, use transpose in the particular described cases.<\/p>\n\n<p>After an element wise multiplication you can <code>sum<\/code> over some dimensions.<\/p>\n"}]}
{"QuestionId":42108566,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2017-02-08T08:54:56.193","AcceptedAnswerId":42121982.0,"OwnerUserId":7533175.0,"Title":"Re-using Caffe files after building it","Body":"<p>I succesfully builded Caffe on  10 by following . <\/p>\n\n<p>Then, after copying  folder to my Python site_packages folder, I can use \"import caffe\" in Python.<\/p>\n\n<p>My question is that if I want to use Caffe on other computers without building Caffe from the beginning, do I just need to copy the files in  to the other computers' site_packages folder?<\/p>\n\n<p>Actually, I am not familiar with 'build' and 'compile' and I don't know what 'build' is doing during the Caffe installation.<\/p>\n","answers":[{"AnswerId":"42121982","CreationDate":"2017-02-08T19:28:32.940","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>You can do the copy provided that the destination computer uses the same support libraries (for software linkage to your Python and Caffe) and has bit-compatible hardware (to support the binary code).<\/p>\n\n<p>I do this regularly, within a controlled cluster.  All the nodes have the same OS configuration.  I build on one node, and then copy the entire directory to the same place on the other nodes.  Then I run multi-node applications on the result.<\/p>\n\n<p>If you switch to a different core processor, or a different OS implementation, your attempts to use the result will likely crash.  I've tried that one, too.  :-)<\/p>\n"}]}
{"QuestionId":42108829,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-08T09:08:07.767","AcceptedAnswerId":42116264.0,"OwnerUserId":2547570.0,"Title":"slice_input_producer(q.dequeue()) not working","Body":"<p>The argument to slice_input_producer() is Tensor, so I gave q.dequeue() to slice_input_producer(), but slice_input_producer immediately closes the queue (I don't know if it's an internal queue or my q).<\/p>\n\n<p>I expected it to hang until something is enqueued.<\/p>\n\n<pre><\/pre>\n\n<p>But this got an error:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>This code reproduces an undefined behavior of input_producer with dynamic tensor.<\/p>\n\n<pre><\/pre>\n\n<p>Result:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42116264","CreationDate":"2017-02-08T14:48:39.250","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><strong>TL;DR:<\/strong> The tensors in the <code>tensor_list<\/code> argument to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#slice_input_producer\" rel=\"nofollow noreferrer\"><code>tf.train.slice_input_producer(tensor_list)<\/code><\/a> must have a rank of 1 (i.e. a vector) or greater.<\/p>\n\n<p>The <code>tf.train.slice_input_producer()<\/code> function splits its input tensors into slices along dimension 0&amp;mdash, and produces one slice at a time. For example, if the input tensor is a matrix, it produces rows of the matrix.<\/p>\n\n<p>In your example, the input to <code>tf.train.slice_input_producer()<\/code> is a scalar tensor, which has no dimension 0 along which to slice it. The fact that this value comes from a queue is irrelevant (although since there are no shapes declared on the queue, you get a runtime error rather than a graph construction&ndash;time error). The following definition would be equivalent, and also produce an error:<\/p>\n\n<pre><code>producer = tf.train.slice_input_producer([3], shuffle=False)\n<\/code><\/pre>\n\n<p><strong>N.B.<\/strong> You might find that defining a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/input_pipeline#slice_input_producer\" rel=\"nofollow noreferrer\"><code>tf.train.slice_input_producer()<\/code><\/a> from a queue has unexpected results. In particular, the current implementation evaluates the tensors in <code>tensor_list<\/code> every time you consume a slice. This works fine with static tensors, but in your program it would dequeue a new element from the queue for each slice, which is unlikely to be what you intended.<\/p>\n"}]}
{"QuestionId":42109084,"AnswerCount":1,"Tags":"<lua><neural-network><torch>","CreationDate":"2017-02-08T09:20:09.337","AcceptedAnswerId":null,"OwnerUserId":3809516.0,"Title":"lua torch, how to get the print() output into a string","Body":"<pre><\/pre>\n\n<p>this outputs <\/p>\n\n<pre><\/pre>\n\n<p>I would like to get this print output for logging purposes. Does anyone know how to do that? <\/p>\n","answers":[{"AnswerId":"42131717","CreationDate":"2017-02-09T08:37:16.147","ParentId":null,"OwnerUserId":"6073048","Title":null,"Body":"<p>Hmm I'm not sure how you get this output. When I run this code I get <code>nn.ClassNLLCriterion<\/code>. Maybe different Lua\/Torch versions?<\/p>\n\n<p>Anyway, if you want to have this info in a string, you might have to extract it yourself. This can easily be done doing a simple loop:<\/p>\n\n<pre><code>for k,v in pairs(criterion) do\n    print(k,v)\nend\n<\/code><\/pre>\n\n<p>If you want the fancy print output then I suggest you look at TREPL's code (like @nobody suggested in the comments). It's all in Lua so it's very easy to replicate. More precisely, I recommend their <a href=\"https:\/\/github.com\/torch\/trepl\/blob\/master\/init.lua#L146\" rel=\"nofollow noreferrer\">sizestr()<\/a>, <a href=\"https:\/\/github.com\/torch\/trepl\/blob\/master\/init.lua#L260\" rel=\"nofollow noreferrer\">print_new()<\/a> and <a href=\"https:\/\/github.com\/torch\/trepl\/blob\/master\/init.lua#L167\" rel=\"nofollow noreferrer\">printvar()<\/a> functions. Simply change them so instead of printing, they construct a string.<\/p>\n\n<p>A quick example using their sizestr function:<\/p>\n\n<pre><code>-- Copy\/Paste from trepl\/init.lua\nlocal function sizestr(x)\n  local strt = {}\n  if _G.torch.typename(x):find('torch.*Storage') then\n    return _G.torch.typename(x):match('torch%.(.+)') .. ' - size: ' .. x:size()\n  end\n  if x:nDimension() == 0 then\n    table.insert(strt, _G.torch.typename(x):match('torch%.(.+)') .. ' - empty')\n  else\n    table.insert(strt, _G.torch.typename(x):match('torch%.(.+)') .. ' - size: ')\n    for i=1,x:nDimension() do\n      table.insert(strt, x:size(i))\n      if i ~= x:nDimension() then\n        table.insert(strt, 'x')\n      end\n    end\n  end\n  return table.concat(strt)\nend\n\n\nlocal function sutoringu(elem)\n  local str = ''\n  if torch.isTensor(elem) then\n    str = sizestr(elem)\n  else\n    str = tostring(elem)\n  end\n  return str\nend\n\nlocal str = '{\\n'\nlocal tab = '  '\nfor k,v in pairs(criterion) do\n  str = str .. tab .. k .. ' : ' .. sutoringu(v) .. '\\n'\nend\nstr = str .. '}'\n\nprint(str)\n<\/code><\/pre>\n\n<p>This outputs the same thing as what you wished for, constructing a string in the process. It's far from optimal but it's a start.<\/p>\n"}]}
{"QuestionId":42109590,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-08T09:44:28.037","AcceptedAnswerId":42125244.0,"OwnerUserId":1817595.0,"Title":"Mixture of experts - Train best model only at each iteration","Body":"<p>I am trying to implement a crude method based on the Mixture-of-Experts paper in <strong>tensorflow<\/strong> - <\/p>\n\n<p>There would be  models defined:<\/p>\n\n<pre><\/pre>\n\n<p>At every iteration, I want to train the model with the least loss only while keeping the other variables constant. Is it possible to place a switch to execute one of the optimizer only?<\/p>\n\n<p>P.S: This base of this problem is similar to one I had asked previously. <\/p>\n\n<p>Since the suggestion there did not work, I am trying to approach the problem differently.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42125244","CreationDate":"2017-02-08T22:58:02.417","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>This seems to be doable with <code>tf.cond<\/code>:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef make_conditional_train_op(\n    should_update, optimizers, variable_lists, losses):\n  \"\"\"Conditionally trains variables.\n\n  Each argument is a Python list of Tensors, and each list must have the same\n  length. Variables are updated based on their optimizer only if the\n  corresponding `should_update` boolean Tensor is True at a given step.\n\n  Returns a single train op which performs the conditional updates.\n  \"\"\"\n  assert len(optimizers) == len(variable_lists)\n  assert len(variable_lists) == len(losses)\n  assert len(should_update) == len(variable_lists)\n  conditional_updates = []\n  for model_number, (update_boolean, optimizer, variables, loss) in enumerate(\n      zip(should_update, optimizers, variable_lists, losses)):\n    conditional_updates.append(\n        tf.cond(update_boolean,\n                lambda: tf.group(\n                    optimizer.minimize(loss, var_list=variables),\n                    tf.Print(0, [\"Model {} updating\".format(model_number), loss])),\n                lambda: tf.no_op()))\n  return tf.group(*conditional_updates)\n<\/code><\/pre>\n\n<p>The basic strategy is to make sure the optimizer's variable updates are defined in the <code>lambda<\/code> of one of the <code>cond<\/code> branches, in which case there is true conditional op execution, meaning that the assignment to variables (and optimizer accumulators) only happens if that branch of the <code>cond<\/code> is triggered.<\/p>\n\n<p>As an example, we can construct some models:<\/p>\n\n<pre><code>def make_model_and_optimizer():\n  scalar_variable = tf.get_variable(\"scalar\", shape=[])\n  vector_variable = tf.get_variable(\"vector\", shape=[3])\n  loss = tf.reduce_sum(scalar_variable * vector_variable)\n  optimizer = tf.train.AdamOptimizer(0.1)\n  return optimizer, [scalar_variable, vector_variable], loss\n\n# Construct each model\noptimizers = []\nvariable_lists = []\nlosses = []\nfor i in range(10):\n  with tf.variable_scope(\"model_{}\".format(i)):\n    optimizer, variables, loss = make_model_and_optimizer()\n  optimizers.append(optimizer)\n  variable_lists.append(variables)\n  losses.append(loss)\n<\/code><\/pre>\n\n<p>Then determine a conditional update strategy, in this case only training the model with the maximum loss (just because that results in more switching; the output is rather boring if only one model ever updates):<\/p>\n\n<pre><code># Determine which model should be updated (in this case, the one with the\n# maximum loss)\ninteger_one_hot = tf.one_hot(\n    tf.argmax(tf.stack(losses),\n              axis=0),\n    depth=len(losses))\nis_max = tf.equal(\n    integer_one_hot,\n    tf.ones_like(integer_one_hot))\n<\/code><\/pre>\n\n<p>Finally, we can call the <code>make_conditional_train_op<\/code> function to create the train op, then do some training iterations:<\/p>\n\n<pre><code>train_op = make_conditional_train_op(\n    tf.unstack(is_max), optimizers, variable_lists, losses)\n\n# Repeatedly call the conditional train op\nwith tf.Session():\n  tf.global_variables_initializer().run()\n  for i in range(20):\n    print(\"Iteration {}\".format(i))\n    train_op.run()\n<\/code><\/pre>\n\n<p>This is printing the index which is updated and its loss at each iteration, confirming the conditional execution:<\/p>\n\n<pre><code>Iteration 0\nI tensorflow\/core\/kernels\/logging_ops.cc:79] [Model 6 updating][2.7271919]\nIteration 1\nI tensorflow\/core\/kernels\/logging_ops.cc:79] [Model 6 updating][2.1755948]\nIteration 2\nI tensorflow\/core\/kernels\/logging_ops.cc:79] [Model 2 updating][1.9858969]\nIteration 3\nI tensorflow\/core\/kernels\/logging_ops.cc:79] [Model 6 updating][1.6859927]\n<\/code><\/pre>\n"}]}
{"QuestionId":42109668,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><deep-learning><feature-extraction>","CreationDate":"2017-02-08T09:48:30.787","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Feature Representation using CNN (Image Aesthetics)","Body":"<p>I'm new to TensorFlow and i was able to use the below mentioned program for image classification. As Convolutional neural networks are arranged in layers, and each different layer captures information of a different kind. The first layers, which are close to the input, usually detect low-level information and deeper layers extract richer information. <\/p>\n\n<p>Now, my <strong>objective<\/strong> is to compare images and learn the commonalities and differences between three images, that is, i want to learn a feature representation (say D) such as for 3 images (say I1,I2 and I3 - where I1 &amp; I2 are aesthetically pleasing and I3 is not)<\/p>\n\n<pre><\/pre>\n\n<p>What i figured out from this is pixel correlation plays an important role (correct me if i'm wrong).<\/p>\n\n<p><strong>How can i achieve this? What are the steps involved in this process?<\/strong> <\/p>\n\n<p><em>Code for Image Classification:<\/em><\/p>\n\n<pre><\/pre>\n\n<p><em>More clarification:<\/em><\/p>\n\n<p>how to write the code? because the aim is to find a mathematical space in which the distance between the two images that are aesthetically pleasing is less than the distance between an aesthetically pleasing one and one which is not. I'm not able to compute that.<\/p>\n\n<p>Attributes to judge image aesthetics: interesting content, object emphasis, good lighting, color harmony, vivid color, shallow depth of f ield, motion blur, rule of thirds, balancing element, repetition, and symmetry.<\/p>\n","answers":[]}
{"QuestionId":42110091,"AnswerCount":1,"Tags":"<python><theano><elementwise-operations>","CreationDate":"2017-02-08T10:07:07.947","AcceptedAnswerId":null,"OwnerUserId":245549.0,"Title":"Are there something in between tensordot and element-wise multiplication in Theano?","Body":"<p>Let's assume that we have two tensors ( and ) with the same number of dimensions. We can multiply them with . For example:<\/p>\n\n<pre><\/pre>\n\n<p>In this case we <strong>\"pair\"<\/strong> axes of the first tensor with some axes of the second tensor and then we sum over these \"paired\" axes:<\/p>\n\n<pre><\/pre>\n\n<p>In the above example the first and the last axis of the first tensor is paired with the first and the last axis of the second tensor, respectively.<\/p>\n\n<p>Alternatively, we can multiply the two tensors element-wise:<\/p>\n\n<pre><\/pre>\n\n<p>In this case we \"pair\" all the axes of the first tensor with the all corresponding axes of the second tensor (first with first, second with second and so on).<\/p>\n\n<p>Now, I want to do something that is in between the two above described operations. In more details:<\/p>\n\n<ol>\n<li>I want to pair some axis of the first tensor with some axis of the second tensor (like w do it in ). So, I do not want to pair all the axes of A with all the axis of B.<\/li>\n<li>I do not want to sum up over all the paired axes (like we do in pairwise multiplication, there is no summation over the paired axes).<\/li>\n<\/ol>\n\n<p>Here is what I want written in a \"mathematical\" form:<\/p>\n\n<pre><\/pre>\n\n<p>What is the best way to do it in theano?<\/p>\n","answers":[{"AnswerId":"42163869","CreationDate":"2017-02-10T16:14:17.270","ParentId":null,"OwnerUserId":"245549","Title":null,"Body":"<p>The way to approach the described problem is to use the pairwise multiplication <code>*<\/code>. Pairwise multiplication \"pairs\" all the axes of the first tensor with the corresponding axes of the second tensor (first with first, second with second, ..., last with last). Therefore we need to solve two problems: (1) shuffle the axes of the two tensors such that the proper axes are paired with each other, (2) add \"dummy\" axes to prevent pairing where it is not needed. Finally, we do summation over the exes we want.<\/p>\n\n<p>The particular problem mentioned in the question<\/p>\n\n<pre><code>C[a, b, c, i] = sum_d A[a, b, c, d] * B[i, b, c, d]\n<\/code><\/pre>\n\n<p>is solved in the following way:<\/p>\n\n<pre><code>T.sum(A.dimshuffle(0, 1, 2, 3, 'x') * B.dimshuffle('x', 1, 2, 3, 0), axis=4)\n<\/code><\/pre>\n"}]}
{"QuestionId":42110293,"AnswerCount":3,"Tags":"<python><matlab><tensorflow>","CreationDate":"2017-02-08T10:16:26.130","AcceptedAnswerId":42116077.0,"OwnerUserId":5540592.0,"Title":"How will I integrate MATLAB to TensorFlow?","Body":"<p>I want to integrate MATLAB and TensorFlow, although I can run TensorFlow native in python but I am required to use MATLAB for image processing. Can someone please help me out with this one?<\/p>\n","answers":[{"AnswerId":"54422825","CreationDate":"2019-01-29T14:04:20.337","ParentId":null,"OwnerUserId":"8731154","Title":null,"Body":"<p>I used a mex function for inference via the C++ API of TensorFlow once. That's pretty straight forward. I had to link the required TensorFlow libs statically from source though.<\/p>\n"},{"AnswerId":"42116077","CreationDate":"2017-02-08T14:40:00.610","ParentId":null,"OwnerUserId":"6950549","Title":null,"Body":"<p>could this work?<\/p>\n\n<p>A MATLAB implementation of the TensorFlow Neural Networks Playground.<\/p>\n\n<p><a href=\"https:\/\/github.com\/StackOverflowMATLABchat\/NeuralNetPlayground\" rel=\"nofollow noreferrer\">https:\/\/github.com\/StackOverflowMATLABchat\/NeuralNetPlayground<\/a><\/p>\n"},{"AnswerId":"58890917","CreationDate":"2019-11-16T12:49:45.910","ParentId":null,"OwnerUserId":"12382943","Title":null,"Body":"<p>Similar to @HansQ's suggestion, I am currently writing a somewhat extensive wrapper called <em>tensorflow.m<\/em> for Matlab, see its <a href=\"https:\/\/github.com\/asteinh\/tensorflow.m\" rel=\"nofollow noreferrer\">repo on GitHub<\/a>. I cannot provide a stable release yet, but simple functionality like importing a frozen graph and running an inference is already possible (see the examples).<\/p>\n\n<p>I'd be glad if the package can be of use for someone looking for similar solutions; even more so, in case you extend\/implement something and open a PR.<\/p>\n"}]}
{"QuestionId":42111350,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-08T11:03:15.167","AcceptedAnswerId":42121323.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Why is there a need to reshape non-sparse elements once when parsing a TF-example from TFRecord files?","Body":"<p>In the TensorFlow documentation at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/data\/tfexample_decoder.py\" rel=\"nofollow noreferrer\">GitHub<\/a>, there is this following code:<\/p>\n\n<pre><\/pre>\n\n<p>I am wondering why there is a need to reshape a FixedLenFeature tensor after parsing it from a TFRecord file.<\/p>\n\n<p>In fact, what is the difference between a FixedLenFeature and VarLenFeature and what is their relevance to a Tensor? I am loading images in this case, so why would all of them be classified as a FixedLenFeature? What is an example of a VarLenFeature?<\/p>\n","answers":[{"AnswerId":"42121323","CreationDate":"2017-02-08T18:51:27.193","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>Tensors are stored on disk without shape information in an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto\" rel=\"nofollow noreferrer\">Example protocol buffer format<\/a> (TFRecord files are collections of Examples). The documentation in the <code>.proto<\/code> file describes things fairly well, but the basic point is that Tensor entries are stored in row-major order with no shape information, so that must be provided when the Tensors are read. Note that the situation is similar for storing Tensors in memory: the shape information is kept separately, and just reshaping a Tensor changes only metadata (things like transpose, on the other hand, can be expensive).<\/p>\n\n<p>VarLenFeatures are sequences such as sentences which would be difficult to batch together as regular Tensors, since the resulting shape would be ragged. The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/io_ops\/converting#parse_example\" rel=\"nofollow noreferrer\">parse_example documentation<\/a> has some good examples. Images are fixed length in that, if you load a batch of them, they'll all have the same shape (e.g. they're all 32x32 pixels, so a batch of 10 can have shape 10x32x32).<\/p>\n"}]}
{"QuestionId":42111458,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2017-02-08T11:08:06.553","AcceptedAnswerId":42113796.0,"OwnerUserId":2182857.0,"Title":"Using model.pop() changes the model's summary but does not effect the output","Body":"<p>I am trying to remove the top layers from a model I have previously trained.\nThis is the code I use:<\/p>\n\n<pre><\/pre>\n\n<p>Judging comparing the results of the two calls to model.summary(), I can see that indeed the 6 topmost layers were removed.<\/p>\n\n<p>However, the model's output (saved to A and B) does not change after discarding these layers.<\/p>\n\n<p>What is the source of that discrepancy? \nHow can I retrieve the output of the desired layer instead of that of the entire model?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42113796","CreationDate":"2017-02-08T12:57:16.430","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You can't drop layers like that, in order for it to have an effect, you need to recompile the model (AKA model.compile).<\/p>\n\n<p>But that's not the best way to obtain outputs from intermediate layers, you can just use K.function (where K is keras.backend) to build a function from the input to one of the layers and then call the function. More details are available in <a href=\"https:\/\/stackoverflow.com\/a\/41712013\/349130\">this<\/a> answer.<\/p>\n"}]}
{"QuestionId":42111925,"AnswerCount":0,"Tags":"<python><tensorflow><anaconda><conda><tflearn>","CreationDate":"2017-02-08T11:30:27.480","AcceptedAnswerId":null,"OwnerUserId":7022170.0,"Title":"Installation of TFLearn on windows system","Body":"<blockquote>\n  <blockquote>\n    <p>pip install tflearn, works, meaning tflearn gets installed, and I was able to import it, but a complex program doesn't run.<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>Error is related to _curses.<\/p>\n\n<p>I tried installing _curses using whl files.<\/p>\n\n<p>I get not a supported wheel file.<\/p>\n\n<p>I tried installing anaconda. conda command is not recognized.<\/p>\n\n<p>What should I do? Any full proof answers?<\/p>\n","answers":[]}
{"QuestionId":42112260,"AnswerCount":10,"Tags":"<keras><tensorboard>","CreationDate":"2017-02-08T11:47:12.067","AcceptedAnswerId":42112935.0,"OwnerUserId":7387749.0,"Title":"How do I use the Tensorboard callback of Keras?","Body":"<p>I have built a neural network with Keras. I would visualize its data by Tensorboard, therefore I have utilized:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>as explained in <a href=\"https:\/\/keras.io\/callbacks\/#tensorboard\" rel=\"noreferrer\">keras.io<\/a>. When I run the callback I get , but I don't get any file in my folder \"Graph\". Is there something wrong in how I have used this callback?<\/p>\n","answers":[{"AnswerId":"49450153","CreationDate":"2018-03-23T12:54:36.677","ParentId":null,"OwnerUserId":"9527366","Title":null,"Body":"<p>If you are working with Keras library and want to use tensorboard to print your graphs of accuracy and other variables, Then below are the steps to follow.<\/p>\n\n<p>step 1: Initialize the keras callback library to import tensorboard by using below command <\/p>\n\n<pre><code>from keras.callbacks import TensorBoard\n<\/code><\/pre>\n\n<p>step 2: Include the below command in your program just before \"model.fit()\" command. <\/p>\n\n<pre><code>tensor_board = TensorBoard(log_dir='.\/Graph', histogram_freq=0, write_graph=True, write_images=True)\n<\/code><\/pre>\n\n<p>Note: Use \".\/graph\". It will generate the graph folder in your current working directory, avoid using \"\/graph\".<\/p>\n\n<p>step 3: Include Tensorboard callback in \"model.fit()\".The sample is given below.<\/p>\n\n<pre><code>model.fit(X_train,y_train, batch_size=batch_size, epochs=nb_epoch, verbose=1, validation_split=0.2,callbacks=[tensor_board])\n<\/code><\/pre>\n\n<p>step 4 : Run your code and check whether your graph folder is there in your working directory. if the above codes work correctly you will have \"Graph\"\nfolder in your working directory.<\/p>\n\n<p>step 5 : Open Terminal in your working directory and type the command below.<\/p>\n\n<pre><code>tensorboard --logdir .\/Graph\n<\/code><\/pre>\n\n<p>step 6: Now open your web browser and enter the address below.<\/p>\n\n<pre><code>http:\/\/localhost:6006\n<\/code><\/pre>\n\n<p>After entering, the Tensorbaord page will open where you can see your graphs of different variables. <\/p>\n"},{"AnswerId":"42328391","CreationDate":"2017-02-19T14:26:05.600","ParentId":null,"OwnerUserId":"7588691","Title":null,"Body":"<p>Change <\/p>\n\n<pre><code>keras.callbacks.TensorBoard(log_dir='\/Graph', histogram_freq=0,  \n          write_graph=True, write_images=True)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>tbCallBack = keras.callbacks.TensorBoard(log_dir='Graph', histogram_freq=0,  \n          write_graph=True, write_images=True)\n<\/code><\/pre>\n\n<p>and set your model<\/p>\n\n<pre><code>tbCallback.set_model(model)\n<\/code><\/pre>\n\n<p>Run in your terminal<\/p>\n\n<pre><code>tensorboard  --logdir Graph\/\n<\/code><\/pre>\n"},{"AnswerId":"42343678","CreationDate":"2017-02-20T11:43:47.030","ParentId":null,"OwnerUserId":"1498447","Title":null,"Body":"<p>You wrote <code>log_dir='\/Graph'<\/code> did you mean <code>.\/Graph<\/code> instead? You sent it to <code>\/home\/user\/Graph<\/code> at the moment.<\/p>\n"},{"AnswerId":"45384794","CreationDate":"2017-07-29T03:00:31.113","ParentId":null,"OwnerUserId":"2476920","Title":null,"Body":"<p>Here is some code: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>K.set_learning_phase(1)\nK.set_image_data_format('channels_last')\n\ntb_callback = keras.callbacks.TensorBoard(\n    log_dir=log_path,\n    histogram_freq=2,\n    write_graph=True\n)\ntb_callback.set_model(model)\ncallbacks = []\ncallbacks.append(tb_callback)\n\n# Train net:\nhistory = model.fit(\n    [x_train],\n    [y_train, y_train_c],\n    batch_size=int(hype_space['batch_size']),\n    epochs=EPOCHS,\n    shuffle=True,\n    verbose=1,\n    callbacks=callbacks,\n    validation_data=([x_test], [y_test, y_test_coarse])\n).history\n\n# Test net:\nK.set_learning_phase(0)\nscore = model.evaluate([x_test], [y_test, y_test_coarse], verbose=0)\n<\/code><\/pre>\n\n<p>Basically, <code>histogram_freq=2<\/code> is the most important parameter to tune when calling this callback: it sets an interval of epochs to call the callback, with the goal of generating fewer files on disks.<\/p>\n\n<p>So here is an example visualization of the evolution of values for the last convolution throughout training once seen in TensorBoard, under the \"histograms\" tab (and I found the \"distributions\" tab to contain very similar charts, but flipped on the side):<\/p>\n\n<p><img src=\"https:\/\/raw.githubusercontent.com\/Vooban\/Hyperopt-Keras-CNN-CIFAR-100\/master\/tensorboard_histogram_example.png\" alt=\"tensorboard weights monitoring\"><\/p>\n\n<p>In case you would like to see a full example in context, you can refer to this open-source project: <a href=\"https:\/\/github.com\/Vooban\/Hyperopt-Keras-CNN-CIFAR-100\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Vooban\/Hyperopt-Keras-CNN-CIFAR-100<\/a><\/p>\n"},{"AnswerId":"47039589","CreationDate":"2017-10-31T15:55:47.583","ParentId":null,"OwnerUserId":"277345","Title":null,"Body":"<p>There are few things.<\/p>\n\n<p>First, not <code>\/Graph<\/code> but <code>.\/Graph<\/code><\/p>\n\n<p>Second, when you use the TensorBoard callback, always pass validation data, because without it, it wouldn't start.<\/p>\n\n<p>Third, if you want to use anything except scalar summaries, then you should only use the <code>fit<\/code> method because <code>fit_generator<\/code> will not work. Or you can rewrite the callback to work with <code>fit_generator<\/code>.<\/p>\n\n<p>To add callbacks, just add it to <code>model.fit(..., callbacks=your_list_of_callbacks)<\/code><\/p>\n"},{"AnswerId":"42112935","CreationDate":"2017-02-08T12:18:22.737","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<pre class=\"lang-py prettyprint-override\"><code>keras.callbacks.TensorBoard(log_dir='.\/Graph', histogram_freq=0,  \n          write_graph=True, write_images=True)\n<\/code><\/pre>\n\n<p>This line creates a Callback Tensorboard object, you should capture that object and give it to the <code>fit<\/code> function of your model.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>tbCallBack = keras.callbacks.TensorBoard(log_dir='.\/Graph', histogram_freq=0, write_graph=True, write_images=True)\n...\nmodel.fit(...inputs and parameters..., callbacks=[tbCallBack])\n<\/code><\/pre>\n\n<p>This way you gave your callback object to the function. It will be run during the training and will output files that can be used with tensorboard.<\/p>\n\n<p>If you want to visualize the files created during training, run in your terminal<\/p>\n\n<pre><code>tensorboard --logdir path_to_current_dir\/Graph \n<\/code><\/pre>\n\n<p>Hope this helps !<\/p>\n"},{"AnswerId":"55996756","CreationDate":"2019-05-05T21:55:08.650","ParentId":null,"OwnerUserId":"2285255","Title":null,"Body":"<p>Create the Tensorboard callback:<\/p>\n\n<pre><code>from keras.callbacks import TensorBoard\nfrom datetime import datetime\nlogDir = \".\/Graph\/\" + datetime.now().strftime(\"%Y%m%d-%H%M%S\") + \"\/\"\ntb = TensorBoard(log_dir=logDir, histogram_freq=2, write_graph=True, write_images=True, write_grads=True)\n<\/code><\/pre>\n\n<p>Pass the Tensorboard callback to the fit call:<\/p>\n\n<pre><code>history = model.fit(X_train, y_train, epochs=200, callbacks=[tb])\n<\/code><\/pre>\n\n<p>When running the model, if you get a Keras error of <\/p>\n\n<blockquote>\n  <p>\"You must feed a value for placeholder tensor\"<\/p>\n<\/blockquote>\n\n<p>try reseting the Keras session before the model creation by doing:<\/p>\n\n<pre><code>import keras.backend as K\nK.clear_session()\n<\/code><\/pre>\n"},{"AnswerId":"56302587","CreationDate":"2019-05-25T07:08:43.307","ParentId":null,"OwnerUserId":"9039646","Title":null,"Body":"<p>If you are using <strong>google-colab<\/strong> simple visualization of the graph would be :<\/p>\n\n<pre><code>import tensorboardcolab as tb\n\ntbc = tb.TensorBoardColab()\ntensorboard = tb.TensorBoardColabCallback(tbc)\n\n\nhistory = model.fit(x_train,# Features\n                    y_train, # Target vector\n                    batch_size=batch_size, # Number of observations per batch\n                    epochs=epochs, # Number of epochs\n                    callbacks=[early_stopping, tensorboard], # Early stopping\n                    verbose=1, # Print description after each epoch\n                    validation_split=0.2, #used for validation set every each epoch\n                    validation_data=(x_test, y_test)) # Test data-set to evaluate the model in the end of training\n<\/code><\/pre>\n"},{"AnswerId":"43457523","CreationDate":"2017-04-17T18:45:09.190","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>This is how you use the <a href=\"https:\/\/keras.io\/callbacks\/#tensorboard\" rel=\"noreferrer\">TensorBoard callback<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.callbacks import TensorBoard\n\ntensorboard = TensorBoard(log_dir='.\/logs', histogram_freq=0,\n                          write_graph=True, write_images=False)\n# define model\nmodel.fit(X_train, Y_train,\n          batch_size=batch_size,\n          epochs=nb_epoch,\n          validation_data=(X_test, Y_test),\n          shuffle=True,\n          callbacks=[tensorboard])\n<\/code><\/pre>\n"},{"AnswerId":"46996697","CreationDate":"2017-10-29T03:51:25.590","ParentId":null,"OwnerUserId":"2559219","Title":null,"Body":"<p>You should check out Losswise (<a href=\"https:\/\/losswise.com\" rel=\"nofollow noreferrer\">https:\/\/losswise.com<\/a>), it has a plugin for Keras that's easier to use than Tensorboard and has some nice extra features.  With Losswise you'd just use <code>from losswise.libs import LosswiseKerasCallback<\/code> and then <code>callback = LosswiseKerasCallback(tag='my fancy convnet 1')<\/code> and you're good to go (see <a href=\"https:\/\/docs.losswise.com\/#keras-plugin\" rel=\"nofollow noreferrer\">https:\/\/docs.losswise.com\/#keras-plugin<\/a>).<\/p>\n"}]}
{"QuestionId":42112407,"AnswerCount":2,"Tags":"<python><tensorflow><pyspark>","CreationDate":"2017-02-08T11:54:14.603","AcceptedAnswerId":42120646.0,"OwnerUserId":2563522.0,"Title":"Transform map to mapPartition using pyspark","Body":"<p>I am trying to load a tensorflow model from disk and predicting the values.<\/p>\n\n<p>Code<\/p>\n\n<pre><\/pre>\n\n<p>I have a RDD which consists of a tuple (rowkey, input_vector). I want to use the loaded model to predict the score\/class of the input. <\/p>\n\n<p>Code to call get_value()<\/p>\n\n<pre><\/pre>\n\n<p>The problem is every time I call the map, the model is loaded everytime for each tuple and it takes a lot of time.<\/p>\n\n<p>I am thinking of loading the model using <strong>mapPartitions<\/strong> and then use map to call <strong>get_value<\/strong> function.\nI have no clue as how to convert the code to a mapPartition where I load the tensorflow model only once per parition and reduce the running time.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42120646","CreationDate":"2017-02-08T18:14:30.840","ParentId":null,"OwnerUserId":"2563522","Title":null,"Body":"<p>I guess that the below code is a huge improvement as it uses mapPartitions.<\/p>\n\n<p>Code<\/p>\n\n<pre><code>def predict(rows):\n    graph = tf.Graph()\n    checkpoint_file = \"\/home\/sahil\/Desktop\/Relation_Extraction\/data\/1485336002\/checkpoints\/model-300\"\n    print(\"Loading model................................\")\n    with graph.as_default():\n        session_conf = tf.ConfigProto(\n            allow_soft_placement=allow_soft_placement,\n            log_device_placement=log_device_placement)\n        sess = tf.Session(config=session_conf)\n        with sess.as_default():\n            # Load the saved meta graph and restore variables\n            saver = tf.train.import_meta_graph(\"{}.meta\".format(checkpoint_file))\n            saver.restore(sess, checkpoint_file)\n        print(\"**********************************************\")\n        # Get the placeholders from the graph by name\n        input_x = graph.get_operation_by_name(\"X_train\").outputs[0]\n        dropout_keep_prob = graph.get_operation_by_name(\"dropout_keep_prob\").outputs[0]\n        # Tensors we want to evaluate\n        predictions = graph.get_operation_by_name(\"output\/predictions\").outputs[0]\n\n        # Generate batches for one epoch\n        for row in rows:\n            X_test = [row[1]]\n            batch_predictions = sess.run(predictions, {input_x: X_test, dropout_keep_prob: \n            yield (row[0], batch_predictions)\n\n\nresult = data_rdd.mapPartitions(lambda iter: predict(iter))\nresult.foreach(print)\n<\/code><\/pre>\n"},{"AnswerId":"42117987","CreationDate":"2017-02-08T16:04:10.390","ParentId":null,"OwnerUserId":"1928294","Title":null,"Body":"<p>I am not sure if I get your question correctly, but we can optimise your code a bit here. <\/p>\n\n<pre><code>graph = tf.Graph()\n\ncheckpoint_file = \"\/home\/sahil\/Desktop\/Relation_Extraction\/data\/1485336002\/checkpoints\/model-300\"\n\nwith graph.as_default():\n        session_conf = tf.ConfigProto(\n            allow_soft_placement=allow_soft_placement,\n            log_device_placement=log_device_placement)\n        sess = tf.Session(config=session_conf)\n\ns = sess.as_default()\nsaver = tf.train.import_meta_graph(\"{}.meta\".format(checkpoint_file))\nsaver.restore(sess, checkpoint_file)\n\n\ninput_x = graph.get_operation_by_name(\"X_train\").outputs[0]\ndropout_keep_prob = graph.get_operation_by_name(\"dropout_keep_prob\").outputs[0]\npredictions = graph.get_operation_by_name(\"output\/predictions\").outputs[0]\n\nsession_pickle = cPickle.dumps(sess)\n\ndef get_value(key, vector, session_pickle):\n    sess = cPickle.loads(session_pickle)\n    rowkey = key\n    batch_predictions = sess.run(predictions, {input_x: [vector], dropout_keep_prob: 1.0})\n    print(batch_predictions)\n    return (rowkey, batch_predictions\n\n\n\nresult = data_rdd.map(lambda (key, row): get_value(key=key, vector = row ,  session_pickle = session_pickle))\nresult.foreach(print)\n<\/code><\/pre>\n\n<p>So you can serialize your tensorflow session. Though I haven't tested your code here. Run this and leave a comment.<\/p>\n"}]}
{"QuestionId":42113176,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-08T12:29:25.867","AcceptedAnswerId":42118826.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Is dequeueing single examples a bottleneck?","Body":"<p>In the TF-slim <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/data\/parallel_reader.py#L105\" rel=\"nofollow noreferrer\">documentation<\/a> (with reference to the  function), I realized there are multiple readers that feed (enqueue) the parsed TF-examples into a queue, but there is a function  that only dequeues one element at a time. Would this cause a bottleneck in my training when I create my batches? Is it better to use  instead?<\/p>\n\n<p>While I was training my model previously, I noticed on TensorBoard that the parallelreader queue is always full - is this a cause of concern or is the enqueue op supposed to be faster than the dequeue op? Generally speaking, should the dequeue operation dequeue as many examples as the number of readers that enqueue the operations? <\/p>\n\n<p>Here are my visualizations for the queues:\n<a href=\"https:\/\/i.stack.imgur.com\/FtIDn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FtIDn.png\" alt=\"Queue Visualization\"><\/a><\/p>\n\n<p>My guess is that it is probably good to have more examples dequeued at any one time, so long as there is a min_after_dequeue argument that ensures enough examples in the queue to be shuffled at any one point (in fact, how often does the shuffling take place?). But what is the trade off of dequeuing many examples at one go?<\/p>\n","answers":[{"AnswerId":"42118826","CreationDate":"2017-02-08T16:41:26.780","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>This <a href=\"https:\/\/stackoverflow.com\/questions\/41647784\/tfrecordreader-seems-extremely-slow-and-multi-threads-reading-not-working\/41688401#41688401\">answer<\/a> shows how to do profiling, and for that example, each dequeue op took 60 microseconds on one thread. If you use <code>tf.batch<\/code> on top of dequeue, it'll run multiple dequeue ops in parallel, so it may go down to 12 microseconds per deqeueue on average. This mean that it will only be a bottleneck if your computation takes less than 12 microseconds. A while back when I checked, it took 5 microseconds to schedule a single GPU kernel call, so any network with more than 2 GPU ops will take longer to evaluate, and <code>dequeue<\/code> will not be a bottleneck.<\/p>\n"}]}
{"QuestionId":42113592,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-08T12:49:04.297","AcceptedAnswerId":null,"OwnerUserId":5107084.0,"Title":"TensorFlow: Queue visualization mismatch in number in TensorBoard?","Body":"<p>I am wondering why the names of the scalar summaries for the queue as stated in the left graph of the photo has a number 32.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/O7cyC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/O7cyC.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Firstly, the reader (a TFRecord reader in this case) has no attribute or whatsoever that causes it to have a number 32. The graph on the right, however, is correct as the RandomShuffleQueue that is created does have a capacity of 256. In my code, I only have 1 reader created. So why is there a  in my code and what does it mean?<\/p>\n\n<p>I have always give a different number from what I expected. For instance, when my batch is 8, the FIFO queue for the  scalar summary always give me 16. Is that because the queue capacity is always capped at 16?<\/p>\n","answers":[]}
{"QuestionId":42114187,"AnswerCount":0,"Tags":"<numpy><tensorflow>","CreationDate":"2017-02-08T13:14:19.107","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"Tensorflow load numpy array - shape error","Body":"<p>So when I save a int16 numpy array and reload it I get the same shape as expected. In this particular case the shape is (335,306,306) yielding a size of 31368060. <\/p>\n\n<p>Now I am attempting to load such arrays into a tensorflow pipeline as follows;<\/p>\n\n<pre><\/pre>\n\n<p>Evaluating the same example in a session I get;<\/p>\n\n<pre><\/pre>\n\n<p>Yielding (31368100,) which is not the same as (31368060,). What is going wrong?<\/p>\n","answers":[]}
{"QuestionId":42115480,"AnswerCount":1,"Tags":"<python><tensorflow><converter><deep-learning><caffe>","CreationDate":"2017-02-08T14:14:27.383","AcceptedAnswerId":42179514.0,"OwnerUserId":235472.0,"Title":"Error converting a model from Caffe to Tensorflow","Body":"<p>ImportError: No module named kaffe.tensorflow<\/p>\n\n<p>Trying to convert a model from Caffe to Tensorflow with the <a href=\"https:\/\/github.com\/ethereon\/caffe-tensorflow\" rel=\"nofollow noreferrer\">caffe-tensorflow<\/a> converter.<\/p>\n\n<p>With this command:<\/p>\n\n<blockquote>\n  <p>convert.py --caffemodel .\/VGG_ILSVRC_19_layers.caffemodel\n  --data-output-path .\/TF_conv\/dataOutput.npz --code-output-path .\/TF_conv\/codeOutput.py .\/VGG_ILSVRC_19_layers_deploy.prototxt<\/p>\n<\/blockquote>\n\n<p>I get this code output file:<\/p>\n\n<pre><\/pre>\n\n<p>With: \nI get this error:<\/p>\n\n<pre><\/pre>\n\n<p>Do I have to put the codeOutput.py file in the directory where the kaffe.tensorflow module is?<br>\nDo I have to copy the kaffe.tensorflow module where my file is?<br>\nCan I connect the two in some way?<\/p>\n","answers":[{"AnswerId":"42179514","CreationDate":"2017-02-11T18:19:31.760","ParentId":null,"OwnerUserId":"7550917","Title":null,"Body":"<p>caffe-tensorflow has to be inside your PYTHONPATH so python can find the modules.\nYou can either edit the PYTHONPATH enviroment variable inside your .bashrc. For example:<\/p>\n\n<pre><code>export PYTHONPATH='\/path\/to\/caffe-tensorflow'\n<\/code><\/pre>\n\n<p>or you add the path inside your script:<\/p>\n\n<pre><code>import sys\nsys.path.insert(0, '\/path\/to\/caffe-tensorflow')\n<\/code><\/pre>\n"}]}
{"QuestionId":42115585,"AnswerCount":1,"Tags":"<python><tensorflow><keras><stateful>","CreationDate":"2017-02-08T14:18:56.093","AcceptedAnswerId":42116152.0,"OwnerUserId":6950549.0,"Title":"Input to reshape is a tensor with 2 * \"batch_size\" values, but the requested shape has \"batch_size\"","Body":"<p>I want to make a RNN using a Keras sequential model with a tensorflow backend. When I implement the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I get the follow error message:<\/p>\n\n<p><em>Input to reshape is a tensor with 16 values, but the requested shape has 8<\/em><\/p>\n\n<p>But whatever I change the batch_size to the error will just follow the following formula:<\/p>\n\n<p><em>Input to reshape is a tensor with    values, but the requested shape has <\/em><\/p>\n\n<p>I have looked at other <a href=\"https:\/\/stackoverflow.com\/a\/37435874\/6950549\">Q&amp;A<\/a>, but I do not think they help me much. Or I dont understand the answers well enough.<\/p>\n\n<p>Any help would be much appreciated!<\/p>\n\n<p>EDIT:\nas requested the shape of input and target:<\/p>\n\n<pre><\/pre>\n\n<p>EDIT 2: <\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT 3: UPDATE<\/strong><\/p>\n\n<p>With help of @Nassim Ben, he figured out that the problem is in the loss function. He posted code with a regular loss function and then it works just fine. However with the custom loss function that code does not work. As any readers of this question can see I posted my costum loss function above and there is the problem. Currently I do not yet know why this error exist but this is the current status.<\/p>\n","answers":[{"AnswerId":"42116152","CreationDate":"2017-02-08T14:43:34.520","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>EDIT :\nThis code works for me, I have only changed the loss for simplicity.<\/p>\n\n<pre><code>import keras\nfrom keras.layers import *\nfrom keras.models import Sequential\nfrom keras.objectives import *\nimport numpy as np\n\nx_train = np.random.random((512,600, 103))\ny_train = np.random.random((512,600,2))\nbatch_size = 8\nbatch_inputshape = (batch_size,x_train.shape[1],x_train.shape[2]) \nprint(batch_inputshape) #(8, 600, 103)\n\nmodel = Sequential()\nmodel.add(LSTM(103,\n           batch_input_shape = batch_inputshape,\n           return_sequences = True,\n          stateful = True))\nmodel.add(Dropout(0.2))\nmodel.add(LSTM(50,\n           return_sequences = True,\n          stateful = True))\nmodel.add(Dropout(0.2))\n\n\nmodel.add(TimeDistributed(Dense(10)))\nmodel.add(TimeDistributed(Dense(2)))\nmodel.add(Activation('softmax'))\nmodel.compile(loss= \"mse\", optimizer='adam')\n\nprint (model.output_shape) #(8, 600, 2)\n\nmodel.fit(x_train,y_train, batch_size = batch_size,\n                       nb_epoch = 1, validation_split=0.25)\n<\/code><\/pre>\n\n<p>EDIT 2:<\/p>\n\n<p>So the error was coming from the loss function. In the code you copied from github for ncce loss, they had outputs of shape (batch,10). You have outputs of shape (batch, 600, 2). So here is my edit of the function :<\/p>\n\n<pre><code>def w_categorical_crossentropy(y_true, y_pred, weights):\n# https:\/\/github.com\/fchollet\/keras\/issues\/2115#issuecomment-274101310 #\n    nb_cl = len(weights)\n    # Create a mask with zeroes\n    final_mask = K.zeros_like(y_pred[:,:,0])\n    # get the maximum probability value for every output (shape = (batch,600,1))\n    y_pred_max = K.max(y_pred, axis=2, keepdims=True)\n    # Get the actual predictions for every output (shape = (batch,600,2))\n    # This K.equal uses broadcasting, we compare two tensors of different sizes but it works (magic)\n    y_pred_max_mat = K.equal(y_pred, y_pred_max)\n    for c_p, c_t in product(range(nb_cl), range(nb_cl)):\n        # Create the mask of weights to apply to the result of the cat_crossentropy\n        final_mask += (weights[c_t, c_p] * K.cast(y_pred_max_mat[:,:, c_p], K.floatx()) * y_true[:,:, c_t])\n    return K.categorical_crossentropy(y_pred, y_true) * final_mask\n\nw_array = np.ones((2,2))\nw_array[1, 0] = 100\n<\/code><\/pre>\n\n<p>As you can see, I just modified the index play because of your particular shape. \nThe mask has to be of shape (batch, 600). \nThe max has to be done on the 3rd dimension because there lie the probabilities that you want to output. \nThe matrix multiplication to build the max needed to be updated too because of the shape of your tensors again.<\/p>\n\n<p>This should work.<\/p>\n\n<p>If you need more detailed explaination feel free to ask :-)<\/p>\n"}]}
{"QuestionId":42115952,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-08T14:34:26.593","AcceptedAnswerId":42116756.0,"OwnerUserId":2191652.0,"Title":"Tensorflow: What is the equivalent of weight_filler \"gaussian\" from Caffe in Tensorflow?","Body":"<p>I'm trying to convert a Caffemodel to tensorflow. In caffe my convolutions are initialized like this:<\/p>\n\n<pre><\/pre>\n\n<p>How can I get the same initialization in Tensorflow?<\/p>\n","answers":[{"AnswerId":"42116756","CreationDate":"2017-02-08T15:08:46.030","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>With <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/api_docs\/python\/constant_op\/random_tensors#truncated_normal\" rel=\"nofollow noreferrer\"><code>tf.truncated_normal<\/code><\/a> you can get random values from a truncated random distribution, thus:<\/p>\n\n<pre><code>shape = [10,10] # variable shape\nweight = tf.Variable(tf.truncated_normal(shape, mean=0.0, stddev=0.01))\nbias = tf.Variable(0.)\n<\/code><\/pre>\n"}]}
{"QuestionId":42116597,"AnswerCount":3,"Tags":"<python><macos><docker><machine-learning><tensorflow>","CreationDate":"2017-02-08T15:01:48.563","AcceptedAnswerId":42118421.0,"OwnerUserId":5588107.0,"Title":"How to run Python Scripts on Mac Terminal using Docker with Tensorflow?","Body":"<p>I just got a Mac and shifted from Windows and installed Tensorflow using Docker and everything is working fine, but I want to run a python script that I have from before. Is there any way to run a python script in docker on a Mac, using the terminal?<\/p>\n","answers":[{"AnswerId":"42118421","CreationDate":"2017-02-08T16:23:44.200","ParentId":null,"OwnerUserId":"3087106","Title":null,"Body":"<p>Something like the below should do...<\/p>\n\n<p>Create a Dockerfile with the below content:<\/p>\n\n<pre><code>FROM gcr.io\/tensorflow\/tensorflow:latest\nCOPY script.py \/usr\/bin\/\nCMD [\"python\", \"\/usr\/bin\/script.py\"]\n<\/code><\/pre>\n\n<p>build the image:<\/p>\n\n<pre><code>$ docker build -t mytensorflow .\n<\/code><\/pre>\n\n<p>run it:<\/p>\n\n<pre><code>$ docker run -it --rm mytensorflow\n<\/code><\/pre>\n\n<p>if you want to keep the container going after the script is run don't --rm it...<\/p>\n"},{"AnswerId":"42117407","CreationDate":"2017-02-08T15:39:12.653","ParentId":null,"OwnerUserId":"7001235","Title":null,"Body":"<p>More information would be very helpful, but perhaps this is useful for you.<\/p>\n\n<p>It depends on a lot of different factors, but assuming a couple of things: <\/p>\n\n<p><strong>The docker containing Tensorflow has a name like 'tensorflow' already contains the python script:<\/strong> \nyou can use: \ndocker run tensorflow 'python '<\/p>\n\n<p><strong>If the script is not yet present, you can either use build a docker based on that image using a docker file, something along the lines of:<\/strong> <\/p>\n\n<pre><code>FROM tensorflow:latest\nADD \/some\/local\/path.py \/the\/path\/on\/my\/docker\n<\/code><\/pre>\n\n<p>Or you can create\/run the docker interactively by doing something along the lines of: <\/p>\n\n<p>Docker run -ti tensorflow \/bin\/bash\nWhich will start an instance of the tensorflow container and start \/bin\/bash on it. \nYou now have an interactive shell so can place the python script any way you want and start it.<\/p>\n"},{"AnswerId":"44015407","CreationDate":"2017-05-17T04:12:12.337","ParentId":null,"OwnerUserId":"8435658","Title":null,"Body":"<p>Lets assume you have a script <code>my_script.py<\/code> located at <code>\/Users\/awesome_user\/python_scripts\/<\/code> on your mac<\/p>\n\n<p>By default the tensorFlow image bash will locate you at \/notebooks.<\/p>\n\n<p>Run this command in your terminal: <pre>docker run --rm -it -v \/Users\/awesome_user\/python_scripts\/:\/notebooks gcr.io\/tensorflow\/tensorflow bash<\/pre><\/p>\n\n<p>This will map your local mac folder <code>\/Users\/awesome_user\/python_scripts\/<\/code> to the docker's local folder <code>\/notebooks<\/code><\/p>\n\n<p>then just run from the bash <code>python my_script.py<\/code>. Also running <code>ls<\/code> should reveal your folder content<\/p>\n"}]}
{"QuestionId":42117254,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2017-02-08T15:33:02.300","AcceptedAnswerId":null,"OwnerUserId":848691.0,"Title":"Tensorflow: Single GPU vs. Multi GPU","Body":"<p>Will it be better to use one good GPU like the <a href=\"https:\/\/www.nvidia.com\/en-us\/geforce\/products\/10series\/titan-x-pascal\/\" rel=\"nofollow noreferrer\">Titan X<\/a> with 12 GB Memory or two <a href=\"https:\/\/www.nvidia.com\/en-us\/geforce\/products\/10series\/geforce-gtx-1080\" rel=\"nofollow noreferrer\">GTX 1080<\/a> with 8 GB Memory each to compute large CNNs.\nBoth will cost about the same.<\/p>\n\n<p>Nvidia suggests the Titan X <a href=\"http:\/\/www.nvidia.com\/object\/gpu-accelerated-applications-tensorflow-configurations.html\" rel=\"nofollow noreferrer\">here<\/a>, but why? What are the advantages having one GPU with much Memory than having two smaller GPUs which have more Memory when combined?<\/p>\n","answers":[{"AnswerId":"43114809","CreationDate":"2017-03-30T10:11:44.213","ParentId":null,"OwnerUserId":"3571644","Title":null,"Body":"<p>So from my personal experience is tensorflow tends to only use one GPU. That might not always the case, but for the algorithms I'm running it seems to use only one.<\/p>\n\n<p>Personally I would go for two gtx 1080 ti's or one gtx 1080ti. The cost for the titan for one more GB or gddr and a slightly faster bus doesn't make since.<\/p>\n\n<p>I would go for two just so you can run multiple experiments at once.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ITbLA.png\" rel=\"nofollow noreferrer\">8 Nvidia tesala's running tensorflow.<\/a><\/p>\n"}]}
{"QuestionId":42117366,"AnswerCount":1,"Tags":"<python><tensorflow><python-imaging-library><deep-learning>","CreationDate":"2017-02-08T15:37:22.067","AcceptedAnswerId":42243466.0,"OwnerUserId":7535253.0,"Title":"(Modify)Tensorflow tflearn Binary image learning questions","Body":"<p>I would like to learn fingerprint images that have been binarized using PIL for the tensor flow.\nI'm trying to learn a binarized image, so the shape is not right.<\/p>\n\n<pre><\/pre>\n\n<p>I have tried using numpy reshape the dimensions change.\nHowever, the following error is repeated.<\/p>\n\n<p>The error is as follows.\nValueError: Can not feed value of shape (64,224,224) for Tensor u'input \/ X: 0 ', which has shape (?, 224, 224, 3)<\/p>\n\n<p>What is the problem?<\/p>\n","answers":[{"AnswerId":"42243466","CreationDate":"2017-02-15T07:51:24.147","ParentId":null,"OwnerUserId":"1338797","Title":null,"Body":"<p>The problem is with your input shape - it doesn't match the input layer.<\/p>\n\n<p>The input layer is defined in <code>create_vggnet()<\/code>:<\/p>\n\n<pre><code>def create_vggnet(num_classes):\n    # Building 'VGGNet'\n    network = input_data(shape=[None, 224, 224, 3], name='input')\n<\/code><\/pre>\n\n<p>So you expect <code>None<\/code> (== any) times (224, 224, 3), that is 224x224 x RGB (3 channels). And you pass 64 (your batch size) times 224x224.<\/p>\n\n<p>There are two fixes:<\/p>\n\n<p>1) (probably more wasteful) - extend the images to RGB.<\/p>\n\n<p>So, after you convert the image to <code>'L'<\/code> (lightness, that is gray levels) and then binarize, convert it to RGB first. Then you can convert it to <code>'F'<\/code><\/p>\n\n<p>(See: <a href=\"http:\/\/effbot.org\/imagingbook\/image.htm\" rel=\"nofollow noreferrer\">http:\/\/effbot.org\/imagingbook\/image.htm<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/5414586\/how-do-i-save-a-mode-f-image-python-pil\">How do I save a mode &#39;F&#39; image? (Python\/PIL)<\/a>)<\/p>\n\n<pre><code>def binarization(in_img, threshold):\n    im = in_img.convert('L')\n    for i in range(im.size[0]):\n        for j in range(im.size[1]):\n            if im.getpixel((i, j)) &gt; threshold:\n                im.putpixel((i, j), 255)\n            else:\n                im.putpixel((i, j), 0)\n    return im.convert('RGB').convert('F')\n<\/code><\/pre>\n\n<p>2) (less wasteful, but you're changing your network a bit (just the input layer) - so it can be argued, that this \"isn't VGG 16 anymore\") You can change the input layer to 1-channel.<\/p>\n\n<pre><code>def create_vggnet(num_classes):\n    # Building 'VGGNet'\n    network = input_data(shape=[None, 224, 224, 1], name='input')\n<\/code><\/pre>\n\n<p>Unfortunately, <code>shape=[None, 224, 224]<\/code> doesn't work (the error is something about \"The Tensor needs to be 4D\"). So we have a shape of (224, 224, 1) for a single input value.<\/p>\n\n<p>So you need to make the images have an extra dimension:<\/p>\n\n<pre><code>def pil_to_nparray(pil_image):\n    pil_image.load()\n\n    return np.expand_dims(np.asarray(pil_image, dtype=\"float32\"), 2)\n<\/code><\/pre>\n\n<p>or (maybe even better):<\/p>\n\n<pre><code>def pil_to_nparray(pil_image):\n    pil_image.load()\n\n    return np.asarray(pil_image, dtype=\"float32\").reshape((224, 224, 1))\n<\/code><\/pre>\n\n<p>(the latter version looks more direct, you know exactly what it does)\nBut this only works if the input image is 224x224, while the <code>expand_dims<\/code> would always add the extra dimension, for any size.<\/p>\n"}]}
{"QuestionId":42117693,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><lstm>","CreationDate":"2017-02-08T15:51:18.363","AcceptedAnswerId":null,"OwnerUserId":5082661.0,"Title":"Acessing Weights of LSTM Cell in Tensorflow","Body":"<p>How do i get the weights of an LSTM Cell. I would like to assign the weights of this to another LSTM Cell in another network, i.e copy contents.<\/p>\n\n<p>If this was normal weights, i would do sess.run(w1.assign(w2),..)<\/p>\n","answers":[{"AnswerId":"42120089","CreationDate":"2017-02-08T17:43:14.970","ParentId":null,"OwnerUserId":"6791223","Title":null,"Body":"<p>You could do this using <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a0d784bdd31b27e013a7eac58a86ba62e86db299\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard9\/tf.get_variable.md\" rel=\"nofollow noreferrer\">tf.get_variable<\/a> function. \nAll you need is the names of those variables. <\/p>\n\n<p>You can get them by running something like this:<\/p>\n\n<pre><code>for v in tf.global_variables():\n    print(v.name)\n<\/code><\/pre>\n"}]}
{"QuestionId":42117777,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><theano><keras>","CreationDate":"2017-02-08T15:54:53.733","AcceptedAnswerId":null,"OwnerUserId":7343062.0,"Title":"Keras with Theano BackEnd","Body":"<p>im new to Keras in python, i got this warning message when after executing my code. I tried to search on google, but still didnt manage to solve this problem. Thank you in advance.<\/p>\n\n<blockquote>\n  <p>UserWarning: he \"show_accuracy\" argument is deprecated, instead you\n  should pass the \"accurac \" metric to the model at compile time:\n  model.compile(optimizer, loss, metrics=[\"accuracy\"])\n  warnings.warn('The \"show_accuracy\" argument is deprecated, '<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42120658","CreationDate":"2017-02-08T18:15:20.940","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>In Keras &lt; 1.0 (I believe), one would pass the show_accuracy argument to model.fit in order to display the accuracy during training.<\/p>\n\n<p>This method has been replaced by metrics, as you can now define custom metrics that can help you during training. One of the metrics is of course, accuracy. The changes to your code to keep the same behavior are minimum:<\/p>\n\n<ul>\n<li>Remove show_accuracy from the model.fit call.<\/li>\n<li>Add metrics = [\"accuracy\"] to the model.compile call.<\/li>\n<\/ul>\n\n<p>And that's it.<\/p>\n"}]}
{"QuestionId":42117782,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2017-02-08T15:55:18.063","AcceptedAnswerId":null,"OwnerUserId":245549.0,"Title":"How to do the element-wise multiplication over selected dimensions in Theano?","Body":"<p>I have two tensors and I need to do the following:<\/p>\n\n<pre><\/pre>\n\n<p>So, as you can see there is not summation over  and . So, I cannot apply .<\/p>\n","answers":[{"AnswerId":"42160527","CreationDate":"2017-02-10T13:26:58.653","ParentId":null,"OwnerUserId":"245549","Title":null,"Body":"<p>Here is the solution:<\/p>\n\n<pre><code>C = a.dimshuffle(0, 1, 2, 3, 'x', 'x') * b.dimshuffle('x', 'x', 2, 3, 0, 1)\n<\/code><\/pre>\n\n<p>And here is the explanation:<\/p>\n\n<p>You try to archive two goals:<\/p>\n\n<ol>\n<li>Pair only some indices of the first tensor with some indices of the second tensor.<\/li>\n<li>Do not sum up over the paired indices.<\/li>\n<\/ol>\n\n<p>If you use pairwise multiplication, then there will be no summation over the paired indices (unlike in <code>tensordot<\/code>, which sums up over the paired indices \/ axes). So, the second goal is achieved by using pairwise multiplication.<\/p>\n\n<p>Now you need to achieve the first goal. Per default pairwise multiplication pairs all the indices of the first tensor with all the corresponding indices of the second tensor (first with first, second with second and so on). To avoid that use <code>dimshuffle<\/code> to align indexes that have to be paired and use 'x' to introduce \"dummy\" axes for the indices that should not be paired. In your particular example it should like like that:<\/p>\n"}]}
{"QuestionId":42118099,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-08T16:08:47.037","AcceptedAnswerId":null,"OwnerUserId":5951780.0,"Title":"Restore tf-Model for CPU saved under GPU","Body":"<p>I saved a model (tf version 0.12), trained on a machine with GPU. Now I need to restore it on a machine that only uses the CPU. I don't have access to the GPU-machine as I am on travel and would have 2 days to work on the scripts, but need the model in order to do so.<\/p>\n\n<p>Any ideas how to restore the GPU-model on the CPU-only machine?<\/p>\n\n<p>Error message on the CPU machine is:<\/p>\n\n<pre><\/pre>\n\n<p>Looking for a hack. Can I rename the tensornames to pretend they come from a CPU?<\/p>\n","answers":[]}
{"QuestionId":42118197,"AnswerCount":1,"Tags":"<android><react-native><tensorflow><bazel>","CreationDate":"2017-02-08T16:13:36.107","AcceptedAnswerId":null,"OwnerUserId":2190503.0,"Title":"tensorflow android with react-native can't work normally","Body":"<p>I put react-native.jar to tensorflow android's directory and add configuration to  file like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now run  work well. But then use  in java file and run  again, it throws a nonexistent error, can't find  package.<\/p>\n","answers":[{"AnswerId":"42304860","CreationDate":"2017-02-17T18:18:30.200","ParentId":null,"OwnerUserId":"5020937","Title":null,"Body":"<p>In order to expose the classes in the JAR to the Java code in the Android build, you need to add a dependency in the android_binary that you are building on the java_import you have created.<\/p>\n\n<p>For example:<\/p>\n\n<pre>\n    # tensorflow\/examples\/android\/BUILD\n\n    java_import(\n        name = \"react-native\",\n        jars = [\n            \"react-native-0.39.2-sources.jar\",\n        ],\n    )\n\n    android_binary(\n        name = \"tensorflow_demo\",\n        srcs = glob([\"src\/**\/*.java\"]),\n        deps = [\n            \":tensorflow_native_libs\",\n            \"\/\/tensorflow\/contrib\/android:android_tensorflow_inference_java\",\n            \"<b>:react-native<\/b>\",\n        ],\n    )\n\n<\/pre>\n"}]}
{"QuestionId":42118610,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-08T16:31:53.760","AcceptedAnswerId":null,"OwnerUserId":4811003.0,"Title":"Tensorflow stucks at run(init_op) (In distributed environment)","Body":"<p>When I run a program using Estimator in a distributed environment but it stucks at sess.run(init_op) where init_op is just . More specifically, it stucks at  of  which is <\/p>\n\n<pre><\/pre>\n\n<p>The graph_def.SerializeToString() is quite strange and I take part of them here. <\/p>\n\n<p>b'\\ng\\n\\x1dglobal_step\/Initializer\/Const\\x12\\x05Const*\\x12\\n\\x05value\\x12\\tB\\x07\\x08\\t\\x12\\x00R\\x01\\x00*\\x1e\\n\\x06_class\\x12\\x14\\n\\x12\\x12\\x10loc:@global_step*\\x0b\\n\\x05dtype\\x12\\x020\\t\\n\\x87\\x01\\n\\x0bglobal_step\\x12\\nVariableV2\"\\x0e\/job:ps\/task:0*\\x0b\\n\\x05shape\\x12\\x02:\\x00*\\x0b\\n\\x05dtype\\x12\\x020\\t*\\x1e\\n\\x06_class\\x12\\x14\\n\\x12\\x12\\x10loc:@global_step*\\x11\\n\\x0bshared_name\\x12\\x02\\x12\\x00*\\x0f\\n\\tcontainer\\x12\\x02\\x12\\x00\\n\\xaa\\x01\\n\\x12global_step\/Assign\\x12\\x06Assign\\x1a\\x0bglobal_step\\x1a\\x1dglobal_step\/Initializer\/Const\"\\x0e\/job:ps\/task:0*\\x11\\n\\x0buse_locking\\x12\\x02(\\x01*\\x07\\n\\x01T\\x12\\x020\\t*\\x1e\\n\\x06_class\\x12\\x14\\n\\x12\\x12\\x10loc:@global_step*\\x14\\n\\x0evalidate_shape\\x12\\x02(\\x01\\nb\\n\\x10global_step\/read\\x12\\x08Identity\\x1a\\x0bglobal_step\"\\x0e\/job:ps\/task:0*\\x1e\\n\\x06_class\\x12\\x14\\n\\x12\\x12\\x10loc:@global_step*\\x07\\n\\x01T\\x12\\x020\\t\\n\\xdc\\xf3\\x0f\\n\\x05Const\\x12\\x05Const\"\\x12\/job:master\/task:0*\\xa9\\xf3\\x0f\\n\\x05value\\x12\\x9e\\xf3\\x0fB\\x9a\\xf3\\x0f\\x08\\t\\x12\\n\\x12\\x04\\x08\\xb1\\xfe\\x01\\x12\\x02\\x08\\x01\"\\x88\\xf3\\x0f\\'\\x00\\x00\\x00\\x00\\x00\\x00\\x002\\x00\\x00\\x00\\x00\\x00\\x00\\x00&amp;\\x00\\x00\\x00\\x00\\x00\\x00\\x005\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x1c\\x00\\x00\\x00\\x00\\x00\\x00\\x00%\\x00\\x00\\x00\\x00\\x00\\x00\\x001\\x00\\x00\\x00\\x00\\x00\\x00\\x004\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x1f\\x00\\x00\\x00\\x00\\x00\\x00\\x00*\\x00\\x00\\x00\\x00\\x00\\x00\\x00%\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x1e\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x17\\x00\\x00\\x00\\x00\\x00\\x00\\x00 \\x00\\x00\\x00\\x00\\x00\\x00\\x00(\\x00\\x00\\x00\\x00\\x00\\x00\\x00\"\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x19\\x00\\x00\\x00\\x00\\x00\\x00\\x00 \\x00\\x00\\x00\\x00\\x00\\x00\\x00&amp;\\x00\\x00\\x00\\x00\\x00\\x00\\x00+\\x00\\x00\\x00\\x00\\x00\\x00\\x00(\\x00\\x00\\x00\\x00\\x00\\x00\\x006\\x00\\x00\\x00\\x00\\x00\\x00\\x00#\\x00\\x00\\x00\\x00\\x00\\x00\\x00+\\x00\\x00\\x00\\x00\\x00\\x00\\x00;\\x00\\x00\\x00\\x00\\x00\\x00\\x008\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\x13\\x00\\x00\\x00\\x00\\x00\\x00\\x006\\x00\\x00\\x00\\x00\\x00\\x00\\x00\\'\\x00\\x00<\/p>\n\n<p>What are the potential reasons for these output? And how can I possibly solve this problem?<\/p>\n\n<p>Thanks.<\/p>\n\n<hr>\n\n<p>My distributed settings for estimator is <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42118821,"AnswerCount":2,"Tags":"<image-processing><keras><deep-learning><image-segmentation>","CreationDate":"2017-02-08T16:41:06.410","AcceptedAnswerId":42127607.0,"OwnerUserId":4631796.0,"Title":"Cross Entropy Loss for Semantic Segmentation Keras","Body":"<p>I'm pretty sure this is a silly question but I can't find it anywhere else so I'm going to ask it here.<\/p>\n\n<p>I'm doing semantic image segmentation using a cnn (unet) in keras with 7 labels. So my label for each image is (7,n_rows,n_cols) using the theano backend. So across the 7 layers for each pixel, it's one-hot encoded. In this case, is the correct error function to use categorical cross-entropy? It seems that way to me but the network seems to learn better with binary cross-entropy loss. Can someone shed some light on why that would be and what the principled objective is?<\/p>\n","answers":[{"AnswerId":"42127607","CreationDate":"2017-02-09T03:15:01.843","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Binary cross-entropy loss should be used with <code>sigmod<\/code> activation in the last layer and it severely penalizes opposite predictions. It does not take into account that the output is a one-hot coded and the sum of the predictions should be 1. But as mis-predictions are severely penalizing the model somewhat learns to classify properly. <\/p>\n\n<p>Now to enforce the prior of one-hot code is to use <code>softmax<\/code> activation with categorical cross-entropy. This is what you should use. <\/p>\n\n<p>Now the problem is using the <code>softmax<\/code> in your case as Keras don't support softmax on each pixel. <\/p>\n\n<p>The easiest way to go about it is permute the dimensions to (n_rows,n_cols,7) using <code>Permute<\/code> layer and then reshape it to (n_rows*n_cols,7) using <code>Reshape<\/code> layer. Then you can added the <code>softmax<\/code> activation layer and use crossentopy loss. The data should also be reshaped accordingly.<\/p>\n\n<p>The other way of doing so will be to implement depth-softmax :<\/p>\n\n<pre><code>def depth_softmax(matrix):\n    sigmoid = lambda x: 1 \/ (1 + K.exp(-x))\n    sigmoided_matrix = sigmoid(matrix)\n    softmax_matrix = sigmoided_matrix \/ K.sum(sigmoided_matrix, axis=0)\n    return softmax_matrix\n<\/code><\/pre>\n\n<p>and use it as a lambda layer:<\/p>\n\n<pre><code>model.add(Deconvolution2D(7, 1, 1, border_mode='same', output_shape=(7,n_rows,n_cols)))\nmodel.add(Permute(2,3,1))\nmodel.add(BatchNormalization())\nmodel.add(Lambda(depth_softmax))\n<\/code><\/pre>\n\n<p>If <code>tf<\/code> <code>image_dim_ordering<\/code> is used then you can do way with the <code>Permute<\/code> layers.<\/p>\n\n<p>For more reference check <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1169\" rel=\"noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"53224171","CreationDate":"2018-11-09T10:42:38.833","ParentId":null,"OwnerUserId":"4444978","Title":null,"Body":"<p>I tested the solution of @indraforyou and think that the proposed method has some mistakes. As the commentsection does not allow for proper code segments, here is what I think would be the fixed version:<\/p>\n\n<pre><code>def depth_softmax(matrix):\n\n    from keras import backend as K\n\n    exp_matrix = K.exp(matrix)\n    softmax_matrix = exp_matrix \/ K.expand_dims(K.sum(exp_matrix, axis=-1), axis=-1)\n    return softmax_matrix\n<\/code><\/pre>\n\n<p>This method will expect the ordering of the matrix to be (height, width, channels). <\/p>\n"}]}
{"QuestionId":42118850,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><generator><keras>","CreationDate":"2017-02-08T16:42:43.060","AcceptedAnswerId":null,"OwnerUserId":1982806.0,"Title":"How to retrieve the filename of an image with keras flow_from_directory shuffled method?","Body":"<p>If I don't shuffle my files, I can get the file names with . But when the generator shuffles the images, filenames isn't shuffled, so I don't know how to get the file names back.<\/p>\n","answers":[{"AnswerId":"56845128","CreationDate":"2019-07-02T03:59:44.673","ParentId":null,"OwnerUserId":"150851","Title":null,"Body":"<p>Internally, the <code>DirectoryIterator<\/code> will iterate over an <code>index_arrray<\/code> which is shuffled when you ask it to.<\/p>\n\n<p>You just need to index the filename array using the current indexes of the batch:<\/p>\n\n<pre><code>it = ImageDataGenerator().flow_from_directory(shuffle=True,...)\nfor img in it:\n    idx = (it.batch_index - 1) * it.batch_size\n    fnames = [it.filenames[it.index_array[i]] for i in range(idx, idx + it.batch_size)]\n<\/code><\/pre>\n"},{"AnswerId":"46009804","CreationDate":"2017-09-02T01:59:16.343","ParentId":null,"OwnerUserId":"2403473","Title":null,"Body":"<p>I think the only option here is to NOT shuffle the files. I have been wondering this myself and this is the only thing I could find in the docs. Seems odd and not correct...<\/p>\n"}]}
{"QuestionId":42118910,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2017-02-08T16:45:30.000","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Theano TensorVariable object has no attribute get_value","Body":"<p>I have the following theano object:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I try to slice it then get the shape again<\/p>\n\n<pre><\/pre>\n\n<p>Can someone explain to me why after slicing, i get this error?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42330796","CreationDate":"2017-02-19T17:59:05.943","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p><code>train_set_x<\/code> is a theano shared variable, so you can't set it like <code>train_set_x = train_set_x[:100]<\/code>.<\/p>\n\n<p>You should try using set_value():<\/p>\n\n<pre><code>train_set_x.set_value(train_set_x.get_value()[:100])\n<\/code><\/pre>\n"}]}
{"QuestionId":42119238,"AnswerCount":1,"Tags":"<tensorflow><queue>","CreationDate":"2017-02-08T16:58:46.430","AcceptedAnswerId":42119835.0,"OwnerUserId":3002273.0,"Title":"TensorFlow: FIFOQueue's DequeueMany and DequeueUpTo require the components to have specified shapes","Body":"<p>This Tensorflow code using FIFOQueue causes the below error<\/p>\n\n<pre><\/pre>\n\n<p>The error<\/p>\n\n<pre><\/pre>\n\n<p>Can you please tell me what I am doing wrong?<\/p>\n","answers":[{"AnswerId":"42119835","CreationDate":"2017-02-08T17:30:52.140","ParentId":null,"OwnerUserId":"3002273","Title":null,"Body":"<p>Asked too soon. Perhaps I will save future generations<\/p>\n\n<pre><code>with tf.Session() as sess:\n    queue = tf.FIFOQueue(100, dtypes=[tf.float32, tf.float32], shapes=[(), ()])\n    enqueue_op = queue.enqueue_many([[1.2, 2.3], [4.5, 6.7]])\n    inputs = queue.dequeue_many(4)\n    sess.run(enqueue_op)\n    sess.run(enqueue_op)\n    print sess.run(inputs)\n<\/code><\/pre>\n\n<p>prints out,<\/p>\n\n<pre><code>[array([ 1.20000005,  2.29999995,  1.20000005,  2.29999995], dtype=float32), \n array([ 4.5       ,  6.69999981,  4.5       ,  6.69999981], dtype=float32)]\n<\/code><\/pre>\n"}]}
{"QuestionId":42119900,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-08T17:34:07.920","AcceptedAnswerId":42125544.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Is dequeue_up_to slower than dequeue_many in tf.train.batch?","Body":"<p>In the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/input.py#L744\" rel=\"nofollow noreferrer\">source code<\/a> on GitHub, If I use  in ,  is used instead of  for all batches. Is  slower? I cannot find the source code for this somehow even after a search in the TensorFlow repository. I have traced the  and  functions up till this file <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a0d784bdd31b27e013a7eac58a86ba62e86db299\/tensorflow\/python\/ops\/data_flow_ops.py#L440\" rel=\"nofollow noreferrer\">here<\/a> but I cannot find what is  and its functions, and a search in the repo only returns results of  getting imported. Why is this the case?<\/p>\n","answers":[{"AnswerId":"42125544","CreationDate":"2017-02-08T23:27:41.727","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>Difficulty tracing code Python code paths through to C++ ops is an unfortunate consequence of TensorFlow's op wrapping technique. Generally the C++ implementations are named FooBarOp, while the Python ends up calling foo_bar in generated code.<\/p>\n\n<p>In this case <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/data_flow_ops.py#L509\" rel=\"nofollow noreferrer\">gen_data_flow_ops._queue_dequeue_up_to_v2<\/a> is an automatically generated Python wrapper for the registration of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/queue_ops.cc#L372\" rel=\"nofollow noreferrer\">QueueDequeueUpToV2<\/a>, which is an alias for the C++ <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/queue_ops.cc#L318\" rel=\"nofollow noreferrer\">DequeueUpToOp<\/a>.<\/p>\n\n<p>To answer your original question, there is unlikely to be any significant performance difference from the queue itself (the UpTo version of dequeue only does something different once the queue is closed). Enabling <code>allow_small_batch<\/code> will remove some static shape information from the graph (batch sizes), however, which could conceivably make some ops downstream a bit slower if they optimize based on static shapes.<\/p>\n"}]}
{"QuestionId":42120723,"AnswerCount":1,"Tags":"<python><tensorflow><attributeerror>","CreationDate":"2017-02-08T18:19:04.870","AcceptedAnswerId":42310732.0,"OwnerUserId":1514682.0,"Title":"Tensorflow: 'module' object has no attribute 'RunMetaData'","Body":"<p>I am trying to get GPU\/CPU time stats using tensorflow timeline module but on <\/p>\n\n<pre><\/pre>\n\n<p>It gives following error.<\/p>\n\n<pre><\/pre>\n\n<p>I am using  version. \nI also go to  to actually see the function definition but It didn't exist there. \nP.S: I have imported \n <br\/> [Code snippet]<\/p>\n\n<pre><\/pre>\n\n<p><br\/>\n[ERROR SNIPPET]<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42310732","CreationDate":"2017-02-18T03:27:14.267","ParentId":null,"OwnerUserId":"1514682","Title":null,"Body":"<p>I needed to install TensorFlow 0.12.0 or higher.<\/p>\n"}]}
{"QuestionId":42120995,"AnswerCount":1,"Tags":"<deep-learning><caffe><concat>","CreationDate":"2017-02-08T18:33:38.807","AcceptedAnswerId":42182564.0,"OwnerUserId":6786718.0,"Title":"Does the Concatenation Layer produce really a 4D output in caffe?","Body":"<p>I read the documentation about the concatination layer here: <a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers\/concat.html\" rel=\"nofollow noreferrer\">Layer Catalogue Concat<\/a>. It states: <\/p>\n\n<blockquote>\n  <p>input:<\/p>\n  \n  <p><\/p>\n  \n  <p>Output:<\/p>\n  \n  <p>if axis = 0: , and all input c_i\n  should be the same. <\/p>\n  \n  <p>if axis = 1: , and all input n_i should be the same.<\/p>\n<\/blockquote>\n\n<p>However, I am having a hard time imagine this, like how can there be a 4 dimensional output, when all layers accept 3D input? Is there some kind of trick to read the 4D output as a 3D output?<\/p>\n","answers":[{"AnswerId":"42182564","CreationDate":"2017-02-11T23:42:20.333","ParentId":null,"OwnerUserId":"2878130","Title":null,"Body":"<p>Actually input and output have all 4D: batch dimension, number of channels, height and width. You can obtain a different number of dimensions in special cases (e.g. 5D for RGB-D input), but for standard RGB images, 4D is kept everywhere (except for fully connected layers). <\/p>\n"}]}
{"QuestionId":42121114,"AnswerCount":0,"Tags":"<python><keras><convolution><bounding-box>","CreationDate":"2017-02-08T18:40:10.197","AcceptedAnswerId":null,"OwnerUserId":7446360.0,"Title":"How to create bounding boxes around the ROIs using keras","Body":"<p>I have to questions regarding  and Convolutional Neural Networks:<\/p>\n\n<p>1.\n How to create bounding boxes using  and convolution neural network ?<\/p>\n\n<p>2.\n How to use keras  application retraining?<\/p>\n","answers":[]}
{"QuestionId":42121596,"AnswerCount":1,"Tags":"<python><tensorflow><entropy><mnist>","CreationDate":"2017-02-08T19:06:44.250","AcceptedAnswerId":42122337.0,"OwnerUserId":3813674.0,"Title":"Entropy of a MNIST image with Tensorflow","Body":"<p>I want to measure the entropy of an MNIST image. However most of the intensity values are 0, and  returns . I know that I can simply calculate the entropy:<\/p>\n\n<pre><\/pre>\n\n<p>where X is a placeholder the MNIST image batch, defined as . However, the result is  (many pixel values are 0, so  will give ). Is there a way to work around this, or another method to calculate the entropy of one image?<\/p>\n","answers":[{"AnswerId":"42122337","CreationDate":"2017-02-08T19:49:28.657","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>Entropy depends on the <em>probability<\/em> of each intensity value, not on the value itself.  A value with a probability of 0 doesn't enter into the calculations.<\/p>\n\n<p>Write code to compute the proportion of each intensity value that appears in the photo.  <em>Those<\/em> form your X vector.<\/p>\n\n<hr>\n\n<p>I think you misunderstand the concept of entropy.  You can look up <a href=\"https:\/\/en.wikipedia.org\/wiki\/Entropy_%28information_theory%29\" rel=\"nofollow noreferrer\">that part<\/a>, as well as the calculation.<\/p>\n\n<p>In general, it's the average amount of surprise you get with any one transmitted sampling of the distribution.  Another way to interpret it is to design the best binary coding, given the frequency distribution (Huffman code).  The average number of bits transmitted is the entropy.<\/p>\n\n<p>Now, back to your case.  To simplify the numbers, let's consider a 10x10 image, with only four intensity levels, 0-3.  70 of the values are 0, 20 are 1, with six 2's and four 3's filling out the set.  In this case, your <strong>X<\/strong> array above has only four values: [0.70, 0.20, 0.06, 0.04].  There are no zero or negative values to foul your log calculations.<\/p>\n\n<p>To do this by hand, let's consider the Huffman coding.  The simple-minded case is to encode each value with its 2-bit binary equivalent: 00, 01, 10, 11.  This gives us a flat rate of 2 bits per value transmitted.<\/p>\n\n<p>However, we can do better.  If we encode them as<\/p>\n\n<pre><code>0: 0\n1: 10\n2: 110\n3: 111\n<\/code><\/pre>\n\n<p>Then our average number of bits transmitted is:<\/p>\n\n<pre><code>0.70*1 + 0.20*2 + 0.06*3 + 0.04*3\n= 0.70 + 0.40 + 0.18 + 0.12\n= 1.40 bits\n<\/code><\/pre>\n\n<p>... so 1.40 is the entropy for this image.  Note that we didn't take the log of any of the <em>values<\/em>, merely of the <em>frequencies<\/em>.<\/p>\n"}]}
{"QuestionId":42121914,"AnswerCount":0,"Tags":"<python><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-02-08T19:25:03.420","AcceptedAnswerId":null,"OwnerUserId":7249443.0,"Title":"Issue with Caffe's Input layer. Could not open or find file","Body":"<p>I'm fine-tuning Alexnet caffe model on my data. I'm using \"ImageData\" type for Input layer. In my train.txt file (the path of which I gave to \u201csource\u201d in input layer), there aren\u2019t written extensions after image names (.jpg). But in my images directory names of the images are ending with .jpg. How can I fix this? Btw, I'm not permitted to make changes on data. <\/p>\n\n<pre><\/pre>\n\n<p>Below is the path to my images and some examples of files.<\/p>\n\n<pre><\/pre>\n\n<p>And this is the path to train.txt file, and some lines from it:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you very much!<\/p>\n","answers":[]}
{"QuestionId":42122168,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><keras>","CreationDate":"2017-02-08T19:39:12.273","AcceptedAnswerId":42123881.0,"OwnerUserId":6380584.0,"Title":"Keras - How to construct a shared Embedding() Layer for each Input-Neuron","Body":"<p>I want to create a deep neural network in keras, where each element of the input layer is \"encoded\" using the same, shared Embedding()-layer, before it is fed into the deeper layers.<\/p>\n\n<p>Each input would be a number that defines the type of an object, and the network should learn an embedding that encapsulates some internal representation of \"what this object is\".<\/p>\n\n<p>So, if the input layer has X dimensions, and the embedding has Y dimensions, the first hidden layer should consist of X*Y neurons (each input neuron embedded).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Quise.png\" rel=\"noreferrer\">Here is a little image that should show the network architecture that I would like to create, where each input-element is encoded using a 3D-Embedding<\/a><\/p>\n\n<p>How can I do this?<\/p>\n","answers":[{"AnswerId":"42123881","CreationDate":"2017-02-08T21:25:01.213","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<pre><code>from keras.layers import Input, Embedding\n\nfirst_input = Input(shape = (your_shape_tuple) )\nsecond_input = Input(shape = (your_shape_tuple) )\n...\n\nembedding_layer = Embedding(embedding_size)\n\nfirst_input_encoded = embedding_layer(first_input)\nsecond_input_encoded = embedding_layer(second_input)\n...\n\nRest of the model....\n<\/code><\/pre>\n\n<p>The emnedding_layer will have shared weights. You can do this in form of lists of layers if you have a lot of inputs. <\/p>\n\n<p>If what you want is transforming a tensor of inputs, the way to do it is :<\/p>\n\n<pre><code>from keras.layers import Input, Embedding\n\n# If your inputs are all fed in one numpy array :\ninput_layer = Input(shape = (num_input_indices,) )\n\n# the output of this layer will be a 2D tensor of shape (num_input_indices, embedding_size)\nembedded_input = Embedding(embedding_size)(input_layer)\n<\/code><\/pre>\n\n<p>Is this what you were looking for?<\/p>\n"}]}
{"QuestionId":42122343,"AnswerCount":0,"Tags":"<python-3.x><indexing><tensorflow>","CreationDate":"2017-02-08T19:49:52.983","AcceptedAnswerId":null,"OwnerUserId":5818295.0,"Title":"Using tf.gather_nd() to select elements of a tensor","Body":"<p>I have 2 tensors:<\/p>\n\n<ul>\n<li>A of shape [146, 33, 559]<\/li>\n<li>B of shape [146, 33]<\/li>\n<\/ul>\n\n<p>B contains integers between 0 and 559, which serve as indeces.<\/p>\n\n<p>What I'm after is a tensor C of shape [146, 33] where each element corresponds to the index given by B.<\/p>\n\n<p>I tried  which gives me the error<\/p>\n\n<pre><\/pre>\n\n<p>I also tried  which gives me the error<\/p>\n\n<pre><\/pre>\n\n<p>Any idea how to resolve this?<\/p>\n","answers":[]}
{"QuestionId":42122460,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-08T19:57:12.070","AcceptedAnswerId":null,"OwnerUserId":987397.0,"Title":"Sharing weights in Tensorflow between two subgraphs","Body":"<p>I have the following setup, where each input consists of two trajectories. I want that the left graph has the same weight as the right graph<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qo3Aj.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qo3Aj.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I tried to follow the approach described here for sharing variables, <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/how_tos\/variable_scope\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/r1.0\/how_tos\/variable_scope\/<\/a>, however it is not working. Two different graphs are created. What am I doing wrong?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42123168","CreationDate":"2017-02-08T20:40:21.390","ParentId":null,"OwnerUserId":"987397","Title":null,"Body":"<p>Okay, use tf.get_variable instead of tf.Variable for this. This works<\/p>\n\n<pre><code>def build_t_model(trajectories):\n            \"\"\"\n            Build the trajectory network\n            \"\"\"\n            with tf.name_scope('h1_t'):\n                weights = tf.get_variable(\n                    'weights1', \n                    shape=[150, h1_t_units],\n                    initializer=tf.truncated_normal_initializer(\n                        stddev=1.0\/math.sqrt(float(150))))\n                biases = tf.get_variable(\n                    'biases1', \n                    initializer=tf.zeros_initializer(shape=[h1_t_units]))\n                h1_t = tf.nn.relu(tf.matmul(trajectories, weights) + biases)\n\n            with tf.name_scope('h2_t'):\n                weights = tf.get_variable(\n                    'weights2', \n                    shape=[h1_t_units, h2_t_units],\n                    initializer=tf.truncated_normal_initializer(\n                        stddev=1.0\/math.sqrt(float(h1_t_units))))\n                biases = tf.get_variable(\n                    'biases2', \n                    initializer=tf.zeros_initializer(shape=[h2_t_units]))\n                h2_t = tf.nn.relu(tf.matmul(h1_t, weights) + biases)\n\n            with tf.name_scope('h3_t'):\n                weights = tf.get_variable(\n                    'weights3', \n                    shape=[h2_t_units, M],\n                    initializer=tf.truncated_normal_initializer(\n                        stddev=1.0\/math.sqrt(float(h2_t_units))))\n                biases = tf.get_variable(\n                    'biases3', \n                    initializer=tf.zeros_initializer(shape=[M]))\n                h3_t = tf.nn.relu(tf.matmul(h2_t, weights) + biases)\n            return h3_t\n<\/code><\/pre>\n"}]}
{"QuestionId":42122539,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><neural-network>","CreationDate":"2017-02-08T20:01:02.770","AcceptedAnswerId":43237703.0,"OwnerUserId":4632965.0,"Title":"tensorflow stop optimizing when activation step is in a function","Body":"<p>I'm trying to reproduce the result of <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/mnist\/beginners\/<\/a><\/p>\n\n<p>So I designed several functions to take care of the training step such as these two:<\/p>\n\n<pre><\/pre>\n\n<p>By using these two functions for the training<\/p>\n\n<pre><\/pre>\n\n<p>The optimization stops at a cost of 2.3... and the overall accuracy is of 10% whereas in the example it get closer to zero and the accuracy is close to 96%. Does anyone have an explanation for this peculiar behavior?<\/p>\n\n<p>PS when I'm using <i>layer_computation<\/i> in the source code in the example I also get stuck at a cost of 2.3.<\/p>\n","answers":[{"AnswerId":"43237703","CreationDate":"2017-04-05T17:18:58.390","ParentId":null,"OwnerUserId":"4632965","Title":null,"Body":"<p>I caught the error, I was trying to perform back-propagation on the last layer. This question may have been better on cross-validated.<\/p>\n"}]}
{"QuestionId":42122826,"AnswerCount":3,"Tags":"<python><ubuntu><numpy><docker><tensorflow>","CreationDate":"2017-02-08T20:19:19.597","AcceptedAnswerId":null,"OwnerUserId":1601580.0,"Title":"Can one use python 3.5 in a docker container based out of the TensorFlow docker image?","Body":"<p>I was trying to use python 3.5 with my docker container. I tried:<\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre>\n\n<p>but it seems that both images only have python version up to 3.4. Is it possible to have as base image the docker container but also have python 3.5? Or even better, is it possible to have the base image from the official tensorflow image have python 3.5 itself?<\/p>\n\n<p>I know its possible to pip install it in the Dockerfile as in (as shown in the <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#pip_installation\" rel=\"nofollow noreferrer\">tf download page<\/a>):<\/p>\n\n<pre><\/pre>\n\n<p>however that seems that would <strong>not<\/strong> get me the <strong>latest<\/strong> tensorflow version. If one can pip install the most recent TensorFlow version why does the latest base image not a way to get the most recent TensorFlow build <strong>and<\/strong> have it in python3.5?<\/p>\n\n<hr>\n\n<p>I have definitively tried installing python 3.5 as suggested by <a href=\"https:\/\/askubuntu.com\/questions\/682869\/how-do-i-install-newer-python-versions-using-apt-get\">here<\/a> however, even though the installation of python 3.5 is successful, it breaks numpy in a way I can't fix (as explained <a href=\"https:\/\/stackoverflow.com\/questions\/42122639\/how-does-one-install-fix-a-failed-numpy-installation-that-works-on-python-3-4-bu\">here<\/a>). Honestly, the best solution would be to just have python 3.5 automatically available on the image but for some reason its not there. I have done some research on this and it seems to install python 3.5 its a little difficult. Why is that? Is the reason python 3.5 is missing is because of tensorflow or because of ubuntu? My ideal solution would be to not have me install python 3.5 and that it comes, but it seems there might be a fundamental issue with this. What is it? Is it just because it has not been installed for tensorflow docker image and ubuntu, or am I over complicated a simple problem?<\/p>\n\n<hr>\n\n<p>as another solution, I was thinking maybe to install anaconda or something and then do that, but I wanted to have tensorflow as my base image and it seems anaconda suggests to have their image as base. Since there isn't an easy way to install anaconda with apt-install I am still working to see how I can programatically install anaconda so that there can be a tensorflow image as base and then install as instructed in a Dockerfile, some version of anaconda.<\/p>\n\n<hr>\n\n<p>There is now a git issue ine official tensorflow for this:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7368\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7368<\/a><\/p>\n\n<hr>\n\n<p>I mentioned that one can just install TensorFlow in the DockerFile directly so here is an example docker file that worked for me:<\/p>\n\n<pre><\/pre>\n\n<p>I think the only interesting thing to note is that I installed pip directly because that package\/intallation of python3.5 doesn't come with pip for it for some reason. That lead to me to install python packages to use:<\/p>\n\n<pre><\/pre>\n\n<p>instead of<\/p>\n\n<pre><\/pre>\n\n<p>you can see more of those details here: <a href=\"https:\/\/stackoverflow.com\/questions\/42122639\/how-does-one-install-fix-a-failed-numpy-installation-that-works-on-python-3-4-bu\/42124828#42124828\">How does one install\/fix a failed numpy installation that works on python 3.4 but not in 3.5?<\/a><\/p>\n\n<p>Also note that I had issues installing python the \"official way\" (i.e. with apt-get or something like that) so I resorted to what the following question\/answer suggested: <a href=\"https:\/\/askubuntu.com\/questions\/682869\/how-do-i-install-newer-python-versions-using-apt-get\">https:\/\/askubuntu.com\/questions\/682869\/how-do-i-install-newer-python-versions-using-apt-get<\/a><\/p>\n","answers":[{"AnswerId":"54245703","CreationDate":"2019-01-17T23:21:18.010","ParentId":null,"OwnerUserId":"10893536","Title":null,"Body":"<p>I have tried to pull in a python:3.5 image and install tensorflow basing on this image, which works.<\/p>\n\n<p>what I have in the dockerfile:<\/p>\n\n<p>FROM python:3.5<\/p>\n\n<p>RUN pip install tensorflow<\/p>\n"},{"AnswerId":"42144504","CreationDate":"2017-02-09T18:41:59.960","ParentId":null,"OwnerUserId":"5087125","Title":null,"Body":"<p>With a plain 'ubuntu' docker image from DockerHub and relying on pip to do its own dependency resolution (I wonder if <em>not<\/em> doing that is what causes your numpy problems) I ran: <\/p>\n\n<pre><code>apt-get install python3\napt-get install python3-pip \npip install tensorflow\n<\/code><\/pre>\n\n<p>As far as I can tell, this gave me python3.5 with the latest tensorflow - like their docker image but with python3.5.<\/p>\n\n<p>To me, the provided docker image is something that's intended to work 'as is' and presumably the bits and pieces provided are intended to convey the developer's higher confidence that they all work correctly together. If you need to make substantial changes it seems easier and simpler to just start from scratch. <\/p>\n"},{"AnswerId":"45763434","CreationDate":"2017-08-18T18:59:44.127","ParentId":null,"OwnerUserId":"1999265","Title":null,"Body":"<p>Yes. Python 3 docker images are available on dockerhub nightly builds.<\/p>\n\n<h2>CPU only<\/h2>\n\n<p><code>docker pull tensorflow\/tensorflow:nightly-py3<\/code><\/p>\n\n<h2>with GPU support<\/h2>\n\n<p><code>docker pull tensorflow\/tensorflow:nightly-gpu-py3<\/code><\/p>\n\n<hr>\n\n<p><a href=\"https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/tags\/\" rel=\"nofollow noreferrer\">https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/tags\/<\/a>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3467\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3467<\/a><\/p>\n"}]}
{"QuestionId":42123196,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-08T20:41:48.643","AcceptedAnswerId":42172678.0,"OwnerUserId":1524166.0,"Title":"ValueError: No variables to save when saving Tensorflow checkpoint","Body":"<p>I've been trying to use the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"nofollow noreferrer\">freeze_graph<\/a> function to get a model + weights\/biases, but in the process, I found that my inception network does not seem to have any variables, despite being able to correctly classify image. My code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><\/pre>\n\n<p>If you don't already have the inception network downloaded, this will get you the .pb file:<\/p>\n\n<p><\/p>\n\n<p>Also, here's a gist of my full code, if anyone is curious: <a href=\"https:\/\/gist.github.com\/mckeesh\/01af030b147e809b0edbc9f67104e859\" rel=\"nofollow noreferrer\">gist<\/a><\/p>\n\n<p>Does anyone know what's going on here? <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42172678","CreationDate":"2017-02-11T05:41:13.343","ParentId":null,"OwnerUserId":"6106106","Title":null,"Body":"<p>What the <code>freeze_graph<\/code> tool does is convert all of the <code>Variable<\/code> nodes in a graph into <code>tf.constant<\/code> nodes- this allows you to save all the information needed to run inference on a graph in a single protobuf file (as opposed to saving the <code>GraphDef<\/code> and <code>Variable<\/code> checkpoint data separately). <\/p>\n\n<p>What you're experiencing here is the result of a successfully frozen graph (<code>tensorflow_inception_graph.pb<\/code>): there are no variables to save because they have all been converted into constants!<\/p>\n"}]}
{"QuestionId":42124000,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-08T21:31:18.573","AcceptedAnswerId":null,"OwnerUserId":6487788.0,"Title":"Reuse graph, and use it multiple times","Body":"<p>I was toying around with style transfer. An important part is re-using an existing model. Currently I import the vgg graph like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now I would like to re-use the vgg model multiple times (as I need to get the output for multiple images). <\/p>\n\n<p>Unfortunately, running this piece of code multiple times makes me import the whole graph again multiple times (waste of RAM and time). <\/p>\n\n<p>So my question is: how can I import this graph once, and re-use it with multiple inputs. <\/p>\n","answers":[]}
{"QuestionId":42124029,"AnswerCount":2,"Tags":"<python><cuda><tensorflow><nvprof>","CreationDate":"2017-02-08T21:33:10.353","AcceptedAnswerId":42310744.0,"OwnerUserId":1514682.0,"Title":"Unable to import nvprof generated profile data","Body":"<p>I am trying to profile a TensorFlow based code using <strong>nvprof<\/strong>. I am using following command for this<\/p>\n\n<pre><\/pre>\n\n<p>The program runs successfully but at the end it shows following error.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42849155","CreationDate":"2017-03-17T04:19:00.707","ParentId":null,"OwnerUserId":"7725034","Title":null,"Body":"<p>use <code>\/usr\/local\/cuda\/bin\/nvprof xxx<\/code>,  maybe you have install two versions of cuda<\/p>\n\n<p>OR \nyou can add <code>\/usr\/local\/cuda\/bin<\/code> to <code>PATH<\/code> enviroment.<\/p>\n\n<pre><code>vim ~\/.bashrc\nexport PATH=$PATH:\/usr\/local\/cuda\/bin\nsource ~\/.bashrc\n<\/code><\/pre>\n"},{"AnswerId":"42310744","CreationDate":"2017-02-18T03:28:42.583","ParentId":null,"OwnerUserId":"1514682","Title":null,"Body":"<p>I don't know the exact reason but using the full path of nvprof <code>\/usr\/bin\/nvprof<\/code> solved the problem.<\/p>\n"}]}
{"QuestionId":42124395,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network><training-data>","CreationDate":"2017-02-08T21:55:36.887","AcceptedAnswerId":42124738.0,"OwnerUserId":365925.0,"Title":"What are 'batches' and 'steps' in TensorFlow Estimators and how do they differ from epochs?","Body":"<p>I am attempting to use TensorFlow's Estimators. In the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/estimators\/#running_the_abalone_model\" rel=\"nofollow noreferrer\">documentation<\/a> the following code is used to train and evaluate the network.<\/p>\n\n<pre><\/pre>\n\n<p>The whole training set is passed in, but we have . Does this mean that only the first 5000 examples from the set are considered? <\/p>\n\n<p>What does the  parameter mean in this context, and how does it interact with ?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42124738","CreationDate":"2017-02-08T22:21:06.777","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p><strong>batch_size<\/strong> is the number of examples processed at once.  TF pushes all of those through one forward pass (in parallel) and follows with a back-propagation on the same set.  This is one <em>iteration<\/em>, or <em>step<\/em>.<\/p>\n\n<p>The <strong>steps<\/strong> parameter tells TF to run 5000 of these iterations to train the model.<\/p>\n\n<p>One <strong>epoch<\/strong> is treating each example in the training set exactly once.  For instance, if you have one million examples and a batch size of 200, then you need 5000 steps to one epoch: 200 * 5.000 = 1.000.000<\/p>\n\n<p>Does that clear up the terminology?<\/p>\n"}]}
{"QuestionId":42124562,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-08T22:08:41.663","AcceptedAnswerId":null,"OwnerUserId":7465271.0,"Title":"tensorflow attributeerror moddule has no attribute per_image_standardization","Body":"<p>I'm trying to go through the tutorial on convolutional neural nets using cifar10.  The cnn is being built (cifar10.py) but when I try to run cifar10_train.py I'm getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>According to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/image.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/image.md<\/a>, there is indeed a per_image_standardization attribute but it looks like my tensorflow doesn't have it.  I'm not sure what version I have and not sure where to find it, but I built it from source from the repository so I imagine it's the current one.<\/p>\n\n<p>I can't find anyone else who is having this problem so I'm stymied.  Maybe I have to write my own?<\/p>\n","answers":[{"AnswerId":"42660863","CreationDate":"2017-03-08T00:29:12.090","ParentId":null,"OwnerUserId":"7465271","Title":null,"Body":"<p>I reinstalled tensorflow and solved the problem.  Thanks, all!<\/p>\n"}]}
{"QuestionId":42124927,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-08T22:34:56.837","AcceptedAnswerId":42126240.0,"OwnerUserId":7474000.0,"Title":"tensorflow's name convention?","Body":"<p>Whenever using tensorflow variables's name, I have had a question why they ended by . So I should use split function like .<\/p>\n\n<p>Despite there is same name's variables, the tail's name wouldn't be changed, like my expectation , .<\/p>\n\n<p>Does anybody know why tensorflow does this?<\/p>\n\n<p>I couldn't find this reason on tensorflow.org or via googling.<\/p>\n","answers":[{"AnswerId":"42126240","CreationDate":"2017-02-09T00:38:23.780","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>The offending integer is added in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/ops.py#L327\" rel=\"nofollow noreferrer\">Tensor class<\/a>, and is the \"value index\":<\/p>\n\n<blockquote>\n  <p>value_index: An <code>int<\/code>. Index of the operation's endpoint that produces\n          this tensor.<\/p>\n<\/blockquote>\n\n<p>It's not always zero! For example:<\/p>\n\n<pre><code>&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; tf.unique([1,2,3])\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>Unique(y=&lt;tf.Tensor 'Unique:0' shape=(?,) dtype=int32&gt;, idx=&lt;tf.Tensor 'Unique:1' shape=(3,) dtype=int32&gt;)\n<\/code><\/pre>\n\n<p>So the first output has the \":0\" suffix, but the second has a \":1\" suffix since it's the second output of the op.<\/p>\n"}]}
{"QuestionId":42124976,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-08T22:37:39.487","AcceptedAnswerId":42126330.0,"OwnerUserId":1044011.0,"Title":"TensorFlow's Estimator froze with low CPU usage","Body":"<p>I updated my TF to v1.0rc1, and  does not work anymore because it froze at . I tried to reproduce this problem and the following sample code will make TF froze with a 220% (2CPU) CPU usage and no output at all. Any idea why this happen? Thanks!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42126330","CreationDate":"2017-02-09T00:49:01.403","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>By default <code>Estimator.evaluate<\/code> assumes queue-based input, and will continue evaluating until the input pipeline is exhausted. When there is no queue-based input, this means it will loop forever. The fix is easy: simply provide a <code>steps<\/code> argument to <code>evaluate<\/code>.<\/p>\n"}]}
{"QuestionId":42125046,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-08T22:42:59.037","AcceptedAnswerId":44272990.0,"OwnerUserId":2910092.0,"Title":"How to append data to TensorFlow tfrecords file","Body":"<p>How to append new data (e.g. pairs of images and labels) to an already existing tfrecord file?<\/p>\n\n<p>The class <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/lib\/io\/tf_record.py\" rel=\"noreferrer\">tf.python_io.TFRecordWriter<\/a> does not seem to have any option for that.<\/p>\n\n<p>This question may also be reformulated to how to concatenate tfrecord files.<\/p>\n","answers":[{"AnswerId":"44272990","CreationDate":"2017-05-30T22:37:51.740","ParentId":null,"OwnerUserId":"2910092","Title":null,"Body":"<p>According to the comments in the ticket I opened this won't be implemented, soon.<\/p>\n"}]}
{"QuestionId":42125239,"AnswerCount":0,"Tags":"<neural-network><deep-learning><theano>","CreationDate":"2017-02-08T22:57:52.217","AcceptedAnswerId":null,"OwnerUserId":null,"Title":"Building a neural network chess engine","Body":"<p>I am building a single layer NN to play chess.<\/p>\n\n<p>The input is a matrix[n,m] of n training set, 64 features each where each feature represents a square.<\/p>\n\n<p>The training output is a matrix[n,1] where each output is a centipawn score obtained from stockfish. For example, the score can be 100, -200, 1405, etc etc.<\/p>\n\n<p>I want to build this NN that has 64 input nodes, 500 hidden nodes, and 1 output node for outputting a score.<\/p>\n\n<p>I know that NN is used for classification, but I was wondering if I can build a NN that can output arbitrary integers as well? <\/p>\n\n<p>Thanks <\/p>\n","answers":[]}
{"QuestionId":42125502,"AnswerCount":0,"Tags":"<tensorflow><visualization><tensorboard>","CreationDate":"2017-02-08T23:24:30.337","AcceptedAnswerId":null,"OwnerUserId":4888655.0,"Title":"tensorboard visualization site can not be reached","Body":"<p>I am following this website's tutorial to learn tensorboard. But it does not show anything in <a href=\"http:\/\/172.17.0.2:6006\" rel=\"nofollow noreferrer\">http:\/\/172.17.0.2:6006<\/a> I have no idea why this happens. I try to debug with this command line: . Info as following<\/p>\n\n<pre><\/pre>\n\n<p>Any idea how to deal with this problem? Thx <\/p>\n","answers":[]}
{"QuestionId":42125884,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-09T00:00:22.640","AcceptedAnswerId":null,"OwnerUserId":2391777.0,"Title":"one read_and_decode function for different training data","Body":"<p>I'm new to TensorFlow, and here's what I'm trying to do: save training data from different scenarios and then read them back. The sizes of feature and output may be different for different scenarios. <\/p>\n\n<p>The issue is when I tried to read the data back, I got an exception that looks like the following:<\/p>\n\n<pre><\/pre>\n\n<p>The function for saving data looks like below:<\/p>\n\n<pre><\/pre>\n\n<p>The function to read the data back is as follows:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42125955,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-02-09T00:06:16.717","AcceptedAnswerId":null,"OwnerUserId":5434466.0,"Title":"Tensorflow inference for modified Inception V3 model","Body":"<p>Went thru <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"nofollow noreferrer\">https:\/\/tensorflow.github.io\/serving\/serving_basic<\/a> and was able to run inference MNIST example using Tensorflow Serving server.<\/p>\n\n<p>Now, I would like to use a trained modified InceptionV3 model (which generates 2 files: .pb and .txt) to export and use it for inference.<\/p>\n\n<p>Serving Basic tutorial use mnist_saved_model.py for training and exporting the model. Is this file to be modified for the trained modified InceptionV3 model? Also, what is the difference between mnist_saved_model.py and mnist_export.py?<\/p>\n\n<p>Looked at <a href=\"https:\/\/stackoverflow.com\/questions\/37237940\/how-to-serve-the-tensorflow-graph-file-output-graph-pb-via-tensorflow-serving\">How to serve the Tensorflow graph file (output_graph.pb) via Tensorflow serving?<\/a> but the example mnist_saved_model.py creates a directory called 1 and subdirectories as shown below<\/p>\n\n<p>$> ls \/tmp\/mnist_model\/1<\/p>\n\n<p>saved_model.pb  variables<\/p>\n\n<p>Thoughts??<\/p>\n","answers":[]}
{"QuestionId":42127156,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-02-09T02:23:31.587","AcceptedAnswerId":null,"OwnerUserId":3810848.0,"Title":"tensorflow TF-slim inceptionv3 training loss curve is strange","Body":"<p>TF-slim inceptionv3 train from scratch<\/p>\n\n<p>I use slim\/train_image_classifier.py to train a inception_v3 model on my own dataset:\npython train_image_classifier.py --train_dir=${TRAIN_DIR} --dataset_name=mydataset --dataset_split_name=train --dataset_dir=${DATASET_DIR} --model_name=inception_v3 --num_clones=2<\/p>\n\n<p>the loss curve was strange as the tensorboard shown, it was a linearly decreasing straight line with a little bump part.\n<a href=\"https:\/\/i.stack.imgur.com\/NWnm4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NWnm4.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>the following are the last outputs, the loss was decreased 0.0001 every 20 or 30 steps\uff1a<\/p>\n\n<pre><\/pre>\n\n<p>mydataset.py is the same as flowers.py except:<\/p>\n\n<pre><\/pre>\n\n<p>Is it normal? Thanks for any help.<\/p>\n","answers":[{"AnswerId":"47542705","CreationDate":"2017-11-28T23:34:53.213","ParentId":null,"OwnerUserId":"1531462","Title":null,"Body":"<p>You are plotting the graph for total loss after n steps(which is probably number_of_steps if you are using tf.contrib.slim train method) in your training while the loss thats logged is for every 10 steps. \nHope this helps! <\/p>\n"}]}
{"QuestionId":42127447,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-09T02:56:16.233","AcceptedAnswerId":null,"OwnerUserId":1044011.0,"Title":"TensorFlow's Estimator can only get N-1 batches from tf.train.limit_epochs","Body":"<p>Hi this is a follow up question from <a href=\"https:\/\/stackoverflow.com\/questions\/42124976\/tensorflows-estimator-froze-with-low-cpu-usage\">TensorFlow&#39;s Estimator froze with low CPU usage<\/a>.<\/p>\n\n<p>The following code works fine if the  steps is , but if it is empty or , which should be the correct number of steps because there are four rows in  and  and the batch_size is 2, it will throw an  Error. I suppose the Estimator should catch this OutOfRange and use it to stop the evaluation but it does not and the program exits.<\/p>\n\n<pre><\/pre>\n\n<p>The error stack looks like this<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42127505,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-09T03:02:13.593","AcceptedAnswerId":42127784.0,"OwnerUserId":7537870.0,"Title":"Tensorflow dense_to_sparse","Body":"<p>I am trying to convert a uncompressed sparse array into a format accepted by tf.SparseTensor. There is an inbuilt function tf.sparse_to_dense that does exactly the opposite I am trying to do. So my question is there any inbuilt function in Tensorflow or Python to do this conversion?<\/p>\n","answers":[{"AnswerId":"51484756","CreationDate":"2018-07-23T17:59:39.257","ParentId":null,"OwnerUserId":"7930290","Title":null,"Body":"<p>tf.contrib.layers.dense_to_sparse does dense tensor to sparse conversion. termination is detected by presence of zeros in the end of array. Please visit <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/dense_to_sparse\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/dense_to_sparse<\/a> for more details.<\/p>\n"},{"AnswerId":"42127784","CreationDate":"2017-02-09T03:38:56.043","ParentId":null,"OwnerUserId":"3159795","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/questions\/39838234\/sparse-matrix-from-a-dense-one-tensorflow\">according to this question:<\/a> <\/p>\n\n<p>you can do it with this:<\/p>\n\n<p>You can use tf.where and tf.gather_nd to do that:<\/p>\n\n<pre><code>a = np.reshape(np.arange(24), (3, 4, 2))\nwith tf.Session() as sess:\n    a_t = tf.constant(a)\n    idx = tf.where(tf.not_equal(a_t, 0))\n    # Use tf.shape(a_t, out_type=tf.int64) instead of a_t.get_shape() if tensor shape is dynamic\n    sparse = tf.SparseTensor(idx, tf.gather_nd(a_t, idx), a_t.get_shape())\n    dense = tf.sparse_tensor_to_dense(sparse)\n    b = sess.run(dense)\nnp.all(a == b)\n&gt;&gt;&gt; True\n<\/code><\/pre>\n"}]}
{"QuestionId":42128580,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-09T04:58:13.003","AcceptedAnswerId":null,"OwnerUserId":1544674.0,"Title":"Tensorflow - single placeholder or list of them","Body":"<p>i'm working in a code using Tensorflow,  and i came to this problem. the neural network i have to model, has many inputs ( lets say something like a image 512x512 ) annd those inputs are connected to other neurons in the network ( in a feedforward architecture ) , i have to test many of these networks and in each case the connections between  de input neurons and the other neurons is different.  <\/p>\n\n<p>So far i have done this by creating a placeholder( one dimensional)  for each input, store them in a python list, and save in a graph structure the relationships between neurons, but i'm concerned about performance, considering that in each training try, i will have to create something like 100 or more networks. So i end up with 512x512x100 placeholder objects. i'm wondering if i'll have memory issues,  in that case, running the code in Amazon Web Services or Google Cloud may help? <\/p>\n\n<p>So i'm wondering if there is a better way to do it. maybe a way to create a single placeholder object for each network, and access the components in a \"array like\" way. but so far, i can't figure it out. <\/p>\n\n<p>Any advise in how to approach this, is welcome :) \nThanks. <\/p>\n","answers":[]}
{"QuestionId":42128605,"AnswerCount":1,"Tags":"<python><classification><probability><keras><predict>","CreationDate":"2017-02-09T05:00:42.620","AcceptedAnswerId":42129318.0,"OwnerUserId":7421340.0,"Title":"Keras classifier predict_proba() doesn't match predict()","Body":"<p>I'm working with Keras NN with Theanos backend, i'm working on a classification problem with 14 output classes. I want the predicted class plus the associated probabilities. The problem is the probabilities from predict_proba() don't appear to match the predicted class from predict(), here is the code plus the resulting output of 1 sample.<\/p>\n\n<pre><\/pre>\n\n<h1>probabilities<\/h1>\n\n<p>[0.00000, 0.00030, 0.02360, 0.04329, 0.00019, 0.00069, 0.00120, 0.00030, 0.00559, 0.00410, 0.00510, 0.91549, 0.0, 0.0]<\/p>\n\n<h1>predicted class<\/h1>\n\n<p>11<\/p>\n\n<h1>actual class<\/h1>\n\n<p>13<\/p>\n\n<p>It shows 12 having the highest probability from predict_proba() rather than 11 from predict(). Thank you for any help.<\/p>\n","answers":[{"AnswerId":"42129318","CreationDate":"2017-02-09T06:01:54.847","ParentId":null,"OwnerUserId":"1828289","Title":null,"Body":"<p>The index of python arrays (and here classes) counts up from 0, not from 1.  Have another look, 0.91 is the 12th value as people count things, but it's at index = 11 so predict and predict_proba are consistent <\/p>\n\n<p>As for why not 13, the prediction may be wrong (but check that you don't have the same kind of error there)<\/p>\n"}]}
{"QuestionId":42128830,"AnswerCount":3,"Tags":"<python><python-3.x><numpy><tensorflow><mnist>","CreationDate":"2017-02-09T05:21:23.950","AcceptedAnswerId":42444003.0,"OwnerUserId":6833276.0,"Title":"TypeError: only integer scalar arrays can be converted to a scalar index","Body":"<p>I am trying a simple demo code of tensorflow from <a href=\"https:\/\/github.com\/llSourcell\/tensorflow_demo\" rel=\"noreferrer\">github link<\/a>.<br>\nI'm currently using python version 3.5.2 <br><br><\/p>\n\n<pre><\/pre>\n\n<p>I ran into this error when I tried board.py in command-line. I have installed all the dependencies that are required for this to run.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42444003","CreationDate":"2017-02-24T16:46:54.677","ParentId":null,"OwnerUserId":"7618181","Title":null,"Body":"<p>you can modify the function:<\/p>\n\n<pre><code>def _read32(bytestream):\n    dt = numpy.dtype(numpy.uint32).newbyteorder('&gt;')\n    return numpy.frombuffer(bytestream.read(4), dtype=dt)\n<\/code><\/pre>\n\n<p>new version:<\/p>\n\n<pre><code>def _read32(bytestream):\n    dt = numpy.dtype(numpy.uint32).newbyteorder('&gt;')\n    return numpy.frombuffer(bytestream.read(4), dtype=dt)[0]\n<\/code><\/pre>\n\n<p>add <code>[0]<\/code> in the end.<\/p>\n\n<p>This appears to be an issue with the latest version of Numpy. A recent change made it an error to treat a single-element array as a scalar for the purposes of indexing.<\/p>\n"},{"AnswerId":"42139831","CreationDate":"2017-02-09T14:52:09.840","ParentId":null,"OwnerUserId":"7311767","Title":null,"Body":"<p>This file is likely corrupt:<\/p>\n\n<pre><code>Z:\/downloads\/MNIST dataset\\train-images-idx3-ubyte.gz\n<\/code><\/pre>\n\n<p>Let's analyze the error you posted.<\/p>\n\n<p>This, indicates that code is currently working with the file in question:<\/p>\n\n<pre><code>Extracting  Z:\/downloads\/MNIST dataset\\train-images-idx3-ubyte.gz\n<\/code><\/pre>\n\n<p><code>Traceback<\/code> indicates that a stack trace follows:<\/p>\n\n<pre><code>Traceback (most recent call last):\n<\/code><\/pre>\n\n<p>This, indicates that you read your data sets from <code>'Z:\/downloads\/MNIST dataset'<\/code>:<\/p>\n\n<pre><code>File \"board.py\", line 3, in &lt;module&gt;\n    mnist = input_data.read_data_sets(r'Z:\/downloads\/MNIST dataset', one_hot=True)\n<\/code><\/pre>\n\n<p>This, indicates that the code is extracting images:<\/p>\n\n<pre><code>File \"Z:\\downloads\\tensorflow_demo-master\\tensorflow_demo-master\\input_data.py\", line 150, in read_data_sets\n    train_images = extract_images(local_file)\n<\/code><\/pre>\n\n<p>This, indicates that the code is expected to read <code>rows * cols * num_images<\/code> bytes:<\/p>\n\n<pre><code>File \"Z:\\downloads\\tensorflow_demo-master\\tensorflow_demo-master\\input_data.py\", line 40, in extract_images\n    buf = bytestream.read(rows * cols * num_images)\n<\/code><\/pre>\n\n<p>This is the line that errors:<\/p>\n\n<pre><code>File \"C:\\Users\\surak\\AppData\\Local\\Programs\\Python\\Python35\\lib\\gzip.py\", line 274, in read\n    return self._buffer.read(size)\nTypeError: only integer scalar arrays can be converted to a scalar index\n<\/code><\/pre>\n\n<p>I expect <code>size<\/code> is the problematic value and was calculated on the previous line of the stacktrace.<\/p>\n\n<p>I can see at least two ways to proceed.<\/p>\n\n<ol>\n<li><p>Delete the offending file and see if the problem goes away.  This would allow you to verify that the file is somehow corrupt.<\/p><\/li>\n<li><p>Use a debugger to step into the code and then inspect the values used to calculate the offending variable.  Use the knowledge gained to proceed from there.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"42429141","CreationDate":"2017-02-24T01:20:11.007","ParentId":null,"OwnerUserId":"2348704","Title":null,"Body":"<p>The code link you have provided uses a separate file named <code>input_data.py<\/code> to download data from MNIST using the following two lines in <code>board.py<\/code><\/p>\n\n<pre><code>import input_data \nmnist = input_data.read_data_sets(\"\/tmp\/data\/\",one_hot=True)\n<\/code><\/pre>\n\n<p>Since MNIST data is so frequently used for demonstration purposes, Tensorflow provides a way to automatically download it.<\/p>\n\n<p>Replace the above two lines in <code>board.py<\/code> with the following two lines and the error will disappear.<\/p>\n\n<pre><code>from tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets('MNIST_data', one_hot=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":42130119,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2017-02-09T06:53:02.983","AcceptedAnswerId":null,"OwnerUserId":1279459.0,"Title":"What dataset are used for tests in Caffe?","Body":"<p>When I run,<\/p>\n\n<p> or <\/p>\n\n<p>the tests proceed even for the newly imported network architectures. What data is used for those tests? <\/p>\n\n<p>Update:<\/p>\n\n<p>Here is a snippet from the data layer from the Caffenet, where there is no reference to source. It comes from:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/deploy.prototxt\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_reference_caffenet\/deploy.prototxt<\/a><\/p>\n\n<pre><\/pre>\n\n<p>Update 2:<\/p>\n\n<p>But still the following command works and runs the test, despite the fact we are using :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42139105","CreationDate":"2017-02-09T14:19:35.527","ParentId":null,"OwnerUserId":"249226","Title":null,"Body":"<p>As @Shai suggested, you need to look at the input layer, as defined in the prototxt.<\/p>\n\n<p>Since you don't give any details about your models, I will use the interfaces tutorial as an example:\n<a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/interfaces.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/tutorial\/interfaces.html<\/a><\/p>\n\n<p>which has the following:<\/p>\n\n<pre><code>caffe test -model examples\/mnist\/lenet_train_test.prototxt -weights examples\/mnist\/lenet_iter_10000.caffemodel -gpu 0 -iterations 100\n<\/code><\/pre>\n\n<p>You can find this prototxt here:\nFor example:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/mnist\/lenet_train_test.prototxt<\/a><\/p>\n\n<p>The data source (that is included for testing) is given as:<\/p>\n\n<pre><code>source: \"examples\/mnist\/mnist_test_lmdb\"\n<\/code><\/pre>\n\n<p>This is the data that is used.<\/p>\n\n<p>You can read about how to download this data in the <code>examples\/mnist<\/code> directory:<\/p>\n\n<p><a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/examples\/mnist\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/tree\/master\/examples\/mnist<\/a><\/p>\n\n<p>Update: if the Data layer doesn't have a source parameter, then it is likely just used for \"deployment\". In this setup the data can be passed to the model live without the ground truth data. <code>caffe test<\/code> should not be called on this type of prototxt. Look for the training\/testing prototxt instead (most models come with both).<\/p>\n"}]}
{"QuestionId":42130293,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-09T07:05:25.180","AcceptedAnswerId":null,"OwnerUserId":4086817.0,"Title":"Inception retrain error : tensorflow.python.framework.errors_impl.FailedPreconditionError","Body":"<p>I am trying to retrain the iception model using my own classes.<\/p>\n\n<p><strong>What I am doing<\/strong><\/p>\n\n<p>I have directory containing images like <\/p>\n\n<ul>\n<li>Water Glass<\/li>\n<li>Water Bottle<\/li>\n<li>Person<\/li>\n<li>Chair<\/li>\n<li>.......<\/li>\n<\/ul>\n\n<p>I run following command <\/p>\n\n<pre><\/pre>\n\n<p><strong>Results<\/strong><\/p>\n\n<p>I got network trained with test accuracy of 94.6 % with an error.<\/p>\n\n<p><strong>Traceback<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>What I have tried<\/strong><\/p>\n\n<p>I have uninstall the tensorflow and installed again from sources, working fine. I have searched about the error, only thing I have understand is error I have encountered is due to some un initialization of variables. But from stacktrace I am not able to resolve the error.<\/p>\n","answers":[]}
{"QuestionId":42130491,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><lstm><recurrent-neural-network>","CreationDate":"2017-02-09T07:18:48.877","AcceptedAnswerId":null,"OwnerUserId":6195240.0,"Title":"Batch-major vs time-major LSTM","Body":"<p>Do RNNs learn different dependency patterns when the input is batch-major as opposed to time-major? <\/p>\n","answers":[{"AnswerId":"46544816","CreationDate":"2017-10-03T12:39:29.993","ParentId":null,"OwnerUserId":"507062","Title":null,"Body":"<p>There is no difference in what the model learns. <\/p>\n\n<p>At timestep t, RNNs need results from t-1, therefore we need to compute things time-major. If <code>time_major=False<\/code>, TensorFlow transposes batch of sequences from <code>(batch_size, max_sequence_length)<\/code> to <code>(max_sequence_length, batch_size)<\/code>*. It processes the transposed batch one row at a time: at t=0, the first element of each sequence is processed, hidden states and outputs calculated; at t=max_sequence_length, the last element of each sequence is processed.<\/p>\n\n<p>So if your data is already time-major, use <code>time_major=True<\/code>, which avoids a transpose. But there isn't much point in manually transposing your data before feeding it to TensorFlow.<\/p>\n\n<p>*If you have multidimensional inputs (e.g. sequences of word embeddings: <code>(batch_size, max_sequence_length, embedding_size)<\/code>), axes 0 and 1 are transposed, leading to <code>(max_sequence_length, batch_size, embedding_size)<\/code><\/p>\n"},{"AnswerId":"42191600","CreationDate":"2017-02-12T18:36:17.687","ParentId":null,"OwnerUserId":"4305018","Title":null,"Body":"<p>(Edit: sorry my initial argument was why it <em>makes sense<\/em> but I realized that it <em>doesn't<\/em> so this is a little OT.)<\/p>\n\n<p>I haven't found the TF-groups reasoning behind this but it <s>does<\/s> <strong>does not<\/strong> make computational sense as ops are written in C++.<\/p>\n\n<p>Intuitively, we want to mash up (multiply\/add etc) different features from the same sequence on the same timestep. Different timesteps can\u2019t be done in parallell while batch\/sequences can so feature>batch\/sequence>timestep.<\/p>\n\n<p>By <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/glossary.html#glossary\" rel=\"noreferrer\">default Numpy and C++ uses row-major<\/a> (C-like) memory layout so <\/p>\n\n<pre><code>[[ 0.  1.  2.]\n [ 3.  4.  5.]\n [ 6.  7.  8.]]\n<\/code><\/pre>\n\n<p>Is laying like <code>[0,1,2,3,4,5,6,7,8]<\/code> in memory. This means that if we have <\/p>\n\n<pre><code>x = np.zeros([time,batch,feature])\n<\/code><\/pre>\n\n<p>(<code>time_major=True<\/code> in tensorflow)<\/p>\n\n<p>In Row-major memory we get a layout like <code>x[0,0,0],x[0,0,1],x[0,0,2],\u2026,x[0,1,0],...<\/code> so ex. dot product of weights and vectors from the same sequence and timestep (<code>w*x[t,b,:]<\/code>) is the most contiguous operation followed by next sequence <code>w*x[t,b+1,:]<\/code> etc. This is what we want during training.<\/p>\n\n<p>With <code>time_major=False<\/code> which is default we have [batch,time,feature] so ex features from same sequence but different timesteps are more contiguous i.e <code>w*x[batch,t,:]<\/code> followed by <code>w*x[batch,t+1,:]<\/code> etc. This might be faster for prediction of one sequence at a time if RNN is rolled out but this is speculation.<\/p>\n\n<p>If you came to this question for the same reason I did, I learned to be careful with the slightly unintuitive Numpy-indexing which is meant to be pythonic, not necessarily Row Major. Look at this. As expected:<\/p>\n\n<pre><code>x = np.zeros([3,3])\nx[0:9].flat = np.arange(10)\nprint x\n&gt;   [[ 0.  1.  2.]\n&gt;    [ 3.  4.  5.]\n&gt;    [ 6.  7.  8.]]\n<\/code><\/pre>\n\n<p>We would also expect <code>x[1] == x[0,1]<\/code> but<\/p>\n\n<pre><code>print x[1]\n&gt; [ 3.  4.  5.]\n\nprint x[np.arange(10)&lt;=4]\n&gt; IndexError: index 3 is out of bounds for axis 0 with size 3\n<\/code><\/pre>\n"}]}
{"QuestionId":42130662,"AnswerCount":2,"Tags":"<python><opencv><import><keras>","CreationDate":"2017-02-09T07:29:17.090","AcceptedAnswerId":null,"OwnerUserId":7526688.0,"Title":"Importing Images based on a list of paths","Body":"<p>So I am trying to import some images.  All the images have their respective file paths in the first column of a .csv file. Below is the approach I have taken so far.  The idea is make a list of the paths then use opencv to \"imread\" the images to a list \"cv_img\"<\/p>\n\n<p>The problem at hand is only 1 image gets added to the list cv_img.  Any idea why this is happening?  <\/p>\n\n<p>Is this the best approach? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42184096","CreationDate":"2017-02-12T04:05:36.323","ParentId":null,"OwnerUserId":"7526688","Title":null,"Body":"<p>So I am trying a different approach and still having some issues.  The goal here is to have an array of the following dimensions.  (length, 160,329,3).  As you can see my reshape function is commented out. The \"print(images.shape) line returns (8037,).  Not sure how to proceed to get the right array dimensions.  For reference the 1st column in the csv file is a list of paths to the image in question.  This method also fixed the error I was getting before with pandas.  <\/p>\n\n<pre><code>import csv\nimport matplotlib.pyplot as plt\nf = open('\/Users\/username\/Desktop\/data\/driving_log.csv')\ncsv_f = csv.reader(f)\n\nm=[]\nfor row in csv_f:\n  n=(row)\n  m.append(n)\n\nimages=[]\nfor i in range(len(m)):\n    img=(m[i][1])\n    img=img.lstrip()\n    path='\/Users\/username\/Desktop\/data\/'\n    img=path+img\n    image=cv2.imread(img)\n    images.append(image)\nitem_num = len(images)\nimages=np.array(images)\n#images=np.array(images).reshape(item_num, 160, 320, 3)\nprint(images.shape)\n<\/code><\/pre>\n"},{"AnswerId":"42131084","CreationDate":"2017-02-09T07:58:10.940","ParentId":null,"OwnerUserId":"6885902","Title":null,"Body":"<p>I presume there is an error in the <code>for<\/code> loop. I think your <code>for<\/code> loop is just reading the first image in the list. <\/p>\n\n<p>Try modifying it to the following:<\/p>\n\n<pre><code>import cv2\nfor img in range(len(list1)):\n    n= cv2.imread(list1[img])\n    cv_img.append(n)\n<\/code><\/pre>\n"}]}
{"QuestionId":42130673,"AnswerCount":1,"Tags":"<c++><cmake><caffe>","CreationDate":"2017-02-09T07:29:57.110","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Proper way to link to Caffe library","Body":"<p>I am using <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow noreferrer\">caffe library<\/a> for my project.<\/p>\n\n<p>My  file is as follow<\/p>\n\n<pre><\/pre>\n\n<p>When I run  , I have error as<\/p>\n\n<pre><\/pre>\n\n<p>How can I solve the problem?<\/p>\n\n<p>Is there a better way to link to caffe library and its accessories?<\/p>\n","answers":[{"AnswerId":"42195541","CreationDate":"2017-02-13T01:55:21.620","ParentId":null,"OwnerUserId":"2467772","Title":null,"Body":"<p>For those who have difficulty to link to caffe library from your c++ project using cmake, the following is the CMakeLists.txt I could successfully used. Please change to the correct path according to your setting.<\/p>\n\n<pre><code>cmake_minimum_required(VERSION 2.6)\ncmake_policy(SET CMP0012 OLD)\n\nPROJECT(endtoendlib)\n\n\nif(ON)\n  if(NOT OpenCV_FOUND)\n    set(Caffe_OpenCV_CONFIG_PATH \"\/usr\/local\/share\/OpenCV\")\n    if(Caffe_OpenCV_CONFIG_PATH)\n      get_filename_component(Caffe_OpenCV_CONFIG_PATH ${Caffe_OpenCV_CONFIG_PATH} ABSOLUTE)\n\n      if(EXISTS ${Caffe_OpenCV_CONFIG_PATH} AND NOT TARGET opencv_core)\n        message(STATUS \"Caffe: using OpenCV config from ${Caffe_OpenCV_CONFIG_PATH}\")\n        include(${Caffe_OpenCV_CONFIG_PATH}\/OpenCVModules.cmake)\n      endif()\n\n    else()\n      find_package(OpenCV REQUIRED)\n    endif()\n    unset(Caffe_OpenCV_CONFIG_PATH)\n  endif()\nendif()\n\n# Compute paths\nget_filename_component(Caffe_CMAKE_DIR \"\/home\/Softwares\/Libraries\/caffe-master\/build\/cmake\" PATH)\nset(Caffe_INCLUDE_DIRS \"\/home\/Softwares\/Libraries\/caffe-master\/include\")\nlist(APPEND Caffe_INCLUDE_DIRS \"\/usr\/include\")\nlist(APPEND Caffe_INCLUDE_DIRS \"\/usr\/local\/cuda-8.0\/include\")\nlist(APPEND Caffe_INCLUDE_DIRS \"\/usr\/local\/include\/opencv\")\nlist(APPEND Caffe_INCLUDE_DIRS \"\/usr\/include\/atlas\")\nlist(APPEND Caffe_INCLUDE_DIRS \"\/home\/Softwares\/Libraries\/caffe-master\/build\/src\/\")\n\n# Our library dependencies\n#if(NOT TARGET caffe AND NOT caffe_BINARY_DIR)\ninclude(\"\/home\/Softwares\/Libraries\/caffe-master\/build\/ccmake\/CaffeTargets.cmake\")\n#endif()\n\n# List of IMPORTED libs created by CaffeTargets.cmake\nset(Caffe_LIBRARIES \"\/home\/Softwares\/Libraries\/caffe-master\/build\/lib\/libcaffe.so\")\n\n# Definitions\nset(Caffe_DEFINITIONS \"-DUSE_OPENCV;-DUSE_LMDB;-DUSE_LEVELDB\")\n\n# Cuda support variables\nset(Caffe_CPU_ONLY OFF)\nset(Caffe_HAVE_CUDA TRUE)\nset(Caffe_HAVE_CUDNN FALSE)\n\n\ninclude_directories( ${PROJECT_BINARY_DIR} ${Caffe_INCLUDE_DIRS})\n\n# adding sources to target\nadd_executable(endtoendlib endtoendLstm.cpp)\ntarget_link_libraries(endtoendlib ${OpenCV_LIBS} ${Caffe_LIBRARIES} -lglog -lprotobuf -lboost_system)\n<\/code><\/pre>\n\n<p>If you have error like follow, that means you didn't build caffe with <code>WITH_PYTHON_LAYER := 1 in Makefile.config<\/code><\/p>\n\n<pre><code>Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type: ImageData (known types: AbsVal, Accuracy, ArgMax, BNLL, BatchNorm, BatchReindex, Bias, Concat, ContrastiveLoss, Convolution, Crop, Data, Deconvolution, Dropout, DummyData, ELU, Eltwise, Embed, EuclideanLoss, Exp, Filter, Flatten, HDF5Data, HDF5Output, HingeLoss, Im2col, InfogainLoss, InnerProduct, Input, LRN, LSTM, LSTMUnit, Log, MVN, MemoryData, MultinomialLogisticLoss, PReLU, Parameter, Pooling, Power, RNN, ReLU, Reduction, Reshape, SPP, Scale, Sigmoid, SigmoidCrossEntropyLoss, Silence, Slice, Softmax, SoftmaxWithLoss, Split, TanH, Threshold, Tile)\n<\/code><\/pre>\n"}]}
{"QuestionId":42131517,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-09T08:27:16.113","AcceptedAnswerId":42250107.0,"OwnerUserId":4174853.0,"Title":"Tensorflow TFLearn working example","Body":"<p>I'm approaching to the world of deep learning, and the framework that I'm using is Tensorflow. In order to start quickly, I've seen that there are high level api called TFLearn, wich makes the creation of a network a lot easier. Unfortunaly, there are no working examples. In particular, I'm trying this example:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/examples\/extending_tensorflow\/layers.py\" rel=\"nofollow noreferrer\">Example<\/a><\/p>\n\n<p>but it can't find the layers. I've tried to import tflearn in this way:<\/p>\n\n<pre><\/pre>\n\n<p>and the dataset in this way<\/p>\n\n<pre><\/pre>\n\n<p>and they works, but I'm still not able to use the layers.<\/p>\n\n<p><strong><em>EDIT<\/em><\/strong><\/p>\n\n<p>I also tried to import in this way (as reported on tflearn github page):<\/p>\n\n<pre><\/pre>\n\n<p>and I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>What can I do?<\/p>\n","answers":[{"AnswerId":"42147744","CreationDate":"2017-02-09T22:03:54.843","ParentId":null,"OwnerUserId":"7402023","Title":null,"Body":"<p>I also meet some strange errors with TFlearn... it seems it is not properly maintained and updated, so now it is not in sync with Tensorflow...<\/p>\n"},{"AnswerId":"42250107","CreationDate":"2017-02-15T12:59:09.087","ParentId":null,"OwnerUserId":"7096238","Title":null,"Body":"<p>Don't use <\/p>\n\n<pre><code>import tensorflow.contrib.learn as tflearn\n<\/code><\/pre>\n\n<p>Instead install tflearn in your terminal, using installation guide <a href=\"http:\/\/tflearn.org\/installation\/\" rel=\"nofollow noreferrer\">http:\/\/tflearn.org\/installation\/<\/a><\/p>\n\n<p>Then, import like:<\/p>\n\n<pre><code>import tflearn\nfrom tflearn.layers.core import input_data, dropout, fully_connected\nfrom tflearn.layers.conv import conv_2d, max_pool_2d\nfrom tflearn.layers.estimator import regression\n<\/code><\/pre>\n\n<p>This way you can use the layers.<\/p>\n"}]}
{"QuestionId":42131532,"AnswerCount":1,"Tags":"<tensorflow><distributed>","CreationDate":"2017-02-09T08:27:57.770","AcceptedAnswerId":42138402.0,"OwnerUserId":4394807.0,"Title":"Distributed Tensorflow: TypeError: __init__() got multiple values for keyword argument 'dtype'","Body":"<p>When running distributed tensorflow following the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">inception distributed training tutorial<\/a>, I encountered the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I think the problem is due to api mismatch somewhere in . The tensorflow version I used is . Could anyone help me on this? Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42138402","CreationDate":"2017-02-09T13:46:08.383","ParentId":null,"OwnerUserId":"4394807","Title":null,"Body":"<p>Problem solved. It may be due to api mismatch. The solution is to change line 243 in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/slim\/variables.py\" rel=\"nofollow noreferrer\">inception\/slim\/variables.py<\/a> to <code>initializer=tf.zeros_initializer()<\/code>.<\/p>\n\n<p>Please tell me if there is anything wrong.<\/p>\n"}]}
{"QuestionId":42131681,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2017-02-09T08:35:16.490","AcceptedAnswerId":null,"OwnerUserId":6497293.0,"Title":"Theano installation:Error:error: '::hypot' has not been declared using ::hypot;","Body":"<p>I cloned theano from Git and just ran from the directory python setup.py install.\nBut I am  getting the following error.\n'error: '::hypot' has not been declared\n   using ::hypot;'\nI am using windows 8.1 with Python 3.5.\nPlease  anybody help.<\/p>\n","answers":[]}
{"QuestionId":42132086,"AnswerCount":1,"Tags":"<deep-learning><theano><conv-neural-network><image-recognition>","CreationDate":"2017-02-09T08:55:49.097","AcceptedAnswerId":null,"OwnerUserId":7105352.0,"Title":"Pickle multy layer CNN in Theano, Lenet5","Body":"<p>I am new to deep learning, and I got a problem. \nI am using Theano for image recognition, and I want to create a prediction system using trained model.\nI referenced LeNet5 <a href=\"http:\/\/deeplearning.net\/tutorial\/lenet.html\" rel=\"nofollow noreferrer\">Convolutional Neural Networks (LeNet)<\/a> and trained my own data, now I want to use the trained model for predicting new images.\nIn <a href=\"http:\/\/deeplearning.net\/tutorial\/logreg.html\" rel=\"nofollow noreferrer\">Classifying MNIST digits using Logistic Regression<\/a> it described the way to pickle trained models, but it is just a logistic regression, not a multilayered CNN. In the same way I saved each layer, but I cant use it for prediction.\nPlease help me! \nHere is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42149588","CreationDate":"2017-02-10T00:40:45.347","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>There are many ways to save your model. The one I often to use is by pickle the weight and bias of each layer (the order is up to you):<\/p>\n\n<pre><code>f = file('Models\/bestmodel.pickle','wb')\ncPickle.dump(layer0.W.get_value(borrow=True),f,protocol=cPickle.HIGHEST_PROTOCOL)\ncPickle.dump(layer1.W.get_value(borrow=True),f,protocol=cPickle.HIGHEST_PROTOCOL)\ncPickle.dump(layer2.W.get_value(borrow=True),f,protocol=cPickle.HIGHEST_PROTOCOL)\n...\ncPickle.dump(layer0.b.get_value(borrow=True),f,protocol=cPickle.HIGHEST_PROTOCOL)            \ncPickle.dump(layer1.b.get_value(borrow=True),f,protocol=cPickle.HIGHEST_PROTOCOL)\ncPickle.dump(layer2.b.get_value(borrow=True),f,protocol=cPickle.HIGHEST_PROTOCOL)\n...\nf.close()\n<\/code><\/pre>\n\n<p>then for prediction system, create a same model architecture and use saved model as initial value (with the same order as you saved):<\/p>\n\n<pre><code>f=file('Models\/bestmodel.pickle','rb')\nlayer0.W.set_value(cPickle.load(f), borrow=True)\nlayer1.W.set_value(cPickle.load(f), borrow=True)\nlayer2.W.set_value(cPickle.load(f), borrow=True)\n...\nlayer0.b.set_value(cPickle.load(f), borrow=True)\nlayer1.b.set_value(cPickle.load(f), borrow=True)\nlayer2.b.set_value(cPickle.load(f), borrow=True)\n...\nf.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":42133101,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2017-02-09T09:42:00.467","AcceptedAnswerId":null,"OwnerUserId":3305212.0,"Title":"How to improve neural network accuracy for a random X array where the y is modulus(X,2)?","Body":"<p>I am learning about neural networks and I found the following code which works well with the iris dataset reaching around 99% accuracy.<\/p>\n\n<p>However when I try to run the same network on a dummy dataset randomly generated using  where there is correlation between the input and the output as follows, it doesn't work so well. Would anyone be able to shed a light why the neural network struggles with this type of training input?<\/p>\n\n<pre><\/pre>\n\n<p>Here is the output from the program:<\/p>\n\n<pre><\/pre>\n\n<p>For comparison, here is the output from the iris dataset:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42139040","CreationDate":"2017-02-09T14:16:47.723","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>A network that works with one dataset doesn't mean that will work with a different dataset. Each network is designed with a specific learning capacity (approximately the number of parameters in the network) to match a given task or dataset.<\/p>\n\n<p>If the learning capacity in the model is too low, then the model will underfit, and this is what you are seeting. Too much learning capacity and the model will overfit.<\/p>\n\n<p>You need to adjust the learning capacity of the model (by adding or removing layers). In this case it seems you are underfitting so adding layers would help.<\/p>\n"}]}
{"QuestionId":42133165,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-02-09T09:44:53.790","AcceptedAnswerId":42169898.0,"OwnerUserId":7539391.0,"Title":"How to set thread number when running tensorflow sample on Android?","Body":"<p>As we know, the android demo of tensorflow is multi-thread. So how can I change it to the single-thread mode?<\/p>\n","answers":[{"AnswerId":"42169898","CreationDate":"2017-02-10T22:49:11.550","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>There's no way to do this from Java right now. A value for intra-op parallelism is automatically picked in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/common_runtime\/local_device.cc#L38\" rel=\"nofollow noreferrer\">local_device.cc<\/a> matching the number of cores detected on your device. If you just want to test, you could overwrite the value there.<\/p>\n\n<p>Inter-op parallelism is currently always 1 on mobile -- only one op runs at a time.<\/p>\n\n<p>More generally in native code intra and inter op parallism can be set when creating your Session object. e.g.:<\/p>\n\n<pre><code>tensorflow::SessionOptions options;\ntensorflow::ConfigProto&amp; config = options.config;\nconfig.set_intra_op_parallelism_threads(num_threads);\nSession session = tensorflow::NewSession(options); \n<\/code><\/pre>\n\n<p>This will override the automatically picked value. See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/benchmark\/benchmark_model.cc#L56\" rel=\"nofollow noreferrer\">benchmark_model.cc<\/a> for an example.<\/p>\n\n<p>If you want to set this from Java, you would have to add a parameter to  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/android\/jni\/tensorflow_inference_jni.cc#L100\" rel=\"nofollow noreferrer\">TensorFlowInferenceInterface.initializeTensorFlow()<\/a>.<\/p>\n\n<p>We're switching to the Java API soon, though, so it will be easier to access the SessionOptions to set the value then.<\/p>\n"}]}
{"QuestionId":42133661,"AnswerCount":1,"Tags":"<tensorflow><lstm><recurrent-neural-network>","CreationDate":"2017-02-09T10:06:30.433","AcceptedAnswerId":null,"OwnerUserId":2061800.0,"Title":"Tensorflow - LSTM state reuse within batch","Body":"<p>I am working on a Tensorflow NN which uses an LSTM to track a parameter (time series data regression problem). A batch of training data contains a batch_size of <em>consecutive<\/em> observations. I would like to use the LSTM state as input to the next sample. So, if I have a batch of data observations, I would like to feed the state of the first observation as input to the second observation and so on. Below I define the lstm state as a tensor of size = batch_size. I would like to reuse the state <em>within<\/em> a batch:<\/p>\n\n<pre><\/pre>\n\n<p>In the API there is a tf.nn.state_saving_rnn but the documentation is kinda vague. <strong>My question<\/strong>: How to reuse curr_state <em>within<\/em> a training batch.<\/p>\n","answers":[{"AnswerId":"45482821","CreationDate":"2017-08-03T11:25:17.370","ParentId":null,"OwnerUserId":"1782792","Title":null,"Body":"<p>You are basically there, just need to update <code>state<\/code> with <code>curr_state<\/code>:<\/p>\n\n<pre><code>state_update = tf.assign(state, curr_state)\n<\/code><\/pre>\n\n<p>Then, make sure you either call <code>run<\/code> on <code>state_update<\/code> itself or an operation that has <code>state_update<\/code> as a dependency, or the assignment will not actually happen. For example:<\/p>\n\n<pre><code>with tf.control_dependencies([state_update]):\n    model_output = ...\n<\/code><\/pre>\n\n<p>As suggested in the comments, the typical case for RNNs is that you have a batch where the first dimension (0) is the number of sequences and the second dimension (1) is the maximum length of each sequence (if you pass <code>time_major=True<\/code> when you build the RNN these two are swapped). Ideally, in order to get good performance, you stack multiple sequences into one batch, and then split that batch time-wise. But that's all a different topic really.<\/p>\n"}]}
{"QuestionId":42134360,"AnswerCount":8,"Tags":"<tensorflow>","CreationDate":"2017-02-09T10:36:21.010","AcceptedAnswerId":42135964.0,"OwnerUserId":7539585.0,"Title":"Tensorflow - ValueError: Parent directory of trained_variables.ckpt doesn't exist, can't save","Body":"<p>I want to save my tensorflow session  but i have the following error<\/p>\n\n<p><\/p>\n\n<p>This is my line of code :<\/p>\n\n<p><\/p>\n\n<p>I've also tried to change the file name and put  instead of  but i get the same problem.<\/p>\n\n<p>Following this tutorial <a href=\"http:\/\/jrmeyer.github.io\/tutorial\/2016\/02\/01\/TensorFlow-Tutorial.html\" rel=\"noreferrer\">A TensorFlow Tutorial: Email Classification<\/a><\/p>\n","answers":[{"AnswerId":"58592141","CreationDate":"2019-10-28T13:50:03.020","ParentId":null,"OwnerUserId":"5079812","Title":null,"Body":"<p>For Linux: <\/p>\n\n<p>saver.save(sess, \"trained_variables.ckpt\")<\/p>\n\n<p>For Windows:<\/p>\n\n<p>saver.save(sess, \".\/trained_variables.ckpt\")<\/p>\n"},{"AnswerId":"48015268","CreationDate":"2017-12-28T22:40:20.937","ParentId":null,"OwnerUserId":"1284607","Title":null,"Body":"<p>save_path should always be a path instead of a filename here. I wouldn't call it a bug, but it is better that TensorFlow could add some code to auto convert a filename \"abc\" to \".\/abc\" to make it more convenient. <\/p>\n\n<p>In <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/saver.py\" rel=\"nofollow noreferrer\">tensorflow\/python\/training\/saver.py<\/a>:<\/p>\n\n<p>Line:<\/p>\n\n<pre><code>save_path_parent = os.path.dirname(save_path) \n<\/code><\/pre>\n\n<p>will try to get the pathname, if you feed in only a file name without a path, <em>save_path_parent<\/em> will become empty and fail the later test: <\/p>\n\n<pre><code>gfile.IsDirectory(save_path_parent)\n<\/code><\/pre>\n\n<p>So you have to always provide a path + filename when using saver.save() method. Using only file name will leads to this error.<\/p>\n"},{"AnswerId":"49193580","CreationDate":"2018-03-09T12:12:35.497","ParentId":null,"OwnerUserId":"9467481","Title":null,"Body":"<p>You can use your current working directory:<\/p>\n\n<pre><code>import os\u200b\u200b\ncheckpoint = os.path.join(os.getcwd(), 'chatbot_weights.ckpt')\u200b\n<\/code><\/pre>\n"},{"AnswerId":"48554312","CreationDate":"2018-02-01T02:08:13.757","ParentId":null,"OwnerUserId":"2506738","Title":null,"Body":"<p>Here is what I did<\/p>\n\n<pre><code> modelName = \"YourModelName\"\n saver.save(session, '.\/'+modelName) \n<\/code><\/pre>\n\n<p>I believe this should work for you.<\/p>\n"},{"AnswerId":"42136420","CreationDate":"2017-02-09T12:12:13.273","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>I'm not sure I understand. You are trying to store the variables into the current directory, so it's weird that it's saying the parent of the current directory does not exist. Can you try this<\/p>\n\n<pre><code>v = tf.Variable(tf.constant(0))\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    saver = tf.train.Saver()\n    saver.save(sess, 'trained_variables.ckpt')\n<\/code><\/pre>\n\n<p>This worked for me. Generally, you should always make sure that the path to where you want to store the checkpoints exist. Something like this:<\/p>\n\n<pre><code>save_path = 'checkpoints\/'\nmodel_name = 'my_model'\nif not os.path.exists(save_path):\n    os.makedirs(save_path)\n...\nsaver = tf.train.Saver()\nsave_path_full = os.path.join(save_path, model_name)\nwith tf.Session() as sess:\n   ...\n   saver.save(sess, save_path_full)\n<\/code><\/pre>\n"},{"AnswerId":"42135964","CreationDate":"2017-02-09T11:49:50.327","ParentId":null,"OwnerUserId":"2790546","Title":null,"Body":"<p>I would guess that you're trying to save the file in a folder (directory) that doesn't exist...<\/p>\n\n<p>Try using an absolute path for the file instead of just the bare filename. <\/p>\n\n<p>You might want to check what your current working directory is... that could clear up things.<\/p>\n\n<p>Does that help?<\/p>\n\n<p>-josh<\/p>\n"},{"AnswerId":"43153826","CreationDate":"2017-04-01T06:58:19.640","ParentId":null,"OwnerUserId":"3027221","Title":null,"Body":"<pre><code>saver.save(sess, \".\/trained_variables.ckpt\")\n<\/code><\/pre>\n"},{"AnswerId":"42138084","CreationDate":"2017-02-09T13:32:37.457","ParentId":null,"OwnerUserId":"7539585","Title":null,"Body":"<p>I've put the absoluth path of the file instead of the bare filename and it worked.\nThis is the final code <\/p>\n\n<p><code>saver.save(sess, os.path.join(os.getcwd(), 'trained_variables2.ckpt'))<\/code><\/p>\n"}]}
{"QuestionId":42135056,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-09T11:07:56.097","AcceptedAnswerId":null,"OwnerUserId":504634.0,"Title":"How to get only nonzero values from Sparse Tensor","Body":"<p>Utilizing TensorFlow's HashTable lookup implementation i get my SparseTensor back with the default value supplied.  I'd like to clean that off and get a final SparseTensor without the default value.<\/p>\n\n<p>How can I clean that default value? I don't mind what the default value will be in order to make this happen.  0 is fine and so is -1.<\/p>\n","answers":[{"AnswerId":"42143842","CreationDate":"2017-02-09T18:04:16.160","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p><code>tf.sparse_retain<\/code> should work:<\/p>\n\n<pre><code>def sparse_remove(sparse_tensor, remove_value=0.):\n  return tf.sparse_retain(sparse_tensor, tf.not_equal(a.values, remove_value))\n<\/code><\/pre>\n\n<p>As an example:<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.SparseTensor(indices=[[1, 2], [2, 2]], values=[0., 1.], shape=[3, 3])\nwith tf.Session() as session:\n  print(session.run([a, sparse_remove(a)]))\n<\/code><\/pre>\n\n<p>Prints (I've reformatted it slightly):<\/p>\n\n<pre><code>[SparseTensorValue(indices=array([[1, 2], [2, 2]]), values=array([ 0.,  1.], dtype=float32), shape=array([3, 3])), \n SparseTensorValue(indices=array([[2, 2]]), values=array([ 1.], dtype=float32), shape=array([3, 3]))]\n<\/code><\/pre>\n"}]}
{"QuestionId":42137112,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-09T12:45:36.197","AcceptedAnswerId":null,"OwnerUserId":482601.0,"Title":"Process data in batches to prevent Out Of Memory exceptions","Body":"<p>In my TensorFlow network, I calculate<\/p>\n\n<pre><\/pre>\n\n<p>where: <\/p>\n\n<ul>\n<li> is ,<\/li>\n<li> is ,<\/li>\n<li> is ,<\/li>\n<li> is ,<\/li>\n<li> is .<\/li>\n<\/ul>\n\n<p>The problem is that  and , and so  does not fit into memory, but , so  would fit into memory. I don't really need to store .<\/p>\n\n<p> could be calculated in batches, i.e., <\/p>\n\n<pre><\/pre>\n\n<p>Is there a neat way to tell TensorFlow to do this? Or what is a standard procedure here?<\/p>\n","answers":[{"AnswerId":"42163219","CreationDate":"2017-02-10T15:39:01.517","ParentId":null,"OwnerUserId":"482601","Title":null,"Body":"<p>The following seems to work but has the undesired side effect of creating a variable!<\/p>\n\n<p>Any ideas on how to improve on this?<\/p>\n\n<pre><code>n, d, L = 40, 2, 3\nnum_batches = 10\nrows_per_batch = n \/\/ num_batches\n\nX = tf.constant(np.arange(n*d, dtype=np.float32).reshape(n, d))\nC = tf.constant(np.arange(d*L, dtype=np.float32).reshape(d, L))\nR = tf.constant(np.arange(L*d, dtype=np.float32).reshape(L, d))\n\nY = tf.matmul(tf.matmul(X, C), R)\n\nyvar = tf.Variable(tf.constant(0, shape=Y.get_shape(), dtype=tf.float32))\n\ndef condition(current_batch):\n    return current_batch &lt; num_batches\n\ndef body(current_batch):\n    current_rows_to_update = tf.range(rows_per_batch * current_batch,\n                                      rows_per_batch * (current_batch + 1))\n    xbatch = tf.gather(X, current_rows_to_update)\n    ybatch = tf.matmul(tf.matmul(xbatch, C), R, name='matmul_da')\n\n    yvar_update = tf.scatter_update(yvar, current_rows_to_update, ybatch)\n\n    with tf.control_dependencies([yvar_update]):\n        return current_batch + 1\n\nloop = tf.while_loop(condition, body, loop_vars=[0])\nwith tf.control_dependencies([loop]):\n    ybatched = tf.identity(yvar)\n<\/code><\/pre>\n"}]}
{"QuestionId":42138462,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-09T13:48:57.697","AcceptedAnswerId":null,"OwnerUserId":833530.0,"Title":"Accuracy drops after changing training data with one iteration","Body":"<p>I have been working on using LeNet in tensorflow to train and classify the german traffic signs. I have modified the LeNet first and final layers to accept both the 1 and 3 channel color images (layer 1) and the number of classes to 43\n(layer 6). <\/p>\n\n<pre><\/pre>\n\n<p>Since the network is configured to accept both 1 and 3 channel images, (with the  argument, I am trying to use various preprocessing methods (grayscale conversion, normalization between (0,1) and scaling with [-0.5,0.5]) on the input training images and trying to evaluate the accuracy in each step. I have 6 kinds of processed data<\/p>\n\n<ol>\n<li>Original RGB image<\/li>\n<li>Converting to grayscale<\/li>\n<li>Normalization of grayscale image between [0,1]<\/li>\n<li>Scaling of grayscale image with zero mean and unit variance [-0.5,0.5]<\/li>\n<li>Normalization on RGB image between [0,1]<\/li>\n<li>Scaling on RGB with zero mean and unit variance [-0.5,0.5] <\/li>\n<\/ol>\n\n<p>I wanted to create a pipeline in a loop that takes one type of preprocessed data in one iteration and performs training and validation. My code as follows<\/p>\n\n<pre><\/pre>\n\n<p>the input data is a list of tuples where  in each tuple, elem[0] represents the name, elem[1] represents the training set and elem[2] represents the validation set. Now my pipeline is as follows<\/p>\n\n<pre><\/pre>\n\n<p>if I try to train the network with the input data without any preprocessing, the accuracy ranges between 80-90%. However keeping the network under a loop shows strange behavior of dropper accuracy as follows <\/p>\n\n<pre><\/pre>\n\n<p>Any idea where I am making a mistake?<\/p>\n","answers":[{"AnswerId":"46228903","CreationDate":"2017-09-14T22:03:43.983","ParentId":null,"OwnerUserId":"5376867","Title":null,"Body":"<p>One reason is maybe that data is not shuffled properly. So the network is not properly training on the images and doesn't matter which preprocessing techniques you're working with.<\/p>\n"}]}
{"QuestionId":42138926,"AnswerCount":0,"Tags":"<tensorflow><distribute>","CreationDate":"2017-02-09T14:11:40.650","AcceptedAnswerId":null,"OwnerUserId":4394807.0,"Title":"Distributed TensorFlow: error occurred when assigning ps and worker to the same machine","Body":"<p>I am running distributed tensorflow on 2 machines:  and  following the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">inception distributed training tutorial<\/a>. I found if I use one machine as ps and the other as worker, it runs well. The scripts are as follows: <\/p>\n\n<pre><\/pre>\n\n<p>However, if I run a worker on each machine, i.e. 2 workers in all and run ps on one of those 2 machines, the program crashes. The scripts are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The errors occurred on  are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I also try to use 1 worker and 1 ps on only one machine on different ports, but the error remains. <\/p>\n\n<p>So my question is: does that mean ps and worker must be assigned to different machines, even they are assigned to different ports? The reason for using one machine to be worker and ps simultaneously is to maximize the utilization of computing resources. Could anyone show me how to use one machine to be both worker and ps?<\/p>\n","answers":[]}
{"QuestionId":42139446,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-09T14:35:38.310","AcceptedAnswerId":42140230.0,"OwnerUserId":5350430.0,"Title":"TensorArray TensorArray_1_0: Could not read from TensorArray index 0 because it has not yet been written to","Body":"<p>I don't know how to use tensorarray. Here's the code. What's the bug in that?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42140230","CreationDate":"2017-02-09T15:10:14.720","ParentId":null,"OwnerUserId":"5350430","Title":null,"Body":"<p>I solved. It seems that inside the <code>body<\/code> of <code>while_loop<\/code> we should reassign the old <code>TensorArray<\/code> <code>aO<\/code> with the returned value of <code>aO.write()<\/code>:<\/p>\n\n<pre><code>def body(i,aI,aO):\n    aO=aO.write(i, aI.read(i)+1)\n    return (i+1, aI, aO)\n<\/code><\/pre>\n\n<p>Whole code:<\/p>\n\n<pre><code>import tensorflow as tf\n\naI=tf.TensorArray(tf.int32, 2)\naO=tf.TensorArray(tf.int32, 2)\naI=aI.unpack([[1,2],[1,2]])\ndef body(i,aI,aO):\n    aO=aO.write(i, aI.read(i)+1)\n    return (i+1, aI, aO)\ncond=lambda i, *_ : i&lt;2\n_, _, aO=tf.while_loop(cond, body, [0,aI,aO])\nr=aO.pack()\nwith tf.Session() as sess:\n    res=sess.run(r)\n    print('done!')\n<\/code><\/pre>\n"}]}
{"QuestionId":42140211,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-09T15:09:21.300","AcceptedAnswerId":42261893.0,"OwnerUserId":1157751.0,"Title":"Tensorflow - eval() error: You must feed a value for placeholder tensor","Body":"<p>I'm trying to use eval() to understand what is happening in each learning step.<\/p>\n\n<p>However, if I use eval() on an tf.matmul operation, then I would get an error .<\/p>\n\n<p>If I removed the eval(), then everything would work properly as expected. <\/p>\n\n<pre><\/pre>\n\n<p>The exact error is:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone have a better way to log the variables? I've tried tensor_summary, but it doesn't show it on the website.<\/p>\n\n<p>Thanks all<\/p>\n","answers":[{"AnswerId":"42261893","CreationDate":"2017-02-15T23:13:51.097","ParentId":null,"OwnerUserId":"1157751","Title":null,"Body":"<p>Apart from AllenLavoie's comment, you can actually feed the dictionary through eval.<\/p>\n\n<pre><code>loss.eval(feed_dict=feed_dict)\n<\/code><\/pre>\n\n<p>TensorFlow's weird API does not know that I've already fed the dictionary beforehand.<\/p>\n\n<p>Hence: <code>_, l, predictions, summary = session.run([optimizer, loss, train_prediction, summary_op], feed_dict=feed_dict)<\/code><\/p>\n\n<p>Does not work even though it is placed <strong>before<\/strong> <code>loss.eval()<\/code><\/p>\n"}]}
{"QuestionId":42140284,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-09T15:12:22.833","AcceptedAnswerId":null,"OwnerUserId":1660762.0,"Title":"Tensorflow data readers: readning the same files repeatedly until training ends","Body":"<p>I've created a model in tensorflow using <a href=\"https:\/\/www.tensorflow.org\/how_tos\/reading_data\/\" rel=\"nofollow noreferrer\">this<\/a> tutorial, on an example data set I've created. The model feeds the data in the input files until they end and then stops.<\/p>\n\n<p>This is how I read the files:<\/p>\n\n<pre><\/pre>\n\n<p>This is the session.run call:<\/p>\n\n<pre><\/pre>\n\n<p>The inner loop stops when the files end.<\/p>\n\n<p><strong>I want the model to continue until  is reached, repeatedly reading the same files as necessary.<\/strong><\/p>\n\n<p>I know this can be done manually, by passing the data to the session in a , but I'm hoping there's an easier way that allows me to still use the built-in file readers.<\/p>\n","answers":[]}
{"QuestionId":42140922,"AnswerCount":1,"Tags":"<numpy><neural-network><keras><lstm>","CreationDate":"2017-02-09T15:42:04.627","AcceptedAnswerId":null,"OwnerUserId":3682697.0,"Title":"Keras LSTM training data format","Body":"<p>I am trying to use LSTM neural network  (using Keras) to predict opponent's next move in the game Rock-Paper-Scissor. <\/p>\n\n<p>I have encode the inputs as Rock: [1 0 0], Paper: [0 1 0], Scissor: [0 0 1]. Now I want to train the neural network but I am a bit confused of the data structure of my training data. <\/p>\n\n<p>I have stored an opponent's game history in a .csv file with the following structure:<\/p>\n\n<pre><\/pre>\n\n<p>And I am trying to use every 5th data as my training label, and the previous 4 data as the training input. In another word, at each time step, a vector with dimension 3 is sent to the network, and we have 4 time steps. <\/p>\n\n<p>For example, the following is the input data<\/p>\n\n<pre><\/pre>\n\n<p>And the fifth one is the training label<\/p>\n\n<pre><\/pre>\n\n<p>My question is what type of data format does Keras' LSTM network accept? What would be an optimum way to rearrange my data for this purpose? My incomplete code is attached as following if it helps:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42141725","CreationDate":"2017-02-09T16:19:00.217","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>The input format for the LSTM should have a shape (sequence_length, input_dim).\nSo in your case, numpy arrays of shape (4,3) should do it. <\/p>\n\n<p>What you will feed to the model will then be a numpy array of shape (number_of_train_examples, sequence_length,  input_dim).\nIn other words, you will feed number_of_train_examples tables of shape (4,3). \nBuild a list of :<\/p>\n\n<pre><code>1,0,0\n0,1,0\n0,1,0\n0,0,1\n<\/code><\/pre>\n\n<p>and then do np.array(list_of_train_example).<\/p>\n\n<p>However, I don't understand why you return the whole sequence for the second LSTM? It will output you something with the shape (4,4), the Dense layer will probably fail on that. Return sequence means that you will return the whole sequence, so every hidden output at each step of LSTM. I would set this to False for the second LSTM to only get a \"summary\" vector of shape (4,) that your Dense layer can read.\nAnyway, even for the first LSTM it means that with an input of shape (4,3), you output something which has shape (4,4), so you will have more parameters than input data for this layer... Can't be really good.<\/p>\n\n<p>Regarding the activations, I would also use softmax but only on the last layer, softmax is used to get probabilities as output of the layer. It doesn't make really sense to use a softmax out of LSTM's and the Dense before the last. Go for some other non linearity like \"sigmoid\" or \"tanh\".<\/p>\n\n<p>This is what I would do model-wise<\/p>\n\n<pre><code>def createNet(summary=False):\n    print(\"Start Initialzing Neural Network!\")\n    model = Sequential()\n    model.add(LSTM(4,input_dim=input_dim,input_length=input_length,\n            return_sequences=True,activation='tanh'))\n    model.add(Dropout(0.1))\n    # output shape : (4,4)\n    model.add(LSTM(4,\n            return_sequences=False,activation='tanh'))\n    model.add(Dropout(0.1))\n    # output shape : (4,)\n    model.add(Dense(3,activation='tanh'))\n    model.add(Dropout(0.1))\n    # output shape : (3,)\n    model.add(Dense(3,activation='softmax'))\n    # output shape : (3,)\n    model.compile(loss='categorical_crossentropy',optimizer='Adam',metrics=['accuracy'])\n    if summary:\n        print(model.summary())\n    return model\n<\/code><\/pre>\n"}]}
{"QuestionId":42141054,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-02-09T15:47:49.793","AcceptedAnswerId":42144422.0,"OwnerUserId":3681744.0,"Title":"BatchNormalization in TensorFlow","Body":"<p>What is the difference between  and ? I have checked the documentation but did not understand if one is better than the other.<\/p>\n","answers":[{"AnswerId":"42144422","CreationDate":"2017-02-09T18:37:29.073","ParentId":null,"OwnerUserId":"3260576","Title":null,"Body":"<p>tf.contrib refers to a high level machine learning API over Tensorflow. So tf.contrib.layers.batch_norm basically acts as a wrapper over tf.nn.batch_normalization making it easier to use but less flexible than the latter. So, depending upon the application you can use whatever fits your needs. For example in case of tf.nn.batch_norm you will have to give a symbolic link to mean and variance tensors by yourself, but tf.contrib.layers.batch_norm will take care of that for you.<\/p>\n"}]}
{"QuestionId":42141072,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-02-09T15:48:37.090","AcceptedAnswerId":null,"OwnerUserId":6069513.0,"Title":"Tensorflow scalibility","Body":"<p>I am using tensorflow to train DNN, my network structure is very simple, each minibatch takes about 50ms when only one parameter server and one worker.  In order to process huge samples, I am using distributed ASGD training, however, I found that increasing worker count could not increase throughput, for example, 40 machines could achieve 1.5 million samples per second, after doubling parameter server machine count and worker machine count, cluster still could only process 1.5 million samples per second or even worse. The reason is each step takes much longer when cluster is large. Does tensorflow have good scalibility, and any advice for speeding up training?<\/p>\n","answers":[{"AnswerId":"42164268","CreationDate":"2017-02-10T16:33:04.103","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>General approach to solving these problems is to find where bottlenecks are. You could be hitting a bottleneck in software or in your hardware.<\/p>\n\n<p>General example of doing the math -- suppose you have 250M parameters, and each backward pass takes 1 second. This means each worker will be sending 1GB\/sec of data and receiving 1GB\/sec of data. If you have 40 machines, that'll be 80GB\/sec of transfer between workers and parameter server. Suppose parameter server machines only have 1GB\/sec fully duplex NIC cards. This means that if you have less than 40 parameter server shards, then your NIC card speed will be the bottleneck.<\/p>\n\n<p>After ruling that out, you should consider interconnect speed. You may have N network cards in your cluster, but the cluster most likely can't handle all network cards sending data to all other network cards. Can your cluster handle 80GB\/sec of data flowing between 80 machines? Google designs their own network hardware to handle their interconnect demands, so this is an important problem constraint.<\/p>\n\n<p>Once you checked that your network hardware can handle the load, I would check software. IE, suppose you have a single worker, how does \"time to send\" scale with the number of parameter server shards? If the scaling is strongly sublinear, this suggests a bottleneck, perhaps some inefficient scheduling of threads or some-such.<\/p>\n\n<p>As an example of finding and fixing a software bottleneck, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6116\" rel=\"nofollow noreferrer\">grpc RecvTensor is slow<\/a> issue. That issue involved gRPC layer become inefficient if you are trying to send more than 100MB messages. This issue was fixed in upstream gRPC release, but not integrated into TensorFlow release\u00a0yet, so current work-around is to break messages into pieces 100MB or smaller.<\/p>\n\n<p>The general approach to finding these is to write lots of benchmarks to validate your assumptions about the speed. <\/p>\n\n<p>Here are some examples:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/gist.github.com\/yaroslavvb\/e196107b5e0afc834652bd3153030c42\" rel=\"nofollow noreferrer\">benchmark<\/a> sending messages between workers(local) <\/p><\/li>\n<li><p><a href=\"https:\/\/gist.github.com\/yaroslavvb\/ea1b1bae0a75c4aae593df7eca72d9ca\" rel=\"nofollow noreferrer\">benchmark<\/a> sharded PS benchmark (local)<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":42141540,"AnswerCount":1,"Tags":"<tensorflow><computer-vision><deep-learning><caffe><keras>","CreationDate":"2017-02-09T16:10:25.507","AcceptedAnswerId":null,"OwnerUserId":288609.0,"Title":"image augmentation make performance worse","Body":"<p>I am working on an image segmentation project, and have been trying to adopt the image augmentation technique to increase the training set size. At first, I just try to use the horizontal flip to double the image size, but I found the performance was much worse than not using it. Are there any insight that can be shared. Thanks.<\/p>\n","answers":[{"AnswerId":"42143362","CreationDate":"2017-02-09T17:39:06.383","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So basically you need to answer yourself one important question: <strong>Is a flipped image a valid image in your domain?<\/strong><\/p>\n\n<ol>\n<li>If not - then it may harm your training process simply because you are providing a network an invalid input which may learn your network spurious patterns in your data. It's not so rare that flips might harm your training - e.g. in logo recognition it's important to not change the orientation of your data in order to learn logos correctly.<\/li>\n<li><p>If yes - then there might be loads of different reason why your model started to behave worse. One of them might be that it has simply too small capacity and it's not able to learn all the patterns in your data. Second - that you have not enough examples - and when you add the flipped image it turned out that it in fact memoized loads of your traning cases. Another thing is that maybe you learnt it for a too small amount of time and setting the number of iterations to a bigger value might be a good idea. <\/p>\n\n<p>One thing is sure - your model is not generalizing well since your flipped data is valid. <\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42141960,"AnswerCount":1,"Tags":"<build><tensorflow>","CreationDate":"2017-02-09T16:29:22.623","AcceptedAnswerId":42142451.0,"OwnerUserId":1326770.0,"Title":"tensorflow new op that uses external libraries","Body":"<p>I wrote an op that reads data from EXR files.<\/p>\n\n<p>The code is able to compile, but when I call  i have an  error.<\/p>\n\n<p>I suspect this is because my cc file requires  library. My question is how to tell tensorflow to load the additional libraries as well?<\/p>\n\n<p>An alternative may be, to include the additional library in my . <\/p>\n\n<p>For this, I tried calling in CMake<\/p>\n\n<pre><\/pre>\n\n<p>but it gives the same error.<\/p>\n","answers":[{"AnswerId":"42142451","CreationDate":"2017-02-09T16:53:06.690","ParentId":null,"OwnerUserId":"1326770","Title":null,"Body":"<p>I'm very sorry. The problem seems to be solved by adding the following to the CMakeLists.txt<\/p>\n\n<pre><code>target_link_libraries(custom_ops ${OpenEXR_LIBS})\n<\/code><\/pre>\n"}]}
{"QuestionId":42142074,"AnswerCount":0,"Tags":"<tensorflow><gpu>","CreationDate":"2017-02-09T16:35:12.360","AcceptedAnswerId":null,"OwnerUserId":7520802.0,"Title":"tensorflow training is slow, stops running on gpu?","Body":"<p>I'm running the cifar_train program from tensorflow's <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\/\" rel=\"nofollow noreferrer\">tutorial<\/a>. The tutorial says it reaches a peak performance of 86% after a few hours. I tried running this on my laptop, and after about 4 days it was still training. I'm wondering if I'm misreading the tutorial by thinking that it'll be finished after a few hours, or if that's the accuracy at that time. Also I'm looking at my gpu usage via afterburner and when I started it, it was around 30-40% usage, but after looking at it overnight the gpu usage is 0 and the cpu is at 100%. Is this normal?<\/p>\n","answers":[]}
{"QuestionId":42142774,"AnswerCount":1,"Tags":"<deep-learning><caffe><concat>","CreationDate":"2017-02-09T17:08:54.547","AcceptedAnswerId":42182439.0,"OwnerUserId":6786718.0,"Title":"What is the output dimension in the googlenet after the concat layer?","Body":"<p>When looking throught the prototxt of the googlenet one finds that the inception layers have a concat layer at the end which takes several bottom inputs.<\/p>\n\n<p>e.g:<\/p>\n\n<pre><\/pre>\n\n<p>As it can be seen, there is one 1x1 conv-layer, one 3x3 conv-layer , one 5x5 conv-layer and finally a pooling layer. These layers are described as following:<\/p>\n\n<pre><\/pre>\n\n<p>It can be seen that these have different numbers of output and also different filter size, anyhow the documentation on the concat layer is the following:<\/p>\n\n<blockquote>\n  <p>input:<\/p>\n  \n  <p><\/p>\n  \n  <p>Output:<\/p>\n  \n  <p>if axis = 0: , and all input c_i\n  should be the same. <\/p>\n  \n  <p>if axis = 1: , and all input n_i should be the same.<\/p>\n<\/blockquote>\n\n<p>Firstly, I am not sure what the default is and secondly I am not sure which Dimensions will have the output Volume, since width and height should stay the same but all thre conv layer produce different outputs. Any pointers would be really appreciated<\/p>\n","answers":[{"AnswerId":"42182439","CreationDate":"2017-02-11T23:25:53.590","ParentId":null,"OwnerUserId":"2878130","Title":null,"Body":"<p>The default value for 'Concat' axis is 1, thus concatenating through channel dimension. In order to do this, all the layers that are concatenated, should have the same height and width. Looking to the log, the dimensions are (assuming batch size 32):<\/p>\n\n<p>inception_3a\/1x1 -> [32, 64, 28, 28]<br\/>\ninception_3a\/3x3 -> [32, 128, 28, 28]<br\/>\ninception_3a\/5x5 -> [32, 32, 28, 28]<br\/>\ninception_3a\/pool_proj -> [32, 32, 28, 28]<\/p>\n\n<p>Thus the final output will have dimension:<br\/>\ninception_3a\/output -> [32 (64+128+32+32) 28, 28] -> [32, 256, 28, 28]<\/p>\n\n<p>As expected from the Caffe log:<\/p>\n\n<pre><code>Creating Layer inception_3a\/output\ninception_3a\/output &lt;- inception_3a\/1x1\ninception_3a\/output &lt;- inception_3a\/3x3\ninception_3a\/output &lt;- inception_3a\/5x5\ninception_3a\/output &lt;- inception_3a\/pool_proj\ninception_3a\/output -&gt; inception_3a\/output\nSetting up inception_3a\/output\nTop shape: 32 256 28 28 (6422528)\n<\/code><\/pre>\n"}]}
{"QuestionId":42142913,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2017-02-09T17:15:38.807","AcceptedAnswerId":42163485.0,"OwnerUserId":57215.0,"Title":"Tensorflow not using GPU for one dataset, where it does for a very similar dataset","Body":"<p>I'm using TensorFlow to train a model using data originating from two sources. For both sources the training and validation data shape are almost identical and the dtypes throughout are np.float32.<\/p>\n\n<p>The strange thing is, when I use the first data set the GPU on my machine is used, but when using the second data set the GPU is not used.<\/p>\n\n<p>Does anyone have some suggestions on how to investigate? <\/p>\n\n<pre><\/pre>\n\n<p>Edit. Here's a snip of the log with log_device_placement=True. <\/p>\n\n<pre><\/pre>\n\n<p>It does seem to be placing the tasks on the GPU, however I still see almost entirely 0% GPU-Util in the nvidia-smi monitor. <\/p>\n\n<p>The pandas dataframe is of course in memory. Is there any other IO that could be impacting this process?<\/p>\n\n<p>Edit 2: I captured the log_device_placement logs for both the fast and slow data sets. They are identical, even though in one case the GPU usage is 25%, and the other 0%. Really scratching my head now.... <\/p>\n","answers":[{"AnswerId":"42163485","CreationDate":"2017-02-10T15:53:14.697","ParentId":null,"OwnerUserId":"57215","Title":null,"Body":"<p>The cause of the slowness was the memory layout of the ndarray backing the DataFrame. The s2 data was column-major meaning that each row of features and target was not contiguous.<\/p>\n\n<p>This operation changes the memory layout:<\/p>\n\n<pre><code>s2_train_data = s2_train_data.values.copy(order='C')\n<\/code><\/pre>\n\n<p>and now the GPU is running at 26% utilisation. Happy days :)<\/p>\n"}]}
{"QuestionId":42143650,"AnswerCount":3,"Tags":"<python><keras>","CreationDate":"2017-02-09T17:54:13.497","AcceptedAnswerId":null,"OwnerUserId":6115356.0,"Title":"ImageDataGenerator.flow() gives IOError: [Errno 2] No such file or directory","Body":"<p>I was following the tutorial given here: <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>On executing the first half i.e. the data preprocessing code : <\/p>\n\n<pre><\/pre>\n\n<p>I get the error : <\/p>\n\n<pre><\/pre>\n\n<p>IOError: [Errno 2] No such file or directory: 'preview\/cat_0_7886.jpeg`<\/p>\n\n<p>How do I solve this?<\/p>\n","answers":[{"AnswerId":"43361458","CreationDate":"2017-04-12T06:01:45.383","ParentId":null,"OwnerUserId":"7830827","Title":null,"Body":"<p>I was having the same issue and I fixed using<\/p>\n\n<pre><code>folder = \"training_blond_sep\" (no slash or backslash)\n(...)\nfor batch in datagen.flow_from_directory(folder, classes=['cat1','cat2'], target_size=(180,180), batch_size=100,  shuffle = False, seed=7, save_to_dir=folder+'\/aug', save_prefix=\"changed\"):\n<\/code><\/pre>\n\n<p>but I had to create this folder manually.<\/p>\n\n<p>Hope this helps. <\/p>\n"},{"AnswerId":"42628005","CreationDate":"2017-03-06T14:22:51.050","ParentId":null,"OwnerUserId":"4265409","Title":null,"Body":"<p>Try using an absolute path like <code>\/tmp\/preview<\/code> after making sure that path exists<\/p>\n"},{"AnswerId":"49825819","CreationDate":"2018-04-13T22:18:13.223","ParentId":null,"OwnerUserId":"1007845","Title":null,"Body":"<p><code>save_to_dir<\/code> must have the full path and you have to create all directories in that path before hand<\/p>\n\n<p>Example:<br>\n<code>save_to_dir=r'C:\\ML\\augImage\\train'<\/code><\/p>\n"}]}
{"QuestionId":42144191,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning><keras>","CreationDate":"2017-02-09T18:23:30.730","AcceptedAnswerId":42144407.0,"OwnerUserId":297850.0,"Title":"the Keras layers(functions) corresponding to tf.nn.conv2d_transpose","Body":"<p>In , what are the layers(functions) corresponding to  in ?  I once saw the comment that we can . Is that right?<\/p>\n\n<p>In the following two examples, they all use this kind of combination.<\/p>\n\n<p>1) In <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">Building Autoencoders in Keras<\/a>, author builds decoder as follows.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0twc6.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0twc6.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>2) In an <a href=\"https:\/\/github.com\/jocicmarko\/ultrasound-nerve-segmentation\/blob\/master\/train.py\" rel=\"nofollow noreferrer\">u-uet implementation<\/a>, author builds deconvolution as follows<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42144407","CreationDate":"2017-02-09T18:36:01.140","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The corresponding layers in <code>Keras<\/code> are <a href=\"https:\/\/keras.io\/layers\/convolutional\/#deconvolution2d\" rel=\"nofollow noreferrer\">Deconvolution2D<\/a> layers.<\/p>\n\n<p>It's worth to mention that you should be really careful with them because they sometimes might behave in unexpected way. I strongly advise you to read <a href=\"https:\/\/stackoverflow.com\/questions\/39018767\/deconvolution2d-layer-in-keras\">this<\/a> Stack Overflow question (and its answer) before you start to use this layer.<\/p>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<ol>\n<li>Deconvolution is a layer which was add relatively recently - and maybe this is the reason why people advise you to use <code>Convolution2D * UpSampling2D<\/code>.<\/li>\n<li>Because it's relatively new - it may not work correctly in some cases. It also need some experience to use them properly.<\/li>\n<li>In fact - from a mathematical point of view - every Deconvolution might be presented as a composition of <code>Convolution2D<\/code> and <code>UpSampling2D<\/code> - so maybe this is the reason why it was mentioned in texts you provided.<\/li>\n<\/ol>\n\n<p><strong>UPDATE 2:<\/strong><\/p>\n\n<p>Ok. I think I found an easy explaination why <code>Deconvolution2D<\/code> might be presented in a form of a composition of <code>Convolution2D<\/code> and <code>UpSampling2D<\/code>. We would use a definition that <code>Deconvolution2D<\/code> is a gradient of some convolution layer. Let's consider three most common cases:<\/p>\n\n<ol>\n<li>The easiest one is a <code>Convolutional2D<\/code> without any pooling. In this case - as it's the linear operation - its gradient is a function itself - so <code>Convolution2D<\/code>.<\/li>\n<li>The more tricky one is a gradient of <code>Convolution2D<\/code> with <code>AveragePooling<\/code>. So: <code>(AveragePooling2D * Convolution2D)' = AveragePooling2D' * Convolution2D'<\/code>. But a gradient of <code>AveragePooling2D = UpSample2D * constant<\/code> - so it's also in this case when the preposition is true.<\/li>\n<li>The most tricky one is one with <code>MaxPooling2D<\/code>. In this case still <code>(MaxPooling2D * Convolution2D)' = MaxPooling2D' * Convolution2D'<\/code> But <code>MaxPooling2D' != UpSample2D<\/code>. But in this case one can easily find an easy <code>Convolution2D<\/code> which makes <code>MaxPooling2D' = Convolution2D * UpSample2D<\/code> (intuitively - a gradient of <code>MaxPooling2D<\/code> is a zero matrix with only one 1 on its diagonal. As <code>Convolution2D<\/code> might express a matrix operation - it may also represent the injection from a identity matrix to a <code>MaxPooling2D<\/code> gradient). So:\n<code>(MaxPooling2D * Convolution2D)' = UpSampling2D * Convolution2D * Convolution2D = UpSampling2D * Convolution2D'<\/code>.<\/li>\n<\/ol>\n\n<p>The final remark is that all parts of the proof have shown that <code>Deconvolution2D<\/code> is a composition of <code>UpSampling2D<\/code> and <code>Convolution2D<\/code> instead of opposite. One can easily proof that every function of a form a composition of <code>UpSampling2D<\/code> and <code>Convolution2D<\/code> might be easily presented in a form of a composition of <code>UpSampling2D<\/code> and <code>Convolution2D<\/code>. So basically - the proof is done :)<\/p>\n"}]}
{"QuestionId":42144496,"AnswerCount":1,"Tags":"<eclipse><lua><ide><torch>","CreationDate":"2017-02-09T18:41:33.980","AcceptedAnswerId":null,"OwnerUserId":1082019.0,"Title":"Missing dependencies installing Lua Development Tools IDE Eclipse on Linux CentOS 7","Body":"<p>I would like to use Eclipse for my Torch\/Lua project, on my CentOS Linux release 7.3.1611 (Core) 64-bit notebook.<\/p>\n\n<p>I installed the latest Eclipse Luna IDE and then I tried to install the Lua Development Tools (LDT) package by following their \"Existing Eclipse Installation\" on the <a href=\"http:\/\/www.eclipse.org\/ldt\/\" rel=\"nofollow noreferrer\">website<\/a>.<\/p>\n\n<p>Unfortunately, after trying to install the package, Eclipse says:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone have any idea about how to solve this problem?<\/p>\n","answers":[{"AnswerId":"42145084","CreationDate":"2017-02-09T19:16:48.450","ParentId":null,"OwnerUserId":"2670892","Title":null,"Body":"<p>Looks like the install wants version 5.5 or later of <code>org.eclipse.dltk.debug.ui<\/code> part of the Eclipse Dynamic Languages Toolkit.<\/p>\n\n<p>It looks like version 5.5 is part of Eclipse Neon, so Luna is too old.<\/p>\n"}]}
{"QuestionId":42144915,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2017-02-09T19:07:18.017","AcceptedAnswerId":null,"OwnerUserId":6842947.0,"Title":"Convert TensorFlow string to python string","Body":"<p>I am aware that in TensorFlow, a tf.string tensor is basically a byte string. I need to do some operation with a filename which is stored in a queue using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/string_input_producer\" rel=\"nofollow noreferrer\">tf.train.string_input_producer()<\/a>.<\/p>\n\n<p>A small snippet is shown below :<\/p>\n\n<pre><\/pre>\n\n<p>However as a byte string it gives an output like the following :<\/p>\n\n<pre><\/pre>\n\n<p>I tried to convert using <\/p>\n\n<pre><\/pre>\n\n<p>However Tensor objects are not iterable and hence this fails.<\/p>\n\n<p>Where am I going wrong ?<\/p>\n\n<p>Is it a missing feature in TensorFlow that tf.string be converted to a Python string easily , or is there some other feature I am not aware about ?<\/p>\n\n<p><strong>More Info<\/strong><\/p>\n\n<p>The filename_queue has been prepared as follows :<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"53416706","CreationDate":"2018-11-21T16:37:37.333","ParentId":null,"OwnerUserId":"1466792","Title":null,"Body":"<p>If I understand your issue correctly you can try:<\/p>\n\n<pre><code>print(filename.decode())\n<\/code><\/pre>\n"},{"AnswerId":"56654238","CreationDate":"2019-06-18T17:27:27.107","ParentId":null,"OwnerUserId":"3211036","Title":null,"Body":"<p>In tensorflow 2.0.0, it can be done in the following way:<\/p>\n\n<pre><code>import tensorflow as tf\n\nmy_str = tf.constant('Hello World')\nmy_str_npy = my_str.numpy()\n\nprint(my_str_npy)\ntype(my_str_npy)\n<\/code><\/pre>\n\n<p>This converts a string tensor into a string of 'bytes' class<\/p>\n"},{"AnswerId":"54997921","CreationDate":"2019-03-05T07:58:30.357","ParentId":null,"OwnerUserId":"3381961","Title":null,"Body":"<p>Use the as_text function in compat (from tensorflow.python.util) to convert the byte string of tensorflow. I.e. <\/p>\n\n<p>filename = compat.as_text(filename)<\/p>\n"},{"AnswerId":"50001474","CreationDate":"2018-04-24T12:18:03.733","ParentId":null,"OwnerUserId":"7006377","Title":null,"Body":"<pre><code>key, value = reader.read(filename_queue)\n<\/code><\/pre>\n\n<p>In this, the Reader just read the file you give, so value is the content of the file, not the filename, but you can output key, then you get filename<\/p>\n"}]}
{"QuestionId":42144956,"AnswerCount":1,"Tags":"<tensorflow><nlp>","CreationDate":"2017-02-09T19:09:59.477","AcceptedAnswerId":null,"OwnerUserId":860340.0,"Title":"Double Batching Tensorflow Input Data","Body":"<p>I am implementing an convnet for token classification of string data. I \nneed to take in string data from a TFRecord, batch shuffled, then perform some processing which expands the data, and batch that again.  Is this possible with two batch_shuffle operations? <\/p>\n\n<p>This is what I need to do: <\/p>\n\n<ol>\n<li>enqueue filenames into a filequeue<\/li>\n<li>for each serialized Example, put onto a shuffle_batch<\/li>\n<li>When I pull each example off the shuffle batch, I need to PAD it, replicate it by the sequence length, concatinating position vector, this creates multiple examples for each original example from the first batch.  I need to batch it again. <\/li>\n<\/ol>\n\n<p>Of course, one solution is to just preprocess the data before loading it into TF, but that will take up way more diskspace than is necessary.  <\/p>\n\n<p><strong>DATA<\/strong> <\/p>\n\n<p>Here is some sample data.  I have two \"Examples\".  Each Example contains features of a tokenized sentences and labels for each token: <\/p>\n\n<pre><\/pre>\n\n<p>Each \"Example\"  Now has features as below (some reducution for clarity): <\/p>\n\n<pre><\/pre>\n\n<p><strong>Transformation<\/strong><\/p>\n\n<p>After Batching the sparse data, I receive a sentence as list of tokens: <\/p>\n\n<p>['the', 'quick', 'brown', 'fox', ...] <\/p>\n\n<p>I need to PAD the list first to a predetermined SEQ_LEN, and then insert\nposition indices into each example, rotating the positions such that the \ntoke I want to classify is at pos 0, and every position token is relative\nto the 0 position: <\/p>\n\n<pre><\/pre>\n\n<p><strong>Batching and ReBatching The Data<\/strong><\/p>\n\n<p>Here is a simplified version of what I'm trying to do: <\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I'm now trying to use the RandomShuffleQueue.  On each enqueue, I would like to enqueue a batch with shape(None, SEQ_LEN). I've modified the code above to reflect this. <\/p>\n\n<p>I no longer get complaints about the input shapes, but the queuing does hang at   <\/p>\n","answers":[{"AnswerId":"42165351","CreationDate":"2017-02-10T17:31:52.077","ParentId":null,"OwnerUserId":"860340","Title":null,"Body":"<p>I was approaching the entire problem incorrectly.  I mistakenly was thinking that I had to define the complete shape of the batch while inserting into <code>tf.batch_shuffle<\/code>, but I actually needed to define  only the shape of each element that I was inputing, and set <code>enqueue_many=True<\/code>. <\/p>\n\n<p>Here is the the correct code: <\/p>\n\n<pre><code>single_batch=1\ninput_batch_size = 64\nmin_after_dequeue = 10 \ncapacity = min_after_dequeue + 3 * input_batch_size\nnum_epochs=2\nSEQ_LEN = 10\nfilenames =[outfilepath]\n\nfq = tf.train.string_input_producer(filenames, num_epochs=num_epochs, shuffle=True)\nreader = tf.TFRecordReader()\nkey, serialized_example = reader.read(fq)\n\n# Dequeue examples of batch_size == 1. Because all examples are Sparse Tensors, do 1 at a time\nfirst_batch = tf.train.shuffle_batch([serialized_example], ONE, capacity, min_after_dequeue)\n\n# Get a single sentence and preprocess it shape=(sent_len)\nsingle_sentence = tf.parse_example(first_batch, features=feature_mapping)\n\n# Preprocess Sentence.  shape=(sent_len, SEQ_LEN * 2). Each row is example\nprocessed_inputs = preprocess(single_sentence)\n\n# Re batch\ninput_batch = tf.train.shuffle_batch([processed_inputs],\n              batch_size=input_batch_size, \n              capacity=capacity, min_after_dequeue=min_after_dequeue,\n              shapes=[SEQ_LEN * 2],  enqueue_many=True) #&lt;- This is the fix\n\n\ninit_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer(), tf.initialize_all_tables())\n\nsess = create_session()\nsess.run(init_op)\n\n#tf.get_default_graph().finalize()\ncoord = tf.train.Coordinator()\nthreads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\ntry:\n  i = 0  \n  while True:\n    print i    \n    print sess.run(input_batch)\n    i += 1\nexcept tf.errors.OutOfRangeError as e:\n  print \"No more inputs.\"\n<\/code><\/pre>\n"}]}
{"QuestionId":42145668,"AnswerCount":2,"Tags":"<python-2.7><machine-learning><neural-network><theano><keras>","CreationDate":"2017-02-09T19:53:57.090","AcceptedAnswerId":null,"OwnerUserId":7343062.0,"Title":"Keras Theano BackEnd","Body":"<p>im new to Keras in Python. I just created my first convolution neural network for number recognition using MNIST datasets. However, i got this warning message which i cannot figure out the solution. <\/p>\n\n<blockquote>\n  <p>UserWarning: Method on_batch_end() is slow compared to the batch\n  update. Check your callbacks.<\/p>\n<\/blockquote>\n\n<p>The full code is provided in this <a href=\"http:\/\/pastebin.com\/PjnRK1H2\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n\n<p>By the way, im using windows 10 and . My  version is 1.2.1 and  is 0.8.2. Thank you in advance.<\/p>\n","answers":[{"AnswerId":"42800679","CreationDate":"2017-03-15T03:38:47.683","ParentId":null,"OwnerUserId":"7349628","Title":null,"Body":"<p>model.fit(X_train, Y_train,\n          batch_size=32, nb_epoch=10, verbose=1)<\/p>\n\n<p>Change verbose to higher amount. Like 2 or more.<\/p>\n\n<p>What verbose=1 does is printing a log line after every batch. Printing, itself is not an issue, but you don't want it to happen millions of times per second. <\/p>\n\n<p>Your computer is simply not good enough for your code. <\/p>\n\n<p>So yes, you can ignore this warning, or change verbose and get rid of it.<\/p>\n"},{"AnswerId":"42335562","CreationDate":"2017-02-20T02:31:49.110","ParentId":null,"OwnerUserId":"6222724","Title":null,"Body":"<p>I run your code in my PC and get the result follow. using windows 10 and python 2.7. My keras version is 1.0.5 and theano is 0.8.2. \nwarning dosen't appear, and no error you can ignore your warning I think.<\/p>\n\n<hr>\n\n<ol>\n<li><p>C:\\Anaconda\\python.exe H:\/keras-master\/examples\/download.py Using\nTheano backend.<\/p>\n\n<p>Using gpu device 0: GeForce GTX 745 (CNMeM is disabled, cuDNN 5103)\nC:\\Anaconda\\lib\\site-packages\\theano\\sandbox\\cuda__init__.py:600:\nUserWarning: Your cuDNN version is more recent than the one Theano\nofficially supports. If you see any problems, try updating Theano or\ndowngrading cuDNN to version 5.   warnings.warn(warn) ('X_train\nshape:', (50000L, 1L, 28L, 28L)) (50000L, 'train samples') (10000L,\n'test samples')<\/p><\/li>\n<\/ol>\n\n<p>Epoch 1\/10 50000\/50000 [==============================] - 44s -\n    loss: 0.9157 - acc: 0.6879<br>\nEpoch 2\/10 50000\/50000\n    [==============================] - 43s - loss: 0.3536 - acc: 0.8903 <\/p>\n\n<p>Epoch 3\/10 50000\/50000 [==============================] - 44s -\n    loss: 0.3032 - acc: 0.9065<br>\nEpoch 4\/10 50000\/50000\n    [==============================] - 44s - loss: 0.2753 - acc: 0.9150 <\/p>\n\n<p>Epoch 5\/10 50000\/50000 [==============================] - 43s -\n    loss: 0.2526 - acc: 0.9203<br>\nEpoch 6\/10 50000\/50000\n    [==============================] - 43s - loss: 0.2391 - acc: 0.9257 <\/p>\n\n<p>Epoch 7\/10 50000\/50000 [==============================] - 45s -\n    loss: 0.2285 - acc: 0.9296<br>\nEpoch 8\/10 50000\/50000\n    [==============================] - 43s - loss: 0.2155 - acc: 0.9322 <\/p>\n\n<p>Epoch 9\/10 50000\/50000 [==============================] - 43s -\n    loss: 0.2104 - acc: 0.9347<br>\nEpoch 10\/10 50000\/50000\n    [==============================] - 44s - loss: 0.1963 - acc: 0.9392 <\/p>\n\n<pre><code>('Test score:', 0.1083350375296548) ('Test accuracy:',\n0.96709999999999996) 4\/4 [==============================] - 0s [2 1 0 4] [[ 0.  0.  1.  0.  0.  0.  0.  0.  0.  0.]  [ 0.  1.  0.  0. \n0.  0.  0.  0.  0.  0.]  [ 1.  0.  0.  0.  0.  0.  0.  0.  0.  0.]  [ 0.  0.  0.  0.  1.  0.  0.  0.  0.  0.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":42145977,"AnswerCount":0,"Tags":"<tensorflow><distributed><keras>","CreationDate":"2017-02-09T20:11:54.873","AcceptedAnswerId":null,"OwnerUserId":3399275.0,"Title":"Loading keras (tensorflow) model from .h5 file","Body":"<p>I saved model trained in Keras as .h5 file. \nWhen trying to load it in distributed setting, I got the error:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): Cannot assign a device\n  to node 'policy\/dense_2_b': Could not satisfy explicit device\n  specification '\/job:ps\/task:0' because no devices matching that\n  specification are registered in this process; available devices:\n  \/job:localhost\/replica:0\/task:0\/cpu:0<\/p>\n<\/blockquote>\n\n<p>It seems that somehow variables in .h5 file are assigned to specific device with job \"localhost\". So when I am trying to load it on parameter server, I get the error. <\/p>\n\n<p>Could anyone clarify how to address this? I probably should load keras model first without starting servers and then to reload it on parameter server. But details are unclear for me..<\/p>\n","answers":[]}
{"QuestionId":42147362,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-09T21:38:00.350","AcceptedAnswerId":42149956.0,"OwnerUserId":7542570.0,"Title":"Create SparseTensor from dequeued batch of values, indices, shapes","Body":"<p>I am attempting to feed tensors, some of which are sparse, to a model from RAM.  I have created a PaddingFIFOQueue into which I am queuing the indices, values, and shapes for my sparse tensors separately under the assumption that sparse values cannot be batched from RAM by some other method (If that's not the case please let me know).  They need to be padded as the sequences are all of different lengths.<\/p>\n\n<p>I am dequeuing the following...<\/p>\n\n<pre><\/pre>\n\n<p>I have attempted to use these values to create a SparseTensor but am getting the following error.<\/p>\n\n<pre><\/pre>\n\n<p>The main portion of the code is as follows...<\/p>\n\n<pre><\/pre>\n\n<p>I assume this is because SparseTensor is expecting a rank 2 tensor and not a batch of rank 2 tensors (as indicated by the error message) but I am not sure how to convert the batch.<\/p>\n","answers":[{"AnswerId":"42149956","CreationDate":"2017-02-10T01:20:21.947","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>This is possible with a bit of tiling and reshaping:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef sparse_tensor_merge(indices, values, shape):\n  \"\"\"Creates a SparseTensor from batched indices, values, and shapes.\n\n  Args:\n    indices: A [batch_size, N, D] integer Tensor.\n    values: A [batch_size, N] Tensor of any dtype.\n    shape: A [batch_size, D] Integer Tensor.\n  Returns:\n    A SparseTensor of dimension D + 1 with batch_size as its first dimension.\n  \"\"\"\n  merged_shape = tf.reduce_max(shape, axis=0)\n  batch_size, elements, shape_dim = tf.unstack(tf.shape(indices))\n  index_range_tiled = tf.tile(tf.range(batch_size)[..., None],\n                              tf.stack([1, elements]))[..., None]\n  merged_indices = tf.reshape(\n      tf.concat([tf.cast(index_range_tiled, tf.int64), indices], axis=2),\n      [-1, 1 + tf.size(merged_shape)])\n  merged_values = tf.reshape(values, [-1])\n  return tf.SparseTensor(\n      merged_indices, merged_values,\n      tf.concat([[tf.cast(batch_size, tf.int64)], merged_shape], axis=0))\n<\/code><\/pre>\n\n<p>So for example:<\/p>\n\n<pre><code>batch_indices = tf.constant(\n    [[[0, 0], [0, 1]],\n     [[0, 0], [1, 1]]], dtype=tf.int64)\nbatch_values = tf.constant(\n    [[0.1, 0.2],\n     [0.3, 0.4]])\nbatch_shapes = tf.constant(\n    [[2, 2],\n     [3, 2]], dtype=tf.int64)\n\nmerged = sparse_tensor_merge(batch_indices, batch_values, batch_shapes)\n\nwith tf.Session():\n  print(merged.eval())\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>SparseTensorValue(indices=array([[0, 0, 0],\n       [0, 0, 1],\n       [1, 0, 0],\n       [1, 1, 1]]), \n  values=array([ 0.1       ,  0.2       ,  0.30000001,  0.40000001],\n      dtype=float32), \n  dense_shape=array([2, 3, 2]))\n<\/code><\/pre>\n\n<p>Notice that the shape of the combined SparseTensor is the original batch dimension followed by the max across the batch for each of the other dimensions.<\/p>\n"}]}
{"QuestionId":42147427,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-09T21:44:03.367","AcceptedAnswerId":42149498.0,"OwnerUserId":2191652.0,"Title":"Tensorflow: How to randomly crop input images and labels in the same way?","Body":"<p>I'm trying to take random crops from an image. Just like it's done in Caffe for the sake of data augmentation.<\/p>\n\n<p>I know that tensorflow already has a function<\/p>\n\n<pre><\/pre>\n\n<p>But I'm not sure whether it takes it takes the same crop for image and label. Also this function cannot automatically 0-pad images with one or two dimensions smaller than the crop size [h,w].<\/p>\n\n<p>That again is done by<\/p>\n\n<pre><\/pre>\n\n<p>But it takes only center crops not random crops.<\/p>\n\n<p>Edit:<br>\nHere is some code how the padding could be done: <\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately one cannot use the python  conditional here so one has to go with the ugly  instead.<\/p>\n","answers":[{"AnswerId":"42149498","CreationDate":"2017-02-10T00:30:56.550","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>I would suggest combining the image with the labels and randomly cropping them together:<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef random_crop_and_pad_image_and_labels(image, labels, size):\n  \"\"\"Randomly crops `image` together with `labels`.\n\n  Args:\n    image: A Tensor with shape [D_1, ..., D_K, N]\n    labels: A Tensor with shape [D_1, ..., D_K, M]\n    size: A Tensor with shape [K] indicating the crop size.\n  Returns:\n    A tuple of (cropped_image, cropped_label).\n  \"\"\"\n  combined = tf.concat([image, labels], axis=2)\n  image_shape = tf.shape(image)\n  combined_pad = tf.image.pad_to_bounding_box(\n      combined, 0, 0,\n      tf.maximum(size[0], image_shape[0]),\n      tf.maximum(size[1], image_shape[1]))\n  last_label_dim = tf.shape(labels)[-1]\n  last_image_dim = tf.shape(image)[-1]\n  combined_crop = tf.random_crop(\n      combined_pad,\n      size=tf.concat([size, [last_label_dim + last_image_dim]],\n                     axis=0))\n  return (combined_crop[:, :, :last_image_dim],\n          combined_crop[:, :, last_image_dim:])\n<\/code><\/pre>\n\n<p>As an example:<\/p>\n\n<pre><code>cropped_image, cropped_labels = random_crop_and_pad_image_and_labels(\n    image=tf.reshape(tf.range(4*4*3), [4, 4, 3]),\n    labels=tf.reshape(tf.range(4*4), [4, 4, 1]),\n    size=[2, 2])\n\nwith tf.Session() as session:\n  print(session.run([cropped_image, cropped_labels]))\n<\/code><\/pre>\n\n<p>Prints something like:<\/p>\n\n<pre><code>[array([[[30, 31, 32],\n        [33, 34, 35]],\n\n       [[42, 43, 44],\n        [45, 46, 47]]], dtype=int32), array([[[10],\n        [11]],\n\n       [[14],\n        [15]]], dtype=int32)]\n<\/code><\/pre>\n\n<p>And a second example with an under-sized image:<\/p>\n\n<pre><code>cropped_image, cropped_labels = random_crop_and_pad_image_and_labels(\n    image=tf.reshape(tf.range(4*1*3), [4, 1, 3]),\n    labels=tf.reshape(tf.range(4*1), [4, 1, 1]),\n    size=[2, 2])\n\nwith tf.Session() as session:\n  print(session.run([cropped_image, cropped_labels]))\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>[array([[[3, 4, 5],\n        [0, 0, 0]],\n\n       [[6, 7, 8],\n        [0, 0, 0]]], dtype=int32), array([[[1],\n        [0]],\n\n       [[2],\n        [0]]], dtype=int32)]\n<\/code><\/pre>\n"}]}
{"QuestionId":42147441,"AnswerCount":0,"Tags":"<deep-learning><caffe><openblas>","CreationDate":"2017-02-09T21:44:51.030","AcceptedAnswerId":null,"OwnerUserId":4241727.0,"Title":"Caffe always uses single core in CPU mode","Body":"<p>I've built caffe from source (based on the BVLC installation guide). The problem is that in CPU mode it always uses single core, although I set  to different values. <\/p>\n\n<p>On my CentOS box, I installed , , and  rpm packages and set  in . Also, under  I see both single and parallel shared library, but  output on caffe binary shows that it loads  and not  (presumably the parallel version).<\/p>\n\n<p>I'm not sure if there's anything else that I am missing or it's a bug in caffe.<\/p>\n","answers":[]}
{"QuestionId":42147539,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-02-09T21:51:35.583","AcceptedAnswerId":null,"OwnerUserId":2638485.0,"Title":"Force set_shape on Tensorflow Tensor","Body":"<p>I would like to change the shape of a tensorflow tensor.  does not work in this case because I am trying to re-define previously defined dimensions:<\/p>\n\n<pre><\/pre>\n\n<p>this seems to work, but required accessing a protected field:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a better way to do this? I know what I am doing here and <em>would<\/em> like to forcibly change the shape of the tensor.<\/p>\n","answers":[]}
{"QuestionId":42148770,"AnswerCount":1,"Tags":"<python><tensorflow><protocol-buffers>","CreationDate":"2017-02-09T23:19:00.403","AcceptedAnswerId":42151635.0,"OwnerUserId":3002273.0,"Title":"Getting Protobuf Message for every TensorFlow Operation","Body":"<p>From <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/language_bindings\/\" rel=\"nofollow noreferrer\">here<\/a>,<\/p>\n\n<p>There are a few ways to get a list of the OpDefs for the registered ops:<\/p>\n\n<blockquote>\n  <ul>\n  <li>TF_GetAllOpList in the C API retrieves all registered OpDef protocol    messages. This can be used to write the generator in the\n  client    language. This requires that the client language have\n  protocol buffer    support in order to interpret the OpDef messages. <\/li>\n  <li>The C++ function    OpRegistry::Global()->GetRegisteredOps() returns the same list of all    registered OpDefs (defined in\n  [tensorflow\/core\/framework\/op.h]). This    can be used to write the\n  generator in C++ (particularly useful for    languages that do not\n  have protocol buffer support). <\/li>\n  <li>The    ASCII-serialized version of that list is periodically checked in to    [tensorflow\/core\/ops\/ops.pbtxt] by an automated\n  process.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>But alas, I want to do this in Python like,<\/p>\n\n<pre><\/pre>\n\n<p>I want a way to get the Protobuf message for every operation in Tensorflow so that I can dump it as JSON via <\/p>\n","answers":[{"AnswerId":"42151635","CreationDate":"2017-02-10T04:39:03.583","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>It's in <code>ops.txt<\/code>. Here's an example of listing all <code>OpDef<\/code> messages for ops that produce strings outputs.<\/p>\n\n<pre><code>import tensorflow as tf\n\nfrom tensorflow.core.framework import op_def_pb2\nfrom google.protobuf import text_format\n\ndef get_op_types(op):\n    for attr in op.attr:\n        if attr.type != 'type':\n            continue\n        return list(attr.allowed_values.list.type)\n    return []\n\n# directory where you did \"git clone\"\ntensorflow_git_base = \"\/Users\/yaroslav\/tensorflow.git\"\nops_file = tensorflow_git_base+\"\/tensorflow\/tensorflow\/core\/ops\/ops.pbtxt\"\nops = op_def_pb2.OpList()\ntext_format.Merge(open(ops_file).read(), ops)\n\nfor op in ops.op:\n    # get templated string types\n    if tf.string in get_op_types(op):\n        print(op.name, op.summary)\n    #for arg in op.input_arg:\n    for arg in op.output_arg:\n        if arg.type == tf.string:\n            print(op.name, op.summary)\n            break\n<\/code><\/pre>\n\n<p>** Added **\nIf you want to be sensitive to new ops being added you could reverse engineer how current Python wrappers do it. For instance, consider <code>gen_array_ops.py<\/code> file. It has the following snippet<\/p>\n\n<pre><code>def _InitOpDefLibrary():\n  op_list = _op_def_pb2.OpList()\n  _text_format.Merge(_InitOpDefLibrary.op_list_ascii, op_list)\n  _op_def_registry.register_op_list(op_list)\n  op_def_lib = _op_def_library.OpDefLibrary()\n  op_def_lib.add_op_list(op_list)\n  return op_def_lib\n\n\n_InitOpDefLibrary.op_list_ascii = \"\"\"op {\n  name: \"BatchMatrixBandPart\"\n  input_arg {\n    name: \"input\"\n    type_attr: \"T\"\n  }\n  input_arg {\n    name: \"num_lower\"\n    type: DT_INT64\n  }\n  input_arg {\n    name: \"num_upper\"\n    type: DT_INT64\n  }\n  output_arg {\n    name: \"band\"\n    type_attr: \"T\"\n  }\n  attr {\n    name: \"T\"\n    type: \"type\"\n  }\n  deprecation {\n    version: 14\n    explanation: \"Use MatrixBandPart\"\n  }\n}\n<\/code><\/pre>\n\n<p>So those messages protobufs are generated from underlying C code during generation of <code>gen_array_ops<\/code>. For tracking down how they were generated, see <a href=\"https:\/\/stackoverflow.com\/a\/41149557\/419116\">https:\/\/stackoverflow.com\/a\/41149557\/419116<\/a><\/p>\n"}]}
{"QuestionId":42148871,"AnswerCount":0,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-02-09T23:27:26.957","AcceptedAnswerId":null,"OwnerUserId":2638485.0,"Title":"tensor.get_shape() vs gen_array_ops.shape(tensor) in Tensorflow","Body":"<p>What is the difference between calling  vs using  on a tensor?<\/p>\n","answers":[]}
{"QuestionId":42149128,"AnswerCount":0,"Tags":"<c++><deep-learning><caffe>","CreationDate":"2017-02-09T23:54:05.957","AcceptedAnswerId":null,"OwnerUserId":7337951.0,"Title":"Use MemoryDataLayer to test in caffe","Body":"<p>Now I want to load a image from memory then pass it to network to test on it.<\/p>\n\n<p>So I write code like this(judging <a href=\"https:\/\/medium.com\/@shiyan\/caffe-c-helloworld-example-with-memorydata-input-20c692a82a22#.nqoycpgsn\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@shiyan\/caffe-c-helloworld-example-with-memorydata-input-20c692a82a22#.nqoycpgsn<\/a> ):<\/p>\n\n<pre><\/pre>\n\n<p>But it core dump.I check the data and ensure the data is passed OK.\nSo I check into the caffe and find that in Net class:<\/p>\n\n<pre><\/pre>\n\n<p>The size of net_input_blobs_ is 0, so <\/p>\n\n<pre><\/pre>\n\n<p>in Forward function in Net class goes wrong.<\/p>\n\n<p>this is my proto about MemoryDataLayer:<\/p>\n\n<pre><\/pre>\n\n<p>I am confused for this question for several days, Can you give me some advice or some example of MemoryDataLayer?<\/p>\n","answers":[]}
{"QuestionId":42149649,"AnswerCount":0,"Tags":"<python-3.x><theano><keras>","CreationDate":"2017-02-10T00:46:33.537","AcceptedAnswerId":null,"OwnerUserId":7486546.0,"Title":"TypeError: ('Shape must be integers', TensorConstant{[ 1. 8. 1. 1.]}, 'float64')","Body":"<p>I am getting following error in Python 3.5 with theano and keras. I am running a CNN model in keras.<\/p>\n\n<pre><\/pre>\n\n<p>I am also getting following warning:<\/p>\n\n<pre><\/pre>\n\n<p>I think this warning is causing problem. I tried everything like casting to int etc.\nPlease help me resolve this problem.<\/p>\n\n<p>code link : <a href=\"http:\/\/ideone.com\/Y02erf\" rel=\"nofollow noreferrer\">http:\/\/ideone.com\/Y02erf<\/a><\/p>\n","answers":[]}
{"QuestionId":42149679,"AnswerCount":0,"Tags":"<python><numpy><machine-learning><tensorflow><data-science>","CreationDate":"2017-02-10T00:49:27.530","AcceptedAnswerId":null,"OwnerUserId":5368816.0,"Title":"Calculate linear indices Tensorflow","Body":"<p>Good afternoon.\nI continue to have issues with updating random elements in tensorflow by index.\nI want to randomly choose indices (half of all, for instance), and then set to zero elements correspond to that indices.\nHere's the problematic part:<\/p>\n\n<pre><\/pre>\n\n<p>It gives me ValueError: Cannot reshape a tensor with 10 elements to shape [784,10] (7840 elements) for 'foo_1\/Reshape_1' (op: 'Reshape') with input shapes: [10], [2]., but I don't know how to be here without reshaping (I tried to reshape to both s1 and s2, no use)<\/p>\n\n<p>I have already read these topics:<a href=\"https:\/\/stackoverflow.com\/questions\/40568572\/update-values-of-a-matrix-variable-in-tensorflow-advanced-indexing?noredirect=1&amp;lq=1\">Update values of a matrix variable in tensorflow, advanced indexing<\/a> (feed_dict doesn't seem to work in my case), <a href=\"https:\/\/stackoverflow.com\/questions\/24560298\/python-numpy-valueerror-operands-could-not-be-broadcast-together-with-shapes\">python numpy ValueError: operands could not be broadcast together with shapes<\/a> and practically everything on the subject on stackoverflow =(<\/p>\n","answers":[]}
{"QuestionId":42149892,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-10T01:13:35.813","AcceptedAnswerId":42163515.0,"OwnerUserId":1710775.0,"Title":"Convert binary to int in tensorflow","Body":"<p>I wish to convert a tensor of binary numbers, for example, [0,1,0,1,1] to an integer in tensorflow. In my case, the tensor is quite large, with a length of approximately 80 elements. Is there a way to do this efficiently? <\/p>\n","answers":[{"AnswerId":"42163515","CreationDate":"2017-02-10T15:54:33.540","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>This will almost certainly lead to overflows with length-80 binary Tensors, but the basic strategy would be to do a vectorized multiplication with a Tensor which has powers of two:<\/p>\n\n<pre><code>import tensorflow as tf\n\nbinary_string = tf.constant([1, 0, 0, 1, 1])\n\nresult = tf.reduce_sum(\n    tf.cast(tf.reverse(tensor=binary_string, axis=[0]), dtype=tf.int64)\n    * 2 ** tf.range(tf.cast(tf.size(binary_string), dtype=tf.int64)))\n\nwith tf.Session():\n    print(result.eval())\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>19\n<\/code><\/pre>\n"}]}
{"QuestionId":42150087,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-10T01:36:29.183","AcceptedAnswerId":null,"OwnerUserId":1601580.0,"Title":"What does it mean to import TensorFlow from source directory?","Body":"<p>I was trying to use TensorFlow in my Anaconda docker container and I got the error:<\/p>\n\n<blockquote>\n  <p>Error importing tensorflow.  Unless you are using bazel, you should\n  not try to import tensorflow from its source directory; please exit\n  the tensorflow source tree, and relaunch your python interpreter from\n  there.<\/p>\n<\/blockquote>\n\n<p>however, I have no idea what the source directory even means. What does the source directory mean and what does it mean to import tensorflow from it? Why is it bad to do that?<\/p>\n\n<p>I am only doing:<\/p>\n\n<pre><\/pre>\n\n<p>then inside of the container I am doing:<\/p>\n\n<pre><\/pre>\n\n<p>and then I go to the python3 interpreter and get that error once I import tensorflow.<\/p>\n\n<hr>\n\n<p>For completeness, the full error message is: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42151448","CreationDate":"2017-02-10T04:19:05.023","ParentId":null,"OwnerUserId":"5017858","Title":null,"Body":"<p>The source directory is the directory containing your TensorFlow installation.  Python is getting confused by the fact that it sees TensorFlow source files locally and not just in your <code>PYTHONPATH<\/code>.  The error message asks you to run the interpreter from a location other than the directory in which TensorFlow is installed.<\/p>\n"},{"AnswerId":"42159273","CreationDate":"2017-02-10T12:20:39.930","ParentId":null,"OwnerUserId":"5317732","Title":null,"Body":"<p>try with <code>export LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64<\/code><\/p>\n\n<p>source: <a href=\"http:\/\/withr.me\/install-tensorflow-ubuntu-16-04\/\" rel=\"nofollow noreferrer\">http:\/\/withr.me\/install-tensorflow-ubuntu-16-04\/<\/a><\/p>\n"},{"AnswerId":"45024094","CreationDate":"2017-07-11T01:47:20.130","ParentId":null,"OwnerUserId":"6009249","Title":null,"Body":"<p>Make sure you are using python 3.5.3! Other version may not work!<\/p>\n"}]}
{"QuestionId":42151044,"AnswerCount":1,"Tags":"<tensorflow><edward>","CreationDate":"2017-02-10T03:32:34.370","AcceptedAnswerId":46339520.0,"OwnerUserId":7345996.0,"Title":"where is the location of class Normal in edward's source code?","Body":"<p>When using edward, we always use  , but i didn't find the declaration of  in <a href=\"https:\/\/github.com\/blei-lab\/edward\/tree\/master\/edward\/models\" rel=\"nofollow noreferrer\">github<\/a><\/p>\n\n<p>anybody who can tell me where is it<\/p>\n","answers":[{"AnswerId":"46339520","CreationDate":"2017-09-21T08:54:56.367","ParentId":null,"OwnerUserId":"7008416","Title":null,"Body":"<p>They are defined in <a href=\"https:\/\/github.com\/blei-lab\/edward\/blob\/ec44461ecc7887d3dedf1cd4f1c821bab912b194\/edward\/models\/random_variables.py\" rel=\"nofollow noreferrer\"><code>edward\/models\/random_variables.py<\/code><\/a>.<\/p>\n\n<p>You import the <code>Normal<\/code> class like this:<\/p>\n\n<pre><code>from edward.models import Normal\n<\/code><\/pre>\n\n<p>This suggests looking in <a href=\"https:\/\/github.com\/blei-lab\/edward\/blob\/ec44461ecc7887d3dedf1cd4f1c821bab912b194\/edward\/models\/__init__.py\" rel=\"nofollow noreferrer\"><code>edward\/models\/__init__.py<\/code><\/a>, which has this line:<\/p>\n\n<pre><code>from edward.models.random_variables import *\n<\/code><\/pre>\n\n<p>Looking in <code>edward\/models\/random_variables.py<\/code> we find <a href=\"https:\/\/github.com\/blei-lab\/edward\/blob\/ec44461ecc7887d3dedf1cd4f1c821bab912b194\/edward\/models\/random_variables.py#L7-L27\" rel=\"nofollow noreferrer\">this code<\/a>:<\/p>\n\n<pre><code>from edward.models.random_variable import RandomVariable as _RandomVariable\nfrom tensorflow.contrib import distributions as _distributions\n\n# Automatically generate random variable classes from classes in\n# tf.contrib.distributions.\n_globals = globals()\nfor _name in sorted(dir(_distributions)):\n  _candidate = getattr(_distributions, _name)\n  if (_inspect.isclass(_candidate) and\n          _candidate != _distributions.Distribution and\n          issubclass(_candidate, _distributions.Distribution)):\n\n    # to use _candidate's docstring, must write a new __init__ method\n    def __init__(self, *args, **kwargs):\n      _RandomVariable.__init__(self, *args, **kwargs)\n    __init__.__doc__ = _candidate.__init__.__doc__\n    _params = {'__doc__': _candidate.__doc__,\n               '__init__': __init__}\n    _globals[_name] = type(_name, (_RandomVariable, _candidate), _params)\n\ndel _candidate\n<\/code><\/pre>\n\n<p>This goes through the <code>tensorflow.contrib.distributions<\/code> module looking for classes derived from <code>tensorflow.contrib.distributions.Distribution<\/code> (ignoring other attributes like e.g. the <code>__file__<\/code> member of the module, or the base <code>Distribution<\/code> class itself). For each one, it does a bit of hacking (which only affects the generated documentation) then executes this key line:<\/p>\n\n<pre><code>_globals[_name] = type(_name, (_RandomVariable, _candidate), _params)\n<\/code><\/pre>\n\n<p>The <a href=\"https:\/\/docs.python.org\/3.4\/library\/functions.html#type\" rel=\"nofollow noreferrer\"><code>type()<\/code><\/a> built-in function creates a new type i.e. declares a new class. The second parameter is the list of base classes, which here is edward's <code>RandomVariable<\/code> class and the TensorFlow random variable class. Earlier it defined <code>_globals<\/code> to be <a href=\"https:\/\/docs.python.org\/3.4\/library\/functions.html#globals\" rel=\"nofollow noreferrer\"><code>globals()<\/code><\/a>, which is a built-in function returning the dictionary of the module's variables. Therefore, in the case you're interested in, the line above is equivalent to the following:<\/p>\n\n<pre><code>from edward.models.random_variable import RandomVariable as EdRandVar\nfrom tensorflow.contrib.distributions import Normal as TfNormal\nNormal = type(\"Normal\", (EdRandVar, TfNormal), {...})\n<\/code><\/pre>\n\n<p>Which in turn is equivalent to this (if you ignore the docstring stuff):<\/p>\n\n<pre><code>from edward.models.random_variable import RandomVariable as EdRandVar\nfrom tensorflow.contrib.distributions import Normal as TfNormal\nclass Normal(EdRandVar, TfNormal):\n    pass\n<\/code><\/pre>\n"}]}
{"QuestionId":42151292,"AnswerCount":1,"Tags":"<python><theano><pymc3>","CreationDate":"2017-02-10T04:00:33.447","AcceptedAnswerId":42214002.0,"OwnerUserId":3615378.0,"Title":"Posterior Predictive Check on PyMC3 Deterministic Variable","Body":"<p><strong>TL; DR<\/strong><\/p>\n\n<p>What's the right way to do posterior predictive checks on  variables that take stochastics (rendering the deterministic also stochastic) as input?<\/p>\n\n<p><strong>Too Short; Didn't Understand<\/strong><\/p>\n\n<p>Say we have a  model like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I may be misunderstanding, but my intention is for the posteriors of  and  to be sampled, and for those samples to fed into the deterministic variables. Is the posterior predictive check only possible on observed random variables?<\/p>\n\n<p>It's straightforward to get posterior predictive samples from  and other random variables using , but the majority of my model's value is derived from the state of  and , given those samples.<\/p>\n\n<p>The problem arises in that  seems to return a . So, when this is called:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>...and  attempts this block of code in :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>...it complains because a  obviously doesn't have a .<\/p>\n\n<p>So, what is the right way to carry the posterior samples of stochastics through deterministics? Do I need to explicitly create a  that takes stochastic posterior samples and calculates the deterministic values? That seems silly given the fact that  already has the graph in place.<\/p>\n","answers":[{"AnswerId":"42214002","CreationDate":"2017-02-13T21:36:37.467","ParentId":null,"OwnerUserId":"3615378","Title":null,"Body":"<p>Yes, I was misunderstanding the purpose of <code>.sample_ppc<\/code>. You don't need it for unobserved variables because those have samples in the trace. Observed variables aren't sampled from, because their data is observed, thus you need <code>sample_ppc<\/code> to generate samples.<\/p>\n\n<p>In short, I can gather samples of the <code>pm.Deterministic<\/code> variables from the trace.<\/p>\n"}]}
{"QuestionId":42151994,"AnswerCount":1,"Tags":"<deep-learning><keras><lstm><recurrent-neural-network>","CreationDate":"2017-02-10T05:15:18.737","AcceptedAnswerId":null,"OwnerUserId":6685656.0,"Title":"Use Stateful LSTM with mini-batching and input with variable time steps, in Keras?","Body":"<p>I'm new in Keras, and trying to implement this network\n<a href=\"https:\/\/i.stack.imgur.com\/MHyVi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/MHyVi.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>this network takes a video frames as x = {x1,........,xT} where T is the number of the frames in the video and x is the visual features of the frames\nof size 2048<\/p>\n\n<p>i tried to use stateful LSTM as each sample have a number of frames as refereed <a href=\"http:\/\/philipperemy.github.io\/keras-stateful-lstm\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>and this is my model<\/p>\n\n<pre><\/pre>\n\n<p>and tried to train the model using mini-batching<\/p>\n\n<pre><\/pre>\n\n<p>but it seems like the model cannot converge as it gives 0.3 accuracy<\/p>\n\n<p>I also tried to do it with stateless LSTM with input shape (None,1024) but it didn't converge too<\/p>\n","answers":[{"AnswerId":"48850780","CreationDate":"2018-02-18T10:46:29.853","ParentId":null,"OwnerUserId":"9373938","Title":null,"Body":"<p>I think your LSTM is not able to extract relevant features from the video frames in order to achieve a good accuracy.<\/p>\n\n<p>The approach that usually gives the best results when dealing with images (or video frames) is extracting features with a stack of convolution + relu + max pooling layers (see <a href=\"https:\/\/arxiv.org\/abs\/1612.02903\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/abs\/1612.02903<\/a> which is a survey on facial expression recognition, they all use convolutions to extract useful features from images).<\/p>\n\n<p>These work best with 2-dimensional input, but i see that you represent a video frame with an array of size 2048 instead of a matrix. Usually images are represented with a shape similar to <code>(rows, cols, color_channels)<\/code>.<\/p>\n\n<p>In your case, the input would have shape <code>(1, None, rows, cols, color_channels)<\/code>, then the convolutions would look something like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras.layers import Input, LSTM, Conv2D, MaxPool2D, TimeDistributed, Flatten\n\nx = Input(batch_shape=(1, None, rows, cols, color_channels), name='x')\nconvs = TimeDistributed(Conv2D(16, kernel_size=(3,3), activation='relu', padding='same'))(x)\nconvs = TimeDistributed(MaxPool2D(pool_size=(2,2)))(convs)\nconvs = TimeDistributed(Conv2D(32, kernel_size=(3,3), activation='relu', padding='same'))(convs)\nconvs = TimeDistributed(MaxPool2D(pool_size=(2,2)))(convs)\nlstm_input = TimeDistributed(Flatten())(convs)\nlstmR = LSTM(256, return_sequences=True, name='lstmR', stateful=True)(lstm_input)\nlstmL = LSTM(256, return_sequences=True, go_backwards=True, name='lstmL', stateful=True)(lstm_input)\n...\n<\/code><\/pre>\n\n<p>Where <code>TimeDistrubuted<\/code> applies the given layer to each time step.<\/p>\n"}]}
{"QuestionId":42153339,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2017-02-10T07:01:31.767","AcceptedAnswerId":null,"OwnerUserId":5159608.0,"Title":"Recognizing specified objects via Android camera using TensorFlow","Body":"<p>I stuck with the next problem: is it possible to train TensorFlow model to \"remember\" any new object that is visible from Android device camera and recognize it next time it will be in the camera focus? I've tried to find tutorial and I've read many of them, but they only describe how to recognize object category with pre-trained model. Thanks for any help!<\/p>\n","answers":[{"AnswerId":"42169712","CreationDate":"2017-02-10T22:32:44.037","ParentId":null,"OwnerUserId":"5696243","Title":null,"Body":"<p>TensorFlow does not support on-device training at this point, only inference. So you can't easily update the model in real time.<\/p>\n\n<p>However, you might try treating the penultimate layer as a embedding space vector and, and then compute distance between frames to see how related they are. <\/p>\n\n<p>Alternatively if you have an example of a frame with the object and one without you can get then see the delta that signifies when that object has appeared in the frame.<\/p>\n\n<p>If you have a labeled dataset, you can retrain the graph offline: see <a href=\"https:\/\/petewarden.com\/2016\/09\/27\/tensorflow-for-mobile-poets\/\" rel=\"nofollow noreferrer\">https:\/\/petewarden.com\/2016\/09\/27\/tensorflow-for-mobile-poets\/<\/a>. It will use the knowledge already present in the Inception graph to quickly retrain just the top layer to recognize new classes of objects.<\/p>\n"}]}
{"QuestionId":42154108,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-10T07:48:48.837","AcceptedAnswerId":42157627.0,"OwnerUserId":2784721.0,"Title":"Tensorflow : Shape mismatch issue with one dimensional data","Body":"<p>I am trying to pass x_data as feed_dict but getting below error, I am not sure that what is wrong in the code.<\/p>\n\n<pre><\/pre>\n\n<p>My Code:<\/p>\n\n<pre><\/pre>\n\n<p>I checked the shape of the  and  and it is same<\/p>\n\n<pre><\/pre>\n\n<p>I am working with one dimensional data.\nAny help is appreciated, Thanks!<\/p>\n","answers":[{"AnswerId":"42157627","CreationDate":"2017-02-10T10:55:21.417","ParentId":null,"OwnerUserId":"1771479","Title":null,"Body":"<p>To make it work I have changed two things, first I changed <code>y<\/code> to be a <code>Tensor<\/code>. And secondly I have not changed the <code>x_data<\/code> to <code>Tensor<\/code>, as commented <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/client\/session_management#Session.run\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<blockquote>\n  <p>The optional feed_dict argument allows the caller to override the value of tensors in the graph. Each key in feed_dict can be one of the following types:<\/p>\n  \n  <p>If the key is a Tensor, the value may be a Python scalar, string, list, or numpy ndarray that can be converted to the same dtype as that tensor. Additionally, if the key is a placeholder, the shape of the value will be checked for compatibility with the placeholder.<\/p>\n<\/blockquote>\n\n<p>The changed code which works for me:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nmodel = tf.global_variables_initializer()\n#define x and y\nx = tf.placeholder(shape=[1000],dtype=tf.int32,name=\"x\")\ny = 5*x**2-3*x+15 # without tf.Variable, making it a tf.Tensor\nx_data = np.random.randint(0,100,size=1000) # without tf.pack\nprint(x_data)\nprint(x)\nwith tf.Session() as sess:\n    sess.run(model)\n    print(sess.run(y,feed_dict={x:x_data}))\n<\/code><\/pre>\n"}]}
{"QuestionId":42154190,"AnswerCount":1,"Tags":"<python><neural-network><theano><non-linear-regression>","CreationDate":"2017-02-10T07:53:35.623","AcceptedAnswerId":42163617.0,"OwnerUserId":4586728.0,"Title":"Theano Neural Net All Outputs Converge to Same Value For All Inputs","Body":"<p>I've been struggling to get the implementation of a neural net to converge to meaningful values. I have black and white images. Each image is either 40% black and 60% white or 60% white and 40% black. Classifying for more black or white.<\/p>\n\n<p>I break the images into arrays of pixel values and feed them through the network. The issues is that it converges to the same constant value for all images. I am using 1000 images to train. 25*25 pixels for input and a hidden layer of 20.<\/p>\n\n<p>CODE:<\/p>\n\n<pre><\/pre>\n\n<p>Cost Coverages to Single value as well as any inputs having same output:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42163617","CreationDate":"2017-02-10T16:00:52.263","ParentId":null,"OwnerUserId":"6169238","Title":null,"Body":"<p>Just an idea:<br>\nI have seen examples where the entire image was presented to the NN same way as you do it. However those networks were designed for character recognition and similar image processing. So if you feed the entire image to the network it will try to find similar images. I understood that your images are random and that can be the reason why it fails to train. Actually there may be no similarities between training images and there is nothing to learn. I would present the picture to the program this way if I want to distinguish between images of circles and squares. However for deciding if a picture is rather dark or light I would simply feed the network the count of black pixels and white pixels. Some linear pre-processing can be very beneficial.<\/p>\n"}]}
{"QuestionId":42154391,"AnswerCount":1,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2017-02-10T08:05:53.203","AcceptedAnswerId":42157415.0,"OwnerUserId":2508729.0,"Title":"TensorFlow cost function sorting","Body":"<p>I'm pretty new to tensor flow and i'm trying to make my network sort an series of 8 numbers. <\/p>\n\n<p>the way i would like to evaluate how well it has done is by subtracting x by y, invert the negative numbers and multiply them by a weight if thats necessary. And i was wondering if this is possible in tensorflow. Or if in any case there is better way to do this.<\/p>\n\n<p>Something that is not tensoflow would look something like:<\/p>\n\n<pre><\/pre>\n\n<p>In this case, if the total is a negative it will be inverted, however i would preferably do each element individually.<\/p>\n\n<p>I'm currently stuck after:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42157415","CreationDate":"2017-02-10T10:45:07.853","ParentId":null,"OwnerUserId":"2508729","Title":null,"Body":"<p>i found a solution that works and looks like :<\/p>\n\n<pre><code>prediction = neural_network_model(x)\nsub = tf.subtract(prediction,y)\nsign = tf.sign(sub)\ncost = tf.multiply(sub,sign)\n<\/code><\/pre>\n\n<p>this does not yet include the weight but for now its working fine<\/p>\n"}]}
{"QuestionId":42154464,"AnswerCount":1,"Tags":"<eclipse><tensorflow><bazel><tensorflow-serving>","CreationDate":"2017-02-10T08:11:07.247","AcceptedAnswerId":43071478.0,"OwnerUserId":7544641.0,"Title":"How do you debug tensorflow serving, what IDE do you use?","Body":"<p>I have used eclipse with the bazel plugin but cannot seem to bind the binaries to the gdb debugger. How do you develop your own sources and servables without the ability to debug them?<\/p>\n\n<p>Just looking for options on how this is done by other devs<\/p>\n","answers":[{"AnswerId":"43071478","CreationDate":"2017-03-28T13:52:45.677","ParentId":null,"OwnerUserId":"7778502","Title":null,"Body":"<p>You need to build using <code>bazel build -c dbg &lt;target&gt;<\/code> to build a binary with debug info, then you should be able to attach to the binaries with <code>gdb<\/code>.<\/p>\n"}]}
{"QuestionId":42154766,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-10T08:30:51.013","AcceptedAnswerId":null,"OwnerUserId":4489467.0,"Title":"Tensorflow InvalidArgumentError Invalid JPEG data, size 4096, Not Jpeg file 0x00 0x05","Body":"<p>I am using the Tensorflow Image-retraining Tutorial, managed to get the Flower_photos working, but when i put in my photos for a test train, it got the error <\/p>\n\n<p>I am wondering if its because of an image that too large? If there is a set-Max size for an image? Or if there is a fixed height\/width? <\/p>\n\n<p>Because i see that on the flower_photos example, the pictures in there has all different height and width, so i am assuming there is no fixed height\/width required? <\/p>\n\n<p>Because i made sure the images are jpeg, and i changed the retrain.py in the image_retraining folder to tf.image.decode_image....but no luck<\/p>\n\n<p>Edit \nNow i see that Not a JPEG file: starts with 0x00 0x05 not sure what it means<\/p>\n\n<p>Edit: I couldnt find any threads\/info or anything to help but after 14 hours of systematic testing, i have just copied the data set from Tensorflow retraining tutorial the flower_photos exactly as it is(from a Mac to Window 10 and onto VBox Ubuntu), and that also ran into the same problem. Now i believe its the copy\/transfering process causes a JPEG to corrupt. <\/p>\n","answers":[{"AnswerId":"42175204","CreationDate":"2017-02-11T11:15:49.547","ParentId":null,"OwnerUserId":"4489467","Title":null,"Body":"<p>After 18 hours of repeat testing and trying many things out. If a dataset is transfer via USB (in my case from a Mac to Window onto Vbox), it will corrupt the JPEG, even if it doesnt seem like its corrupted, the image can be viewed with no problem etc....but when the decoder tries to read it, it will start with 0x00 0x05. <\/p>\n"}]}
{"QuestionId":42154792,"AnswerCount":2,"Tags":"<machine-learning><neural-network><keras><keras-layer>","CreationDate":"2017-02-10T08:32:38.177","AcceptedAnswerId":42160691.0,"OwnerUserId":710734.0,"Title":"Keras Custom Layer 2D input -> 2D output","Body":"<p>I have an 2D input (or 3D if one consider the number of samples) and I want to apply a keras layer that would take this input and outputs another 2D matrix. So, for example, if I have an input with size (ExV), the learning weight matrix would be (SxE) and the output (SxV). Can I do this with Dense layer? <\/p>\n\n<p><strong>EDIT<\/strong> (Nassim request):<\/p>\n\n<p>The first layer is doing nothing. It's just to give an input to Lambda layer: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42160691","CreationDate":"2017-02-10T13:35:07.823","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>It depends on what you want to do. Is it 2D because it's a sequence? Then LSTM are made for that and will return a sequence if desired size if you set return_sequence=True.<\/p>\n\n<p>CNN's can also work on 2D inputs and will output something of variable size depending on the number of kernels you use.<\/p>\n\n<p>Otherwise you can reshape it to a (E x V, ) 1D tensor, use a Dense layer with SxV dimension and reshape the output to a (S,V) 2D tensor...<\/p>\n\n<p>I can not help you more, we need to know your usecase :-) there are too many possibilities with neural networks.<\/p>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>You can use TimeDistributed(Dense(S)).\nIf your input has a shape (E,V), you reshape to (V,E) to have V as the \"time dimension\". Then you apply TimeDistributed(Dense(S)) which will be a dense layer with weights (ExS), the output will have the shape (V,S) so you can reshape it to (S,V).<\/p>\n\n<p>Does that make what you want ? The TimeDistributed() layer will apply the same Dense(S) layer to every V lines of your input with shared weights.<\/p>\n\n<p><strong>EDIT 2:<\/strong><\/p>\n\n<p>After looking at the code of keras backend, it turns out that to use the transpose from tensorflow with 'permutation patterns' option available, you need to use K.permute_dimensions(x,pattern). The batch dimension must be included. In your case :<\/p>\n\n<pre><code>Lambda(lambda x: K.permute_dimensions(x,[0,2,1]))\n<\/code><\/pre>\n\n<p>K.transpose(x) uses the same function internally (for tf backend) but permutations is set to the default value which is [n,n-1,...,0].<\/p>\n"},{"AnswerId":"42160176","CreationDate":"2017-02-10T13:07:31.253","ParentId":null,"OwnerUserId":"2300116","Title":null,"Body":"<p>What you want is probably autoencoder.<\/p>\n\n<p><a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-autoencoders-in-keras.html<\/a><\/p>\n"}]}
{"QuestionId":42155234,"AnswerCount":1,"Tags":"<keras><gaussian>","CreationDate":"2017-02-10T08:59:16.520","AcceptedAnswerId":42585010.0,"OwnerUserId":7544784.0,"Title":"Does the GaussianDropout Layer in Keras retain probability like the Dropout Layer?","Body":"<p>I was wondering, if the GaussianDropout Layer in Keras retains probability like the Dropout Layer.\nThe Dropout Layer is implemented as an Inverted Dropout which retains probability.<\/p>\n\n<p>If you aren't aware of the problem you may have a look at the <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/3305\" rel=\"nofollow noreferrer\">discussion<\/a> and specifically at the linxihui's answer.\nThe crucial point which makes the Dropout Layer retaining the probability is the call of K.dropout, which isn't called by a GaussianDropout Layer.<\/p>\n\n<p>Is there any reason why GaussianDropout Layer does not retain probability?\nOr is it retaining but in another way being unseen?<\/p>\n\n<p>Similar - referring to the Dropout Layer: <a href=\"https:\/\/stackoverflow.com\/questions\/38592943\/is-the-keras-implementation-of-dropout-correct\">Is the Keras implementation of dropout correct?<\/a><\/p>\n","answers":[{"AnswerId":"42585010","CreationDate":"2017-03-03T17:25:01.110","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So Gaussian dropout doesn't need to retain probability as its origin comes from applying <em>Central Limit Theorem<\/em> to inverted dropout. The details might be found <a href=\"http:\/\/jmlr.org\/papers\/volume15\/srivastava14a\/srivastava14a.pdf\" rel=\"nofollow noreferrer\">here<\/a> in 2nd and 3rd paragraph of Multiplicative Gaussian Noise chapter (chapter 10, p. 1951).<\/p>\n"}]}
{"QuestionId":42156296,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-10T09:53:53.690","AcceptedAnswerId":44355224.0,"OwnerUserId":5080995.0,"Title":"DropoutWrapper being non-deterministic across runs?","Body":"<p>In the beginning of my code, (outside the scope of a ), I've set my random seed -<\/p>\n\n<pre><\/pre>\n\n<p>This is what my dropout definition looks like -<\/p>\n\n<pre><\/pre>\n\n<p>In my first experiment, I kept . All results obtained were deterministic. I'm running this on a multicore CPU.<\/p>\n\n<p>In my second experiment, I set  and I ran the same code two times. Each code had these statements,<\/p>\n\n<pre><\/pre>\n\n<p>Results for first code run -<\/p>\n\n<pre><\/pre>\n\n<p>Expected behaviour, since  uses .<\/p>\n\n<p>Results for second code run -<\/p>\n\n<pre><\/pre>\n\n<p>Why is this sequence not identical to the first output despite defining an operation and graph seed?<\/p>\n","answers":[{"AnswerId":"44355224","CreationDate":"2017-06-04T14:36:02.207","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>The answer was already provided in the comments, but no-one has written it explicitly yet, so here it is:<\/p>\n\n<p><code>dynamic_rnn<\/code> will internally use <code>tf.while_loop<\/code>, which can actually evaluate multiple iterations in parallel (see documentation on <code>parallel_iterations<\/code>). In practice, if everything inside the loop-body or loop-cond depends on the previous values, it cannot run anything in parallel but there could be computations which don't depend on the previous values. These will be evaluated in parallel. In your case, inside the <code>DropoutWrapper<\/code>, you have at some point sth like this:<\/p>\n\n<pre><code>random_ops.random_uniform(noise_shape, ...)\n<\/code><\/pre>\n\n<p>This operation is independent from the previous values of the loop, so it can be calculated in parallel for all time-steps. If you do such parallel execution, it will be non-deterministic which time-frame gets which dropout mask.<\/p>\n"}]}
{"QuestionId":42156316,"AnswerCount":1,"Tags":"<python><theano><nvcc><theano-cuda>","CreationDate":"2017-02-10T09:54:53.200","AcceptedAnswerId":42292149.0,"OwnerUserId":4342311.0,"Title":"Error to load compiler libraries while importing Theano: fatal error C1083","Body":"<p>I'm trying to use Theano with my GPU for several days in my Python environment.\nWhile importing theano, I obtain this error: <\/p>\n\n<pre><\/pre>\n\n<p>I tried to include \"D:\\Anaconda2\\MinGW\\x86_64-w64-mingw32\\include\" (this folder contains \"pyconfig.h\") in PYTHONPATH, in PATH and in .theanorc.txt, with the same error message.<\/p>\n\n<p>Do you have any ideas to connect anaconda\/cuda\/nvcc to the compiler in Theano?<\/p>\n\n<p>Here is my configuration:<\/p>\n\n<ul>\n<li>OS: Windows 7<\/li>\n<li>GPU: GeForce GTX 950<\/li>\n<li>cl.exe : D:\\Program Files (x86)\\Microsoft Visual Studio 11.0\\VC\\bin<\/li>\n<li>nvcc.exe : C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\bin<\/li>\n<li>Python IDE: Spyder from Anaconda 2, in Python 2.<\/li>\n<li>Theano 0.8.2<\/li>\n<\/ul>\n\n<p>I can \"import theano\" with the CPU, but I want to use GPU for computations.<\/p>\n\n<p>After typing \"import theano\" (with device=gpu), I can see a console window launching \"nvcc.exe\" before showing the error.<\/p>\n\n<p>I tested CUDA with Visual Studio 2012 and it's working, for example \"bilateralFilter\" works in Visual Studio without error:<\/p>\n\n<pre><\/pre>\n\n<ul>\n<li><p>Here is my theanorc.txt file (the same error occurs with or without [nvcc] and [cuda])<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Here is my PATH<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Here is my PYTHONPATH<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Here is the complete error message after typing \"import theano\" in Spyder (with device=gpu in theanorc.txt):<\/p>\n\n<pre><\/pre><\/li>\n<\/ul>\n","answers":[{"AnswerId":"42292149","CreationDate":"2017-02-17T07:42:20.167","ParentId":null,"OwnerUserId":"4342311","Title":null,"Body":"<p>I found a solution for my problem.<\/p>\n\n<p><strong>Short answer<\/strong><\/p>\n\n<ol>\n<li>Add C:\\Windows\\System32 to PATH<\/li>\n<li>Launch C:\\Program Files (x86)\\Microsoft Visual Studio 11.0\\VC\\bin\\amd64\\vcvars64.bat<\/li>\n<\/ol>\n\n<p>End.<\/p>\n\n<p>For information, here is my new current PATH:<\/p>\n\n<pre><code>C:\\Windows\\System32\nC:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\bin\nC:\\Program Files (x86)\\Microsoft Visual Studio 11.0\\VC\\bin\\amd64\nD:\\Anaconda2\nD:\\Anaconda2\\Scripts\nD:\\Anaconda2\\Library\\bin\n<\/code><\/pre>\n\n<p>Here is my PYTHONPATH:<\/p>\n\n<p><em>nothing<\/em><\/p>\n\n<p>Here is my .theanorc.txt:<\/p>\n\n<pre><code>[global]\ndevice=gpu\nfloatx = float32\n<\/code><\/pre>\n\n<p><strong>Some details<\/strong><\/p>\n\n<p>In addition with PATH, one needs to configure Windows registry, as explained in this post: <a href=\"https:\/\/stackoverflow.com\/questions\/10773732\/compilation-error-in-visual-studio-linked-with-python26\">compilation error in visual studio linked with python26<\/a> \n. It is done by launching vcvars64.bat<\/p>\n\n<p>I did this before, but I forgot to notice that a message appeared: <\/p>\n\n<pre><code>ERROR: Cannot determine the location of the VS Common Tools folder.\n<\/code><\/pre>\n\n<p>According to <a href=\"https:\/\/social.msdn.microsoft.com\/Forums\/en-US\/78703f6b-f610-456c-b770-76a12be3e1ae\/error-cannot-determine-the-location-of-the-vs-common-tools-folder?forum=vssetup\" rel=\"nofollow noreferrer\">https:\/\/social.msdn.microsoft.com\/Forums\/en-US\/78703f6b-f610-456c-b770-76a12be3e1ae\/error-cannot-determine-the-location-of-the-vs-common-tools-folder?forum=vssetup<\/a> , the solution is to add C:\\Windows\\System32 to PATH.<\/p>\n"}]}
{"QuestionId":42156479,"AnswerCount":1,"Tags":"<tensorflow><sparse-matrix><categorical-data>","CreationDate":"2017-02-10T10:02:38.560","AcceptedAnswerId":42168658.0,"OwnerUserId":4362500.0,"Title":"Can tensorflow handle categorical features with multiple inputs within one column?","Body":"<p>For example, I have data in the following csv format:<\/p>\n\n<blockquote>\n  <p>1, 2, 1:3:4, 2<\/p>\n  \n  <p>0, 1, 3:5, 1<\/p>\n  \n  <p>...<\/p>\n<\/blockquote>\n\n<p>Each column seperated by comma represent one feature. Normally, a feature is one-hot(<em>e.g. col0, col1, col3<\/em>), but in this case, the feature for <em>col2<\/em> has multiple inputs(seperated by colon). <\/p>\n\n<p>I'm sure tensorflow can handle one-hot feature with <strong>sparse tensor<\/strong>, but I'm not sure if it could handle feature with multiple inputs like <em>col2<\/em>?<\/p>\n\n<p>And if ok, how should it be represented in tensorflow's <strong>sparse tensor<\/strong>?<\/p>\n","answers":[{"AnswerId":"42168658","CreationDate":"2017-02-10T21:04:41.877","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>TensorFlow has some string processing ops which can handle lists within CSVs. I'd read the list as a string column first, the process it like this:<\/p>\n\n<pre><code>def process_list_column(list_column, dtype=tf.float32):\n  sparse_strings = tf.string_split(list_column, delimiter=\":\")\n  return tf.SparseTensor(indices=sparse_strings.indices,\n                         values=tf.string_to_number(sparse_strings.values,\n                                                    out_type=dtype),\n                         dense_shape=sparse_strings.dense_shape)\n<\/code><\/pre>\n\n<p>An example of using this function:<\/p>\n\n<pre><code># csv_input.csv contains:\n# 1,2,1:3:4,2\n# 0,1,3:5,1\nfilename_queue = tf.train.string_input_producer([\"csv_input.csv\"])\n# Read two lines, batched\n_, lines = tf.TextLineReader().read_up_to(filename_queue, 2)\ncolumns = tf.decode_csv(lines, record_defaults=[[0], [0], [\"\"], [0]])\ncolumns[2] = process_list_column(columns[2], dtype=tf.int32)\n\nwith tf.Session() as session:\n  coordinator = tf.train.Coordinator()\n  tf.train.start_queue_runners(session, coord=coordinator)\n\n  print(session.run(columns))\n\n  coordinator.request_stop()\n  coordinator.join()\n<\/code><\/pre>\n\n<p>Outputs:<\/p>\n\n<pre><code>[array([1, 0], dtype=int32), \n array([2, 1], dtype=int32), \n SparseTensorValue(indices=array([[0, 0],\n       [0, 1],\n       [0, 2],\n       [1, 0],\n       [1, 1]]), \n     values=array([1, 3, 4, 3, 5], dtype=int32), \n     dense_shape=array([2, 3])),\n array([2, 1], dtype=int32)]\n<\/code><\/pre>\n"}]}
{"QuestionId":42156833,"AnswerCount":1,"Tags":"<python><theano><symbolic-math>","CreationDate":"2017-02-10T10:18:02.747","AcceptedAnswerId":42171097.0,"OwnerUserId":6897811.0,"Title":"How do we get the list of variables that a theano expression depends on?","Body":"<p>In sympy, I would something like:<\/p>\n\n<pre><\/pre>\n\n<p>to get the variables X depends on. This is super convenient, because if we want to do a lambdify afterwards:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to do something similar with theano:<\/p>\n\n<pre><\/pre>\n\n<p>But, instead of providing  I would like to access directly the list of variables needed to create the <\/p>\n\n<p>Is it possible ? If so, I did not find it in the theano doc, so any help or a link would be appreciated :) <\/p>\n","answers":[{"AnswerId":"42171097","CreationDate":"2017-02-11T01:12:10.417","ParentId":null,"OwnerUserId":"1950580","Title":null,"Body":"<p>Some theano functions isn't well doc-ed since mostly for internal use.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nx, y = T.vectors('xy')\nz = x+y\ntheano.gof.graph.inputs([z])\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[x, y]\n<\/code><\/pre>\n"}]}
{"QuestionId":42156957,"AnswerCount":5,"Tags":"<python><tensorflow><gradient>","CreationDate":"2017-02-10T10:23:40.963","AcceptedAnswerId":null,"OwnerUserId":7545284.0,"Title":"How to update model parameters with accumulated gradients?","Body":"<p>I'm using TensorFlow to build a deep learning model. And new to TensorFlow.<\/p>\n\n<p>Due to some reason, my model has limited batch size, then this limited batch-size will make the model has a high variance.<\/p>\n\n<p>So, I want to use some trick to make the batch size larger. My idea is to store the gradients of each mini-batch, for example 64 mini-batches, and then sum the gradients together, use the mean gradients of this 64 mini batches of training data to update the model's parameters.<\/p>\n\n<p>This means that for the first 63 mini-batches, do not update the parameters, and after the 64 mini batch, update the model's parameters only once.<\/p>\n\n<p>But as TensorFlow is graph based, do anyone know how to implement this wanted feature?<\/p>\n\n<p>Thanks very much.<\/p>\n","answers":[{"AnswerId":"44599599","CreationDate":"2017-06-17T00:23:38.113","ParentId":null,"OwnerUserId":"5128473","Title":null,"Body":"<p>I had the same problem and just figured it out. <\/p>\n\n<p>First get symbolic gradients, then define accumulated gradients as tf.Variables. (It seems that <code>tf.global_variables_initializer()<\/code> has to be run before defining <code>grads_accum<\/code>. I got errors otherwise, not sure why.)<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>tvars = tf.trainable_variables()\noptimizer = tf.train.GradientDescentOptimizer(lr)\ngrads = tf.gradients(cost, tvars)\n\n# initialize\ntf.local_variables_initializer().run()\ntf.global_variables_initializer().run()\n\ngrads_accum = [tf.Variable(tf.zeros_like(v)) for v in grads] \nupdate_op = optimizer.apply_gradients(zip(grads_accum, tvars)) \n<\/code><\/pre>\n\n<p>In training you can accumulate gradients (saved in <code>gradients_accum<\/code>) at each batch, and update the model after running the 64-th batch:<\/p>\n\n<pre><code>feed_dict = dict()\nfor i, _grads in enumerate(gradients_accum):\n    feed_dict[grads_accum[i]] = _grads\nsess.run(fetches=[update_op], feed_dict=feed_dict) \n<\/code><\/pre>\n\n<p>You can refer to <a href=\"https:\/\/github.com\/DjangoPeng\/tensorflow\/blob\/06210b87746ae38351f0472041e47a26c4940d07\/tensorflow\/python\/training\/optimizer_test.py\" rel=\"nofollow noreferrer\">tensorflow\/tensorflow\/python\/training\/optimizer_test.py<\/a> for example usage, particularly this function: <code>testGradientsAsVariables()<\/code>. <\/p>\n\n<p>Hope it helps. <\/p>\n"},{"AnswerId":"56578297","CreationDate":"2019-06-13T10:15:52.047","ParentId":null,"OwnerUserId":"11641757","Title":null,"Body":"<p>You can use Pytorch instead of Tensorflow as it allows the user to accumulate gradients during training<\/p>\n"},{"AnswerId":"42222976","CreationDate":"2017-02-14T09:58:08.280","ParentId":null,"OwnerUserId":"7545284","Title":null,"Body":"<p>I found a solution here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3994#event-766328647\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3994#event-766328647<\/a><\/p>\n\n<pre><code>opt = tf.train.AdamOptimizer()\ntvs = tf.trainable_variables()\naccum_vars = [tf.Variable(tf.zeros_like(tv.initialized_value()), trainable=False) for tv in tvs]                                        \nzero_ops = [tv.assign(tf.zeros_like(tv)) for tv in accum_vars]\ngvs = opt.compute_gradients(rmse, tvs)\naccum_ops = [accum_vars[i].assign_add(gv[0]) for i, gv in enumerate(gvs)]\ntrain_step = opt.apply_gradients([(accum_vars[i], gv[1]) for i, gv in enumerate(gvs)])\n<\/code><\/pre>\n\n<p>In the training loop:<\/p>\n\n<pre><code>while True:\n    sess.run(zero_ops)\n    for i in xrange(n_minibatches):\n        sess.run(accum_ops, feed_dict=dict(X: Xs[i], y: ys[i]))\n    sess.run(train_step)\n<\/code><\/pre>\n\n<p>But this code seems not very clean and pretty, does anyone know how to optimize these code?<\/p>\n"},{"AnswerId":"56434311","CreationDate":"2019-06-03T20:42:25.877","ParentId":null,"OwnerUserId":"11595744","Title":null,"Body":"<p>The method you posted seems to fail if I dont give the feed_dict again in the sess.run(train_step). I don't know why require the feed_dict, but it is possible that run again all the accumulator adding with the last example repeated. This is what I had to do in my case:<\/p>\n\n<pre><code>            self.session.run(zero_ops)\n            for i in range(0, mini_batch):\n\n                self.session.run(accum_ops, feed_dict={self.ph_X: imgs_feed[np.newaxis, i, :, :, :], self.ph_Y: flow_labels[np.newaxis, i, :, :, :], self.keep_prob: self.dropout})\n\n            self.session.run(norm_acums, feed_dict={self.ph_X: imgs_feed[np.newaxis, i, :, :, :], self.ph_Y: flow_labels[np.newaxis, i, :, :, :], self.keep_prob: self.dropout})\n            self.session.run(train_op, feed_dict={self.ph_X: imgs_feed[np.newaxis, i, :, :, :], self.ph_Y: flow_labels[np.newaxis, i, :, :, :], self.keep_prob: self.dropout})\n<\/code><\/pre>\n\n<p>And for normalize the gradient I understand that is only divide the accumulated gradietn by the batchsize so I only add a new op<\/p>\n\n<pre><code>norm_accums = [accum_op\/float(batchsize) for accum_op in accum_ops]\n<\/code><\/pre>\n\n<p>Did someone have that same issue of feed_dict?<\/p>\n\n<p>*UPDATE\nAs I suppused that is wrong, it runs the all graph again with the last example in the batch.\nThis little code test that<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\nph = tf.placeholder(dtype=tf.float32, shape=[])\nvar_accum = tf.get_variable(\"acum\", shape=[], \ninitializer=tf.zeros_initializer())\nacum = tf.assign_add(var_accum, ph)\ndivide = acum\/5.0\ninit = tf.global_variables_initializer()\n    with tf.Session() as sess:\n    sess.run(init)\n    for i in range(5):\n         sess.run(acum, feed_dict={ph: 2.0})\n\nc = sess.run([divide], feed_dict={ph: 2.0})\n#10\/5 = 2\nprint(c)\n#but it gives 2.4, that is 12\/5, so sums one more time\n<\/code><\/pre>\n\n<p>I figured out how to solve this. So, tensorflow has conditional operations. I put\nthe accumulation in one branch and the last accumulation with normalization and update in another branch. My code is a mess, but for fast check of I'm saying I let a little code of an example of use.<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nph = tf.placeholder(dtype=tf.float32, shape=[])\n#placeholder for conditional braching in the graph\ncondph = tf.placeholder(dtype=tf.bool, shape=[])\n\nvar_accum = tf.get_variable(\"acum\", shape=[], initializer=tf.zeros_initializer())\n\naccum_op = tf.assign_add(var_accum, ph)\n\n#function when condition of condph is True\ndef truefn():\n   return accum_op\n#function when condtion of condph is False\ndef falsefn():\n   div = accum_op\/5.0\n   return div\n\n#return the conditional operation\ncond = tf.cond(condph, truefn, falsefn)\n\ninit = tf.global_variables_initializer()\n\nwith tf.Session() as sess:\n   sess.run(init)\n   for i in range(4):\n       #run only accumulation\n       sess.run(cond, feed_dict={ph: 2.0, condph: True})\n   #run acumulation and divition\n   c = sess.run(cond, feed_dict={ph: 2.0, condph: False})\n\nprint(c)\n#now gives 2\n<\/code><\/pre>\n\n<p>*IMPORTANT NOTE: Forget everything didnt work. The optimizers drop a failure.<\/p>\n"},{"AnswerId":"53056101","CreationDate":"2018-10-30T01:24:53.397","ParentId":null,"OwnerUserId":"10577637","Title":null,"Body":"<p>The previous solutions do not compute the average of the accumulated gradients, which may lead to instability in training. I've modified the above code, which should solve this problem.<\/p>\n\n<pre><code># Fetch a list of our network's trainable parameters.\ntrainable_vars = tf.trainable_variables()\n\n# Create variables to store accumulated gradients\naccumulators = [\n    tf.Variable(\n        tf.zeros_like(tv.initialized_value()),\n        trainable=False\n    ) for tv in trainable_vars\n]\n\n# Create a variable for counting the number of accumulations\naccumulation_counter = tf.Variable(0.0, trainable=False)\n\n# Compute gradients; grad_pairs contains (gradient, variable) pairs\ngrad_pairs = optimizer.compute_gradients(loss, trainable_vars)\n\n# Create operations which add a variable's gradient to its accumulator.\naccumulate_ops = [\n    accumulator.assign_add(\n        grad\n    ) for (accumulator, (grad, var)) in zip(accumulators, grad_pairs)\n]\n\n# The final accumulation operation is to increment the counter\naccumulate_ops.append(accumulation_counter.assign_add(1.0))\n\n# Update trainable variables by applying the accumulated gradients\n# divided by the counter. Note: apply_gradients takes in a list of \n# (grad, var) pairs\ntrain_step = optimizer.apply_gradients(\n    [(accumulator \/ accumulation_counter, var) \\\n        for (accumulator, (grad, var)) in zip(accumulators, grad_pairs)]\n)\n\n# Accumulators must be zeroed once the accumulated gradient is applied.\nzero_ops = [\n    accumulator.assign(\n        tf.zeros_like(tv)\n    ) for (accumulator, tv) in zip(accumulators, trainable_vars)\n]\n\n# Add one last op for zeroing the counter\nzero_ops.append(accumulation_counter.assign(0.0))\n<\/code><\/pre>\n\n<p>This code is used in the same manner as that provided by @weixsong.<\/p>\n"}]}
{"QuestionId":42157781,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-10T11:02:42.723","AcceptedAnswerId":42166910.0,"OwnerUserId":2503930.0,"Title":"Block Diagonal Matrices in Tensorflow","Body":"<p>Let's say I have a few tensors A_i of varying shape [N_i, N_i]. Is it possible in tensorflow to create a block diagonal matrix with these matrices on the diagonal? The only way I can think of now is completely construct it myself by stacking and adding tf.zeros.<\/p>\n","answers":[{"AnswerId":"42166910","CreationDate":"2017-02-10T19:08:14.333","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>I agree that it would be nice to have a C++ op that does this. In the meantime, here's what I do (getting the static shape information right is a bit fiddly):<\/p>\n\n<pre><code>import tensorflow as tf\n\ndef block_diagonal(matrices, dtype=tf.float32):\n  r\"\"\"Constructs block-diagonal matrices from a list of batched 2D tensors.\n\n  Args:\n    matrices: A list of Tensors with shape [..., N_i, M_i] (i.e. a list of\n      matrices with the same batch dimension).\n    dtype: Data type to use. The Tensors in `matrices` must match this dtype.\n  Returns:\n    A matrix with the input matrices stacked along its main diagonal, having\n    shape [..., \\sum_i N_i, \\sum_i M_i].\n\n  \"\"\"\n  matrices = [tf.convert_to_tensor(matrix, dtype=dtype) for matrix in matrices]\n  blocked_rows = tf.Dimension(0)\n  blocked_cols = tf.Dimension(0)\n  batch_shape = tf.TensorShape(None)\n  for matrix in matrices:\n    full_matrix_shape = matrix.get_shape().with_rank_at_least(2)\n    batch_shape = batch_shape.merge_with(full_matrix_shape[:-2])\n    blocked_rows += full_matrix_shape[-2]\n    blocked_cols += full_matrix_shape[-1]\n  ret_columns_list = []\n  for matrix in matrices:\n    matrix_shape = tf.shape(matrix)\n    ret_columns_list.append(matrix_shape[-1])\n  ret_columns = tf.add_n(ret_columns_list)\n  row_blocks = []\n  current_column = 0\n  for matrix in matrices:\n    matrix_shape = tf.shape(matrix)\n    row_before_length = current_column\n    current_column += matrix_shape[-1]\n    row_after_length = ret_columns - current_column\n    row_blocks.append(tf.pad(\n        tensor=matrix,\n        paddings=tf.concat(\n            [tf.zeros([tf.rank(matrix) - 1, 2], dtype=tf.int32),\n             [(row_before_length, row_after_length)]],\n            axis=0)))\n  blocked = tf.concat(row_blocks, -2)\n  blocked.set_shape(batch_shape.concatenate((blocked_rows, blocked_cols)))\n  return blocked\n<\/code><\/pre>\n\n<p>As an example:<\/p>\n\n<pre><code>blocked_tensor = block_diagonal(\n    [tf.constant([[1.]]),\n     tf.constant([[1., 2.], [3., 4.]])])\n\nwith tf.Session():\n  print(blocked_tensor.eval())\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>[[ 1.  0.  0.]\n [ 0.  1.  2.]\n [ 0.  3.  4.]]\n<\/code><\/pre>\n"},{"AnswerId":"57330303","CreationDate":"2019-08-02T16:23:24.083","ParentId":null,"OwnerUserId":"4615566","Title":null,"Body":"<p>For anyone visiting this now - tensorflow now has <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/linalg\/LinearOperatorBlockDiag\" rel=\"nofollow noreferrer\">tf.linalg.LinearOperatorBlockDiag<\/a>. Following the example from Allen above:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\ntfl = tf.linalg\n\nblocks = [tf.constant([[1.0]]), tf.constant([[1.0, 2.0], [3.0, 4.0]])]\n\nlinop_blocks = [tfl.LinearOperatorFullMatrix(block) for block in blocks]\nlinop_block_diagonal = tfl.LinearOperatorBlockDiag(linop_blocks)\n\n&gt;&gt;&gt; print(linop_block_diagonal.to_dense())\ntf.Tensor(\n[[1. 0. 0.]\n [0. 1. 2.]\n [0. 3. 4.]], shape=(3, 3), dtype=float32)\n<\/code><\/pre>\n"}]}
{"QuestionId":42158505,"AnswerCount":1,"Tags":"<python><neural-network><keras><conv-neural-network><face-recognition>","CreationDate":"2017-02-10T11:39:43.760","AcceptedAnswerId":null,"OwnerUserId":2756668.0,"Title":"Implementing facenet in keras","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/RT3TZ.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RT3TZ.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Here I am trying to implement open face in keras.\nbut i am confused about that how to do triplet embedding (As Image in above link)\nI know about triplet selection and convolution neural network<\/p>\n\n<p>can someone help to figure out:<\/p>\n\n<p>1> model structure for triplet training<\/p>\n\n<p>2> triplet loss function<\/p>\n","answers":[{"AnswerId":"47704330","CreationDate":"2017-12-07T21:43:58.820","ParentId":null,"OwnerUserId":"5175433","Title":null,"Body":"<p>The model can be anything, but generally it's a convolutional net. The triplet loss is computed on one of the late dense layer outputs, according to the formula you linked. Basically, if d(A, B) is the euclidean distance between outputs A and B (treated as vectors), then take <code>max(0, d(A, P)^2 - d(A, N)^2 + alpha)<\/code><\/p>\n"}]}
{"QuestionId":42158866,"AnswerCount":1,"Tags":"<machine-learning><neural-network><classification><keras>","CreationDate":"2017-02-10T11:58:49.570","AcceptedAnswerId":null,"OwnerUserId":2794119.0,"Title":"Neural network for multi label classification with large number of classes outputs only zero","Body":"<p>I am training a neural network for multilabel classification, with a large number of classes (1000). Which means more than one output can be active for every input. On an average, I have two classes active per output frame. On training with a cross entropy loss the neural network resorts to outputting only zeros, because it gets the least loss with this output since 99.8% of my labels are zeros. Any suggestions on how I can push the network to give more weight to the positive classes? <\/p>\n","answers":[{"AnswerId":"47313183","CreationDate":"2017-11-15T16:58:23.463","ParentId":null,"OwnerUserId":"979377","Title":null,"Body":"<p>Tensorflow has a loss function <code>weighted_cross_entropy_with_logits<\/code>, which can be used to give more weight to the 1's. So it should be applicable to a sparse multi-label classification setting like yours.<\/p>\n\n<p>From the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/weighted_cross_entropy_with_logits\" rel=\"noreferrer\">documentation<\/a>:<\/p>\n\n<blockquote>\n  <p>This is like sigmoid_cross_entropy_with_logits() except that pos_weight, allows one to trade off recall and precision by up- or down-weighting the cost of a positive error relative to a negative error.<\/p>\n  \n  <p>The argument pos_weight is used as a multiplier for the positive targets<\/p>\n<\/blockquote>\n\n<p>If you use the tensorflow backend in Keras, you can use the loss function like this (Keras 2.1.1):<\/p>\n\n<pre><code>import tensorflow as tf\nimport keras.backend.tensorflow_backend as tfb\n\nPOS_WEIGHT = 10  # multiplier for positive targets, needs to be tuned\n\ndef weighted_binary_crossentropy(target, output):\n    \"\"\"\n    Weighted binary crossentropy between an output tensor \n    and a target tensor. POS_WEIGHT is used as a multiplier \n    for the positive targets.\n\n    Combination of the following functions:\n    * keras.losses.binary_crossentropy\n    * keras.backend.tensorflow_backend.binary_crossentropy\n    * tf.nn.weighted_cross_entropy_with_logits\n    \"\"\"\n    # transform back to logits\n    _epsilon = tfb._to_tensor(tfb.epsilon(), output.dtype.base_dtype)\n    output = tf.clip_by_value(output, _epsilon, 1 - _epsilon)\n    output = tf.log(output \/ (1 - output))\n    # compute weighted loss\n    loss = tf.nn.weighted_cross_entropy_with_logits(targets=target,\n                                                    logits=output,\n                                                    pos_weight=POS_WEIGHT)\n    return tf.reduce_mean(loss, axis=-1)\n<\/code><\/pre>\n\n<p>Then in your model: <\/p>\n\n<pre><code>model.compile(loss=weighted_binary_crossentropy, ...)\n<\/code><\/pre>\n\n<p>I have not found many resources yet which report well working values for the <code>pos_weight<\/code> in relation to the number of classes, average active classes, etc.<\/p>\n"}]}
{"QuestionId":42158867,"AnswerCount":2,"Tags":"<tensorflow><scikit-learn><neural-network><regression><tflearn>","CreationDate":"2017-02-10T11:58:52.830","AcceptedAnswerId":42339584.0,"OwnerUserId":5895305.0,"Title":"tflearn multi layer perceptron with unexpected prediction","Body":"<p>I would like to rebuild a MLP I implemented first with scikit-learn's MLPRegressor with tflearn.<\/p>\n\n<p>sklearn.neural_network.MLPRegressor implementation:<\/p>\n\n<pre><\/pre>\n\n<p>The model worked and I got an accuracy of . Now I would like to build a similar MLP with tflearn. I started with the following code:<\/p>\n\n<pre><\/pre>\n\n<p>At some point I definitely configured something the wrong way because the prediction is way off. The range of Y_train is between  and  and the prediction shows numbers around . In the tflearn documentation I just found examples for classification.<\/p>\n\n<h2>UPDATE 1:<\/h2>\n\n<p>I realized that the regression layer uses by default  as loss-function which is for classification problems. So I selected  instead. I also tried to normalize . The prediction still not even matches the range of . Any thoughts?<\/p>\n\n<h2>FINAL UPDATE:<\/h2>\n\n<p>Take a look at the accepted answer.<\/p>\n","answers":[{"AnswerId":"42339584","CreationDate":"2017-02-20T08:23:16.137","ParentId":null,"OwnerUserId":"5895305","Title":null,"Body":"<p>I made a couple of actually really dumb mistakes. <\/p>\n\n<p>First of all I scalled the output to the interval <code>0<\/code> to <code>1<\/code> but used in the output-layer the activatuion function <code>tanh<\/code> which delivers values from <code>-1<\/code> to <code>1<\/code>. So I had to use either an activation function that outputs values between <code>0<\/code> and <code>1<\/code> (like e.g. <code>sigmoid<\/code>) or <code>linear<\/code> without any scaling applied.<\/p>\n\n<p>Secondly and most importantly, for my data I chose a pretty bad combination for <code>learning rate<\/code> and <code>n_epoch<\/code>. I didn't specify any learning rate and the default one is <code>0.1<\/code>, I think. In any case it was too small (I end up using <code>3.0<\/code>). At the same time the epoch count (<code>10<\/code>) was also far too small, with <code>200<\/code> it worked fine.<\/p>\n\n<p>I also explicitly chose <code>sgd<\/code> as <code>optimizer<\/code> (default: <code>adam<\/code>), which turned out to work a lot better.<\/p>\n\n<p>I updated the code in my question.<\/p>\n"},{"AnswerId":"42247802","CreationDate":"2017-02-15T11:16:09.807","ParentId":null,"OwnerUserId":"7096238","Title":null,"Body":"<p>One step should be not to scale the output.\nI am also working on regression problem and I scale only the inputs and it work fine with some neural networks. Although if I use tflearn I get wrong predictions. <\/p>\n"}]}
{"QuestionId":42159404,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2017-02-10T12:26:37.843","AcceptedAnswerId":42161648.0,"OwnerUserId":3667569.0,"Title":"How to access class probabilities in keras?","Body":"<p>I am training a model for which  I need to report class probabilities instead of a single classification. I have three classes and each training instance has either of the three classes assigned to it.<\/p>\n\n<p>I am trying to use Keras to create an MLP. But I can't figure how to extract the final class probabilities for each class. I am using this as my base example: <a href=\"http:\/\/machinelearningmastery.com\/regression-tutorial-keras-deep-learning-library-python\/\" rel=\"noreferrer\">http:\/\/machinelearningmastery.com\/regression-tutorial-keras-deep-learning-library-python\/<\/a><\/p>\n\n<p>Thanks !<\/p>\n","answers":[{"AnswerId":"42161648","CreationDate":"2017-02-10T14:21:12.683","ParentId":null,"OwnerUserId":"4955956","Title":null,"Body":"<p>You could use the <strong>predict<\/strong> method of your trained model<\/p>\n\n<blockquote>\n  <p>predict<\/p>\n  \n  <p>predict(self, x, batch_size=32, verbose=0)<\/p>\n  \n  <p>Generates output  predictions for the input samples, processing the samples in a batched\n  way.<\/p>\n  \n  <p>Arguments<\/p>\n  \n  <p>x: the input data, as a Numpy array (or list of Numpy arrays if the\n  model has multiple outputs). batch_size: integer. verbose: verbosity\n  mode, 0 or 1.<\/p>\n  \n  <p>Returns a Numpy array of predictions.<\/p>\n<\/blockquote>\n\n<pre><code>model.predict(input_to_your_network)\n<\/code><\/pre>\n"},{"AnswerId":"42176971","CreationDate":"2017-02-11T14:22:55.540","ParentId":null,"OwnerUserId":"7544784","Title":null,"Body":"<p>In order to perform multi-class classification (nb_classes > 1) you have to prepare your model in a specific manner. <\/p>\n\n<ol>\n<li>Make sure your labels are well-designed for multi-class classification. Have a look at the  <a href=\"https:\/\/keras.io\/utils\/np_utils\/\" rel=\"noreferrer\">numpy_utils<\/a><\/li>\n<li>You have to use the categorical_crossentropy as objective function for multi-class classification (see <a href=\"https:\/\/keras.io\/objectives\/\" rel=\"noreferrer\">Keras objectives<\/a>)<\/li>\n<li>Your last layer must have the softmax activation function (guarantees the output to be between 0 and 1) and nb_classes neurons.<\/li>\n<li>Train your model as usual<\/li>\n<li>Use the predict function. You will receive a vector of size (nb_classes,1) containing the probabilities of each class.<\/li>\n<\/ol>\n"}]}
{"QuestionId":42159518,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-10T12:33:30.817","AcceptedAnswerId":null,"OwnerUserId":6770703.0,"Title":"How to include the .so of custom ops in the pip wheel and organize the sources of custom ops?","Body":"<p>Following <a href=\"https:\/\/www.tensorflow.org\/how_tos\/adding_an_op\/\" rel=\"nofollow noreferrer\">the documentation<\/a>, I put  and  under , and created  which contains<\/p>\n\n<pre><\/pre>\n\n<p>Then I run the following commands under the root of tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>After building and installing the pip wheel, I want to use  in the project <em><\/em>.<\/p>\n\n<p>I think I should create something like  and , which calls  like <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/how_tos\/adding_an_op\/cuda_op.py\" rel=\"nofollow noreferrer\">the example cuda_op.py<\/a>. However, the  is not included in the installed pip wheel. How can I generate the input (the path of ) for ?<\/p>\n\n<p>Is there any better way to organized , ,  with <em><\/em>?<\/p>\n","answers":[{"AnswerId":"52606531","CreationDate":"2018-10-02T10:27:34.073","ParentId":null,"OwnerUserId":"5265615","Title":null,"Body":"<p>You can preserve directory structure of your project and create setup.py such that it also include .so files. You can also add other non-python files of your project same way.<br>\nExample Directory Structure:<\/p>\n\n<p><code>my_package\n   my_project\n   __init__.py\n   setup.py<\/code><\/p>\n\n<p>You can install 'my_project' package while in my_package directory using command:<br>\n<code>pip install . --user<\/code>  (Avoid --user argument if you install packages with root access) <\/p>\n\n<pre><code>from setuptools import setup, find_packages\n\nsetup(name='my_project',\n      version='1.0',\n      description='Project Details',\n      packages=find_packages(),\n      include_package_data=True,\n      package_data = {\n           '': ['*.so', '*.txt', '*.csv'],\n      },\n      zip_safe=False) \n<\/code><\/pre>\n\n<p>Don't forget to add <code>__init__.py<\/code> in all folders containing python modules you want to import.\nReference: <a href=\"https:\/\/docs.python.org\/2\/distutils\/setupscript.html#installing-package-data\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/2\/distutils\/setupscript.html#installing-package-data<\/a><\/p>\n"}]}
{"QuestionId":42159830,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-10T12:50:04.643","AcceptedAnswerId":null,"OwnerUserId":6306884.0,"Title":"How to use tf.compute_gradient_error","Body":"<p>I want to use  to check if all gradients are reasonable (almost equal zero).   <\/p>\n\n<p>I did a small test, It seems this function accept placeholder.<br>\nCompute gradient error without feed placeholder data is ok<\/p>\n\n<pre><\/pre>\n\n<p>But when I start check, It shows error message <\/p>\n\n<pre><\/pre>\n\n<p>lossL2 is weight l2 loss<\/p>\n\n<pre><\/pre>\n\n<p>If I use parameter  to feed placeholder data, the pycharm debuger show <br>\nI think timeout is a acceptable result, because my model use almost 2 mins for each iteraction, I need to find some way to speed up    <\/p>\n\n<p><strong>Edit 02\/11<\/strong>    <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42159937,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-10T12:55:35.530","AcceptedAnswerId":42165077.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Does each session run initiate a different batch of data in a graph?","Body":"<p>Meaning to say if I have the following graph like:<\/p>\n\n<pre><\/pre>\n\n<p>and in a supervisor  as  within the graph context, I run the following every once in a while:<\/p>\n\n<pre><\/pre>\n\n<p>Do they actually call in different batches for each sess run, given that the batch tensor must actually start from  onwards before they ever get to , , or ? Because now when I run each of these sessions, I get a very confusing result in that even the predictions do not match , and despite a high accuracy in the model, I do not get any predictions that remotely even match the labels to give that kind of high accuracy. Therefore I suspect that each of the result from  probably come from a different batch of data.<\/p>\n\n<p>This brings me to my next question: Is there a way to inspect the results of different parts of a graph when a batch from load_batch goes all the way till a train_op, where the  is actually run instead? In other words, is there a way to do what I want to do without calling for another ?<\/p>\n\n<p>Also, if I were to check the results using sess.run in such a way, would it affect my training in that some batches of data will be skipped and not reach the train_op?<\/p>\n","answers":[{"AnswerId":"42165077","CreationDate":"2017-02-10T17:15:35.980","ParentId":null,"OwnerUserId":"5107084","Title":null,"Body":"<p>I realized there is a problem with running using separate <code>sess.run<\/code> in that the data loaded is always different. Instead, when I did something like:<\/p>\n\n<pre><code>logits, probabilities, predictions, labels = sess.run([logits, probabilities, predictions, labels])\nprint 'logits: \\n', logits\nprint 'Probabilities: \\n', probabilities\nprint 'predictions: \\n', predictions\nprint 'Labels:\\n:', labels\n<\/code><\/pre>\n\n<p>All the quantities coincide very well as what I had expected. I had also tried using <code>tf.Print<\/code> by writing something like:<\/p>\n\n<pre><code>logits = tf.Print(logits, [logits], message = 'logits: \\n', summarize = 100)\n<\/code><\/pre>\n\n<p>immediately after defining the <code>logits<\/code>, so that they can get printed within the same session I run the <code>train_op<\/code>. However, the printing is rather messy and so I would prefer the first method of running everything in a session to obtain the values first and then printing them normally like numpy arrays. <\/p>\n"}]}
{"QuestionId":42159946,"AnswerCount":1,"Tags":"<tensorflow><conditional>","CreationDate":"2017-02-10T12:56:18.133","AcceptedAnswerId":42164494.0,"OwnerUserId":1876627.0,"Title":"How to use tf.cond for batch processing","Body":"<p>I want to use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#case\" rel=\"noreferrer\">tf.cond(pred, fn1, fn2, name=None)<\/a> for conditional branching. Let say I have two tensors: . Each tensor is a batch of 0\/1 and I want to use this tensors compression  as the source for \n<strong>tf.cond<\/strong>  argument:<\/p>\n\n<blockquote>\n  <p>pred: A scalar determining whether to return the result of fn1 or fn2.<\/p>\n<\/blockquote>\n\n<p>But if I am working with batches then it looks like I need to iterate over the source tensor inside the graph and make slices for every item in batch and apply tf.cond for every item. Looks suspiciously as for me. Why tf.cond not accept batch and only scalar? Can you advise what is the right way to use it with batch?<\/p>\n","answers":[{"AnswerId":"42164494","CreationDate":"2017-02-10T16:44:52.893","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/where\" rel=\"noreferrer\">tf.where<\/a> sounds like what you want: a vectorized selection between Tensors.<\/p>\n\n<p><code>tf.cond<\/code> is a control flow modifier: it determines which ops are <em>executed<\/em>, and so it's difficult to think of useful batch semantics.<\/p>\n\n<p>We can also put together a mixture of these operations: an operation which slices based on a condition and passes those slices to two branches.<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.python.util import nest\n\ndef slicing_where(condition, full_input, true_branch, false_branch):\n  \"\"\"Split `full_input` between `true_branch` and `false_branch` on `condition`.\n\n  Args:\n    condition: A boolean Tensor with shape [B_1, ..., B_N].\n    full_input: A Tensor or nested tuple of Tensors of any dtype, each with\n      shape [B_1, ..., B_N, ...], to be split between `true_branch` and\n      `false_branch` based on `condition`.\n    true_branch: A function taking a single argument, that argument having the\n      same structure and number of batch dimensions as `full_input`. Receives\n      slices of `full_input` corresponding to the True entries of\n      `condition`. Returns a Tensor or nested tuple of Tensors, each with batch\n      dimensions matching its inputs.\n    false_branch: Like `true_branch`, but receives inputs corresponding to the\n      false elements of `condition`. Returns a Tensor or nested tuple of Tensors\n      (with the same structure as the return value of `true_branch`), but with\n      batch dimensions matching its inputs.\n  Returns:\n    Interleaved outputs from `true_branch` and `false_branch`, each Tensor\n    having shape [B_1, ..., B_N, ...].\n  \"\"\"\n  full_input_flat = nest.flatten(full_input)\n  true_indices = tf.where(condition)\n  false_indices = tf.where(tf.logical_not(condition))\n  true_branch_inputs = nest.pack_sequence_as(\n      structure=full_input,\n      flat_sequence=[tf.gather_nd(params=input_tensor, indices=true_indices)\n                     for input_tensor in full_input_flat])\n  false_branch_inputs = nest.pack_sequence_as(\n      structure=full_input,\n      flat_sequence=[tf.gather_nd(params=input_tensor, indices=false_indices)\n                     for input_tensor in full_input_flat])\n  true_outputs = true_branch(true_branch_inputs)\n  false_outputs = false_branch(false_branch_inputs)\n  nest.assert_same_structure(true_outputs, false_outputs)\n  def scatter_outputs(true_output, false_output):\n    batch_shape = tf.shape(condition)\n    scattered_shape = tf.concat(\n        [batch_shape, tf.shape(true_output)[tf.rank(batch_shape):]],\n        0)\n    true_scatter = tf.scatter_nd(\n        indices=tf.cast(true_indices, tf.int32),\n        updates=true_output,\n        shape=scattered_shape)\n    false_scatter = tf.scatter_nd(\n        indices=tf.cast(false_indices, tf.int32),\n        updates=false_output,\n        shape=scattered_shape)\n    return true_scatter + false_scatter\n  result = nest.pack_sequence_as(\n      structure=true_outputs,\n      flat_sequence=[\n          scatter_outputs(true_single_output, false_single_output)\n          for true_single_output, false_single_output\n          in zip(nest.flatten(true_outputs), nest.flatten(false_outputs))])\n  return result\n<\/code><\/pre>\n\n<p>Some examples:<\/p>\n\n<pre><code>vector_test = slicing_where(\n    condition=tf.equal(tf.range(10) % 2, 0),\n    full_input=tf.range(10, dtype=tf.float32),\n    true_branch=lambda x: 0.2 + x,\n    false_branch=lambda x: 0.1 + x)\n\ncross_range = (tf.range(10, dtype=tf.float32)[:, None]\n               * tf.range(10, dtype=tf.float32)[None, :])\nmatrix_test = slicing_where(\n    condition=tf.equal(tf.range(10) % 3, 0),\n    full_input=cross_range,\n    true_branch=lambda x: -x,\n    false_branch=lambda x: x + 0.1)\n\nwith tf.Session():\n  print(vector_test.eval())\n  print(matrix_test.eval())\n<\/code><\/pre>\n\n<p>Prints:<\/p>\n\n<pre><code>[ 0.2         1.10000002  2.20000005  3.0999999   4.19999981  5.0999999\n  6.19999981  7.0999999   8.19999981  9.10000038]\n[[  0.           0.           0.           0.           0.           0.\n    0.           0.           0.           0.        ]\n [  0.1          1.10000002   2.0999999    3.0999999    4.0999999\n    5.0999999    6.0999999    7.0999999    8.10000038   9.10000038]\n [  0.1          2.0999999    4.0999999    6.0999999    8.10000038\n   10.10000038  12.10000038  14.10000038  16.10000038  18.10000038]\n [  0.          -3.          -6.          -9.         -12.         -15.\n  -18.         -21.         -24.         -27.        ]\n [  0.1          4.0999999    8.10000038  12.10000038  16.10000038\n   20.10000038  24.10000038  28.10000038  32.09999847  36.09999847]\n [  0.1          5.0999999   10.10000038  15.10000038  20.10000038\n   25.10000038  30.10000038  35.09999847  40.09999847  45.09999847]\n [  0.          -6.         -12.         -18.         -24.         -30.\n  -36.         -42.         -48.         -54.        ]\n [  0.1          7.0999999   14.10000038  21.10000038  28.10000038\n   35.09999847  42.09999847  49.09999847  56.09999847  63.09999847]\n [  0.1          8.10000038  16.10000038  24.10000038  32.09999847\n   40.09999847  48.09999847  56.09999847  64.09999847  72.09999847]\n [  0.          -9.         -18.         -27.         -36.         -45.\n  -54.         -63.         -72.         -81.        ]]\n<\/code><\/pre>\n"}]}
{"QuestionId":42160421,"AnswerCount":1,"Tags":"<ipython-notebook><keras>","CreationDate":"2017-02-10T13:20:57.810","AcceptedAnswerId":42162955.0,"OwnerUserId":826983.0,"Title":"Verbosity is killing IPython Notebook","Body":"<p>I am currently working with  in an IPython 3 notebook, however, if I use the  option of the  function, the notebook starts to freeze. I am not sure why this is the case. I have to wait a considerable amount of time and maybe after a while it's coming back but most of the time I simply kill it. For now I don't use it at all but I'd loke to. Any idea what the problem could be?<\/p>\n\n<p>Example using :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42162955","CreationDate":"2017-02-10T15:26:02.470","ParentId":null,"OwnerUserId":"3427404","Title":null,"Body":"<p>This issue has been pointed out before: <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/5165\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/pull\/5165<\/a><\/p>\n\n<p>As far as I can tell, the issue is with data accumulation in iPython, so it's pretty much unavoidable if you are running many epochs and long epochs.<\/p>\n\n<p>Either decrease the update rate in <code>keras\/utils\/generic_utils.py<\/code>, or set <code>verbosity=0<\/code><\/p>\n"}]}
{"QuestionId":42160431,"AnswerCount":0,"Tags":"<tensorflow><distributed>","CreationDate":"2017-02-10T13:21:33.123","AcceptedAnswerId":null,"OwnerUserId":4394807.0,"Title":"Distributed TensorFlow: run multi-gpu on each machine: ran out of memory error occurred","Body":"<p>I'm running <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">inception distributed training code<\/a>. Since the official version only supports single gpu on each machine, I modified the code to use multi-gpu, multi-machines. <\/p>\n\n<p>After looking into the <a href=\"https:\/\/github.com\/tensorflow\/models\/issues\/54#issuecomment-274722215\" rel=\"nofollow noreferrer\">Issue 54<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/39991238\/distributed-tensorflow-with-multiple-gpu\">distribute tensorflow with multiple gpu<\/a>, I found the most convenient way is to treat each gpu as a , which can be easily implemented with the help of . Here I use <a href=\"https:\/\/github.com\/ZhuFengdaaa\/models\/commit\/57b3afaf7c811c3d13548eab4cedd202add20016#diff-2bbdfc8464d99bf28f8fbf62126fbf50R114\" rel=\"nofollow noreferrer\">ZhuFengdaaa's codes<\/a> (thanks the author!). The main logic is adding  and using . <\/p>\n\n<p>Specifically, I use two machines:  and , each of which has 4 GPUs (Titan X with 12 GB memory) and use 2 GPUs in each machine, resulting overall 4 workers. I also use those two machines as parameter servers. The TensorFlow version I'm using is . The scripts I used for training are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Then I encountered the following  error on  and  and the program crashes:<\/p>\n\n<pre><\/pre>\n\n<p>I checked the usage of GPU memories on each machine using . I found even when I use only one gpu on each machine (in this case, program runs well), all four gpus' memories in each machine are almost fully occupied, i.e. reaching ~12GB. Based on this observation, I wonder the  error is due to there is not enough space left for running one more worker on each machine. I have tried to make the program uses as much GPU memory based on runtime allocations by setting , but the error remains.<\/p>\n\n<p>Could anyone help me on this problem? Thanks in advance.<\/p>\n","answers":[]}
{"QuestionId":42160731,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-10T13:37:30.697","AcceptedAnswerId":42161303.0,"OwnerUserId":null,"Title":"Tensorfow: the best way to convert a scalar to a 1-dimensional tensor","Body":"<p>I have to convert tensorflow scalars into 1-dimensional tensors with 1 element. \nI wonder what is the best way to do this?<\/p>\n","answers":[{"AnswerId":"42161303","CreationDate":"2017-02-10T14:03:58.180","ParentId":null,"OwnerUserId":"6770703","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\"><code>tf.reshape<\/code><\/a><\/p>\n\n<p>For example:<\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.constant(2)\nx_1d = tf.reshape(x, [1])\n<\/code><\/pre>\n"}]}
{"QuestionId":42160782,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-02-10T13:39:42.933","AcceptedAnswerId":42161461.0,"OwnerUserId":2191652.0,"Title":"Tensorflow: How to subtract a mean image?","Body":"<p>I'm looking for a code snippet on how to subtract a mean image from an input image. I have the image mean as a numpy array:<\/p>\n\n<pre><\/pre>\n\n<p>I guess the subtraction part will be something like <\/p>\n\n<pre><\/pre>\n\n<p>but how to create the  with the same shape as img?<\/p>\n","answers":[{"AnswerId":"42161461","CreationDate":"2017-02-10T14:11:56.130","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>assume img is all one matrix, size is (100, 100, 3)\nrgb mean is [1, 2, 3]<\/p>\n\n<pre><code>img = tf.ones([100, 100, 3], dtype=tf.float32) #(100, 100, 3)\nmean = tf.constant([1, 2, 3], dtype=tf.float32) # (3)\nmean = tf.reshape(mean, [1, 1, 3])\n\nimg_m = img - mean\n\nwith tf.Session() as sess:\n    a = sess.run(img_m) # shape of a (100, 100, 3)\n    # a[:,:,0] = 0\n    # a[:,:,1] = -1\n    # a[:,:,1] = -2\n<\/code><\/pre>\n"},{"AnswerId":"42161575","CreationDate":"2017-02-10T14:17:52.727","ParentId":null,"OwnerUserId":"6770703","Title":null,"Body":"<p><code>img = tf.subtract(img, mean_img)<\/code> is correct as long as the last dimension of <code>img<\/code> (<code>tf.shape(img)[-1]<\/code>) is 3. <\/p>\n\n<p>You can check <a href=\"https:\/\/www.tensorflow.org\/resources\/glossary\" rel=\"nofollow noreferrer\">Broadcasting operation in tensorflow Glossary<\/a> and <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">numpy Broadcasting<\/a> for further information.<\/p>\n\n<blockquote>\n  <p>how to create the mean_image with the same shape as img?<\/p>\n<\/blockquote>\n\n<p>You can try:<\/p>\n\n<pre><code>tf.reshape(tf.tile(m, [tf.reduce_prod(tf.shape(img)[:-1])]),\n           tf.shape(img)).eval()\n<\/code><\/pre>\n"},{"AnswerId":"44526767","CreationDate":"2017-06-13T16:17:41.807","ParentId":null,"OwnerUserId":"7886651","Title":null,"Body":"<p>It would be better to use <code>tf.image.per_image_standardization<\/code> as in: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/per_image_standardization\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/per_image_standardization<\/a>. <\/p>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":42161103,"AnswerCount":0,"Tags":"<python><video><tensorflow><sequence><recurrent-neural-network>","CreationDate":"2017-02-10T13:55:23.717","AcceptedAnswerId":null,"OwnerUserId":2706744.0,"Title":"How to feed sequence of images to a tf.placeholder in tensorflow?","Body":"<p>I am trying to make an RNN that can classify sequences of images. I load my data into an ndarray which has the shape<\/p>\n\n<pre><\/pre>\n\n<p>and I have 2 classes,  and .<\/p>\n\n<p>I create a placeholder to feed the data to during training and I reshape it somewhat:<\/p>\n\n<pre><\/pre>\n\n<p>For now I set  and for each training step I get the next sample as:<\/p>\n\n<pre><\/pre>\n\n<p>And then I run <\/p>\n\n<pre><\/pre>\n\n<p>And I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I tried fixing this by casting  into a  which is an immutable datatype, but I still get the same error even though everything is a  when I debug:<\/p>\n\n<pre><\/pre>\n\n<p>I also read <a href=\"https:\/\/stackoverflow.com\/questions\/33684657\/issue-feeding-a-list-into-feed-dict-in-tensorflow\">this<\/a> and tried it, which gave me other issues regarding dimensions. When I fix the dimension issues it throws the exact same errors as before. I think I am doing something wrong, but I have no idea what :\\<\/p>\n","answers":[]}
{"QuestionId":42161861,"AnswerCount":1,"Tags":"<ubuntu><caffe><amd>","CreationDate":"2017-02-10T14:32:47.117","AcceptedAnswerId":null,"OwnerUserId":7441967.0,"Title":"can I install \"Caffe , Deep Learning\" without dependencies with a Nvida GPU","Body":"<p>I want to install caffe on my laptop with AMD GPU ON UBUNTU 16 and all installation I find is an installation with NVIDA dependencies !\nI have tried to install NVIDA driver but I end up with losing my ubuntu OS \nis there any other solution please !!<\/p>\n","answers":[{"AnswerId":"42205744","CreationDate":"2017-02-13T13:54:53.293","ParentId":null,"OwnerUserId":"7441262","Title":null,"Body":"<p>From the official <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html\" rel=\"nofollow noreferrer\">Caffe installation documentation<\/a> :<\/p>\n\n<blockquote>\n  <p>CPU-only Caffe: for cold-brewed CPU-only Caffe uncomment the CPU_ONLY\n  := 1 flag in Makefile.config to configure and build Caffe without\n  CUDA. This is helpful for cloud or cluster deployment.<\/p>\n<\/blockquote>\n\n<p>This will allow you to compile Caffe without any GPU dependency.<\/p>\n\n<p>Alternatively, there is an experimental <a href=\"https:\/\/github.com\/amd\/OpenCL-caffe\" rel=\"nofollow noreferrer\">Open-CL Caffe fork<\/a> developed by AMD.<\/p>\n"}]}
{"QuestionId":42161916,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-02-10T14:35:36.467","AcceptedAnswerId":42162483.0,"OwnerUserId":2191652.0,"Title":"Tensorflow: How to switch channels of a tensor from RGB to BGR?","Body":"<p>Is there a quick way to convert tensors which encode an image from RGB to BGR?<\/p>\n\n<p>In python that would be something like:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44075469","CreationDate":"2017-05-19T17:05:58.870","ParentId":null,"OwnerUserId":"1351629","Title":null,"Body":"<p>You can use <code>tf.split<\/code> as well.<br>\nSnippet from <a href=\"https:\/\/github.com\/ry\/tensorflow-vgg16\/blob\/master\/vgg16.py#L5\" rel=\"nofollow noreferrer\">tensorflow-vgg16\/blob\/master\/vgg16.py#L5<\/a>:<\/p>\n\n<pre><code>red, green, blue = tf.split(3, 3, rgb_scaled)\nbgr = tf.concat(3, [blue, green, red])\n<\/code><\/pre>\n"},{"AnswerId":"43124498","CreationDate":"2017-03-30T17:22:41.147","ParentId":null,"OwnerUserId":"1760830","Title":null,"Body":"<p>Assuming the channel is the last dimension.<\/p>\n\n<pre><code>channels = tf.unstack (image, axis=-1)\nimage    = tf.stack   ([channels[2], channels[1], channels[0]], axis=-1)\n<\/code><\/pre>\n"},{"AnswerId":"42162483","CreationDate":"2017-02-10T15:02:56.633","ParentId":null,"OwnerUserId":"6770703","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#strided_slice\" rel=\"nofollow noreferrer\"><code>tf.strided_slice<\/code><\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.0\/api_docs\/python\/array_ops\/slicing_and_joining#reverse\" rel=\"nofollow noreferrer\"><code>tf.reverse<\/code><\/a>.<\/p>\n\n<p>For example,<\/p>\n\n<pre><code>import tensorflow as tf\n\nimg = tf.reshape(tf.range(30), [2, 5, 3])\n\n# the following two lines produce equivalent result:\nimg_channel_swap = img[..., ::-1]\nimg_channel_swap_1 = tf.reverse(img, axis=[-1])\n<\/code><\/pre>\n\n<p>Note that the api of <code>tf.reverse<\/code> is changed from tensorflow r1.0.<\/p>\n"}]}
{"QuestionId":42161960,"AnswerCount":2,"Tags":"<image-processing><neural-network><deep-learning><caffe><convolution>","CreationDate":"2017-02-10T14:37:58.837","AcceptedAnswerId":42163494.0,"OwnerUserId":1965914.0,"Title":"Caffe Copy PreTrained Weights of AlexNet to custom network that has Two AlexNets","Body":"<p>I am trying to build a network that contains two image inputs. Each image will go through a network simultaneous with late fusion that will merge and give one outputs. I use a diagram below to show what I need (ps: sorry my english is not so good)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/z6Mwn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/z6Mwn.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My network defined in caffe prototxt model definition file which contains exact AlexNet defined twice upto the Pool 5. For first network the layers has same name as in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/models\/bvlc_alexnet\/train_val.prototxt\" rel=\"nofollow noreferrer\">AlexNet<\/a> while for the second net  I added a <em>\"_1\"<\/em> suffix to each layer name. \nMy question is how can i copy the pretained weight? <\/p>\n\n<p>For Eg: my convolution layer 1 of each network is as follows. Note that for  the pretrained weights can easily be copied since the layer name is same as one in pretrained model. However of  the same is different so I am afraid I cannot copy the pretrained weights? Or is there a way to do this even is the layer names are different?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42163494","CreationDate":"2017-02-10T15:53:45.613","ParentId":null,"OwnerUserId":"1484481","Title":null,"Body":"<p>I presume that you are trying to finetune the whole network after initialization, otherwise you could simply use features extracted from AlexNet and start your training at the FC layer. For finetuning, you will need to copy the weights on the first network (one with the same names), and have second network share the weights with the first one. Take a look at <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/jc0Jx4JjWwo\" rel=\"nofollow noreferrer\">this<\/a> thread. Or rather <a href=\"https:\/\/groups.google.com\/forum\/#!msg\/caffe-users\/jc0Jx4JjWwo\/4yhRN6o5FgAJ\" rel=\"nofollow noreferrer\">this<\/a> reply from Evan Shelmar.<\/p>\n"},{"AnswerId":"42271449","CreationDate":"2017-02-16T10:37:58.653","ParentId":null,"OwnerUserId":"7396273","Title":null,"Body":"<p>I did something similar, here you can see the Siamese Network with Identical AlexNet. <a href=\"https:\/\/i.stack.imgur.com\/Q7QaJ.png\" rel=\"nofollow noreferrer\">Identical AlexNet for Siamese Network<\/a>. Here is <a href=\"https:\/\/www.dropbox.com\/s\/oej5c7sq5ti1f9x\/ferrmenta_siamese_alexnet.prototxt?dl=0\" rel=\"nofollow noreferrer\">prototxt file<\/a><\/p>\n"}]}
{"QuestionId":42162709,"AnswerCount":0,"Tags":"<tensorflow><theano><keras>","CreationDate":"2017-02-10T15:13:24.170","AcceptedAnswerId":null,"OwnerUserId":1245262.0,"Title":"Should Keras with Theano backend be over 18X slower than with Tensorflow backend?","Body":"<p>I've just installed keras, tensorflow and theano on my machine and did a quick comparison of keras using tensorflow as a backend to using theano as a backend. The results were more extreme than I expected.<\/p>\n\n<p>I am using the following versions of the 3 packages:<\/p>\n\n<pre><\/pre>\n\n<p>To compare the two backends, I used the  When I use tensorflow as my backend, I get this result:<\/p>\n\n<pre><\/pre>\n\n<p>and the whole thing takes just over  to run.<\/p>\n\n<p>When I use Theano as my backend, I get this result:<\/p>\n\n<pre><\/pre>\n\n<p>This takes about 15.75 hours to run !?<\/p>\n\n<p>Should I be surprised that the keras theano backend is about <strong><em>18X - 19X<\/em><\/strong> slower than the keras tensorflow backend? (I'm including the version of cifar10_cnn.py that I was using) IN convering from one backend to the other, I just changed the backend specification in keras.json. I didn't adjust the image_dim_ordering, since that seemed to be specified by the dataset.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42163697,"AnswerCount":1,"Tags":"<tensorflow><visualization><keras><backpropagation>","CreationDate":"2017-02-10T16:04:52.457","AcceptedAnswerId":null,"OwnerUserId":6900889.0,"Title":"Multi-input gradients calculation on keras with K.function","Body":"<h3>This model.input is a four head placeholder with size (3, 227, 227)<\/h3>\n\n<pre><\/pre>\n\n<h3>compute the gradient of the input picture wrt this loss<\/h3>\n\n<pre><\/pre>\n\n<h3>this function returns the loss and grads given the input picture<\/h3>\n\n<pre><\/pre>\n\n<p>I want to compute gradients for a siamese NN with fout head inputs, but when I apply the above K function I get a tensor of gradient with shape , when what I want is .<\/p>\n\n<p>What can I do to get the correct gradients?<\/p>\n","answers":[{"AnswerId":"42166967","CreationDate":"2017-02-10T19:11:50.587","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>I see you are trying to generate saliency maps. <\/p>\n\n<p>Check whats the shape of input_imgs tensor. Thats what determines the shape your gradient. My guess will be <code>input_imgs<\/code> is a list of tensors and in the <code>grads = K.gradients(loss, input_imgs)[0]<\/code> since you are indexing to <code>0<\/code> you are dropping all other gradients in the list. So probably this will fix your issue:<\/p>\n\n<pre><code>grads = K.gradients(loss, input_imgs)\niterate = K.function(input_imgs, grads)\n<\/code><\/pre>\n\n<p>Now when you use <code>iterate<\/code> it will give you a list of 4 each of shape <code>(1, 3, 227, 227)<\/code><\/p>\n"}]}
{"QuestionId":42164355,"AnswerCount":0,"Tags":"<python><django><tensorflow>","CreationDate":"2017-02-10T16:37:35.840","AcceptedAnswerId":null,"OwnerUserId":1849805.0,"Title":"TensorFlow: error in use of tf.app.flags","Body":"<p>I faced the same duplicate problem as <a href=\"https:\/\/stackoverflow.com\/questions\/38267771\/error-in-use-of-tf-app-flags\">error in use of tf.app.flags<\/a>, but cannot understand how to solve<\/p>\n\n<pre><\/pre>\n\n<p>On server i see this problem, but in local PC all is working. I use Django 1.8.12, trying to invoke neural model from view.\nthis is how it working on local PC:<\/p>\n\n<pre><\/pre>\n\n<p>this is how i tried to set model dir path on server:<\/p>\n\n<pre><\/pre>\n\n<p>but this line throw error<\/p>\n","answers":[]}
{"QuestionId":42164741,"AnswerCount":0,"Tags":"<tensorflow><profiling>","CreationDate":"2017-02-10T16:59:36.933","AcceptedAnswerId":null,"OwnerUserId":6618884.0,"Title":"Specify ops for profiling","Body":"<p><a href=\"https:\/\/stackoverflow.com\/questions\/34293714\/can-i-measure-the-execution-time-of-individual-operations-with-tensorflow\">Here<\/a> excellent example how to profile Tensorflow computation graph, but output is very clutter, there is way to specify which <strong>op<\/strong> to trace?<\/p>\n","answers":[]}
{"QuestionId":42164772,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-10T17:00:55.463","AcceptedAnswerId":42169408.0,"OwnerUserId":5184894.0,"Title":"Tensorflow Estimator API: Summaries","Body":"<p>I can't achieve to make summaries work with the Estimator API of Tensorflow.<\/p>\n\n<p>The Estimator class is very useful for many reasons: I have already implemented my own classes which are really similar but I am trying to switch to this one.<\/p>\n\n<p>Here is the code sample:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see, I can create an Estimator and use it but I can achieve to add hooks to the fitting process.<\/p>\n\n<p>The logging hooks works just fine but the others require both <strong>tensors<\/strong> and a <strong>saver<\/strong> which I can't provide.<\/p>\n\n<p>The tensors are defined in the model function, thus I can't pass them to the <strong>SummaryHook<\/strong> and the <strong>Saver<\/strong> can't be initialized because there is no tensor to save...<\/p>\n\n<p>Is there a solution to my problem? (I am guessing yes but there is a lack of documentation of this part in the tensorflow documentation)<\/p>\n\n<ul>\n<li>How can I initialized my <strong>saver<\/strong>? Or should I use other objects such as <em>Scaffold<\/em>?<\/li>\n<li>How can I pass <strong>summaries<\/strong> to the <strong>SummaryHook<\/strong> since they are defined in my model function?<\/li>\n<\/ul>\n\n<p>Thanks in advance.<\/p>\n\n<p><em>PS: I have seen the DNNClassifier API but I want to use the estimator API for Convolutional Nets and others. I need to create summaries for any estimator.<\/em><\/p>\n","answers":[{"AnswerId":"42169408","CreationDate":"2017-02-10T22:03:50.883","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>The intended use case is that you let the Estimator save summaries for you. There are options in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/run_config.py#L182\" rel=\"noreferrer\">RunConfig<\/a> for configuring summary writing. RunConfigs get passed when <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py#L347\" rel=\"noreferrer\">constructing the Estimator<\/a>.<\/p>\n"},{"AnswerId":"50930835","CreationDate":"2018-06-19T14:26:50.823","ParentId":null,"OwnerUserId":"6370088","Title":null,"Body":"<p>Just have <code>tf.summary.scalar(\"loss\", loss)<\/code> in the <code>model_fn<\/code>, and run the code without <code>summary_hook<\/code>. The loss is recorded and shown in the tensorboard.<\/p>\n\n<hr>\n\n<p>See also:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/45086109\/tensorflow-using-tf-summary-with-1-2-estimator-api\">Tensorflow - Using tf.summary with 1.2 Estimator API<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":42164983,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-10T17:10:52.150","AcceptedAnswerId":null,"OwnerUserId":3678911.0,"Title":"Running a tensorflow program multiple times each time afresh","Body":"<p>I have the following tensorflow code in a script named to fit to a given data. I reuse the variables  and  for the whole sequence length. <\/p>\n\n<pre><\/pre>\n\n<p>I am calling this script from a different script called  to collect the  for several runs:<\/p>\n\n<pre><\/pre>\n\n<p>However I am receiving this following error after I called the script once:<\/p>\n\n<pre><\/pre>\n\n<p>I think the issue is because it's using the same graph as in iteration  even for iteration . How do I make sure that the entire script  is called afresh each time I call ?<\/p>\n","answers":[]}
{"QuestionId":42165043,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-10T17:14:08.040","AcceptedAnswerId":null,"OwnerUserId":987397.0,"Title":"Tensorflow: is this normal behaviour of Batch normalization?","Body":"<p>See the screenshot. Without batch normalization is the blue line. Accuracy in the upper plot, loss in the lower one. So without BN the loss decreases slowly and the accuracy increases slowly, expected behaviour.<\/p>\n\n<p>But then I try batch normalization. Training loss converges quite fast to values near zero. So much better than without, but testing accuracy on both train and test sets give worse results.\n<a href=\"https:\/\/i.stack.imgur.com\/9ZaS8.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9ZaS8.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I based my implementation on this: <a href=\"http:\/\/ruishu.io\/2016\/12\/27\/batchnorm\/\" rel=\"noreferrer\">http:\/\/ruishu.io\/2016\/12\/27\/batchnorm\/<\/a>\nSo adding batch norm to a layer like this:<\/p>\n\n<pre><\/pre>\n\n<p>with model.is_training a placeholder that is set to zero when testing accuracy (upper plot). <\/p>\n\n<p>I do this as well:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas or suggestions?<\/p>\n","answers":[]}
{"QuestionId":42166475,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-10T18:41:29.870","AcceptedAnswerId":42166591.0,"OwnerUserId":573082.0,"Title":"How to reuse model in Tensorflow","Body":"<p>I'm new to TensorFlow. I have the following graph. The test accuracy that I get is 90%. I'd like to reuse the model. One way that I figured it out is to initiate my variables from the learned weights (look in REUSE_MODEL below). But, when I run the test dataset through the model I get now accuracy of 2.0%.<\/p>\n\n<p>What is the problem in the way I'm doing it and what's best way to do it?<\/p>\n\n<p><strong>GRAPH BUILD AND RUN<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>REUSE MODEL<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42166591","CreationDate":"2017-02-10T18:48:21.253","ParentId":null,"OwnerUserId":"6471140","Title":null,"Body":"<p>I think the correct way is to save and restore the metagraph ,this is the official documentation on this:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/exporting_and_importing_meta_graphs\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/exporting_and_importing_meta_graphs<\/a><\/p>\n"}]}
{"QuestionId":42168599,"AnswerCount":1,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2017-02-10T21:00:27.277","AcceptedAnswerId":null,"OwnerUserId":7547765.0,"Title":"TFLearn produces same results on every prediction","Body":"<p>My network produces the same output for every prediction. I have about 49,000 data samples in a pandas data frame.<br>\nHow can I fix this?  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42255060","CreationDate":"2017-02-15T16:35:05.970","ParentId":null,"OwnerUserId":"7096238","Title":null,"Body":"<p>Try using sigmoid or relu instead of softmax. I got better predictions with those 2. Maybe you want to use sigmoid for the first and relu for the second layer. Just play with them and combine them, so that you have better predictions. Also try other loss functions.<\/p>\n"}]}
{"QuestionId":42168977,"AnswerCount":1,"Tags":"<tensorflow><protocol-buffers>","CreationDate":"2017-02-10T21:31:01.530","AcceptedAnswerId":null,"OwnerUserId":3002273.0,"Title":"TensorFlow Manual Construction of GraphDef","Body":"<p>In TensorFlow, I found that I can do the following,<\/p>\n\n<pre><\/pre>\n\n<p>Which prints<\/p>\n\n<pre><\/pre>\n\n<p>Now, my node is not actually set to an Operation. I can't figure out how to make node_def be an operation. I can construct OpDefs via,<\/p>\n\n<pre><\/pre>\n\n<p>Which is of class <\/p>\n\n<p>I would like to make my NodeDef register as this OpDef.<\/p>\n\n<p>EDIT:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"51883395","CreationDate":"2018-08-16T18:35:31.927","ParentId":null,"OwnerUserId":"1497311","Title":null,"Body":"<p>I believe you are looking for a way to set message fields inside of the <code>GraphDef<\/code> and <code>NodeDef<\/code> protos. How to modify ProtoBuf objects in Python is detailed <a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/reference\/python-generated\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
{"QuestionId":42169300,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-10T21:54:16.660","AcceptedAnswerId":null,"OwnerUserId":860340.0,"Title":"Tensorflow - Choose TFRecord Features at Graph Build Tim","Body":"<p>Let's say I have TFRecord Examples with the following feature mapping: <\/p>\n\n<pre><\/pre>\n\n<p>I always need  and , but sometimes I want 0,1, or more of the remaining features.  I know what features I want at Graph build time. <\/p>\n\n<p>How can I choose multiple features at graph build time? <\/p>\n\n<p>For instance, sentence and labels, no problem: <\/p>\n\n<pre><\/pre>\n\n<p>But can I pull out multiple features? I.e.:<\/p>\n\n<pre><\/pre>\n\n<p>I'm willing to change my TFRecord representation as well.  Any help would be appreciated.  <\/p>\n\n<p>Thanks! <\/p>\n","answers":[{"AnswerId":"42207054","CreationDate":"2017-02-13T14:58:07.793","ParentId":null,"OwnerUserId":"860340","Title":null,"Body":"<p>Thanks to @Allen Lavoie for pointing out that it's a simple solution.  I needed to put them in a list, and then according to this post <a href=\"https:\/\/stackoverflow.com\/questions\/37697747\/typeerror-tensor-object-does-not-support-item-assignment-in-tensorflow\">here<\/a>, I just needed to tf.pack into a list.  Below is the solution. <\/p>\n\n<pre><code># get features\nFEATURES = ['labels', 'caps']\noutput_list = [] \nfor f in range(len(FEATURES)):\n    feats = parsed[FEATURES[f]]\n    dense_feats = tf.sparse_tensor_to_dense(feats, default_value='&lt;PAD&gt;')\n    output_list.append(dense_feats)\n\nfeatures = tf.pack(output_list)\n<\/code><\/pre>\n"}]}
{"QuestionId":42169421,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-10T22:05:13.230","AcceptedAnswerId":null,"OwnerUserId":7241513.0,"Title":"Read mixed data types from CSV row via tf.TextLineReader and tf.decode_csv","Body":"<p>I want to read more than just numeric features and labels from each line of the input CSV data file. My features and labels are both numeric, but I also want to use a date-like or string-like identifier for each line so I can refer back to the line after a random-shuffle batch read.<\/p>\n\n<p>Since TensorFlow requires a  array that I pre-populate programmatically, how can I make TF read mixed types in one-line? If I pre-populate  with floats, I get an error for cells that contain dates\/strings. If I pre-populate  with strings, with the view of casting the numeric parts later into floats, I get an error that TF doesn't support string-to-float casting.<\/p>\n\n<p>The docs state:<\/p>\n\n<p><em>A list of Tensor objects with types from: float32, int32, int64, string. One tensor per column of the input record, with either a scalar default value for that column or empty if the column is required<\/em><\/p>\n\n<p>As per above, how can I specify 'empty' if I just want all columns as required and skip the whole defaults headache? And in that case how would TF know what data types to pick automatically for each cell? I have lots of columns, so it's not practical to write it out one-by-one as is done in the one example in the docs:<\/p>\n\n<pre><\/pre>\n\n<p>(I have hundreds of cells!)<\/p>\n\n<p>My code looks something like this:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for any input!<\/p>\n","answers":[]}
{"QuestionId":42169776,"AnswerCount":1,"Tags":"<python><numpy><matrix><theano>","CreationDate":"2017-02-10T22:38:22.307","AcceptedAnswerId":42194662.0,"OwnerUserId":3615378.0,"Title":"numpy.matmul in Theano","Body":"<p><strong>TL;DR<\/strong><br>\nI want to replicate the functionality of  in . What's the best way to do this?<\/p>\n\n<p><strong>Too Short; Didn't Understand<\/strong><br>\nLooking at  and , I'm not seeing an easy way to do a straightforward batch matrix multiplication. i.e. treat the last two dimensions of N dimensional tensors as matrices, and multiply them. Do I need to resort to some goofy usage of ? Or *shudder* loop them myself without broadcasting!?<\/p>\n","answers":[{"AnswerId":"42194662","CreationDate":"2017-02-12T23:50:31.627","ParentId":null,"OwnerUserId":"3615378","Title":null,"Body":"<p>The current pull requests don't support broadcasting, so I came up with this for now. I may clean it up, add a little more functionality, and submit my own PR as a temporary solution. Until then, I hope this helps someone!\nI included the test to show it replicates numpy.matmul, given that the input complies with my more strict (temporary) assertions.<\/p>\n\n<p>Also, .scan stops iterating the sequences at <code>argmin(*sequencelengths)<\/code> iterations. So, I believe that mismatched array shapes won't raise any exceptions.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import theano as th\nimport theano.tensor as tt\nimport numpy as np\n\n\ndef matmul(a: tt.TensorType, b: tt.TensorType, _left=False):\n    \"\"\"Replicates the functionality of numpy.matmul, except that\n    the two tensors must have the same number of dimensions, and their ndim must exceed 1.\"\"\"\n\n    # TODO ensure that broadcastability is maintained if both a and b are broadcastable on a dim.\n\n    assert a.ndim == b.ndim  # TODO support broadcasting for differing ndims.\n    ndim = a.ndim\n    assert ndim &gt;= 2\n\n    # If we should left multiply, just swap references.\n    if _left:\n        tmp = a\n        a = b\n        b = tmp\n\n    # If a and b are 2 dimensional, compute their matrix product.\n    if ndim == 2:\n        return tt.dot(a, b)\n    # If they are larger...\n    else:\n        # If a is broadcastable but b is not.\n        if a.broadcastable[0] and not b.broadcastable[0]:\n            # Scan b, but hold a steady.\n            # Because b will be passed in as a, we need to left multiply to maintain\n            #  matrix orientation.\n            output, _ = th.scan(matmul, sequences=[b], non_sequences=[a[0], 1])\n        # If b is broadcastable but a is not.\n        elif b.broadcastable[0] and not a.broadcastable[0]:\n            # Scan a, but hold b steady.\n            output, _ = th.scan(matmul, sequences=[a], non_sequences=[b[0]])\n        # If neither dimension is broadcastable or they both are.\n        else:\n            # Scan through the sequences, assuming the shape for this dimension is equal.\n            output, _ = th.scan(matmul, sequences=[a, b])\n        return output\n\n\ndef matmul_test() -&gt; bool:\n    vlist = []\n    flist = []\n    ndlist = []\n    for i in range(2, 30):\n        dims = int(np.random.random() * 4 + 2)\n\n        # Create a tuple of tensors with potentially different broadcastability.\n        vs = tuple(\n            tt.TensorVariable(\n                tt.TensorType('float64',\n                              tuple((p &lt; .3) for p in np.random.ranf(dims-2))\n                              # Make full matrices\n                              + (False, False)\n                )\n            )\n            for _ in range(2)\n        )\n        vs = tuple(tt.swapaxes(v, -2, -1) if j % 2 == 0 else v for j, v in enumerate(vs))\n\n        f = th.function([*vs], [matmul(*vs)])\n\n        # Create the default shape for the test ndarrays\n        defshape = tuple(int(np.random.random() * 5 + 1) for _ in range(dims))\n        # Create a test array matching the broadcastability of each v, for each v.\n        nds = tuple(\n            np.random.ranf(\n                tuple(s if not v.broadcastable[j] else 1 for j, s in enumerate(defshape))\n            )\n            for v in vs\n        )\n        nds = tuple(np.swapaxes(nd, -2, -1) if j % 2 == 0 else nd for j, nd in enumerate(nds))\n\n        ndlist.append(nds)\n        vlist.append(vs)\n        flist.append(f)\n\n    for i in range(len(ndlist)):\n        assert np.allclose(flist[i](*ndlist[i]), np.matmul(*ndlist[i]))\n\n    return True\n\n\nif __name__ == \"__main__\":\n    print(\"matmul_test -&gt; \" + str(matmul_test()))\n<\/code><\/pre>\n"}]}
{"QuestionId":42169830,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-10T22:43:16.230","AcceptedAnswerId":null,"OwnerUserId":1214630.0,"Title":"How to save two different checkpoints during training in Tensorflow","Body":"<p>Assume we have test\/val\/train splits. During training, we want to save some model checkpoints [save_1] that can be used to restart the training later.<\/p>\n\n<p>In addition, we want to save another model during training that shows the best performance on the validation sets [save_2]. After done with training, we use save_2 to report the performance on the test data. <\/p>\n\n<p>My question is that how we can have two different tf.savers during training in TensorFlow? whatever examples that I have seen, only save [save_1].\nPointer to any codes would be appreciated.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"42170949","CreationDate":"2017-02-11T00:48:20.990","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>You can get quite close by using an <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py#L984\" rel=\"nofollow noreferrer\">Estimator<\/a> to wrap your model. Specifically see the options surrounding saving multiple checkpoints in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/run_config.py#L182\" rel=\"nofollow noreferrer\">RunConfig<\/a> (you don't have to throw any away). Could be combined with a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/monitors.py#L609\" rel=\"nofollow noreferrer\">ValidationMonitor<\/a> to find the lowest validation error.<\/p>\n"}]}
{"QuestionId":42170998,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2017-02-11T00:56:54.450","AcceptedAnswerId":44296777.0,"OwnerUserId":330561.0,"Title":"Tensorflow embedding_lookup gradients registering on CPU?","Body":"<p>I have something equivalent to a sparse softmax:<\/p>\n\n<pre><\/pre>\n\n<p>However, when I enable placement logging, I see multiple instances of the gradients being placed on the CPU, e.g.:<\/p>\n\n<pre><\/pre>\n\n<p>This happens no matter the optimizer I choose. Am I missing something here?<\/p>\n","answers":[{"AnswerId":"44296777","CreationDate":"2017-06-01T00:46:33.130","ParentId":null,"OwnerUserId":"1937197","Title":null,"Body":"<p>If you use<\/p>\n\n<pre><code>tf.Session(config=tf.ConfigProto(allow_soft_placement=False))\n<\/code><\/pre>\n\n<p>you should get an error. That's because <code>embedding_lookup<\/code> isn't implemented on the GPU currently.<\/p>\n"}]}
{"QuestionId":42171299,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-11T01:46:28.020","AcceptedAnswerId":null,"OwnerUserId":957011.0,"Title":"Why does tf.nn.seq2seq.embedding_attention_decoder use tf.nn.embeddin_lookup without tf.device('\/cpu:0')?","Body":"<p>From the following commit,  uses  without .<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/8cc567bf9703e14f0e16595eb9f220859a5eff20#diff-cf94d06c60c56f954b1926fb6e53cb11L86\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/commit\/8cc567bf9703e14f0e16595eb9f220859a5eff20#diff-cf94d06c60c56f954b1926fb6e53cb11L86<\/a><\/p>\n\n<p>Why  no longer use ?\nIs  currently implemented on the GPU?<\/p>\n","answers":[]}
{"QuestionId":42172072,"AnswerCount":1,"Tags":"<go><tensorflow>","CreationDate":"2017-02-11T04:00:48.560","AcceptedAnswerId":null,"OwnerUserId":2330178.0,"Title":"Error Building Go Bindings for Tensorflow","Body":"<p>I am trying to get the Go bindings for Tensorflow going. I've followed the QuickStart instructions from the repo. <\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/go\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/go<\/a><\/p>\n\n<p>However the build command:<\/p>\n\n<p><\/p>\n\n<p>fails with the following error:<\/p>\n\n<pre><\/pre>\n\n<p>The lib has been downloaded and installed in \/usr\/local\/lib<\/p>\n\n<pre><\/pre>\n\n<p>It appears to all be setup as per instructions. Not sure how else to indicate to the Go build where to find the library. Anyone else got the Go wrapper to build?<\/p>\n","answers":[{"AnswerId":"42173093","CreationDate":"2017-02-11T06:36:19.483","ParentId":null,"OwnerUserId":"2330178","Title":null,"Body":"<p>Found the solution.<\/p>\n\n<p>On both Linux and OSX to get it to build and test I needed to explicitly set the library paths.<\/p>\n\n<p>On an Ubuntu VM I spun up for testing build:<\/p>\n\n<pre><code>~\/work$ export LD_LIBRARY_PATH=\/usr\/local\/lib\n~\/work$ export LIBRARY_PATH=\/usr\/local\/lib\n~\/work$ go get github.com\/tensorflow\/tensorflow\/tensorflow\/go\n~\/work# go test github.com\/tensorflow\/tensorflow\/tensorflow\/go\nok github.com\/tensorflow\/tensorflow\/tensorflow\/go   0.210s\n~\/work$  \n<\/code><\/pre>\n\n<p>On my local OSX machine:<\/p>\n\n<pre><code>pendragon:go sjc$ export LIBRARY_PATH=$LIBRARY_PATH:\/usr\/local\/lib\npendragon:go sjc$ export DYLD_LIBRARY_PATH=$DYLD_LIBRARY_PATH:\/usr\/local\/lib\npendragon:go sjc$ go get github.com\/tensorflow\/tensorflow\/tensorflow\/go\npendragon:go sjc$ go test github.com\/tensorflow\/tensorflow\/tensorflow\/go\nok      github.com\/tensorflow\/tensorflow\/tensorflow\/go  0.237s\n<\/code><\/pre>\n\n<p>So it seems the instructions about setting those path environment variables seem to apply even if you have installed <code>libtensorflow<\/code> into <code>\/usr\/local\/lib<\/code>.<\/p>\n"}]}
{"QuestionId":42172489,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network><artificial-intelligence>","CreationDate":"2017-02-11T05:14:12.110","AcceptedAnswerId":42176675.0,"OwnerUserId":4093621.0,"Title":"Tensorflow doesn't converge for XOR","Body":"<p>I tried writing a simple XOR neural network but it never converges<\/p>\n\n<p>I create a NN with 2 inputs, 2 hidden nodes and 1 output.<\/p>\n\n<p>I use a relu on the first hidden layer and a softmax at the end to get the outputs.<\/p>\n\n<p>In theory it should learn how to solve it and converge?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42176675","CreationDate":"2017-02-11T13:54:26.917","ParentId":null,"OwnerUserId":"4093621","Title":null,"Body":"<p>The mistakes...<\/p>\n\n<ol>\n<li><p>The W2 weights should be between -1 and 1 since they're not using a ReLu. Also the first layer weights use ReLu so I set them as positive to try avoid dead neurons.<\/p><\/li>\n<li><p>Softmax makes no sense unless it's a 1 hot vector layer. Sigmoid makes more sense. Reading up on how Softmax works helps out.<\/p><\/li>\n<li><p>Reduce sum should be done on y not hidden2<\/p><\/li>\n<li><p><code>hidden2 = tf.matmul(hidden, W2 + b2)<\/code> has brackets incorrectly. Should be <code>hidden2 = tf.matmul(hidden, W2) + b2<\/code><\/p><\/li>\n<li><p>-Log as an error function only works if you're trying to get the output to 1 not 0. This is because -log(1) = 0, and infinity when -log(0). This encourages output to go to 1 but not 0. Good for 1 hot vector not if you're trying to push an output to 0 for one input and 1 for another.<\/p><\/li>\n<li><p>2 neurons in the hidden layer does work. But it is very susceptible to randomness in initialization. Using extra neurons (10 instead of 2) makes this less susceptible to initialization errors.<\/p><\/li>\n<\/ol>\n\n<p>The code below works. It uses a cost function that helps converge to 0 and to 1 for different inputs.<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\n\n# define placeholder for input, None as first argument means tensor can be any length\nx_ = tf.placeholder(tf.float32, shape=[4,2], name=\"x-input\")\ny_ = tf.placeholder(tf.float32, shape=[4,1], name=\"y-input\")\n\n# Configure weights and layers\nW = tf.Variable(tf.random_uniform([2, 10], 0.001, .01))\nb = tf.Variable(tf.zeros([10]))\nhidden  = tf.nn.relu(tf.matmul(x_,W) + b) # first layer.\n\nW2 = tf.Variable(tf.random_uniform([10,1], -1, 1))\nb2 = tf.Variable(tf.zeros([1]))\nhidden2 = tf.matmul(hidden, W2) + b2\ny = tf.nn.sigmoid(hidden2)\n\n# Training function + data\ncost = tf.reduce_mean(( (y_ * tf.log(y)) +\n((1 - y_) * tf.log(1.0 - y)) ) * -1)\ntrain_step = tf.train.GradientDescentOptimizer(0.01).minimize(cost)\n\nXOR_X = [[0,0],[0,1],[1,0],[1,1]]\nXOR_Y = [[0],[1],[1],[0]]\n\ninit = tf.global_variables_initializer()\nsess.run(init)\n# Train on the input data\nfor i in range(100000):\n    sess.run(train_step, feed_dict={x_: XOR_X, y_: XOR_Y})\n    if i % 2000 == 0:\n        print ('W1', sess.run(W))\n        print('Output ', sess.run(y, feed_dict={x_: XOR_X, y_: XOR_Y}))\n<\/code><\/pre>\n"}]}
{"QuestionId":42172871,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2017-02-11T06:06:52.390","AcceptedAnswerId":42182253.0,"OwnerUserId":5009875.0,"Title":"Explain silence layer in caffe","Body":"<pre><\/pre>\n\n<p>It just sets the diff to zero.\nWhat is the use of this layer?<\/p>\n","answers":[{"AnswerId":"42182253","CreationDate":"2017-02-11T23:01:24.507","ParentId":null,"OwnerUserId":"2878130","Title":null,"Body":"<p>The use of this layer is simply to avoid that the output of unused blobs is reported in the log. Being an output manager layer, it is obviously zero its gradient.<\/p>\n\n<p>For instance, let us assume we are using AlexNet and we change the bottom of the 'fc7' layer to 'pool5' instead of 'fc6'. If we do not delete the 'fc6' blob declaration, this layer is not used anymore but its ouput will be printed in stderr: it is considered as an output of the whole architecture. If we want to keep 'fc6' for some reasons, but without showing its values, we can use the 'SilenceLayer'. <\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/tutorial\/layers\/silence.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/tutorial\/layers\/silence.html<\/a><\/p>\n\n<p>See also <a href=\"http:\/\/caffe.help\/manual\/layers\/silence.html\" rel=\"nofollow noreferrer\">caffe.help<\/a>.<\/p>\n"}]}
{"QuestionId":42172967,"AnswerCount":2,"Tags":"<python><artificial-intelligence><deep-learning><caffe>","CreationDate":"2017-02-11T06:20:26.307","AcceptedAnswerId":null,"OwnerUserId":3886530.0,"Title":"How to preprocess image before feeding into LeNet CNN in Caffe?","Body":"<p>I am having issue in feeding my own image into LeNet using Caffe library. I have deployed and initialised the weight obtained through training with no difficulties. As the net is trained using input size of 28x28, I tried resizing the input image to 28x28 and feed into the deployed LeNet, but it gave me <em>\"unhashable numpy array\"<\/em> error.<\/p>\n\n<p>Not only that, I also tried to transpose it with  after resizing but it gave me <em>\"TypeError: only length-1 arrays can be converted to Python scalars\"<\/em><\/p>\n\n<p>Below are the python codes I tried so far in pre-processing my image :<\/p>\n\n<pre><\/pre>\n\n<p>I am a beginner in using Caffe and still in the process of learning. I hope someone can give me some example or insights in how to pre-process an image before feeding into the net.<\/p>\n\n<p>Thank You.<\/p>\n\n<p>Best regards.<\/p>\n","answers":[{"AnswerId":"42240895","CreationDate":"2017-02-15T04:42:50.743","ParentId":null,"OwnerUserId":"5009875","Title":null,"Body":"<p>You can use <code>caffe.io.Transformer<\/code>\nThis is used to preprocess caffe 'data' blob.\nDefine it as <\/p>\n\n<pre><code>transformer = caffe.io.Transformer({'data':net.blobs['data'].data.shape})\ntransformer.set_transpose('data',(2,0,1))\n<\/code><\/pre>\n\n<p>then you can,<\/p>\n\n<pre><code>img = caffe.io.load_image('number5.png')\nimg = caffe.io.resize_image(img, (28,28), interp_order=3)\nimg_transposed=transformer.preprocess('data',img)\n<\/code><\/pre>\n"},{"AnswerId":"42179445","CreationDate":"2017-02-11T18:13:36.033","ParentId":null,"OwnerUserId":"7550917","Title":null,"Body":"<p>Just use this instead:<\/p>\n\n<pre><code>img = img.transpose((2,0,1))\n<\/code><\/pre>\n"}]}
{"QuestionId":42173255,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network>","CreationDate":"2017-02-11T07:01:29.867","AcceptedAnswerId":42173927.0,"OwnerUserId":4910036.0,"Title":"How does batching interact with the loss function in TensorFlow?","Body":"<p>I'm training a multi-objective neural net in TensorFlow with my own loss function and can't find documentation regarding how batching interacts with that functionality.<\/p>\n\n<p>For example, I have snippet of my loss function below, which takes the tensor\/list of predictions and makes sure that their absolute value sums to no more than one:<\/p>\n\n<pre><\/pre>\n\n<p>But when I'm passing in a batch of estimates I feel like this loss function is normalising the whole set of inputs to sum to 1 at this point, rather than each individual set summing to 1. I.e.<br>\n[[.8,.8],[.8,.8]] -> [[.25,.25],[.25,25]]<br>\nrather than the desired<br>\n[[.8,.8],[.8,.8]] -> [[.5,.5],[.5,.5]]<\/p>\n\n<p>Can anybody clarify or put to rest my suspicions? If this is how my function is currently working, how do I change that?<\/p>\n","answers":[{"AnswerId":"42173927","CreationDate":"2017-02-11T08:38:20.507","ParentId":null,"OwnerUserId":"2130551","Title":null,"Body":"<p>You must specify a reduction axis for reduction ops, otherwise all axes will be reduces. Traditionally this is the first dimension of your tensor. So, line 2 should look like this:<\/p>\n\n<pre><code>sumTensor = tf.reduce_sum(absTensor, 0)\n<\/code><\/pre>\n\n<p>After you make that change you will run into another problem. <code>sumTensor<\/code> will no longer be a scalar and will thus no longer make sense as a condition for tf.cond (i.e. what does it mean to branch per entry of a batch?). What you really want is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/comparison_operators#select\" rel=\"nofollow noreferrer\">tf.select<\/a> since you don't really want to branch logic per batch entry. Like this:<\/p>\n\n<pre><code>isGTOne = tf.greater(sumTensor,oneTensor)\n\nnorm = predictions\/sumTensor\n\npredictions = tf.select(isGTOne,norm,predictions)\n<\/code><\/pre>\n\n<p>But, looking at this now, I wouldn't even bother conditionally normalizing the entries. Since you are operating at the granularity of a batch now, I don't think you can gain performance from normalizing an entry of a batch one at a time. Especially, since dividing is not really an expensive side effect. Might as well just do:<\/p>\n\n<pre><code>def fitness(predictions,actual):\n\n  absTensor = tf.abs(predictions)\n  sumTensor = tf.reduce_sum(absTensor, 0)\n\n  predictions = predictions\/sumTensor\n\n  etc...\n<\/code><\/pre>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":42173704,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><keras><cross-validation><tflearn>","CreationDate":"2017-02-11T08:03:24.353","AcceptedAnswerId":null,"OwnerUserId":562769.0,"Title":"How do I make a cross-validation with tflearn?","Body":"<p>I want to make k-fold cross-validation with tflearn. Hence I want to reset the network k times. However, I think I need to reset the graph (e.g. ), but I'm not sure and I don't know how to do this with tflearn.<\/p>\n\n<h2>My code<\/h2>\n\n<p>For the following, you need <a href=\"https:\/\/github.com\/MartinThoma\/HASY\/blob\/master\/hasy_tools.py\" rel=\"nofollow noreferrer\"><\/a><\/p>\n\n<pre><\/pre>\n\n<h2>The error<\/h2>\n\n<p>Running the code with only one fold works fine, but with multiple folds I get:<\/p>\n\n<pre><\/pre>\n\n<p>edit: I get the same problem with Keras.<\/p>\n","answers":[{"AnswerId":"57670380","CreationDate":"2019-08-27T08:14:36.327","ParentId":null,"OwnerUserId":"11982604","Title":null,"Body":"<p>I put in my code <code>tf.reset_default_graph()<\/code> right before building the neural network, before the <code>for<\/code> loop.<\/p>\n\n<p>You should also define your model outside the for loop, right before it. Only the training should be in the <code>kfolds<\/code> loop.<\/p>\n"}]}
{"QuestionId":42174411,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><matrix-factorization>","CreationDate":"2017-02-11T09:41:50.737","AcceptedAnswerId":42185121.0,"OwnerUserId":7421488.0,"Title":"memory overflow when processing eval() in Tensorflow","Body":"<p>I'm using Tensorflow to process a simple matrix factorization algorithm. Every step went correct but at the last step, where I want to  a Tensor to store it, the program didn't work and only occupied more and more memory. So is there something wrong in my code? I'm a beginner in Tensorflow and I'm don't know where the problem is. Below is my code.<\/p>\n\n<pre><\/pre>\n\n<p>The program stopped and occupying memory at the last two lines in , which contains  and . So what should I do to fix it? Can someone help?<\/p>\n","answers":[{"AnswerId":"42185121","CreationDate":"2017-02-12T07:16:47.603","ParentId":null,"OwnerUserId":"7421488","Title":null,"Body":"<p>Solved. You cannot add iteration of symbolic operations in Tensorflow. Instead, you should build the data flow first, which means you should define your operations in the initialize step.<\/p>\n"}]}
{"QuestionId":42174875,"AnswerCount":1,"Tags":"<python><django><tensorflow><subprocess>","CreationDate":"2017-02-11T10:37:16.413","AcceptedAnswerId":null,"OwnerUserId":1849805.0,"Title":"Cannot get subprocess output in Django view","Body":"<p>I want to emulate command line call of my script (TensorFlow neaural chatbot model) in Django view and get output from console to variable.<\/p>\n\n<p>When i do manually in terminal of my server:<\/p>\n\n<pre><\/pre>\n\n<p>the output is good and printed out.<\/p>\n\n<p>So in my Django view i do:<\/p>\n\n<pre><\/pre>\n\n<p>And  variable is printed as  in Apache error log.<\/p>\n\n<p>I cannot figure out what's wrong in my call.<\/p>\n","answers":[{"AnswerId":"42175093","CreationDate":"2017-02-11T11:03:58.713","ParentId":null,"OwnerUserId":"1849805","Title":null,"Body":"<p>The answer is to use .communicate() and PIPE:<\/p>\n\n<pre><code>        from subprocess import Popen, PIPE\n        proc = Popen(\n            \"python3 var\/www\/engine\/chatbot\/udc_predict.py --model_dir=var\/www\/engine\/chatbot\/runs\/1486057482\/\",\n            shell=True,\n            stdout=PIPE, stderr=PIPE\n        )\n        proc.wait()    \n        res = proc.communicate() \n        if proc.returncode:\n            print(res[1])\n        print('result:', res[0])\n        answer = res[0]\n<\/code><\/pre>\n"}]}
{"QuestionId":42175033,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-11T10:56:15.750","AcceptedAnswerId":null,"OwnerUserId":2492058.0,"Title":"Scale-out performance limitations for scatter-gather pattern in Tensorflow","Body":"<p>Scale-out performance for scatter-gather dataflow pipelines is limited, and I am not sure if this is<\/p>\n\n<ul>\n<li>a bug in Tensorflow (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7415#event-957548310\" rel=\"nofollow noreferrer\">I was told that it wasn't<\/a>)<\/li>\n<li>an underperformant way to design a graph for a distributed session<\/li>\n<\/ul>\n\n<p>I use Tensorflow to build custom pipelines (i.e. I write my own OpKernels) of \"embarrassingly parallel\" problems (no coordination between \"local\" pipelines required). Typically these involve the local pipeline being replicated across all the machines in my cluster, and having a source and sink queue to feed input and receive output, respectively.<\/p>\n\n<p>There is <em>no<\/em> coordination between the pipelines on individual machines, except for<\/p>\n\n<ul>\n<li>dequeuing input from the \"source\" queue<\/li>\n<li>enqueuing output to the \"sink\" queue<\/li>\n<\/ul>\n\n<p>As the exact codebase I use is still closed source, I created a <a href=\"https:\/\/gist.github.com\/samwhitlock\/3b099f8489909a3293b3e4239a7188d0\" rel=\"nofollow noreferrer\">similar example in vanilla Tensorflow<\/a>. I installed via pip (v0.12.1).<\/p>\n\n<p>This script replicates my \"local\" pipeline by a while loop that induces a certain delay. The efficiency of this while loop is besides the point. The point is that the speedup I see when I scale this out across machines is not very good. <a href=\"https:\/\/docs.google.com\/spreadsheets\/d\/1Ht8O9OT4csfYVOxSDybojXnZ-_uV2A60CF9Jt0jhps0\/edit#gid=492424317\" rel=\"nofollow noreferrer\">Here is the data I collected on a cluster of 10 machines<\/a> (only goes up to 9 because the machine hosting the source\/sink queues is \"unloaded\").<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8Lome.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8Lome.png\" alt=\"scale-out performance of nodes\"><\/a><\/p>\n\n<p>The ideal time is flat because the workload scales up linearly with the number of nodes (and should therefore take a constant time regardless of the number of nodes).<\/p>\n\n<p>My question is<\/p>\n\n<ul>\n<li>Is this a known limitation of Tensorflow? I currently sidestep this by using ZeroMQ to coordinate between local nodes (i.e. I <em>don't<\/em> use the distributed session), and I know others who use MPI or other such non-TF systems.<\/li>\n<li>Is there a way to make this scatter-gather pattern faster by some primitive in Tensorflow that I'm currently not using, or perhaps by a way to construct graphs to require less coordination when executing?<\/li>\n<\/ul>\n","answers":[]}
{"QuestionId":42175232,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2017-02-11T11:18:37.323","AcceptedAnswerId":42175397.0,"OwnerUserId":6404493.0,"Title":"Cost at all epochs are zero, even before training","Body":"<p>I wrote simple multi-layer perceptron program using TensorFlow.\nThis program was made to predict following number after 5 sequence. (e.g. 1 4 9 14 19 [24]) Yes, it is very simple.<\/p>\n\n<p>But I wander to death during at least 4 hours. Because cost at all epochs are zero, even if what I do. Surprisingly, I ensured that weights and biases are initialized to non-zero (using ), it didn't be helped.<\/p>\n\n<p>How can I not see zero-valued cost, anymore?<\/p>\n\n<p><strong>Code<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Console output<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42175397","CreationDate":"2017-02-11T11:35:45.357","ParentId":null,"OwnerUserId":"6791223","Title":null,"Body":"<p>The problem is you are using loss function for classification (softmax is generally used for classification) while your network could produce an arbitrary single real number, so it's regression, not classification. Use proper cost (say, mean squared error) and your network will start to converge. <\/p>\n\n<p>In this particular case, just change this line:<\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y))\n<\/code><\/pre>\n\n<p>to something like this:<\/p>\n\n<pre><code>cost = tf.reduce_mean(tf.squared_difference(y, pred))\n<\/code><\/pre>\n"}]}
{"QuestionId":42175360,"AnswerCount":3,"Tags":"<python><python-3.x><caffe><pycaffe>","CreationDate":"2017-02-11T11:32:12.890","AcceptedAnswerId":null,"OwnerUserId":4663495.0,"Title":"How to install caffe for python3 in ubuntu","Body":"<p>I followed the steps on <a href=\"http:\/\/caffe.berkeleyvision.org\/installation.html#compilation\" rel=\"nofollow noreferrer\">caffe<\/a> and changed the configure file:<\/p>\n\n<pre><\/pre>\n\n<p>Then did:<\/p>\n\n<pre><\/pre>\n\n<p>These run OK. But when I run:<\/p>\n\n<pre><\/pre>\n\n<p>I get an error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I solve this problem?<\/p>\n","answers":[{"AnswerId":"49020079","CreationDate":"2018-02-28T00:19:35.810","ParentId":null,"OwnerUserId":"7305537","Title":null,"Body":"<p>All these libs should already have been configured in your Make.config file<\/p>\n\n<pre><code>$ sudo apt-get install libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler\n$ sudo apt-get install -y python-numpy python-scipy\n$ sudo apt-get install -y python3-dev\n$ sudo apt-get install -y python3-numpy python3-scipy\n$ sudo apt-get install -y libopencv-dev\n$ sudo apt-get install libboost-all-dev \n$ apt-cache search gflags\n$ sudo apt-get install libgflags2 libgflags-dev\n$ sudo apt-get install libgflags libgflags-dev\n$ sudo apt-get install libgflags-dev\n$ sudo apt-get install libgflags-dev libgoogle-glog-dev liblmdb-dev\n$ apt-cache search openblas\n$ sudo apt-get install libopenblas-dev\n<\/code><\/pre>\n\n<p>Also refer: <br>\n1. <a href=\"https:\/\/gist.github.com\/arundasan91\/b432cb011d1c45b65222d0fac5f9232c\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/arundasan91\/b432cb011d1c45b65222d0fac5f9232c<\/a> <br>\n2. <a href=\"https:\/\/askubuntu.com\/questions\/629654\/building-caffe-failed-to-see-hdf5-h\">https:\/\/askubuntu.com\/questions\/629654\/building-caffe-failed-to-see-hdf5-h<\/a> <br>\n3. <a href=\"https:\/\/gist.github.com\/victoriastuart\/fb2cb22209ccb2771963a25c06221213\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/victoriastuart\/fb2cb22209ccb2771963a25c06221213<\/a><\/p>\n"},{"AnswerId":"42179374","CreationDate":"2017-02-11T18:07:40.840","ParentId":null,"OwnerUserId":"7550917","Title":null,"Body":"<p>Probably you have to link against a different python_boost version. <\/p>\n\n<p>In the Makefile find the line:<\/p>\n\n<pre><code>PYTHON_LIBRARIES ?= boost_python....\n<\/code><\/pre>\n\n<p>Probably you have to change that to <code>boost_python-py35<\/code>:<\/p>\n\n<pre><code>PYTHON_LIBRARIES ?= boost_python-py35 python3.5m\n<\/code><\/pre>\n\n<p>Look in \/usr\/lib\/x86_64-linux-gnu for the libboost_python files or wherever your libs are located:<\/p>\n\n<pre><code>find \/usr\/lib\/x86_64-linux-gnu\/ -name libboost_python*so\nfind \/usr\/lib\/x86_64-linux-gnu\/ -name libpython*so\n<\/code><\/pre>\n\n<p>will show you the libraries you can choose from. (If file is libboost_python-py35.so write boost_python-py35) The boost and python libs have to match  in version.<\/p>\n"},{"AnswerId":"45563730","CreationDate":"2017-08-08T08:52:37.207","ParentId":null,"OwnerUserId":"7290990","Title":null,"Body":"<p>Linker looks for libboost_python3, but in different systems this lib can have different names, i.e. ubuntu 16.04 have boost_python-py35 so you may create symlink <\/p>\n"}]}
{"QuestionId":42175609,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-11T12:00:52.487","AcceptedAnswerId":42177088.0,"OwnerUserId":7549930.0,"Title":"Using multiple input pipelines in TensorFlow","Body":"<p>I know how to use an input pipeline to read data from files:<\/p>\n\n<pre><\/pre>\n\n<p>But how can I switch between multiple input pipelines? Say, if I want to train a network for 1000 batches on the training set from the training pipeline, then validate it on a validation set from another pipeline, then keep training, then validate, then train, ..., and so forth.<\/p>\n\n<p>It's easy to implement this with feed_dict. I also know how to use checkpoints to achieve this, just like in the cifar-10 example. But it's kind of cumbersome: I need to dump the model to disk then read it from disk again.<\/p>\n\n<p>Can I just switch between two input pipelines (one for training data, one for validation data) to achieve this? Reading 1000 batches from the training data queue, then a few batched from the validation data queue, and so forth. If it is possible, how to do it?<\/p>\n","answers":[{"AnswerId":"42177088","CreationDate":"2017-02-11T14:35:53.037","ParentId":null,"OwnerUserId":"7241513","Title":null,"Body":"<p>Not sure if this is exactly what you are looking for, but I am doing training and validation in the same code in two separate loops. My code reads numeric and string data from .CSV files, not images. I am reading from two separate CSV files, one for training and one for validation. I'm sure you can generalize it to read from two 'sets' of files, rather than just single files, as the code is there.<\/p>\n\n<p>Here are the code snippets in case it helps. Note that this code first reads everything as string and then converts the necessary cells into floats, just given my own requirements. If your data is purely numeric, you should just set the defaults to floats and all should be easier. Also, there are a couple of lines in there that drop Weights and Biases into a CSV file AND serialize them into the TF checkpoint file, depending on which way you'd prefer.<\/p>\n\n<pre><code>        #first define the defaults:\n        rDefaults = [['a'] for row in range((TD+TS+TL))]\n\n    # this function reads line-by-line from CSV and separates cells into chunks:   \n        def read_from_csv(filename_queue):\n            reader = tf.TextLineReader(skip_header_lines=False)\n            _, csv_row = reader.read(filename_queue)\n            data = tf.decode_csv(csv_row, record_defaults=rDefaults)\n            dateLbl = tf.slice(data, [0], [TD])\n            features = tf.string_to_number(tf.slice(data, [TD], [TS]), tf.float32)\n            label = tf.string_to_number(tf.slice(data, [TD+TS], [TL]), tf.float32)\n            return dateLbl, features, label\n\n    #this function loads the above lines and spits them out as batches of N:\n        def input_pipeline(fName, batch_size, num_epochs=None):\n            filename_queue = tf.train.string_input_producer(\n                [fName],\n                num_epochs=num_epochs,\n                shuffle=True)  \n            dateLbl, features, label = read_from_csv(filename_queue)\n            min_after_dequeue = 10000 \n            capacity = min_after_dequeue + 3 * batch_size # max of how much to load into memory\n            dateLbl_batch, feature_batch, label_batch = tf.train.shuffle_batch(\n                [dateLbl, features, label], \n                batch_size=batch_size,\n                capacity=capacity,\n                min_after_dequeue=min_after_dequeue)\n            return dateLbl_batch, feature_batch, label_batch\n\n    # These are the TRAINING features, labels, and meta-data to be loaded from the train file:    \n        dateLbl, features, labels = input_pipeline(fileNameTrain, batch_size, try_epochs)\n    # These are the TESTING features, labels, and meta-data to be loaded from the test file:\n        dateLblTest, featuresTest, labelsTest = input_pipeline(fileNameTest, batch_size, 1) # 1 epoch here regardless of training\n\n    # then you define the model, start the session, blah blah    \n\n    # fire up the queue:        \n            coord = tf.train.Coordinator()\n            threads = tf.train.start_queue_runners(coord=coord)\n\n    #This is the TRAINING loop:\ntry:            \n                while not coord.should_stop():\n\n                    dateLbl_batch, feature_batch, label_batch = sess.run([dateLbl, features, labels])      \n\n                   _, acc, summary = sess.run([train_step, accuracyTrain, merged_summary_op], feed_dict={x: feature_batch, y_: label_batch, \n                keep_prob: dropout, \n                learning_rate: lRate})\n\n            except tf.errors.OutOfRangeError: # (so done reading the file(s))\n\n    # by the way, this dumps weights and biases into a CSV file, since you asked for that\n                np.savetxt(fPath + fIndex + '_weights.csv', sess.run(W), \n    # and this serializes weight and biases into the TF-formatted protobuf:\n        #        tf.train.Saver({'varW': W, 'varB': b}).save(sess, fileNameCheck)\n\n            finally:\n                coord.request_stop()\n\n    # now re-start the runners for the testing file:   \n            coord = tf.train.Coordinator()\n            threads = tf.train.start_queue_runners(coord=coord)\n\n            try:\n\n                while not coord.should_stop():\n    # so now this line reads features, labels, and meta-data, but this time from the training file:                \n                    dateLbl_batch, feature_batch, label_batch = sess.run([dateLblTest, featuresTest, labelsTest])\n\n                    guessY = tf.argmax(y, 1).eval({x: feature_batch, keep_prob: 1})\n                    trueY = tf.argmax(label_batch, 1).eval()\n\n                    accuracy = round(tf.reduce_mean(tf.cast(tf.equal(guessY, trueY), tf.float32)).eval(), 2)\n\n            except tf.errors.OutOfRangeError:\n                acCumTest \/= i\n            finally:\n                coord.request_stop()\n\n\n            coord.join(threads)\n<\/code><\/pre>\n\n<p>This may differ from what you are trying to do in the sense that it first completes the Training loop and THEN restarts the queues for the Testing loop. Not sure how you'd do this if you want to go back and fourth, but you can try to experiment with the two functions defined above by passing them the relevant file names (or lists) interchangeably.<\/p>\n\n<p>Also I'm not sure if re-starting the queues after training is the best way to go, but it works for me. Would love to see a better example out there, as most TF examples use some built-in wrappers around the MNIST dataset to do the training in one go...<\/p>\n"}]}
{"QuestionId":42175635,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-11T12:03:36.870","AcceptedAnswerId":42178448.0,"OwnerUserId":1683542.0,"Title":"Gather elements along second dimension of tensor","Body":"<p>Assume  and tensor  both have shape . Now if we think of them in terms of matrices, I would like for each row of  to get the row element corresponding to the index where  has it's maximum. I can easily find those indices with <\/p>\n\n<pre><\/pre>\n\n<p>which returns a tensor of shape . Now, how can I gather up these indices from  such that I get something of shape ? I tried <\/p>\n\n<pre><\/pre>\n\n<p>but it doesn't do the right thing - it returns something of shape  which means that it didn't gather up anything.<\/p>\n","answers":[{"AnswerId":"42178448","CreationDate":"2017-02-11T16:43:57.220","ParentId":null,"OwnerUserId":"6770703","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#gather_nd\" rel=\"nofollow noreferrer\">tf.gather_nd<\/a>.<\/p>\n\n<p>For example,<\/p>\n\n<pre><code>import tensorflow as tf\n\nsess = tf.InteractiveSession()\n\nvalues = tf.constant([[0, 0, 0, 1],\n                      [0, 1, 0, 0],\n                      [0, 0, 1, 0]])\n\nT = tf.constant([[0, 1, 2 ,  3],\n                 [4, 5, 6 ,  7],\n                 [8, 9, 10, 11]])\n\nmax_indices = tf.argmax(values, axis=1)\n# If T.get_shape()[0] is None, you can replace it with tf.shape(T)[0].\nresult = tf.gather_nd(T, tf.stack((tf.range(T.get_shape()[0], \n                                            dtype=max_indices.dtype),\n                                   max_indices),\n                                  axis=1))\n\nprint(result.eval())\n<\/code><\/pre>\n\n<p>However when the ranks of <code>values<\/code> and <code>T<\/code> are higher, the use of <code>tf.gather_nd<\/code> will be a little awkward. I posted my current solution on <a href=\"https:\/\/stackoverflow.com\/questions\/41897212\/how-to-sort-a-multi-dimensional-tensor-using-the-returned-indices-of-tf-nn-top-k\">this question<\/a>. There might be a better solution in case of high dimensional <code>values<\/code> and <code>T<\/code>.<\/p>\n"}]}
{"QuestionId":42176122,"AnswerCount":0,"Tags":"<theano>","CreationDate":"2017-02-11T12:56:59.463","AcceptedAnswerId":null,"OwnerUserId":4603642.0,"Title":"Selectively disabling optimizations in Theano","Body":"<p>Is it possible to \"protect\" expressions from optimizations in Theano?<\/p>\n\n<p>In a few places I'm relying on the fact that floating point numbers are not associative. For instance, I'm computing  and I don't want Theano to simplify it as .<\/p>\n\n<p>Is this possible without disabling optimizations entirely?<\/p>\n","answers":[]}
{"QuestionId":42176186,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><chatbot>","CreationDate":"2017-02-11T13:04:05.623","AcceptedAnswerId":null,"OwnerUserId":7291516.0,"Title":"Tensorflow models error","Body":"<p>I am testing some chatbots from Github, I seem to get the following error whenever I try running chatbots using tensorflow.<\/p>\n\n<p>after typing <\/p>\n\n<pre><\/pre>\n\n<p>Source <a href=\"https:\/\/github.com\/AhmedAbdalazeem\/ChatBot\" rel=\"nofollow noreferrer\">https:\/\/github.com\/AhmedAbdalazeem\/ChatBot<\/a><\/p>\n","answers":[{"AnswerId":"42660554","CreationDate":"2017-03-07T23:54:24.653","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>You need to install the <a href=\"https:\/\/github.com\/tensorflow\/models\" rel=\"nofollow noreferrer\">tensorflow.models repository<\/a>.<\/p>\n"}]}
{"QuestionId":42176824,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-02-11T14:08:07.860","AcceptedAnswerId":null,"OwnerUserId":3659972.0,"Title":"How to save neural network model with .ckpt extension instead of.meta extnsion in tensorflow using python?","Body":"<p>I am new to tensorflow and python. I am trying to save my trained neural network model with  extension. For that purpose, I have used code below:<\/p>\n\n<pre><\/pre>\n\n<p>But this saves model with  extenssion. Can anyone help me save model with  extension? <\/p>\n","answers":[{"AnswerId":"42177664","CreationDate":"2017-02-11T15:32:49.743","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>Your code is fine: you don't need to save the checkpoint with the <code>.ckpt<\/code> extension.<\/p>\n\n<p>The <code>.ckpt<\/code> extension alone was used in the version 1 (V1) of the checkpoint files.\nFrom Tensorflow <code>0.12<\/code> the checkpoint V1 can be read, but when you save a new checkpoint it will use the V2 format, where the name you choose is just a prefix for the <code>.meta<\/code> and <code>.index<\/code> files.<\/p>\n\n<p>If you really need to write a <code>.ckpt<\/code> file you have to downgrade your Tensorflow installation to something <code>&lt;0.12<\/code><\/p>\n"}]}
{"QuestionId":42176873,"AnswerCount":1,"Tags":"<tensorflow><visualization>","CreationDate":"2017-02-11T14:13:57.893","AcceptedAnswerId":null,"OwnerUserId":3174413.0,"Title":"Where is the source to embedding-projector-standalone?","Body":"<p>I'm able to run the Tensorflow Projector as a <a href=\"https:\/\/github.com\/tensorflow\/embedding-projector-standalone\" rel=\"noreferrer\">standalone 3d viewer<\/a>. There is some very minimal documentation <a href=\"https:\/\/www.tensorflow.org\/how_tos\/embedding_viz\/\" rel=\"noreferrer\">here<\/a>. Running standalone is as simple as<\/p>\n\n<pre><\/pre>\n\n<p>However,<\/p>\n\n<ol>\n<li>I don't see any source code in this repo. How is <a href=\"https:\/\/github.com\/tensorflow\/embedding-projector-standalone\/blob\/master\/oss_demo_bin.js\" rel=\"noreferrer\">oss_demo_bin.js<\/a> generated, and how can I fork it?<\/li>\n<li>I don't understand the tensors () file format which are associated with each labels () file. What is the spec for the  files in the <a href=\"https:\/\/github.com\/tensorflow\/embedding-projector-standalone\/tree\/master\/oss_data\" rel=\"noreferrer\">sample data directory<\/a>?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"59041237","CreationDate":"2019-11-25T22:48:50.367","ParentId":null,"OwnerUserId":"12432870","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/tree\/master\/tensorboard\/plugins\/projector\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/tree\/master\/tensorboard\/plugins\/projector<\/a><\/p>\n\n<p>This can launch the standalone, so I assume it's possible to also build it from this.<\/p>\n"}]}
{"QuestionId":42177658,"AnswerCount":8,"Tags":"<backend><theano><keras>","CreationDate":"2017-02-11T15:32:24.477","AcceptedAnswerId":42177775.0,"OwnerUserId":1426385.0,"Title":"How to switch Backend with Keras (from TensorFlow to Theano)","Body":"<p>I tried to switch Backend with Keras (from TensorFlow to Theano) but did not manage.\nI followed the temps described <a href=\"https:\/\/keras.io\/backend\/\" rel=\"noreferrer\">here<\/a> but it doesn't work. I created a keras.json in the keras' directory (as it did not exist) but it doesn't change anything when I import it from Python.<\/p>\n","answers":[{"AnswerId":"47806100","CreationDate":"2017-12-14T04:49:47.273","ParentId":null,"OwnerUserId":"5783902","Title":null,"Body":"\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras import backend as K\nfrom os import environ\n\n# user defined function to change keras backend\ndef set_keras_backend(backend):\n    if K.backend() != backend:\n       environ['KERAS_BACKEND'] = backend\n       reload(K)\n       assert K.backend() == backend\n\n# call the function with \"theano\"\nset_keras_backend(\"theano\")\n<\/code><\/pre>\n"},{"AnswerId":"48607089","CreationDate":"2018-02-04T10:26:21.160","ParentId":null,"OwnerUserId":"662770","Title":null,"Body":"<p>If you're using windows you can run from command line:<\/p>\n\n<p><code>set \"KERAS_BACKEND=theano\"<\/code><\/p>\n"},{"AnswerId":"48330757","CreationDate":"2018-01-18T21:40:02.040","ParentId":null,"OwnerUserId":"2464597","Title":null,"Body":"\n\n<p>I had an issue where I could not <code>from keras import backend<\/code> at all until I set the backend to theano. The provided answers should work if you can import backend, but if not, just use:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nos.environ['KERAS_BACKEND'] = 'theano'\nimport keras as ks\n# Using Theano backend.\n<\/code><\/pre>\n"},{"AnswerId":"44446831","CreationDate":"2017-06-08T23:00:02.337","ParentId":null,"OwnerUserId":"2608199","Title":null,"Body":"\n\n<p>In case you want to change the config permanently, the json is available here: <code>~\/.keras\/keras.json<\/code> and you can change the backend.<\/p>\n\n<p>To do this dynamically in python 2.7 you can run:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>from keras import backend as K\nimport os\n\ndef set_keras_backend(backend):\n\n    if K.backend() != backend:\n        os.environ['KERAS_BACKEND'] = backend\n        reload(K)\n        assert K.backend() == backend\n\nset_keras_backend(\"theano\")\n<\/code><\/pre>\n"},{"AnswerId":"45795798","CreationDate":"2017-08-21T11:20:13.943","ParentId":null,"OwnerUserId":"7657597","Title":null,"Body":"<p>Type following on command prompt and press enter:<\/p>\n\n<pre><code>%USERPROFILE%\/.keras\/keras.json\n<\/code><\/pre>\n\n<p>Change backend in the opened text file and save it. You are done. <\/p>\n"},{"AnswerId":"45911427","CreationDate":"2017-08-28T03:27:38.003","ParentId":null,"OwnerUserId":"8328970","Title":null,"Body":"<p>For Linux systems, the hidden .keras directory will be created in the user\u2019s home directory. To observe whether or not it has been created, run the following command from your home directory (the -a allows you to see hidden files and directories).<\/p>\n\n<pre><code>ls \u2013a \n<\/code><\/pre>\n\n<p>If the directory is there, then cd into it and modify the keras.json file. If it\u2019s not there, then create the directory with<\/p>\n\n<pre><code>mkdir .keras\n<\/code><\/pre>\n\n<p>Then create the file with<\/p>\n\n<pre><code>touch keras.json \n<\/code><\/pre>\n\n<p>Then edit the file to make the config changes you referenced to change the backend engine to Theano.<\/p>\n\n<p>This process is covered fully in <a href=\"https:\/\/youtu.be\/sRF_-bwxru4\" rel=\"nofollow noreferrer\">this video<\/a>.<\/p>\n"},{"AnswerId":"42177775","CreationDate":"2017-02-11T15:43:38.783","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>Create a <code>.keras<\/code> (note the <code>.<\/code> in front) folder in you home directory and put the <code>keras.json<\/code> file there.<\/p>\n\n<p>For example, <code>\/home\/DaniPaniz\/.keras\/keras.json<\/code> (or <code>~\/.keras\/keras.json<\/code> in short) if you are on a UNIX like system (MacOS X, Linux, *BSD). On Windows you want to create the folder <code>%USERPROFILE%\/.keras<\/code> and put the JSON file there.<\/p>\n\n<p>Alternatively, you can also set the environment variable <code>KERAS_BACKEND<\/code>:<\/p>\n\n<pre><code>KERAS_BACKEND=theano python mymodel.py\n<\/code><\/pre>\n"},{"AnswerId":"47489718","CreationDate":"2017-11-25T19:23:50.740","ParentId":null,"OwnerUserId":"2323612","Title":null,"Body":"<p>In windows, you need to find .keras folder in your C drive. Most probably, it will be somewhere in <strong>C:\/users\/username\/<\/strong>. There you will find .keras folder, it contains a json file, <strong>keras.json<\/strong>, open it.\nYou will see:<\/p>\n\n<pre><code>{\n\u201cbackend\u201d: \u201ctensorflow\u201d,\n\u201cfloatx\u201d: \u201cfloat32\u201d,\n\u201cepsilon\u201d: 1e-07\n}\n<\/code><\/pre>\n\n<p>more or less.\n<strong>replace 'tensorflow' with 'theano'.<\/strong> and save the file.<\/p>\n"}]}
{"QuestionId":42178490,"AnswerCount":0,"Tags":"<tensorflow><machine-learning><deep-learning>","CreationDate":"2017-02-11T16:48:40.053","AcceptedAnswerId":null,"OwnerUserId":4489467.0,"Title":"Tensorflow Retrain the retrained model","Body":"<p>I am very new to Neural network and tensorflow, just starting on the retrain image tutorial. I have successfully completed the flower_photos training and i have 2 questions. <\/p>\n\n<p>1.) Is it a good\/bad idea to keep building upon a retrained model many times over and over? Or would it be a lot better to train a model fresh everytime? That leads to my second question<\/p>\n\n<p>2.) If it is ok to retrain a model over and over, for the retrain model tutorial in Tensorflow (Image_retraining), in the retrain.py would i simply replace the classify_image_graph_def.pb and imagenet_synset_to_human_label_map.txt with the one outputted from my retraining? But i see that there is also a imagenet_2012_challenge_label_map_proto.pbtxt, would i have to replace that one with something else? <\/p>\n\n<p>Thanks for your time<\/p>\n","answers":[]}
{"QuestionId":42179213,"AnswerCount":1,"Tags":"<neural-network><keras>","CreationDate":"2017-02-11T17:54:23.260","AcceptedAnswerId":42223240.0,"OwnerUserId":5695300.0,"Title":"How to use weights to manually predict data in neural network with karas","Body":"<p>I've done my best to follow online guides regarding the structure of neural networks, but I must be missing something fundamental. Given a set of trained weights along with their bias, I'd like to simply predict an input manually with those weights without using the predict method.<\/p>\n\n<p>Using MNIST images with keras I've attempted to manually edit my data to include an extra feature for the bias, however this effort seems to offer no better image accuracy than using no bias at all, and definitely far less accuracy than using the keras predict method. My code is below along with my attempt.<\/p>\n\n<p>Please note the two comments near the bottom for using the keras method prediction for an accurate image representation, and then my poor attempt from getting the weights manually and adding the bias.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42223240","CreationDate":"2017-02-14T10:08:53.280","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>The manual prediction calculation seems to be correct other than the missing activation function like <code>activation='relu'<\/code> after first layer and <code>activation='sigmoid'<\/code> in the final layer. <\/p>\n\n<p>Do the following changes to the manual prediction code and the prediction should work fine:<\/p>\n\n<pre><code>from scipy.stats import logistic\n\nweights=np.vstack([model.get_weights()[0],model.get_weights()[1]]) \nprediction=np.dot(x_train,weights) \n\nprediction[prediction&lt;0]=0              ### RELU after 1st layer\n\nprediction=np.hstack([prediction,np.zeros(prediction.shape[0]).reshape(prediction.shape[0],1)])\nprediction[:,-1]=1\nweights=np.vstack([model.get_weights()[2],model.get_weights()[3]])\nprediction=np.dot(prediction,weights)\n\nprediction=logistic.cdf(prediction)     ### Sigmoid after 2nd layer\n\nprediction=prediction.reshape(priorShape_x_train)\n<\/code><\/pre>\n"}]}
{"QuestionId":42179732,"AnswerCount":1,"Tags":"<tensorflow><gpu><keras>","CreationDate":"2017-02-11T18:39:13.393","AcceptedAnswerId":null,"OwnerUserId":381397.0,"Title":"How do I call model.predict with data stored in GPU in Keras","Body":"<p>I have built and trained a Sequential model.<\/p>\n\n<p>Now before each model.predict call I want to upload the data into GPU, do some operations and then call model.predict using the output stored in GPU without downloading to memory and handover to keras model for it to upload to gpu again.<\/p>\n\n<p>Edit:\nI would like to use opencv operations on the input image in gpu and use the output directly to call model.predict if possible.<\/p>\n","answers":[{"AnswerId":"42187259","CreationDate":"2017-02-12T11:55:47.367","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can easily achieve this by adding the operations as <code>Lambda<\/code> layers on the top of the model. <\/p>\n\n<p>Here is a very simple example.. you can extend from here:<\/p>\n\n<pre><code>import numpy as np\nfrom keras import backend as K\nfrom keras.models import Sequential, Model\nfrom keras.layers import Dense, Lambda, Input, merge\n\nX = np.random.random((1000,5))\nY = np.random.random((1000,1))\n\ninp = Input(shape = (5,))\nd1 = Dense(60, input_dim=5, init='normal', activation='relu')\nd2 = Dense(1, init='normal', activation='sigmoid')\nout = d2(d1(inp))\n\nmodel = Model(input=[inp], output=[out])\nmodel.compile(loss='binary_crossentropy', optimizer='adam')\nmodel.summary()\nmodel.fit(X, Y, nb_epoch=1)\n\n\n\nX1 = np.random.random((10,3))\nX2 = np.random.random((10,2))\n\ninp1 = Input(shape = (3,))\ninp2 = Input(shape = (2,))\n\np1 = Lambda(lambda x: K.sqrt(x))(inp1)\np2 = Lambda(lambda x: K.tf.exp(x))(inp2)\nmer = merge([p1, p2], mode='concat')\nout2 = d2(d1(mer))\n\nmodel2 = Model(input=[inp1, inp2], output=[out2])\nmodel2.summary()\n\nypred = model2.predict([X1, X2])\nprint ypred.shape\n<\/code><\/pre>\n\n<p>Here from <code>model.summary()<\/code> you can see both the models are sharing the upper layers so in essence is using the weights already learnt during the training of the first model<\/p>\n"}]}
{"QuestionId":42180348,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2017-02-11T19:36:08.137","AcceptedAnswerId":null,"OwnerUserId":7551239.0,"Title":"Keras - How to fit an object in model","Body":"<p>i'm trying to fit an object into the keras model,<\/p>\n\n<p>i have a structure similar to this (as input) with variable length:<\/p>\n\n<pre><\/pre>\n\n<p>as you can see Item_(N) can be empty or contain another object with a variable length, the content of category can also vary<\/p>\n\n<p>is it possible somehow to fit this object into keras to correctly train it?<\/p>\n\n<p>if a reshape is needed how i can not loose the \"Item_(N)\" key -> array since the (N) is not an index (it can be \"Alpha\", \"Bravo\" etc..) ?<\/p>\n","answers":[]}
{"QuestionId":42180906,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-02-11T20:31:32.183","AcceptedAnswerId":42181095.0,"OwnerUserId":826983.0,"Title":"How to feed a different Dropout p value after training?","Body":"<p>In Tensorflow I would have a  such that I can feed it to the network as required:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, I am not sure how to do this in Keras:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42181095","CreationDate":"2017-02-11T20:51:40.900","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>In keras <code>Dropout<\/code> layers behave differently in training and test phase, that is its only enabled in training phase. <\/p>\n\n<p>To use dropout in through training\/test phase you have to replace the Dropout layers with <code>Lambda<\/code> layers using <code>dropout<\/code> function from keras backend. <\/p>\n\n<pre><code>from keras.layers.core import Lambda\nfrom keras import backend as K\n\nmodel.add(Lambda(lambda x: K.dropout(x, level=0.5)))\n<\/code><\/pre>\n\n<p>For more reference check: <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1606\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":42182233,"AnswerCount":1,"Tags":"<tensorflow><neural-network><gradient-descent>","CreationDate":"2017-02-11T22:58:57.560","AcceptedAnswerId":null,"OwnerUserId":1825034.0,"Title":"The gradient of an output w.r.t network weights that holds another output constant","Body":"<p>Let's assume I have a simple MLP<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DxKhw.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DxKhw.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And I have a gradient of some loss function with respect to the output layer to get G = [0, -1] (that is, increasing the second output variable decreases the loss function). <\/p>\n\n<p>If I take the gradient of G with respect to my network parameters and apply a gradient decent weight update the second output variable should increase, but nothing is said about the first output variable, and a scaled application of the gradient will almost certainly change the output variable (be it increasing it or decreasing it)<\/p>\n\n<p>How can I modify my loss function, or any of the gradient calculations, to ensure that the first output does not change?<\/p>\n","answers":[{"AnswerId":"42290921","CreationDate":"2017-02-17T06:32:56.990","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"\n\n<p>Update: I misunderstood the question. This is the new answer.<\/p>\n\n<p>For this purpose, you need to update connections between the hidden layer and the second output unit only, while keep those between the hidden layer and the first output unit intact. <\/p>\n\n<p><strong>The first approach is to introduce two sets of variables<\/strong>: one for connections between the hidden layer and the first output unit, one for the rest. Then you can combine them using <code>tf.stack<\/code>, and pass a <code>var_list<\/code> to get the corresponding derivatives. It's like (Just for illustration. Not tested. Use with care):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>out1 = tf.matmul(hidden, W_h_to_out1) + b_h_to_out1\nout2 = tf.matmul(hidden, W_h_to_out2) + b_h_to_out2\nout = tf.stack([out1, out2])\nout = tf.transpose(tf.reshape(out, [2, -1]))\nloss = some_function_of(out)\noptimizer = tf.train.GradientDescentOptimizer(0.1)\ntrain_op_second_unit = optimizer.minimize(loss, var_list=[W_h_to_out2, b_h_to_out2])\n<\/code><\/pre>\n\n<p><strong>Another approach is to use a mask.<\/strong> This is easier to implement and more flexible when you work with some frameworks (say, slim, Keras, etc.), and I'll recommend this way. The idea to hide the first output unit to the loss function, while do not change the second output unit. <strong>This can be done using a binary variable: multiply something by 1 if you want to keep it, and multiply it by 0 to drop it.<\/strong> Here's the code:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\n# let's make our tiny dataset: (x, y) pairs, where x = (x1, x2, x3), y = (y1, y2),\n# and y1 = x1+x2+x3, y2 = x1^2+x2^2+x3^2\n\n# n_sample data points\nn_sample = 8\ndata_x = np.random.random((n_sample, 3))\ndata_y = np.zeros((n_sample, 2))\ndata_y[:, 0] += np.sum(data_x, axis=1)\ndata_y[:, 1] += np.sum(data_x**2, axis=1)\ndata_y += 0.01 * np.random.random((n_sample, 2))  # add some noise\n\n\n# build graph\n# suppose we have a network of shape [3, 4, 2], i.e.: one hidden layer of size 4.\n\nx = tf.placeholder(tf.float32, shape=[None, 3], name='x')\ny = tf.placeholder(tf.float32, shape=[None, 2], name='y')\nmask = tf.placeholder(tf.float32, shape=[None, 2], name='mask')\n\nW1 = tf.Variable(tf.random_normal(shape=[3, 4], stddev=0.1), name='W1')\nb1 = tf.Variable(tf.random_normal(shape=[4], stddev=0.1), name='b1')\nhidden = tf.nn.sigmoid(tf.matmul(x, W1) + b1)\nW2 = tf.Variable(tf.random_normal(shape=[4, 2], stddev=0.1), name='W2')\nb2 = tf.Variable(tf.random_normal(shape=[2], stddev=0.1), name='b2')\nout = tf.matmul(hidden, W2) + b2\nloss = tf.reduce_mean(tf.square(out - y))\n\n# multiply out by mask, thus out[0] is \"invisible\" to loss, and its gradient will not be propagated\nmasked_out = mask * out\nloss2 = tf.reduce_mean(tf.square(masked_out - y))\n\noptimizer = tf.train.GradientDescentOptimizer(0.1)\ntrain_op_all = optimizer.minimize(loss)  # update all variables in the network\ntrain_op12 = optimizer.minimize(loss, var_list=[W2, b2])  # update hidden -&gt; output layer\ntrain_op2 = optimizer.minimize(loss2, var_list=[W2, b2])  # update hidden -&gt; second output unit\n\n\nsess = tf.InteractiveSession()\nsess.run(tf.global_variables_initializer())\nmask_out1 = np.zeros((n_sample, 2))\nmask_out1[:, 1] += 1.0\n# print(mask_out1)\nprint(sess.run([hidden, out, loss, loss2], feed_dict={x: data_x, y: data_y, mask: mask_out1}))\n\n# In this case, only out2 is updated. You see the loss and loss2 decreases.\nsess.run(train_op2, feed_dict={x: data_x, y:data_y, mask: mask_out1})\nprint(sess.run([hidden, out, loss, loss2], feed_dict={x: data_x, y:data_y, mask: mask_out1}))\n\n# In this case, both out1 and out2 is updated. You see the loss and loss2 decreases.\nsess.run(train_op12, feed_dict={x: data_x, y:data_y, mask: mask_out1})\nprint(sess.run([hidden, out, loss, loss2], feed_dict={x: data_x, y:data_y, mask: mask_out1}))\n\n# In this case, everything is updated. You see the loss and loss2 decreases.\nsess.run(train_op_all, feed_dict={x: data_x, y:data_y, mask: mask_out1})\nprint(sess.run([hidden, out, loss, loss2], feed_dict={x: data_x, y:data_y, mask: mask_out1}))\nsess.close()\n<\/code><\/pre>\n\n<p>=======================Below is the old answer==============================<\/p>\n\n<p>To get derivatives w.r.t. different variables, you can pass a <code>var_list<\/code> to decide which variable to update. Here is an example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\n# let's make our tiny dataset: (x, y) pairs, where x = (x1, x2, x3), y = (y1, y2),\n# and y1 = x1+x2+x3, y2 = x1^2+x2^2+x3^2\n\n# n_sample data points\nn_sample = 8\ndata_x = np.random.random((n_sample, 3))\ndata_y = np.zeros((n_sample, 2))\ndata_y[:, 0] += np.sum(data_x, axis=1)\ndata_y[:, 1] += np.sum(data_x**2, axis=1)\ndata_y += 0.01 * np.random.random((n_sample, 2))  # add some noise\n\n\n# build graph\n# suppose we have a network of shape [3, 4, 2], i.e.: one hidden layer of size 4.\n\nx = tf.placeholder(tf.float32, shape=[None, 3], name='x')\ny = tf.placeholder(tf.float32, shape=[None, 2], name='y')\n\nW1 = tf.Variable(tf.random_normal(shape=[3, 4], stddev=0.1), name='W1')\nb1 = tf.Variable(tf.random_normal(shape=[4], stddev=0.1), name='b1')\nhidden = tf.nn.sigmoid(tf.matmul(x, W1) + b1)\nW2 = tf.Variable(tf.random_normal(shape=[4, 2], stddev=0.1), name='W2')\nb2 = tf.Variable(tf.random_normal(shape=[2], stddev=0.1), name='b2')\nout = tf.matmul(hidden, W2) + b2\n\nloss = tf.reduce_mean(tf.square(out - y))\noptimizer = tf.train.GradientDescentOptimizer(0.1)\n# You can pass a variable list to decide which variable(s) to minimize.\ntrain_op_second_layer = optimizer.minimize(loss, var_list=[W2, b2])\n# If there is no var_list, all variables will be updated.\ntrain_op_all = optimizer.minimize(loss)\n\nsess = tf.InteractiveSession()\nsess.run(tf.global_variables_initializer())\nprint(sess.run([W1, b1, W2, b2, loss], feed_dict={x: data_x, y:data_y}))\n\n# In this case, only W2 and b2 are updated. You see the loss decreases.\nsess.run(train_op_second_layer, feed_dict={x: data_x, y:data_y})\nprint(sess.run([W1, b1, W2, b2, loss], feed_dict={x: data_x, y:data_y}))\n\n# In this case, all variables are updated. You see the loss decreases.\nsess.run(train_op_all, feed_dict={x: data_x, y:data_y})\nprint(sess.run([W1, b1, W2, b2, loss], feed_dict={x: data_x, y:data_y}))\nsess.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":42183154,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-12T01:09:54.413","AcceptedAnswerId":42183315.0,"OwnerUserId":6815505.0,"Title":"ValueError when trying to evaluate a new data instance in tensorflow","Body":"<p>I am trying to predict the classification of a new instance of data using the feed forward DNN graph in tensorflow.<\/p>\n\n<p>The code is:<\/p>\n\n<pre><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><\/pre>\n\n<p>on this line<\/p>\n\n<pre><\/pre>\n\n<p>I'm not sure where to begin as I don't want to change around the values in my placeholders as I think they are where they need to be. Any help is appreciated. Thanks!<\/p>\n","answers":[{"AnswerId":"42183315","CreationDate":"2017-02-12T01:36:16.153","ParentId":null,"OwnerUserId":"5788391","Title":null,"Body":"<p>Adding another [] around the input should resolve your problem. <code>[51.0,10.0,71.0,65.0,5.0,70.0,30.06]<\/code> is shape (7,) but <code>[[51.0,10.0,71.0,65.0,5.0,70.0,30.06]]<\/code> is (1,7)<\/p>\n\n<pre><code>classification = y.eval(feed_dict={x: [[51.0,10.0,71.0,65.0,5.0,70.0,30.06]]})\n<\/code><\/pre>\n"}]}
{"QuestionId":42183255,"AnswerCount":2,"Tags":"<python><c++><tensorflow><embedding>","CreationDate":"2017-02-12T01:25:52.167","AcceptedAnswerId":null,"OwnerUserId":7551991.0,"Title":"Importing tensorflow when embedding python in c++ returns null","Body":"<p>I've a question regarding embedding python into a C++ application. The setup is as follows: I have a large C++ application which generates some data (renders images in real-time) and displays them. I have also trained a neural network in python using tensorflow which will accept these images. <\/p>\n\n<p>My idea was to embed python and send data as a numpy array, predict using the neural network and get back another processed numpy array to display (in C++). I've made some basic tests without tensorflow on python side to get a feel for embedding python in c and such and it seems to work.<\/p>\n\n<p>However, once i place \"import tensorflow\" into any python script that I want to import, I'll get a NULL from PyImport_ImportModule in c++ part.<\/p>\n\n<p>e.g.<\/p>\n\n<pre><\/pre>\n\n<p>works fine. But the following doesn't:<\/p>\n\n<pre><\/pre>\n\n<p>In the second case, I still get the message in stdout from tensorflow that it has found cuda etc, but then the module import fails.<\/p>\n\n<p>My setup is on Windows 10 x64, Anaconda Python 3.5, tensorflow-0.12 and CUDA 8. Has anyone ran across a similar problem? Other modules that I've tested (numpy, pil, scipy) seem to load fine.<\/p>\n\n<p>If it looks like it cannot be resolved, I'll resort to IPC of some sort between the c++ part and python.<\/p>\n","answers":[{"AnswerId":"42183354","CreationDate":"2017-02-12T01:43:01.920","ParentId":null,"OwnerUserId":"7551991","Title":null,"Body":"<p>I've resolved the issue. I needed to set argc and argv with PySys_SetArgv. I've uncovered this using PyErr_Occurred() and PyErr_Print() right after the failed import to see the problem.<\/p>\n"},{"AnswerId":"43014614","CreationDate":"2017-03-25T09:16:08.237","ParentId":null,"OwnerUserId":"4401462","Title":null,"Body":"<p>ref: <a href=\"https:\/\/docs.python.org\/3.5\/extending\/embedding.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3.5\/extending\/embedding.html<\/a><\/p>\n\n<p>main.cpp<\/p>\n\n<pre><code>#include &lt;Python.h&gt;\n#include &lt;iostream&gt;\n#include &lt;QString&gt;\n#include &lt;QDir&gt;\n#include &lt;cstring&gt;\n\nusing namespace std;\n\nint main(int argc, char *argv[])\n{\n    PyObject *pName, *pModule, *pDict, *pFunc;\n    PyObject *pArgs, *pValue;\n    int i;\n\n    if (argc &lt; 3) {\n        fprintf(stderr,\"Usage: call pythonfile funcname [args]\\n\");\n        return 1;\n    }\n\n    Py_SetProgramName((wchar_t*)L\"test\");\n\n    Py_Initialize();\n\n    PySys_SetArgv(argc, (wchar_t**)argv);\n    PyRun_SimpleString(\"import tensorflow as tf\\n\"\n                       \"print(tf.__version__)\\n\");\n\n    PyRun_SimpleString(\"import cv2\\n\"\n                       \"print(cv2.__version__)\\n\");\n\n    QString qs = QDir::currentPath();\n    std::wstring ws = qs.toStdWString();\n    PySys_SetPath(ws.data());\n    pName = PyUnicode_DecodeFSDefault(argv[1]);\n    \/* Error checking of pName left out *\/\n\n    pModule = PyImport_Import(pName);\n    Py_DECREF(pName);\n\n    if (pModule != NULL) {\n        pFunc = PyObject_GetAttrString(pModule, argv[2]);\n        \/* pFunc is a new reference *\/\n\n        if (pFunc &amp;&amp; PyCallable_Check(pFunc)) {\n            pArgs = PyTuple_New(argc - 3);\n            for (i = 0; i &lt; argc - 3; ++i) {\n                pValue = PyLong_FromLong(atoi(argv[i + 3]));\n                if (!pValue) {\n                    Py_DECREF(pArgs);\n                    Py_DECREF(pModule);\n                    fprintf(stderr, \"Cannot convert argument\\n\");\n                    return 1;\n                }\n                \/* pValue reference stolen here: *\/\n                PyTuple_SetItem(pArgs, i, pValue);\n            }\n            pValue = PyObject_CallObject(pFunc, pArgs);\n            Py_DECREF(pArgs);\n            if (pValue != NULL) {\n                printf(\"Result of call: %ld\\n\", PyLong_AsLong(pValue));\n                Py_DECREF(pValue);\n            }\n            else {\n                Py_DECREF(pFunc);\n                Py_DECREF(pModule);\n                PyErr_Print();\n                fprintf(stderr,\"Call failed\\n\");\n                return 1;\n            }\n        }\n        else {\n            if (PyErr_Occurred())\n                PyErr_Print();\n            fprintf(stderr, \"Cannot find function \\\"%s\\\"\\n\", argv[2]);\n        }\n        Py_XDECREF(pFunc);\n        Py_DECREF(pModule);\n    }\n    else {\n        PyErr_Print();\n        fprintf(stderr, \"Failed to load \\\"%s\\\"\\n\", argv[1]);\n        return 1;\n    }\n    Py_Finalize();\n    return 0;\n}\n<\/code><\/pre>\n\n<p>multiply.py<\/p>\n\n<pre><code>import tensorflow as tf\nimport cv2\n\ndef multiply(a,b):\n    print(tf.__version__) \n    print(cv2.__version__) \n    print(\"Will compute\", a, \"times\", b)\n    c = 0\n    for i in range(0, a):\n        c = c + b\n    return c\n<\/code><\/pre>\n"}]}
{"QuestionId":42183553,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-12T02:20:14.283","AcceptedAnswerId":null,"OwnerUserId":1601580.0,"Title":"Why would a TensorFlow restored checkpoint run out of memory but the original script wouldn't?","Body":"<p>I am running some TensorFlow code that restores and re-starts training from a checkpoint. Whenever I restore from a cpu build it seems to work perfectly fine. But if I try to restore when I run my code with gpu it seems to not work. In particular I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>I see that it says I am running out of memory, but when I increase the memory to say 10GBs it doesn't really change anything. This only happens with my gpu build because the cpu one restores perfectly fine.<\/p>\n\n<p>Anyway have any idea or starting ideas of what might be causing this? <\/p>\n\n<p>The gpu's are essentially assigned automatically so I'm not quite sure what might be causing it or what are the starting steps to even debug this.<\/p>\n\n<hr>\n\n<p>full error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42183661","CreationDate":"2017-02-12T02:39:20.660","ParentId":null,"OwnerUserId":"5788391","Title":null,"Body":"<p>Tensorflow CPU use benefits from both physical and virtual memory giving you almost unlimited memory to manipulate your models. Your first step in debugging is to build a smaller model by simply removing some weights\/layers and running on the GPU to ensure you have no coding errors. Then slowly increase the layers\/weights until the you run out of memory again. This will confirm that you have a memory issue on the GPU. I would recommend initially building your graph on the GPU that way you know it will fit later when you train on it. If you need the large graph then consider allocating parts of the graph to different GPU if you have them. <\/p>\n"}]}
{"QuestionId":42184277,"AnswerCount":0,"Tags":"<tensorflow><magenta>","CreationDate":"2017-02-12T04:41:12.453","AcceptedAnswerId":null,"OwnerUserId":6102900.0,"Title":"error when running bazel test of magenta","Body":"<p>My server is Ubuntu 14.04 CPU \nAfter installing magenta, I run the \n     bazel test magenta\/... ,\n and it gives me this error:<\/p>\n\n<p>ERROR: \/home\/jcc\/.cache\/bazel\/_bazel_jcc\/ed5599de3378c6e0a5a71ebe1650713e\/external\/protobuf\/protobuf.bzl:91:19: name 'HOST_CFG' is not defined.<\/p>\n\n<p>ERROR: package contains errors: magenta\/protobuf.\nERROR: error loading package 'magenta\/protobuf': Extension 'protobuf.bzl' has errors.<\/p>\n\n<p>INFO: Elapsed time: 0.274s<\/p>\n\n<p>ERROR: Couldn't start the build. Unable to run tests.<\/p>\n\n<p>Any advice?<\/p>\n","answers":[]}
{"QuestionId":42184663,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-02-12T05:54:09.013","AcceptedAnswerId":42190780.0,"OwnerUserId":7552393.0,"Title":"How to find an index of the first matching element in TensorFlow","Body":"<p>I am looking for a TensorFlow way of implementing something similar to Python's list.index() function.<\/p>\n\n<p>Given a matrix and a value to find, I want to know the first occurrence of the value in each row of the matrix.<\/p>\n\n<p>For example,<\/p>\n\n<pre><\/pre>\n\n<p>I cannot assume that 'val' appears only once in each row, otherwise I would have implemented it using tf.argmax(m == val). In my case, it is important to get the index of the <em>first<\/em> occurrence of 'val' and not any.<\/p>\n","answers":[{"AnswerId":"53368397","CreationDate":"2018-11-19T04:38:23.647","ParentId":null,"OwnerUserId":"2341325","Title":null,"Body":"<p>Here is another solution to the problem, assuming there is a hit on every row.<\/p>\n\n<pre><code>import tensorflow as tf\n\nval = 3\nm = tf.constant([\n    [0  ,   0,   val,   0, val],\n    [val,   0,   val, val,   0],\n    [0  , val,     0,   0,   0]])\n\n# replace all entries in the matrix either with its column index, or out-of-index-number\nmatch_indices = tf.where(                          # [[5, 5, 2, 5, 4],\n    tf.equal(val, m),                              #  [0, 5, 2, 3, 5],\n    x=tf.range(tf.shape(m)[1]) * tf.ones_like(m),  #  [5, 1, 5, 5, 5]]\n    y=(tf.shape(m)[1])*tf.ones_like(m))\n\nresult = tf.reduce_min(match_indices, axis=1)\n\nwith tf.Session() as sess:\n    print(sess.run(result)) # [2, 0, 1]\n<\/code><\/pre>\n"},{"AnswerId":"42187741","CreationDate":"2017-02-12T12:48:33.680","ParentId":null,"OwnerUserId":"6791223","Title":null,"Body":"<p>Looks a little ugly but works (assuming <code>m<\/code> and <code>val<\/code> are both tensors):<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>idx = list()\nfor t in tf.unpack(m, axis=0):\n    idx.append(tf.reduce_min(tf.where(tf.equal(t, val))))\nidx = tf.pack(idx, axis=0)\n<\/code><\/pre>\n\n<p><strong>EDIT:<\/strong>\nAs <a href=\"https:\/\/stackoverflow.com\/users\/419116\/yaroslav-bulatov\">Yaroslav Bulatov<\/a> mentioned, you could achieve the same result with <code>tf.map_fn<\/code>:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>def index1d(t):\n    return tf.reduce_min(tf.where(tf.equal(t, val)))\n\nidx = tf.map_fn(index1d, m, dtype=tf.int64)\n<\/code><\/pre>\n"},{"AnswerId":"42190780","CreationDate":"2017-02-12T17:24:12.143","ParentId":null,"OwnerUserId":"6770703","Title":null,"Body":"<p>It seems that <code>tf.argmax<\/code> works like <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.argmax.html\" rel=\"noreferrer\"><code>np.argmax<\/code><\/a> (according to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/kernel_tests\/argmax_op_test.py#L67\" rel=\"noreferrer\">the test<\/a>), which will return the first index when there are multiple occurrences of the max value.\nYou can use <code>tf.argmax(tf.cast(tf.equal(m, val), tf.int32), axis=1)<\/code> to get what you want. However, currently the behavior of <code>tf.argmax<\/code> is undefined in case of multiple occurrences of the max value.<\/p>\n\n<p>If you are worried about undefined behavior, you can apply <code>tf.argmin<\/code> on the return value of <code>tf.where<\/code> as @Igor Tsvetkov suggested.\nFor example,<\/p>\n\n<pre><code># test with tensorflow r1.0\nimport tensorflow as tf\n\nval = 3\nm = tf.placeholder(tf.int32)\nm_feed = [[0  ,   0, val,   0, val],\n          [val,   0, val, val,   0],\n          [0  , val,   0,   0,   0]]\n\ntmp_indices = tf.where(tf.equal(m, val))\nresult = tf.segment_min(tmp_indices[:, 1], tmp_indices[:, 0])\n\nwith tf.Session() as sess:\n    print(sess.run(result, feed_dict={m: m_feed})) # [2, 0, 1]\n<\/code><\/pre>\n\n<p>Note that <code>tf.segment_min<\/code> will raise <code>InvalidArgumentError<\/code> when there is some row containing no <code>val<\/code>. In your code <code>row_elems.index(val)<\/code> will raise exception too when <code>row_elems<\/code> don't contain <code>val<\/code>.<\/p>\n"}]}
{"QuestionId":42184791,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><keras><dropout>","CreationDate":"2017-02-12T06:18:55.240","AcceptedAnswerId":null,"OwnerUserId":297850.0,"Title":"dropout and data.split in model.fit","Body":"<p>As we know, dropout is a mechanism to help control overfitting. In the training process of Keras, we can conduct online cross-validation by monitoring the validation loss, and setup data split in . <\/p>\n\n<p>Generally, do I need to use both of these mechanisms? Or if I setup data split in , then I do not need to use dropout.<\/p>\n","answers":[{"AnswerId":"42193385","CreationDate":"2017-02-12T21:19:55.880","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>Dropout is a regularization technique, i.e. it prevents the network from overfitting on your data quickly. The validation loss just gives you an indication of when your network is overfitting. These are two completely different things and having a validation loss does not help you when your model is overfitting, it just shows you that it is.<\/p>\n\n<p>I would say that having a validation loss is valuable information to have during training and you should never go without it. Whether you need regularization techniques such as noise, dropout or batch normalization depends on how your network learns. If you see that it overfits then you should attempt to employ regularization techniques.<\/p>\n"}]}
{"QuestionId":42184863,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><deep-learning><keras><keras-layer>","CreationDate":"2017-02-12T06:33:52.660","AcceptedAnswerId":44771313.0,"OwnerUserId":99379.0,"Title":"How do you make TensorFlow + Keras fast with a TFRecord dataset?","Body":"<p><strong>What is an example of how to use a TensorFlow TFRecord with a Keras Model and tf.session.run() while keeping the dataset in tensors w\/ queue runners?<\/strong><\/p>\n\n<p>Below is a snippet that works but it needs the following improvements:<\/p>\n\n<ul>\n<li>Use the <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">Model API<\/a><\/li>\n<li>specify an Input()<\/li>\n<li>Load a dataset from a TFRecord<\/li>\n<li>Run through a dataset in parallel (such as with a queuerunner)<\/li>\n<\/ul>\n\n<p>Here is the snippet, there are several TODO lines indicating what is needed:<\/p>\n\n\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p><strong>Why is this question relevant?<\/strong><\/p>\n\n<ul>\n<li>For high performance training without going back to python\n\n<ul>\n<li>no <a href=\"https:\/\/stackoverflow.com\/questions\/36026892\/how-can-i-convert-tfrecords-into-numpy-arrays\">TFRecord to numpy<\/a> to tensor conversions<\/li>\n<\/ul><\/li>\n<li><a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/5358\" rel=\"nofollow noreferrer\">Keras will soon be part of tensorflow<\/a><\/li>\n<li>Demonstrate how Keras Model() classes can accept tensors for input data correctly.<\/li>\n<\/ul>\n\n<p><strong>Here is some starter information for a semantic segmentation problem example:<\/strong><\/p>\n\n<ul>\n<li>example unet Keras model <a href=\"https:\/\/github.com\/ahundt\/tf-image-segmentation\/blob\/cd7df53e6f59ef9e3dff1ed0119e12922ae98a3a\/tf_image_segmentation\/models\/unet.py\" rel=\"nofollow noreferrer\">unet.py<\/a>, happens to be for semantic segmentation.<\/li>\n<li><a href=\"https:\/\/blog.keras.io\/keras-as-a-simplified-interface-to-tensorflow-tutorial.html\" rel=\"nofollow noreferrer\">Keras + Tensorflow Blog Post<\/a><\/li>\n<li>An <a href=\"https:\/\/github.com\/ahundt\/tf-image-segmentation\/blob\/cd7df53e6f59ef9e3dff1ed0119e12922ae98a3a\/tf_image_segmentation\/recipes\/pascal_voc\/FCNs\/densenet_fcn.py#L78\" rel=\"nofollow noreferrer\">attempt at running the unet model a tf session with TFRecords and a Keras model<\/a> (not working)<\/li>\n<li>Code to create the TFRecords: <a href=\"https:\/\/github.com\/ahundt\/tf-image-segmentation\/blob\/cd7df53e6f59ef9e3dff1ed0119e12922ae98a3a\/tf_image_segmentation\/utils\/tf_records.py\" rel=\"nofollow noreferrer\">tf_records.py<\/a><\/li>\n<li>An attempt at running the unet model a tf session with TFRecords and a Keras model is in <a href=\"https:\/\/github.com\/ahundt\/tf-image-segmentation\/blob\/cd7df53e6f59ef9e3dff1ed0119e12922ae98a3a\/tf_image_segmentation\/recipes\/pascal_voc\/FCNs\/densenet_fcn.py#L78\" rel=\"nofollow noreferrer\">densenet_fcn.py<\/a> (not working)<\/li>\n<\/ul>\n","answers":[{"AnswerId":"42324697","CreationDate":"2017-02-19T07:21:08.013","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>I don't use tfrecord dataset format so won't argue on the pros and cons, but I got interested in extending Keras to support the same. <\/p>\n\n<p><a href=\"https:\/\/github.com\/indraforyou\/keras_tfrecord\" rel=\"noreferrer\">github.com\/indraforyou\/keras_tfrecord<\/a> is the repository. Will briefly explain the main changes.<\/p>\n\n<blockquote>\n  <p>Dataset creation and loading<\/p>\n<\/blockquote>\n\n<p><code>data_to_tfrecord<\/code> and <code>read_and_decode<\/code> <a href=\"https:\/\/github.com\/indraforyou\/keras_tfrecord\/blob\/4fa7e5d454dd4f3f33f1d756a2a8659f2e789141\/keras_tfrecord.py#L12-L65\" rel=\"noreferrer\">here<\/a> takes care of creating tfrecord dataset and loading the same. Special care must be to implement the <code>read_and_decode<\/code> otherwise you will face cryptic errors during training. <\/p>\n\n<blockquote>\n  <p>Initialization and Keras model <\/p>\n<\/blockquote>\n\n<p>Now both <code>tf.train.shuffle_batch<\/code> and Keras <code>Input<\/code> layer returns tensor. But the one returned by <code>tf.train.shuffle_batch<\/code> don't have metadata needed by Keras internally. As it turns out, any tensor can be easily turned into a tensor with keras metadata by calling <code>Input<\/code> layer with <code>tensor<\/code> param. <\/p>\n\n<p>So this takes care of initialization:<\/p>\n\n<pre><code>x_train_, y_train_ = ktfr.read_and_decode('train.mnist.tfrecord', one_hot=True, n_class=nb_classes, is_train=True)\n\nx_train_batch, y_train_batch = K.tf.train.shuffle_batch([x_train_, y_train_],\n                                                batch_size=batch_size,\n                                                capacity=2000,\n                                                min_after_dequeue=1000,\n                                                num_threads=32) # set the number of threads here\n\nx_train_inp = Input(tensor=x_train_batch)\n<\/code><\/pre>\n\n<p>Now with <code>x_train_inp<\/code> any keras model can be developed. <\/p>\n\n<blockquote>\n  <p>Training (simple)<\/p>\n<\/blockquote>\n\n<p>Lets say <code>train_out<\/code> is the output tensor of your keras model. You can easily write a custom training loop on the lines of:<\/p>\n\n<pre><code>loss = tf.reduce_mean(categorical_crossentropy(y_train_batch, train_out))\ntrain_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss)\n\n\n# sess.run(tf.global_variables_initializer())\nsess.run(tf.initialize_all_variables())\n\nwith sess.as_default():\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\n    try:\n      step = 0\n      while not coord.should_stop():\n        start_time = time.time()\n\n        _, loss_value = sess.run([train_op, loss], feed_dict={K.learning_phase(): 0})\n\n        duration = time.time() - start_time\n\n        if step % 100 == 0:\n          print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value,\n                                                     duration))\n        step += 1\n    except tf.errors.OutOfRangeError:\n      print('Done training for %d epochs, %d steps.' % (FLAGS.num_epochs, step))\n    finally:\n      coord.request_stop()\n\n    coord.join(threads)\n    sess.close()\n<\/code><\/pre>\n\n<blockquote>\n  <p>Training (keras style)<\/p>\n<\/blockquote>\n\n<p>One of the features of keras that makes it so lucrative is its generalized training mechanism with the callback functions. <\/p>\n\n<p>But to support tfrecords type training there are several changes that are need in the <code>fit<\/code> function<\/p>\n\n<ul>\n<li>running the queue threads<\/li>\n<li>no feeding in batch data through <code>feed_dict<\/code><\/li>\n<li>supporting validation becomes tricky as the validation data will also be coming in through another tensor an different model needs to be internally created with shared upper layers and validation tensor fed in by other tfrecord reader. <\/li>\n<\/ul>\n\n<p>But all this can be easily supported by another flag parameter. What makes things messing are the keras features <code>sample_weight<\/code> and <code>class_weight<\/code> they are used to weigh each sample and weigh each class. For this in <code>compile()<\/code> keras creates placeholders (<a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/4fa7e5d454dd4f3f33f1d756a2a8659f2e789141\/keras\/engine\/training.py#L592-L640\" rel=\"noreferrer\">here<\/a>) and placeholders are also implicitly created for the targets (<a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/4fa7e5d454dd4f3f33f1d756a2a8659f2e789141\/keras\/engine\/training.py#L643-L650\" rel=\"noreferrer\">here<\/a>) which is not needed in our case the labels are already fed in by tfrecord readers. These placeholders needs to be fed in during session run which is unnecessary in our cae. <\/p>\n\n<p>So taking into account these changes, <code>compile_tfrecord<\/code>(<a href=\"https:\/\/github.com\/indraforyou\/keras_tfrecord\/blob\/4fa7e5d454dd4f3f33f1d756a2a8659f2e789141\/keras_tfrecord.py#L68\" rel=\"noreferrer\">here<\/a>) and <code>fit_tfrecord<\/code>(<a href=\"https:\/\/github.com\/indraforyou\/keras_tfrecord\/blob\/4fa7e5d454dd4f3f33f1d756a2a8659f2e789141\/keras_tfrecord.py#L220\" rel=\"noreferrer\">here<\/a>) are the extension of <code>compile<\/code> and <code>fit<\/code> and shares say 95% of the code. <\/p>\n\n<p>They can be used in the following way:<\/p>\n\n<pre><code>import keras_tfrecord as ktfr\n\ntrain_model = Model(input=x_train_inp, output=train_out)\nktfr.compile_tfrecord(train_model, optimizer='rmsprop', loss='categorical_crossentropy', out_tensor_lst=[y_train_batch], metrics=['accuracy'])\n\ntrain_model.summary()\n\nktfr.fit_tfrecord(train_model, X_train.shape[0], batch_size, nb_epoch=3)\ntrain_model.save_weights('saved_wt.h5')\n<\/code><\/pre>\n\n<p>You are welcome to improve on the code and pull requests.<\/p>\n"},{"AnswerId":"44771313","CreationDate":"2017-06-27T02:11:14.433","ParentId":null,"OwnerUserId":"99379","Title":null,"Body":"<p>Update 2018-08-29 this is now directly supported in keras, see the following example:<\/p>\n\n<p><a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/mnist_tfrecord.py\" rel=\"noreferrer\">https:\/\/github.com\/keras-team\/keras\/blob\/master\/examples\/mnist_tfrecord.py<\/a><\/p>\n\n<p>Original Answer:<\/p>\n\n<p>TFRecords are supported by using an external loss. Here are the key lines constructing an external loss:<\/p>\n\n<pre><code># tf yield ops that supply dataset images and labels\nx_train_batch, y_train_batch = read_and_decode_recordinput(...)\n\n# create a basic cnn\nx_train_input = Input(tensor=x_train_batch)\nx_train_out = cnn_layers(x_train_input)\n\nmodel = Model(inputs=x_train_input, outputs=x_train_out)\nloss = keras.losses.categorical_crossentropy(y_train_batch, x_train_out)\nmodel.add_loss(loss)\n\nmodel.compile(optimizer='rmsprop', loss=None)\n<\/code><\/pre>\n\n<p>Here is an example for Keras 2. It works after applying the small patch <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/7060\" rel=\"noreferrer\">#7060<\/a>:<\/p>\n\n<pre><code>'''MNIST dataset with TensorFlow TFRecords.\n\nGets to 99.25% test accuracy after 12 epochs\n(there is still a lot of margin for parameter tuning).\n'''\nimport os\nimport copy\nimport time\n\nimport numpy as np\n\nimport tensorflow as tf\nfrom tensorflow.python.ops import data_flow_ops\nfrom keras import backend as K\nfrom keras.models import Model\nfrom keras.layers import Dense\nfrom keras.layers import Dropout\nfrom keras.layers import Flatten\nfrom keras.layers import Input\nfrom keras.layers import Conv2D\nfrom keras.layers import MaxPooling2D\nfrom keras.callbacks import EarlyStopping\nfrom keras.callbacks import TensorBoard\nfrom keras.objectives import categorical_crossentropy\nfrom keras.utils import np_utils\nfrom keras.utils.generic_utils import Progbar\nfrom keras import callbacks as cbks\nfrom keras import optimizers, objectives\nfrom keras import metrics as metrics_module\n\nfrom keras.datasets import mnist\n\nif K.backend() != 'tensorflow':\n    raise RuntimeError('This example can only run with the '\n                       'TensorFlow backend for the time being, '\n                       'because it requires TFRecords, which '\n                       'are not supported on other platforms.')\n\n\ndef images_to_tfrecord(images, labels, filename):\n    def _int64_feature(value):\n        return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))\n\n    def _bytes_feature(value):\n        return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))\n\n    \"\"\" Save data into TFRecord \"\"\"\n    if not os.path.isfile(filename):\n        num_examples = images.shape[0]\n\n        rows = images.shape[1]\n        cols = images.shape[2]\n        depth = images.shape[3]\n\n        print('Writing', filename)\n        writer = tf.python_io.TFRecordWriter(filename)\n        for index in range(num_examples):\n            image_raw = images[index].tostring()\n            example = tf.train.Example(features=tf.train.Features(feature={\n                'height': _int64_feature(rows),\n                'width': _int64_feature(cols),\n                'depth': _int64_feature(depth),\n                'label': _int64_feature(int(labels[index])),\n                'image_raw': _bytes_feature(image_raw)}))\n            writer.write(example.SerializeToString())\n        writer.close()\n    else:\n        print('tfrecord %s already exists' % filename)\n\n\ndef read_and_decode_recordinput(tf_glob, one_hot=True, classes=None, is_train=None,\n                                batch_shape=[1000, 28, 28, 1], parallelism=1):\n    \"\"\" Return tensor to read from TFRecord \"\"\"\n    print 'Creating graph for loading %s TFRecords...' % tf_glob\n    with tf.variable_scope(\"TFRecords\"):\n        record_input = data_flow_ops.RecordInput(\n            tf_glob, batch_size=batch_shape[0], parallelism=parallelism)\n        records_op = record_input.get_yield_op()\n        records_op = tf.split(records_op, batch_shape[0], 0)\n        records_op = [tf.reshape(record, []) for record in records_op]\n        progbar = Progbar(len(records_op))\n\n        images = []\n        labels = []\n        for i, serialized_example in enumerate(records_op):\n            progbar.update(i)\n            with tf.variable_scope(\"parse_images\", reuse=True):\n                features = tf.parse_single_example(\n                    serialized_example,\n                    features={\n                        'label': tf.FixedLenFeature([], tf.int64),\n                        'image_raw': tf.FixedLenFeature([], tf.string),\n                    })\n                img = tf.decode_raw(features['image_raw'], tf.uint8)\n                img.set_shape(batch_shape[1] * batch_shape[2])\n                img = tf.reshape(img, [1] + batch_shape[1:])\n\n                img = tf.cast(img, tf.float32) * (1. \/ 255) - 0.5\n\n                label = tf.cast(features['label'], tf.int32)\n                if one_hot and classes:\n                    label = tf.one_hot(label, classes)\n\n                images.append(img)\n                labels.append(label)\n\n        images = tf.parallel_stack(images, 0)\n        labels = tf.parallel_stack(labels, 0)\n        images = tf.cast(images, tf.float32)\n\n        images = tf.reshape(images, shape=batch_shape)\n\n        # StagingArea will store tensors\n        # across multiple steps to\n        # speed up execution\n        images_shape = images.get_shape()\n        labels_shape = labels.get_shape()\n        copy_stage = data_flow_ops.StagingArea(\n            [tf.float32, tf.float32],\n            shapes=[images_shape, labels_shape])\n        copy_stage_op = copy_stage.put(\n            [images, labels])\n        staged_images, staged_labels = copy_stage.get()\n\n        return images, labels\n\n\ndef save_mnist_as_tfrecord():\n    (X_train, y_train), (X_test, y_test) = mnist.load_data()\n    X_train = X_train[..., np.newaxis]\n    X_test = X_test[..., np.newaxis]\n    images_to_tfrecord(images=X_train, labels=y_train, filename='train.mnist.tfrecord')\n    images_to_tfrecord(images=X_test, labels=y_test, filename='test.mnist.tfrecord')\n\n\ndef cnn_layers(x_train_input):\n    x = Conv2D(32, (3, 3), activation='relu', padding='valid')(x_train_input)\n    x = Conv2D(64, (3, 3), activation='relu')(x)\n    x = MaxPooling2D(pool_size=(2, 2))(x)\n    x = Dropout(0.25)(x)\n    x = Flatten()(x)\n    x = Dense(128, activation='relu')(x)\n    x = Dropout(0.5)(x)\n    x_train_out = Dense(classes,\n                        activation='softmax',\n                        name='x_train_out')(x)\n    return x_train_out\n\n\nsess = tf.Session()\nK.set_session(sess)\n\nsave_mnist_as_tfrecord()\n\nbatch_size = 100\nbatch_shape = [batch_size, 28, 28, 1]\nepochs = 3000\nclasses = 10\nparallelism = 10\n\nx_train_batch, y_train_batch = read_and_decode_recordinput(\n    'train.mnist.tfrecord',\n    one_hot=True,\n    classes=classes,\n    is_train=True,\n    batch_shape=batch_shape,\n    parallelism=parallelism)\n\nx_test_batch, y_test_batch = read_and_decode_recordinput(\n    'test.mnist.tfrecord',\n    one_hot=True,\n    classes=classes,\n    is_train=True,\n    batch_shape=batch_shape,\n    parallelism=parallelism)\n\n\nx_batch_shape = x_train_batch.get_shape().as_list()\ny_batch_shape = y_train_batch.get_shape().as_list()\n\nx_train_input = Input(tensor=x_train_batch, batch_shape=x_batch_shape)\nx_train_out = cnn_layers(x_train_input)\ny_train_in_out = Input(tensor=y_train_batch, batch_shape=y_batch_shape, name='y_labels')\ncce = categorical_crossentropy(y_train_batch, x_train_out)\ntrain_model = Model(inputs=[x_train_input], outputs=[x_train_out])\ntrain_model.add_loss(cce)\n\ntrain_model.compile(optimizer='rmsprop',\n                    loss=None,\n                    metrics=['accuracy'])\ntrain_model.summary()\n\ntensorboard = TensorBoard()\n\n# tensorboard disabled due to Keras bug\ntrain_model.fit(batch_size=batch_size,\n                epochs=epochs)  # callbacks=[tensorboard])\n\ntrain_model.save_weights('saved_wt.h5')\n\nK.clear_session()\n\n# Second Session, pure Keras\n(X_train, y_train), (X_test, y_test) = mnist.load_data()\nX_train = X_train[..., np.newaxis]\nX_test = X_test[..., np.newaxis]\nx_test_inp = Input(batch_shape=(None,) + (X_test.shape[1:]))\ntest_out = cnn_layers(x_test_inp)\ntest_model = Model(inputs=x_test_inp, outputs=test_out)\n\ntest_model.load_weights('saved_wt.h5')\ntest_model.compile(optimizer='rmsprop', loss='categorical_crossentropy', metrics=['accuracy'])\ntest_model.summary()\n\nloss, acc = test_model.evaluate(X_test, np_utils.to_categorical(y_test), classes)\nprint('\\nTest accuracy: {0}'.format(acc))\n<\/code><\/pre>\n\n<p>I've also been working to improve the support for TFRecords in the following issue and pull request:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/6928\" rel=\"noreferrer\">#6928<\/a> Yield Op support: High Performance Large Datasets via TFRecords, and RecordInput<\/li>\n<li><a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/7102\" rel=\"noreferrer\">#7102<\/a> Keras Input Tensor API Design Proposal<\/li>\n<\/ul>\n\n<p>Finally, it is possible to use <code>tf.contrib.learn.Experiment<\/code> to train Keras models in TensorFlow.<\/p>\n"}]}
{"QuestionId":42185084,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-12T07:11:01.070","AcceptedAnswerId":42185954.0,"OwnerUserId":6815505.0,"Title":"InvalidArgumentError error when evaluating new data in tensorflow","Body":"<p>I am trying to predict the classification of a new instance of data using the feed forward DNN graph in tensorflow. <\/p>\n\n<p>This is a continuation of this thread, which was solved <a href=\"https:\/\/stackoverflow.com\/questions\/42183154\/valueerror-when-trying-to-evaluate-a-new-data-instance-in-tensorflow\/42183315#42183315\">ValueError when trying to evaluate a new data instance in tensorflow<\/a><\/p>\n\n<p>The code is:<\/p>\n\n<pre><\/pre>\n\n<p>The error is:<\/p>\n\n<pre><\/pre>\n\n<p>on this line<\/p>\n\n<pre><\/pre>\n\n<p>Code and data is here: <a href=\"https:\/\/github.com\/jhsmith12345\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jhsmith12345\/tensorflow<\/a><\/p>\n\n<p>I see that this is asking for float values, which I thought I had fed it. Any help is appreciated!<\/p>\n","answers":[{"AnswerId":"42185954","CreationDate":"2017-02-12T09:23:15.127","ParentId":null,"OwnerUserId":"5788391","Title":null,"Body":"<p>So assuming that y is your target data, during training you need to provide a value for the target data so that the error can be calculated. The line <code>classification = y.eval<\/code> is obtaining a prediction, so there should be no requirement to provide training data - so should be  <code>classification = prediction.eval ...<\/code><\/p>\n\n<pre><code> def train_neural_network(x):\n  prediction = neural_network_model(x)\n  cost =tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(prediction,y))\noptimizer = tf.train.AdamOptimizer().minimize(cost)\n\n  hm_epochs = 5\n\n  with tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n\n    for epoch in range(hm_epochs):\n        epoch_loss = .1\n        for _ in range(399):\n            _, c = sess.run([optimizer, cost], feed_dict = {x: trainX, y: trainY})\n            epoch_loss += c\n        print('Epoch', epoch, 'completed out of', hm_epochs, 'loss:', epoch_loss)\n\n    correct = tf.equal(tf.argmax(prediction,1), tf.argmax(y,1))\n    accuracy = tf.reduce_mean(tf.cast(correct, 'float'))\n    print('Accuracy:',accuracy.eval({x: testX, y: testY}))\n    classification = prediction.eval(feed_dict={x: [[51.0,10.0,71.0,65.0,5.0,70.0,30.06]]})\n    print (classification)\n\n train_neural_network(x)\n<\/code><\/pre>\n"}]}
{"QuestionId":42185198,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-12T07:29:00.307","AcceptedAnswerId":null,"OwnerUserId":5350430.0,"Title":"function tf.slice_input_producer error using tensorflow","Body":"<p>I'm using the tfrecord to deal with sequences dataset. When one example is readed into the memory, I'll get two context features and one sequence feature. The context features are the length of the sequences and the label of the sequences. The sequence feature is a list of bytes with represent the feature of every time step.\nSo I have three tensor for every example:<\/p>\n\n<pre><\/pre>\n\n<p>I want to use every sequence feature to predict the label, so I have to make the label be the same length as the frame.<\/p>\n\n<pre><\/pre>\n\n<p>This time I get the following res:<\/p>\n\n<pre><\/pre>\n\n<p>And I have to push them into a queue so that I can get one single frame and label.<\/p>\n\n<pre><\/pre>\n\n<p>Ant then batch them then do the net training routine.<\/p>\n\n<pre><\/pre>\n\n<p>It should work, however, error happened in function tf.train.slice_input_producer Here's the error info:<\/p>\n\n<pre><\/pre>\n\n<p>the name of the slice_input_producer is slice_timestep<\/p>\n\n<p>the name of the shuffle_batch is batch_ex<\/p>\n\n<p>Here's the graph showed in my tensorboard.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DANSy.png\" rel=\"nofollow noreferrer\">the whole graph<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/M4NJw.png\" rel=\"nofollow noreferrer\">local zoomed graph<\/a><\/p>\n\n<p>Below is the simplified code which reproduce the error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42195962","CreationDate":"2017-02-13T02:53:37.280","ParentId":null,"OwnerUserId":"5350430","Title":null,"Body":"<p>I've solved. It seems that the slice_input_producer is static. I use<\/p>\n"}]}
{"QuestionId":42185312,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-02-12T07:45:14.913","AcceptedAnswerId":42185390.0,"OwnerUserId":5745846.0,"Title":"Caffe - Forward multiple images through a net in parallel","Body":"<p>How exactly would one do this. Right now, I have to iterate through each image and forward it. I would like to know if I can set multiple images at once and forward them through<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42185390","CreationDate":"2017-02-12T07:57:19.220","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You can do this by putting all images into a single batch and then running <code>net.forward()<\/code> once for the entire batch.<\/p>\n\n<pre><code>bs = len(fnames)  # batch size\nin_shape = net.blobs['data'].data.shape\nin_shape[0] = bs # set new batch size\nnet.blobs['data'].reshape(*in_shape)\nnet.reshape()\nfor i, f in enumerate(fnames):\n    img = Image.open(f)\n    # scale all images to 256x256\n    img = img.resize((256,256), PIL.Image.ANTIALIAS)\n    img = numpy.array(img).astype(numpy.float32)\n\n    transformed_image = transformer.preprocess('data', img)\n    #print transformed_image.shape\n\n    # put the image into i-th place in batch\n    net.blobs['data'].data[i,:,:,:] = transformed_image   \n\n# after reading all images into batch, forward once:\nnet.forward()\n<\/code><\/pre>\n"}]}
{"QuestionId":42186543,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><gensim>","CreationDate":"2017-02-12T10:38:17.083","AcceptedAnswerId":42194067.0,"OwnerUserId":5417769.0,"Title":"Training wordvec in Tensorflow, importing to Gensim","Body":"<p>I am training a word2vec model from the tensorflow tutorial.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py<\/a><\/p>\n\n<p>After training I get the embedding matrix. I would like to save this and import it as a trained model in gensim.<\/p>\n\n<p>To load a model in gensim, the command is:<\/p>\n\n<pre><\/pre>\n\n<p>But how do I generate the  file from Tensorflow?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42194067","CreationDate":"2017-02-12T22:31:27.647","ParentId":null,"OwnerUserId":"690430","Title":null,"Body":"<p>One way to is save the file in the non-binary Word2Vec format, which essentially looks like this:<\/p>\n\n<pre><code>num_words vector_size  # this is the header\nlabel0 x00 x01 ... x0N\nlabel1 x10 x11 ... x1N\n...\n<\/code><\/pre>\n\n<p>Example:<\/p>\n\n<pre><code>2 3\nword0 -0.000737 -0.002106 0.001851\nword1 -0.000878 -0.002106 0.002834\n<\/code><\/pre>\n\n<p>Save the file and then load with kwarg <code>binary=False<\/code>:<\/p>\n\n<pre><code>model = Word2Vec.load_word2vec_format(filename, binary=False)\n\nprint(model['word0'])\n<\/code><\/pre>\n\n<h2>Update<\/h2>\n\n<p>New way to load model is:<\/p>\n\n<pre><code>from gensim.models.keyedvectors import KeyedVectors\n\nmodel = KeyedVectors.load_word2vec_format(model_path, binary=False)\n<\/code><\/pre>\n"}]}
{"QuestionId":42187056,"AnswerCount":2,"Tags":"<tensorflow><bazel>","CreationDate":"2017-02-12T11:32:12.897","AcceptedAnswerId":43071369.0,"OwnerUserId":5440222.0,"Title":"tf_cc_test vs cc_test in the tensorflow\/core\/BUILD","Body":"<p>I mainly  concerned the difference of tf_cc_test and cc_test here.(I'm the new player of bazel due to tensorflow)<\/p>\n\n<p>I use the following command:<\/p>\n\n<blockquote>\n  <p>bazel build -c dbg \/\/tensorflow\/core:lib_random_weighted_picker_test<\/p>\n<\/blockquote>\n\n<p>to build lib_random_weighted_picker_test, all thing goes well. <\/p>\n\n<p>while I use <\/p>\n\n<blockquote>\n  <p>bazel build -c dbg \/\/tensorflow\/core:higher_level_tests<\/p>\n<\/blockquote>\n\n<p>to get the test instance about graph_partition_test which is in the tensorflow\/core\/graph\/graph_partition_test.cc, it reported the following error:<\/p>\n\n<blockquote>\n  <p>ERROR: no such target '\/\/tensorflow\/core:higher_level_tests': target\n  'higher_level_tests' not declared in package 'tensorflow\/core' defined\n  by \/home\/pzz\/workspace\/tensorflow\/tensorflow\/core\/BUILD. INFO: Elapsed\n  time: 0.151s<\/p>\n<\/blockquote>\n\n<p>And finally I found that I can use bazel build -c dbg \/\/tensorflow\/core:graph_graph_partition_test to get the test instance under the  graph_partition_test.cc.\nand all the tests in the higher_level_test can get from the above way.<\/p>\n\n<p>The higher_level_test is come from the tensorflow\/core\/BUILD#1705,<\/p>\n\n<pre><\/pre>\n\n<p><strong><em>Question:<\/em><\/strong>\nBut I still cannot make sense the difference about tf_cc_test and cc_test in the BUILD !!\nIf I want to write some test, which way should I take in the BUILD file?<\/p>\n\n<p>Any advice will be good, thanks.<\/p>\n","answers":[{"AnswerId":"42381421","CreationDate":"2017-02-22T02:06:32.293","ParentId":null,"OwnerUserId":"5440222","Title":null,"Body":"<p>answer can be got in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/tensorflow.bzl\" rel=\"nofollow noreferrer\">file<\/a>.<\/p>\n"},{"AnswerId":"43071369","CreationDate":"2017-03-28T13:48:53.917","ParentId":null,"OwnerUserId":"7778502","Title":null,"Body":"<p><code>tf_cc_test<\/code> is a macro that is just a wrapper for <code>cc_test<\/code>, as <a href=\"https:\/\/stackoverflow.com\/users\/5440222\/zhenzhong-pan\">Zhenzhong Pan<\/a> points out.<\/p>\n\n<p>Concrete location: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/tensorflow.bzl#L321-L332\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/tensorflow.bzl#L321-L332<\/a><\/p>\n\n<p>You can read about build rule macros here: <a href=\"https:\/\/bazel.build\/versions\/master\/docs\/skylark\/macros.html\" rel=\"nofollow noreferrer\">https:\/\/bazel.build\/versions\/master\/docs\/skylark\/macros.html<\/a><\/p>\n"}]}
{"QuestionId":42187425,"AnswerCount":3,"Tags":"<keras><keras-layer>","CreationDate":"2017-02-12T12:17:06.203","AcceptedAnswerId":42188039.0,"OwnerUserId":6293886.0,"Title":"How to change input shape in Sequential model in Keras","Body":"<p>I have a sequential model that I built in Keras.\nI try to figure out how to change the shape of the input. In the following example<\/p>\n\n<pre><\/pre>\n\n<p>let's say that I want to build a new model with different input shape, conceptual this should looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>is there a way to modify the model input shape?<\/p>\n","answers":[{"AnswerId":"55146575","CreationDate":"2019-03-13T16:16:53.333","ParentId":null,"OwnerUserId":"2135504","Title":null,"Body":"<p>Here is another solution without defining each layer of the model from scratch. The key for me was to use \"_layers\" instead of \"layers\". The latter only seems to return a copy.<\/p>\n\n<pre><code>import keras\nimport numpy as np\n\ndef get_model():\n    old_input_shape = (20, 20, 3)\n    model = keras.models.Sequential()\n    model.add(keras.layers.Conv2D(9, (3, 3), padding=\"same\", input_shape=old_input_shape))\n    model.add(keras.layers.MaxPooling2D((2, 2)))\n    model.add(keras.layers.Flatten())\n    model.add(keras.layers.Dense(1, activation=\"sigmoid\"))\n    model.compile(loss='binary_crossentropy', optimizer=keras.optimizers.Adam(lr=0.0001), metrics=['acc'], )\n    model.summary()\n    return model\n\ndef change_model(model, new_input_shape=(None, 40, 40, 3)):\n    # replace input shape of first layer\n    model._layers[1].batch_input_shape = new_input_shape\n\n    # feel free to modify additional parameters of other layers, for example...\n    model._layers[2].pool_size = (8, 8)\n    model._layers[2].strides = (8, 8)\n\n    # rebuild model architecture by exporting and importing via json\n    new_model = keras.models.model_from_json(model.to_json())\n    new_model.summary()\n\n    # copy weights from old model to new one\n    for layer in new_model.layers:\n        try:\n            layer.set_weights(model.get_layer(name=layer.name).get_weights())\n        except:\n            print(\"Could not transfer weights for layer {}\".format(layer.name))\n\n    # test new model on a random input image\n    X = np.random.rand(10, 40, 40, 3)\n    y_pred = new_model.predict(X)\n    print(y_pred)\n\n    return new_model\n\nif __name__ == '__main__':\n    model = get_model()\n    new_model = change_model(model)\n<\/code><\/pre>\n"},{"AnswerId":"42188039","CreationDate":"2017-02-12T13:19:09.903","ParentId":null,"OwnerUserId":"4955956","Title":null,"Body":"<p>Think about what changing the input shape in that situation would mean.<\/p>\n\n<p>Your first model <\/p>\n\n<pre><code>model.add(Dense(32, input_shape=(500,)))\n<\/code><\/pre>\n\n<p>Has a dense layer that really is a 500x32 matrix.<\/p>\n\n<p>If you changed your input to 250 elements, your layers's matrix and input dimension would mismatch.<\/p>\n\n<p>If, however,  what you were trying to achieve was to reuse your last layer's trained parameters from your first 500 element input model, you could get those weights by get_weights. Then you could rebuild a new model and set values at the new model with set_weights.<\/p>\n\n<pre><code>model1 = Sequential()\nmodel1.add(Dense(32, input_shape=(250,)))\nmodel1.add(Dense(10, activation='softmax'))\nmodel1.layers[1].set_weights(model1.layers[1].get_weights())\n<\/code><\/pre>\n\n<p>Keep in mind that model1 first layer (aka model1.layers[0]) would still be untrained<\/p>\n"},{"AnswerId":"46653912","CreationDate":"2017-10-09T19:45:28.070","ParentId":null,"OwnerUserId":"4735180","Title":null,"Body":"<p>Somewhat related, so hopefully someone will find this useful: If you have an existing model where the input is a placeholder that looks like (None, None, None, 3) for example, you can load the model, replace the first layer with a concretely shaped input. Transformation of this kind is very useful when for example you want to use your model in iOS CoreML (In my case the input of the model was a MLMultiArray instead of CVPixelBuffer, and the model compilation failed)<\/p>\n\n<pre><code>from keras.models import load_model\nfrom keras import backend as K\nfrom keras.engine import InputLayer\nimport coremltools\n\nmodel = load_model('your_model.h5')\n\n# Create a new input layer to replace the (None,None,None,3) input layer :\ninput_layer = InputLayer(input_shape=(272, 480, 3), name=\"input_1\")\n\n# Save and convert :\nmodel.layers[0] = input_layer\nmodel.save(\"reshaped_model.h5\")    \ncoreml_model = coremltools.converters.keras.convert('reshaped_model.h5')    \ncoreml_model.save('MyPredictor.mlmodel')\n<\/code><\/pre>\n"}]}
{"QuestionId":42188563,"AnswerCount":1,"Tags":"<theano><lasagne>","CreationDate":"2017-02-12T14:08:45.750","AcceptedAnswerId":42368796.0,"OwnerUserId":2280645.0,"Title":"How to reset weights of lasagne network?","Body":"<p>I want to reset the weights of my convolutional neural network if a \"nan\" is detected.<\/p>\n\n<p>Im not sure how to do it.<\/p>\n\n<p>Im also confused if i should change the seed as well in this case.<\/p>\n\n<pre><\/pre>\n\n<p><strong>How should i implement reset_network() ?<\/strong><\/p>\n","answers":[{"AnswerId":"42368796","CreationDate":"2017-02-21T13:34:57.797","ParentId":null,"OwnerUserId":"1391609","Title":null,"Body":"<p>I'm not sure this is the intended way of resetting network weights, but here's how I did it. In the following code <code>network<\/code> is a reference to a CNN with 2 convolutional layers followed by max pooling layers. I believe it should work with other architectures as well.<\/p>\n\n<p>The trick here is to update all trainable parameters of the network with initializer functions. <\/p>\n\n<pre><code>def reset_weights(network):\n    params = lasagne.layers.get_all_params(network, trainable=True)\n    for v in params:\n        val = v.get_value()\n        if(len(val.shape) &lt; 2):\n            v.set_value(lasagne.init.Constant(0.0)(val.shape))\n        else:\n            v.set_value(lasagne.init.GlorotUniform()(val.shape))\n<\/code><\/pre>\n\n<p>I hope it helps!<\/p>\n"}]}
{"QuestionId":42188656,"AnswerCount":0,"Tags":"<c++><opencv><caffe>","CreationDate":"2017-02-12T14:16:38.857","AcceptedAnswerId":null,"OwnerUserId":4720028.0,"Title":"OpenCV Caffe C++ - Forwarding multiple Images","Body":"<p>In the tutorial listed on the OPENCV website:<\/p>\n\n<p><a href=\"http:\/\/docs.opencv.org\/trunk\/d5\/de7\/tutorial_dnn_googlenet.html\" rel=\"nofollow noreferrer\">http:\/\/docs.opencv.org\/trunk\/d5\/de7\/tutorial_dnn_googlenet.html<\/a><\/p>\n\n<p>It seems we can easily forward single images through the network. I would however like to forward a batch of images through. This is possible in python but I have not found any C++ example online. I would like to avoid using a for loop to iterate and forward each image.<\/p>\n","answers":[]}
{"QuestionId":42188952,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-12T14:43:25.087","AcceptedAnswerId":45602524.0,"OwnerUserId":6102784.0,"Title":"Tensorflow: How to train for TF Detect","Body":"<p>The Tensorflow TF Detect for android example draws boxes around humans seen by the camera, however I would like to train a model to detect things other than humans.   <\/p>\n\n<p>How was the example's model trained? <\/p>\n","answers":[{"AnswerId":"45602524","CreationDate":"2017-08-10T00:15:43.960","ParentId":null,"OwnerUserId":"448547","Title":null,"Body":"<p>Now you have the tools on the TensorFlow repo!<\/p>\n\n<p><a href=\"https:\/\/research.googleblog.com\/2017\/06\/supercharge-your-computer-vision-models.html\" rel=\"nofollow noreferrer\">Supercharge your Computer Vision models with the TensorFlow Object Detection API<\/a><\/p>\n"}]}
{"QuestionId":42189401,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><hdf5>","CreationDate":"2017-02-12T15:26:11.503","AcceptedAnswerId":null,"OwnerUserId":6147229.0,"Title":"How to load multiple files to caffe model","Body":"<p>I am trying to use the following caffe model for prediction \"<a href=\"https:\/\/github.com\/amandajshao\/www_deep_crowd\" rel=\"nofollow noreferrer\">https:\/\/github.com\/amandajshao\/www_deep_crowd<\/a>\". this model don't take as input an image but it takes some motion and appearance features. I have those features for every frame as .mat files and converted them to hdf5.\nbut i couldn't figure out how to load them to caffe.\nany help please? <\/p>\n","answers":[{"AnswerId":"42192737","CreationDate":"2017-02-12T20:17:37.493","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Converting to HDF5 is more than half the way there...<\/p>\n\n<p><a href=\"http:\/\/caffe.help\/manual\/layers\/hdf5data.html\" rel=\"nofollow noreferrer\"><code>\"HDF5Data\"<\/code><\/a> layer can produce several <code>\"top\"<\/code>s: one for each dataset stored in the hdf5 file. So, suppose you <a href=\"https:\/\/stackoverflow.com\/a\/34261942\/1714410\">used matlab to write your hdf5 files<\/a> <\/p>\n\n<pre><code>hdf5write('my_data.h5', 'motion', single( motion )  );\nhdf5write('my_data.h5', 'appearance', single(appearnace), 'WriteMode', 'append' );\n<\/code><\/pre>\n\n<p>Then you can have an <code>\"HDF5Data\"<\/code> layer with two <code>\"top<\/code>\"s:<\/p>\n\n<pre><code>layer {\n  name: \"data\"\n  type: \"HDF5Data\"\n  top: \"motion\"\n  top: \"appearance\"\n  hdf5_data_param { source: \"path\/to\/hdf_list.txt\" }\n}\n<\/code><\/pre>\n\n<p>You might want to pay close attention to the order the data is stored, as Matlab stores arrays in \"Fortran\" fashion, while caffe expects its data ordered in a \"C\" fashion. See <a href=\"https:\/\/stackoverflow.com\/a\/34431532\/1714410\">this post for more information<\/a>.<\/p>\n"}]}
{"QuestionId":42189517,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-12T15:37:04.427","AcceptedAnswerId":null,"OwnerUserId":541106.0,"Title":"Using Tensor Flow CNN for prediction","Body":"<p>I am new to tensorflow. I have manged to train and validate a CNN, saved the session through the Saver object into a CPKT file and loaded it back. Now I'd like to use the trained model in order to check how it performs against a photo I took myself. I can't figure out how to do it.<\/p>\n","answers":[{"AnswerId":"48914808","CreationDate":"2018-02-21T20:22:32.427","ParentId":null,"OwnerUserId":"5361479","Title":null,"Body":"<ul>\n<li><p>Use a <code>tf.train.Saver()<\/code> object<\/p>\n\n<pre><code>saver = tf.train.Saver()\nwith tf.Session() as sess:\n    saver.restore(sess, tf.train.latest_checkpoint('.\/'))\n<\/code><\/pre><\/li>\n<\/ul>\n"},{"AnswerId":"42210526","CreationDate":"2017-02-13T18:01:23.553","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>Great to hear! All you need to do is to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables\" rel=\"nofollow noreferrer\">load the checkpoint<\/a>, and evaluate the network against it. That is typically done using the same script as the training script, but only evaluating the loss function with a batch size of <code>1<\/code>.<\/p>\n"}]}
{"QuestionId":42190777,"AnswerCount":0,"Tags":"<python><hadoop><apache-spark><neural-network><keras>","CreationDate":"2017-02-12T17:24:01.617","AcceptedAnswerId":null,"OwnerUserId":4338302.0,"Title":"Training Neural Networks on Hadoop Cluster","Body":"<p>I have been studying neural networks for some weeks. Furthermore even if I always used R the <em>Keras<\/em> library in <em>Python<\/em> was really helpful with someone with a <strong>small programming background like me<\/strong>. <\/p>\n\n<p><strong>Keras<\/strong> it's a very nice interface which <strong>allows the customization<\/strong> I need without even invoking the backend, unless for some custom loss metrics I used. <\/p>\n\n<p>Being that straightforward is also the <strong>Hardware specification<\/strong>, which for example <strong>allows you to switch from the CPU<\/strong> of the machine where you have your Python+Keras installed <strong>to<\/strong> the machine <strong>(compatible) GPU<\/strong>, <strong>allowing to exploit the strong parallelization of neural networks when training them<\/strong>. <\/p>\n\n<p><strong>I was wondering if there is anything which allows you to switch to hadoop cluster training of neural networks with the same kind of ease.\nMoreover is there some hadoop open source cluster available to do so?<\/strong><\/p>\n\n<p>Thank you for your help<\/p>\n","answers":[]}
{"QuestionId":42191267,"AnswerCount":0,"Tags":"<python><neural-network><keras><conv-neural-network>","CreationDate":"2017-02-12T18:08:05.053","AcceptedAnswerId":null,"OwnerUserId":3840626.0,"Title":"How to detect that there is no item of predefined categories on photo with keras NN?","Body":"<p>I try to implement face recognition for a set of faces. For solution I use a basic neural network in Keras.<\/p>\n\n<pre><\/pre>\n\n<p>There is set of grayscale images with size of 50x50px on input layer, vectors like [0 1 0 ...] on output layer.<\/p>\n\n<p>Everything works fine when I test model with photos with persons from training set. But if I call  with a photo of random stranger, I always get decision that there is one person from training set.For example for two persons I always get answers  or  and there is no way to get answer  or at least  (is last one possible sometimes?).<\/p>\n\n<p>I tried also to add to training set strangers and to assign new category for them (training output for them is ). But in this case my NN can't compute appropriate weights for ages.<\/p>\n\n<p>So is there any way to recognize persons from predefined set and ignore others?<\/p>\n","answers":[]}
{"QuestionId":42191362,"AnswerCount":2,"Tags":"<tensorflow><computer-vision><deep-learning><keras><image-segmentation>","CreationDate":"2017-02-12T18:15:45.343","AcceptedAnswerId":42193434.0,"OwnerUserId":297850.0,"Title":"loss function design to incorporate different weight for false positive and false negative","Body":"<p>I am trying to solve a semantic segmentation problem. In accordance with the real constraints, the criteria for false positive and the criteria for false negative is different. For instance, if a pixel is miscorrected as foreground is less desirable than a pixel is miscorrected as background. How to handle this kind of constraint in setting up the loss function. <\/p>\n","answers":[{"AnswerId":"42193434","CreationDate":"2017-02-12T21:24:06.077","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>You can use the <code>class_weight<\/code> parameter of <a href=\"https:\/\/keras.io\/models\/model\/#fit\" rel=\"noreferrer\"><code>model.fit<\/code><\/a> to weight your classes and, as such, punish misclassifications differently depending on the class.<\/p>\n\n<blockquote>\n  <p><code>class_weight<\/code>: optional dictionary mapping class indices (integers) to a weight (float) to apply to the model's loss for the samples from this class during training. 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>For example:<\/p>\n\n<pre><code>out = Dense(2, activation='softmax')\nmodel = Model(input=..., output=out)\nmodel.fit(X, Y, class_weight={0: 1, 1: 0.5})\n<\/code><\/pre>\n\n<p>This would punish the second class less than the first.<\/p>\n"},{"AnswerId":"52259231","CreationDate":"2018-09-10T13:47:14.650","ParentId":null,"OwnerUserId":"177449","Title":null,"Body":"<p>Check out the jaccard distance (or IOU) loss function in keras-contrib:<\/p>\n\n<blockquote>\n  <p>This loss is useful when you have unbalanced numbers of pixels within an image\n      because it gives all classes equal weight. However, it is not the defacto\n      standard for image segmentation.\n      For example, assume you are trying to predict if each pixel is cat, dog, or background.\n      You have 80% background pixels, 10% dog, and 10% cat. If the model predicts 100% background\n      should it be be 80% right (as with categorical cross entropy) or 30% (with this loss)?<\/p>\n<\/blockquote>\n\n<p>Source:\n<a href=\"https:\/\/github.com\/keras-team\/keras-contrib\/blob\/master\/keras_contrib\/losses\/jaccard.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/keras-team\/keras-contrib\/blob\/master\/keras_contrib\/losses\/jaccard.py<\/a><\/p>\n"}]}
{"QuestionId":42191525,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-12T18:29:59.773","AcceptedAnswerId":42544831.0,"OwnerUserId":5864449.0,"Title":"How to restore tensor summary to a variable","Body":"<p>I trained my network and saved the final weights as summaries like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now I want to use my classifier. I'm trying to load weights from summary:<\/p>\n\n<pre><\/pre>\n\n<p>So, how should I load the weights? I know about the \"global\" model saver but I would prefer saving only the data I need.<\/p>\n\n<p>Thanks in advance,\nAlexander<\/p>\n","answers":[{"AnswerId":"42210577","CreationDate":"2017-02-13T18:03:56.683","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>If you want a subset of the model, I would suggest you use the <code>var_list<\/code> parameter in the <code>tf.train.Saver<\/code> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables\" rel=\"nofollow noreferrer\">constructor<\/a>.<\/p>\n"},{"AnswerId":"42544831","CreationDate":"2017-03-02T00:39:08.520","ParentId":null,"OwnerUserId":"5864449","Title":null,"Body":"<p>Finally I've found the solution:<\/p>\n\n<pre><code>def tensor_summary_value_to_variable(value):\n    fb = numpy.frombuffer(v.tensor.tensor_content, dtype = numpy.float32)\n    shape = []\n    for d in v.tensor.tensor_shape.dim:\n        shape.append(d.size)\n    fb = fb.reshape(shape)\n    var = tf.Variable(fb)\n    return var\n<\/code><\/pre>\n"}]}
{"QuestionId":42191656,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-02-12T18:41:09.660","AcceptedAnswerId":42191829.0,"OwnerUserId":1128571.0,"Title":"Tensorflow learning XOR with linear function even though it shouldn't","Body":"<p>I am implementing a simple network in tensorflow and for pedagogical purposes, I am trying to show that the linear transformation:<\/p>\n\n<pre><\/pre>\n\n<p>cannot learn XOR. But the problem right now is that with my current implementation, it does! This suggests a bug in the code. Please elucidate?<\/p>\n\n<pre><\/pre>\n\n<p>Right now accuracy is at , even though it should be at .<\/p>\n","answers":[{"AnswerId":"42191829","CreationDate":"2017-02-12T18:56:24.780","ParentId":null,"OwnerUserId":"5799031","Title":null,"Body":"<p>tf.argmax(Y,1) will return [0,0,0,0]. This is not what you want.<\/p>\n"}]}
{"QuestionId":42192700,"AnswerCount":0,"Tags":"<python><tensorflow><jpeg>","CreationDate":"2017-02-12T20:13:53.417","AcceptedAnswerId":null,"OwnerUserId":7520802.0,"Title":"Tensorflow: Jpeg input for model","Body":"<p>I'm currently working with the transfer learning retrain.py file and have a basic VGG16 model I made with pretrained weights. I was looking at the inception v3 graph in tensorboard to see how it accepts jpeg images of varying values.\n<a href=\"https:\/\/i.stack.imgur.com\/qFSPI.png\" rel=\"nofollow noreferrer\">Inception V3 input<\/a> I went through what each operation is and I think I have it set up correctly, but apparently not.<\/p>\n\n<pre><\/pre>\n\n<p>This is what I use to test it.<\/p>\n\n<pre><\/pre>\n\n<p>So I'm just passing it a string of the name of the jpg in the folder. It's telling me that...<\/p>\n\n<pre><\/pre>\n\n<p>So what am I doing wrong? What can I do to get so I can transplant my model into the retrain.py file?<\/p>\n","answers":[]}
{"QuestionId":42193055,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-12T20:47:52.053","AcceptedAnswerId":42193508.0,"OwnerUserId":2148352.0,"Title":"Can't save tensorflow image to file","Body":"<p>I need to resize an image to certain size and save it to file, so I chose tf.image.resize_image_with_crop_or_pad function:<\/p>\n\n<pre><\/pre>\n\n<p>Failed with errors:<\/p>\n\n<pre><\/pre>\n\n<p>I tried to convert the Tensor to string using <strong>tf.as_string()<\/strong> but crashed with:<\/p>\n\n<pre><\/pre>\n\n<p>I use Tensorflow v0.12.0-rc0 on Linux Mint<\/p>\n","answers":[{"AnswerId":"42193508","CreationDate":"2017-02-12T21:30:58.123","ParentId":null,"OwnerUserId":"2476373","Title":null,"Body":"<p>You first need to encode the image from a tensor to a jpeg and then save it. Moreover, you should execute a session to evaluate your code:<\/p>\n\n<pre><code>import tensorflow as tf\n\nimage_decoded = tf.image.decode_jpeg(tf.read_file('1.jpg'), channels=3)\ncropped       = tf.image.resize_image_with_crop_or_pad(image_decoded, 200, 200)\nenc = tf.image.encode_jpeg(cropped)\nfname = tf.constant('2.jpg')\nfwrite = tf.write_file(fname, enc)\n\nsess = tf.Session()\nresult = sess.run(fwrite)\n<\/code><\/pre>\n"}]}
{"QuestionId":42193436,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><theano><keras>","CreationDate":"2017-02-12T21:24:12.260","AcceptedAnswerId":null,"OwnerUserId":497600.0,"Title":"Keras: convert pretrained weights between theano and tensorflow","Body":"<p>I would like to use <a href=\"https:\/\/gist.github.com\/baraldilorenzo\/07d7802847aaad0a35d3\" rel=\"nofollow noreferrer\">this pretrained model<\/a>.<\/p>\n\n<p>It is in theano layout, my code depends on tensorflow image dimension ordering.<\/p>\n\n<p>There is a guide on <a href=\"https:\/\/github.com\/fchollet\/keras\/wiki\/Converting-convolution-kernels-from-Theano-to-TensorFlow-and-vice-versa\" rel=\"nofollow noreferrer\">converting weights between the formats<\/a>.<\/p>\n\n<p>But this seems broken. In the section to convert theano to tensorflow, the first instruction is to load the weights into the tensorflow model.<\/p>\n\n<blockquote>\n  <p>Keras backend should be TensorFlow in this case.\n  First, load the Theano-trained weights into your TensorFlow model:<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>This raises an exception, the weight layouts would be incompatible.\nAnd if the  function would take theano weights for a tensorflow model, there wouldn't be any need to convert them.<\/p>\n\n<p>I took a look at the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/utils\/np_utils.py\" rel=\"nofollow noreferrer\"><\/a> function to see, whether I could do the necessary steps myself.<\/p>\n\n<p>The code is rather simple - I don't understand why the guide makes use of a tensorflow session. That seems unnecessary.<\/p>\n\n<p>I've copied the code from the pretrained model to create a model with tensorflow layer. This just meant changing the input shape and the  before adding any Convolutions.\nThen I used this loop:<\/p>\n\n<p> is the original model, created from the code I linked at the beginning.\n is the exact same model, but with tensorflow layout.<\/p>\n\n<pre><\/pre>\n\n<p>In the original code, there is a testcase: Prediction ran on the image of a cat.\nI've downloaded a cat image and tried the testcase with the original model: 285.\nThe converted model predicts 585.<\/p>\n\n<p>I don't know whether 285 is the correct label for a cat, but even if it isn't, the two models should be broken in the same way, I would expect the same prediction.<\/p>\n\n<p>What is the correct way of converting weights between models ?<\/p>\n","answers":[{"AnswerId":"57560299","CreationDate":"2019-08-19T16:01:35.790","ParentId":null,"OwnerUserId":"11530462","Title":null,"Body":"<p>You are right. The code is broken. As of now, there is a work around for this issue and the solution is described <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/5374#issuecomment-279256632\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>I have tested it myself and it worked for me.<\/p>\n\n<p>If you feel the answer is useful, please upvote it. Thanks!<\/p>\n"}]}
{"QuestionId":42193592,"AnswerCount":3,"Tags":"<python><numpy><matrix><tensorflow>","CreationDate":"2017-02-12T21:40:57.317","AcceptedAnswerId":null,"OwnerUserId":766551.0,"Title":"Difference in matrix multiplication tensorflow vs numpy","Body":"<p>I have a case where matrix multiplication of two matrices with certain dimensions work in numpy, but doesn't work in tensorflow.<\/p>\n\n<pre><\/pre>\n\n<p>This works as expected and prints:<\/p>\n\n<pre><\/pre>\n\n<p>However in tensorflow when I try to multiply placeholder and variable of the same shapes as the numpy arrays above I get an error<\/p>\n\n<pre><\/pre>\n\n<p>Results in<\/p>\n\n<pre><\/pre>\n\n<p>Why does this operation fail?<\/p>\n","answers":[{"AnswerId":"42193699","CreationDate":"2017-02-12T21:52:11.493","ParentId":null,"OwnerUserId":"6791223","Title":null,"Body":"<p>Don't know why <code>tf.matmul<\/code> does not support this kind of multiplication (may be one of the core developers could provide a meaningful answer). <\/p>\n\n<p>But if you just want to be able to multiply tensors in this way, take a look at <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/einsum\" rel=\"nofollow noreferrer\">tf.einsum<\/a> function. It could operate with tensors of arbitrary rank.<\/p>\n"},{"AnswerId":"43433213","CreationDate":"2017-04-16T02:46:24.037","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>People already told you that you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/einsum\" rel=\"nofollow noreferrer\">tf.einsum() <\/a> to get the result you want.<\/p>\n\n<pre><code>import tensorflow as tf\nx = tf.random_normal([10, 20, 30])\ny = tf.random_normal([30, 40])\nz = tf.einsum('ijk,kl-&gt;ijl', x, y)\n<\/code><\/pre>\n\n<p>The reason why <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul\" rel=\"nofollow noreferrer\">tf.matmul()<\/a> does not work the way you expected is written in the documentation.<\/p>\n\n<blockquote>\n  <p>The inputs must be matrices (or tensors of rank > 2, representing\n  batches of matrices), with matching inner dimensions, possibly after\n  transposition.<\/p>\n<\/blockquote>\n\n<p>In your case you have a matrix <code>y<\/code> and a tensor <code>x<\/code> (rank 3 > 2). In your case inner dimensions do not match. If you want, them to match, you will need to have something like this:<\/p>\n\n<pre><code>import tensorflow as tf\na, b, c = 12, 50, 20\nx = tf.random_normal([a, b, c])\ny = tf.random_normal([a, c, b])\nz = tf.matmul(x, y)\n<\/code><\/pre>\n\n<p>But clearly it calculates not the stuff you want.<\/p>\n"},{"AnswerId":"42200034","CreationDate":"2017-02-13T08:54:52.173","ParentId":null,"OwnerUserId":"766551","Title":null,"Body":"<p>As suggested by Dmytro <code>tf.einsum<\/code> can be used to multiply these two arrays. <\/p>\n\n<pre><code>x = np.ndarray(shape=(10,20,30), dtype = float)\ny = np.ndarray(shape=(30,40), dtype = float)\n<\/code><\/pre>\n\n<p>These two operations produce exactly the same result:<\/p>\n\n<pre><code>np.einsum('ijk,kl-&gt;ijl', x, y)\nnp.matmul(x,y)\n<\/code><\/pre>\n\n<p>And corresponding tensorflow operation also works<\/p>\n\n<pre><code>tf.einsum('ijk,kl-&gt;ijl', tf_x,tf_y)\n<\/code><\/pre>\n"}]}
{"QuestionId":42193779,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-02-12T22:00:17.623","AcceptedAnswerId":null,"OwnerUserId":3047361.0,"Title":"Can't import legacy_seq2seq from tensorflow.contrib","Body":"<p>I'm using tensorflow 0.12.1 on Python 3.5.2 on a Windows 10 64bit computer. For some reason, whenever I try to import  from , it always occurs the error: .<\/p>\n\n<p>What causes the problem and how can I fix it?<\/p>\n","answers":[{"AnswerId":"47625593","CreationDate":"2017-12-04T02:14:49.840","ParentId":null,"OwnerUserId":"2279150","Title":null,"Body":"<p>I meet this problem recently, and you should change from<\/p>\n\n<pre><code>from tensorflow.contrib import legacy_seq2seq\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>from tensorflow.contrib import seq2seq\n<\/code><\/pre>\n"},{"AnswerId":"42228500","CreationDate":"2017-02-14T14:18:31.583","ParentId":null,"OwnerUserId":"3237238","Title":null,"Body":"<p>The below code should work for your version. <\/p>\n\n<pre><code>from tensorflow.python.ops.seq2seq import sequence_loss\n<\/code><\/pre>\n\n<p>In general, download the code for your tensorflow version from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/releases\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/releases<\/a> and then read the code to find out where your function is located.<\/p>\n"},{"AnswerId":"42195147","CreationDate":"2017-02-13T01:00:01.560","ParentId":null,"OwnerUserId":"2283175","Title":null,"Body":"<p>I'm using <code>tf.nn.seq2seq.sequence_loss_by_example<\/code> - they've moved a lot of stuff from <code>tf.contrib<\/code> to main packages. This is because they updated their code, but not their examples - if you open github - you'll see a lot of requests to fix issues related to that!<\/p>\n"}]}
{"QuestionId":42194051,"AnswerCount":1,"Tags":"<python><python-2.7><python-3.x><numpy><tensorflow>","CreationDate":"2017-02-12T22:29:25.337","AcceptedAnswerId":42194197.0,"OwnerUserId":3813674.0,"Title":"Filter out non-zero values in a tensor","Body":"<p>Suppose I have an array: , and I want to filter out .<\/p>\n\n<p>I know that you can use  with a condition but the returned value still has 0's in it. Output of the following snippet is  . I also don't understand why  needs both  and . <\/p>\n\n<p>Is there anyway I can get rid of the 0's in the resulting tensor? <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42194197","CreationDate":"2017-02-12T22:49:46.723","ParentId":null,"OwnerUserId":"5788391","Title":null,"Body":"<p>First create a boolean mask to identify where your condition is true; then apply the mask to your tensor, as shown below. You can if you want use tf.where to index - however it returns a tensor using x&amp;y with the same rank as the input so without further work the best you could achieve would be something like [[[1 -1 3 5 -1 8 6]]] changing -1 with something that you would identify to remove later. Just using where (without x&amp;y) will give you the index of all values where your condition is true so a solution can be created using indexes if that is what you prefer. My recommendation is below for the most clarity.  <\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\ninput = np.array([[1,0,3,5,0,8,6]])\nX = tf.placeholder(tf.int32,[None,7])\nzeros = tf.cast(tf.zeros_like(X),dtype=tf.bool)\nones = tf.cast(tf.ones_like(X),dtype=tf.bool)\nloc = tf.where(input!=0,ones,zeros)\nresult=tf.boolean_mask(input,loc)\nwith tf.Session() as sess:\n out = sess.run([result],feed_dict={X:input})\n print (np.array(out))\n<\/code><\/pre>\n"}]}
{"QuestionId":42195414,"AnswerCount":1,"Tags":"<python><numpy><theano><pymc><pymc3>","CreationDate":"2017-02-13T01:38:40.693","AcceptedAnswerId":42216531.0,"OwnerUserId":7313632.0,"Title":"how to add deterministic vector operations in PYMC3?","Body":"<p>how can deterministic vector operations be implemented in PYMC3? for example the model:<\/p>\n\n<pre><\/pre>\n\n<p>M is mean of Gaussian observations and S is standard deviation. The standard deviation is assumed to be uniformly distributed in [0, |1\/M|] (abs value necessary when M is negative). <\/p>\n\n<p>this code:<\/p>\n\n<pre><\/pre>\n\n<p>has the error:<\/p>\n\n<pre><\/pre>\n\n<p>do i need to use theano to achieve this operation on vectors?<\/p>\n","answers":[{"AnswerId":"42216531","CreationDate":"2017-02-14T01:51:10.073","ParentId":null,"OwnerUserId":"815542","Title":null,"Body":"<p>I guess the problematic part of your model is <code>1\/M<\/code>. Because this goes to infinity as M approach to 0. In fact in your example, the first proposed value for M is 0 (the mean of the lower and upper boundaries) hence the error of \"no finite value\" you are getting (the error comes from the variable S).<\/p>\n\n<p>One way to fix this is to pass a reasonable <code>testval<\/code> value to M (in this case anything but 0), for example:<\/p>\n\n<pre><code>M = pm.Uniform(\"M\", -5., 5., shape=size, testval=0.1)\n<\/code><\/pre>\n\n<p>Now M will be initialized at 0.1 (instead of 0.) <\/p>\n\n<p>As a side note, intuitively I think NUTS will have trouble to properly sample from this model, thus I think Metropolis could be a better choice here, but feel free to experiment with different samplers.<\/p>\n"}]}
{"QuestionId":42195922,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-13T02:48:31.927","AcceptedAnswerId":null,"OwnerUserId":3930944.0,"Title":"Tensorflow CIFAR10 Multi GPU - Why Combined Loss?","Body":"<p>In the TensorFlow <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py#L80-L98\" rel=\"nofollow noreferrer\">CIFAR10 example<\/a>, trained over multiple GPUs, the loss seems to be combined for each \"tower\", and the gradient is calculated from this combined loss.<\/p>\n\n<pre><\/pre>\n\n<p>I'm new to TensorFlow, but from my understanding, every time  is called,  is run and the loss from the current batch is being stored in the collection.<\/p>\n\n<p>Then  is called, and <em>all<\/em> the losses are being retrieved from the collection. Then  op is adding all the retrieved loss tensors from this \"tower\" together.<\/p>\n\n<p>I expected the loss to be <em>just<\/em> from the current training step\/batch, not <em>all<\/em> batches.<\/p>\n\n<p>Am I misunderstanding something? Or is there a reason for combining the losses together?<\/p>\n","answers":[{"AnswerId":"43768033","CreationDate":"2017-05-03T19:00:07.250","ParentId":null,"OwnerUserId":"4970778","Title":null,"Body":"<p>If weight decay is enabled, it will also add it to the losses collection. \nTherefore, for each tower(scope), it will add_n all the losses: cross_entropy_mean and weight_decay. <\/p>\n\n<p>Then Gradients are calculated for each tower(scope). At the end all the gradients for different towers (scopes) will get averaged in the average_gradients. <\/p>\n"},{"AnswerId":"45719277","CreationDate":"2017-08-16T17:03:35.397","ParentId":null,"OwnerUserId":"1896918","Title":null,"Body":"<blockquote>\n  <p>Why combined loss<\/p>\n<\/blockquote>\n\n<p>The example you are referring is a example of data parallelism over multiple gpus. Data parallelism helps towards training deeper model with bigger batch_size. In this setting you need to combine loss from the gpus as each of the gpus is holding one part of the input batch (loss and gradients corresponding to that input part).  One illustration is provided in the following example from <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">tensorflow data parallism example<\/a>. <\/p>\n\n<p>Note: In case of model parallelism different subgraph of the model run on separate gpus and intermediate outputs are collected by the master. <\/p>\n\n<h1>example<\/h1>\n\n<p>if you want to train the model using a batch size of 256, for a deeper model (for example, resnet\/inception)that mayn't fit into one single gpu (for example a 8 GB memory), so you can split the batch into two batches of size 128 and do forward pass of the model using the two batches on separate gpus and compute loss and gradients. The computed (loss. gradients) from each of the gpus are collected and averaged over. the averaged gradient is used to update the model parameters. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xTPHn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xTPHn.png\" alt=\"enter image description here\"><\/a>   <\/p>\n"}]}
{"QuestionId":42196007,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-13T03:00:28.943","AcceptedAnswerId":null,"OwnerUserId":5350430.0,"Title":"error happened during using TensorArray and while_loop to do dynamic rnn","Body":"<p>My intention is to do CNN (VGG16) and then push the output of it to two-layer lstm for every frame. However, error happens when I try to use sess.run(). That suggests the graph is constructed correctly. So where's my bug?\nHere's the error info.<\/p>\n\n<pre><\/pre>\n\n<p>And here's some of my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42196387","CreationDate":"2017-02-13T03:46:50.683","ParentId":null,"OwnerUserId":"5350430","Title":null,"Body":"<p>Problem solved. It seems if the VggNet is defined in the while loop, the loss has to be computed inside the loop.\nHere's the changed code:<\/p>\n\n<pre><code>lstm=tf.nn.rnn_cell.BasicLSTMCell(128)\ncell=tf.nn.rnn_cell.MultiRNNCell([lstm]*2)\nstate=cell.zero_state(1, tf.float32)\n\ninputImgA=tf.TensorArray(tf.string, length)\noutputLSTM=tf.TensorArray(tf.float32, length)\nlossLSTM=tf.TensorArray(tf.float32, length)\n\nimg=sequence_parsed['imgs_list']\ninputImgA=inputImgA.unpack(img)\n\ni=tf.constant(0)\n\n\ndef cond(i, state, inputImgA, outputLSTM, lossLSTM):\n    return tf.less(i, length)\n\ndef body(i, state, inputImgA, outputLSTM, lossLSTM):\n    imcontent=inputImgA.read(i)\n    image=tf.image.decode_jpeg(imcontent, 3, name='decode_image')\n    with tf.variable_scope('Image_Process'):\n        image=tf.image.resize_images(image, [224, 224])\n        channels = tf.split(2, 3, image)\n        channels[0] -= _R_MEAN\n        channels[1] -= _G_MEAN\n        channels[2] -= _B_MEAN\n        image=tf.concat(2, channels)\n        images=tf.expand_dims(image, 0)\n\n    net, end = VggNet(images, is_training=True)\n    output, state=cell(net, state)\n    outputLSTM=outputLSTM.write(i, output)\n\n    loss2=tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)\n    loss2=tf.add_n(loss2)\n    loss1=tf.nn.sparse_softmax_cross_entropy_with_logits(output, label)\n    lossLSTM=lossLSTM.write(i, loss1+loss2)\n    return (i+1, state, inputImgA, outputLSTM, lossLSTM) \n\n\n_, _, _, outputLSTM, lossLSTM=tf.while_loop(cond, body, [i, state, inputImgA, outputLSTM, lossLSTM])\n\noutput=outputLSTM.pack()\ntotal_loss=lossLSTM.pack()\ntotal_loss=tf.reduce_mean(total_loss)\n<\/code><\/pre>\n"}]}
{"QuestionId":42196271,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-13T03:32:50.683","AcceptedAnswerId":null,"OwnerUserId":2638485.0,"Title":"Gaussian Kernel in Tensorflow","Body":"<p><a href=\"https:\/\/github.com\/torch\/image\/blob\/master\/doc\/tensorconstruct.md#res-imagegaussiansize-sigma-amplitude-normalize-\" rel=\"nofollow noreferrer\">Torch7 has a library for generating Gaussian Kernels<\/a> over a fixed support. Is there anything comparable in Tensorflow? I see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/contrib.distributions\/multivariate_distributions#MultivariateNormalDiag\" rel=\"nofollow noreferrer\">these distributions in TF contrib<\/a>, but I am not sure how I would use that to get a kernel (on a fixed grid with weights summing to 1).<\/p>\n","answers":[]}
{"QuestionId":42196427,"AnswerCount":0,"Tags":"<caffe><lmdb><imagenet>","CreationDate":"2017-02-13T03:51:36.660","AcceptedAnswerId":null,"OwnerUserId":7555683.0,"Title":"caffe convert_imageset pending","Body":"<p>When I convert ImageNet dataset to lmdb, convert_imageset will suspend and no longer process image data.Can you tell me how to solve this problem? I have tried it many time.I checked the log.\nEvery time the number of images that was processed is different.<\/p>\n","answers":[]}
{"QuestionId":42197235,"AnswerCount":1,"Tags":"<python-3.x><keras>","CreationDate":"2017-02-13T05:32:27.397","AcceptedAnswerId":42214366.0,"OwnerUserId":5636645.0,"Title":"How to fix the following issue, plotting a keras model?","Body":"<p>Hello I am working with keras, my code is fine, until the point where I tried to graph as follows:<\/p>\n\n<pre><\/pre>\n\n<p>this is a part of my code:<\/p>\n\n<pre><\/pre>\n\n<p>and this is the output:<\/p>\n\n<pre><\/pre>\n\n<p>I tried:<\/p>\n\n<pre><\/pre>\n\n<p>however I got:<\/p>\n\n<pre><\/pre>\n\n<p>I also tried:<\/p>\n\n<pre><\/pre>\n\n<p>and I got:<\/p>\n\n<pre><\/pre>\n\n<p>So I would like to receive support to overcome this situation, <\/p>\n","answers":[{"AnswerId":"42214366","CreationDate":"2017-02-13T22:02:57.637","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>You have to install the graphviz binaries, you cannot install these with pip (that instead installs a python wrapper). Use your Linux distribution's package manager to install graphviz and then it should work.<\/p>\n"}]}
{"QuestionId":42197450,"AnswerCount":2,"Tags":"<image><python-3.x><csv><keras>","CreationDate":"2017-02-13T05:52:10.467","AcceptedAnswerId":null,"OwnerUserId":7526688.0,"Title":"Importing images to prep for keras","Body":"<p>I am trying to import a bunch of images and get them ready for keras. The goal here is to have an array of the following dimensions. (length, 160,329,3). As you can see my reshape function is commented out. The \"print(images.shape) line returns (8037,). Not sure how to proceed to get the right array dimensions. For reference the 1st column in the csv file is a list of paths to the image in question. I have a function below that combines the path of the image inside the folder and the path to the folder. <\/p>\n\n<p>When I run the commented out reshape function I get the following error.  \"ValueError: cannot reshape array of size 8037 into shape (8037,160,320,3)\"<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42206801","CreationDate":"2017-02-13T14:45:49.497","ParentId":null,"OwnerUserId":"7558284","Title":null,"Body":"<p>Can you print the shape of an <code>image<\/code> before it is appended to <code>images<\/code> to verify it is what you expect? Even better would be adding an <code>imshow<\/code> in the loop to make sure you're loading the images you expect (only need to do for one or two). <code>cv2.imread<\/code> does not throw an error if there isn't an image at the file path you give it, so your array might be all <code>None<\/code> which would yield the exact behavior you've described.<\/p>\n\n<p>If that is the problem, check the <code>img<\/code> variable and make sure it's pointing exactly where you want it to.<\/p>\n"},{"AnswerId":"42240428","CreationDate":"2017-02-15T03:54:20.773","ParentId":null,"OwnerUserId":"7526688","Title":null,"Body":"<p>Turns out it was including the first line of the CSV file which was heading.  After I sorted that out it ran great.  It gave me the requested shape.  <\/p>\n\n<pre><code>images=[]\nfor i in range(1,len(labels)):\n    img=(m[i][1])\n    img=img.lstrip()\n    path='\/Users\/user\/Desktop\/data\/'\n    img=path+img\n\n    image=cv2.imread(img)\n    images.append(image)\n<\/code><\/pre>\n"}]}
{"QuestionId":42198839,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><linear-regression><gradient-descent>","CreationDate":"2017-02-13T07:40:34.337","AcceptedAnswerId":null,"OwnerUserId":1616677.0,"Title":"Pure Stohastic Gradient Descent scaling anomaly","Body":"<p>I train linear regression (y = b0 + b1 * x) with SGD, batch_size=1 and one data pass (epoch). It converges when scaling the input to unreasonable values (see below). When scaling to \"standard\" values (calculating unit variance and mean on the input dataset) I cannot get convergence to any reasonable error value, no matter how I change learninig rate\/iterations. Both x and y are divided to scale_ and centered to mean_.<\/p>\n\n<p>Training data:\nx = [1 .. 70] y = [15.39 ... 356.76]<\/p>\n\n<p>Standard scaling:\niterations = 500 learning_rate=0.00003\nx.scale_ [ 20.20519735] x.mean_ [ 35.5] y.scale_ [ 100.3283148] y.mean_ [ 187.43898263]<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/k0U9g.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/k0U9g.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Abnoraml scaling: iterations=500 learning_rate=0.001 x.scale_ [ 2.5] x.mean_ [ 2.5] y.scale_ [ 3.] y.mean_ [ 3.]<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vvxj4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vvxj4.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>You can see that the learning rate is adjusted according to the scales quotient, so this is not just a learning rate problem (I've tried different values as well).<\/p>\n","answers":[]}
{"QuestionId":42201368,"AnswerCount":1,"Tags":"<python><ubuntu><tensorflow><pycharm>","CreationDate":"2017-02-13T10:05:53.430","AcceptedAnswerId":42373797.0,"OwnerUserId":4795896.0,"Title":"Error using Tensorflow virtualenv in Pycharm (Ubuntu 16)","Body":"<p>I've installed Tensorflow within a virtual environment on my machine. It works from the command line but when I try to run it from Pycharm it gives the following error: <a href=\"https:\/\/i.stack.imgur.com\/raJWe.png\" rel=\"nofollow noreferrer\">Pycharm Console Output<\/a><\/p>\n\n<p>The project interpreter is set to virtualenv at ~\/tensorflow and for the interpreter paths the following ones are set:<\/p>\n\n<ul>\n<li>\/home\/user\/tensorflow\/lib\/python3.5<\/li>\n<li>\/home\/user\/tensorflow\/lib\/python3.5\/plat-x86_64-linux-gpu<\/li>\n<li>\/home\/user\/tensorflow\/lib\/python3.5\/lib-dynload<\/li>\n<li>\/usr\/lib\/python3.5<\/li>\n<li>\/usr\/lib\/python3.5\/plat-x86_64-linux-gnu<\/li>\n<li>\/home\/user\/tensorflow\/lib\/python3.5\/site-packages<\/li>\n<li>\/usr\/lib\/python3.5\/site-packages<\/li>\n<li>\/usr\/local\/lib\/python3.5\/dist-packages<\/li>\n<li>\/usr\/lib\/python3\/dist-packages<\/li>\n<\/ul>\n\n<p>The code (from a tutorial) I've tried to run is the following:<\/p>\n\n<pre><\/pre>\n\n<p>So what could this error be related to?<\/p>\n","answers":[{"AnswerId":"42373797","CreationDate":"2017-02-21T17:14:46.577","ParentId":null,"OwnerUserId":"4795896","Title":null,"Body":"<p>Ok, so it works if you call Pycharm from the terminal. \nI guess that the terminal context is given to pycharm the moment you load it from the command line. \n(Pycharm is executed from the command line by executing the pycharm.sh script. If you don't know how where it is then type 'locate pycharm.sh')<\/p>\n"}]}
{"QuestionId":42201565,"AnswerCount":1,"Tags":"<tensorflow><convolution><tf-slim><deconvolution>","CreationDate":"2017-02-13T10:14:57.997","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Using Tensorflow.slim to apply convolution2d_transpose","Body":"<p>I am trying to apply 2 convolutional layers with the tf.slim.conv2d function, they basically reduce the size of my input image by half each time. Then I want to apply the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L906\" rel=\"nofollow noreferrer\">convolution2d_transpose<\/a> to get my original image shape back. The problem is I don't exactly know how to use the transpose convolution function, and the documentation is not much help. <\/p>\n\n<p>I am using a custom wrapper, but here is what I have so far:<\/p>\n\n<pre><\/pre>\n\n<p>How can I apply the convolution_transpose function to reverse the effect of these two layers now ? <\/p>\n","answers":[{"AnswerId":"43852112","CreationDate":"2017-05-08T15:54:35.120","ParentId":null,"OwnerUserId":"7981678","Title":null,"Body":"<p>According to the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/layers.py#L1068\" rel=\"nofollow noreferrer\">tensorflow api-docs<\/a> link you provided above:<\/p>\n\n<pre><code>def convolution2d_transpose(\ninputs,\nnum_outputs,\nkernel_size,\nstride=1,\npadding='SAME',\ndata_format=DATA_FORMAT_NHWC,\nactivation_fn=nn.relu,\nnormalizer_fn=None,\nnormalizer_params=None,\nweights_initializer=initializers.xavier_initializer(),\nweights_regularizer=None,\nbiases_initializer=init_ops.zeros_initializer(),\nbiases_regularizer=None,\nreuse=None,\nvariables_collections=None,\noutputs_collections=None,\ntrainable=True,\nscope=None):\n<\/code><\/pre>\n\n<p>you can, for your example, utilize it like this:<\/p>\n\n<pre><code>slim.convolution2d_transpose(input_tensor, 32, [4,4], [2,2], scope='output')\n<\/code><\/pre>\n"}]}
{"QuestionId":42202440,"AnswerCount":0,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-02-13T11:00:44.863","AcceptedAnswerId":null,"OwnerUserId":5027315.0,"Title":"Feed Slim Training Loop with Data Batches from Keras ImageDataGenerator","Body":"<p>I have an ImageDataGenerator object from Keras which yields batches indefinitely. How can I feed n batches of size s each to training loop of tensorflow-slim.<\/p>\n\n<pre><\/pre>\n\n<p>Please note that I want to use training loop provided by tensor-slim rather than writing my own routine for training using native tensorflow.<\/p>\n","answers":[]}
{"QuestionId":42203190,"AnswerCount":0,"Tags":"<python><input><tensorflow><shape>","CreationDate":"2017-02-13T11:40:21.730","AcceptedAnswerId":null,"OwnerUserId":7499201.0,"Title":"Tensorflow: Input shape change when RNN cell is called","Body":"<p>I was digging into the Tensorflows source code in order to fix a input shape problem I have with their Translation model. The model creates a seq2seq model, which again calls a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r0.12\/tensorflow\/python\/ops\/rnn.py\" rel=\"nofollow noreferrer\">rnn model<\/a> . On line 226, my RNNCell is called, but the input sent in is not the same as the one caught by the <strong>call<\/strong> method in my cell. Am I missing something? (Note that call_cell() is a lambda function)<\/p>\n\n<p> in the rnn.py gives the following output:<\/p>\n\n<blockquote>\n  <p>Tensor(\"encoder0:0\", shape=(8, 8), dtype=int32)<\/p>\n<\/blockquote>\n\n<p>while  in the RNNCell gives:<\/p>\n\n<blockquote>\n  <p>Tensor(\"model_with_buckets\/embedding_attention_seq2seq\/RNN\/EmbeddingWrapper\/embedding_lookup:0\", shape=(64, 1000), dtype=float32, device=\/device:CPU:0)<\/p>\n<\/blockquote>\n\n<p>As you can see, the input is not the same, but I cannot see what else that is happening. My code breaks when call_cell is called in the rnn.py file. Is the tensor affected by some operations I can't see, or is it simply another tensor the RNNCell receives?<\/p>\n","answers":[]}
{"QuestionId":42203760,"AnswerCount":0,"Tags":"<tensorflow><lstm>","CreationDate":"2017-02-13T12:12:37.310","AcceptedAnswerId":null,"OwnerUserId":6777494.0,"Title":"Tensorflow LSTM memory usage","Body":"<p>I'm trying  sequence labeling. Before add embedding layer I put input data directly to  layer. But it gives me GPU memory error even if batch size is 1.<\/p>\n\n<p> is 330, should I have to change model or adding embedding layer will work? I'm using Titan X GPU with 12 GB RAM.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42203806,"AnswerCount":1,"Tags":"<deep-learning><caffe><convolution><dimensions>","CreationDate":"2017-02-13T12:14:54.260","AcceptedAnswerId":42241003.0,"OwnerUserId":6786718.0,"Title":"Is the dimensions of layers rounded up or rounded down when it is not an integer?","Body":"<p>Sometimes it may happen that the filter mask of a conv or pool layer does not fit into the input volume.  For example I have a  input and a  filter mask and do not use any padding, while stride has the d value of .. <\/p>\n\n<p>Using the formula:<\/p>\n\n<pre><\/pre>\n\n<p>Where W is the input volume, F is the filter size, P is padding and S is stride - all one dimensional. <\/p>\n\n<pre><\/pre>\n\n<p>What will be the used output dimensions? Will it get floored to \"2\" or ceiled to \"3\" ?<\/p>\n","answers":[{"AnswerId":"42241003","CreationDate":"2017-02-15T04:53:42.623","ParentId":null,"OwnerUserId":"5009875","Title":null,"Body":"<p>It will get floored to 2 considering caffe implementation.\nLine 18 of <code>$CAFFE_ROOT\/src\/caffe\/layers\/conv_layer.cpp<\/code> <\/p>\n\n<pre><code> const int output_dim = (input_dim + 2 * pad_data[i] - kernel_extent)\n        \/ stride_data[i] + 1;\n<\/code><\/pre>\n\n<p>Basically ,consider a filter of size SxS which is sliding on input.\nAs it moves at the end ,it might have to face an input equal to its size or less than its size.\nIf it is less than S ,then that part is ignored.This is how caffe implements it.<\/p>\n"}]}
{"QuestionId":42204436,"AnswerCount":0,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-02-13T12:47:39.993","AcceptedAnswerId":null,"OwnerUserId":2116638.0,"Title":"Python 3.5 and Tensorflow","Body":"<p>Good day,<\/p>\n\n<p>I am have installed Python 3.5 with all the extra libs required to make Tensorflow work in Windows 10. The IDE I'm using is PyCharm community edition and Python seems to be running fine.<\/p>\n\n<p>However, I'm getting the following error from Python:<\/p>\n\n<p>AttributeError: 'numpy.ndarray' object has no attribute 'values'<\/p>\n\n<p>My code is taken from the Tensorflow Examples page:\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/monitors\/\" rel=\"nofollow noreferrer\">Tensorflow Page<\/a>\nWhich was giving depreciated problems in the first case. So I have gone ahead and started fixing the problems but I am starting to get a little frustrated banging my head against a wall.<\/p>\n\n<p>Here is my code so far:<\/p>\n\n<pre><\/pre>\n\n<p>======= TRACE\/ERROR LOG ======<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42204950,"AnswerCount":1,"Tags":"<tensorflow><normalization>","CreationDate":"2017-02-13T13:13:31.343","AcceptedAnswerId":null,"OwnerUserId":7557760.0,"Title":"Tensorflow How to apply batch normalization to inference","Body":"<p>Hi. I am quite new to tf and the question is how to apply the batch normalization to inference. I applied batch normalization with tensorflow during training and testing. The code is as following and as you can see, I treat training and testing with different mean and var value. Now the training is completed and I want to apply the model to practice. Such as use a demo load the ckpt file and test just one example. In this case, how can I normalize this one case? Is there any approach I can save the mean and var of BN after each epoch of training and restore later? Thank you very very much!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42205690","CreationDate":"2017-02-13T13:51:56.673","ParentId":null,"OwnerUserId":"7557760","Title":null,"Body":"<p>I am not sure but do you guys think this gonna be right? I saved the shift(offset) and scale, which yield during training. And when I need to apply the model to practice, I just use shift and scale to replace the mean and var.And set the shift and var of test case as zero.<\/p>\n\n<pre><code>    if not test:\n        input_BN = tf.nn.batch_normalization(input, mean, var, shift, scale, epsilon)\n    else:\n        input_BN = tf.nn.batch_normalization(input, shift, scale, 0, 0, epsilon)\n<\/code><\/pre>\n"}]}
{"QuestionId":42205205,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-13T13:27:00.773","AcceptedAnswerId":42278731.0,"OwnerUserId":6183298.0,"Title":"TensorFlow Python script getting killed","Body":"<p>I am a beginner in TensorFlow. My TensorFlow script abruptly exits saying . My code is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I have omitted the parts which I know for sure are not causing problems. I know which parts are fine because the script was not giving any troubles earlier. After adding certain parts (mainly the <strong>summary writer operations<\/strong>), the script abruptly says  and exits after executing a certain number of training operations. I suspect this is due to a memory leak but I can't detect it.<\/p>\n","answers":[{"AnswerId":"42278731","CreationDate":"2017-02-16T15:54:45.333","ParentId":null,"OwnerUserId":"6746595","Title":null,"Body":"<p>I ran into a similar problem just a few days ago. In my case I had some operations that turned out to be computationally very heavy, as I learnt later. As soon as I reduced the size of my tensors, the message disappeared and my code ran. \nI can't tell exactly what is the cause of the issue in your case, but from my experience and from what you say (that error only comes up when adding summary) I would suggest to fiddle with the size of your X_valid, Y_valid. It might just be that the writer can't cope with too much data...<\/p>\n"}]}
{"QuestionId":42206870,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><deep-learning><softmax>","CreationDate":"2017-02-13T14:49:35.543","AcceptedAnswerId":null,"OwnerUserId":1861627.0,"Title":"NaN in custom sparse softmax for pointing networks","Body":"<p>I am trying to implement some sort of <a href=\"https:\/\/arxiv.org\/abs\/1506.03134\" rel=\"nofollow noreferrer\">pointing network<\/a>. Since the output of the pointing layer has no fixed dimension, I had to implement some sort of masking of the output, and I came out with this:<\/p>\n\n<pre><\/pre>\n\n<p>After some training steps, I have some  occurring the  tensor. I couldn't figure out <em>why<\/em> this happens, but googling around, I came up to <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/3162\" rel=\"nofollow noreferrer\">this discussion about something similar<\/a> that might happen in . Could it be caused by some saturation in the softmax implementation I had to mimic in order not to keep in account the zeroes? <\/p>\n\n<p>Thanks in advance,\nGiulio<\/p>\n","answers":[]}
{"QuestionId":42206906,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-13T14:50:56.690","AcceptedAnswerId":42210345.0,"OwnerUserId":6306884.0,"Title":"Alternative plan of tf.floor","Body":"<p>One of my operaction need integer, but output of convolution is float.<br>\nIt means I need to use , , ...etc to handle it.<br>\nBut these operactions cause  gradients, since operactions like  are not differentiable    <\/p>\n\n<p>So, I tried something like below   <\/p>\n\n<p><strong>First<\/strong>. detour<\/p>\n\n<pre><\/pre>\n\n<p>But output of  is 500 or 0, I don't think this is a reasonable gradient.    <\/p>\n\n<p><strong>Second<\/strong>. override gradient function of floor  <\/p>\n\n<pre><\/pre>\n\n<p>output of  is 500 or 1, doesn't work too. <\/p>\n\n<p><strong>Question<\/strong>: A way to get integer and keep back propagation work fine (value like 2.0, 5.0 is ok)<\/p>\n","answers":[{"AnswerId":"42210345","CreationDate":"2017-02-13T17:51:07.457","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>In general, it's not inadvisable to solve discrete problem with gradient descent. You should be able express, to some extent integer solvers in TF but you're more or less on your own.<\/p>\n\n<p>FWIW, the floor function looks like a saw. Its derivative is a constant function at 1 with little holes at every integer. At these positions you have a Dirac functional pointing downwards, like a rake if you wish. The Dirac functional has finite energy but no finite value.<\/p>\n\n<p>The canonical way to tackle these problems is to relax the problem by \"relaxiing\" the hard floor constraint with something that is (at least once) differentiable (smooth).<\/p>\n\n<p>There are multiple ways to do this. Perhaps the most popular are:<\/p>\n\n<ol>\n<li>Hack up a function that looks like what you want. For instance a piece-wise linear function that slopes down quickly, but not vertically.<\/li>\n<li>Replace step functions by sigmoids<\/li>\n<li>Use a filter approximation which is well understood if it's a time series<\/li>\n<\/ol>\n"}]}
{"QuestionId":42206915,"AnswerCount":1,"Tags":"<theano><keras><theano-cuda>","CreationDate":"2017-02-13T14:51:22.433","AcceptedAnswerId":null,"OwnerUserId":1190427.0,"Title":"Keras with Theano on GPU","Body":"<p>While trying to run my Keras code on GPU (CUDA installed), I am not able to execute the following statement, as has been suggested on many online references.<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the following error. <\/p>\n\n<pre><\/pre>\n\n<p>I have tried the other mode suggested as well from inside the code.<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error.<\/p>\n\n<pre><\/pre>\n\n<p>Apart from knowing how to make it run, I would also take this opportunity to ask a simpler question. How to know in Windows what is my device i.e. whether 'gpu' or 'gpu1' or 'gpu0'? I have tried all 3 for my case but it hasn't yielded result. <\/p>\n\n<p>Any suggestions will be appreciated.<\/p>\n","answers":[{"AnswerId":"44331861","CreationDate":"2017-06-02T14:55:51.503","ParentId":null,"OwnerUserId":"2787380","Title":null,"Body":"<p>The best way is using <code>THEANO_FLAGS<\/code> before run code, because the config variables cannot be changed after importing Theano, try this:<\/p>\n\n<pre><code>import os\nos.environ['THEANO_FLAGS'] = \"device=cuda,force_device=True,floatX=float32\"\nimport theano\n<\/code><\/pre>\n"}]}
{"QuestionId":42207055,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-13T14:58:09.337","AcceptedAnswerId":null,"OwnerUserId":7019234.0,"Title":"Topic Modelling with tensorflow","Body":"<p>I want to do Topic Modelling with Tensorflow, I know there are algorithms like LDA, but is there any neural network representation in Tensorflow which can be used to do the same thing better or atleast on par with LDA.\nAny tips \/answers will be greatly appreciated. Also if you can please provide examples of source code etc? <\/p>\n","answers":[]}
{"QuestionId":42207266,"AnswerCount":2,"Tags":"<python><tensorflow><classification>","CreationDate":"2017-02-13T15:09:34.437","AcceptedAnswerId":null,"OwnerUserId":7558325.0,"Title":"Tensorflow python classification training with subcategories","Body":"<p>The problem im facing is that i need to train a tensorflow graph with \"maincategories\" and \"subcategories\". <\/p>\n\n<p>I already know how to fully train a graph with \"maincategories\".<\/p>\n\n<p>For example i have 3 different animal types: cat, dog and elephant. <\/p>\n\n<p>If i train a graph with pictures to those 3 types, it would return scores for the 3 types like: <\/p>\n\n<p>cat 0,7 <\/p>\n\n<p>dog 0,3<\/p>\n\n<p>elephant 0,02<\/p>\n\n<p>What i want tho is to train those 3 types and subcategories for those types, for example: <\/p>\n\n<p>cat: white, fat, ...<\/p>\n\n<p>dog: husky, ...<\/p>\n\n<p>And when i now classify an image with this graph, i would like to get an output with the maincategories and the subcategories, like this: <\/p>\n\n<p>cat: 0,7 -> white:0,3 -> fat:0,5<\/p>\n\n<p>dog: 0,2 .....<\/p>\n\n<p>currently the only way i see is to train a graph with the maincategories and to train individual graphs for all the maincategories with the subcategories, then code something like:<\/p>\n\n<pre><\/pre>\n\n<p>In the example the catgraph would be a trained with the different cattypes as categories.<\/p>\n\n<p>This however isnt very ellegant and im hoping there is better way of doing this. <\/p>\n\n<p>Im thankfull for any help you could give me. <\/p>\n\n<p>~Meow<\/p>\n","answers":[{"AnswerId":"45690192","CreationDate":"2017-08-15T09:31:51.587","ParentId":null,"OwnerUserId":"4825076","Title":null,"Body":"<p>You have to train the subcategories individually. Point the <code>--image_dir<\/code> to the root of your subcategories. NOTE: the subcategories here am refering to are the subfolders. Then get the output of each subcategory as you would with the main category. I hope this sheds some light. You can refer to <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">Tensorflow image retraining<\/a><\/p>\n"},{"AnswerId":"42219837","CreationDate":"2017-02-14T07:03:27.757","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>You can train a graph with two outputs, sharing everything but the last layer (or up to a layer you like). Then you would have a (say) logloss for each, and optimize against the sum of the losses. If you are bottlenecked by the learning capacity of the network, you may want to weigh one loss more than the other.<\/p>\n\n<p>Remember to track both losses in the summary for easy visualization. You can add the losses to the losses collection.<\/p>\n"}]}
{"QuestionId":42207554,"AnswerCount":2,"Tags":"<python><matrix><tensorflow>","CreationDate":"2017-02-13T15:22:54.787","AcceptedAnswerId":null,"OwnerUserId":1057098.0,"Title":"Swapping elements within a matrix rows and columns - TensorFlow scatter_nd","Body":"<p>I am trying to use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/slicing_and_joining#scatter_nd\" rel=\"nofollow noreferrer\">scatter_nd<\/a> function in TensorFlow to reorder elements within rows of a Matrix.  For example, suppose I have the code:<\/p>\n\n<pre><\/pre>\n\n<p>This reorders the rows of the  matrix.  <\/p>\n\n<p>Instead of only being able to reorder the rows, I'd like to reorder the individual elements within each row as well.  If I just have a vector (Tensor of rank 1), then this example works:<\/p>\n\n<pre><\/pre>\n\n<p>What I really care about is to be able to swap elements within each row in , as I had done in , but do it for each row of .  I've tried various combinations of  but keep getting errors that the sizes are inconsistent thrown by the  function.  <\/p>\n","answers":[{"AnswerId":"48019530","CreationDate":"2017-12-29T08:12:45.517","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>Suppose you want to swap elements in the second dimension either keeping the first dimension order or not. <\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\n\n\ndef prepare_fd(fd_indices, sd_dims):\n    fd_indices = tf.expand_dims(fd_indices, 1)\n    fd_indices = tf.tile(fd_indices, [1, sd_dims])\n    return fd_indices\n\n# define the updates\nupdates = tf.constant([[11, 12, 13, 14],\n                       [21, 22, 23, 24],\n                       [31, 32, 33, 34]])\nsd_dims = tf.shape(updates)[1]\n\nsd_indices = tf.constant([[1, 0, 2, 3], [0, 2, 1, 3], [0, 1, 3, 2]])\nfd_indices_range = tf.range(0, limit=tf.shape(updates)[0])\nfd_indices_custom = tf.constant([2, 0, 1])\n\n# define the indices\nindices1 = tf.stack((prepare_fd(fd_indices_range, sd_dims), sd_indices), axis=2)\nindices2 = tf.stack((prepare_fd(fd_indices_custom, sd_dims), sd_indices), axis=2)\n\n# define the shape\nshape = tf.shape(updates)\n\nscatter1 = tf.scatter_nd(indices1, updates, shape)\nscatter2 = tf.scatter_nd(indices2, updates, shape)\n\nprint(scatter1.eval())\n\n# array([[12, 11, 13, 14],\n#        [21, 23, 22, 24],\n#        [31, 32, 34, 33]], dtype=int32)\n\nprint(scatter2.eval())\n\n# array([[21, 23, 22, 24],\n#        [31, 32, 34, 33],\n#        [12, 11, 13, 14]], dtype=int32)\n<\/code><\/pre>\n\n<p>May this example help. <\/p>\n"},{"AnswerId":"43953233","CreationDate":"2017-05-13T12:36:05.643","ParentId":null,"OwnerUserId":"1035460","Title":null,"Body":"<p>The following swaps the elements of each row of each row using scatter_nd<\/p>\n\n<pre><code>indices = tf.constant([[[0, 1], [0, 0], [0, 2], [0, 3]], \n                      [[1, 1], [1, 0], [1, 2], [1, 3]]])\nupdates = tf.constant([ [5, 6, 7, 8],\n                        [1, 2, 3, 4] ])\nshape = tf.constant([2, 4])\nscatter1 = tf.scatter_nd(indices, updates, shape)\nwith tf.Session() as sess:\n    print(sess.run(scatter1))\n<\/code><\/pre>\n\n<p>Giving an output of:<br>\n<code>[[6 5 7 8]\n [2 1 3 4]]\n<\/code><\/p>\n\n<p>The locations of the coordinate in <code>indices<\/code> define where the values are being taken from in <code>updates<\/code> and the actual cordinates define where the values will be placed in <code>scatter1<\/code>. <\/p>\n\n<p>This answer is a few months late but hopefully still helpful. <\/p>\n"}]}
{"QuestionId":42207928,"AnswerCount":3,"Tags":"<tensorflow><deep-learning><multilabel-classification>","CreationDate":"2017-02-13T15:40:57.537","AcceptedAnswerId":null,"OwnerUserId":6806378.0,"Title":"How to perform a multi label classification with tensorflow?","Body":"<p>I'm new to tensorflow and would like to know if there is any tutorial or example of a multilabel classification with multiple network outputs.<\/p>\n\n<p>I'm asking this because I have a collection of images, in which, each image can belong to several classes and my output needs to have a score of each class.<\/p>\n\n<p>I also do not know if the tensorflow follows some file pattern of the images and the classes, so if someone has some example it would facilitate a lot.<\/p>\n\n<p>Thank you.<\/p>\n","answers":[{"AnswerId":"42209441","CreationDate":"2017-02-13T16:58:19.950","ParentId":null,"OwnerUserId":"2476373","Title":null,"Body":"<p>The most naive (and reasonable) approach would be to train a classification network, and remove the softmax layer and replace it with a vector of sigmoids. This way you can have multiple units with an activation of 1.<\/p>\n\n<p>You can see <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim#Training\" rel=\"nofollow noreferrer\">on TF-slim<\/a> examples for classification networks. Under the path <code>datasets<\/code> you will find examples on how to prepare the <code>TFExample<\/code> \"file pattern\" for images and classes<\/p>\n"},{"AnswerId":"42213883","CreationDate":"2017-02-13T21:27:38.457","ParentId":null,"OwnerUserId":"223443","Title":null,"Body":"<p>You can also try transforming your problem from a multi-label to multi-class classification using a Label Powerset approach. Label Powerset  transformation treats every label combination attested in the training set as a different class and constructs one instance of a multi-class clasifier - and after prediction converts the assigned classes back to multi-label case. It is provided in <a href=\"http:\/\/scikit.ml\" rel=\"nofollow noreferrer\">scikit-multilearn<\/a> and scikit-compatibility wrapper over the tensorflow Estimator or via an <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/input_fn\/\" rel=\"nofollow noreferrer\">input_fn<\/a> or use <a href=\"http:\/\/terrytangyuan.github.io\/2016\/03\/14\/scikit-flow-intro\/\" rel=\"nofollow noreferrer\">skflow<\/a>. Then just plug it into an instance of <a href=\"http:\/\/scikit.ml\/api\/api\/skmultilearn.problem_transform.html#skmultilearn.problem_transform.lp.LabelPowerset\" rel=\"nofollow noreferrer\">LabelPowerset<\/a>. <\/p>\n\n<p>The code could go as follows:<\/p>\n\n<pre><code>from skmultilearn.problem_transform import LabelPowerset\nimport tensorflow.contrib.learn as skflow\n\n# assume data is loaded using\n# and is available in X_train\/X_test, y_train\/y_test\n\n# initialize LabelPowerset multi-label classifier\n# with tensor flow DNN base classifier\nclassifier = LabelPowerset(skflow.TensorFlowDNNClassifier(OPTIONS))\n\n# train\nclassifier.fit(X_train, y_train)\n\n# predict\npredictions = classifier.predict(X_test)\n<\/code><\/pre>\n"},{"AnswerId":"50229812","CreationDate":"2018-05-08T09:00:15.447","ParentId":null,"OwnerUserId":"7819691","Title":null,"Body":"<p>Most solutions refer to sigmoid loss, and sigmoid do solve multi-label classification well in my case by <code>tf.nn.sigmoid_cross_entropy_with_logits(labels,logits)<\/code> in tensorflow.<\/p>\n\n<p>However, when I handled class unbalance problem, where negative cases is much more than positive cases, I found my edited softsign loss worked much better than sigmoid. The adjust coefficient <code>gamma<\/code> is added to label to lower negative class's gradient by 3\/4. <\/p>\n\n<pre><code>def unbalance_softsign_loss(labels, logits):\n    gamma = 1.25 *labels - 0.25 \n    res = 1 - tf.log1p( gamma*logits\/(1+ tf.abs(logits)) )\n\n    return res \n<\/code><\/pre>\n\n<p>where labels is multi-hot encoding vectors like [0, 1, 0, 1, 0], logits ~ (-inf, inf)<\/p>\n"}]}
{"QuestionId":42208067,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><deep-learning>","CreationDate":"2017-02-13T15:48:19.397","AcceptedAnswerId":null,"OwnerUserId":1861627.0,"Title":"Convert tensor of unknown shape to a SparseTensor in tensorflow","Body":"<p>I have a tensor of partially unknown shape and a mask -- a tensor of same shape filled with  or  -- and I want to convert it into a SparseTensor, considering only the items corresponding to  in the mask. So, I think I have to go with something like:<\/p>\n\n<pre><\/pre>\n\n<p>At this point, I don't know how to proceed, since, both the way that I tried ended up in different errors, as follows. The first:<\/p>\n\n<pre><\/pre>\n\n<p>The second:<\/p>\n\n<pre><\/pre>\n\n<p>The third:<\/p>\n\n<pre><\/pre>\n\n<p>Any hint? Thanks!<\/p>\n","answers":[{"AnswerId":"46326873","CreationDate":"2017-09-20T16:04:34.923","ParentId":null,"OwnerUserId":"8642832","Title":null,"Body":"<p>In my case type casting the shape as <code>shape=tf.cast(tf.shape(dense), tf.int64)<\/code> in your first approach resolved the mentioned error.<\/p>\n"}]}
{"QuestionId":42209606,"AnswerCount":0,"Tags":"<visualization><reshape><conv-neural-network><torch>","CreationDate":"2017-02-13T17:07:18.233","AcceptedAnswerId":null,"OwnerUserId":6463339.0,"Title":"Torch: How to reshape a 4D tensor in 3D one","Body":"<p>I need to reshape a 4D tensor of dimension a x b x c x d into a list of 3D ones, a * b x c x d. I need them to visualize filters after each convolutional layer, using the following code:<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, due to the way the tensor are set up I receive this error:<\/p>\n\n<blockquote>\n  <p> packed must be a HxW or KxHxW or Kx3xHxW tensor, or a list of tensors<\/p>\n<\/blockquote>\n\n<p>Do you have some suggestion or a better way to do what I need to do?\nThanks in advance<\/p>\n","answers":[]}
{"QuestionId":42209854,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><summary>","CreationDate":"2017-02-13T17:21:39.757","AcceptedAnswerId":null,"OwnerUserId":1861627.0,"Title":"The node 'Merge\/MergeSummary' has inputs from different frames: what does it mean?","Body":"<p>trying to merge all my summaries, I have an error saying that the inputs of  comes from different frames. So, first of all: what is a <em>frame<\/em>? Could you please point me somewhere in the TF documentation about such stuff? -- of course, I googled a bit but could find almost nothing. How can I fix this issue? Below the code to reproduce the error. Thanks in advance.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"51509634","CreationDate":"2018-07-25T01:26:48.407","ParentId":null,"OwnerUserId":"6780025","Title":null,"Body":"<p>The error message was indeed not user friendly. It has been updated to<\/p>\n\n<pre><code>ValueError: Cannot use 'map\/while\/summaries\/histogram' as input to 'Merge\/MergeSummary' because 'map\/while\/summaries\/histogram' is in a while loop. See info log for more details.\n<\/code><\/pre>\n\n<p>As the new message says, the problem is that you cannot produce summaries from inside of the while loop. The <code>frame<\/code> that the original message referred to is the \"execution frame\" of the while loop - all the state for each iteration of the while loop is kept in a <code>frame<\/code>.<\/p>\n\n<p>In this case, the <code>while_loop<\/code> is created by <code>tf.map_fn<\/code> and the summary inside it is <code>tf.summary.histogram('histogram', activations)<\/code>.<\/p>\n\n<p>There are a couple of ways to deal with this. You can take the summary out of the <code>get_weights<\/code>, have the <code>get_weights<\/code> return activations as well, create the summary using the newly returned activations from <code>tf.map_fn<\/code> call.<\/p>\n\n<p>Another approach, if NUM_QUERIES is constant and small, can be to statically unroll the loop instead of using <code>tf.map_fn<\/code>. Here is the code to do this:<\/p>\n\n<pre><code># TOP PART OF THE CODE IS THE SAME\n\nstates = tf.placeholder(tf.float32, shape=[BATCH, None, SIZE])  # unknown length\nqueries = tf.placeholder(tf.float32, shape=[NUM_QUERIES, BATCH, ATT_SIZE])\nattention = Attention(states, ATT_SIZE)\nfunc = lambda x: attention.get_weights(x, \"Softmax\")\n\n# NEW CODE BEGIN\nsplit_queries = tf.split(queries, NUM_QUERIES)\nweights = []\nfor query in split_queries:\n    weights.append(func(query))\n# NEW CODE END\n\nfor var in tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES):\n    name = var.name.replace(':', '_')\n    tf.summary.histogram(name, var)\nsummary_op = tf.summary.merge_all()\n\nstates_np = np.random.rand(BATCH, LENGTH, SIZE)\nqueries_np = np.random.rand(NUM_QUERIES, BATCH, ATT_SIZE)\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    # NEW CODE BEGIN\n    results = sess.run(weights + [summary_op], {states: states_np, queries: queries_np})\n    weights_np, summary_str = results[:-1], results[-1]\n    # NEW CODE END\n    print weights_np\n<\/code><\/pre>\n"}]}
{"QuestionId":42211214,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-02-13T18:41:10.057","AcceptedAnswerId":null,"OwnerUserId":7559305.0,"Title":"Tensorflow jobs causing high disk load on the Root Filesystem","Body":"<p>When running training using Tensorflow, the root filesytem, which is on two SSDs, is seeing extremely high disk utilization, with a wait as high as 39min. Trying to figure out what is causing it.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42211473,"AnswerCount":1,"Tags":"<python><matrix><tensorflow><matrix-multiplication>","CreationDate":"2017-02-13T18:56:31.790","AcceptedAnswerId":null,"OwnerUserId":3245539.0,"Title":"multiplication of numerous matrices with vectors and summation of the resulting vector in tensor flow","Body":"<p>I am trying to multiply a set of vectors with corresponding matrices and would like to sum the resulting vectors at the end. As a numpy example let's assume we have 20 vectors and matrices of sizes 10x1 and and 150x1 respectively:<\/p>\n\n<pre><\/pre>\n\n<p>So far with tensorflow 0.10 I got:<\/p>\n\n<pre><\/pre>\n\n<p>But this symbolic multiplication gives me the error \"ValueError: Shape (?, 150, 10) must have rank 2\".<\/p>\n\n<p>Does anyone know why I am getting such an error message? How can I get tf_allMuls correctly?<\/p>\n","answers":[{"AnswerId":"42223643","CreationDate":"2017-02-14T10:26:32.903","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>From the documentation on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/math_ops\/matrix_math_functions#matmul\" rel=\"nofollow noreferrer\">tf.matmul<\/a>:<\/p>\n\n<blockquote>\n  <p>The inputs must be matrices (or tensors of rank > 2, representing\n  batches of matrices), with matching inner dimensions, possibly after\n  transposition.<\/p>\n<\/blockquote>\n\n<p>Considering that you are using <code>None<\/code> as the first argument to your placeholders, the second option is relevant for you, i.e. \"batches of matrices\". But your <code>tf_b<\/code> is a batch of vectors, not matrices, so the ranks of the two matrices are not the same, which is why you get the error. You should use instead:<\/p>\n\n<pre><code>tf_b = tf.placeholder(\"float\", [None, 10, 1])\ntf_A = tf.placeholder(\"float\", [None, 150, 10])\ntf_allMuls = tf.matmul(tf_A, tf_b)\n<\/code><\/pre>\n\n<p>It seems thus that <code>matmul<\/code> is not able to broadcast (may be check this <a href=\"https:\/\/stackoverflow.com\/questions\/38051143\/no-broadcasting-for-tf-matmul-in-tensorflow\/38056381\">post<\/a>) and I agree that the error message you get is a bit misleading.<\/p>\n\n<p>Here is a simple example:<\/p>\n\n<pre><code>tf_b = tf.placeholder(\"float\", [None, 3, 1])\ntf_A = tf.placeholder(\"float\", [None, 3, 3])\ntf_allMuls = tf.matmul(tf_A, tf_b)\n\nwith tf.Session() as sess:\n    b = np.array([1, 2, 3])[np.newaxis, :, np.newaxis]\n    A = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])[np.newaxis, ...]\n    muls = sess.run(tf_allMuls, feed_dict={tf_b: b, tf_A: A})\n    print muls\n<\/code><\/pre>\n\n<p>which prints<\/p>\n\n<pre><code>[[[ 14.]\n  [ 32.]\n  [ 50.]]]\n<\/code><\/pre>\n\n<p>Also note that the order of the arguments for <code>tf.matmul<\/code> matters, much like you are used to from actual matrix multiplication. So while this<\/p>\n\n<pre><code>tf_b = tf.placeholder(\"float\", [None, 1, 3])\ntf_A = tf.placeholder(\"float\", [None, 3, 3])\ntf_allMuls = tf.matmul(tf_A, tf_b)\n<\/code><\/pre>\n\n<p>NOT work, the following would (of course, it is not computing the same thing, but it does not raise an error):<\/p>\n\n<pre><code>tf_b = tf.placeholder(\"float\", [None, 1, 3])\ntf_A = tf.placeholder(\"float\", [None, 3, 3])\ntf_allMuls = tf.matmul(tf_b, tf_A)\n<\/code><\/pre>\n"}]}
{"QuestionId":42211485,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-platform><google-cloud-ml>","CreationDate":"2017-02-13T18:57:39.707","AcceptedAnswerId":null,"OwnerUserId":54858.0,"Title":"How to collect performance metrics from running google cloud ml training instances?","Body":"<p>I'm running a model on google cloud ml training, and it's taking about 10 hours with some naive guesses at the shapes of the machine. I'd like to optimize it a bit to cut down on running time and overall cost.<\/p>\n\n<p>What's the best way to determine if I'm using the resources effectively? I'd like cpu measurements, memory pressure, and GPU usage (whenever they are available). I suspect I'd need to either 1) log these or 2) install a monitoring agent like stack driver, and assume things like nvidia-smi are locatable, but I'm curious if any one has tried.<\/p>\n","answers":[{"AnswerId":"42781196","CreationDate":"2017-03-14T08:37:29.823","ParentId":null,"OwnerUserId":"40999","Title":null,"Body":"<p>This feature is now built-in into the product -- the CPU and RAM usage metrics (for now) are published as a Stackdriver metrics.<\/p>\n\n<p>A view of the metrics is also displayed in the console on the job detail page.<\/p>\n\n<p>Hope that helps.<\/p>\n"}]}
{"QuestionId":42211619,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-02-13T19:05:00.563","AcceptedAnswerId":42212349.0,"OwnerUserId":826983.0,"Title":"How to set weights for Convolution2D?","Body":"<p>I would like to set the weights of a  layer:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>but I am not sure what's expected here. I've tried several thing but most of the time I get<\/p>\n\n<pre><\/pre>\n\n<p>Not sure what this exactly means. <\/p>\n","answers":[{"AnswerId":"42212349","CreationDate":"2017-02-13T19:49:03.770","ParentId":null,"OwnerUserId":"4955956","Title":null,"Body":"<p>You should pass a numpy array to your convolutional layer through the set_weights method.<\/p>\n\n<p>Remember that the weights of a convolutional layer are not only the weights of each individual filter, but also the <strong>bias<\/strong>. So if you want to set your weights you need to add an extra dimension.<\/p>\n\n<p>For example, if you want to set a 1x3x3 filter with all weights zero except for the central element, you should make it:<\/p>\n\n<pre><code>w = np.asarray([ \n    [[[\n    [0,0,0],\n    [0,2,0],\n    [0,0,0]\n    ]]]\n    ])\n<\/code><\/pre>\n\n<p>And then set it.<\/p>\n\n<p>For a code you could run:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n# -*- coding: utf-8 -*-\nfrom __future__ import print_function\nimport numpy as np\nnp.random.seed(1234)\nfrom keras.layers import Input\nfrom keras.layers.convolutional import Convolution2D\nfrom keras.models import Model\nprint(\"Building Model...\")\ninp = Input(shape=(1,None,None))\noutput   = Convolution2D(1, 3, 3, border_mode='same', init='normal',bias=False)(inp)\nmodel_network = Model(input=inp, output=output)\nprint(\"Weights before change:\")\nprint (model_network.layers[1].get_weights())\nw = np.asarray([ \n    [[[\n    [0,0,0],\n    [0,2,0],\n    [0,0,0]\n    ]]]\n    ])\ninput_mat = np.asarray([ \n    [[\n    [1.,2.,3.],\n    [4.,5.,6.],\n    [7.,8.,9.]\n    ]]\n    ])\nmodel_network.layers[1].set_weights(w)\nprint(\"Weights after change:\")\nprint(model_network.layers[1].get_weights())\nprint(\"Input:\")\nprint(input_mat)\nprint(\"Output:\")\nprint(model_network.predict(input_mat))\n<\/code><\/pre>\n\n<p>Try changing the central element in the convolutional fillter (2 in the example).<\/p>\n\n<p>What the code does:<\/p>\n\n<p>At first build a model.<\/p>\n\n<pre><code>inp = Input(shape=(1,None,None))\noutput   = Convolution2D(1, 3, 3, border_mode='same', init='normal',bias=False)(inp)\nmodel_network = Model(input=inp, output=output)\n<\/code><\/pre>\n\n<p>Print your original weights (initialized with normal distribution, init='normal' )<\/p>\n\n<pre><code>print (model_network.layers[1].get_weights())\n<\/code><\/pre>\n\n<p>Create your desired weight tensor w and some input input_mat<\/p>\n\n<pre><code>w = np.asarray([ \n    [[[\n    [0,0,0],\n    [0,2,0],\n    [0,0,0]\n    ]]]\n    ])\ninput_mat = np.asarray([ \n    [[\n    [1.,2.,3.],\n    [4.,5.,6.],\n    [7.,8.,9.]\n    ]]\n    ])\n<\/code><\/pre>\n\n<p>set your weights and print them<\/p>\n\n<pre><code>model_network.layers[1].set_weights(w)\nprint(\"Weights after change:\")\nprint(model_network.layers[1].get_weights())\n<\/code><\/pre>\n\n<p>Finally, use it to generate output with predict (predict automatically compiles your model)<\/p>\n\n<pre><code>print(model_network.predict(input_mat))\n<\/code><\/pre>\n\n<p>Example Output:<\/p>\n\n<pre><code>Using Theano backend.\nBuilding Model...\nWeights before change:\n[array([[[[ 0.02357176, -0.05954878,  0.07163535],\n         [-0.01563259, -0.03602944,  0.04435815],\n         [ 0.04297942, -0.03182618,  0.00078482]]]], dtype=float32)]\nWeights after change:\n[array([[[[ 0.,  0.,  0.],\n         [ 0.,  2.,  0.],\n         [ 0.,  0.,  0.]]]], dtype=float32)]\nInput:\n[[[[ 1.  2.  3.]\n   [ 4.  5.  6.]\n   [ 7.  8.  9.]]]]\nOutput:\n[[[[  2.   4.   6.]\n   [  8.  10.  12.]\n   [ 14.  16.  18.]]]]\n<\/code><\/pre>\n"}]}
{"QuestionId":42211833,"AnswerCount":1,"Tags":"<python><tensorflow><classification><deep-learning><prediction>","CreationDate":"2017-02-13T19:17:47.957","AcceptedAnswerId":42212638.0,"OwnerUserId":6039807.0,"Title":"TensorFlow - How to predict with trained model on a different test dataset?","Body":"<p>I am new using TensorFlow and I don't know how to classify a picture with a trained model. I have already built a train, validation and test dataset for my training and all works but I want to predict on second test dataset (called test2). I am classifying pictures of digits.<\/p>\n\n<p>I have tried this but it doesn't work :<\/p>\n\n<pre><\/pre>\n\n<p>So the first time, I trained the model and test, then I want to predict on the other test set:<\/p>\n\n<pre><\/pre>\n\n<p>Functions , and  give me datasets with shape : (nb_pictures, 28, 28, 1) and labels with shape : (nb_pictures, 10).<\/p>\n\n<p>Thank you very much for any help<\/p>\n","answers":[{"AnswerId":"42212638","CreationDate":"2017-02-13T20:07:41.893","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You have everything that you need there. If you just want to predict you can extract the  function:<\/p>\n\n<pre><code> with tf.Session(graph=graph) as session:\n      ckpt = tf.train.get_checkpoint_state('.\/model\/')\n      saver.restore(session, ckpt.model_checkpoint_path)\n      feed_dict = {tf_train_dataset : batch_data}\n      predictions = session.run([test_prediction], feed_dict)\n<\/code><\/pre>\n"}]}
{"QuestionId":42212206,"AnswerCount":1,"Tags":"<theano><keras><convolution><assertion><subsampling>","CreationDate":"2017-02-13T19:40:04.633","AcceptedAnswerId":null,"OwnerUserId":6421501.0,"Title":"Keras Convolution3D subsample error","Body":"<p>I was trying to build a 3D convolutional layer using keras. It works fine, but when I added a subsample parameter it crashed. The code:<\/p>\n\n<pre><\/pre>\n\n<p>the error is:<\/p>\n\n<pre><\/pre>\n\n<p>I am using theano 0.8.2.<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42297579","CreationDate":"2017-02-17T12:06:47.047","ParentId":null,"OwnerUserId":"1150636","Title":null,"Body":"<p>You cannot use the <code>subsample<\/code> parameter with <code>border_mode='same'<\/code>. Use <code>'valid'<\/code> or <code>'full'<\/code><\/p>\n\n<p>Check out the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/backend\/theano_backend.py#L1685\" rel=\"nofollow noreferrer\">line of code where the assertion error happens<\/a><\/p>\n"}]}
{"QuestionId":42213270,"AnswerCount":2,"Tags":"<python><windows><numpy><tensorflow><package>","CreationDate":"2017-02-13T20:48:49.533","AcceptedAnswerId":45978311.0,"OwnerUserId":2707197.0,"Title":"Why can I not import TensorFlow package in Python Idle?","Body":"<p>I installed Tensorflow on Windows, using pip (Python 3.5)<\/p>\n\n<pre><\/pre>\n\n<p>Afterwards, I went to my command line and succesfully ran:<\/p>\n\n<pre><\/pre>\n\n<p>However, then I opened up Python IDLE (3.5) and tried to import tensorflow, with no luck:<\/p>\n\n<pre><\/pre>\n\n<p>Same story with numpy. My Numpy and tensorflow packages are installed here:<\/p>\n\n<pre><\/pre>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"52579839","CreationDate":"2018-09-30T16:24:59.857","ParentId":null,"OwnerUserId":"10410840","Title":null,"Body":"<p>I also faced the same problem in Python version 3.6. Finally, I managed to solve it by accessing Windows Command Prompt in Admin mode and type \nto enter to python IDLE<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>python<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>and successfully import tensorflow<\/p>\n\n<blockquote>\n  <blockquote>\n    <p>import tensorflow as tf<\/p>\n  <\/blockquote>\n<\/blockquote>\n\n<p>and checked the version with <\/p>\n\n<blockquote>\n  <blockquote>\n    <p>print(tf.<strong>version<\/strong>)<\/p>\n  <\/blockquote>\n<\/blockquote>\n"},{"AnswerId":"45978311","CreationDate":"2017-08-31T09:44:57.180","ParentId":null,"OwnerUserId":"4982729","Title":null,"Body":"<p>I think you got your answer. As asked here is how you can check your python version.<br>\n<code>import sys\nprint (sys.version)<\/code><\/p>\n"}]}
{"QuestionId":42213637,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-13T21:11:40.263","AcceptedAnswerId":null,"OwnerUserId":350080.0,"Title":"Inject function into tensorflow training loop?","Body":"<p>I have a training loop that I would like inject a function, forexample:<\/p>\n\n<pre><\/pre>\n\n<p>Here is my model<\/p>\n\n<pre><\/pre>\n\n<p>I tried adding in a function called tester:<\/p>\n\n<pre><\/pre>\n\n<p>And just before # Run training, I put<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42213775,"AnswerCount":0,"Tags":"<python><docker><flask><tensorflow><flask-restful>","CreationDate":"2017-02-13T21:20:39.713","AcceptedAnswerId":null,"OwnerUserId":654203.0,"Title":"Docker container with Flask server stops responding after around 20 requests","Body":"<p>I have a docker container that takes in URL to an image, downloads the image to memory using , does some processing to it using Tensorflow, and returns back with JSON data. <\/p>\n\n<p>The code is structured as follows:<\/p>\n\n<p><strong>myapp.py<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>my_tensorflow.py<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The container works fine for about 20 requests. After about 20 requests it stops responding and just responds with \"internal server error\". I can't seem to troubleshoot this issue. I put in print statements in my code to see where it becomes unresponsive or errors out but none of the print statements show up in  <\/p>\n\n<p> doesn't show anything interesting. Below is what it shows:<\/p>\n\n<pre><\/pre>\n\n<p>When the container becomes unresponsive (responds with ) the  returns this:<\/p>\n\n<pre><\/pre>\n\n<p>Everything starts working again once I restart the container with .<\/p>\n\n<p>I'm not sure what could be going wrong and how I can troubleshoot the issue?<\/p>\n","answers":[]}
{"QuestionId":42214228,"AnswerCount":2,"Tags":"<python><numpy><matrix><tensorflow>","CreationDate":"2017-02-13T21:53:10.907","AcceptedAnswerId":42214263.0,"OwnerUserId":766551.0,"Title":"Adding matrices with different dimensions","Body":"<p>In numpy and tensorflow it's possible to add matrices (or tensors) of different dimensionality if the shape of smaller matrix is a suffix of bigger matrix. This is an example:<\/p>\n\n<pre><\/pre>\n\n<p>For these two matrices operation  is a shortcut for:<\/p>\n\n<pre><\/pre>\n\n<p>In my case however I have matrices of shape  and  and likewise I would like to perform the  operation using similar logic:<\/p>\n\n<pre><\/pre>\n\n<p>In this case however  operation fails as neither numpy nor tensorflow understands what I want to do. Is there any way I can perform this operation efficiently (without writing python loop myself)?<\/p>\n\n<p>In so far I have figured that I could introduce a temporary matrix  of shape  using einsum like this:<\/p>\n\n<pre><\/pre>\n\n<p>but this creates an explicit three-dimensional matrix (or tensor) and if possible I would prefer to avoid that.<\/p>\n","answers":[{"AnswerId":"42214263","CreationDate":"2017-02-13T21:55:40.067","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>In NumPy, you could extend <code>y<\/code> to <code>3D<\/code> and then add -<\/p>\n\n<pre><code>x + y[:,None,:]\n<\/code><\/pre>\n\n<p>Haven't dealt with <code>tensorflow<\/code> really, but looking into its docs, it seems, we could use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/array_ops\/shapes_and_shaping#expand_dims\" rel=\"nofollow noreferrer\"><code>tf.expand_dims<\/code><\/a> -<\/p>\n\n<pre><code>x + tf.expand_dims(y, 1)\n<\/code><\/pre>\n\n<p>The extended version would still be a view into <code>y<\/code> and as such won't occupy anymore of the memory, as tested below -<\/p>\n\n<pre><code>In [512]: np.may_share_memory(y, y[:,None,:])\nOut[512]: True\n<\/code><\/pre>\n"},{"AnswerId":"42220994","CreationDate":"2017-02-14T08:18:31.470","ParentId":null,"OwnerUserId":"766551","Title":null,"Body":"<p>As correctly pointed out in the accepted answer the solution is to expand dimensions using available construct.<\/p>\n\n<p>The point is to understand how numpy is doing the broadcasting of matrices in case of adding matrices if their dimensions don't match. The rule is that two matrices must have exactly the same dimensions with an exception that some dimensions in either of matrices can be replaced by 1.<\/p>\n\n<p>E.g.<\/p>\n\n<pre><code>A      (4d array):  8 x 1 x 6 x 1\nB      (3d array):      7 x 1 x 5\nResult (4d array):  8 x 7 x 6 x 5\n<\/code><\/pre>\n\n<p>This example and detailed explanation can be found in scipy docs\n<a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.10.0\/user\/basics.broadcasting.html\" rel=\"nofollow noreferrer\">https:\/\/docs.scipy.org\/doc\/numpy-1.10.0\/user\/basics.broadcasting.html<\/a> <\/p>\n"}]}
{"QuestionId":42214934,"AnswerCount":3,"Tags":"<tensorflow><neural-network><deep-learning><caffe><conv-neural-network>","CreationDate":"2017-02-13T22:45:43.287","AcceptedAnswerId":null,"OwnerUserId":6258504.0,"Title":"How to define a layer only for training phase in TensorFlow?","Body":"<p>I wanted to know if it's possible to define a layer (convolution, element-wise summation, etc.) only for the training phase in TensorFlow. <\/p>\n\n<p>For example, I want to have an element-wise summation layer in my network only for the training phase and I want to ignore this layer in the test phase. <\/p>\n\n<p>This is easily doable in Caffe, I wanted to know if it's possible to do so in TensorFlow as well. <\/p>\n","answers":[{"AnswerId":"42241359","CreationDate":"2017-02-15T05:28:25.123","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>I think you can use a boolean placeholder with tf.cond().\nJust like this:<\/p>\n\n<pre><code>train_phase = tf.placeholder(tf.bool, [])\nx = tf.constant(2)\ndef f1(): return tf.add(x, 1)\ndef f2(): return tf.identity(x)\nr = tf.cond(train_phase, f1, f2)\nsess.run(r, feed_dict={train_phase: True})  # training phase, r = tf.add(x, 1) = x + 1\nsess.run(r, feed_dict={train_phase: False})  # testing phase, r = tf.identity(x) = x\n<\/code><\/pre>\n"},{"AnswerId":"42221671","CreationDate":"2017-02-14T08:56:42.917","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>You might want to do this with the \"tf.cond\" control_flow operation.  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#cond\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/control_flow_operations#cond<\/a><\/p>\n"},{"AnswerId":"42216859","CreationDate":"2017-02-14T02:29:38.393","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>I think you can do this by <code>if<\/code>     <\/p>\n\n<pre><code>Train = False\n\nx = tf.constant(5.)\ny = x + 1\nif Train:\n    y = y + 2\ny = y + 3\n\nwith tf.Session() as sess:\n    res = sess.run(y) # 11 if Train else 9\n<\/code><\/pre>\n"}]}
{"QuestionId":42215574,"AnswerCount":1,"Tags":"<graph><tensorflow><gradient-descent>","CreationDate":"2017-02-13T23:51:02.937","AcceptedAnswerId":null,"OwnerUserId":568986.0,"Title":"Exclude specific tensors being updated by optimizer in TensorFlow","Body":"<p>I have two graphs, which I suppose to train them independently, which means I have two different optimizers, but at the same time one of them is using the tensor values of the other graph. As a result, I need to be able to stop specific tensors being updated while training one of the graphs. I have assigned two different namescopes two my tensors and using this code to control updates over tensors for different optimizers:<\/p>\n\n<pre><\/pre>\n\n<p>the vars variable is being used like below, in the training method of mnist object:<\/p>\n\n<pre><\/pre>\n\n<p>I'm using the <strong>var_list<\/strong> attribute of the <strong>optimizer class<\/strong> in order to control vars being updated by the optimizer. <\/p>\n\n<p>Right now I'm confused whether I have done what I supposed to do appropriately, and even if there is anyway to check if any optimizer would only update partial of a graph?<\/p>\n\n<p>I would appreciate if anyone can help me with this issue.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42223250","CreationDate":"2017-02-14T10:09:17.937","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>I have had a similar problem and used the same approach as you, i.e. via the <code>var_list<\/code> argument of the optimizer. I then checked whether the variables not intended for training stayed the same using:<\/p>\n\n<pre><code>the_var_np = sess.run(tf.get_default_graph().get_tensor_by_name('the_var:0'))\nassert np.equal(the_var_np, pretrained_weights['the_var']).all()\n<\/code><\/pre>\n\n<p><code>pretrained_weights<\/code> is a dictionary returned by <code>np.load('some_file.npz')<\/code> which I used to store the pre-trained weights to disk.<\/p>\n\n<p>Just in case you need that as well, here is how you can override a tensor with a given value:<\/p>\n\n<pre><code>value = pretrained_weights['the_var']\nvariable = tf.get_default_graph().get_tensor_by_name('the_var:0')\nsess.run(tf.assign(variable, value))\n<\/code><\/pre>\n"}]}
{"QuestionId":42216208,"AnswerCount":1,"Tags":"<python><c++><tensorflow>","CreationDate":"2017-02-14T01:06:27.627","AcceptedAnswerId":43531616.0,"OwnerUserId":3627572.0,"Title":"Should TensorFlow users prefer SavedModel over Checkpoint or GraphDef?","Body":"<p>From <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/python\/saved_model\/\" rel=\"noreferrer\">SavedModel Docs<\/a>,<\/p>\n\n<blockquote>\n  <p>SavedModel, the universal serialization format for TensorFlow models.<\/p>\n<\/blockquote>\n\n<p>and<\/p>\n\n<blockquote>\n  <p>SavedModel wraps a TensorFlow Saver. The Saver is primarily used to generate the variable checkpoints.<\/p>\n<\/blockquote>\n\n<p>From my understanding,  is must if someone wants use TensorFlow Serving. However, I can deploy Tensorflow Model to service server without : Freeze graph and export it as , and load graph into Session using  and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/session#create\" rel=\"noreferrer\">Create<\/a> in C++ or <a href=\"https:\/\/godoc.org\/github.com\/tensorflow\/tensorflow\/tensorflow\/go#Graph.Import\" rel=\"noreferrer\">Import<\/a> in Go.<\/p>\n\n<p>What is the purpose of SavedModel? Should users prefer SavedModel over Checkpoint or GraphDef to aggregate more data related to the model?<\/p>\n","answers":[{"AnswerId":"43531616","CreationDate":"2017-04-20T23:41:14.613","ParentId":null,"OwnerUserId":"1661471","Title":null,"Body":"<p>A checkpoint contains the value of (some of the) variables in a TensorFlow model. It is created by a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"noreferrer\"><code>Saver<\/code><\/a>, which is either given specific <code>Variable<\/code>s to save, or by default saves all (non-local) Variables.<\/p>\n\n<p>To use a checkpoint, you need to have a compatible TensorFlow <code>Graph<\/code>, whose <code>Variable<\/code>s have the same names as the <code>Variable<\/code>s in the checkpoint. (If you don't have a compatible <code>Graph<\/code>, you can still load the values stored in a checkpoint into selected <code>Variable<\/code>s using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/framework\/init_from_checkpoint\" rel=\"noreferrer\"><code>init_from_checkpoint<\/code><\/a> utilities in contrib.)<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/saved_model\" rel=\"noreferrer\"><code>SavedModel<\/code><\/a> is much more comprehensive: It contains a set of <code>Graph<\/code>s (<a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/meta_graph\" rel=\"noreferrer\"><code>MetaGraph<\/code><\/a>s, in fact, saving collections and such), as well as a checkpoint which is supposed to be compatible with these <code>Graph<\/code>s, and any asset files that are needed to run the model (e.g. Vocabulary files). For each <code>MetaGraph<\/code> it contains, it also stores a set of signatures. Signatures define (named) input and output tensors. <\/p>\n\n<p>This means that given only a SavedModel, you can write tools (such as <code>tensorflow\/serving<\/code>, or the new <code>saved_model<\/code> command line utility that will appear in <code>tools\/<\/code> shortly) that interpret or execute the graphs inside. All you have to provide is the data. <\/p>\n\n<p>If in doubt, I would always err on the side of writing a <code>SavedModel<\/code>, not just a checkpoint. Not only does this allow you to use tensorflow\/serving (and other neat utilities that will grow in number), it makes sure that you have all the information necessary to run the model. Nothing is more frustrating than a checkpoint you cannot use any more because you modified your model and now it is incompatible with checkpoint files and all you want to do is run some predictions through it for comparison.<\/p>\n"}]}
{"QuestionId":42217059,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-02-14T02:52:26.753","AcceptedAnswerId":42218732.0,"OwnerUserId":7560632.0,"Title":"tensorflow:AttributeError: 'module' object has no attribute 'mul'","Body":"<p>I have used tensorflow for ONE day,but there comes some troubles,when I import tensorflow, there would be AttributeError: 'module' object has no attribute 'XXXXXX'<\/p>\n\n<h2>Environment<\/h2>\n\n<p>I use ubuntu14.04, python2.7, CUDA toolkit 8.0 and CuDNN v5.\nAnd versions of my six and protobuf are:\nName: six\nVersion: 1.10.0\nLocation: \/usr\/local\/lib\/python2.7\/dist-packages\nRequires: \nName: protobuf\nVersion: 3.2.0\nLocation: \/usr\/local\/lib\/python2.7\/dist-packages\nRequires: six, setuptools<\/p>\n\n<p>here is my test code:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I get this output:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LCHw1.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LCHw1.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Is there any problem with the tensorflow installation? or any other problems?<\/p>\n","answers":[{"AnswerId":"42218732","CreationDate":"2017-02-14T05:45:20.440","ParentId":null,"OwnerUserId":"6408452","Title":null,"Body":"<p>According to the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md\" rel=\"noreferrer\">tensorflow 1.0.0 release notes<\/a>, <\/p>\n\n<blockquote>\n  <p><code>tf.mul<\/code>, <code>tf.sub<\/code> and <code>tf.neg<\/code> are deprecated in favor of <code>tf.multiply<\/code>, <code>tf.subtract<\/code> and <code>tf.negative<\/code>.<\/p>\n<\/blockquote>\n\n<p>You'll need to replace <code>tf.mul<\/code> with <code>tf.multiply<\/code>.<\/p>\n"},{"AnswerId":"43362710","CreationDate":"2017-04-12T07:10:22.667","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>This operation was previously available in 0.x versions. With the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md#breaking-changes-to-the-api\" rel=\"nofollow noreferrer\">release of TF 1.0 they introduced breaking changes to the API<\/a>. In addition to <\/p>\n\n<blockquote>\n  <p><code>tf.mul<\/code>, <code>tf.sub<\/code> and <code>tf.neg<\/code> are deprecated in favor of <code>tf.multiply<\/code>,\n  <code>tf.subtract<\/code> and <code>tf.negative<\/code><\/p>\n<\/blockquote>\n\n<p>many other functions were renamed and changed with the following justification:<\/p>\n\n<blockquote>\n  <p>Several python API calls have been changed to resemble NumPy more\n  closely.<\/p>\n<\/blockquote>\n\n<p>So a lot of the scripts that you already found on the web or from the books will not work. Good thing is that majority of them can be fixed with their migration <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/compatibility\" rel=\"nofollow noreferrer\">script<\/a>. It can be run with <code>tf_upgrade.py --infile foo.py --outfile foo-upgraded.py<\/code>. It will not be able to solve everything (limitations are listed <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"nofollow noreferrer\">here<\/a>), but will save you a lot of work.<\/p>\n"},{"AnswerId":"51621482","CreationDate":"2018-07-31T20:20:55.983","ParentId":null,"OwnerUserId":"10162556","Title":null,"Body":"<p>In python-3 use <code>tf.multiply<\/code> instead of <code>tf.mul<\/code>.<\/p>\n"}]}
{"QuestionId":42217320,"AnswerCount":1,"Tags":"<python><variables><tensorflow><restore>","CreationDate":"2017-02-14T03:22:56.977","AcceptedAnswerId":42226102.0,"OwnerUserId":3935993.0,"Title":"Restore variables that are a subset of new model in Tensorflow?","Body":"<p>I am doing an example for boosting(4 layers DNN to 5 layers DNN) via Tensorflow. I am making it with save session and restore in TF because there is a brief paragraph in TF tute:\n'For example, you may have trained a neural net with 4 layers, and you now want to train a new model with 5 layers, restoring the parameters from the 4 layers of the previously trained model into the first 4 layers of the new model.', where tensorflow tute inspires in <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variables\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/variables\/<\/a>.<\/p>\n\n<p>However, I found that nobody has asked how to use 'restore' when the checkpoint saves parameters of 4 layers but we need to put that into 5 layers, raising a red flag.<\/p>\n\n<p>Making this in real code, I made<\/p>\n\n<pre><\/pre>\n\n<p>where variables inside(or called from) 'fcl1' - so that I have 'fcl1\/Variable' and 'fcl1\/Variable_1' for weight and bias - 'fcl2', 'fclf', and 'outputl' are stored by saver.save() in the script without 'boosting' layer. However, as we now have 'boosting' layer, saver.restore(sess, \"saved_models\/model_list.ckpt\") does not work as<\/p>\n\n<pre><\/pre>\n\n<p>I really hope to hear about this problem. Thank you.\nBelow code is the main part of the code I am in trouble.<\/p>\n\n<pre><\/pre>\n\n<p>For clarity, the checkpoint file has <\/p>\n\n<pre><\/pre>\n\n<p>As the original 4 layers model does not have 'boosting' layer.<\/p>\n","answers":[{"AnswerId":"42226102","CreationDate":"2017-02-14T12:23:38.650","ParentId":null,"OwnerUserId":"6717469","Title":null,"Body":"<p>It doesn't look right to read values for boosting from the checkpoint in this case and I think that's not what you want to do. Obviously you're getting error, since while restoring the variables you are first catching the list of all of the variables in your model and then you look for corresponding variables in your checkpoint, which doesn't have them. <\/p>\n\n<p>You can restore only part of your model by defining a subset of your model variables. For example you can do it using tf.slim library. Getting the list of variables in your models:<\/p>\n\n<pre><code>variables = slim.get_variables_to_restore()\n<\/code><\/pre>\n\n<p>Now variables is a list of tensors, but for each element you can access its name attribute. Using that you can specify that you only want to restore layers other than boosting, e.g.:<\/p>\n\n<pre><code>variables_to_restore = [v for v in variables if v.name.split('\/')[0]!='boosting'] \nmodel_path = 'your\/model\/path'\n\nsaver = tf.train.Saver(variables_to_restore)\n\nwith tf.Session() as sess:\n    saver.restore(sess, model_path)\n<\/code><\/pre>\n\n<p>This way you will have your 4 layers restored. Theoretically you could try to catch values of one of your variables from checkpoint by creating another server that will only have boosting in variables list and renaming the chosen variable from the checkpoint, but I really don't think it's what you need here.<\/p>\n\n<p>Since this is a custom layer for your model and you don't have this variable anywhere, just initialize it within your workflow instead of trying to import it. You can do for example by passing this argument while calling a function fully_connected:<\/p>\n\n<pre><code>weights_initializer = slim.variance_scaling_initializer()\n<\/code><\/pre>\n\n<p>You need to check details yourself though, since I'm not sure what your imports are and which function are you using here.<\/p>\n\n<p>Generally I'd advice you to take a look at slim library, which will make it easier for you to define a model and scopes for layers (instead of defining it by with you can rather pass a scope argument while calling a function). It would look something like that with slim:<\/p>\n\n<pre><code>boost = slim.fully_connected(input, number_of_outputs, activation_fn=None, scope='boosting', weights_initializer=slim.variance_scaling_initializer())\n<\/code><\/pre>\n"}]}
{"QuestionId":42217532,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2017-02-14T03:45:51.093","AcceptedAnswerId":42262984.0,"OwnerUserId":395857.0,"Title":"TensorFlow version 1.0.0-rc2 on Windows: \"OpKernel ('op: \"BestSplits\" device_type: \"CPU\"') for unknown op: BestSplits\" with test code","Body":"<p>I installed TensorFlow version 1.0.0-rc2 on Windows 7 SP1 x64 Ultimate (Python 3.5.2 |Anaconda custom (64-bit)) using:<\/p>\n\n<pre><\/pre>\n\n<p>When I try running the test script from <a href=\"https:\/\/web.archive.org\/web\/20170214034751\/https:\/\/www.tensorflow.org\/get_started\/os_setup#test_the_tensorflow_installation\" rel=\"noreferrer\">https:\/\/web.archive.org\/web\/20170214034751\/https:\/\/www.tensorflow.org\/get_started\/os_setup#test_the_tensorflow_installation<\/a> in Eclipse 4.5 or in the console:<\/p>\n\n<pre><\/pre>\n\n<p>I obtain some error message:<\/p>\n\n<pre><\/pre>\n\n<p>Why?<\/p>\n\n<p>I didn't have such issues with TensorFlow 0.12.1  (installed with ):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43169316","CreationDate":"2017-04-02T14:11:55.640","ParentId":null,"OwnerUserId":"6799699","Title":null,"Body":"<p>Seems the issue is fixed in the version 1.1.0rc0 and later.<\/p>\n\n<p>Find the latest version of tensorflow:<\/p>\n\n<pre><code>pip search --version tensorflow\n<\/code><\/pre>\n\n<p>Upgrade tensorflow:<\/p>\n\n<pre><code>pip install --upgrade tensorflow==1.1.0rc1\n<\/code><\/pre>\n"},{"AnswerId":"43249084","CreationDate":"2017-04-06T07:56:11.523","ParentId":null,"OwnerUserId":"6902409","Title":null,"Body":"<p>Refering to suggestions above, I think doing 2 steps is helpful:<\/p>\n\n<p>1st, upgrade tensorflow:<\/p>\n\n<pre><code>pip install --upgrade tensorflow==1.1.0rc1\n<\/code><\/pre>\n\n<p>then, error logs changes into warn logs:<\/p>\n\n<pre><code>W c:\\tf_jenkins\\home\\workspace\\release-win\\device\\cpu\\os\\windows\\tensorflow\\core\\platform\\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations.\n<\/code><\/pre>\n\n<p>2nd, you may be able to suppress the warning filter at level 2.<\/p>\n\n<pre><code>os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\n<\/code><\/pre>\n\n<p>I think it works well without filtering 'error' logs.<\/p>\n"},{"AnswerId":"42686552","CreationDate":"2017-03-09T04:10:59.337","ParentId":null,"OwnerUserId":"189218","Title":null,"Body":"<p>You may be able to suppress the warning filter at level 2. This worked for me with TensorFlow 1.0.1 in a virtualenv install. <\/p>\n\n<p>os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'<\/p>\n\n<p>Sry about the additional answer, but I'm not worthy of commenting.<\/p>\n"},{"AnswerId":"42262984","CreationDate":"2017-02-16T01:01:47.037","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>Installing <a href=\"http:\/\/ci.tensorflow.org\/view\/Nightly\/job\/nightly-win\/85\/\" rel=\"nofollow noreferrer\">today's nightly build<\/a> (CPU version):<\/p>\n\n<pre><code>pip install --upgrade http:\/\/ci.tensorflow.org\/view\/Nightly\/job\/nightly-win\/85\/DEVICE=cpu,OS=windows\/artifact\/cmake_build\/tf_python\/dist\/tensorflow-1.0.0rc2-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>fixed the issue (no more  <code>\u201cOpKernel ('op: \u201dBestSplits\u201c device_type: \u201dCPU\u201c') for unknown op: BestSplits\u201d<\/code> etc.).<\/p>\n\n<p>There are now some SSE warnings:<\/p>\n\n<pre><code>TensorFlow version: 1.0.0-rc2\nb'Hello, TensorFlow!'\n2017-02-15 19:56:22.688266: W c:\\tf_jenkins\\home\\workspace\\nightly-win\\device\\cpu\\os\\windows\\tensorflow\\core\\platform\\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations.\n2017-02-15 19:56:22.688266: W c:\\tf_jenkins\\home\\workspace\\nightly-win\\device\\cpu\\os\\windows\\tensorflow\\core\\platform\\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE2 instructions, but these are available on your machine and could speed up CPU computations.\n2017-02-15 19:56:22.689266: W c:\\tf_jenkins\\home\\workspace\\nightly-win\\device\\cpu\\os\\windows\\tensorflow\\core\\platform\\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE3 instructions, but these are available on your machine and could speed up CPU computations.\n2017-02-15 19:56:22.689266: W c:\\tf_jenkins\\home\\workspace\\nightly-win\\device\\cpu\\os\\windows\\tensorflow\\core\\platform\\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.\n2017-02-15 19:56:22.689266: W c:\\tf_jenkins\\home\\workspace\\nightly-win\\device\\cpu\\os\\windows\\tensorflow\\core\\platform\\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.\n2017-02-15 19:56:22.689266: W c:\\tf_jenkins\\home\\workspace\\nightly-win\\device\\cpu\\os\\windows\\tensorflow\\core\\platform\\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.\n<\/code><\/pre>\n\n<p>in which case you can try <a href=\"https:\/\/stackoverflow.com\/q\/41293077\/395857\">How to compile Tensorflow with SSE4.2 and AVX instructions?<\/a><\/p>\n\n<hr>\n\n<p><a href=\"https:\/\/web.archive.org\/web\/20170219153708\/https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md\" rel=\"nofollow noreferrer\">TensorFlow 1.0.0<\/a> was released a few days ago. However, it has the same issue. A more recent nightly build has <a href=\"https:\/\/web.archive.org\/web\/20170219153938\/https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7500#issuecomment-280576326\" rel=\"nofollow noreferrer\">different warnings<\/a>:<\/p>\n\n<pre><code>sess = tf.Session()\n2017-02-17 13:01:59.790943: W c:\\tf_jenkins\\home\\workspace\\nightly-win\\device\\cpu\\os\\windows\\tensorflow\\core\\platform\\cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE instructions, but these are available on your machine and could speed up CPU computations.\n<\/code><\/pre>\n\n<p>FYI:  <a href=\"https:\/\/github.com\/Thunderbottom\/Tensorflow-macOS-binary\" rel=\"nofollow noreferrer\">Tensorflow macOS binary, compiled with SSE4.1, SSE4.2 and AVX optimizations.<\/a><\/p>\n\n<hr>\n\n<p>To hide the warnings\/errors, you can <a href=\"https:\/\/stackoverflow.com\/q\/35911252\/395857\">use<\/a> <code>os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'<\/code>, e.g.:<\/p>\n\n<pre><code>import tensorflow as tf\nimport os\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'\nprint('TensorFlow version: {0}'.format(tf.__version__))\nhello = tf.constant('Hello, TensorFlow!')\nsess = tf.Session()\nprint(sess.run(hello))\n<\/code><\/pre>\n\n<p><code>TF_CPP_MIN_LOG_LEVEL<\/code>:<\/p>\n\n<ul>\n<li><code>0<\/code>: all logs shown (that's the default setting)<\/li>\n<li><code>1<\/code>: filter out <code>INFO<\/code> logs<\/li>\n<li><code>2<\/code>: additionally filter out <code>WARNING<\/code> logs<\/li>\n<li><code>3<\/code>: additionally filter out <code>ERROR<\/code> logs.<\/li>\n<\/ul>\n"}]}
{"QuestionId":42217777,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><conv-neural-network>","CreationDate":"2017-02-14T04:17:32.743","AcceptedAnswerId":null,"OwnerUserId":7521228.0,"Title":"Implementing dropout in my CNN makes Training accuracy drops","Body":"<p>Implementing a Convolutional neural network for image classification, I have good results on my training set accuracy (98%), but testing set accuracy remains low (around 82%). I am facing an overfitting issue. \nI tried to solve it by implementing a dropout function (keeping 50% of neurons), but it turns out that this solution makes my training accuracy drops to 80%! Does it mean that I should have more neurons in my CNN?\nWhat could I do to prevent this issue?<\/p>\n\n<p>I am using Tensorflow library and here is the piece of code for the training loop : (I am not sure I implemented it correctly)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42217952","CreationDate":"2017-02-14T04:36:52.700","ParentId":null,"OwnerUserId":"3260576","Title":null,"Body":"<p>If your test accuracy is also around the same mark as traning accuracy in case of dropout, then you can increase number of neurons. Also, 0.50 as initial dropout is pretty high. Start with 0.25(i.e. keep probability of 0.75 for dropout layer.).Also, I would recommend using some data augmentation like rotation, distorting brightness, swapping color channels depending upon nature of the data and using some regularization too. Also plot learning curves and check how test accuracy changes with training accuracy.<\/p>\n"}]}
{"QuestionId":42218771,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-14T05:48:56.483","AcceptedAnswerId":42360539.0,"OwnerUserId":1043993.0,"Title":"How do I load images from URLs into tensorflow?","Body":"<p>I'm currently running this:<\/p>\n\n<pre><\/pre>\n\n<p>and I get this error:<\/p>\n\n<blockquote>\n  <p>Unable to get element from the feed as bytes.<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42258541","CreationDate":"2017-02-15T19:37:53.167","ParentId":null,"OwnerUserId":"1067799","Title":null,"Body":"<p>Just use a method to load it into NumPy and then load that into TensorFlow. This is likely more reliable, more flexible.<\/p>\n\n<p><a href=\"http:\/\/www.pyimagesearch.com\/2015\/03\/02\/convert-url-to-image-with-python-and-opencv\/\" rel=\"nofollow noreferrer\">http:\/\/www.pyimagesearch.com\/2015\/03\/02\/convert-url-to-image-with-python-and-opencv\/<\/a><\/p>\n"},{"AnswerId":"42360539","CreationDate":"2017-02-21T06:52:08.570","ParentId":null,"OwnerUserId":"1043993","Title":null,"Body":"<p>The solution was quite simple... All I had to do was call the <code>read<\/code> method on response from urlopen. The following works like a charm:<\/p>\n\n<pre><code>image = urllib.urlopen(imgUrl)\npool3_features = sess.run(pool3,{'incept\/DecodeJpeg\/contents:0': image.read()})\n<\/code><\/pre>\n"}]}
{"QuestionId":42219140,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-14T06:16:05.780","AcceptedAnswerId":null,"OwnerUserId":7421488.0,"Title":"Initialize error in Tensorflow 0.12.1","Body":"<p>I run my code well under Tensorflow 0.10, but after I update the version to 0.12.1, all the variables throws a . And my code is unchanged except I use  instead of . I tried to add  under  but it didn't work. So I open this question because I don't know where the bug is.<\/p>\n\n<p>My code is as below. Omitted some details to be more readable.\nAnd the error is different every time. All  defined variables threw error.<\/p>\n\n<pre><\/pre>\n\n<p>Can anybody help? The program worked well in 0.10 but crashed after I updated to 0.12.1. I changed nothing but  to .<\/p>\n","answers":[]}
{"QuestionId":42219296,"AnswerCount":1,"Tags":"<python><csv><tensorflow><neural-network>","CreationDate":"2017-02-14T06:25:51.683","AcceptedAnswerId":42221515.0,"OwnerUserId":2707197.0,"Title":"Neural network with a CSV file in TensorFlow","Body":"<p>I am using two tutorials to figure out how to take a CVS file of format:<\/p>\n\n<pre><\/pre>\n\n<p>and train a neural network on it. What I do in the code below is read in the CVS file and group 100 lines at a time into batches: x_batch and y_batch. Next, i try to have the NN learn in batches. However, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I am wondering what i am doing wrong and what another approach might be. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42221515","CreationDate":"2017-02-14T08:48:48.787","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Your placeholder y specifies you input an array of unknown length, with arrays of length \"n_classes\" (which is 4). In your feed_dict you give the array batch_y, which is an array of length 99 (your batch_size) with numbers. <\/p>\n\n<p>What you want to do is change your batch_y variable to have one-hot vectors as input. Please let me know if this works!<\/p>\n"}]}
{"QuestionId":42219848,"AnswerCount":1,"Tags":"<python><tensorflow><gpu>","CreationDate":"2017-02-14T07:04:19.433","AcceptedAnswerId":42219996.0,"OwnerUserId":1625695.0,"Title":"Tensorflow automatically choose least loaded GPU","Body":"<p>We specify the GPU device to be used using:<\/p>\n\n<pre><\/pre>\n\n<p>gpu_id is a string variable where I manually set the GPU id to be used.\nI need to run several experiments, each on a different GPU. So, I manually change the value of gpu_id before running a code instance.<\/p>\n\n<p>Can I write some code that automatically detects the first unutilized GPU and set it to gpu_id ?<\/p>\n","answers":[{"AnswerId":"42219996","CreationDate":"2017-02-14T07:15:25.003","ParentId":null,"OwnerUserId":"499581","Title":null,"Body":"<p>There's already a function that allows you to know which GPU is being used for a tensor:<\/p>\n\n<pre><code># Creates a session with log_device_placement set to True.\nsess = tf.Session(config=tf.ConfigProto(log_device_placement=True))\n<\/code><\/pre>\n\n<p>Setting <code>log_device_placement<\/code> to <code>True<\/code> will return data similar to this:<\/p>\n\n<pre><code>Device mapping:\n\/job:localhost\/replica:0\/task:0\/gpu:0 -&gt; device: 0, name: Tesla K40c, pci bus\nid: 0000:05:00.0\nb: \/job:localhost\/replica:0\/task:0\/gpu:0\na: \/job:localhost\/replica:0\/task:0\/gpu:0\nMatMul: \/job:localhost\/replica:0\/task:0\/gpu:0\n[[ 22.  28.]\n [ 49.  64.]]\n<\/code><\/pre>\n\n<p>\u21b3 <a href=\"https:\/\/www.tensorflow.org\/how_tos\/using_gpu\/\" rel=\"nofollow noreferrer\">Using GPUs<\/a><\/p>\n"}]}
{"QuestionId":42221348,"AnswerCount":1,"Tags":"<tensorflow><neural-network><regression><conv-neural-network>","CreationDate":"2017-02-14T08:40:01.517","AcceptedAnswerId":null,"OwnerUserId":2513126.0,"Title":"Loss suddenly increases with Adam Optimizer in Tensorflow","Body":"<p>I am using a CNN for a regression task. I use Tensorflow and the optimizer is Adam. The network seems to converge perfectly fine till one point where the loss suddenly increases along with the validation error. Here are the loss plots of the labels and the weights separated (Optimizer is run on the sum of them)\n<img src=\"https:\/\/i.imgur.com\/KTGvMVV.jpg\" alt=\"label loss\" title=\"loss\">\n<img src=\"https:\/\/imgur.com\/auF2ef5.jpg\" alt=\"weight loss\"><\/p>\n\n<p>I use l2 loss for weight regularization and also for the labels. I apply some randomness on the training data. I am currently trying RSMProp to see if the behavior changes but it takes at least 8h to reproduce the error.<\/p>\n\n<p>I would like to understand how this can happen. Hope you can help me.<\/p>\n","answers":[{"AnswerId":"44984818","CreationDate":"2017-07-08T09:50:54.923","ParentId":null,"OwnerUserId":"2513126","Title":null,"Body":"<p>My experience over the last months is the following:\nAdam is very easy to use because you don't have to play with initial learning rate very much and it almost always works. However, when coming to convergence Adam does not really sattle with a solution but jiggles around at higher iterations. While SGD gives an almost perfectly shaped loss plot and seems to converge much better in higher iterations.\nBut changing litte parts of the setup requires to adjust the SGD parameters or you will end up with NaNs... For experiments on architectures and general approaches I favor Adam, but if you want to get the best version of one chosen architecture you should use SGD and at least compare the solutions.<\/p>\n\n<p>I also noticed that a good initial SGD setup (learning rate, weight decay etc.) converges as fast as using Adam, at leas for my setup.\nHope this may help some of you!<\/p>\n\n<p>EDIT: Please note that the effects in my initial question are <strong>NOT<\/strong> normal even with Adam. Seems like I had a bug but I can't really remember the issue there.<\/p>\n"}]}
{"QuestionId":42221374,"AnswerCount":0,"Tags":"<python><caffe>","CreationDate":"2017-02-14T08:41:38.300","AcceptedAnswerId":null,"OwnerUserId":4952954.0,"Title":"What does [...] mean in Python code?","Body":"<p>sample:<\/p>\n\n<pre><\/pre>\n\n<p>This code always appears in caffe predict code. But I don't know what the  means. <\/p>\n","answers":[]}
{"QuestionId":42222197,"AnswerCount":1,"Tags":"<neural-network><protocol-buffers><deep-learning><caffe>","CreationDate":"2017-02-14T09:23:34.410","AcceptedAnswerId":42223172.0,"OwnerUserId":2467772.0,"Title":"Understanding caffe library","Body":"<p>I am trying to understand the <a href=\"http:\/\/caffe.berkeleyvision.org\/\" rel=\"nofollow noreferrer\">caffe library<\/a>. For that I run through step by step for feature_extraction.cpp and classification.cpp.<\/p>\n\n<p>In those cpp files, I found out files.<\/p>\n\n<p>I know caffe is formed using different layers. So those layer files inside layer folder are used.<\/p>\n\n<p>prototxt file is meant for the structure of a particular network such as googlenet, alexnet etc. Different net has different structure.<\/p>\n\n<p>caffemodel is the trained model using caffe library for a specific type of net structure.<\/p>\n\n<p>What do those net.cpp, caffe.pb.cc do? I mean how to understand their roles in forming this caffe deep learning network.<\/p>\n","answers":[{"AnswerId":"42223172","CreationDate":"2017-02-14T10:06:12.203","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You understand correctly that caffe implements deep learning by stacking \"layers\" one on top of the other to form a \"net\". <\/p>\n\n<p><strong><code>'net.cpp'<\/code><\/strong><br>\nEach layer works as a \"functional block\" and its behavior\/implementation is defined in <code>src\/caffe\/layers\/&lt;layer&gt;.cpp<\/code>, <code>src\/caffe\/layers\/&lt;layer&gt;.cu<\/code> and <code>include\/caffe\/layers\/&lt;layer&gt;.hpp<\/code>.<br>\nThe code that actually \"stack\" all the layers into a net can be found (mostly) in <code>net.cpp<\/code>.<\/p>\n\n<p><strong><code>'caffe.pb.h'<\/code>, <code>'caffe.pb.cc'<\/code><\/strong><br>\nIn order to define the specific structure of a specific deep net architecture (e.g., AlexNet, GoogLeNet, ResNet etc.) caffe uses <a href=\"\/questions\/tagged\/protocol-buffers\" class=\"post-tag\" title=\"show questions tagged &#39;protocol-buffers&#39;\" rel=\"tag\">protocol-buffers<\/a> library. The specific format of caffe protocol buffer is defined in <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto\" rel=\"nofollow noreferrer\"><code>src\/caffe\/proto\/caffe.proto<\/code><\/a>. The <code>caffe.proto<\/code> is \"compiled\" using <a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/overview\" rel=\"nofollow noreferrer\">google protobuffer<\/a> compiler to produce <code>'caffe.pb.h'<\/code> and <code>'caffe.pb.cc'<\/code> c++ code for parsing and processing caffe <code>prototxt<\/code> and <code>caffemodel<\/code> files.<\/p>\n"}]}
{"QuestionId":42222872,"AnswerCount":0,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2017-02-14T09:53:50.507","AcceptedAnswerId":null,"OwnerUserId":6430408.0,"Title":"how to predict with restored model by multi-process","Body":"<p>I have saved the training model.<\/p>\n\n<p>the predict code is as follow:<\/p>\n\n<pre><\/pre>\n\n<p>I used starmap multi process to get predict value:<\/p>\n\n<pre><\/pre>\n\n<p>here I only used one process. Run the project, the predict value for \"1.jpg\" can get successfully, then get the error:<\/p>\n\n<pre><\/pre>\n\n<p>I really do not know why the predict for second picture get the error ,but the first can get the right result?<\/p>\n\n<p>When I add   os._exit(0) in \"finally\",everything goes well, three pictures all can get the right result ,why?<\/p>\n\n<p>If add if I add sess.run(tf.global_variables_initializer()), after predict first picture, then get the log:Traceback (most recent call last):<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42223668,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-14T10:27:45.140","AcceptedAnswerId":42228577.0,"OwnerUserId":3157961.0,"Title":"NotFoundError when restore tensorflow model","Body":"<p>I am trying to save model and restore in other file using tensorflow. I use this code for train and save model.<\/p>\n\n<pre><\/pre>\n\n<p>and try to restore using this code.<\/p>\n\n<pre><\/pre>\n\n<p>In  line I have a NotFoundError. The error look like this<\/p>\n\n<pre><\/pre>\n\n<p>What is the real problem here? <\/p>\n","answers":[{"AnswerId":"43347517","CreationDate":"2017-04-11T13:33:21.457","ParentId":null,"OwnerUserId":"7840495","Title":null,"Body":"<p>In my opinion, you should give the tf.Variable w and b a name. For example,<br>\n<code>W = tf.Variable(tf.zeros([784,10]),name=\"w_var\")<\/code><br>\n<code>b = tf.Variable(tf.zeros([10]),name=\"b_var\")<\/code><br>\nGood luck to you!<\/p>\n"},{"AnswerId":"42228577","CreationDate":"2017-02-14T14:21:43.930","ParentId":null,"OwnerUserId":"4116272","Title":null,"Body":"<p>It looks like you are saving in your \"\/\" directory and loading from \".\/\".<\/p>\n"}]}
{"QuestionId":42223906,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-14T10:38:06.757","AcceptedAnswerId":44914739.0,"OwnerUserId":991519.0,"Title":"tensorflow : How to compute L1 difference between the batch (matrix of size B X DIM) and vocabulary (matrix of size V X DIM)?","Body":"<p>I have two tensorflow objects as follows<\/p>\n\n<pre><\/pre>\n\n<p>batch has 3 vectors, each of 2 dimensions. vocab has 6 vectors, each of 2 dimensions.\nI want to compute L1 difference between the every possible pair of vectors (generating 18 values).\ne.g. L1 differences of [1,2] would be <\/p>\n\n<pre><\/pre>\n\n<p>How can I do this with tensorflow functions ? The resulting matrix could be 6 X 3 (batch X vocab). Matrix multiplication seems to be the computation that will generate the expected order of the matrix. But is there a way to override the underlying operation of matrix multiplication ? (i.e. overriding the component-wise multiplication with absolute value of difference) ?<\/p>\n","answers":[{"AnswerId":"44914739","CreationDate":"2017-07-04T22:32:17.167","ParentId":null,"OwnerUserId":"8256101","Title":null,"Body":"<p>May be this response is not useful to you anymore, but I recently came across to this example of <a href=\"https:\/\/github.com\/PacktPublishing\/TensorFlow-Machine-Learning-Cookbook\/blob\/master\/Chapter%2005\/nearest_neighbor.py\" rel=\"nofollow noreferrer\">k-nearest neighbors<\/a> with L1 and L2 distance.\nSo, I guess, your problem can be solved as:<\/p>\n\n<pre><code>batch = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3,2], name='batch')\nvocab = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0,1.0,2.0,3.0,4.0,5.0,6.0], shape=[6,2], name='vocab')\ndistance_L1 = tf.reduce_sum(tf.abs(tf.subtract(vocab, tf.expand_dims(batch,1))), axis=2)\n<\/code><\/pre>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":42224307,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><prediction>","CreationDate":"2017-02-14T10:54:37.237","AcceptedAnswerId":42224496.0,"OwnerUserId":6925731.0,"Title":"Retrain model in Tensorflow","Body":"<p>I have a simple neural-network using Tensorflow.\nHere is the Session:<\/p>\n\n<pre><\/pre>\n\n<p>at the last 2 rows I saved the model and restore the graph in another class:<\/p>\n\n<pre><\/pre>\n\n<p>I want to re-train the model, which means not initialize the weights, just to update them from the last point that it stopped.<\/p>\n\n<p>How can I do that? <\/p>\n","answers":[{"AnswerId":"42224496","CreationDate":"2017-02-14T11:03:59.007","ParentId":null,"OwnerUserId":"3502070","Title":null,"Body":"<p>From <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/saving_and_restoring_variables<\/a><\/p>\n\n<blockquote>\n  <p>tf.train.Saver.restore(sess, save_path)<\/p>\n  \n  <p>Restores previously saved variables.<\/p>\n  \n  <p>This method runs the ops added by the constructor for restoring\n  variables. It requires a session in which the graph was launched. <strong>The\n  variables to restore do not have to have been initialized, as\n  restoring is itself a way to initialize variables.<\/strong><\/p>\n<\/blockquote>\n\n<p>The following example is from <a href=\"https:\/\/www.tensorflow.org\/how_tos\/variables\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/variables\/<\/a><\/p>\n\n<pre><code># Create some variables.\nv1 = tf.Variable(..., name=\"v1\")\nv2 = tf.Variable(..., name=\"v2\")\n...\n# Add ops to save and restore all the variables.\nsaver = tf.train.Saver()\n\n# Later, launch the model, use the saver to restore variables from disk, and\n# do some work with the model.\nwith tf.Session() as sess:\n  # Restore variables from disk.\n  saver.restore(sess, \"\/tmp\/model.ckpt\")\n  print(\"Model restored.\")\n  # Do some work with the model\n  ...\n<\/code><\/pre>\n"}]}
{"QuestionId":42224327,"AnswerCount":1,"Tags":"<python><tensorflow><queue>","CreationDate":"2017-02-14T10:55:50.237","AcceptedAnswerId":42279347.0,"OwnerUserId":5861256.0,"Title":"OutOfRangeError (see above for traceback): FIFOQueue '_1_batch\/fifo_queue' is closed and has insufficient elements (requested 32, current size 0)","Body":"<p>I m having problem with tensorflow reading images in using queue. Please let me know what mistake i am doing. Below is the code.<\/p>\n\n<pre><\/pre>\n\n<p>I have used different capacity values but none have worked <\/p>\n\n<pre><\/pre>\n\n<p>Please help me <\/p>\n","answers":[{"AnswerId":"42279347","CreationDate":"2017-02-16T16:20:23.423","ParentId":null,"OwnerUserId":"2670000","Title":null,"Body":"<p>You can add a filename queue, e.g., <code>string_input_producer<\/code>, to run over all files in the folder with replacement. And try to comment out the <code>shuffle_batch<\/code> and see if the filename queue is getting any data. This method could run through multiple times if you left <code>num_epoch<\/code> to none.<\/p>\n"}]}
{"QuestionId":42224655,"AnswerCount":1,"Tags":"<python-2.7><import><tensorflow><protocol-buffers><importerror>","CreationDate":"2017-02-14T11:11:52.763","AcceptedAnswerId":42253328.0,"OwnerUserId":7499201.0,"Title":"Tensorflow import error after python update","Body":"<p>I am using tensorflow with python 2.7. However, after updating python 2.7.10 to 2.7.13, I get an import error with tensorflow<\/p>\n\n<pre>File \"\", line 1, in \n  File \"\/Users\/usrname\/Library\/Python\/2.7\/lib\/python\/site-\n                        packages\/tensorflow\/__init__.py\", line 24, in \n    from tensorflow.python import *\n  File \"\/Users\/usrname\/Library\/Python\/2.7\/lib\/python\/site-\n                        packages\/tensorflow\/python\/__init__.py\", line 63, in \n    from tensorflow.core.framework.graph_pb2 import *\n  File \"\/Users\/usrname\/Library\/Python\/2.7\/lib\/python\/site-\n                        packages\/tensorflow\/core\/framework\/graph_pb2.py\", line 6, in \n    from google.protobuf import descriptor as _descriptor\nImportError: No module named google.protobuf<\/pre>\n\n<p>Output from <\/p>\n\n<pre>Requirement already satisfied: protobuf in \/usr\/local\/lib\/python2.7\/site-packages \nRequirement already satisfied: setuptools in \/Users\/usrname\/Library\/Python\/2.7\/lib\/\n    python\/site-packages (from protobuf) \nRequirement already satisfied: six>=1.9 in \/Library\/Python\/2.7\/site-packages\/\n    six-1.10.0-py2.7.egg (from protobuf) \nRequirement already satisfied: appdirs>=1.4.0 in \/usr\/local\/lib\/python2.7\/site-packages\n    (from setuptools->protobuf) \nRequirement already satisfied: packaging>=16.8 in \/usr\/local\/lib\/python2.7\/site-packages \n    (from setuptools->protobuf) \nRequirement already satisfied: pyparsing in \/usr\/local\/lib\/python2.7\/site-packages \n    (from packaging>=16.8->setuptools->protobuf) <\/pre>\n\n<p>Output from :<\/p>\n\n<pre>\/Library\/Frameworks\/Python.framework\/Versions\/2.7\/bin\/python<\/pre>\n\n<p>I believe this path changed after the python update, but not sure. A solution could possibly be to downgrade python, but this seems like a bad solution? As I work in a team, I would like to avoid reinstalling Tensorflow due to end up with different versions, but this is perhaps the way to go? Any advice?<\/p>\n\n<hr>\n\n<p>Update: I tired to install tensorflow all over, but the same error keeps popping up. Maybe the problem is the environment variables as returns (which is different from which python)?<\/p>\n","answers":[{"AnswerId":"42253328","CreationDate":"2017-02-15T15:19:55.980","ParentId":null,"OwnerUserId":"7497449","Title":null,"Body":"<p>I had a similar problem. Make sure that pip and python has the same path when typing <code>which pip<\/code>and <code>which python<\/code>. If they differ, change your ~.bash_profile so that the python path match the pip path, and use <code>source ~\\.bash_profile<\/code>.<\/p>\n\n<p>If that doesn't work, I would try to reinstall pip and tensorflow.<\/p>\n\n<p>I installed pip using this command:<\/p>\n\n<pre>wget https:\/\/bootstrap.pypa.io\/get-pip.py\nsudo python2.7 get-pip.py<\/pre>\n"}]}
{"QuestionId":42224712,"AnswerCount":0,"Tags":"<ios><opencv><tensorflow>","CreationDate":"2017-02-14T11:15:10.433","AcceptedAnswerId":null,"OwnerUserId":4622363.0,"Title":"how to pass extra input parameter in tensorflow in iOS?","Body":"<p>I am sending parameter to tensorflow as following:  <\/p>\n\n<pre><\/pre>\n\n<p>I tried as<\/p>\n\n<pre><\/pre>\n\n<p>but it is not accepting extra parameter \nHow can i send extra input parameter ?<\/p>\n","answers":[]}
{"QuestionId":42225143,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><jupyter-notebook>","CreationDate":"2017-02-14T11:36:26.810","AcceptedAnswerId":42226790.0,"OwnerUserId":4520592.0,"Title":"Tensorflow: why softmax outputs [1, 0, 0..., 0]","Body":"<p>I have a neural net model, it's last layer is fully connected layer with 9 output neurons. <br>\nTo train my network correctly, I'm using softmax_cross_entropy_with_logits. <br>\nIt trains ok, but when I want to evaluate my model, I want probabilities also. <br>\nSo I take an evaluation sample and feed it to the network. <br>\nAfter that I apply softmax to the output and get <br>\n    <\/p>\n\n<p>Here unnormalized probabilites also:<\/p>\n\n<pre><\/pre>\n\n<p>I'm also getting probility of 1 and rest are zeros.\nCould anyone please help to handle this issue?<\/p>\n\n<p>EDIT:\nInput images are of shape 64 * 160. <br>\nAll activation functions are relu. <br>\nMax poolings are 2x2. <br>\nIn conv_plus_max_pool_layer(x_image, 5, 1, 96) 5 is kernel size. <br>\nHere is network layout:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42226790","CreationDate":"2017-02-14T12:56:11.577","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Looks like your network is pretty sure about the output ;)<\/p>\n\n<p>In this case, I don't think we can do a lot for you without your network layout... Some gut feelings from my side: the layer leading up to your output layer has too many nodes (thus giving you these huuuge numbers), and I suspect that you don't use nonlinearities such as RELU, or tanh. Another thing you might want to check are the initial values for the weights (might be too big), and the learning rate you are using (might be too high). <\/p>\n"}]}
{"QuestionId":42225688,"AnswerCount":1,"Tags":"<python><python-imaging-library><caffe><deep-dream>","CreationDate":"2017-02-14T12:02:31.153","AcceptedAnswerId":42292718.0,"OwnerUserId":1542594.0,"Title":"Deepdream python script not showing images","Body":"<p>I have followed the guide at <a href=\"https:\/\/github.com\/google\/deepdream\/blob\/master\/dream.ipynb\" rel=\"nofollow noreferrer\">Link<\/a> up until the<\/p>\n\n<pre><\/pre>\n\n<p>step. Now it runs without errors but it doesn't show the images. This is the output:<\/p>\n\n<pre><\/pre>\n\n<p>It continues with similar output.<\/p>\n\n<p>I can show a file with this code:<\/p>\n\n<pre><\/pre>\n\n<p>But it doesn't work with the code that the guide provides. Only thing it seems strange is that f (StringIO) contains \"garbage\" when I print it using f.getvalue().<\/p>\n\n<p>Any insight on why is it not showing the images?<\/p>\n\n<p><strong>EDIT:<\/strong>\nIf I change the line (in showarray method)<\/p>\n\n<pre><\/pre>\n\n<p>to<\/p>\n\n<pre><\/pre>\n\n<p>It doesn't show the image but at least it is saving it in the specified folder.<\/p>\n\n<p><strong>EDIT 2:<\/strong>\nChanging the line<\/p>\n\n<pre><\/pre>\n\n<p>with<\/p>\n\n<pre><\/pre>\n\n<p>shows the image. Now the problem is that I can't close it so it shows one on top of the others.<\/p>\n","answers":[{"AnswerId":"42292718","CreationDate":"2017-02-17T08:16:14.037","ParentId":null,"OwnerUserId":"1542594","Title":null,"Body":"<p>Changing the line<\/p>\n\n<pre><code>display(Image(data=f.getvalue()))\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>im = PIL.Image.open('\/home\/jose\/Downloads\/deepdream\/test')\n<\/code><\/pre>\n\n<p>shows the image. Now the problem is that I can't close it so it shows one on top of the others.<\/p>\n"}]}
{"QuestionId":42229080,"AnswerCount":1,"Tags":"<tensorflow><restore>","CreationDate":"2017-02-14T14:46:10.393","AcceptedAnswerId":null,"OwnerUserId":3681744.0,"Title":"Initializing variables from a restored model in TensorFlow","Body":"<p>Suppose that I train and save the following simple network (Lets call it network A):<\/p>\n\n<pre><\/pre>\n\n<p>Now, I would like to restore network A, design a new network B and use parameters of network A to initialize the first 3 layers of network B:<\/p>\n\n<pre><\/pre>\n\n<p>How can I do this in TensorFlow?<\/p>\n","answers":[{"AnswerId":"42237288","CreationDate":"2017-02-14T22:17:46.913","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>You can just create network A + B, and remember the Conv3 layer. You can then compute a loss on the output of Conv3, then train for a while. Once you are satisfied, just switch to training on a loss on Conv5.<\/p>\n"}]}
{"QuestionId":42229360,"AnswerCount":0,"Tags":"<tensorflow><tf-slim><deconvolution>","CreationDate":"2017-02-14T14:58:45.953","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"How to perform upsampling using the convolution_transpose function ?","Body":"<p>I want to use the convolution_transpose function of tf.slim and it seems that, no matter how I choose my kernel size and strides, I cannot get the output of the convolution_transpose layer to be larger than that of the initial convolutional layer. Am I missing something here? <\/p>\n","answers":[]}
{"QuestionId":42229376,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><unsupervised-learning><autoencoder>","CreationDate":"2017-02-14T14:59:24.890","AcceptedAnswerId":null,"OwnerUserId":7563536.0,"Title":"TensorFlow network having low epoch loss but still getting low accuracy","Body":"<p>So I'm trying to build a deep net with stacked auto-encoders for training the network with MNIST dataset. I first pre-trained the model (layer-wise) and did a normal backprop for fine tuning. The problem is that when I do a normal backprop on randomly initialized weights, I get very high epoch loss but still accuracy is ~96%., but when I do the same with stacked auto-encoders, I get comparatively low epoch loss and low accuracy. I'm attaching the complete source code and output here.  <\/p>\n\n<p>test2.py<\/p>\n\n<pre><\/pre>\n\n<p>stacked_autoencoder.py<\/p>\n\n<pre><\/pre>\n\n<p>When I run test2.py, I get this<\/p>\n\n<pre><\/pre>\n\n<p>I feel like I'm missing something here. Please help!<\/p>\n","answers":[]}
{"QuestionId":42230120,"AnswerCount":1,"Tags":"<python><tensorflow><mnist>","CreationDate":"2017-02-14T15:32:22.190","AcceptedAnswerId":null,"OwnerUserId":7508642.0,"Title":"TensorFlow, Implementing Solution","Body":"<p>I'm new to Python, Machine Learning and TensorFlow and feeling very overwhelmed!<\/p>\n\n<p>I'm still working on the MNIST dataset and have the following code (which I've complied from the TensorFlow tutorials):<\/p>\n\n<pre><\/pre>\n\n<p>My question is, from here, how do I extract the solution to make a usable application? By which I mean, how do I get to the stage where I can feed it one image and get one prediction of a number out based on the training of the MNIST dataset.<\/p>\n\n<p>Many thanks<\/p>\n","answers":[{"AnswerId":"42231046","CreationDate":"2017-02-14T16:14:08.367","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>With your code it should work like this: <\/p>\n\n<p><code>test_prediction = y.eval(feed_dict={x: mnist.test.images})<\/code><\/p>\n\n<p>If you want to have \"your own images\", you have to make sure they are the same format as your mnist.test.images variable. In this case, you can replace mnist.test.images in the code above by any Numpy array filled with arrays with length 787 (image of 28 by 28 pixels). \nPlease let me know if this works!<\/p>\n"}]}
{"QuestionId":42231022,"AnswerCount":0,"Tags":"<tensorflow><tflearn>","CreationDate":"2017-02-14T16:13:05.450","AcceptedAnswerId":null,"OwnerUserId":1619749.0,"Title":"How to read pretrained weights of google inception v3 model from tensorflow to tflean?","Body":"<p>I am currently trying to train google inception v3 model with TF learn. Tensorflow has provided the trained model over imagenet dataset. But it is not in a usual .tflearn format. I have gone through the documentation of tflearn, but I cannot find any description to convert tensorflow weights into TFlearn format. I know that the right place to ask this question is tflearn gihub page. But it is currently very busy with high number of open issues.\nCould anyone point me to the correct direction?<\/p>\n","answers":[]}
{"QuestionId":42231162,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><computer-vision><deep-learning><caffe>","CreationDate":"2017-02-14T16:21:47.567","AcceptedAnswerId":42241738.0,"OwnerUserId":288609.0,"Title":"semantic segmentation for large images","Body":"<p>I am working on a limited number of large size images, each of which can have  pixels. To train a semantic segmentation model using FCN or U-net, I construct a large sample of training sets, each training image is . <\/p>\n\n<p>In the prediction stage, what I do is to cut a large image into small pieces, the same as trainning set of , and feed these small pieces into the trained model, get the predicted mask. Afterwards, I just stitch these small patches together to get the mask for the whole image. Is this the right mechanism to perform the semantic segmentation against the large images?<\/p>\n","answers":[{"AnswerId":"42241738","CreationDate":"2017-02-15T05:59:19.450","ParentId":null,"OwnerUserId":"5009875","Title":null,"Body":"<p>Input image data:\nI would not advice feeding the big image (3072x3072) directly into the caffe.\nBatch of small images will fit better into the memory and parallel programming will too come into play.\nData Augmentation will also be feasible.<\/p>\n\n<p>Output for big Image:\nAs for the output of big Image, you better recast the input size of FCN to 3072x3072 during test phase. Because, layers of FCN can accept inputs of any size.\nThen you will get 3072x3072 segmented image as output.<\/p>\n"},{"AnswerId":"47771986","CreationDate":"2017-12-12T11:53:46.950","ParentId":null,"OwnerUserId":"5197034","Title":null,"Body":"<p>Your solution is often used for this kind of problem. However, I would argue that it depends on the data if it truly makes sense. Let me give you two examples you can still find on kaggle.<\/p>\n\n<p>If you wanted to mask certain parts of satellite images, you would probably get away with this approach without a drop in accuracy. These images are highly repetitive and there's likely no correlation between the segmented area and where in the original image it was taken from.<\/p>\n\n<p>If you wanted to segment a car from its background, it wouldn't be desirable to break it into patches. Over several layers the network will learn the global distribution of a car in the frame. It's very likely that the mask is positive in the middle and negative in the corners of the image.<\/p>\n\n<p>Since you didn't give any specifics what you're trying to solve, I can only give a general recommendation: <strong>Try to keep the input images as large as your hardware allows.<\/strong> In many situation I would rather downsample the original images than breaking it down into patches.<\/p>\n\n<p>Concerning the recommendation of curio1729, I can only advise against training on small patches and testing on the original images. While it's technically possible thanks to fully convolutional networks, you're changing the data to an extend, that might very likely hurt performance. CNNs are known for their extraction of local features, but there's a large amount of global information that is learned over the abstraction of multiple layers.<\/p>\n"}]}
{"QuestionId":42231473,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-14T16:36:54.540","AcceptedAnswerId":null,"OwnerUserId":5407700.0,"Title":"tensorflow - add tensor by name","Body":"<p>Say I define a variable in a scope such as:<\/p>\n\n<pre><\/pre>\n\n<p>and I want to reference it later to perform an operation (eg addition) with another variable something like;<\/p>\n\n<pre><\/pre>\n\n<p>How do I go about this?<\/p>\n","answers":[{"AnswerId":"42236005","CreationDate":"2017-02-14T20:54:23.820","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>In this case you can remove the \n<code>name=<\/code>\npart. <\/p>\n\n<p>More details can be found here: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sharing_variables#get_variable\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/state_ops\/sharing_variables#get_variable<\/a><\/p>\n"}]}
{"QuestionId":42231821,"AnswerCount":0,"Tags":"<regression><caffe>","CreationDate":"2017-02-14T16:54:16.807","AcceptedAnswerId":null,"OwnerUserId":5009875.0,"Title":"caffe constant output with multilabel regression?","Body":"<p>As discussed in this PR <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/1765\" rel=\"nofollow noreferrer\">#1765<\/a> , I face the same problem.\nIt's not clear what is the solution to this problem\nWhen I try to do multilabel regression with caffe ,always prediction is constant.\nThen after inspecting the weights , I found that the bias of the last fc layer are the exact values of the constant output.\nWhat could be the problem?<\/p>\n","answers":[]}
{"QuestionId":42232341,"AnswerCount":2,"Tags":"<lua><torch>","CreationDate":"2017-02-14T17:19:07.677","AcceptedAnswerId":null,"OwnerUserId":2978677.0,"Title":"Dividing a DoubleTensor matrix by a DoubleTensor number in torch","Body":"<p>I'm trying to update a matrix of real valued numbers in a for loop using .\nHere is what I'd like to do:<\/p>\n\n<pre><\/pre>\n\n<p>Just to be clear:<\/p>\n\n<pre><\/pre>\n\n<p>The error when I do the division is:<\/p>\n\n<pre><\/pre>\n\n<p>I'm a heavy numpy users so I thought \"broadcasting\" was the issue, therefore I tried simulating it through   with no luck. If denom is just a number (not a ) everything works fine anyway. Using the element doesn't work either,  gives the same error.<\/p>\n\n<p>What am I doing wrong?<\/p>\n\n<p>Edit:\nI tried using<\/p>\n\n<pre><\/pre>\n\n<p>which doesn't throw an error but the results are wrong. To see this, try the following:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42232696","CreationDate":"2017-02-14T17:36:59.617","ParentId":null,"OwnerUserId":"2978677","Title":null,"Body":"<p>I managed by using <code>apply<\/code>, <code>repeatTensor<\/code> and finally <code>cmul<\/code> for element-wise multiplication<\/p>\n\n<pre><code>rPr = r:view(1, N) * P * r:view(N, 1)\ndenominator = (1 + rPr):apply(function(x) return x^(-1) end)\ndelta_w:cmul(torch.repeatTensor(denominator, N, N))\n<\/code><\/pre>\n\n<p>though I wonder if this can be transferred to GPU with cutorch.<\/p>\n"},{"AnswerId":"42471535","CreationDate":"2017-02-26T17:05:03.640","ParentId":null,"OwnerUserId":"7596504","Title":null,"Body":"<p>If you do <code>denominator[1][1]<\/code>, you will get a 'number' instead of a 'torch.Tensor'. Then you can just write the division statement normally.<\/p>\n\n<pre><code>-- W and P are of size NxN, r is of size N\ndelta_W = P * r:view(N, 1) * r:view(1, N) * P  -- this is an NxN\ndenominator = 1 + r:view(1, N) * P * r:view(N, 1)  -- this is a 1x1\ndelta_W = delta_w \/ denominator[1][1]\nW = W + delta_W\n<\/code><\/pre>\n\n<p>By the way, in the 1st statement did you want one of the P matrices transposed (<code>P:t()<\/code>)?<\/p>\n"}]}
{"QuestionId":42232803,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><restore>","CreationDate":"2017-02-14T17:42:44.937","AcceptedAnswerId":null,"OwnerUserId":3681744.0,"Title":"Variable names in .ckpt file of TensorFlow","Body":"<p>I have the following variable in my Convnet:<\/p>\n\n<pre><\/pre>\n\n<p>and save the model when it is trained. For figuring out the value of  once training is finished, I use the  function introduced in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/inspect_checkpoint.py\" rel=\"nofollow noreferrer\">here<\/a>. However, in addition to , I see  and  are also saved in the ckpt file. What are these variables exactly? How are they different from ?<\/p>\n","answers":[]}
{"QuestionId":42233250,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-14T18:06:52.813","AcceptedAnswerId":42235017.0,"OwnerUserId":7241513.0,"Title":"Error from tf.reshape() after trying to convert vector into matrix","Body":"<p>I am trying to convert a one-dimensional array into a TF matrix, for use in a convolutional network, similar to how its done in the TF deep MNIST example, by using :<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>5 is my  and 480 is the length of the original array. I want to convert it into a 60x8 tensor, plus 1 output channel, so I am following with the TF MNIST example of using the target shape of .<\/p>\n\n<ol>\n<li>Any idea why I'm getting this? Seems like in the MNIST example they use -1 to adjust the single line into AxA matrix and add one extra dimension for the output channel?<\/li>\n<li>Can I even use non-square matrixes like this, or will it give me an error down the road? I tried doing something like 400=20x20 as well just to see if the error is caused by the uneven shape, but still getting the error...<\/li>\n<\/ol>\n","answers":[{"AnswerId":"42235017","CreationDate":"2017-02-14T19:53:50.603","ParentId":null,"OwnerUserId":"2878004","Title":null,"Body":"<p>On your first line of code you define a placeholder and store it in the python variable X. On the second line of code you define a reshape operator and also store it in the python variable X. You have now overwritten the previous value of X and lost access to the placeholder. You are trying to feed in a value to the placeholder but you are actually feeding it to the reshape operator. If you use different names for these variables then you will not see this error.<\/p>\n"}]}
{"QuestionId":42233963,"AnswerCount":2,"Tags":"<debugging><tensorflow><keras>","CreationDate":"2017-02-14T18:48:37.547","AcceptedAnswerId":42236399.0,"OwnerUserId":4402325.0,"Title":"Debugging keras intermediate layer or objective variables with tensorflow","Body":"<p>I'm writing a custom objective to train a Keras (with tensorflow backend) model but I need to debug some intermediate computation. For simplicity, let's say I have:<\/p>\n\n<pre><\/pre>\n\n<p>I could not find an easy way to access, for example, the intermediate variable diff or its shape during training. In this simple example, I know that I could return diff to print its values, but my actual loss is more complex and I can't return intermediate values without getting compiling errors.<\/p>\n\n<p>Is there an easy way to debug intermediate variables in Keras?<\/p>\n","answers":[{"AnswerId":"57730056","CreationDate":"2019-08-30T16:06:59.430","ParentId":null,"OwnerUserId":"1191087","Title":null,"Body":"<p>In TensorFlow 2, you can now add IDE breakpoints in the TensorFlow Keras models\/layers\/losses, including when using the fit, evaluate, and predict methods. However, you must add <code>model.run_eagerly = True<\/code> <strong>after<\/strong> calling <code>model.compile()<\/code> for the values of the tensor to be available in the debugger at the breakpoint. For example,<\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow.keras.layers import Dense\nfrom tensorflow.keras.models import Model\nfrom tensorflow.keras.optimizers import Adam\n\ndef custom_loss(y_pred, y_true):\n    diff = y_pred - y_true\n    return tf.keras.backend.square(diff)  # Breakpoint in IDE here. =====\n\nclass SimpleModel(Model):\n\n    def __init__(self):\n        super().__init__()\n        self.dense0 = Dense(2)\n        self.dense1 = Dense(1)\n\n    def call(self, inputs):\n        z = self.dense0(inputs)\n        z = self.dense1(z)\n        return z\n\nx = tf.convert_to_tensor([[1, 2, 3], [4, 5, 6]], dtype=tf.float32)\ny = tf.convert_to_tensor([0, 1], dtype=tf.float32)\n\nmodel0 = SimpleModel()\nmodel0.run_eagerly = True\nmodel0.compile(optimizer=Adam(), loss=custom_loss)\ny0 = model0.fit(x, y, epochs=1)  # Values of diff *not* shown at breakpoint. =====\n\nmodel1 = SimpleModel()\nmodel1.compile(optimizer=Adam(), loss=custom_loss)\nmodel1.run_eagerly = True\ny1 = model1.fit(x, y, epochs=1)  # Values of diff shown at breakpoint. =====\n<\/code><\/pre>\n\n<p>This also works for debugging the outputs of intermediate network layers (for example, adding the breakpoint in the <code>call<\/code> of the SimpleModel).<\/p>\n\n<p><em>Note: this was tested in TensorFlow <code>2.0.0-rc0<\/code>.<\/em><\/p>\n"},{"AnswerId":"42236399","CreationDate":"2017-02-14T21:18:11.740","ParentId":null,"OwnerUserId":"1643939","Title":null,"Body":"<p>This is not something that is solved in Keras as far as I know, so you have to resort to backend-specific functionality. Both <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/printing.html\" rel=\"noreferrer\">Theano<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/control_flow_ops\/debugging_operations\" rel=\"noreferrer\">TensorFlow<\/a> have <code>Print<\/code> nodes that are identity nodes (i.e., they return the input node) and have the side-effect of printing the input (or some tensor of the input).<\/p>\n\n<p>Example for Theano:<\/p>\n\n<pre><code>diff = y_pred - y_true\ndiff = theano.printing.Print('shape of diff', attrs=['shape'])(diff)\nreturn K.square(diff)\n<\/code><\/pre>\n\n<p>Example for TensorFlow:<\/p>\n\n<pre><code>diff = y_pred - y_true\ndiff = tf.Print(diff, [tf.shape(diff)])\nreturn K.square(diff)\n<\/code><\/pre>\n\n<p>Note that this only works for intermediate values. Keras expects tensors that are passed to other layers to have specific attributes such as <code>_keras_shape<\/code>. Values processed by the backend, i.e. through <code>Print<\/code>, usually do not have that attribute. To solve this, you can wrap debug statements in a <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"noreferrer\"><code>Lambda<\/code><\/a> layer for example.<\/p>\n"}]}
{"QuestionId":42234535,"AnswerCount":1,"Tags":"<python><neural-network><keras><perceptron><data-formats>","CreationDate":"2017-02-14T19:24:35.843","AcceptedAnswerId":42235801.0,"OwnerUserId":4266946.0,"Title":"Keras correct input shape for multilayer perceptron","Body":"<p>I'm trying to make a basic MLP example in keras. My input data has the shape  and my testing data has the shape .  and  are the numbers of samples of training and test data (in other words, the shape of the data is , but there are 2000 and 500 pieces of training and testing data). The output should have two classes.<\/p>\n\n<p>I'm unsure what value to use for the  parameter on the first layer of the network. Using the code from the mnist example in the keras repository, I have (<strong>updated<\/strong>):<\/p>\n\n<pre><\/pre>\n\n<p>where  is 75 * 75 (emulating the MNIST example). The error I get is:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"42235801","CreationDate":"2017-02-14T20:41:09.740","ParentId":null,"OwnerUserId":"4955956","Title":null,"Body":"<p>From keras MLP example, <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/mnist_mlp.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/mnist_mlp.py<\/a><\/p>\n\n<pre><code># the data, shuffled and split between train and test sets\n(X_train, y_train), (X_test, y_test) = mnist.load_data()\n\nX_train = X_train.reshape(60000, 784)\nX_test = X_test.reshape(10000, 784)\n<\/code><\/pre>\n\n<p>And the model input <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(512, input_shape=(784,)))\n<\/code><\/pre>\n\n<p>So you should reshape your train and test to (2000,75*75) and (500,75*75) with <\/p>\n\n<pre><code>train_data = train_data.reshape(2000, 75*75)\ntest_data = test_data.reshape(500, 75*75)\n<\/code><\/pre>\n\n<p>and then set the model input shape as you did<\/p>\n\n<pre><code>model.add(Dense(512, input_shape=(75*75,)))\n<\/code><\/pre>\n"}]}
{"QuestionId":42234735,"AnswerCount":0,"Tags":"<tensorflow><conv-neural-network><torch>","CreationDate":"2017-02-14T19:37:33.327","AcceptedAnswerId":null,"OwnerUserId":7302262.0,"Title":"how can use torch model?","Body":"<p>I have a <strong><a href=\"https:\/\/github.com\/satoshiiizuka\/siggraph2016_colorization\" rel=\"nofollow noreferrer\">Torch Model<\/a><\/strong> which is trained on a large scale dataset (<a href=\"http:\/\/places.csail.mit.edu\/\" rel=\"nofollow noreferrer\">Places Dataset<\/a>) and it's authors uploaded it on github, i am working on a similar project and i want to make use of it and use it's trained weights instead of use the large dataset to train it and save time and efforts, it is possible ? how can i know the only the trained filters weights? i don't want to copy the code, i only want to make use of it and save time and efforts.<\/p>\n\n<p><strong>NOTE:<\/strong>  I use Tensoflow in my implementation.<\/p>\n","answers":[]}
{"QuestionId":42235298,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-14T20:09:16.883","AcceptedAnswerId":null,"OwnerUserId":7466373.0,"Title":"sequence to sequence models output","Body":"<p>I executed the code translate.py from sequence to sequence models tutorial of tensorflow 60 hours ago and still waiting for the output.\nCan anyone tell me how much more time will it take to display the output\nSystem specifications:\nI installed Ubuntu 16.04 on VMware with a disk space of 35 GB and RAM of 1 GB.<\/p>\n","answers":[{"AnswerId":"45399596","CreationDate":"2017-07-30T12:02:01.137","ParentId":null,"OwnerUserId":"5915270","Title":null,"Body":"<p>The best way is check the code . Normally in the training phase we don't print the progress in every epoch .  So it might take some time . <\/p>\n\n<p>Any way you should see in the following info in the terminal <\/p>\n\n<ol>\n<li>May be something model creation .<\/li>\n<li>Some GPU details when starting a session . <\/li>\n<\/ol>\n\n<p>So as soon as session started it should print GPU details like , \nHow many GPUs , available memory etc . And some configuration details also . Normally for that it won't take much time .<\/p>\n\n<p><strong>Soon after model creation you should able to see something. Maximum with in 10 .(If your machine is really really slow)<\/strong> <\/p>\n\n<p>I think there's an error in your implementation ! <\/p>\n"}]}
{"QuestionId":42235611,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><keras><keras-layer>","CreationDate":"2017-02-14T20:29:54.743","AcceptedAnswerId":42235908.0,"OwnerUserId":7565103.0,"Title":"ValueError: Error when checking model target: expected dense_4 to have shape (None, 4) but got array with shape (13252, 1)","Body":"<p>Hi does anyone have any ideas why this error is happening?\nHere is the error <\/p>\n\n<pre><\/pre>\n\n<p>And here is the code:<\/p>\n\n<pre><\/pre>\n\n<p>The input shape is (X):<\/p>\n\n<pre><\/pre>\n\n<p>And (y):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42235908","CreationDate":"2017-02-14T20:47:53.480","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>Since your network has <strong>four<\/strong> outputs, your <strong>y<\/strong> has to be N x 4 matrix, not a vector of length N. Either change y, or last layer to be <code>Dense(1)<\/code><\/p>\n"}]}
{"QuestionId":42235664,"AnswerCount":0,"Tags":"<c++><tensorflow>","CreationDate":"2017-02-14T20:32:57.307","AcceptedAnswerId":null,"OwnerUserId":3002273.0,"Title":"Tensorflow Modify OpKernelConstruction Attributes","Body":"<p>In kernel operations, an OpKernelConstruction pointer is given to the OpKernel constructor, for <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/kernels\/fifo_queue_op.cc#L44\" rel=\"nofollow noreferrer\">example<\/a><\/p>\n\n<p>I would like to change some attributes in the context before passing it to the parent constructor. <\/p>\n\n<p>For example, something like<\/p>\n\n<pre><\/pre>\n\n<p>But it appears in the description of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/op_kernel.h#L213\" rel=\"nofollow noreferrer\">OpKernelConstruction<\/a>, that no modifying method exists. Is there a way to modify it?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>I want to modify the attributes because I am attempting to make a FIFOQueue that pulls data from a database. I only want the user to have to specify the data stream id and then the queue would hit the database for the shape and type of the data. As such, I want to modify the context to add the shape and type info after I get it from the DB, but before I pass the context onto the base class. The base class makes use of shape and type, so I need to make sure it is in correspondence with the DB.<\/p>\n","answers":[]}
{"QuestionId":42236124,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-14T21:00:49.493","AcceptedAnswerId":42236227.0,"OwnerUserId":7449493.0,"Title":"def extract_features(list_images) ....Gives: NameError: name 'image' is not defined","Body":"<p>This code is producing this error:<\/p>\n\n<blockquote>\n  <p>NameError: name 'image' is not defined<\/p>\n<\/blockquote>\n\n<p>I would appreciate any help:<\/p>\n\n<pre><\/pre>\n\n<p>In a previous section of the code I declare this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42236227","CreationDate":"2017-02-14T21:07:25.867","ParentId":null,"OwnerUserId":"4540123","Title":null,"Body":"<p>Looks like you have indentation bug, you need to add another tab to <\/p>\n\n<pre><code>  if not gfile.Exists(image):\n        tf.logging.fatal('File does not exist %s', image)\n<\/code><\/pre>\n\n<p>lines.<\/p>\n\n<p>The issue: image is defined just if <code>list_images<\/code> if a iterable with at least one value.<\/p>\n"}]}
{"QuestionId":42236574,"AnswerCount":2,"Tags":"<python><ubuntu><tensorflow><pip><deep-learning>","CreationDate":"2017-02-14T21:29:03.120","AcceptedAnswerId":null,"OwnerUserId":2414957.0,"Title":"AttributeError: 'module' object has no attribute 'histogram' when using tf-faster-rcnn","Body":"<p>Please have a look at this issue and let me know if you know how to fix it?\nHere's the link to the original repo:\n<a href=\"https:\/\/github.com\/endernewton\/tf-faster-rcnn\" rel=\"nofollow noreferrer\">https:\/\/github.com\/endernewton\/tf-faster-rcnn<\/a><\/p>\n\n<pre><\/pre>\n\n<p>Operating System:<\/p>\n\n<pre><\/pre>\n\n<p>The output of bazel version<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42237576","CreationDate":"2017-02-14T22:41:16.057","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>You are using an older version of tensorflow. The old version was <code>tf.histogram_summary<\/code>. You can see a list of API changes in our <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/compatibility\/tf_upgrade.py#L116\" rel=\"nofollow noreferrer\">upgrade script<\/a>.<\/p>\n"},{"AnswerId":"42237116","CreationDate":"2017-02-14T22:05:19.953","ParentId":null,"OwnerUserId":"2414957","Title":null,"Body":"<p>According to the answer I got in github issues, I am using a very old version of TensorFlow. This solved the problem:<\/p>\n\n<pre><code>$ sudo pip2 install tensorflow-gpu\n<\/code><\/pre>\n\n<p>Or <\/p>\n\n<pre><code> $ sudo pip2 install tensorflow\n<\/code><\/pre>\n\n<p>For others.<\/p>\n"}]}
{"QuestionId":42236903,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-14T21:50:51.900","AcceptedAnswerId":42325368.0,"OwnerUserId":3681744.0,"Title":"Batch normalization initializer in TensorFlow","Body":"<p>In TensorFlow, batch normalization parameters include , , , and . However, for initializing these parameters there is only one argument in  called  which according to the documents it contains optional initializers for , ,  and .<\/p>\n\n<p>How can we use  to initialize these parameters?<\/p>\n","answers":[{"AnswerId":"42325368","CreationDate":"2017-02-19T09:02:16.230","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>Here is how you use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/layers\/batch_normalization\" rel=\"nofollow noreferrer\">batch normalization<\/a> with Tensorflow 1.0:<\/p>\n\n<pre><code>import tensorflow as tf\nbatch_normalization = tf.layers.batch_normalization\n\n... (define the network)\nnet = batch_normalization(net)\n... (define the network)\n<\/code><\/pre>\n\n<p>If you want to set parameters, just do it like this:<\/p>\n\n<pre><code>net = batch_normalization(net, \n                          beta_initializer=tf.zeros_initializer(), \n                          moving_variance_initializer=tf.ones_initializer())\n<\/code><\/pre>\n\n<h2>*args, **kwargs<\/h2>\n\n<p>This is the python way to pass arbitrary many non-keyword arguments <code>args<\/code> and arbitrary many keyword arguments <code>kwargs<\/code>. For example:<\/p>\n\n<pre><code>def test(*args, **kwargs):\n    print(\"#\" * 80)\n    print(args)\n    print(\"#\" * 80)\n    print(kwargs)\n\ntest(1, 2, 42, 3.141, 'foo', a=7, b=3, c='bla')\n<\/code><\/pre>\n\n<p>gives<\/p>\n\n<pre><code>################################################################################\n(1, 2, 42, 3.141, 'foo')\n################################################################################\n{'a': 7, 'c': 'bla', 'b': 3}\n<\/code><\/pre>\n"}]}
{"QuestionId":42237575,"AnswerCount":1,"Tags":"<python><tensorflow><unique>","CreationDate":"2017-02-14T22:41:09.037","AcceptedAnswerId":42245425.0,"OwnerUserId":2421391.0,"Title":"Find unique pairs of values in Tensorflow","Body":"<p>In Python 3.X + TensorFlow, if I have two TF vectors, point_x and point_y (same shape) that represent the X and Y coordinates of some number of pointers, how do I find all unique points?<\/p>\n\n<p>I was able to hack this together in Theano using a complex vector, with X in the real and Y in the imaginary portion:<\/p>\n\n<pre><\/pre>\n\n<p>The TF equivalent I'm trying is:<\/p>\n\n<pre><\/pre>\n\n<p>TensorFlow fails with something like:<\/p>\n\n<pre><\/pre>\n\n<p>Clearly, no one's implemented\/registered a complex version of the \"unique\" op. Any idea how to accomplish this task?<\/p>\n","answers":[{"AnswerId":"42245425","CreationDate":"2017-02-15T09:30:34.853","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>Well, here's an even hacker solution: use bit-level cast.<\/p>\n\n<p>If you tensor are all of type tf.float32, you can use:<\/p>\n\n<pre><code>xy = tf.transpose(tf.pack([point_x, point_y]))\nxy64 = tf.bitcast(xy, type=tf.float64)\nunique64, idx = tf.unique(xy64)\nunique_points = tf.bitcast(unique64, type=tf.float32)\n<\/code><\/pre>\n\n<p>The principle behind this is to put x and y coordinates together and let TensorFlow treat an (x, y) pair as a longer float, then tf.unique works for this 1-D tensor. Finally, convert the longer float to two genuine floats, as we desired.<\/p>\n\n<p>Note: <strong>This method is really hacky, and you have a risk of suffering from Nan or infinity or some strange values. But the chance is really slim.<\/strong><\/p>\n\n<p>Another possible work around is, if your data type is integer, you can pack two integers into one, like what a compiler does when it convert 2-d indices into 1-d ones. Say, if x = [1, 2, 3, 2], y = [0, 1, 0, 1], you can compress x and y into one tensor by x*10+y (10 is a large enough number. Any value larger than max(y) should work), then find unique values in this compressed array.<\/p>\n\n<p>Lastly, if you don't have any reason to do this inside TensorFlow, it might be better to do it outside, say, in numpy. You can evaluate the tensors, and remove duplicate values in numpy, then use these numpy arrays to generate new tensors and feed to the rest of your network.<\/p>\n"}]}
{"QuestionId":42238711,"AnswerCount":0,"Tags":"<tensorflow><speech-recognition><lstm><recurrent-neural-network>","CreationDate":"2017-02-15T00:26:05.710","AcceptedAnswerId":null,"OwnerUserId":7537870.0,"Title":"Best optimizer to minimize Connectionist Temporal Classification (CTC) loss in Tensorflow","Body":"<p>I am currently trying to minimize the ctc loss from a bidirectional RNN(LSTMCell), for speech recognition. I am using the MomentumOptimizer to do this as it was used in an example I referred, but neither do I have theoretical backing on why I'm using it nor do I know if this is the best optimizer to use to update the weights. A sample code of how I am computing the loss and optimizer is given below.<\/p>\n\n<pre><\/pre>\n\n<p>Please let me know if there is a better optimizer I can use.  <\/p>\n","answers":[]}
{"QuestionId":42239431,"AnswerCount":1,"Tags":"<c++><computer-vision><deep-learning><caffe><opencv3.0>","CreationDate":"2017-02-15T01:56:16.280","AcceptedAnswerId":null,"OwnerUserId":6675900.0,"Title":"How to read \"mean_file_proto\" into OpenCV Mat in order to subtract image from mean to use in Caffe Model?","Body":"<p>I am using OpenCV 3.2.<\/p>\n\n<p>I downloaded a Caffe model containing a mean file . I want to subtract from my input image the mean.\nHow to read  into OpenCV Mat in order to subtract image from mean to use in Caffe Model?<\/p>\n\n<p><strong>Edit<\/strong>: <\/p>\n\n<p>Can I use just OpenCV to do it without using Caffe? I am using OpenCV 3.2 in Visual Studio 2015 on Windows 10 64bit. I get an image from the webcam and I want to subtract the mean then, using , get the result. <\/p>\n","answers":[{"AnswerId":"42240073","CreationDate":"2017-02-15T03:17:50.750","ParentId":null,"OwnerUserId":"5899976","Title":null,"Body":"<p>The mean file is planar, probably either 1 or 3 channels.  The scale is [0, 255].<\/p>\n\n<pre><code>caffe::BlobProto mean_proto;\ncaffe::ReadProtoFromBinaryFile( filename, &amp;mean_proto ); \/\/ check return code\ncaffe::Blob&lt; float &gt; mean_blob;\nmean_blob.FromProto( mean_proto );\nfloat *data = mean_blob.mutable_cpu_data();\n<\/code><\/pre>\n\n<p>You can now create a <code>cv::Mat<\/code> of type <code>CV_32FC1<\/code>, for each channel.<\/p>\n\n<pre><code>cv::Mat channel( mean_blob.height(), mean_blob.width(), CV_32FC1, data );\ndata += mean_blob.height() * mean_blob.width();\n<\/code><\/pre>\n\n<p>Finally, populate a vector with these channels, and use <code>cv::merge()<\/code> to combine them into a single <code>cv::Mat<\/code> that you can subtract from another <code>cv::Mat<\/code> containing your input image.  If your image is already normalized [0, 1), you'll also need to rescale the mean.<\/p>\n"}]}
{"QuestionId":42239980,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-15T03:05:40.040","AcceptedAnswerId":42240039.0,"OwnerUserId":5171213.0,"Title":"Keras: multiclass classification","Body":"<p>I am following the examples from <a href=\"http:\/\/machinelearningmastery.com\/tutorial-first-neural-network-python-keras\/\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"http:\/\/machinelearningmastery.com\/multi-class-classification-tutorial-keras-deep-learning-library\/\" rel=\"nofollow noreferrer\">here<\/a>, and am totally new to Keras. It looks amazing - but I'm running into something I don't understand.<\/p>\n\n<p>I have an 8-class classification problem. My training set has 5120 rows and 62 columns, the last column being the target variable. <\/p>\n\n<p>My target variable is currently encoded as floats, so I convert them to integer and then to a dummy matrix for the model using to_categorical. The result is a numpy.ndarray of shape (num_samples, num_classes+1). Anyone know why?<\/p>\n\n<p>Here's the code:<\/p>\n\n<pre><\/pre>\n\n<p>EDIT<\/p>\n\n<p>The reason I didn't understand the answer given was I didn't understand that Keras was literally interpreting class labels as numbers. For example, since my classes are labelled 1 through 8, Keras look at the label '1' and says 'that's a 1 - I'll put it in the '1' position in the one-hot vector, like this: 0 1 0 0 0 0 0 0 0. It does the same with '2': 0 0 1 0 0 0 0 0 0, up to 8. That's why there's an extra column: to deal with the '0th' case, which doesn't exist in the mapping. Technically the accepted answer explained that, this just gives more detail.<\/p>\n","answers":[{"AnswerId":"42240039","CreationDate":"2017-02-15T03:12:16.280","ParentId":null,"OwnerUserId":"1862500","Title":null,"Body":"<p>Because <code>to_categorical<\/code> converts class vector (integers <strong>from 0<\/strong> to nb_classes) to binary class matrix, for use with categorical_crossentropy as documented in <a href=\"https:\/\/keras.io\/utils\/np_utils\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":42240489,"AnswerCount":1,"Tags":"<neural-network><keras><conv-neural-network><keras-layer>","CreationDate":"2017-02-15T04:00:27.030","AcceptedAnswerId":42591211.0,"OwnerUserId":5051564.0,"Title":"In what ways are the output of neural network layers useful?","Body":"<p>I'm currently working with keras and want to visualize the output of each layer. When having a visualisation of a layer of a neural networks output, like the example below, which is for MNIST handwriting number recognition.<\/p>\n\n<ol>\n<li>What information or insight does a researcher gain from these images<\/li>\n<li>How are these images interpreted<\/li>\n<li>If you would choose to see the output of a layer, what are your criteria for selection?<\/li>\n<\/ol>\n\n<p>Any comment or suggestion is greatly appreciated. Thank you.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/p01cs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/p01cs.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42591211","CreationDate":"2017-03-04T01:59:01.927","ParentId":null,"OwnerUserId":"3618272","Title":null,"Body":"<p><strong>Preface:<\/strong> A convolutional network is a collection of filters applied to sections of an image (strides, which are seen in the gif). They produce true\/false labels for if a given sub-section of the image matches the filter.<\/p>\n\n<p><a href=\"https:\/\/cs231n.github.io\/assets\/cnnvis\/filt1.jpeg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/ujwlkarn.files.wordpress.com\/2016\/08\/giphy.gif?w=748\" alt=\"Striding CNN\"><\/a><\/p>\n\n<p>What you're seeing in the images you provide is not the best representation of how these visualizations work in my opinion, as they visualise how the CNN percieves the whole image, through each neuron. This means all of them look very similar.<\/p>\n\n<p>Here is a better representation of how the basic filters of a network might look like. Some of them will trigger on straight lines, others will trigger on horizontal lines. That is also what the image you linked shows, except it does so for the whole image, on a visually simple object, which makes it a bit more difficult to understand.\nWhen you get to more complex filters which build on top of these basic filters, you might be better off visualizing the entire image. <\/p>\n\n<p><a href=\"https:\/\/ujwlkarn.files.wordpress.com\/2016\/08\/giphy.gif?w=748\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/cs231n.github.io\/assets\/cnnvis\/filt1.jpeg\" alt=\"Simple convolutional filters\"><\/a><\/p>\n\n<p>There is also a concept called transfer learning, where you take existing generalized models that are highly regarded, and try to apply these to your specific problem. These models often need to be tuned, which might mean removing some layers that are not needed (as each layer we keep means it's usually more time-consuming to train), and\/or adding more layers.<\/p>\n\n<p>A researcher will better be able to interpret how each layer in the network builds on the previous layers, and how they contribute to solving the problem at hand. This is often based on gut-feeling (Which can be simplified by good visualizations such as this <a href=\"https:\/\/youtu.be\/AgkfIQ4IGaM?t=30\" rel=\"nofollow noreferrer\">deep visualization toolbox video<\/a>)<\/p>\n\n<p>As an example, let's say I'm using VGG16, which is the name of a general model trained on <a href=\"http:\/\/image-net.org\/\" rel=\"nofollow noreferrer\">image-net<\/a>. I want to change it to classify distinct categories of furniture, instead of the 1000 classes from of completely different things it was originally intended to classify. Because it is such a general model, it can recognize a lot of different things, from humans to animals, to cars, to furniture. But a lot of these things don't make sense for me to incur a performance-penalty for, since they don't really help me classifying my furniture.<\/p>\n\n<p>Since a lot of the most important discoveries we make about these classes happen at different layers in the network, I can then move back up the convolutional layers, and remove everything which seem to be too complex for the task I'm doing. This might mean I remove some layers that seem to have specialized in categorising human features such as ears, mouths, eyes and faces.<\/p>\n\n<p>As far as I know, people visualize as many layers as they find useful, and then usually make a judgement call based on instinct as to which layers to keep or throw away after that.<\/p>\n\n<p>Images borrowed from:<\/p>\n\n<p><a href=\"http:\/\/cs231n.github.io\/understanding-cnn\/\" rel=\"nofollow noreferrer\">Visualizing what ConvNets learn<\/a><\/p>\n\n<p><a href=\"https:\/\/ujjwalkarn.me\/2016\/08\/11\/intuitive-explanation-convnets\/\" rel=\"nofollow noreferrer\">An Intuitive Explanation of Convolutional Neural Networks<\/a><\/p>\n"}]}
{"QuestionId":42240696,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-02-15T04:20:14.810","AcceptedAnswerId":56105187.0,"OwnerUserId":7549930.0,"Title":"How could I use batch normalization with tensorflow slim?","Body":"<p>I've found a few questions on using batch normalization in TensorFlow, but none of them are about it's wrapper in slim.<\/p>\n\n<p>I'm trying to use batch normalization to train a MNIST digit classifier. While the training performance is high enough, the validation or test performance is poor.<\/p>\n\n<p>I built only one graph, and passed  as a tf.placeholder, just like this (BN is used in every conv and fc layer):<\/p>\n\n<pre><\/pre>\n\n<p>I also added control dependencies as following:<\/p>\n\n<pre><\/pre>\n\n<p>For the training phase, I use:<\/p>\n\n<pre><\/pre>\n\n<p>For the validation phase, I use:<\/p>\n\n<pre><\/pre>\n\n<p>For test purpose, I dump the trained model to a checkpoint, then pass  as False. Again, it's performance is poor.<\/p>\n\n<p>So what's wrong with it? Is it about the  parameter? Or I need to maintain  and  variables in BN layer myself?<\/p>\n\n<p>For ease of reproducity, this is my code(set  to  to train a model and validate,  to restore from checkpoint and test):\n<a href=\"https:\/\/github.com\/soloice\/mnist-bn\/blob\/master\/mnist_bn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/soloice\/mnist-bn\/blob\/master\/mnist_bn.py<\/a><\/p>\n","answers":[{"AnswerId":"56105187","CreationDate":"2019-05-13T02:42:55.860","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>Finally I figured out the problem, see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1122#issuecomment-280325584\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1122#issuecomment-280325584<\/a>\n for details. Roughly speaking, one need to use <code>slim.learning.create_train_op<\/code> to create train op, and must be patient to wait for moving mean\/variance parameters to warm up. <\/p>\n"}]}
{"QuestionId":42241001,"AnswerCount":0,"Tags":"<keras><keras-layer>","CreationDate":"2017-02-15T04:53:39.013","AcceptedAnswerId":null,"OwnerUserId":2692779.0,"Title":"How do I train a Gater using custom inputs?","Body":"<p>I am trying to use keras to create a mixture of SVMs based on the paper Bengio et all. <\/p>\n\n<p>I have trained two svm classifiers with their outputs as svm1, svm2. I have written a pseudo code below.<\/p>\n\n<pre><\/pre>\n\n<p>My question<\/p>\n\n<ul>\n<li>How do I add custom inputs to my model. (in this case y_hats from svm experts)?<\/li>\n<\/ul>\n","answers":[]}
{"QuestionId":42241038,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><text-recognition>","CreationDate":"2017-02-15T04:56:59.040","AcceptedAnswerId":42244824.0,"OwnerUserId":6039807.0,"Title":"TensorFlow - Text recognition in image","Body":"<p>I am new to TensorFlow and to Deep Learning.\nI am trying to recognize text in naturel scene images. I used to work with an OCR but I would like to use Deep Learning. The text has always the same format :\n. <\/p>\n\n<p>What I have done is recognize every character\/digit. It means that I cropped the image around every character (so each picture gives me 10 characters) to build my training and test set and they build a two conv neural networks. So my training set was a set of characters pictures and the labels were just characters\/digits.<\/p>\n\n<p>But I want to go further. What I would like to do is just to give the full pictures and output the entire text (not one character such as in my previous model).<\/p>\n\n<p>Thank you in advance for any help.<\/p>\n","answers":[{"AnswerId":"42244824","CreationDate":"2017-02-15T09:02:47.453","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>The difficulty is that you don't know where the text is. The solution is, given an image, you need to use a sliding window to crop different part of the image, then use a classifier to decide if there are texts in the cropped area. If so, use your character\/digit recognizer to tell which characters\/digits they really are.<\/p>\n\n<p>So you need to train another classifer: given a cropped image (the size of cropped images should be slightly larger than that of your text area), decide if there are texts inside.<\/p>\n\n<p>Just construct training set (positive samples are text areas, negative samples are other areas randomly cropped from the big images) and train it~<\/p>\n"}]}
{"QuestionId":42241982,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-02-15T06:17:12.003","AcceptedAnswerId":42244320.0,"OwnerUserId":7124519.0,"Title":"Tensorflow epoch_loss becomes zero on the second epoch","Body":"<p>I am using python 2.7 and I really don't understand why I am getting this.\nI am guessing the problem is my python 2.7 is causing a float type problem.<\/p>\n\n<pre><\/pre>\n\n<p>My code is like the following:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42244320","CreationDate":"2017-02-15T08:37:37.510","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Maybe you fit the data nearly perfect and the rounding makes it go to 0.0. You could try to use your loss function on the test set to see the output or print intermediate outputs of every iteration to see if the loss is decreasing as it should.<\/p>\n\n<p>I would not worry about this though, because your test accuracy is 100% and therefore optimal. What more would you want?<\/p>\n"}]}
{"QuestionId":42242700,"AnswerCount":1,"Tags":"<python><crop><caffe><hdf5><mirror>","CreationDate":"2017-02-15T07:06:01.047","AcceptedAnswerId":null,"OwnerUserId":7349973.0,"Title":"How to read hdf5 with python layer in caffe and do data augmentation?","Body":"<p>I want to read data from  file and do data augmentation.<\/p>\n\n<p>I can do data augmentation in  with mirror and crop.<\/p>\n\n<p>However,I cannot have  in a  layer - caffe does not support this.<\/p>\n\n<p>I consider it might be a good way to write a python layer to solve it. Read data from hdf5 file and do data augmentation at the same time.<\/p>\n\n<p>I have read some relevant answers such as \"<a href=\"https:\/\/stackoverflow.com\/questions\/34996075\/caffe-data-layer-example-step-by-step\">caffe data layer example step by step<\/a>\" and \"<a href=\"https:\/\/stackoverflow.com\/questions\/41344168\/what-is-a-python-layer-in-caffe\">What is a Python layer in caffe?<\/a>\"<\/p>\n\n<p>I still don't know some details such as where should I put the written file? And how to write the codes to do data augmentation?<\/p>\n","answers":[{"AnswerId":"42249614","CreationDate":"2017-02-15T12:35:42.343","ParentId":null,"OwnerUserId":"5009875","Title":null,"Body":"<p>You python layer <code>*.py<\/code> should be in your <code>$PYTHONPATH<\/code>.<\/p>\n\n<p>Consider examples of python layers in <code>$CAFFE_ROOT\/examples\/pycaffe\/layers<\/code>\n,for custom python loss and data layer.<\/p>\n\n<p>Consider particularly in <code>$CAFFE_ROOT\/examples\/pycaffe\/layers\/pascal_multilabel_datalayers.py<\/code><\/p>\n\n<p>Before next batch gets loaded, you can do the data augmentation on the fly.<\/p>\n\n<p>Then in <code>*.prototxt<\/code><\/p>\n\n<pre><code>layer {\n  name: \"data\"\n  type: \"Python\"\n  top: \"data\"\n  top: \"label\"\n  python_param {\n    module: \"pascalcontext_layers\"\n    layer: \"PASCALContextSegDataLayer\"\n    param_str: \"{\\'context_dir\\': \\'..\/..\/data\/pascal-context\\', \\'seed\\': 1337, \\'split\\': \\'train\\', \\'voc_dir\\': \\'..\/..\/data\/pascal\\'}\"\n  }\n}\n<\/code><\/pre>\n\n<p><code>module<\/code> is the file name and <code>layer<\/code> is the class name.<\/p>\n"}]}
{"QuestionId":42243323,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-02-15T07:41:27.787","AcceptedAnswerId":42359135.0,"OwnerUserId":6211276.0,"Title":"Using Keras with Tensorflow as backend to train cifar10 using vgg16.py from keras","Body":"<p>I used a  provided by keras. <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/applications\/vgg16.py\" rel=\"nofollow noreferrer\">vgg16.py<\/a> <br>\nIn vgg16.py, I changed the min input size from 48 to 32 and default from 225 to 32. The dimensions of cifar10 is (nb_samples, 3, 32, 32).<br><\/p>\n\n<p>The following is the code :<\/p>\n\n<pre><\/pre>\n\n<p>The above code works but the weights are initialized randomly.The following is the result : <br><\/p>\n\n<pre><\/pre>\n\n<p>This training seems to .<br>\nThe code for <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py\" rel=\"nofollow noreferrer\">cifar10_cnn.py<\/a> by keras uses real-time data augmentation and speed  Any reasons and also accuracy rises to 80% on training set for later but in above case, it is constant at 9%?<\/p>\n","answers":[{"AnswerId":"42359135","CreationDate":"2017-02-21T05:07:13.970","ParentId":null,"OwnerUserId":"6211276","Title":null,"Body":"<p>What I found after trying different configurations is that <code>VGG16 architecture is too big<\/code> for an image of size 32x32.I tried to use VGG16 until <code>block3_pool<\/code> and then added a <code>dense 512<\/code> fully_connected followed by <code>softmax classifier<\/code> for<code>10 classes<\/code>. Below is modified code:<br><\/p>\n\n<pre><code>base_model = VGG16(weights=None, include_top=False, \n             input_shape=X_train.shape[1:], classes=10)\nx = base_model.get_layer('block3_pool').output\nx = Flatten(name='Flatten')(x)\nx = Dense(512, activation='relu', name='fc1')(x)\npredictions = Dense(nb_classes, activation='softmax')(x)\nmodel = Model(input=base_model.input, output=predictions)\n<\/code><\/pre>\n\n<p>The drawbacks I found in the question model were: <br>(i) Reduce the batch size from 256 to 32.<br>\n(ii) Normalization of data using mean.<br>\n(iii) Data Augmentation should be used using <code>data_gen<\/code> keras function.<br>\n(iv) Smaller architecture VGG16 (16 layer) to  <code>block3_pool<\/code> (3 layer).<br>\n(v) Reduction in time from 2000s to 900s using above architecture.<br><\/p>\n\n<p>The <code>accuracy<\/code> achieved was around <code>75%<\/code>. Is it possible to <code>further reduce time on CPU<\/code> ?<br>\nAfter evaluating the model on test examples of size 10000, I get the following result: <br>\n<code>['loss', 'acc'] : [2.3033507381439211, 0.10000000000000001]<\/code>. <br>\nThis is <code>overfitting<\/code>.<\/p>\n"},{"AnswerId":"44776461","CreationDate":"2017-06-27T09:16:41.630","ParentId":null,"OwnerUserId":"4391129","Title":null,"Body":"<p>I found <a href=\"https:\/\/github.com\/rajatvikramsingh\/cifar10-vgg16\" rel=\"nofollow noreferrer\">this github project<\/a>. Here they are using vgg16 on cifar10 with keras. They reach a validation accuracy of over 0.97.<\/p>\n\n<p><strong>More precisely:<\/strong><\/p>\n\n<ul>\n<li>they use the vgg16 up to the layer <code>'block3_pool'<\/code><\/li>\n<li>than they adds some fully connected layers<\/li>\n<\/ul>\n\n<p>code for the fully connected layers:<\/p>\n\n<pre><code>x = Flatten()(last)\nx = Dense(256, activation='relu')(x)\nx = Dropout(0.5)(x)\npred = Dense(10, activation='sigmoid')(x)\n<\/code><\/pre>\n"}]}
{"QuestionId":42243895,"AnswerCount":1,"Tags":"<python><tensorflow><theano>","CreationDate":"2017-02-15T08:14:38.573","AcceptedAnswerId":42246938.0,"OwnerUserId":4393898.0,"Title":"printing theano.tensor without having any idea about the shape and type of the tensor","Body":"<p>I am debugging a code which I haven't written. I want to print out  variable which is calculated as following:<\/p>\n\n<pre><\/pre>\n\n<p>When I use , I get <\/p>\n\n<blockquote>\n  <p>MissingInputError: (\"An input of the graph, used to compute\n  Reshape{1}(y_sampler, TensorConstant{(1,) of -1}), was not provided\n  and not given a value.Use the Theano flag\n  exception_verbosity='high',for more information on this error.\",\n  y_sampler)<\/p>\n<\/blockquote>\n\n<p>error. How can I print this damn \"state_below_\" value?<\/p>\n\n<p>Thanks,<\/p>\n","answers":[{"AnswerId":"42246938","CreationDate":"2017-02-15T10:36:53.027","ParentId":null,"OwnerUserId":"3212998","Title":null,"Body":"<p>When you are using purely symbolic variable, you have to compile a function with respective inputs and outputs. Then you have to make a call to your function with constant values for respective symbolic variables. \nElse, if you are using a shared variables, the way you have written will work. \nThe documentation to <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/compile\/function.html\" rel=\"nofollow noreferrer\">function<\/a> , <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#converting-from-python-objects\" rel=\"nofollow noreferrer\">shared variable<\/a> <\/p>\n"}]}
{"QuestionId":42244002,"AnswerCount":0,"Tags":"<compilation><tensorflow><bazel>","CreationDate":"2017-02-15T08:21:27.410","AcceptedAnswerId":null,"OwnerUserId":829361.0,"Title":"Bazel fetch issue while compiling Tensorflow","Body":"<p>I'm compiling the latest Tensorflow from source. After the configure step I'm facing the following error:<\/p>\n\n<blockquote>\n  <p>ERROR:\n  com.google.devtools.build.lib.packages.BuildFileContainsErrorsException:\n  error loading package '': Encountered error while reading extension\n  file 'closure\/defs.bzl': no such package\n  '@io_bazel_rules_closure\/\/closure': Error downloading\n  [<a href=\"http:\/\/bazel-mirror.storage.googleapis.com\/github.com\/bazelbuild\/rules_closure\/archive\/5ca1dab6df9ad02050f7ba4e816407f88690cf7d.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/bazel-mirror.storage.googleapis.com\/github.com\/bazelbuild\/rules_closure\/archive\/5ca1dab6df9ad02050f7ba4e816407f88690cf7d.tar.gz<\/a>,\n  <a href=\"https:\/\/github.com\/bazelbuild\/rules_closure\/archive\/5ca1dab6df9ad02050f7ba4e816407f88690cf7d.tar.gz]\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bazelbuild\/rules_closure\/archive\/5ca1dab6df9ad02050f7ba4e816407f88690cf7d.tar.gz]<\/a>\n  to\n  \/home\/xyzuser\/.cache\/bazel\/_bazel_xyzuser\/cb1e63cb5e61cab49a9fd2f5ba92d003\/external\/io_bazel_rules_closure\/5ca1dab6df9ad02050f7ba4e816407f88690cf7d.tar.gz:\n  All mirrors are down: [java.lang.NullPointerException].<\/p>\n<\/blockquote>\n\n<p>This stems from the bazel_clean_and_fetch() in configure file:<\/p>\n\n<pre><\/pre>\n\n<p>Same error replicates if I simply do <\/p>\n\n<p>My system is behind a proxy and I've set HTTP_PROXY and HTTPS_PROXY in the env.<\/p>\n\n<p><strong>bazel version:<\/strong>\nBuild label: 0.4.4- (@non-git)\nBuild target: bazel-out\/local-opt\/bin\/src\/main\/java\/com\/google\/devtools\/build\/lib\/bazel\/BazelServer_deploy.jar<\/p>\n\n<p><strong>Operating System:<\/strong>\nUbuntu 14.04.5 LTS<\/p>\n\n<p>Any insights would be helpful as I'm new to Bazel and Tensorflow.<\/p>\n","answers":[]}
{"QuestionId":42244198,"AnswerCount":16,"Tags":"<python><module><installation><tensorflow><pip>","CreationDate":"2017-02-15T08:30:59.683","AcceptedAnswerId":null,"OwnerUserId":5147241.0,"Title":"ImportError: No module named tensorflow","Body":"<p>Please help me with this error<\/p>\n\n<p>I have installed the tensorflow module on my server and below is it's information<\/p>\n\n<pre><\/pre>\n\n<p>But when I try to import tensorflow I get following error<\/p>\n\n<pre><\/pre>\n\n<p>My python version is as following<\/p>\n\n<pre><\/pre>\n\n<p>I have tried the solutions given in \n<a href=\"https:\/\/stackoverflow.com\/questions\/40472144\/importerror-no-module-named-tensorflow-cant-install-tensorflow\">sol1<\/a><\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/41658793\/tensorflow-installation-issuesimporterror-no-module-named-tensorflow\">Sol2<\/a><\/p>\n\n<p>I do not have sudo access to the server \nI can only use pip to install any module<\/p>\n","answers":[{"AnswerId":"53130962","CreationDate":"2018-11-03T11:40:38.617","ParentId":null,"OwnerUserId":"5235611","Title":null,"Body":"<p><strong>with python2<\/strong><\/p>\n\n<p><code>pip show tensorflow<\/code>  to check install<\/p>\n\n<p><code>python test.py<\/code>  to run test<\/p>\n\n<p><strong>with python3<\/strong><\/p>\n\n<p><code>pip3 show tensorflow<\/code> to check install<\/p>\n\n<p><code>python3 test.py<\/code> to run test<\/p>\n\n<p><strong>test.py<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nc = np.array([[3.,4], [5.,6], [6.,7]])\nstep = tf.reduce_mean(c, 1)                                                                                 \nwith tf.Session() as sess:\n    print(sess.run(step))\n<\/code><\/pre>\n\n<p>Or, if you haven't install tensorflow yet, try the <a href=\"https:\/\/www.tensorflow.org\/install\/#anaconda_installation\" rel=\"nofollow noreferrer\">offical document<\/a><\/p>\n"},{"AnswerId":"54222135","CreationDate":"2019-01-16T17:14:44.933","ParentId":null,"OwnerUserId":"9930898","Title":null,"Body":"<p>you might wanna try this:<\/p>\n\n<pre><code>$conda install -c conda-forge tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"44969816","CreationDate":"2017-07-07T11:36:38.217","ParentId":null,"OwnerUserId":"4414341","Title":null,"Body":"<p>Check if Tensorflow was installed successfully using:<\/p>\n\n<pre><code> pip3 show tensorflow\n<\/code><\/pre>\n\n<p>If you get something like<\/p>\n\n<pre><code>Name: tensorflow\nVersion: 1.2.1\nSummary: TensorFlow helps the tensors flow\nHome-page: http:\/\/tensorflow.org\/\nAuthor: Google Inc.\nAuthor-email: opensource@google.com\nLicense: Apache 2.0\nLocation: \/usr\/local\/lib\/python3.5\/dist-packages\nRequires: bleach, markdown, html5lib, backports.weakref, werkzeug, numpy, protobuf, wheel, six\n<\/code><\/pre>\n\n<p>You may try adding the path of your tensorflow location by:<\/p>\n\n<pre><code>export PYTHONPATH=\/your\/tensorflow\/path:$PYTHONPATH.\n<\/code><\/pre>\n"},{"AnswerId":"56775862","CreationDate":"2019-06-26T15:03:59.183","ParentId":null,"OwnerUserId":"1683053","Title":null,"Body":"<p>I had exactly the same problem.\nIt's because default python is in version 2\nYou need to link to the version 3.<\/p>\n\n<pre><code>&gt;sudo rm -rf \/usr\/bin\/python\n&gt;sudo ln -s \/usr\/bin\/python3.5 \/usr\/bin\/python\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LEk0t.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LEk0t.png\" alt=\"python links\"><\/a><\/p>\n"},{"AnswerId":"52502491","CreationDate":"2018-09-25T16:08:42.423","ParentId":null,"OwnerUserId":"4719725","Title":null,"Body":"<p>You may need this since first one may not work.<\/p>\n\n<p><code>python3 -m pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.12.0-py3-none-any.whl\n<\/code><\/p>\n"},{"AnswerId":"42244427","CreationDate":"2017-02-15T08:43:42.243","ParentId":null,"OwnerUserId":"3972930","Title":null,"Body":"<p>Try <a href=\"https:\/\/www.tensorflow.org\/get_started\/os_setup#anaconda_installation\" rel=\"nofollow noreferrer\">Anaconda install steps<\/a> from TensorFlow docs.<\/p>\n"},{"AnswerId":"58111959","CreationDate":"2019-09-26T07:48:35.540","ParentId":null,"OwnerUserId":"4573839","Title":null,"Body":"<p>In my case, I install <code>32 Bit Python<\/code> so I cannot install Tensorflow, After uninstall <code>32 Bit Python<\/code> and install <code>64 Bit Python<\/code>, I can install tensorflow successfully.<\/p>\n\n<p>After reinstall <code>Python 64 bit<\/code>, you need to check your python install folder path is properly set in Windows Environment Path.<\/p>\n\n<p>You can check Python version by typing <code>python<\/code> in cmd.<\/p>\n"},{"AnswerId":"57729690","CreationDate":"2019-08-30T15:38:40.253","ParentId":null,"OwnerUserId":"6114891","Title":null,"Body":"<p>I ran into the same issue. I simply updated my command to begin with <code>python3<\/code> instead of <code>python<\/code> and it worked perfectly.<\/p>\n"},{"AnswerId":"46107897","CreationDate":"2017-09-08T02:51:27.657","ParentId":null,"OwnerUserId":"1141159","Title":null,"Body":"<p>For Anaconda3, simply install in Anaconda Navigator:\n<a href=\"https:\/\/i.stack.imgur.com\/1VQwu.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1VQwu.jpg\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"45561284","CreationDate":"2017-08-08T06:55:49.137","ParentId":null,"OwnerUserId":"5212057","Title":null,"Body":"<p>Try installing tensorflow in the user site - This installation only works for you.<\/p>\n\n<blockquote>\n  <p>pip install tensorflow --user<\/p>\n<\/blockquote>\n"},{"AnswerId":"47976692","CreationDate":"2017-12-26T09:50:43.450","ParentId":null,"OwnerUserId":"9141194","Title":null,"Body":"<p>Activate the virtualenv environment by issuing one of the following commands:<\/p>\n\n<p>$ source ~\/tensorflow\/bin\/activate # bash, sh, ksh, or zsh<br \/>\n$ source ~\/tensorflow\/bin\/activate.csh  # csh or tcsh<\/p>\n\n<p>Hope this help<\/p>\n"},{"AnswerId":"48021584","CreationDate":"2017-12-29T11:03:53.687","ParentId":null,"OwnerUserId":"9152792","Title":null,"Body":"<p>This Worked for me:<\/p>\n\n<pre><code>$ sudo easy_install pip\n$ sudo easy_install --upgrade six\n$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/mac\/tensorflow-0.9.0-py2-none-any.whl\n$ sudo pip install --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n"},{"AnswerId":"48241805","CreationDate":"2018-01-13T16:25:07.587","ParentId":null,"OwnerUserId":"4427457","Title":null,"Body":"<p>I had a more basic problem when I received this error.<\/p>\n\n<p>The \"Validate your installation\" instructions say to type: <strong>python<\/strong><\/p>\n\n<p>However, I have both 2.7 and 3.6 installed.  Because I used <strong>pip3<\/strong> to install tensorflow, I needed to type: <strong>python3<\/strong><\/p>\n\n<p>Using the correct version, I could import the \"tensorflow\" module.<\/p>\n"},{"AnswerId":"44479696","CreationDate":"2017-06-11T01:53:56.623","ParentId":null,"OwnerUserId":"3893220","Title":null,"Body":"<p>I was trying to install tensorflow GPU for a Windows 7 with pip3 for python3.5.x. \nInstead of doing <code>pip3 install --upgrade tensorflow<\/code> I just did <code>pip install tensorflow<\/code> and after it completed i was finally able to import tensorflow in python3.5.x.<\/p>\n"},{"AnswerId":"48799321","CreationDate":"2018-02-15T02:51:39.230","ParentId":null,"OwnerUserId":"5001404","Title":null,"Body":"<p>Instead of using the doc's command (<code>conda create -n tensorflow pip python=2.7 # or python=3.3, etc.<\/code>) which wanted to install python2.7 in the conda environment, and kept erroring out saying the module can't be found when following the installation validation steps, I used <code>conda create -n tensorflow pip python=3<\/code> to make sure python3 was installed in the environment. <\/p>\n\n<p>Doing this, I only had to type <code>python<\/code> instead of <code>python3<\/code> when validating the installation and the error went away.<\/p>\n"},{"AnswerId":"45551934","CreationDate":"2017-08-07T16:47:51.300","ParentId":null,"OwnerUserId":"2103732","Title":null,"Body":"<p>Try installing tensorflow again with the whatever version you want and with option --ignore-installed like:<\/p>\n\n<pre><code>pip install tensorflow==1.2.0 --ignore-installed\n<\/code><\/pre>\n\n<p>I solved same issue using this command.<\/p>\n"}]}
{"QuestionId":42244222,"AnswerCount":1,"Tags":"<tensorflow><convolution><deconvolution>","CreationDate":"2017-02-15T08:32:12.120","AcceptedAnswerId":null,"OwnerUserId":5016028.0,"Title":"Convolution-Deconvolution pair gives slightly different dimensionality","Body":"<p>I am using a convolution layer followed by a deconvolution layer like so:<\/p>\n\n<pre><\/pre>\n\n<p>My idea is to make the initial image smaller, then bring it to its original size with the deconvolution. I am actually using the tf.slim functions, but the arguments are the ones above. <\/p>\n\n<p>When I look at the input and output, I have a small difference:<\/p>\n\n<pre><\/pre>\n\n<p>I think it could be due to my stride size or kernel size. I've tried multiple values but none seem to reproduce the original dimensions. <\/p>\n","answers":[{"AnswerId":"43862040","CreationDate":"2017-05-09T05:50:28.617","ParentId":null,"OwnerUserId":"7981678","Title":null,"Body":"<p>A popular methodology is to pad the input image such that the output after the convolution and deconvolution would be of the same size as the padded input. Then crop the output to fit the original input without the padding.<\/p>\n"}]}
{"QuestionId":42244476,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><nao-robot>","CreationDate":"2017-02-15T08:46:09.780","AcceptedAnswerId":null,"OwnerUserId":7528619.0,"Title":"How can i install tensorflow framework on robot NAO?","Body":"<p>I want to install TensorFlow framework in robot NAO , \nHow to do it ? <\/p>\n","answers":[{"AnswerId":"42246710","CreationDate":"2017-02-15T10:27:27.973","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Great question! The NAO is a linux machine, so technically it might be possible. Unfortunately, the NAO also has a limited amount of computational power... Depending on the plan you have it might be a better idea to set up an external computer that does the heavy computations for you. This all depends on the application you want to build. <\/p>\n\n<p>If you decide to install tensorflow on Nao: simply try to use SSH (or Putty) to get a console you can use to install tensorflow. <\/p>\n\n<p>If you decide to use an external server: maybe this program I wrote a long time ago helps you: <a href=\"https:\/\/github.com\/rmeertens\/nao-wit\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rmeertens\/nao-wit<\/a> . It is an example of how to send speech to an external server. <\/p>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":42244666,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-02-15T08:55:33.250","AcceptedAnswerId":42257312.0,"OwnerUserId":6219234.0,"Title":"How to build static libraries and linkthem to a c++ project in linux platform?","Body":"<p>I am working on a text-classification project, which is big and doesn't use bazel as its build tool. I want to integrate tensorflow into my project, but I find it is hard to change my build tool to bazel. So I wish to build static libraries on tensorflow and link them into my project.\nDoes anyone know how to build standalone static libraries on tensorflow source and link them in the existing c++ project? Thanks a lot.<\/p>\n","answers":[{"AnswerId":"42257312","CreationDate":"2017-02-15T18:26:56.157","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>The TensorFlow repository has some Makefiles you can use to build a static library (see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/makefile\" rel=\"noreferrer\">tensorflow\/contrib\/Makefile<\/a>).<\/p>\n\n<p>Alternatively, you could use <code>bazel<\/code> to build the TensorFlow C++ shared library and then load and use the shared library in your application (<code>bazel build -c opt \/\/tensorflow:libtensorflow_cc.so<\/code>). Unfortunately, bazel can't yet produce a static library (<a href=\"https:\/\/github.com\/bazelbuild\/bazel\/issues\/1920\" rel=\"noreferrer\">#1920<\/a>).<\/p>\n\n<p>Hope that helps.<\/p>\n"}]}
{"QuestionId":42245001,"AnswerCount":1,"Tags":"<variables><tensorflow>","CreationDate":"2017-02-15T09:10:37.183","AcceptedAnswerId":42245377.0,"OwnerUserId":7474000.0,"Title":"Tensorflow: not shown tensor","Body":"<p>I have a question about not shown tensor.\nI coded as below, and I expected there are tensors having name as 'ttt', 'l1a' or 'l1_maxpool', but there aren't when seeing by tf.all_variables function.\nBut seeing by toggling a breakpoint at the position, they exist.<\/p>\n\n<p>Is there reason why they are not shown or another reason that I should modify the code?\nThanks in advance.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42245377","CreationDate":"2017-02-15T09:28:09.463","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>There is a difference between variables and operations. Convolution, relu and max_pool are operations you perform on variables. If you visualise your graph (<a href=\"https:\/\/www.tensorflow.org\/how_tos\/graph_viz\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/how_tos\/graph_viz\/<\/a>) you can see the operations you added are indeed in there!<\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":42245237,"AnswerCount":0,"Tags":"<model><tensorflow>","CreationDate":"2017-02-15T09:22:17.047","AcceptedAnswerId":null,"OwnerUserId":4362500.0,"Title":"Cannot achieve the reported accuracy with tensorflow wide_n_deep_tutorial?","Body":"<p>I'm learning the google's wide&amp;deep model these days with its tensorflow tutorial. After running the code in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/wide_n_deep_tutorial.py\" rel=\"nofollow noreferrer\">\"wide_n_deep_tutorial.py\"<\/a>!, I achieved accuracy 0.834 for wide-only model, but the achieved accuracy for wide-and-deep model was only about 0.82.<\/p>\n\n<p>I didn't change anything in the source code and just couldn't achieve the accuracy 0.844 as reported on its website. <\/p>\n\n<p>Has anyone met the same problem with me? And does anyone know how to fix it?<\/p>\n","answers":[]}
{"QuestionId":42247905,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-15T11:20:50.270","AcceptedAnswerId":null,"OwnerUserId":6434823.0,"Title":"How to train a set of images (apart form the tutorial) and use it to classify in Tensorflow?","Body":"<p>I am using from the TensorPy GitHub repo (built by Michael Mintz) (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/imagenet\/classify_image.py)to\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/models\/image\/imagenet\/classify_image.py)to<\/a> easily handle image classifications of either individual or multiple images directly from web pages. However i want to train the system with my own set of images and then run the classifier. Could anyone please point me in the right direction for training a new model and how do i use it in classifying a set of bulk images?<\/p>\n\n<p>Thank you.<\/p>\n","answers":[]}
{"QuestionId":42249808,"AnswerCount":0,"Tags":"<python><deep-learning><keras><keras-layer>","CreationDate":"2017-02-15T12:44:53.043","AcceptedAnswerId":null,"OwnerUserId":7568768.0,"Title":"Visualizing the input that maximizes the activation of a layer with dropout in Keras","Body":"<p>I'm trying to replicate the code in this blog article <a href=\"https:\/\/blog.keras.io\/how-convolutional-neural-networks-see-the-world.html\" rel=\"nofollow noreferrer\" title=\"How convolutional neural networks see the world\">How convolutional neural networks see the world<\/a><\/p>\n\n<p>It works well in a CNN where there's no dropout layer but when there's one (or more) dropout layers, I  can't directly use the  line because it expects a learning phase.<\/p>\n\n<p>When I use the recommend way to extract the output of a layer : <\/p>\n\n<pre><\/pre>\n\n<p>The problem is that I can't put a placeholder in  because it expects \"real\" data but if I put directly \"real\" data then the rest of the code doesn't work (creating the loss, gradients and iterating needs a placeholder). <\/p>\n\n<p>Is there a way I can make this work?<\/p>\n\n<p>I'm using the Tensorflow backend.<\/p>\n\n<p>EDIT : I solved my issue by using the  method before doing anything like building my model (I had to start from a new environment and I used the method right after the imports).<\/p>\n","answers":[]}
{"QuestionId":42250205,"AnswerCount":0,"Tags":"<python><ubuntu><deep-learning><caffe><hdf5>","CreationDate":"2017-02-15T13:03:54.953","AcceptedAnswerId":null,"OwnerUserId":4990057.0,"Title":"hdf5 -- ValueError: Not a property list class (Not a property list class)","Body":"<p>When I tried to train a network with caffe, I got the following error information:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the related part of the caffe model file:<\/p>\n\n<pre><\/pre>\n\n<p>To figure out the problem, I tried:<\/p>\n\n<pre><\/pre>\n\n<p>It succeeded.<\/p>\n\n<p>Then I tried:<\/p>\n\n<pre><\/pre>\n\n<p>It also executed.<\/p>\n\n<p>So I have been stuck...  Can anybody please help me?<\/p>\n","answers":[]}
{"QuestionId":42250340,"AnswerCount":1,"Tags":"<python><python-3.x><machine-learning><tensorflow><k-means>","CreationDate":"2017-02-15T13:10:21.530","AcceptedAnswerId":null,"OwnerUserId":5217799.0,"Title":"RuntimeError: No C++ shape function registered for standard op: NearestNeighbors","Body":"<p>Update:-\nPlease try this code:-<\/p>\n\n<pre><\/pre>\n\n<p>The error changes by changing the type of x in the line km.fit(x)\nfor example x.astype('float32')or x.astype('float64')<\/p>\n\n<p>I am trying to implement KMeansClustering using tensorflow.contrib.learn.python.learn.estimators.kmeans<\/p>\n\n<p>But I am getting the following error while using the code :-<\/p>\n\n<pre><\/pre>\n\n<p>\"No C++ shape function registered for standard op: %s\" % op.type)\nRuntimeError: No C++ shape function registered for standard op: NearestNeighbors\"<\/p>\n\n<p>X is a numpy array of shape (2000,1000)<\/p>\n\n<p>Stacktrace:-<\/p>\n\n<pre><\/pre>\n\n<p>I also tried it in a ubuntu virtual box but got the following error:-<\/p>\n\n<pre><\/pre>\n\n<p>Update :- \nAs per the suggestion given by @Changming Sun I tried to update the gen_clustering_ops.py, it solved the first problem but I generated a new error.<\/p>\n\n<pre><\/pre>\n\n<p>I am facing a similar issue with GMM in tensorflow<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/42299378\/valueerror-features-are-incompatible-with-given-information-given-features-te]\">ValueError: Features are incompatible with given information. Given features: Tensor(&quot;input:0&quot;, shape=(?, 198), dtype=float32)<\/a><\/p>\n","answers":[{"AnswerId":"42302028","CreationDate":"2017-02-17T15:47:41.507","ParentId":null,"OwnerUserId":"523503","Title":null,"Body":"<p>Open\n\"Lib\\site-packages\\tensorflow\\contrib\\factorization\\python\\ops\\gen_clustering_ops.py\"<\/p>\n\n<p>Add<\/p>\n\n<pre><code>ops.RegisterShape(\"NearestNeighbors\")(None)\n<\/code><\/pre>\n\n<p>For any error like this, fix it in this way.<\/p>\n"}]}
{"QuestionId":42252040,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-02-15T14:25:44.693","AcceptedAnswerId":42254393.0,"OwnerUserId":5522711.0,"Title":"How to translate(or shift) images in tensorflow","Body":"<p>I want to make my input image (tensor) of my model to shift up\/down or right\/left and then pad.<\/p>\n\n<p>For example, if the original image is 3x3 like below,<\/p>\n\n<pre><\/pre>\n\n<p>Then, if I shift to left,<\/p>\n\n<pre><\/pre>\n\n<p>I found that there is an image rotate function in Tensorflow but I couldn't find translate or shift.\nPlease let me know if there is a built-in function,\nor suggest the way to implement.<\/p>\n","answers":[{"AnswerId":"45662771","CreationDate":"2017-08-13T16:58:59.977","ParentId":null,"OwnerUserId":"1004331","Title":null,"Body":"<p>I wrote a function to do this based on tf.contrib.image.transform:\n<a href=\"https:\/\/gist.github.com\/astromme\/8116a154be8dae5528f33669e490c19a\" rel=\"noreferrer\">https:\/\/gist.github.com\/astromme\/8116a154be8dae5528f33669e490c19a<\/a><\/p>\n\n<pre><code>## Tensorflow image translation op\n# images:        A tensor of shape (num_images, num_rows, num_columns, num_channels) (NHWC),\n#                (num_rows, num_columns, num_channels) (HWC), or (num_rows, num_columns) (HW).\n# tx:            The translation in the x direction.\n# ty:            The translation in the y direction.\n# interpolation: If x or y are not integers, interpolation comes into play. Options are 'NEAREST' or 'BILINEAR'\ndef tf_image_translate(images, tx, ty, interpolation='NEAREST'):\n    # got these parameters from solving the equations for pixel translations\n    # on https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/image\/transform\n    transforms = [1, 0, -tx, 0, 1, -ty, 0, 0]\n    return tf.contrib.image.transform(images, transforms, interpolation)\n<\/code><\/pre>\n\n<p>Use it like this:<\/p>\n\n<pre><code>translation_op = tf_image_translate(images, tx=-5, ty=10)\n\nwith tf.Session() as sess:\n    translated_images = sess.run(translation_op)\n<\/code><\/pre>\n"},{"AnswerId":"49262473","CreationDate":"2018-03-13T17:36:21.943","ParentId":null,"OwnerUserId":"1599349","Title":null,"Body":"<p>There is now a function for this (at least, as of TF v1.6): \n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/image\/translate\" rel=\"nofollow noreferrer\">tf.contrib.image.translate<\/a><\/p>\n"},{"AnswerId":"42254393","CreationDate":"2017-02-15T16:05:41.790","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>I think you can combine <code>tf.image.crop_to_bounding_box<\/code> and <code>tf.image.pad_to_bounding_box<\/code> to achieve this. \nHere's the API: <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/image#Cropping\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/image#Cropping<\/a><\/p>\n"}]}
{"QuestionId":42253032,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-15T15:06:27.867","AcceptedAnswerId":42322918.0,"OwnerUserId":1498389.0,"Title":"AttributeError: module 'tensorflow.contrib.rnn' has no attribute 'stack_bidirectional_dynamic_rnn'","Body":"<p>I've just installed  (1.0) with pip on a new system (after installing cuda and cudnn, obviously).<\/p>\n\n<p>Unfortunately, my code (working well on cpu with tensorflow 0.12), now throws<\/p>\n\n<pre><\/pre>\n\n<p>When trying to call .<br>\nMoreover, the function is present in  and seems to be imported in .<\/p>\n\n<p>Any ideas ?<\/p>\n\n<p>P.S.: Downgrade to 0.12 () resolve my issue, but that a bit disappointing :\/<\/p>\n","answers":[{"AnswerId":"42322918","CreationDate":"2017-02-19T02:12:30.373","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>Oh, that's an error in the tensorflow code. Sorry about that. Please add an entry here listing <code>@@stack_bidirectional_rnn<\/code> <a href=\"https:\/\/www.github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/__init__.py#L49\" rel=\"nofollow noreferrer\">here<\/a> so that <code>stack_bidirectional_rnn<\/code> is exposed. Fix coming up. Feel free to open a github issue on our page and CC drpng there if you want to track progress.<\/p>\n"}]}
{"QuestionId":42253259,"AnswerCount":1,"Tags":"<neural-network><python-3.5><keras><floating-accuracy><loss>","CreationDate":"2017-02-15T15:16:17.570","AcceptedAnswerId":null,"OwnerUserId":7336641.0,"Title":"Accuracy goes down during epoch keras","Body":"<p>I've tried to write a neural network but the accuracy doesn't change each epoch. I'm using keras and I can watch the accuracy change as each epoch is evaluated per se and it will start low, go up a bit, then drop back down to the exact same value each time <a href=\"https:\/\/i.stack.imgur.com\/2ONj9.png\" rel=\"nofollow noreferrer\">example output<\/a>. I've tried changing the batch size, learning rates, changing the data around a bit, but every time it does the same thing, just perhaps with a different accuracy value. I've also tried different optimizers. Any help is appreciated. (Also I was able to get an mnist example working)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42278661","CreationDate":"2017-02-16T15:51:26.323","ParentId":null,"OwnerUserId":"7568768","Title":null,"Body":"<p>Since you only have a single neuron in your output layer, I assume you're doing regression and not classification.<\/p>\n\n<p>If that's the case, then you should change your loss function to <code>'mse'<\/code> and you also should remove the activation in your output layer because the sigmoid function will squash your output between 0 and 1.<\/p>\n"}]}
{"QuestionId":42253585,"AnswerCount":1,"Tags":"<python><keras><chmod>","CreationDate":"2017-02-15T15:32:14.583","AcceptedAnswerId":42253862.0,"OwnerUserId":5171213.0,"Title":"Keras: chmod results in not finding keras module","Body":"<p>I'm running code from <a href=\"http:\/\/machinelearningmastery.com\/tutorial-first-neural-network-python-keras\/\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n\n<p>When I run the code using:<\/p>\n\n<pre><\/pre>\n\n<p>It runs without issue, producing expected output.<\/p>\n\n<p>However, when I add a line to the top:<\/p>\n\n<pre><\/pre>\n\n<p>and run <\/p>\n\n<pre><\/pre>\n\n<p>and then this:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>My end goal is to modify the code to accept things like --epochs from the command line. How can I do this with Keras?<\/p>\n\n<p>EDIT:<\/p>\n\n<p>Here's the code that works (run using python mycode.py). You can <a href=\"http:\/\/archive.ics.uci.edu\/ml\/machine-learning-databases\/pima-indians-diabetes\/pima-indians-diabetes.data\" rel=\"nofollow noreferrer\">obtain the dataset here<\/a> (a csv, not huge). The problem comes up when I add the shebang to the top and chmod and run the alternate way.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42253862","CreationDate":"2017-02-15T15:44:36.327","ParentId":null,"OwnerUserId":"5424277","Title":null,"Body":"<p>Most likely, the reason is here: <code>#!\/usr\/bin\/python<\/code>. When you running <code>python<\/code> from console it can use another python interpreter (e.g. you have it installed at <code>\/usr\/local\/bin\/python<\/code>). So you need to use the same in your script, the easiest way is to declare it like <code>#!\/usr\/bin\/env python<\/code><\/p>\n"}]}
{"QuestionId":42253627,"AnswerCount":1,"Tags":"<tensorflow><keras><keras-layer>","CreationDate":"2017-02-15T15:34:19.013","AcceptedAnswerId":null,"OwnerUserId":1641501.0,"Title":"Classify images with different number of data for each class","Body":"<p>I have 7 class of images which I need to classify. Each class has a different number of sample. <\/p>\n\n<ol>\n<li>75 Images<\/li>\n<li>70 Images<\/li>\n<li>98 Images<\/li>\n<li>182 Images<\/li>\n<li>146 Images<\/li>\n<li>197 Images<\/li>\n<li>150 Images<\/li>\n<\/ol>\n\n<p>I ended up removing all images in all class down to 70 images (50 Training and 20 Validation). Using Keras and the generator and flow_from_directory method, I able to classify all of them, but it's obvious that I won't get a good accuracy. I have a plan to augment the data too later.  <\/p>\n\n<p>My Question is, can I use different number of training data for each class using flow_from_directory method? And is there any potential drawback with this method?<\/p>\n\n<p>For example : <\/p>\n\n<ol>\n<li>55 Training - 20 Validation<\/li>\n<li>50 Training - 20 Validation<\/li>\n<li>78 Training - 20 Validation<\/li>\n<li>......<\/li>\n<\/ol>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42254087","CreationDate":"2017-02-15T15:53:55.983","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>You can, but it's better to equalize the number of training samples in each class, or the classifier may tend to classify an image to the class in which there are most training samples.<\/p>\n\n<p>Say, maybe you can double the training samples in class 2 (by copying and pasting)?<\/p>\n\n<p>And for data augmentation, you can add noises to you training sample, say add Gaussian noises, crop and resize the images, and so on. By doing this, the model will be more robust.<\/p>\n"}]}
{"QuestionId":42253853,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-02-15T15:44:14.303","AcceptedAnswerId":42254973.0,"OwnerUserId":7396273.0,"Title":"Use AlexNet instead of LeNet model in Siamese Network","Body":"<p>I have successfully used  LeNet model to train my own dataset using Siamese Network using this <a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/siamese.html\" rel=\"nofollow noreferrer\">tutorial<\/a>. Now I want to use AlexNet as I believe it is more powerful than LeNet. Can someone provide guidelines or tutorial to use the AlexNet in Siamese Network.<\/p>\n","answers":[{"AnswerId":"42254973","CreationDate":"2017-02-15T16:31:24.687","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>You should go through this <a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow noreferrer\">github<\/a> repository. Specially the models and examples section. You can get the implementation of <a href=\"http:\/\/papers.nips.cc\/paper\/4824-imagenet-classification-with-deep-convolutional-neural-networks\" rel=\"nofollow noreferrer\">AlexNet<\/a> in caffe <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/master\/models\/bvlc_alexnet\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>N.B. Please don't flag this post because of its length or sharing link. Links that I shared in this answer contains large code which I cannot post as an answer.<\/p>\n"}]}
{"QuestionId":42253934,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><keras><lstm>","CreationDate":"2017-02-15T15:47:42.660","AcceptedAnswerId":42266198.0,"OwnerUserId":6380584.0,"Title":"Keras - How to use the learned Embedding() Layer for Input and Output?","Body":"<p>I would like to train a model to generate text, similar to <a href=\"https:\/\/larseidnes.com\/2015\/10\/13\/auto-generating-clickbait-with-recurrent-neural-networks\/\" rel=\"nofollow noreferrer\">this blog post<\/a><\/p>\n\n<p>This model uses - as far as I understand it - the following architecture<br>\n[Sequence of Word Indices] -> [Embedding] -> [LSTM] -> [1 Hot Encoded \"next word\"]<\/p>\n\n<p>Basically, the author models the process as classification problem, where the output layer has as many dimensions as there are words in the corpus. <\/p>\n\n<hr>\n\n<p>I would like to model the process as regression problem, by re-using the learned Embeddings and then minimising the distance between predicted and real embedding. <\/p>\n\n<p>Basically:<\/p>\n\n<p>[Sequence of Word Indices] -> [Embedding] -> [LSTM] -> [Embedding-Vector of the \"next word\"]<\/p>\n\n<p>My problem is, as the model is learning the embeddings on the fly, how could I feed the output in the same way I feed the input (as word indices) and then just tell the model \"But before you use the output, replace it by its embedding vector\" ?<\/p>\n\n<hr>\n\n<p>Thank you very much for all help :-)<\/p>\n","answers":[{"AnswerId":"42266198","CreationDate":"2017-02-16T06:07:56.243","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p><strong>In training phase:<\/strong><\/p>\n\n<p>You can use two inputs (one for target, one for input, there's an offset of 1 between these two sequences) and reuse the embedding layer.\nIf you input sentence is [1, 2, 3, 4], you can generate two sequence from it: in = [1, 2, 3], out = [2, 3, 4]. Then you can use Keras' functional API to reuse embedding layer:<\/p>\n\n<pre><code>emb1 = Embedding(in)\nemb2 = Embedding(out)\npredict_emb = LSTM(emb1)\nloss = mean_squared_error(emb2, predict_emb)\n<\/code><\/pre>\n\n<p>Note it's not Keras code, just pseudo code.<\/p>\n\n<p><strong>In testing phase:<\/strong><\/p>\n\n<p>Typically, you'll need to write your own decode function. Firstly, you choose a word (or a few words) to start from. Then, feed this word (or short word sequence) to network to predict next word's embedding. At this step, you can define your own sample function, say: you may want to choose the word whose embedding is nearest to the predicted one as the next word, or you may want to sample the next word from a distribution in which words with nearer embeddings to the predicted embedding has a larger probability to be chosen. Once you choose the next word, then feed it to network and predict the next one, and so forth.<\/p>\n\n<p><strong>So, you need to generate one word (put it another way, one embedding) at a time rather than input a whole sequence to the network.<\/strong><\/p>\n\n<p>If the above statements are too abstract for you, here's an good example: <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/lstm_text_generation.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/lstm_text_generation.py<\/a><\/p>\n\n<p>Line 85 is the introduction part, which randomly choose a small piece of texts from corpus to work on. From line 90 on there's a loop, in which each step samples a character (This is a char-rnn, so each timestep inputs a char. For your case, it should be a word, not a char): L95 predicts next char's distribution, L96 samples from the distribution. Hope this is clear enough.<\/p>\n"}]}
{"QuestionId":42254154,"AnswerCount":1,"Tags":"<deep-learning><keras><recurrent-neural-network><bidirectional>","CreationDate":"2017-02-15T15:56:18.200","AcceptedAnswerId":null,"OwnerUserId":5373209.0,"Title":"How do stateful bidirectional RNNs work in Keras","Body":"<p>In Keras the  wrapper for RNNs also supports . I don't really understand how this is supposed to work:<\/p>\n\n<p>In a stateful unidirectional model the state of a batch is carried over to the next batch. I guess it works the same for the forward layer in the bidirectional model.<\/p>\n\n<p>But where is the backward layer getting it's states from? If I understand everything correctly it should technically recieve it's state from the \"next\" batch. But obviously the \"next\" batch is not computet yet, so how does it work?<\/p>\n","answers":[{"AnswerId":"42584358","CreationDate":"2017-03-03T16:48:41.687","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>One may think about a <code>Bidirectional<\/code> layer in a following manner:<\/p>\n\n<pre><code>forward = Recurrent(..)(input)\nbackward = Recurrent(..., reverse_input=True)(input)\noutput = merge([forward, backward], ...)\n<\/code><\/pre>\n\n<p>So - as you can see - you are losing the <em>temporal<\/em> orientation. You are analysing the input both from its beginning and end. In this case - setting <code>stateful=True<\/code> simply takes its starting state from a previous sample accordingly to direction of a bidirectional branch (<code>forward<\/code> takes from <code>forward<\/code>, <code>backward<\/code> takes from <code>backward<\/code>).<\/p>\n\n<p>This makes your model losing the interpretation - that samples from concurrent batches might be interpreted as a compact sequence divided into batches.<\/p>\n"}]}
{"QuestionId":42254192,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-15T15:58:08.920","AcceptedAnswerId":null,"OwnerUserId":2476373.0,"Title":"Does TFslim TFExampleDecoder also support decoding sequence example?","Body":"<p>Does TFslim <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/data\/tfexample_decoder.py\" rel=\"nofollow noreferrer\">TFExampleDecoder<\/a> also support decoding sequence examples?<\/p>\n\n<p>If not, how easy do you think it will be to derive a child class to support it?<\/p>\n","answers":[]}
{"QuestionId":42254245,"AnswerCount":0,"Tags":"<python><tensorflow><distributed-computing>","CreationDate":"2017-02-15T15:59:52.177","AcceptedAnswerId":null,"OwnerUserId":1085548.0,"Title":"TensorFlow Distributed training using tf.Supervisor stalls when running init_op","Body":"<p>I'm trying to get some simple distributed training code working using TensorFlow r0.12.1 based on the <a href=\"https:\/\/www.tensorflow.org\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">example on this page from the documentation<\/a>. However, I'm having issues with my distributed sessions never returning.<\/p>\n\n<p>My cluster specification is as follows, and I am running both the ps and worker processes on the same server:<\/p>\n\n<pre><\/pre>\n\n<p>Using this, my ps process is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>And my two workers attempt to train a simple network using MNIST, and are launched using  or :<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, my code appears to freeze on the call to . Further diagnosis shows that the troublesome line is in fact the call to  which is initiated from the chief worker and, within this, the line 235 of :<\/p>\n\n<pre><\/pre>\n\n<p>So it seems like the whole process is stalling when attempting to run the initialisation operation, but does not produce any error messages. Regular training (not via tf.Supervisor) works fine on the same system, and so it seems unlikely it is installation \/ driver issues.<\/p>\n\n<p>Is there any reason why the process should stall when running ? Is there something I am missing about how to properly use ?<\/p>\n","answers":[]}
{"QuestionId":42255618,"AnswerCount":0,"Tags":"<python><indexing><tensorflow><neural-network><symbolic-math>","CreationDate":"2017-02-15T16:58:28.980","AcceptedAnswerId":null,"OwnerUserId":6024751.0,"Title":"Dynamic shape and Indexing with Tensorflow","Body":"<p>I am adapting <a href=\"http:\/\/deeplearning.net\/tutorial\/mlp.html\" rel=\"nofollow noreferrer\">this MLP tutorial<\/a> for Tensorflow.<\/p>\n\n<p>However, I can't figure out how to properly and efficiently compute the negative-log-likelihood. When I try to build the following graph, an error occur :<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 1 but is rank 2 for 'strided_slice_2' (op: 'StridedSlice') with input shapes: [?,10], [2,?], [2,?], [2]<\/p>\n<\/blockquote>\n\n<p><strong>Is it possible to index a tensor with dynamic shape in the graph?<\/strong><\/p>\n\n<p><strong>What am I missing?<\/strong><\/p>\n\n<p>P.s. When I evaluate (sess.run(...)) individually the ,  and  tensors, I can then easily index the resulting ndarray  with the two index vectors. The problem is that I don't know how to do that symbolically in TF.<\/p>\n","answers":[]}
{"QuestionId":42255729,"AnswerCount":1,"Tags":"<python><tensorflow><artificial-intelligence>","CreationDate":"2017-02-15T17:01:39.700","AcceptedAnswerId":42340320.0,"OwnerUserId":4724265.0,"Title":"Tensorflow Error: ValueError: Shapes must be equal rank, but are 2 and 1 From merging shape 1 with other shapes","Body":"<p>I am trying to use tensorflow for implementing a dcgan and have run into this error:<\/p>\n\n<pre><\/pre>\n\n<p>As far as iv gathered it indicates that my tensor shapes are different, but i cannot see what i need to change in order to fix this error. I believe the mistake hangs somewhere in between these methods:<\/p>\n\n<p>First i create a placeholder in a method using:<\/p>\n\n<pre><\/pre>\n\n<p>Then the last statement calls the generator method, this method uses reshape to change the tensor via:<\/p>\n\n<pre><\/pre>\n\n<p>If it helps here is my linear method:<\/p>\n\n<pre><\/pre>\n\n<p>EDIT:<\/p>\n\n<p>I also use these placeholders:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42340320","CreationDate":"2017-02-20T09:06:37.667","ParentId":null,"OwnerUserId":"7045779","Title":null,"Body":"<p><code>linear(z, self.gen_dimension * 8 * sample_H16 * sample_W16, 'gen_h0_lin', with_w=True)<\/code> would be return the tuple <code>(tf.matmul(input_, matrix) + bias, matrix, bias)<\/code>.<\/p>\n\n<p>Therefore, <code>self.z_<\/code> is assigned by the tuple, not the only one tf tensor.<\/p>\n\n<p>Just change <code>linear(z, self.gen_dimension * 8 * sample_H16 * sample_W16, 'gen_h0_lin', with_w=True)<\/code> to <code>linear(z, self.gen_dimension * 8 * sample_H16 * sample_W16, 'gen_h0_lin', with_w=False)<\/code>.<\/p>\n"}]}
{"QuestionId":42255855,"AnswerCount":1,"Tags":"<python><logging><tensorflow><filewriter>","CreationDate":"2017-02-15T17:07:38.087","AcceptedAnswerId":42260302.0,"OwnerUserId":4884986.0,"Title":"TensorFlow FileWriter not writing to file","Body":"<p>I am training a simple TensorFlow model. The training aspect works fine, but no logs are being written to  and I'm not sure why. Could anyone provide some insight? Thank you<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42260302","CreationDate":"2017-02-15T21:21:30.523","ParentId":null,"OwnerUserId":"4884986","Title":null,"Body":"<p>A combination of changing the file path from <code>\/temp\/...<\/code> to <code>temp\/...<\/code> and adding <code>summary_writer.flush()<\/code> and <code>summary_writer.close()<\/code> made the logs be written successfully.<\/p>\n"}]}
{"QuestionId":42256807,"AnswerCount":0,"Tags":"<python><python-2.7><machine-learning><tensorflow><data-science>","CreationDate":"2017-02-15T17:58:22.337","AcceptedAnswerId":null,"OwnerUserId":5368816.0,"Title":"Update part of linear indices in 1D Tensor","Body":"<p>I want to update only part of linear indices (in ones_flat), for example, half or quarter, like that: [0,1,2,3,.....N] changed to [0,1,....,N\/2]<\/p>\n\n<pre><\/pre>\n\n<p>The thing is, after setting to zero 1\/2 of elements in ones_mask and multiplication to W, W doesn't seem to change, since the accuracy of the network doesn't change. And there's no mistakes, just this:<\/p>\n\n<pre><\/pre>\n\n<p>Can you please help me to find the mistake here?<\/p>\n","answers":[]}
{"QuestionId":42256938,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-15T18:05:16.573","AcceptedAnswerId":null,"OwnerUserId":5029595.0,"Title":"What does tf.gfile do in TensorFlow?","Body":"<p>I've seen people using several functions from  such as  or . I have the idea that  deals with files. However, I haven't been able to find the official documentation to see what else it offers. <\/p>\n\n<p>It'd be great if you could help me with it.<\/p>\n","answers":[{"AnswerId":"42936969","CreationDate":"2017-03-21T19:55:12.913","ParentId":null,"OwnerUserId":"2476373","Title":null,"Body":"<p>For anyone landing here, the following answer was provided (by a googler) on: <a href=\"https:\/\/stackoverflow.com\/questions\/42922948\/why-use-tensorflow-gfile-for-file-i-o\/\">Why use tensorflow gfile? (for file I\/O)<\/a><\/p>\n\n<blockquote>\n  <p>The main roles of the tf.gfile module are:<\/p>\n  \n  <ol>\n  <li><p>To provide an API that is close to Python's file objects, and<\/p><\/li>\n  <li><p>To provide an implementation based on TensorFlow's C++ FileSystem API.<\/p><\/li>\n  <\/ol>\n  \n  <p>The C++ FileSystem API supports multiple file system implementations,\n  including local files, Google Cloud Storage (using a <code>gs:\/\/<\/code> prefix),\n  and HDFS (using an <code>hdfs:\/\/<\/code> prefix). TensorFlow exports these as\n  <code>tf.gfile<\/code>, so that you can use these implementations for saving and\n  loading checkpoints, writing TensorBoard logs, and accessing training\n  data (among other uses). However, if all of your files are local, you\n  can use the regular Python file API without any problem.<\/p>\n<\/blockquote>\n"},{"AnswerId":"42259977","CreationDate":"2017-02-15T21:01:56.857","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>As you correctly point out <code>tf.gfile<\/code> is an abstraction for accessing the filesystem and is documented <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gfile\" rel=\"nofollow noreferrer\">here<\/a>. It is recommended over using plain python API since it provides some level of portability.<\/p>\n"}]}
{"QuestionId":42257015,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2017-02-15T18:09:49.077","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"How can I list all Tensorflow variables a node depends on?","Body":"<p>How can I list all Tensorflow variables\/constants\/placeholders a node depends on?<\/p>\n\n<p>Example 1 (addition of constants):<\/p>\n\n<pre><\/pre>\n\n<p>I would like to have a function  such as:<\/p>\n\n<ul>\n<li> returns <\/li>\n<li> returns <\/li>\n<\/ul>\n\n<p>Example 2 (matrix multiplication between a placeholder and a weight matrix, followed by the addition of a bias vector):<\/p>\n\n<pre><\/pre>\n\n<p>I would like to have a function  such as:<\/p>\n\n<ul>\n<li> returns <\/li>\n<li> returns <\/li>\n<\/ul>\n","answers":[{"AnswerId":"56171327","CreationDate":"2019-05-16T14:51:30.933","ParentId":null,"OwnerUserId":"4790871","Title":null,"Body":"<p>These are all excellent answers, I will add a simple approach that produces the dependencies in a less easy to read format, but can be useful for quick debugging.<\/p>\n\n<pre><code>tf.get_default_graph().as_graph_def()\n<\/code><\/pre>\n\n<p>Printing that produces the operation in the graph as a simple dictionary shown below. Each OP is easy to spot by name with its attributes and inputs, allowing you to follow dependencies.<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.placeholder(tf.float32, name='placeholder_1')\nb = tf.placeholder(tf.float32, name='placeholder_2')\nc = a + b\n\ntf.get_default_graph().as_graph_def()\n\nOut[14]: \nnode {\n  name: \"placeholder_1\"\n  op: \"Placeholder\"\n  attr {\n    key: \"dtype\"\n    value {\n      type: DT_FLOAT\n    }\n  }\n  attr {\n    key: \"shape\"\n    value {\n      shape {\n        unknown_rank: true\n      }\n    }\n  }\n}\nnode {\n  name: \"placeholder_2\"\n  op: \"Placeholder\"\n  attr {\n    key: \"dtype\"\n    value {\n      type: DT_FLOAT\n    }\n  }\n  attr {\n    key: \"shape\"\n    value {\n      shape {\n        unknown_rank: true\n      }\n    }\n  }\n}\nnode {\n  name: \"add\"\n  op: \"Add\"\n  input: \"placeholder_1\"\n  input: \"placeholder_2\"\n  attr {\n    key: \"T\"\n    value {\n      type: DT_FLOAT\n    }\n  }\n}\nversions {\n  producer: 27\n}\n<\/code><\/pre>\n"},{"AnswerId":"55322685","CreationDate":"2019-03-24T10:11:42.890","ParentId":null,"OwnerUserId":"7661833","Title":null,"Body":"<p>In some cases one may want to find all the \"input\" variables which are connected to an \"output\" tensor, such as the loss of a graph. To this aim the following code snipped may be useful (inspired by the above code):<\/p>\n\n<pre><code>def findVars(atensor):\n    allinputs=atensor.op.inputs\n    if len(allinputs)==0:\n        if atensor.op.type == 'VariableV2' or atensor.op.type == 'Variable':\n            return set([atensor.op])\n    a=set()\n    for t in allinputs:\n        a=a | findVars(t)\n    return a\n<\/code><\/pre>\n\n<p>This can be used in debugging to find out where a connection in the graph is missing.<\/p>\n"},{"AnswerId":"42257992","CreationDate":"2017-02-15T19:03:51.933","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here are utilities I use for this (from <a href=\"https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/linearize\/linearize.py\" rel=\"noreferrer\">https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/linearize\/linearize.py<\/a>)<\/p>\n\n<pre><code># computation flows from parents to children\n\ndef parents(op):\n  return set(input.op for input in op.inputs)\n\ndef children(op):\n  return set(op for out in op.outputs for op in out.consumers())\n\ndef get_graph():\n  \"\"\"Creates dictionary {node: {child1, child2, ..},..} for current\n  TensorFlow graph. Result is compatible with networkx\/toposort\"\"\"\n\n  ops = tf.get_default_graph().get_operations()\n  return {op: children(op) for op in ops}\n\n\ndef print_tf_graph(graph):\n  \"\"\"Prints tensorflow graph in dictionary form.\"\"\"\n  for node in graph:\n    for child in graph[node]:\n      print(\"%s -&gt; %s\" % (node.name, child.name))\n<\/code><\/pre>\n\n<p>These functions work on ops. To get an op that produces tensor <code>t<\/code>, use <code>t.op<\/code>. To get tensors produced by op <code>op<\/code>, use <code>op.outputs<\/code><\/p>\n"},{"AnswerId":"42262154","CreationDate":"2017-02-15T23:36:39.397","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/a\/42257992\/395857\">Yaroslav Bulatov's answer<\/a> is great, I'll just add one plotting function that uses Yaroslav's <code>get_graph()<\/code>  and <code>children()<\/code> method:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport networkx as nx\ndef plot_graph(G):\n    '''Plot a DAG using NetworkX'''        \n    def mapping(node):\n        return node.name\n    G = nx.DiGraph(G)\n    nx.relabel_nodes(G, mapping, copy=False)\n    nx.draw(G, cmap = plt.get_cmap('jet'), with_labels = True)\n    plt.show()\n\nplot_graph(get_graph())\n<\/code><\/pre>\n\n<p>Plotting the example 1 from the question:<\/p>\n\n<pre><code>import matplotlib.pyplot as plt\nimport networkx as nx\nimport tensorflow as tf\n\ndef children(op):\n  return set(op for out in op.outputs for op in out.consumers())\n\ndef get_graph():\n  \"\"\"Creates dictionary {node: {child1, child2, ..},..} for current\n  TensorFlow graph. Result is compatible with networkx\/toposort\"\"\"\n  print('get_graph')\n  ops = tf.get_default_graph().get_operations()\n  return {op: children(op) for op in ops}\n\ndef plot_graph(G):\n    '''Plot a DAG using NetworkX'''        \n    def mapping(node):\n        return node.name\n    G = nx.DiGraph(G)\n    nx.relabel_nodes(G, mapping, copy=False)\n    nx.draw(G, cmap = plt.get_cmap('jet'), with_labels = True)\n    plt.show()\n\na = tf.constant(1, name = 'a')\nb = tf.constant(3, name = 'b')\nc = tf.constant(9, name = 'c')\nd = tf.add(a, b, name='d')\ne = tf.add(d, c, name='e')\n\nsess = tf.Session()\nprint(sess.run([d, e]))\nplot_graph(get_graph())\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Jcahf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Jcahf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Plotting the example 2 from the question:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XIT8T.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XIT8T.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>If you use Microsoft Windows, you may run into this issue: <a href=\"https:\/\/stackoverflow.com\/a\/42261757\/395857\">Python Error (ValueError: _getfullpathname: embedded null character)<\/a>, in which case you need to patch matplotlib as the link explains.<\/p>\n"}]}
{"QuestionId":42257196,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-02-15T18:19:59.640","AcceptedAnswerId":null,"OwnerUserId":7570624.0,"Title":"Problems reshaping an array in TF","Body":"<p>I'm passing in a [64,240] tf.array.\nI want to reshape this to a [64,10,24] tf.array.\nI've tried various different methods but I always get the same error.<\/p>\n\n<p>... raise ValueError(\"Shape %s must have rank %d\" % (self, rank))\nValueError: Shape (64, 240) must have rank 1<\/p>\n\n<p>The code fails at: \nself.x_expand[i] = tf.reshape(self.input_x[i],[num_classes,data_size]) <\/p>\n\n<p>I guess i'm making a basic mistake, hopefully someone can point out what it is...<\/p>\n\n<p>A larger fragment of the code is:<\/p>\n\n<p>class CNN(object):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42258045","CreationDate":"2017-02-15T19:06:58.467","ParentId":null,"OwnerUserId":"6557716","Title":null,"Body":"<p>I think you should try:<\/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>self.x_expand = tf.reshape(self.input_x, [batch_size, num_classes,data_size]<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"},{"AnswerId":"42257989","CreationDate":"2017-02-15T19:03:42.613","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>It seems you want <code>self.x_expand<\/code> to be the input reshaped I think the following might be more appropriate:<\/p>\n\n<pre><code>with tf.device('\/cpu:0'), tf.name_scope(\"reshaping\"):\n    self.x_expand = tf.reshape(self.input_x, [batch_size, num_classes, data_size])\n<\/code><\/pre>\n\n<p>Two things of note:<\/p>\n\n<ol>\n<li><p>Did not use a <code>tf.placeholder<\/code> for <code>self.x_expand<\/code>. Placeholders are meant for feeding input from the user, but here it seems what you really wanted was to reshape existing feeds <\/p><\/li>\n<li><p>Did not change shape on a per-element basis, but rather with a single operation.<\/p><\/li>\n<\/ol>\n\n<p>NB: I was unable to reproduce the error message you saw, possibly because we're using different versions of TensorFlow and there have been changes to the error message reporting in recent versions. I am using 1.0.0.<\/p>\n"}]}
{"QuestionId":42258101,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-15T19:10:29.180","AcceptedAnswerId":42288391.0,"OwnerUserId":7449493.0,"Title":"NameError in Tensorflow and Python","Body":"<p>This line indicates that the code started the processing part,\n\"Processing images\/Young-Bengal-tiger.jpg...\".<\/p>\n\n<p>I am trying to solve the error below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42288391","CreationDate":"2017-02-17T02:33:00.633","ParentId":null,"OwnerUserId":"7449493","Title":null,"Body":"<p>Thanks for your answers. I followed your recomendations and I solved my question by getting a better identation.<\/p>\n"}]}
{"QuestionId":42258829,"AnswerCount":0,"Tags":"<neural-network><keras>","CreationDate":"2017-02-15T19:55:37.033","AcceptedAnswerId":null,"OwnerUserId":2436425.0,"Title":"Validation MSE loss is different from validation MSE metric","Body":"<p>I have written a model in Keras (with theano backend) and compile my model like this: , i.e. my objective loss function is <em>mean squared error<\/em> and the metrics to report are <em>mean squared error<\/em> and <em>mean absolute error<\/em>.<\/p>\n\n<p>Then I run my model:\n<\/p>\n\n<p>Keras reports results as:<\/p>\n\n<blockquote>\n  <p>Epoch 500\/500:\n  0s - loss: 5.5990 - mean_squared_error: 4.4311 - mean_absolute_error: 0.9511 - val_loss: 7.5573 - val_mean_squared_error: 6.3877 - val_mean_absolute_error: 1.1335<\/p>\n<\/blockquote>\n\n<p>I expected <strong>val_loss<\/strong> to be same as <strong>val_mean_squared_error<\/strong>. What is <strong>val_loss<\/strong> here if not <strong>val_mean_squared_error<\/strong>?<\/p>\n","answers":[]}
{"QuestionId":42258989,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-15T20:05:27.680","AcceptedAnswerId":null,"OwnerUserId":1813487.0,"Title":"Suspicious warnings on Tensorflow 1.0","Body":"<p>Installing the brand new Tensorflow 1.0 from PyPI on Windows, and running the <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows#validate_your_installation\" rel=\"nofollow noreferrer\">Hello World example<\/a>, I get a bunch of suspicious warnings:<\/p>\n\n<pre><\/pre>\n\n<p>But after the warnings, \"Hello World\" does get printed. Should I worry?<\/p>\n","answers":[]}
{"QuestionId":42259071,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-02-15T20:10:41.850","AcceptedAnswerId":42278154.0,"OwnerUserId":826983.0,"Title":"How to feed a model with \"a list of outputs\"?","Body":"<p>Sorry for the title but I could't come up with a better description here.<\/p>\n\n<p>I am trying to apply batches for training on a model which should have 13 fully connected output layers. Each output layer has only two nodes (but are fully connected as stated).<\/p>\n\n<p>Building the model's output looks like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, I can't manage to feed this model. I've tried to go with a  sized output array:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>But for a batch of size 2 I get:<\/p>\n\n<pre><\/pre>\n\n<p>I've tried several other things but it's simply not clear to me how the input for the model looks like. <\/p>\n\n<p>How can I train this model?<\/p>\n\n<hr>\n\n<p>I have also tried to pass a list of lists of numpy arrays:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xZdUa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xZdUa.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>where the first level of the batch represent the sample (here 2) and the second level is the sample with the list of 13 numpy arrays. Yet I am getting:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>As suggested, I also tried to return a  of numpy arrays of size :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DgAjh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DgAjh.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Where the error becomes:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<h3>The code<\/h3>\n\n<p>Below you can find the current code which generates one sample in  and a full batch in  (which uses ).<\/p>\n\n<p>def batch_generator(w2v, file_path, meta_info, batch_size, sample_generator_fn, embedding_size):<\/p>\n\n<p><strong>Please note:<\/strong> The code shows now how I generate a  of  ndarrays whereas the number of such ndarrays in that list is defined by .<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>As requested, the call to <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42278154","CreationDate":"2017-02-16T15:29:15.277","ParentId":null,"OwnerUserId":"3427404","Title":null,"Body":"<p>Your output should be a list as specified in the error. Each element of the list should be a numpy array of size <code>[batch_size, nb_outputs]<\/code>. So a list of 13 elements of size <code>[batch_size,2]<\/code> in your case.<\/p>\n"}]}
{"QuestionId":42259301,"AnswerCount":0,"Tags":"<tensorflow><record>","CreationDate":"2017-02-15T20:23:45.260","AcceptedAnswerId":null,"OwnerUserId":7558108.0,"Title":"Storing a list of list as tf record in tensorflow","Body":"<p>What would be the best way of storing\/reading a list of list in a tf record in TensorFlow?\nI tried to serialize the data to a one dimensional list, then reshape it to its original size when reading. However, the encoding process is taking forever.<\/p>\n\n<p>Writing to :\nVariable in question:  (shape=[nb_channels, 1500]) <\/p>\n\n<pre><\/pre>\n\n<p>Reading from :<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42260147,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2017-02-15T21:12:09.163","AcceptedAnswerId":null,"OwnerUserId":7571333.0,"Title":"ValueError: Error when checking model input: expected input_node to have 4 dimensions, but got array with shape (0, 1)","Body":"<p>Using Keras and getting an error running model.fit, this results in an value error. But not entirely sure how checking the model fails when using fit. <\/p>\n\n<pre><\/pre>\n\n<p>Here is more details from the stack trace:<\/p>\n\n<pre><\/pre>\n\n<p>Additionally here is some more context on the model.<\/p>\n\n<pre><\/pre>\n\n<p>And then..<\/p>\n\n<pre><\/pre>\n\n<p>You can also clone this project via <a href=\"https:\/\/github.com\/metaflow-ai\/neural_style\" rel=\"nofollow noreferrer\">https:\/\/github.com\/metaflow-ai\/neural_style<\/a><\/p>\n","answers":[]}
{"QuestionId":42260167,"AnswerCount":3,"Tags":"<python><tensorflow><neural-network><deep-learning>","CreationDate":"2017-02-15T21:13:24.720","AcceptedAnswerId":null,"OwnerUserId":2414957.0,"Title":"Unsuccessful TensorSliceReader constructor: Failed to find any matching files for bird-classifier.tfl.ckpt-50912","Body":"<p>I am following this tutorial <a href=\"http:\/\/www.bitfusion.io\/2016\/08\/31\/training-a-bird-classifier-with-tensorflow-and-tflearn\/\" rel=\"noreferrer\">http:\/\/www.bitfusion.io\/2016\/08\/31\/training-a-bird-classifier-with-tensorflow-and-tflearn\/<\/a>\nI assume that training was done but the system was restarted so I can't verify if the 100 epochs were done. Can you please suggest fixes?<\/p>\n\n<pre><\/pre>\n\n<p>Here are the dumped training files I have <a href=\"http:\/\/pastebin.com\/9RF58yBB\" rel=\"noreferrer\">http:\/\/pastebin.com\/9RF58yBB<\/a> <\/p>\n","answers":[{"AnswerId":"43283827","CreationDate":"2017-04-07T17:20:35.647","ParentId":null,"OwnerUserId":"3990607","Title":null,"Body":"<p>I just had the same problem. \nThe problem is in this line:<\/p>\n\n<pre><code>model = tflearn.DNN(network, tensorboard_verbose=0, checkpoint_path='bird-classifier.tfl.ckpt')\n<\/code><\/pre>\n\n<p>Giving the checkpoint file name is not enough, you need to give the path. \nIf the checkpoint file is in the working directory, you can simply change it to<\/p>\n\n<pre><code>model = tflearn.DNN(network, tensorboard_verbose=0, checkpoint_path='.\/bird-classifier.tfl.ckpt')\n<\/code><\/pre>\n\n<p>and it should work. <\/p>\n"},{"AnswerId":"48085905","CreationDate":"2018-01-03T22:10:31.737","ParentId":null,"OwnerUserId":"1429847","Title":null,"Body":"<p>It's a bug. Just use \".\/test_images\/\" instead of test_images\/ in any path attributes. This helped for me.<\/p>\n"},{"AnswerId":"48936204","CreationDate":"2018-02-22T20:00:18.050","ParentId":null,"OwnerUserId":"5706413","Title":null,"Body":"<p>As said by <a href=\"https:\/\/stackoverflow.com\/users\/3990607\/patapouf-ai\">patapouf_ai<\/a> and <a href=\"https:\/\/stackoverflow.com\/users\/8588940\/sergey-shitikov\">sergy shitikov<\/a> <\/p>\n\n<p>use path as <\/p>\n\n<blockquote>\n  <p><strong>.<\/strong>\/filename<\/p>\n<\/blockquote>\n\n<p>instead of <\/p>\n\n<blockquote>\n  <p>\/filename<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42260265,"AnswerCount":1,"Tags":"<python><keras><keras-layer><cv2>","CreationDate":"2017-02-15T21:19:09.380","AcceptedAnswerId":42262632.0,"OwnerUserId":1158977.0,"Title":"Resizing an input image in a Keras Lambda layer","Body":"<p>I would like my  model to resize the input image using cv2 or similar.  <\/p>\n\n<p>I have seen the use of , but I would prefer to write my own generator and simply resize the image in the first layer with .<\/p>\n\n<p>How would I do this?<\/p>\n","answers":[{"AnswerId":"42262632","CreationDate":"2017-02-16T00:25:27.073","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>If you are using tensorflow backend then you can use <code>tf.image.resize_images()<\/code> function to resize the images in <code>Lambda<\/code> layer. <\/p>\n\n<p>Here is a small example to demonstrate the same:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import numpy as np\nimport scipy.ndimage\nimport matplotlib.pyplot as plt\n\nfrom keras.layers import Lambda, Input\nfrom keras.models import Model\nfrom keras.backend import tf as ktf\n\n\n# 3 channel images of arbitrary shape\ninp = Input(shape=(None, None, 3))\ntry:\n    out = Lambda(lambda image: ktf.image.resize_images(image, (128, 128)))(inp)\nexcept :\n    # if you have older version of tensorflow\n    out = Lambda(lambda image: ktf.image.resize_images(image, 128, 128))(inp)\n\nmodel = Model(input=inp, output=out)\nmodel.summary()\n\nX = scipy.ndimage.imread('test.jpg')\n\nout = model.predict(X[np.newaxis, ...])\n\nfig, Axes = plt.subplots(nrows=1, ncols=2)\nAxes[0].imshow(X)\nAxes[1].imshow(np.int8(out[0,...]))\n\nplt.show()\n<\/code><\/pre>\n"}]}
{"QuestionId":42263689,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-16T02:24:29.853","AcceptedAnswerId":42264945.0,"OwnerUserId":5522711.0,"Title":"How to translate or shift batches of Tensors randomly in Tensorflow","Body":"<p>I want to make my input image (tensor) to shift up\/down or right\/left randomly in every batch.<\/p>\n\n<p>For example,  I have a batch of grayscale images with size .<\/p>\n\n<p>If there is one image, I know I can use tf.pad and tf.slice(or other built-in functions)<\/p>\n\n<p>But I want to apply random shift to 10 different images with one operation.<\/p>\n\n<p>Is it possible? or should I use loop such as tf.scan?<\/p>\n","answers":[{"AnswerId":"44111564","CreationDate":"2017-05-22T11:20:57.210","ParentId":null,"OwnerUserId":"3053918","Title":null,"Body":"<p>As an alternative, you could also use <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/contrib\/image\/transform\" rel=\"nofollow noreferrer\">tf.contrib.image.transform()<\/a> and use the parameters <em>a2<\/em> and <em>b2<\/em> to translate the image:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nimage1 = np.array([[[.1], [.1], [.1], [.1]],\n                  [[.2], [.2], [.2], [.2]],\n                  [[.3], [.3], [.3], [.3]],\n                  [[.4], [.4], [.4], [.4]]])\nimage2 = np.array([[[.1], [.2], [.3], [.4]],\n                  [[.1], [.2], [.3], [.4]],\n                  [[.1], [.2], [.3], [.4]],\n                  [[.1], [.2], [.3], [.4]]])\nimages = np.stack([image1, image2])\nimages_ = tf.convert_to_tensor(images, dtype=tf.float32)\n\nshift1_x = 1\nshift1_y = 2\nshift2_x = -1\nshift2_y = 0\ntransforms_ = tf.convert_to_tensor([[1, 0, -shift1_x, 0, 1, -shift1_y, 0, 0],\n                                   [1, 0, -shift2_x, 0, 1, -shift2_y, 0, 0]],\n                                   tf.float32)\nshifted_ = tf.contrib.image.transform(images=images_,\n                                      transforms=transforms_)\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n\n    shifted = sess.run([shifted_])\n    print(shifted)\n<\/code><\/pre>\n\n<p>The <em>transforms<\/em> projection matrix can also be a tensor of size N x 8, so it is possible to shift every image of a batch differently. This can be easily extended by <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/random_uniform\" rel=\"nofollow noreferrer\">tf.random_uniform()<\/a> to include some randomness to the x\/y shift of each image.<\/p>\n\n<p><strong>Edit:<\/strong>\nTo use random shifts for every image of the batch:<\/p>\n\n<pre><code>...\nimages_ = tf.convert_to_tensor(images, dtype=tf.float32)\n\nnum_imgs = images.shape[0]\nbase_ = tf.convert_to_tensor(np.tile([1, 0, 0, 0, 1, 0, 0, 0], [num_imgs, 1]), dtype=tf.float32)\nmask_ = tf.convert_to_tensor(np.tile([0, 0, 1, 0, 0, 1, 0, 0], [num_imgs, 1]), dtype=tf.float32)\nrandom_shift_ = tf.random_uniform([num_imgs, 8], minval=-2.49, maxval=2.49, dtype=tf.float32)\ntransforms_ = base_ + random_shift_ * mask_\n\nshifted_ = tf.contrib.image.transform(images=images_,\n                                      transforms=transforms_)\n...\n<\/code><\/pre>\n\n<p><strong>Edit 2:<\/strong>\nFor the sake of completion, here just another helper function with applies random rotation and shift to each single image of a batch:<\/p>\n\n<pre><code>def augment_data(input_data, angle, shift):\n    num_images_ = tf.shape(input_data)[0]\n    # random rotate\n    processed_data = tf.contrib.image.rotate(input_data,\n                                             tf.random_uniform([num_images_],\n                                                               maxval=math.pi \/ 180 * angle,\n                                                               minval=math.pi \/ 180 * -angle))\n    # random shift\n    base_row = tf.constant([1, 0, 0, 0, 1, 0, 0, 0], shape=[1, 8], dtype=tf.float32)\n    base_ = tf.tile(base_row, [num_images_, 1])\n    mask_row = tf.constant([0, 0, 1, 0, 0, 1, 0, 0], shape=[1, 8], dtype=tf.float32)\n    mask_ = tf.tile(mask_row, [num_images_, 1])\n    random_shift_ = tf.random_uniform([num_images_, 8], minval=-shift, maxval=shift, dtype=tf.float32)\n    transforms_ = base_ + random_shift_ * mask_\n\n    processed_data = tf.contrib.image.transform(images=processed_data,\n                                                transforms=transforms_)\n    return processed_data\n<\/code><\/pre>\n"},{"AnswerId":"42264945","CreationDate":"2017-02-16T04:26:17.037","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>Are you looking for <code>tf.random_crop<\/code> and <code>tf.pad<\/code>?<\/p>\n\n<p>Well, when using <code>tf.random_crop<\/code>, a random shift will be applied to all images in the batch. The shift inside a batch is the same, but can be different for different batches.<\/p>\n\n<p>If you want to use different shift within a batch, I think it's better to use a queue\/input pipeline. See <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/reading_data<\/a> for more.\nHere's an example code from part of my own project. <code>self.image_names<\/code> is a Python list which contains paths to all training images. In an input pipeline, the data flow like a stream: you just need to deal with only one image, and the queue automatically takes care of scheduling things (some threads read the data, some process the data, some group single images into batches, others feed the data to GPU, etc., to keep the whole pipeline busy). In the code below, <code>images<\/code> and <code>labels<\/code> are queues. That is to say, when you process this variable (as I do in <code>self.data_augmentation<\/code>), you can think it contains only one image, but actually the queue processes every item in it (It's like an implicit loop), then <code>tf.train.shuffle_batch<\/code> will shuffle training data in the queue and group them into batches.<\/p>\n\n<pre><code>def data_augmentation(images):\n    if FLAGS.random_flip_up_down:\n        images = tf.image.random_flip_up_down(images)\n    if FLAGS.random_brightness:\n        images = tf.image.random_brightness(images, max_delta=0.3)\n    if FLAGS.random_contrast:\n        images = tf.image.random_contrast(images, 0.8, 1.2)\n    return images\n\ndef input_pipeline(self, batch_size, num_epochs=None, aug=False):\n    images_tensor = tf.convert_to_tensor(self.image_names, dtype=tf.string)\n    labels_tensor = tf.convert_to_tensor(self.labels, dtype=tf.int64)\n    input_queue = tf.train.slice_input_producer([images_tensor, labels_tensor], num_epochs=num_epochs)\n\n    labels = input_queue[1]\n    images_content = tf.read_file(input_queue[0])\n    images = tf.image.convert_image_dtype(tf.image.decode_png(images_content, channels=1), tf.float32)\n    if aug:\n        images = self.data_augmentation(images)\n    new_size = tf.constant([FLAGS.image_size, FLAGS.image_size], dtype=tf.int32)\n    images = tf.image.resize_images(images, new_size)\n    image_batch, label_batch = tf.train.shuffle_batch([images, labels], batch_size=batch_size, capacity=50000,\n                                                      min_after_dequeue=10000)\n    # print 'image_batch', image_batch.get_shape()\n    return image_batch, label_batch\n<\/code><\/pre>\n"}]}
{"QuestionId":42263798,"AnswerCount":0,"Tags":"<python><keras>","CreationDate":"2017-02-16T02:35:07.530","AcceptedAnswerId":null,"OwnerUserId":5840158.0,"Title":"Keras LSTM Network Input Format","Body":"<p>I have a time series of regressors  where each  is a vector of dimension . I have another time series  each of which is a scalar. The goal is to predict<\/p>\n\n<p> for some function f. I want to forecast L steps ahead with regressors looking back upto M. <\/p>\n\n<p>How do I create an LSTM network? I am confused by the documentation.<\/p>\n\n<p>I started with<\/p>\n\n<pre><\/pre>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":42264567,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-02-16T03:50:47.063","AcceptedAnswerId":42264773.0,"OwnerUserId":624869.0,"Title":"Keras ML library: how to do weight clipping after gradient updates? TensorFlow backend","Body":"<p>I'm trying to use Keras to implement part of an algorithm that requires weight clipping, i.e. limiting the weight values after a gradient update. I haven't found any solutions through web searches so far.<\/p>\n\n<p>For background, this has to do with the WGANs algorithm: <\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1701.07875.pdf\" rel=\"noreferrer\">https:\/\/arxiv.org\/pdf\/1701.07875.pdf<\/a><\/p>\n\n<p>If you look at algorithm 1 on page 8, you'll see the following: <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/x7ZSN.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/x7ZSN.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I've highlighted the lines that I'm trying to implement in Keras: after computing a gradient to use to update the weights in the network, I want to make sure that all the weights are clipped between some values [-c, c] that I can set. <\/p>\n\n<p>How could I go about doing this in Keras? <\/p>\n\n<p>For reference I am using the TensorFlow backend. I don't mind digging into things and adding messy quick-fixes for now. <\/p>\n\n<p>Thanks! <\/p>\n","answers":[{"AnswerId":"42264773","CreationDate":"2017-02-16T04:10:18.737","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>While creating the optimizer object set param <code>clipvalue<\/code>. <s>It will do precisely what you want.<\/s><\/p>\n\n<pre><code># all parameter gradients will be clipped to\n# a maximum value of 0.5 and\n# a minimum value of -0.5.\nrsmprop = RMSprop(clipvalue=0.5)\n<\/code><\/pre>\n\n<p>and then use this object to for model compiling<\/p>\n\n<pre><code>model.compile(loss='mse', optimizer=rsmprop)\n<\/code><\/pre>\n\n<p>For more reference check: <a href=\"https:\/\/keras.io\/optimizers\/#parameters-common-to-all-keras-optimizers\" rel=\"noreferrer\">here<\/a>. <\/p>\n\n<p>Also, I prefer to use <code>clipnorm<\/code> over <code>clipvalue<\/code> because with <code>clipnorm<\/code> the optimization remains stable. For example say you have 2 parameters and the gradients came out to be <code>[0.1, 3]<\/code>. By using <code>clipvalue<\/code> the gradients will become [0.1, 0.5] ie there are chances that the direction of steepest decent can get changed drastically. While <code>clipnorm<\/code> don't have similar problem as all the gradients will be appropriately scaled and the direction will be preserved and all the while ensuring the constraint on the magnitude of the gradient. <\/p>\n\n<p><strong>Edit: The question asks weights clipping not gradient clipping:<\/strong><\/p>\n\n<p>Gradiant clipping on weights is not part of keras code. But <code>maxnorm<\/code> on weights constraints is. Check <a href=\"https:\/\/keras.io\/constraints\/\" rel=\"noreferrer\">here<\/a>.<\/p>\n\n<p>Having said that it can be easily implemented. Here is a very small example:<\/p>\n\n<pre><code>from keras.constraints import Constraint\nfrom keras import backend as K\n\nclass WeightClip(Constraint):\n    '''Clips the weights incident to each hidden unit to be inside a range\n    '''\n    def __init__(self, c=2):\n        self.c = c\n\n    def __call__(self, p):\n        return K.clip(p, -self.c, self.c)\n\n    def get_config(self):\n        return {'name': self.__class__.__name__,\n                'c': self.c}\n\nimport numpy as np\nfrom keras.models import Sequential\nfrom keras.layers import Dense\n\nmodel = Sequential()\nmodel.add(Dense(30, input_dim=100, W_constraint = WeightClip(2)))\nmodel.add(Dense(1))\n\nmodel.compile(loss='mse', optimizer='rmsprop')\n\nX = np.random.random((1000,100))\nY = np.random.random((1000,1))\n\nmodel.fit(X,Y)\n<\/code><\/pre>\n\n<p>I have tested the running of the above code, but not the validity of the constraints. You can do so by getting the model weights after training using <code>model.get_weights()<\/code> or <code>model.layers[idx].get_weights()<\/code> and checking whether its abiding the constraints. <\/p>\n\n<p>Note: The constrain is not added to all the model weights .. but just to the weights of the specific layer its used and also <code>W_constraint<\/code> adds constrain to <code>W<\/code> param and <code>b_constraint<\/code> to <code>b<\/code> (bias) param<\/p>\n"}]}
{"QuestionId":42264649,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-16T03:58:48.420","AcceptedAnswerId":42275663.0,"OwnerUserId":5171213.0,"Title":"Keras: Binary_crossentropy has negative values","Body":"<p>I'm following <a href=\"http:\/\/machinelearningmastery.com\/tutorial-first-neural-network-python-keras\/\" rel=\"noreferrer\">this tutorial<\/a> (section 6: Tying it All Together), with my own dataset. I can get the example in the tutorial working, no problem, with the sample dataset provided.<\/p>\n\n<p>I'm getting a binary cross-entropy error that is negative, and no improvements as epochs progress. I'm pretty sure binary cross-entropy should always be positive, and I should see some improvement in the loss. I've truncated the sample output (and code call) below to 5 epochs. Others seem to run into similar problems sometimes when training CNNs, but I didn't see a clear solution in my case. Does anyone know why this is happening?<\/p>\n\n<p>Sample output:<\/p>\n\n<pre><\/pre>\n\n<p>My code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42275663","CreationDate":"2017-02-16T13:43:03.520","ParentId":null,"OwnerUserId":"5171213","Title":null,"Body":"<p>I should have printed out my response variable. The categories were labelled as 1 and 2 instead of 0 and 1, which confused the classifier.<\/p>\n"}]}
{"QuestionId":42264716,"AnswerCount":1,"Tags":"<tensorflow><linear-regression><gradient-descent>","CreationDate":"2017-02-16T04:04:56.137","AcceptedAnswerId":42267630.0,"OwnerUserId":7572581.0,"Title":"My TensorFlow Gradient Descent diverges","Body":"<pre><\/pre>\n\n<p>Here's my code. My sat.csv contains a data of reading, writing and math scores at SAT. As you can guess, the difference between the features is not that big. <\/p>\n\n<p>This is a part of sat.csv. <\/p>\n\n<pre><\/pre>\n\n<p>I've only used math, writing and reading scores. My goal for the code above is to predict the writing score if I give math and reading scores. <\/p>\n\n<p>I've never seen Tensorflow's gradient descent model diverges with this such simple data. What'd be wrong?<\/p>\n","answers":[{"AnswerId":"42267630","CreationDate":"2017-02-16T07:35:29.493","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Here are a few options you could try: <\/p>\n\n<ul>\n<li>Normalise you input and output data<\/li>\n<li>Set smaller initial values for your weights<\/li>\n<li>Use a lower learning rate<\/li>\n<li>Divide your loss by the amount of samples you have (not putting your data in a placeholder is already uncommon). <\/li>\n<\/ul>\n\n<p>Let me know what (if any) of these options helped and good luck!<\/p>\n"}]}
{"QuestionId":42264873,"AnswerCount":1,"Tags":"<c++><machine-learning><tensorflow>","CreationDate":"2017-02-16T04:19:52.230","AcceptedAnswerId":null,"OwnerUserId":456329.0,"Title":"C++ equivalent of train.import_meta_graph clear_devices argument?","Body":"<p>I'm trying to train a graph on GPU using python, to load graph from C++ process.<\/p>\n\n<pre><\/pre>\n\n<p>Then I got error message<\/p>\n\n<p>\"Cannot assign a device to node ... because no device matching that specification are registered in this process; available devices: \/job:localhost\/replica:0\/task:0\/cpu:0\"<\/p>\n\n<p>For python train.import_meta_graph API has clear_devices argument, but what is equivalent of it on C++ API?<\/p>\n\n<p>For loading graph, I'm using Tensorflor on Windows built using CMake with -Dtensorflow_ENABLE_GPU=ON, so my vcxproj has GOOGLE_CUDA definition.<\/p>\n\n<p>I've read <a href=\"https:\/\/stackoverflow.com\/questions\/40451974\/tensorflow-restore-variables-in-a-specific-device\">Tensorflow, restore variables in a specific device<\/a> but it's only for python API.<\/p>\n","answers":[{"AnswerId":"42265749","CreationDate":"2017-02-16T05:33:43.307","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>Given that you're exporting the graph from Python anyway, perhaps you can export a graph with the devices cleared out? Something like:<\/p>\n\n<pre><code>meta_graph = tf.train.export_meta_graph()\nwith tf.Graph().as_default():\n    tf.train.import_meta_graph(meta_graph, clear_devices=True)\n    # Export the GraphDef now\n    with open('\/tmp\/model.pb', 'w') as f:\n        f.write(tf.get_default_graph().as_graph_def().SerializeToString())\n<\/code><\/pre>\n\n<p>Alternatively, you could duplicate the behavior of <code>clear_devices=True<\/code> in C++ by clearing out the <code>device<\/code> field of every node in the graph. Something like:<\/p>\n\n<pre><code>status = ReadBinaryProto(Env::Default(), \"model.pb\", &amp;graph_def);\nfor (int n = 0; n &lt; graph_def.node_size(); ++n) {\n  graph_def.mutable_node(n)-&gt;clear_device();\n}\nsession-&gt;Create(graph_def);\n<\/code><\/pre>\n\n<p>But I would advise against that since it is relying on internal details of how <code>GraphDef<\/code>s are consumed by the framework, which might be fragile. <\/p>\n"}]}
{"QuestionId":42265950,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-02-16T05:49:05.730","AcceptedAnswerId":null,"OwnerUserId":7572852.0,"Title":"how to feed DNNClassifier with numpy arrays","Body":"<p>I'm trying to create a DNNClassifier but i don't know how to pass my data into the object.My data files are .npy files created with np.save().<\/p>\n\n<ul>\n<li>Training data: an array of shape (106398,338) where 106398 is the number of instances of data.<\/li>\n<li>Training labels: an array of shape (106398,97) where 97 is the number of classes that i want to predict (in hot encoding)<\/li>\n<\/ul>\n\n<hr>\n\n<pre><\/pre>\n\n<p>and i get: ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()<\/p>\n\n<p>I've tried to make both(feature_columns and feature_tags) into tf.constant bot it don't work.<\/p>\n\n<p>How can i fix it?<\/p>\n\n<hr>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42266133","CreationDate":"2017-02-16T06:03:28.420","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>If look at other SO questions about that <code>ValueError..<\/code> you'll that this arises when you try to do some sort True\/False test on an array.  <\/p>\n\n<pre><code>if X&gt;0:....\n<\/code><\/pre>\n\n<p>produces this error if <code>X<\/code> is a multielement array.  <code>X&gt;0<\/code> is then an array of True\/False values.  That's ambiguous.  <\/p>\n\n<p>Having identified the basic issue, we then need to find where you are doing that sort of test.<\/p>\n\n<p>Another thing - when reporting an error, also report on the stack - <strong>Where exactly does this error occur?<\/strong><\/p>\n\n<p>Looking at your code, I don't see any test that could trigger this error.  That means it is happening deep inside one of the functions that you call.  <strong>Which?<\/strong><\/p>\n\n<p>I'm guessing that one of function arguments has the wrong form, shape or type.<\/p>\n\n<hr>\n\n<p>The error is in the <code>tuple(feature_columns or [])<\/code> expression.  Your <code>feature_columns<\/code> parameter should not be an array.  Check the documentation.<\/p>\n\n<p>From that expression, I'm guessing that <code>feature_columns<\/code> should a default <code>None<\/code>, or a list like <code>[1,2,3]<\/code>:<\/p>\n\n<pre><code>In [110]: [1,2,3] or []\nOut[110]: [1, 2, 3]\nIn [111]: None or []\nOut[111]: []\nIn [112]: np.array([1,2,3]) or []\n....\nValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()\nIn [113]: \n<\/code><\/pre>\n"},{"AnswerId":"43782932","CreationDate":"2017-05-04T12:28:26.850","ParentId":null,"OwnerUserId":"2740947","Title":null,"Body":"<p>Use a function <code>learn.infer_real_valued_columns_from_input<\/code> to create <code>feature_columns<\/code> from the numpy array. The following solution works for me:<\/p>\n\n<pre><code>X = np.load(path_to_file)\nfeature_columns = tf.contrib.learn.infer_real_valued_columns_from_input(X)\nclassifier = DNNClassifier(hidden_units=[10, 20, 10], n_classes=97, feature_columns=feature_columns)\n<\/code><\/pre>\n"}]}
{"QuestionId":42266094,"AnswerCount":9,"Tags":"<tensorflow>","CreationDate":"2017-02-16T06:00:55.223","AcceptedAnswerId":42266598.0,"OwnerUserId":1642057.0,"Title":"Tensorflow 1.0 Windows + 64-bit Anaconda 4.3.0 error","Body":"<p>Following the instructions under \"Installing with Anaconda\" at <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a>, I get to this point and get a failure.  <\/p>\n\n<p><\/p>\n\n<p>This is my environment 64-bit Anaconda3 4.3.0<\/p>\n\n<p><\/p>\n\n<p>I previously successfully installed pre-1.0 tensorflow from pip.<\/p>\n","answers":[{"AnswerId":"42266598","CreationDate":"2017-02-16T06:35:03.053","ParentId":null,"OwnerUserId":"1642057","Title":null,"Body":"<p>Apparently python 3.5 is very important.  Since the latest Anaconda3 distribution is 3.6, you have to do this:<\/p>\n\n<p><code>&gt; conda create -n tensorflow python=3.5<\/code><\/p>\n\n<p>and it seems the 1.0 package is available on pip now, so you can just do this:<\/p>\n\n<p><code>..&gt; activate tensorflow\n...&gt; pip install tensorflow-gpu\nCollecting tensorflow-gpu\n  Downloading tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl (43.1MB)\n    100% |################################| 43.1MB 31kB\/s\nCollecting protobuf&gt;=3.1.0 (from tensorflow-gpu)\n...\n<\/code><\/p>\n\n<p>I've successfully run the 'hello world' commands in the python 3.5 shell and <code>tf.__version__<\/code> shows as <code>'1.0.0'<\/code><\/p>\n"},{"AnswerId":"42266602","CreationDate":"2017-02-16T06:35:29.000","ParentId":null,"OwnerUserId":"7274047","Title":null,"Body":"<p><strong>1) C:>pip uninstall tensorflow <br>\n2) C:>pip install tensorflow<\/strong><\/p>\n"},{"AnswerId":"42269352","CreationDate":"2017-02-16T09:08:12.607","ParentId":null,"OwnerUserId":"5797975","Title":null,"Body":"<p>It looks like <code>tensorflow-1.0.0-cp35-cp35m-win_x86_64.whl<\/code> will fail but <code>tensorflow-1.0.0-cp35-cp35m-win_amd_64.whl<\/code> works fine for me.<\/p>\n"},{"AnswerId":"42287869","CreationDate":"2017-02-17T01:36:47.707","ParentId":null,"OwnerUserId":"5709540","Title":null,"Body":"<p>You can see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7552\" rel=\"nofollow noreferrer\">this issue<\/a>, maybe <code>win_x86_64<\/code> should be changed to <code>win_amd64<\/code><\/p>\n"},{"AnswerId":"42298455","CreationDate":"2017-02-17T12:51:19.100","ParentId":null,"OwnerUserId":"7580851","Title":null,"Body":"<p>I was facing the same issue yesterday with command\n(tensorflow)C:> pip install --ignore-installed --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_x86_64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_x86_64.whl<\/a><\/p>\n\n<p>simply write <\/p>\n\n<blockquote>\n  <p>C:>pip install tensorflow<\/p>\n  \n  <p>C:>pip install tensorflow-gpu #for gpu version<\/p>\n<\/blockquote>\n"},{"AnswerId":"42398001","CreationDate":"2017-02-22T17:09:51.593","ParentId":null,"OwnerUserId":"5501189","Title":null,"Body":"<p>For who using python version 3.6 and anaconda. <\/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>conda create -n tensorflow python=3.5\r\nactivate tensorflow\r\npip install tensorflow-gpu<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"},{"AnswerId":"42463201","CreationDate":"2017-02-26T00:09:17.673","ParentId":null,"OwnerUserId":"4469065","Title":null,"Body":"<p>I also had same problem when I installed anaconda 4.3 version<\/p>\n\n<p>\"tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl is not a supported wheel on this platform\"<\/p>\n\n<p>Here is my solution.<\/p>\n\n<ul>\n<li>Instead of using Anaconda3 4.3, install Anaconda3 4.2(Anaconda3-4.2.0-Windows-x86_64.exe)<\/li>\n<li><p>Type on command line(If you are using GPU version)<\/p>\n\n<pre><code>pip install -U --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre><\/li>\n<li><p>Typeon command line(If you are using CPU only)<\/p>\n\n<pre><code>pip install -U --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre><\/li>\n<\/ul>\n\n<p>I'm working on win10 with python version=3.5.2, 64 bit<\/p>\n"},{"AnswerId":"42291706","CreationDate":"2017-02-17T07:19:10.383","ParentId":null,"OwnerUserId":"5512020","Title":null,"Body":"<p>I have installed Miniconda(Anaconda). I have used pycharm in which I have added the Miniconda(Anaconda) interpreter. TensorFlow can be easily added through this interpreter by searching for the required <code>TensorFlow<\/code> package<\/p>\n"},{"AnswerId":"42303097","CreationDate":"2017-02-17T16:40:13.643","ParentId":null,"OwnerUserId":"3873491","Title":null,"Body":"<p>I was facing the same problem, and this snipped should work for most of people.<\/p>\n\n<pre><code>pip install tensorflow-gpu\n<\/code><\/pre>\n\n<p>Because I have install tensorflow version 0.12 before, it will use cache, so I have to specify the version to be 1.0<\/p>\n\n<pre><code>pip install tensorflow-gpu==1.0\n<\/code><\/pre>\n\n<p>It's worked well for me!!<\/p>\n\n<p>I'm working on win10 with python version=3.5.2, 64 bit<\/p>\n"}]}
{"QuestionId":42266867,"AnswerCount":0,"Tags":"<python><numpy><flask><caffe><gunicorn>","CreationDate":"2017-02-16T06:50:56.950","AcceptedAnswerId":null,"OwnerUserId":5400749.0,"Title":"Python float represented as dictionary","Body":"<p>I have some application developed in python2 using caffe as a framework to classify some types of images. The output from such system should be a numpy array, its values is the probability from softmax layer where I should take the largest value and its index as the prediction.<\/p>\n\n<p>However, sometimes (like 3 times out of 200) the probability value is not printed in decimal representation. Instead it is in the form of a dictionary.<\/p>\n\n<p>For example, instead of , I got\n<\/p>\n\n<p>Anybody have experience the same things? For additional information, the server is implemented using flask with gunicorn as web server.<\/p>\n\n<p>Thanks!<\/p>\n\n<p>==========================================================================<\/p>\n\n<p>Edit: this is snippet of the code that responsible for the value<\/p>\n\n<pre><\/pre>\n\n<p>and this in different files<\/p>\n\n<pre><\/pre>\n\n<p>The full version of related libraries is numpy==1.11.1 flask==0.11.1 gunicorn==19.6.0<\/p>\n","answers":[]}
{"QuestionId":42267185,"AnswerCount":1,"Tags":"<centos><tensorflow><anaconda><environment>","CreationDate":"2017-02-16T07:10:26.570","AcceptedAnswerId":null,"OwnerUserId":7573229.0,"Title":"Can't activate anaconda environment on CentOS","Body":"<p>I'm working on install tensorflow on my CentOS system.<\/p>\n\n<p>I had success install miniconda in my system and created tensorflow environment.\nHowever, I got error message while I run \"<em>source activate tensorflow<\/em>\".<\/p>\n\n<p>I checked on Anaconda page and found out the reason is activate file only support bash and zsh. But my OS centOS is csh (C shell).<\/p>\n\n<p>And I also try to add \"<em>set path<\/em> \" for the tensorflow environment on my <em>~\/.cshrc<\/em> file. But it looks not working, I still get ImportError while I import tensorflow.<\/p>\n\n<p>Does anyone know how to solve this problem? <\/p>\n","answers":[{"AnswerId":"42453033","CreationDate":"2017-02-25T07:01:03.670","ParentId":null,"OwnerUserId":"1005215","Title":null,"Body":"<p>You can try the following (on CentOS 6.8 or CentOS 7):<\/p>\n\n<pre><code>[root@9dc1f60d1dcd ~]# csh\n[root@9dc1f60d1dcd ~]# echo $0\ncsh\n[root@9dc1f60d1dcd ~]# bash Miniconda2-4.3.11-Linux-x86_64.sh -b -p \/m2\n[root@9dc1f60d1dcd ~]# \/m2\/bin\/conda install -y tensorflow\n[root@9dc1f60d1dcd ~]# \/m2\/bin\/python\nPython 2.7.13 |Continuum Analytics, Inc.| (default, Dec 20 2016, 23:09:15) \n[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\nAnaconda is brought to you by Continuum Analytics.\nPlease check out: http:\/\/continuum.io\/thanks and https:\/\/anaconda.org\n&gt;&gt;&gt; import tensorflow\n&gt;&gt;&gt; tensorflow.__version__\n'0.10.0rc0'\n<\/code><\/pre>\n"}]}
{"QuestionId":42267743,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><virtualenv>","CreationDate":"2017-02-16T07:42:19.353","AcceptedAnswerId":null,"OwnerUserId":6714974.0,"Title":"Tensorflow - ( retrain.py ) OSError: [Errno 2] No such file or directory: ''","Body":"<p>I'm running tensorflow in Ubuntu (in VirtualEnv).<\/p>\n\n<p>I want to create training model ...( using retrain.py to train the model )..<\/p>\n\n<p>Github code link ==> <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py<\/a><\/p>\n\n<p>I tried to use command to run in terminal <\/p>\n\n<pre><\/pre>\n\n<p>I'm getting error. How to read that error and how to resolve it..<\/p>\n\n<pre><\/pre>\n\n<p>Note :- I'm new in tensorflow and just started to use it..<\/p>\n","answers":[{"AnswerId":"44685809","CreationDate":"2017-06-21T20:34:55.477","ParentId":null,"OwnerUserId":"7386732","Title":null,"Body":"<p>while running retrain.py, you do not have to specify all the parameters.\nJust specify<\/p>\n\n<pre><code>python \/path_to_file\/retrain.py --img_dir \/dir_containing_subfolders_of_images\n<\/code><\/pre>\n\n<p>while running the code and it will use default paths for remaining arguments. eg. \/tmp\/some_dir_or_file for most cases.<\/p>\n\n<p>If you want to specify custom locations create the desired directories and specify those paths.\neg.<\/p>\n\n<pre><code>python \/path_to_file\/retrain.py \n--img_dir \/dir_containing_subfolders_of_images\n--how_many_training_steps number_of_steps\n--output_graph \/dirPath_to_store_output_graph\/file_name.pb\n--output_labels \/dirPath_to_store_output_labels\/file_name.txt \n<\/code><\/pre>\n"}]}
{"QuestionId":42268180,"AnswerCount":3,"Tags":"<python><macos><curl><tensorflow><deep-learning>","CreationDate":"2017-02-16T08:07:02.647","AcceptedAnswerId":48197901.0,"OwnerUserId":7568128.0,"Title":"tar: Unrecognized archive format error when trying to unpack flower_photos.tgz, TF tutorials on OSX","Body":"<p>I'm trying to unpack the flower_photos.tgz after curling it using <\/p>\n\n<pre><\/pre>\n\n<p>This is from the image retraining tutorial for TensorFlow<\/p>\n\n<p>Results from curling<\/p>\n\n<pre><\/pre>\n\n<p>Then when I try to unpack<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47350025","CreationDate":"2017-11-17T11:50:45.607","ParentId":null,"OwnerUserId":"7353970","Title":null,"Body":"<p>try using the <code>-L<\/code> flag, i.e. <\/p>\n\n<pre><code>curl -OL http:\/\/download.tensorflow.org\/example_..\/images\/flower_photos.tgz\n<\/code><\/pre>\n\n<p>source: <a href=\"https:\/\/github.com\/arq5x\/bedtools2\/issues\/189\" rel=\"nofollow noreferrer\">https:\/\/github.com\/arq5x\/bedtools2\/issues\/189<\/a><\/p>\n"},{"AnswerId":"43269651","CreationDate":"2017-04-07T04:27:54.527","ParentId":null,"OwnerUserId":"5511906","Title":null,"Body":"<p>I solved this problem by downloading the file from safari and then copy the file to your directory. There is something wrong with the file download by command curl. Maybe you can try my method.<\/p>\n"},{"AnswerId":"48197901","CreationDate":"2018-01-10T23:48:05.490","ParentId":null,"OwnerUserId":"7568128","Title":null,"Body":"<p>Apparently the new instructions on TensorFlow website run without issues<\/p>\n\n<p>I just tried the instructions posted on <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">How to Retrain Inception's Final Layer for New Categories<\/a><\/p>\n\n<pre><code>curl -O http:\/\/download.tensorflow.org\/example_images\/flower_photos.tgz\n\ntar xzf flower_photos.tgz\n<\/code><\/pre>\n\n<p>It worked without any problems<\/p>\n"}]}
{"QuestionId":42268463,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-16T08:23:32.877","AcceptedAnswerId":null,"OwnerUserId":4626108.0,"Title":"Different buckets give different outputs at first time step during decoding","Body":"<p><strong>Background<\/strong><\/p>\n\n<p>I'm trying to build a chat bot and have a basic understanding of tensorflow Sequence2Sequence API by reading from here:\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/seq2seq\/<\/a><\/p>\n\n<p>I've also read related papers for Neural translation, attention mechanism during encoding\/decoding etc.<\/p>\n\n<p>RNN translate code is taken from here:\nhttp@s:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/translate\/seq2seq_model.py<\/p>\n\n<p>Bucketing\/Attention mechanism code is picked from here:\nhttp@s:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/legacy_seq2seq\/python\/ops\/seq2seq.py<\/p>\n\n<p>During my training(attention mechanism, GRU cell), I gave my buckets as [(2,2),(4,4),(6,6),(,8,8),(10,10)].<\/p>\n\n<p><strong>Issue<\/strong><\/p>\n\n<p>During decoding, if I force a bucket_index on the same input I'm getting totally different outputs at the first time step. For example:<\/p>\n\n<p>Input : How're you ?\nBucket index : 1<\/p>\n\n<p>Output : NAME UNK NAME<\/p>\n\n<p>Input : How're you ?\nBucket index : 2<\/p>\n\n<p>Output : Hi NAME to the<\/p>\n\n<p><strong>Observations<\/strong><\/p>\n\n<p>I can fix the bad outputs at each time step using beam-search. However, why I'm getting different word as output at the first time step during decoding ? Shouldn't the smaller bucket output be a subset of larger bucket output ?<\/p>\n\n<p>I've tried searching online and everywhere it's mentioned that parameters are shared across various buckets. Buckets are used for training efficiency and not for model tuning. I've also verified that my trainable parameters are common across all the buckets.<\/p>\n\n<p>The biggest difference I can see between encoder inputs for various buckets is the extra padding at the very start. We are doing attention mechanism on encoded_states, and softmax weights are already learned for those encoded states during training. Therefore, Can those softmax weights for those extra padded inputs cause sufficient difference leading to different output during first time step of decoding ?<\/p>\n\n<p>Did any one else also encounter the above issue? Any help for fixing the above error would be greatly appreciated.<\/p>\n\n<p>I've also opened an issue here because I'm unsure whether it is a bug(file on github), or my mis-understanding(Stackoverflow):\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7561\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7561<\/a><\/p>\n","answers":[{"AnswerId":"44597201","CreationDate":"2017-06-16T20:08:07.040","ParentId":null,"OwnerUserId":"7845325","Title":null,"Body":"<p>I am not sure if you have figured out an answer for this. I would like to comment on this in case someone else wants an idea on this.<\/p>\n\n<p>I believe this is the expected behavior. If you check the source implementation of model_with_buckets function in TensorFlow, you see roughly the following lines of code:<\/p>\n\n<pre><code>for j, bucket in enumerate(buckets):\n    with variable_scope.variable_scope(\n        variable_scope.get_variable_scope(), reuse=True if j &gt; 0 else None):\n        bucket_outputs, _ = seq2seq(encoder_inputs[:bucket[0]],\n                                    decoder_inputs[:bucket[1]])\n        outputs.append(bucket_outputs)\n<\/code><\/pre>\n\n<p>If you use 5 buckets, then this above outputs list will contain 5 sub-lists. In each of sub-list, it contains the decoder_outputs for each bucket. This outputs will be returned, and will be used both for training and prediction. You can see that each decoder_output (singular form) in different buckets are normally different as they are trained to generate outputs that match different targets for different buckets (although the variables in different buckets are re-used).<\/p>\n\n<p>One interesting test you can do is: add to your training set a pair with short question but long answer (make sure your model is big enough to remember almost every training samples). This means, this pair will use a larger bucket in the training, but a smaller bucket in the prediction. You also need a mechanism to type in your pad token in string format, say, <em>pad<\/em>. After your model is trained, perform such a prediction test. Type in the original answer, you will very likely get a totally different\/meaningless answer. However, you then type in your question followed with a number of <em>pad<\/em> until the input sentence is long enough so that the desired bucket can be chosen. You will then very likely get the same exact answer in that new training pair.<\/p>\n\n<p>Hope this explanation makes sense.<\/p>\n"}]}
{"QuestionId":42269137,"AnswerCount":1,"Tags":"<spring><jpa><spring-data><caffeine>","CreationDate":"2017-02-16T08:57:45.693","AcceptedAnswerId":null,"OwnerUserId":1115139.0,"Title":"Caffeine cache key in spring data jpa","Body":"<p>I just tried to store the result of items in the cache and all went well except the key. <\/p>\n\n<p>The SimpleKeyGenerator usually save the object or the result based on the arguments and in my case, I want to store an object in the cache and the key should be an attribute of the object inside the list here is an example. <\/p>\n\n<pre><\/pre>\n\n<p>now I want to store each object in the cache based on item.id, I know we can use the key attribute of the annotation but how can I access the Id for each item using SpEL. \nUnfortunately, create custom key generator will not help because all based on the arguments, target class and method. <\/p>\n\n<p>any recommendations?<\/p>\n","answers":[{"AnswerId":"42398534","CreationDate":"2017-02-22T17:36:46.700","ParentId":null,"OwnerUserId":"1125191","Title":null,"Body":"<p>I believe that you want to optimize cache storage to query by item id later.<\/p>\n\n<p>You don't need to do it if you configure JPA second level correctly to your Entity, when it is well configured load entity by ID using JPA will use second level automatically.<\/p>\n\n<p>To single result just use Cache SpEL #result on key attribute:<\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/docs.spring.io\/spring\/docs\/current\/spring-framework-reference\/html\/cache.html#cache-spel-context-tbl\" rel=\"nofollow noreferrer\">https:\/\/docs.spring.io\/spring\/docs\/current\/spring-framework-reference\/html\/cache.html#cache-spel-context-tbl<\/a><\/p>\n<\/blockquote>\n\n<pre><code>@Cacheable(cacheNames=\"item\", key=\"#result.id\")\n<\/code><\/pre>\n"}]}
{"QuestionId":42269153,"AnswerCount":1,"Tags":"<python><matrix><machine-learning><keras>","CreationDate":"2017-02-16T08:58:54.660","AcceptedAnswerId":42270127.0,"OwnerUserId":7507262.0,"Title":"hdf5 files containing matrix N x M as input for machine learning with keras","Body":"<p>I need help\/opinion regarding the input for doing machine learning.<\/p>\n\n<p>I have one hdf5 file <\/p>\n\n<pre><\/pre>\n\n<p>that contains several dataset where each dataset contains matrix of 4 X 125 (4 row and 125 column)<\/p>\n\n<p>I have been searching on how to input the hdf5 file containing matrix as input for machine learning but did not found any answer\/example that good enough for me as reference . Below is my modified script<\/p>\n\n<pre><\/pre>\n\n<p>I ran this script and got error like below<\/p>\n\n<pre><\/pre>\n\n<p>My question is how to load the hdf5 files containing datasets in form of matrix as input data for machine learning?<\/p>\n\n<p>*Additional question, I want to link the output for each dataset. For example, dataset 1 has output data of \"A\", dataset 2 has output data of \"CFL\" and so on...\nIs there any proper way to do this or is it enough with what I wrote in the script?<\/p>\n\n<p>Thank you for your help.<\/p>\n","answers":[{"AnswerId":"42270127","CreationDate":"2017-02-16T09:40:53.473","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>This should be for every named dataset:<\/p>\n\n<pre><code>X = dataset['dataset_name'][:] #input\n<\/code><\/pre>\n\n<p>to find out the names of the dataset you can use the following script:<\/p>\n\n<pre><code>for key in dataset.keys():\n    print key\n<\/code><\/pre>\n\n<p>Here you take the full dataset out of the file. If you do not slice the dataset you only get a pointer to the h5py dataset object. With the slice operator you extract the numpy array. (<a href=\"http:\/\/docs.h5py.org\/en\/latest\/high\/dataset.html#reading-writing-data\" rel=\"nofollow noreferrer\">http:\/\/docs.h5py.org\/en\/latest\/high\/dataset.html#reading-writing-data<\/a>)<\/p>\n\n<p>You just have to imagine your hdf5 file to be like a folder structure you have to traverse the folder structure to get to your data. Then you can use slicing to load the data.<\/p>\n"}]}
{"QuestionId":42269430,"AnswerCount":1,"Tags":"<tensorflow><gpu><cpu>","CreationDate":"2017-02-16T09:11:48.843","AcceptedAnswerId":null,"OwnerUserId":5938762.0,"Title":"Tensorflow doesn't see GPU and uses CPU instead, how come?","Body":"<p>I'm running a Python script using GPU-enabled Tensorflow. However, the program doesn't seem to recognize any GPU and starts using CPU straight away. What could be the cause of this?<\/p>\n","answers":[{"AnswerId":"47712811","CreationDate":"2017-12-08T10:50:20.510","ParentId":null,"OwnerUserId":"712995","Title":null,"Body":"<p>Just want to add to the discussion that tensorflow <em>may stop<\/em> seeing a GPU due to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2882\" rel=\"nofollow noreferrer\">CUDA initialization failure<\/a>, in other words, tensorflow detects the GPU, but can't dispatch any op onto it, so it falls back to CPU. In this case, you should see an error in the log, like this<\/p>\n\n<pre><code>E tensorflow\/stream_executor\/cuda\/cuda_driver.cc:481] failed call to cuInit: CUDA_ERROR_UNKNOWN\n<\/code><\/pre>\n\n<p>The cause is likely to be the conflict between different processes using GPU simultaneously. When that is the case, the most reliable way I found to get tensorflow working is to restart the machine. In the worst case, reinstalling tensorflow and \/ or NVidia driver.<\/p>\n\n<p>See also <a href=\"https:\/\/stackoverflow.com\/q\/46966402\/712995\">one more case<\/a> when GPU suddenly stops working.<\/p>\n"}]}
{"QuestionId":42270132,"AnswerCount":0,"Tags":"<python><ubuntu><import><tensorflow><pycharm>","CreationDate":"2017-02-16T09:41:17.777","AcceptedAnswerId":null,"OwnerUserId":5905757.0,"Title":"TensorFlow not importing when using PyCharm from Desktop Entry or Launcher","Body":"<p>TensorFlow is installed properly on my computer and I can import it and run code with it fine in terminals, PyCharm (opened with pycharm.sh command from a terminal window), and other editors.<\/p>\n\n<p>However, if I start PyCharm from Launcher, it gives me the following error when importing TensorFlow<\/p>\n\n<pre><\/pre>\n\n<p>I have deleted jetbrains-pycharm-ce.desktop and re-created it through Tools -> Create Desktop Entry multiple times.  I also tried opening up PyCharm with pycharm.sh from the terminal and locking the opened application to Launcher.  No matter what I do, PyCharm opened through any of the GUI methods is not able to import TensorFlow.<\/p>\n\n<p>I compared all the run settings and everything in PyCharm and couldn't really find any differences.<\/p>\n\n<p>What could it be that does not allow me to import TensorFlow when PyCharm is not opened through the command line?<\/p>\n\n<p>Below is the entire block of error code I get in PyCharm.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42270739,"AnswerCount":7,"Tags":"<python><tensorflow>","CreationDate":"2017-02-16T10:07:05.053","AcceptedAnswerId":42272349.0,"OwnerUserId":6873695.0,"Title":"How do I resolve these tensorflow warnings?","Body":"<p>I just installed Tensorflow 1.0.0 using pip. When running, I get warnings like the one shown below.<\/p>\n\n<p><\/p>\n\n<p>I get 5 more similar warning for SSE4.1, SSE4.2, AVX, AVX2, FMA.<\/p>\n\n<p>Despite these warnings the program seems to run fine.<\/p>\n","answers":[{"AnswerId":"42975902","CreationDate":"2017-03-23T12:21:41.907","ParentId":null,"OwnerUserId":"5825372","Title":null,"Body":"<blockquote>\n  <p><code>export TF_CPP_MIN_LOG_LEVEL=2<\/code> solved the problem for me on Ubuntu.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7778\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7778<\/a><\/p>\n"},{"AnswerId":"43017673","CreationDate":"2017-03-25T14:36:30.733","ParentId":null,"OwnerUserId":"4190721","Title":null,"Body":"<p>My proposed way to solve the problem:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python3\nimport os\nimport tensorflow as tf\nos.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'\n<\/code><\/pre>\n\n<p>Should work at least on any Debian or Ubuntu systems.<\/p>\n"},{"AnswerId":"42272349","CreationDate":"2017-02-16T11:14:20.617","ParentId":null,"OwnerUserId":"6824263","Title":null,"Body":"<p>I don't know much about C, but I found this<\/p>\n\n<pre><code>bazel build --linkopt='-lrt' -c opt --copt=-mavx --copt=-msse4.2 --copt=-msse4.1 --copt=-msse3-k \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>How you build you program?<\/p>\n"},{"AnswerId":"42539825","CreationDate":"2017-03-01T18:58:20.923","ParentId":null,"OwnerUserId":"137783","Title":null,"Body":"<p>It would seem that the PIP build for the GPU is bad as well as I get the warnings with the GPU version and the GPU installed...<\/p>\n"},{"AnswerId":"42642759","CreationDate":"2017-03-07T07:39:40.243","ParentId":null,"OwnerUserId":"7515188","Title":null,"Body":"<p>Those are simply warnings. \nThey are just informing you if you build TensorFlow from source it can be faster on your machine. <\/p>\n\n<p>Those instructions are not enabled by default on the builds available I think to be compatible with more CPUs as possible.<\/p>\n"},{"AnswerId":"49622736","CreationDate":"2018-04-03T05:19:34.170","ParentId":null,"OwnerUserId":"7745946","Title":null,"Body":"<p>As the warnings say you should only compile TF with these flags if you need to make TF faster.<\/p>\n\n<p>You can use TF environment variable <code>TF_CPP_MIN_LOG_LEVEL<\/code> and it works as follows:<\/p>\n\n<ul>\n<li>It defaults to 0, displaying all logs<\/li>\n<li>To filter out <code>INFO<\/code> logs set it to 1<\/li>\n<li><code>WARNINGS<\/code> additionally, 2<\/li>\n<li>and to additionally filter out <code>ERROR<\/code> logs set it to 3<\/li>\n<\/ul>\n\n<p>So you can do the following to silence the warnings:<\/p>\n\n<pre><code>import os\nos.environ['TF_CPP_MIN_LOG_LEVEL']='2'\nimport tensorflow as tf\n<\/code><\/pre>\n"},{"AnswerId":"42324649","CreationDate":"2017-02-19T07:12:40.190","ParentId":null,"OwnerUserId":"629247","Title":null,"Body":"<p>It seems that even if you don't have a compatible (i.e. Nvidia) GPU, you can actually still install the precompiled package for <a href=\"https:\/\/pypi.python.org\/pypi\/tensorflow-gpu\" rel=\"nofollow noreferrer\">tensorflow-gpu<\/a> via <code>pip install tensorflow-gpu<\/code>. It looks like in addition to the GPU support it also supports (or at least doesn't complain about) the CPU instruction set extensions like SSE3, AVX, etc. The only downside I've observed is that the Python wheel is a fair bit larger: 90MB for tensorflow-gpu instead of 42MB for plain tensorflow.<\/p>\n\n<p>On my machine without an Nvidia GPU I've confirmed that tensorflow-gpu 1.0 runs fine without displaying the cpu_feature_guard warnings.<\/p>\n"}]}
{"QuestionId":42271315,"AnswerCount":0,"Tags":"<python><cuda><neural-network><theano><keras>","CreationDate":"2017-02-16T10:31:48.320","AcceptedAnswerId":null,"OwnerUserId":3982193.0,"Title":"How to get transfer time from cpu to gpu in Python Keras","Body":"<p>I build a neural network model by using Keras in Python 2.7\nand I applied CNN and ANN to this dataset.<\/p>\n\n<p>I got running time of CPU and GPU and Of course, GPU is faster than CPU.\nbut I want to get data transfer time from CPU to GPU and running time of each devices(CPU and GPU).<\/p>\n\n<p>I searched on internet but I couldn't get satisfactory answer.<\/p>\n\n<p>Here is my ANN code<\/p>\n\n<pre><\/pre>\n\n<p>Please, help me out<\/p>\n","answers":[]}
{"QuestionId":42271525,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><conv-neural-network><cross-entropy>","CreationDate":"2017-02-16T10:40:43.557","AcceptedAnswerId":null,"OwnerUserId":4063857.0,"Title":"Relationship between Crossentroy Loss and Accurary","Body":"<p>I am training some CNN on an image classification task.\nOn a simple version this worked fine, but when I made the images more difficult I now encounter this phenomen (I let it train over night):<\/p>\n\n<p>While training, the training crossentropy loss goes down. Also, on my test dataset the crossentropy loss goes down. I am further measuring accuracy on it, which behaves differently. In the beginning it went up, only do go down again, and then it was kind of wavering between 0.1 and 0.3.\nI was expecting the crossentropy loss and the accurary to be somewhat related - since they are both measured on the same, the test dataset.<\/p>\n\n<p>Can somebody explain this to me? Or do I have a mistake in my code?<\/p>\n\n<p>Thanks a lot<\/p>\n","answers":[{"AnswerId":"42282169","CreationDate":"2017-02-16T18:42:10.337","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>The cross-entropy is not always directly related to the error metric. Usually it's correlated well enough to the error rate. Another typical choice is to minimize the Bayes risk. The Bayes risk is simply the expectation, with respect to your model, of the error (conversely accuracy). This is a continuous loss and should correlate better with your error rate. Also, measuring training error is usually a good metric to track.<\/p>\n"}]}
{"QuestionId":42272231,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-16T11:09:29.530","AcceptedAnswerId":null,"OwnerUserId":1960319.0,"Title":"GPU load in tensorflow","Body":"<p>I just built TensorFlow v1.0 and I am trying to run MNIST test just to see if it's working. Seems like it is, but i am observing weird behaiviour. \nMy system has two Tesla P100, and nvidia-smi shows the following:<\/p>\n\n<pre><\/pre>\n\n<p>As it shown, python3 ate all the memory on both GPUs, but computational load are placed only on first.<\/p>\n\n<p>Exporting CUDA_VISIBLE_DEVICES I can limit GPU to be used, but it's not affect computational time. So no gain from adding second GPU. Single GPU<\/p>\n\n<pre><\/pre>\n\n<p>Two GPUs:<\/p>\n\n<pre><\/pre>\n\n<p>So the question is, how to load both GPUs?<\/p>\n","answers":[]}
{"QuestionId":42272503,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-02-16T11:20:43.790","AcceptedAnswerId":42275498.0,"OwnerUserId":2784721.0,"Title":"MNIST neural netowrk : Accuracy is pretty low","Body":"<p>I am using MNIST data set to learn tensorflow and neural network. Below is my code in python.<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the accuracy of 9% to 13% and not more than that. I think I have implemented the code correctly but not able to figure out what is wrong. One thing I figured out is that the accuracy is because the model is predicting only 0s correctly.\nThanks in advance!<\/p>\n","answers":[{"AnswerId":"42275498","CreationDate":"2017-02-16T13:35:23.547","ParentId":null,"OwnerUserId":"2784721","Title":null,"Body":"<p>I have done mistake in calculating the output of the network,<\/p>\n\n<p>Wrong:<\/p>\n\n<pre><code>output = tf.add(tf.matmul( l3, output_layer['weights']), output_layer['biases'])\noutput = tf.nn.relu(output)\n<\/code><\/pre>\n\n<p>Correct:<\/p>\n\n<pre><code>output = tf.add(tf.matmul( l3, output_layer['weights']), output_layer['biases'])\n<\/code><\/pre>\n\n<p>I was normalizing the output again which was messing up all the network. Posting this answer as it may be helpful to someone in future. Thanks!<\/p>\n\n<p>PS : Borrowed code from <a href=\"https:\/\/pythonprogramming.net\/tensorflow-neural-network-session-machine-learning-tutorial\/?completed=\/tensorflow-deep-neural-network-machine-learning-tutorial\/\" rel=\"nofollow noreferrer\">sentdex<\/a><\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I have found that the accuracy can be further improved by using <a href=\"https:\/\/en.wikipedia.org\/wiki\/Convolutional_neural_network\" rel=\"nofollow noreferrer\">CNN<\/a> and even further more by using <a href=\"https:\/\/en.wikipedia.org\/wiki\/Recurrent_neural_network\" rel=\"nofollow noreferrer\">RNN<\/a>. Probably someone will find this useful.<\/p>\n"}]}
{"QuestionId":42273246,"AnswerCount":1,"Tags":"<tensorflow><embedding>","CreationDate":"2017-02-16T11:55:21.150","AcceptedAnswerId":null,"OwnerUserId":4362500.0,"Title":"How does tensorflow implement the embedding_column?","Body":"<p>I'm learning tensorflow's wide_n_deep_tutorial these days, and I'm a little bit confused with the <em>tf.contrib.layers.embedding_column<\/em>. I wonder how does tensorflow implement the embedding column? <\/p>\n\n<p>For example, suppose I have an sparse input with dimension 1000 and I want to embed it into a dense feature with dimension 10. Does it hold a fully connected network with 1000*10 params and train using BP to update the params? Or does it use some other techniques like FM to map the 1000 dim vector to a 10 dim vector?<\/p>\n","answers":[{"AnswerId":"42273512","CreationDate":"2017-02-16T12:07:49.727","ParentId":null,"OwnerUserId":"6824263","Title":null,"Body":"<p>There are 3 combiner in the embedding_column function:<\/p>\n\n<p>\"sum\": do not normalize\n\"mean\": do l1 normalization\n\"sqrtn\": do l2 normalization. see more tf.embedding_lookup_sparse<\/p>\n\n<p>There are not using FM to modulate\/transform the dimensions.<\/p>\n"}]}
{"QuestionId":42273323,"AnswerCount":2,"Tags":"<python><tensorflow><jupyter-notebook>","CreationDate":"2017-02-16T11:58:37.773","AcceptedAnswerId":42357963.0,"OwnerUserId":4838266.0,"Title":"Error importing TensorFlow GPU in Jupyter Notebook","Body":"<p>I am trying to import TensorFlow but end up with this error.<\/p>\n\n<pre><\/pre>\n\n<p>I am able to import the CPU Version of the TensorFlow with no issue. I am also able to import TensorFlow in Python Console launched from Terminal. The path to python executable is same in both Console and Jupyter Notebook ( gave same output in Console and Jupyter Notebook). <\/p>\n\n<p>TensorFlow version - 1.0.0<\/p>\n\n<p>Python Version - 3.5.2<\/p>\n\n<p>Cuda Version - 8.0<\/p>\n","answers":[{"AnswerId":"42357963","CreationDate":"2017-02-21T03:13:38.853","ParentId":null,"OwnerUserId":"4838266","Title":null,"Body":"<p>Looks like the error was when I setup CUDA and CuDNN. Following the guide provided by Nvidia, it told me to setup the path for <code>DYLD_LIBRARY_PATH<\/code> for Mac OS X. Unfortunately, this does not work when using Jupyter Notebook. We need to follow the steps the Nvidia's guide has for Linux and setup the <code>LD_LIBRARY_PATH<\/code>. This solved the issue.<\/p>\n"},{"AnswerId":"47326568","CreationDate":"2017-11-16T09:57:24.993","ParentId":null,"OwnerUserId":"3342058","Title":null,"Body":"<p>In my case when starting Jupyter notebook the <code>LD_LIBRARY_PATH<\/code> was cleared even if exported it in <code>.bashrc<\/code> and <code>.profile<\/code> on Ubuntu. Only starting the notebook server with <code>env VARIABLE=$VARIABLE jupyter notebook<\/code> solved it.<\/p>\n"}]}
{"QuestionId":42274018,"AnswerCount":0,"Tags":"<python><tensorflow><virtualenv><python-3.5>","CreationDate":"2017-02-16T12:29:17.487","AcceptedAnswerId":null,"OwnerUserId":2884988.0,"Title":"Virtualenv setup returns ImportError","Body":"<p>I am trying to install tensorflow with virtualenv, but I get the following error when I run \"virtualenv --system-site-packages tensorflow\". I am currently working this on macOS with python3.5.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42274088,"AnswerCount":0,"Tags":"<parallel-processing><tensorflow>","CreationDate":"2017-02-16T12:32:22.057","AcceptedAnswerId":null,"OwnerUserId":886724.0,"Title":"Tensorflow data parallelism","Body":"<p>I am trying to implement data parallelism with tensorflow using the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">eexample replicated training documentation here<\/a><\/p>\n\n<p>In the example code of the link above, the ps jobs seems to do nothing but blocking on the join operation:<\/p>\n\n<pre><\/pre>\n\n<p>All the data parallelism of the computation can be taken care of by the supervisor. So what is the role of the parameter server? and can it be stopped gracefully after the jobs are done?<\/p>\n","answers":[]}
{"QuestionId":42274599,"AnswerCount":1,"Tags":"<tensorflow><windows-10>","CreationDate":"2017-02-16T12:55:13.833","AcceptedAnswerId":null,"OwnerUserId":7574720.0,"Title":"tensorflow Error after successful installation in windows 10","Body":"<p>After successful installation of Python and TensorFlow, when I ran the given program to test my installation, I got the following errors<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DPboH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DPboH.png\" alt=\"enter image description here\"><\/a> <\/p>\n","answers":[{"AnswerId":"42301878","CreationDate":"2017-02-17T15:39:26.100","ParentId":null,"OwnerUserId":"523503","Title":null,"Body":"<p>There are just warnings. Don't worry, you can ignore them unless you need to use some of these operators: BestSplits,CountExtremelyRandomStats,...<\/p>\n"}]}
{"QuestionId":42274756,"AnswerCount":1,"Tags":"<python><machine-learning><3d><tensorflow><scikit-learn>","CreationDate":"2017-02-16T13:02:01.840","AcceptedAnswerId":null,"OwnerUserId":6205024.0,"Title":"How to analyse 3d mesh data(in .stl) by TensorFlow","Body":"<p>I try to write an script in python for analyse an .stl data file(3d geometry) and say which model is convex or concave and watertight and tell other properties...<\/p>\n\n<p>I would like to use and TensorFlow, scikit-learn or other machine learning library. Create some database with examples of objects with tags and in future add some more examples and just re-train model for better results.<\/p>\n\n<p>But my problem is: I don\u00b4t know how to recalculate or restructure 3d data for working in ML libraries. I have no idea.<\/p>\n\n<p>Thank you for your help.<\/p>\n","answers":[{"AnswerId":"42284733","CreationDate":"2017-02-16T21:13:27.897","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>You have to first extract \"features\" out of your dataset. These are fixed-dimension vectors. Then you have to define labels which define the prediction. Then, you have to define a loss function and a neural network. Put that all together and you can train a classifier.<\/p>\n\n<p>In your example, you would first need to extract a fixed dimension vector out of an object. For instance, you could extract the object and project it on a fixed support on the x, y, and z dimensions. That defines the features.<\/p>\n\n<p>For each object, you'll need to label whether it's convex or concave. You can do that by hand, analytically, or by creating objects analytically that are known to be concave or convex. Now you have a dataset with a lot of sample pairs (object, is-concave).<\/p>\n\n<p>For the loss function, you can simply use the negative log-probability.<\/p>\n\n<p>Finally, a feed-forward network with some convoluational layers at the bottom is probably a good idea.<\/p>\n"}]}
{"QuestionId":42274786,"AnswerCount":0,"Tags":"<tensorflow><neural-network><computer-vision><deep-learning><tensorboard>","CreationDate":"2017-02-16T13:03:34.323","AcceptedAnswerId":null,"OwnerUserId":988709.0,"Title":"how to visualize feature map of CNN for tensorflow?","Body":"<p>I want to see the convoluation result of all the kernals. Another say, I want to see the low level and hight level features of a CNN model. How can do this with tensorflow? Any example for that? <\/p>\n\n<p>Here is a example:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/kRfuH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/kRfuH.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
{"QuestionId":42275324,"AnswerCount":1,"Tags":"<python><flask><tensorflow><flask-restful>","CreationDate":"2017-02-16T13:27:03.660","AcceptedAnswerId":null,"OwnerUserId":654203.0,"Title":"How to avoid reading the graph for each request in Tensorflow","Body":"<p>I have a Tensorflow python class that is called from a REST endpoint and passed in a URL for an image. Each time a new request is initiated, it calls the  method which reads a local  file. This file does not change from request to request. So, I feel it is not a good use of resources and time to read this file at each request. <\/p>\n\n<p>The code is below:<\/p>\n\n<pre><\/pre>\n\n<p>The above code is called from  like this:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Question<\/strong><\/p>\n\n<p>Is there any way to only call  at the first request and then not call it until the service is restarted?<\/p>\n","answers":[{"AnswerId":"42301783","CreationDate":"2017-02-17T15:35:12.883","ParentId":null,"OwnerUserId":"523503","Title":null,"Body":"<p>For serving: Create the session only once per process. You can call Session.run() many times.<\/p>\n"}]}
{"QuestionId":42275815,"AnswerCount":1,"Tags":"<tensorflow><neural-network><keras><deep-learning><theano>","CreationDate":"2017-02-16T13:49:37.900","AcceptedAnswerId":42276427.0,"OwnerUserId":6719992.0,"Title":"Should I substract imagenet pretrained inception_v3 model mean value at inception_v3.py keras?","Body":"<pre><\/pre>\n\n<p>&emsp;I am using <strong>keras<\/strong> inception_v3 imagenet pretrained model(<a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/applications\/inception_v3.py\" rel=\"nofollow noreferrer\">inception_v3.py<\/a>) to finetune on my own dataset.<br>\n&emsp;When I want to <strong>subtract the imagenet mean value [123.68, 116.779, 103.939] and reverse axis RGB to BGR<\/strong> as we often do, I find that the author provided a _preprocess_input()_ function at the end.I am confused about this.  <\/p>\n\n<p>&emsp; Should I use the provided function <em>preprocess_input()<\/em> <strong>or<\/strong> subtract mean value and reverse axis as usual?<br>\n&emsp; Thanks lot.<\/p>\n","answers":[{"AnswerId":"42276427","CreationDate":"2017-02-16T14:14:57.067","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Actually in a original <a href=\"https:\/\/www.cs.unc.edu\/~wliu\/papers\/GoogLeNet.pdf\" rel=\"nofollow noreferrer\">Inception<\/a> paper the autors mention as a data preprocessor the function you provided (one which is zero-centering all channels and resizes it to <code>[-1, 1]<\/code> interval). As in <a href=\"https:\/\/arxiv.org\/abs\/1512.00567\" rel=\"nofollow noreferrer\">InceptionV3<\/a> paper no new data transformation is provided I think that you may assume that you should use the following function:<\/p>\n\n<pre><code>def preprocess_input(x):\n    x \/= 255.\n    x -= 0.5\n    x *= 2.\n    return x\n<\/code><\/pre>\n"}]}
{"QuestionId":42276544,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-16T14:20:20.670","AcceptedAnswerId":null,"OwnerUserId":7575272.0,"Title":"What is the simplest way to convert vector to Toeplitz matrix in TensorFlow?","Body":"<p>I would like to convert a vector to a symmetric Toeplitz matrix using Tensorflow operations like this:<\/p>\n\n<pre><\/pre>\n\n<p>where the shape of A is [vector_size, vector_size]. The relation between the two variables is as below.<\/p>\n\n<pre><\/pre>\n\n<p>What is the simplest way to do it?<\/p>\n","answers":[{"AnswerId":"42281677","CreationDate":"2017-02-16T18:14:43.403","ParentId":null,"OwnerUserId":"4234099","Title":null,"Body":"<p>In case <code>vector_size=3<\/code>:<\/p>\n\n<pre><code>&gt;&gt;&gt; a = tf.placeholder(tf.float32, shape=[vector_size])\n&gt;&gt;&gt; A = [[a[0],a[1],a[2]],[a[1],a[0],a[1]],[a[2],a[1],a[0]]]\n&gt;&gt;&gt; sess = tf.Session()\n&gt;&gt;&gt; sess.run(A, {a: [1, 2, 3]})\n[[1.0, 2.0, 3.0], [2.0, 1.0, 2.0], [3.0, 2.0, 1.0]]\n<\/code><\/pre>\n"}]}
{"QuestionId":42276811,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-16T14:31:31.400","AcceptedAnswerId":42277241.0,"OwnerUserId":5718928.0,"Title":"TensorFlow computation results slightly vary","Body":"<p>Since this <a href=\"https:\/\/stackoverflow.com\/questions\/39429313\/why-do-tensorflow-tf-learn-classification-results-vary-a-lot\">link<\/a> only describes the variety of predictions based on learning approaches, I want to find out of curiosity why TensorFlow computations <strong>slightly vary<\/strong>. <\/p>\n\n<pre><\/pre>\n\n<p>Expected ouput:<\/p>\n\n<pre><\/pre>\n\n<p>Actual output:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42277241","CreationDate":"2017-02-16T14:49:04.723","ParentId":null,"OwnerUserId":"1642057","Title":null,"Body":"<p>This is normal for 32-bit floating point values.  Those 1.1, 2,2 and 3.3 values are not exactly represented in 32-bit floating point.<\/p>\n\n<pre><code>import numpy as np\nx = np.array([1.1, 2.2, 3.3], dtype=np.float32)\ny = np.array([1, 1, 1], dtype=np.float32)\nx+y\n\n&gt;&gt;&gt; array([ 2.0999999 ,  3.20000005,  4.30000019], dtype=float32)\n<\/code><\/pre>\n\n<p>If you haven't read it, you might want to google \"What Every Computer Scientist Should Know About Floating-Point Arithmetic\" to understand better the limitations.<\/p>\n"}]}
{"QuestionId":42276984,"AnswerCount":2,"Tags":"<python><c++><machine-learning><theano>","CreationDate":"2017-02-16T14:38:01.633","AcceptedAnswerId":null,"OwnerUserId":7244583.0,"Title":"'::hypot' has not been declared","Body":"<p>I'm using python3.6 theano, \nwith mingw-w64-x86-64 installed, my os is Win10_64, cuda installed, \nand seems everything is ok <\/p>\n\n<p>the theano.test() is ok, saying my gpu is working,<\/p>\n\n<p>but it just keeps tell me that \"error: '::hypot' has not been declared\"<\/p>\n\n<pre><\/pre>\n\n<p>Any help would be appreciated!<\/p>\n","answers":[{"AnswerId":"42280324","CreationDate":"2017-02-16T17:03:00.317","ParentId":null,"OwnerUserId":"251738","Title":null,"Body":"<p>My guess from your incomplete information is that you aren't compiling in C++11 mode so you aren't picking up <code>::hypot<\/code> from C99.<\/p>\n"},{"AnswerId":"47058517","CreationDate":"2017-11-01T15:40:20.627","ParentId":null,"OwnerUserId":"2602531","Title":null,"Body":"<p>I had this error with building an python file using mingw32 .\nI opened the file that it says (C:\/mingw64\/lib\/gcc\/x86_64-w64-mingw32\/6.3.0\/include\/c++\/cmath:1157:11)<br>\nand changed that line to  <\/p>\n\n<pre><code>using ::_hypot;\n<\/code><\/pre>\n\n<p>or adding this line just before that :  <\/p>\n\n<pre><code>#define hypot _hypot\n<\/code><\/pre>\n\n<p>and after that the problem was solved !!\nI know it's not a basic solution but it is the one that I could find !!<\/p>\n"}]}
{"QuestionId":42277280,"AnswerCount":6,"Tags":"<amazon-web-services><amazon-ec2><tcp><tensorflow><tensorboard>","CreationDate":"2017-02-16T14:50:14.623","AcceptedAnswerId":null,"OwnerUserId":3867406.0,"Title":"Accessing Tensorboard on AWS","Body":"<p>I'm trying to access Tensorboard on AWS. Here is my setting :<\/p>\n\n<ul>\n<li>Tensorboard :  :<\/li>\n<\/ul>\n\n<blockquote>\n  <p>Starting TensorBoard b'39' on port 6006 (You can navigate to\n  <a href=\"http:\/\/172.31.18.170:6006\" rel=\"noreferrer\">http:\/\/172.31.18.170:6006<\/a>)<\/p>\n<\/blockquote>\n\n<ul>\n<li>AWS Security groups (in): \n\n<ul>\n<li>HTTPS      TCP 443 0.0.0.0\/0<\/li>\n<li>Custom_TCP TCP 6006  0.0.0.0\/0<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>However connecting to  I can't see anything, I basically can't connect. <\/p>\n\n<p>Do you have any idea?<\/p>\n","answers":[{"AnswerId":"42476437","CreationDate":"2017-02-27T01:28:53.570","ParentId":null,"OwnerUserId":"7627331","Title":null,"Body":"<p>I suffered from the same problem for several days.<\/p>\n\n<p>Fortunately I solved this issue by adding rule on \"AWS Outbound rule\" as if I had added \"AWS Inbound rule\".<\/p>\n\n<hr>\n\n<p>Regardless of this setting, it works at home.\n<em>The same error is still happening only in the company.<\/em><\/p>\n"},{"AnswerId":"45587724","CreationDate":"2017-08-09T10:11:44.727","ParentId":null,"OwnerUserId":"562222","Title":null,"Body":"<p>I managed to set it up like this:<\/p>\n\n<ul>\n<li>Go to security groups in your ec2 console:<\/li>\n<\/ul>\n\n<p><img src=\"https:\/\/i.imgur.com\/Y7BEcoi.png\" alt=\"\"><\/p>\n\n<ul>\n<li><p>Choose the relevant security group in the table, click edit.<\/p><\/li>\n<li><p>Add a rule like this:<\/p><\/li>\n<\/ul>\n\n<p><img src=\"https:\/\/i.imgur.com\/O76EGzp.png\" alt=\"\"><\/p>\n\n<ul>\n<li><p>Start tensorboard: <code>tensorboard --logdir tf_summary\/ --port 8080\n<\/code><\/p><\/li>\n<li><p>Find out the URL of your instance and visit <code>http:\/\/yourURL:8080<\/code><\/p><\/li>\n<\/ul>\n"},{"AnswerId":"48259449","CreationDate":"2018-01-15T08:44:10.017","ParentId":null,"OwnerUserId":"1100060","Title":null,"Body":"<p>Simply run the tensorboard without the host parameter (which poses restrictions)<\/p>\n\n<pre><code>tensorboard --logdir XXX --port 6006\n<\/code><\/pre>\n"},{"AnswerId":"48649751","CreationDate":"2018-02-06T18:38:16.660","ParentId":null,"OwnerUserId":"3277442","Title":null,"Body":"<ol>\n<li><p>Run tensorboard in your ec2 terminal (you can custom logdir and port)<\/p>\n\n<pre><code>tensorboard --logdir=data\/model --port=8080\n<\/code><\/pre><\/li>\n<li><p>Find your workstations public ip (a.b.c.d) address by visiting <code>http:\/\/ip4.me\/<\/code><\/p><\/li>\n<li><p>Access the security group configuration assigned to your EC2 and add a custom TCP rule to your inbound traffic.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/iTIav.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/iTIav.png\" alt=\"Security group inbound traffic configuration\"><\/a><\/p><\/li>\n<li><p>Outbound should be set to allow traffic from tensorboard port. (In this case 8080). Or you just allow all outgoing traffic from your EC2 instance<\/p>\n\n<pre><code>Protocol       Port Range      Destination      Description   \nAll traffic    All             All               0.0.0.0\/0\n<\/code><\/pre><\/li>\n<li><p>Use your public DNS to access tensorboard from your workstation<\/p>\n\n<p><a href=\"http:\/\/ec2-xx-xxx-xx-xx.compute-1.amazonaws.com:8080\/\" rel=\"noreferrer\">http:\/\/ec2-xx-xxx-xx-xx.compute-1.amazonaws.com:8080\/<\/a><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"44265008","CreationDate":"2017-05-30T14:27:37.323","ParentId":null,"OwnerUserId":"7353970","Title":null,"Body":"<p><strong>Fast (but unsecure) solution:<\/strong><\/p>\n\n<p>Run: <\/p>\n\n<pre><code>tensorboard --logdir=\/training --host=0.0.0.0 --port=8080 \n<\/code><\/pre>\n\n<p>on your AWS instance.<\/p>\n\n<p>Make sure that both your inbound and outbound rules on AWS console (control center) are as unrestricted as possible (allow all types, all ports etc.). However, keep in mind that this solution is not recommendable for environments requiring security (in our case, we didn't consider security for training an NN).<\/p>\n\n<p>An attempt to explain why this works: when the policy is set as described, AWS still seems to prohibit inbound\/outbound connections on the standard tensorboard port 6006. This does not seem to apply to the port 8080.<\/p>\n\n<p><strong>Long (but more secure) solution:<\/strong>\nSee: <a href=\"https:\/\/blog.altoros.com\/getting-started-with-a-cpu-enabled-tensorflow-instance-on-aws.html\" rel=\"nofollow noreferrer\">https:\/\/blog.altoros.com\/getting-started-with-a-cpu-enabled-tensorflow-instance-on-aws.html<\/a>\n(provides explanations for setting ports correctly on AWS)<\/p>\n"},{"AnswerId":"47039100","CreationDate":"2017-10-31T15:31:54.623","ParentId":null,"OwnerUserId":"1644172","Title":null,"Body":"<p>You can use ssh tunneling technique.<\/p>\n\n<p>In your terminal you can write:<\/p>\n\n<pre><code>ssh -i \/path\/to\/your\/AWS\/key\/file -NL 6006:localhost:6006 user@host\n<\/code><\/pre>\n\n<p>where user and host are your aws ec2 user and instance specific.<\/p>\n\n<p>After that you can browse to <a href=\"http:\/\/localhost:6006\/\" rel=\"noreferrer\">http:\/\/localhost:6006\/<\/a><\/p>\n"}]}
{"QuestionId":42278648,"AnswerCount":2,"Tags":"<python><visual-studio><tensorflow>","CreationDate":"2017-02-16T15:50:45.723","AcceptedAnswerId":42278906.0,"OwnerUserId":5480127.0,"Title":"Tensorflow Visual Studio import error","Body":"<p>I've been trying to install Tensorflow for the past 6 hours, and it's still not working. \nGoogling resolved most issues including; the Install page on the website displaying the wrong wheel locations, getting CUDA loaded correctly, putting cuDNN.lib in the correct folders and NOT the ones described by NVidia.<\/p>\n\n<p>When I activate tensorflow in cmd, and run the example code to verify functionality, I get this:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/O3n2G.png\" alt=\"1\"><\/p>\n\n<p>You can see that the CUDA libraries are loaded correctly. If I then try to run this same script in the Spyder IDE that comes with Anaconda, I get:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/M3RuG.png\" alt=\"2]\"><\/p>\n\n<p>So far, so good. I don't want to use Spyder though, I'd prefer to keep using Visual Studio. So I installed the python package for VS, and after a lot of tinkering and googling, managed to get VS to recognize and find the Anaconda compiler. When I run the same code as before, however, this is the result:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/wt9Ne.png\" alt=\"3]\"><\/p>\n\n<p>The last image doesn't show what happens when I F5, but it's as you'd expect if VS doesn't recognize a type: It can't resolve it. This is how I set up Anaconda3 in VS's \"Python Environments\":<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/1z7Aj.png\" alt=\"4]\"><\/p>\n\n<p>Is there anything I can do, or is Visual Studio just not compatible with Anaconda and\/or tensorflow?<\/p>\n","answers":[{"AnswerId":"42278906","CreationDate":"2017-02-16T16:03:06.603","ParentId":null,"OwnerUserId":"6348508","Title":null,"Body":"<p>You have created a separate CONDA environment (tensorflow) to install tensorflow which is used when you are trying Spyder and conda console. <\/p>\n\n<p>But when you are accessing the Anaconda from Visual Studio it is only accessing the default ANACONDA environment. To select the tensorflow environment you need to specify the specific ENV <\/p>\n\n<pre><code>C:\\Users\\Tachyon\\Anaconda3\\envs\\tensorflow\\\nC:\\Users\\Tachyon\\Anaconda3\\envs\\tensorflow\\python.exe\nC:\\Users\\Tachyon\\Anaconda3\\envs\\tensorflow\\pythonw.exe\n<\/code><\/pre>\n"},{"AnswerId":"47416786","CreationDate":"2017-11-21T15:31:13.183","ParentId":null,"OwnerUserId":"8980558","Title":null,"Body":"<p>today I had the same issue and tried to use Tensorflow in VS. Like you I tried it round about 6 h. Then I typed in Tensorflow in the packege search. There is Tensorflow as a seperate installer in VS <a href=\"https:\/\/i.stack.imgur.com\/5YhK6.jpg\" rel=\"nofollow noreferrer\">Package Tensorflow<\/a>. After I installed Tensorflow, the App is running fine.<\/p>\n"}]}
{"QuestionId":42279055,"AnswerCount":1,"Tags":"<python-2.7><scikit-learn><keras><metrics><difference>","CreationDate":"2017-02-16T16:08:35.737","AcceptedAnswerId":null,"OwnerUserId":7544784.0,"Title":"Comparing metrics of Keras with metrics of sklearn.classification_report","Body":"<p>I am struggling with different metrics while evaluating neural networks. \nMy investigations showed that Keras (version 1.2.2) calculates different values for specific metrics (using function evaluate) compared to sklearn.classification report. <\/p>\n\n<p>Specifically, the values for the metric 'precision' (i.e. 'precision' of Keras != 'precision' of sklearn) or 'recall' (i.e. 'recall' of Keras != 'recall' of sklearn) differ. \nFor the following working example the differences seem to be random, but evaluating bigger networks shows that 'precision' of Keras equals (almost) 'recall' of sklearn whereas both 'recall' metrics differ clearly.<\/p>\n\n<p>I appreciate your help!<\/p>\n\n<pre><\/pre>\n\n<p><strong>E D I T<\/strong><\/p>\n\n<p>The output of the script given above:<\/p>\n\n<pre><\/pre>\n\n<p>For another model: <\/p>\n\n<pre><\/pre>\n\n<p>Definition of desired metrics (for model.compile):<\/p>\n\n<pre><\/pre>\n\n<p>Output of model.metrics_names:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42492562","CreationDate":"2017-02-27T17:54:21.437","ParentId":null,"OwnerUserId":"2426324","Title":null,"Body":"<p>Yes, it is different due to the fact that the sklearn classification report gives you the weighted average based on the support.<\/p>\n\n<p>Experiment with: <\/p>\n\n<pre><code>from sklearn.metrics import classification_report\ny_true = [0, 1,2,1]\ny_pred = [0, 0,2,0]\ntarget_names = ['class 0', 'class 1', 'class 2']\nprint(classification_report(y_true, y_pred, target_names=target_names))\n<\/code><\/pre>\n\n<p>Gives you:\n                 precision    recall  f1-score   support<\/p>\n\n<pre><code>    class 0       0.33      1.00      0.50         1\n    class 1       0.00      0.00      0.00         2\n    class 2       1.00      1.00      1.00         1\n\navg \/ total       0.33      0.50      0.38         **4**\n<\/code><\/pre>\n\n<p>However, (1+0+0.33)\/3 = 0.44(3), but as it seems from the support column sklearn returns (1*1+0*2+0.33*1)\/4=0.3325<\/p>\n"}]}
{"QuestionId":42279706,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-16T16:35:07.210","AcceptedAnswerId":null,"OwnerUserId":991703.0,"Title":"where do I find bidirectional_rnn in tensorflow 1.0.0?","Body":"<p>I am using some code from here: <a href=\"https:\/\/github.com\/monikkinom\/ner-lstm\" rel=\"nofollow noreferrer\">https:\/\/github.com\/monikkinom\/ner-lstm<\/a> with tensorflow. I think the code was written for an older version of tensorflow, I am using version 1.0.0. I used tf_upgrade.py to upgrade model.py in that github repos, but I am still getting the error:<\/p>\n\n<pre><\/pre>\n\n<p>this is after I changed the bidirectional_rnn call to use contrib_rnn which is:<\/p>\n\n<pre><\/pre>\n\n<p>The old call was <\/p>\n\n<pre><\/pre>\n\n<p>which also doesn't work.<\/p>\n\n<p>I had to change the LSTMCell, DroputWrapper, etc. to rnn.LSTMCell, but they seem to work fine. It is the bidirectional_rnn that I can't figure out how to change.<\/p>\n","answers":[{"AnswerId":"42280036","CreationDate":"2017-02-16T16:49:09.720","ParentId":null,"OwnerUserId":"4022997","Title":null,"Body":"<p>Maybe you can try to reimplement a bidirectional RNN by simply wrapping into a single class two monodirectional RNNs with the parameter \"go_backwards=True\" set on one of them. Then you can also have control over the type of merge done with the outputs. Maybe taking a look at the implementation in <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/layers\/wrappers.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/layers\/wrappers.py<\/a> (see the class <code>Bidirectional<\/code>) could get you started.<\/p>\n"},{"AnswerId":"42652090","CreationDate":"2017-03-07T15:25:24.083","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In TensorFlow 1.0, you have the choice of two bidirectional RNN functions:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/bidirectional_dynamic_rnn\" rel=\"nofollow noreferrer\"><code>tf.nn.bidirectional_dynamic_rnn()<\/code><\/a><\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\/static_bidirectional_rnn\" rel=\"nofollow noreferrer\"><code>tf.contrib.rnn.static_bidirectional_rnn()<\/code><\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":42279740,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-16T16:36:47.850","AcceptedAnswerId":null,"OwnerUserId":7370471.0,"Title":"Tensorboard Embedding with t-SNE","Body":"<p>When I run t-SNE on my datasets (on smallest or bigger one) with tensorboard embedding , the app crash all the times (after thereabout 500 iterations). \nI can't re-run or stop the operation to change the parameters, the iteration number freeze and the other functionalities on embedding part crash also.<\/p>\n\n<p>Anyone have the same problem ? \nOther curiosity is if I run on <a href=\"http:\/\/projector.tensorflow.org\" rel=\"nofollow noreferrer\">online tensorboard embedding<\/a> I have the same problem.<\/p>\n","answers":[{"AnswerId":"42374755","CreationDate":"2017-02-21T18:06:59.933","ParentId":null,"OwnerUserId":"7370471","Title":null,"Body":"<p>The problem is the browser. Initially I was using Safari and had problems in all tabs of tensorboard. Switch to Chrome solved the problem.<\/p>\n"}]}
{"QuestionId":42280240,"AnswerCount":1,"Tags":"<python><tensorflow><theano>","CreationDate":"2017-02-16T16:58:20.193","AcceptedAnswerId":42292587.0,"OwnerUserId":4393898.0,"Title":"Changing values in a theano tensor vector","Body":"<p>I have a theano tensor vector  something like , which is the output of the following function:<\/p>\n\n<pre><\/pre>\n\n<p> is a vector which has the same dimension as .<\/p>\n\n<p>How can I change one specific value to , and the others to  in  vector?<\/p>\n","answers":[{"AnswerId":"42292587","CreationDate":"2017-02-17T08:09:19.220","ParentId":null,"OwnerUserId":"2147347","Title":null,"Body":"<p>What value specifically you want to change? if you just want have a vector which have same dimension as <code>next_probs<\/code> you can use <code>zeros<\/code> and <code>set_subtensor<\/code> like below<\/p>\n\n<pre><code>ret = T.zeros(next_probs.shape)\nret = T.set_subtensor(ret[index],1)\n<\/code><\/pre>\n\n<p>If you want to use this in classification model, and you need the class with highest probability in <code>next_probs<\/code> become 1 and the others become 0 this is my other answer:<\/p>\n\n<pre><code>ret = T.zeros(next_probs.shape)\nret = T.set_subtensor(ret[T.argmax(next_probs)],1)\n<\/code><\/pre>\n\n<p><code>ret<\/code> is a vector with 1 in the class with highest probability, and 0 for the others<\/p>\n"}]}
{"QuestionId":42280894,"AnswerCount":4,"Tags":"<tensorflow><anaconda>","CreationDate":"2017-02-16T17:29:35.927","AcceptedAnswerId":null,"OwnerUserId":1239984.0,"Title":"Tensorflow Wheel Install not Supported","Body":"<pre><\/pre>\n\n<p>In Anaconda3, Im trying to install the wheel but it is not working. On Windows. Using Python 3.6. There isnt a 3.6 wheel. I get this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44545532","CreationDate":"2017-06-14T12:58:53.593","ParentId":null,"OwnerUserId":"3231635","Title":null,"Body":"<p>Use an alternative tensorflow wheel for Python 3.6: <\/p>\n\n<pre><code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.2.0rc2-cp36-cp36m-win_amd64.whl\n<\/code><\/pre>\n"},{"AnswerId":"43962956","CreationDate":"2017-05-14T10:44:19.370","ParentId":null,"OwnerUserId":"2285474","Title":null,"Body":"<p>Just change your python version to 3.5 and then retry installing tensorflow. Tensorflow is only compatible with Python 3.5<\/p>\n\n<pre><code>conda install python=3.5\n<\/code><\/pre>\n\n<p>and then <\/p>\n\n<pre><code>pip install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-0.12.1-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n"},{"AnswerId":"42283330","CreationDate":"2017-02-16T19:49:58.070","ParentId":null,"OwnerUserId":"7577078","Title":null,"Body":"<p>I also encountered the same problem.\nBefore they update TensorFlow.\nYou can try to download the historical version of his 2016-9-27 on <a href=\"https:\/\/repo.continuum.io\/archive\/index.html\" rel=\"nofollow noreferrer\">this website<\/a>.<\/p>\n"},{"AnswerId":"42356998","CreationDate":"2017-02-21T01:13:29.477","ParentId":null,"OwnerUserId":"4847576","Title":null,"Body":"<p><strong>UPDATE:<\/strong> TensorFlow from version 1.2 forward officially supports Python 3.6<\/p>\n\n<p>There is no wheel for Python 3.6 for currently not being supported on Windows yet.\nAs you can see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6999\" rel=\"nofollow noreferrer\">here<\/a>, Python 3.6 support on Windows is a work in progress. \nThe only alternative to use TensorFlow on Windows with Python 3.6 is to build it from source.<\/p>\n"}]}
{"QuestionId":42281589,"AnswerCount":2,"Tags":"<python><numpy><machine-learning><deep-learning><keras>","CreationDate":"2017-02-16T18:08:52.277","AcceptedAnswerId":42282192.0,"OwnerUserId":5896419.0,"Title":"Keras Image Preprocessing: Tuple Index Out of Range","Body":"<p>The goal of the script is to use Keras' existing image preprocessing module for video data augmentation. In this prototype, a sample video is split into an array of frames and processed, where the final steps involve performing random rotations, shifts, shears, and zooms:<\/p>\n\n<pre><\/pre>\n\n<p>When the script is the run, the following error appears:\n<a href=\"https:\/\/i.stack.imgur.com\/ONl5Y.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ONl5Y.png\" alt=\"Output Error\"><\/a><\/p>\n","answers":[{"AnswerId":"42282192","CreationDate":"2017-02-16T18:43:42.750","ParentId":null,"OwnerUserId":"5896419","Title":null,"Body":"<p>The problem was due to the unequal frame dimensions. The solution was to equalize the frame dimensions first before applying transformations:<\/p>\n\n<pre><code># pre-pad short sequences, equalize frame dimensions, and equalize sequence lengths\nif len(frames) &lt; self_frames_per_sequence:\n    frames = [frames[0]]*(self_frames_per_sequence - len(frames)) + frames\nframes = frames[0:self_frames_per_sequence]\nframes = [cv2.resize(frame, (self_columns, self_rows)).astype('float32') for frame in frames]\nframes = np.asarray(frames)\n\nrotated_frames = random_rotation(frames, rg=45)\nshifted_frames = random_shift(rotated_frames, wrg=0.25, hrg=0.25)\nsheared_frames = random_shear(shifted_frames, intensity=0.79)\nzoomed_frames = random_zoom(sheared_frames, zoom_range=(1.25, 1.25))\n<\/code><\/pre>\n"},{"AnswerId":"42281972","CreationDate":"2017-02-16T18:31:10.630","ParentId":null,"OwnerUserId":"483620","Title":null,"Body":"<p>Look at the parameters: <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/preprocessing\/image.py#L25\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/preprocessing\/image.py#L25<\/a>.<\/p>\n\n<p>You are giving <code>frames<\/code> as a flat array, but it is expecting an array with at least three axes, so it can take <code>row_axis=1, col_axis=2<\/code>, by default. Either specify those parameters correctly, or provide with a correctly shaped array.<\/p>\n"}]}
{"QuestionId":42281844,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-16T18:24:30.847","AcceptedAnswerId":null,"OwnerUserId":49530.0,"Title":"What is the mathematics behind the \"smoothing\" parameter in TensorBoard's scalar graphs?","Body":"<p>I presume it is some kind of moving average, but the valid range is between 0 and 1.<\/p>\n","answers":[{"AnswerId":"49357445","CreationDate":"2018-03-19T07:06:22.470","ParentId":null,"OwnerUserId":"5368083","Title":null,"Body":"<p>It is called <a href=\"https:\/\/en.wikipedia.org\/wiki\/Moving_average#Exponential_moving_average\" rel=\"nofollow noreferrer\">exponential moving average<\/a>, below is a code explanation how it is created.<\/p>\n\n<p>Assuming all the <strong>real<\/strong> scalar values are in a list called <code>scalars<\/code> the smoothing is applied as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>def smooth(scalars: List[float], weight: float) -&gt; List[float]:  # Weight between 0 and 1\n    last = scalars[0]  # First value in the plot (first timestep)\n    smoothed = list()\n    for point in scalars:\n        smoothed_val = last * weight + (1 - weight) * point  # Calculate smoothed value\n        smoothed.append(smoothed_val)                        # Save it\n        last = smoothed_val                                  # Anchor the last smoothed value\n\n    return smoothed\n<\/code><\/pre>\n"}]}
{"QuestionId":42283392,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tflearn>","CreationDate":"2017-02-16T19:53:55.260","AcceptedAnswerId":42283943.0,"OwnerUserId":6824263.0,"Title":"Generate data with neural networks","Body":"<p>I want to generate data with a neural network which I have trained before.\nDid someone know how to make that with TensorFlow of TFlearn?<\/p>\n\n<p>It should generate a image with a network i trained before with a dataset contains pictures. <\/p>\n","answers":[{"AnswerId":"42283943","CreationDate":"2017-02-16T20:27:17.160","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>I don't know how you imagined \"generate an image\", but one way would be trying to see what image maximizes your prediction for one of your labels. To do that you can replace you input image placeholder by a variable you initialize with random noise. Your optimizer will optimize this variable to give a maximum prediction. <\/p>\n\n<p>An example is given in lesson 4 of the Creative Applications Of Deep Learning course on Kadenze. Check it out here: <a href=\"https:\/\/github.com\/pkmital\/CADL\/blob\/master\/session-4\/lecture-4.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pkmital\/CADL\/blob\/master\/session-4\/lecture-4.ipynb<\/a><\/p>\n"}]}
{"QuestionId":42284161,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-16T20:39:05.883","AcceptedAnswerId":null,"OwnerUserId":5587428.0,"Title":"Difference in Behavior with tf.variable_scope in TF Version 1.0.0","Body":"<p>Previously  threw an error if variables were defined with the same name and the  parameter was not set to . This no longer seems to be true in TF 1.0.0. The following minimum example shows this:<\/p>\n\n<pre><\/pre>\n\n<p>TF ver 0.12.0 gives the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: Variable foo\/Conv\/weights already exists, disallowed. Did you mean to set reuse=True in VarScope?<\/p>\n<\/blockquote>\n\n<p>TF ver 1.0.0 executes without error. Is this an intended change in behavior? If so, how should this type of error be raised in ver 1.0?<\/p>\n\n<p>Note that this effect seems to be associated with  and .  doesn't show the same issue. Why???<\/p>\n","answers":[]}
{"QuestionId":42284528,"AnswerCount":1,"Tags":"<python><linux><cuda><tensorflow>","CreationDate":"2017-02-16T21:01:31.517","AcceptedAnswerId":42287754.0,"OwnerUserId":395857.0,"Title":"How can I use TensorFlow without CUDA on Linux?","Body":"<p>I have two computers without CUDA: one runs on Microsoft Windows, the other one runs on Linux (Ubuntu 14.04 64bit \/ Linux 3.13.0-100-generic))<\/p>\n\n<p>I can use TensorFlow without CUDA on Microsoft Windows without any issue: TensorFlow uses the CPU. However, if on the Linux machine I run  in python  , then TensorFlow fails to get imported due to CUDA being not installed:<\/p>\n\n<pre><\/pre>\n\n<p>How can I use TensorFlow without CUDA on Linux?<\/p>\n\n<p>I use .<\/p>\n\n<hr>\n\n<p>I'm aware of the parameter  in  , which allows to disable the GPU, e.g.:<\/p>\n\n<pre><\/pre>\n\n<p>but it doesn't help since   is the issue.<\/p>\n\n<p>I also know how to install CUDA 8:<\/p>\n\n<pre><\/pre>\n\n<p>but would prefer to avoid it on these two machines.<\/p>\n","answers":[{"AnswerId":"42287754","CreationDate":"2017-02-17T01:22:29.513","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>If you build the binary with <code>--config=cuda<\/code> (as was done for <code>tensorflow-gpu<\/code>) then your machine must have GPU drivers. You can install GPU drivers on the machine even if the machine doesn't have GPU which is the common practical solution.<\/p>\n\n<p>What happens is that <code>--config=cuda<\/code> sets <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/search?utf8=%E2%9C%93&amp;q=GOOGLE_CUDA&amp;type=Code\" rel=\"noreferrer\">GOOGLE_CUDA<\/a> macro in the code which changes behavior during runtime. In particular it causes dso_loader to run which you can see by following line printed<\/p>\n\n<pre><code>2017-02-16 17:15:08: I tensorflow\/stream_executor\/dso_loader.cc:125] successfully opened CUDA library libcurand.8.0.dylib locally\n<\/code><\/pre>\n\n<p>A popular approach at Google is to deploy a \"fat\" binary -- ie binary that bundles drivers and client code for all possible hardware accelerators because that simplifies testing and deployment.<\/p>\n\n<p>In open-source release, drivers and client code are separate, but this pattern remains -- the GPU-capable binary is expecting to have GPU drivers available.<\/p>\n"}]}
{"QuestionId":42284619,"AnswerCount":1,"Tags":"<python><numpy><theano>","CreationDate":"2017-02-16T21:06:46.960","AcceptedAnswerId":42286107.0,"OwnerUserId":7577395.0,"Title":"theano batched_dot behaves unexpected","Body":"<p>By using theano's batched_dot, but I found something that looks like a bug to me. Transposition of the numpy array makes batched_dot fail, even the resulting shape is correct. Below is my mwe. My questions: should the last line work and is this a bug? <\/p>\n\n<p>I'm using the bleeding-edge installation which is currently Theano-0.9.0b1.\nIf required, I can make this an issue on github. <\/p>\n\n<p>Thanks.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42286107","CreationDate":"2017-02-16T22:44:27.450","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>I don't know how <code>batched_dot<\/code> works, but the strides of your transposed array may be throwing it off<\/p>\n\n<pre><code>In [128]: a_transposed_np = np.zeros((2,2,4))\nIn [129]: a_transposed_np.shape\nOut[129]: (2, 2, 4)\nIn [130]: a_transposed_np.T.shape\nOut[130]: (4, 2, 2)\nIn [131]: a_transposed_np.strides\nOut[131]: (64, 32, 8)\nIn [132]: a_transposed_np.T.strides\nOut[132]: (8, 32, 64)\nIn [133]: np.zeros((4,2,2)).strides\nOut[133]: (32, 16, 8)\n<\/code><\/pre>\n\n<p><code>np.transpose<\/code> creates a <code>view<\/code>, changing the shape, strides and order (to <code>F<\/code>), without copying the data.<\/p>\n\n<p>Does <code>a_transposed_np.T.copy()<\/code> work?<\/p>\n"}]}
{"QuestionId":42284624,"AnswerCount":1,"Tags":"<python><arrays><tensorflow>","CreationDate":"2017-02-16T21:07:04.063","AcceptedAnswerId":null,"OwnerUserId":4979242.0,"Title":"Converting the shape of a tensorflow matrix from (x,y,z) to (1,x,y,z)","Body":"<p>How to change the matrix from 3-D to 4-D. More specifically how to convert a matrix  with shape  to a matrix  with shape ?<\/p>\n\n<p>Such that for example the element <\/p>\n","answers":[{"AnswerId":"42284736","CreationDate":"2017-02-16T21:13:40.293","ParentId":null,"OwnerUserId":"5393381","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/expand_dims\" rel=\"nofollow noreferrer\"><code>tensorflow.expand_dims<\/code><\/a> function should be a perfect match:<\/p>\n\n<pre><code>new_matrix = tf.expand_dims(old_matrix, axis=0)\n<\/code><\/pre>\n"}]}
{"QuestionId":42284873,"AnswerCount":1,"Tags":"<python><arrays><numpy><keras>","CreationDate":"2017-02-16T21:21:17.803","AcceptedAnswerId":44073986.0,"OwnerUserId":5197034.0,"Title":"Assign ImageDataGenerator result to Numpy array","Body":"<p>I'm using the ImageDataGenerator inside Keras to read a directory of images. I'd like to save the result inside a numpy array, so I can do further manipulations and save it to disk in one file.<\/p>\n\n<p> returns an iterator, which is why I tried the following<\/p>\n\n<pre><\/pre>\n\n<p>but that produced<\/p>\n\n<pre><\/pre>\n\n<p>I think I'm misusing the  function, but I can't figure out where I fail.<\/p>\n","answers":[{"AnswerId":"44073986","CreationDate":"2017-05-19T15:36:05.683","ParentId":null,"OwnerUserId":"8014273","Title":null,"Body":"<p>I had the same problem and solved it the following way:\nitr.next returns the next batch of images as two numpy.ndarray objects: batch_x, batch_y. (Source: <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/preprocessing\/image.py\" rel=\"noreferrer\">keras\/preprocessing\/image.py<\/a>)\nSo what you can do is set the batch_size for flow_from_directory to the size of your whole train dataset.<\/p>\n\n<p>Example, my whole training set consists of 1481 images:<\/p>\n\n<pre><code>train_datagen = ImageDataGenerator(rescale=1. \/ 255)\nitr = train_generator = train_datagen.flow_from_directory(\ntrain_data_dir,\ntarget_size=(img_width, img_height),\nbatch_size=1481,\nclass_mode='categorical')\n\nX, y = itr.next()\n<\/code><\/pre>\n"}]}
{"QuestionId":42284904,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-16T21:22:52.170","AcceptedAnswerId":null,"OwnerUserId":7577537.0,"Title":"Complex gradients on gpu in tensorflow","Body":"<p>Can someone explain to me why the following code does not run:<\/p>\n\n<pre><\/pre>\n\n<p>but this does:<\/p>\n\n<pre><\/pre>\n\n<p>?<\/p>\n\n<p>Is this a bug in tensorflow or am I missing something?<\/p>\n\n<p>Here's the stack trace:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42285200,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2017-02-16T21:41:01.300","AcceptedAnswerId":null,"OwnerUserId":826983.0,"Title":"Why do I get the same prediction for all training samples?","Body":"<p>I have a neural network with  separate outputs where each output consists of a <em>softmax<\/em> layer with two nodes (Yes\/No). <\/p>\n\n<p>I am taking the output of a  and feed it as input for a simple  which I further feed into each of said outputs:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>So far I was able to train the model by <a href=\"https:\/\/stackoverflow.com\/questions\/42259071\/how-to-feed-a-model-with-a-list-of-outputs\">providing a list of training samples<\/a> but now I noticed that I am getting the exact same output for completely different samples in a batch:<\/p>\n\n<blockquote>\n  <p><strong>Please note:<\/strong> Here, each column consists of (2,1) arrays. <em>Each<\/em> column is the prediction for <em>one<\/em> sample.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6wHwG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6wHwG.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I've checked the samples, they <strong>are<\/strong> different. I've also tried to e.g. feed the  into the outputs. In that case the predictions are different. I can only see this outcome if I do it the way shown above.<\/p>\n\n<p>I could live with the fact that the outputs are \"similar\". In that case I'd think that the network is just learning not what I want it to learn but since they are really the <em>same<\/em> I am not quite sure what the problem here is.<\/p>\n\n<hr>\n\n<p>I've tried something similar with a simple feed forward network:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>But it behaves similarly. I can't imagine it's due to imbalanced data. There are 13 classes. There is some imbalance but it's not like that one class has 90% of the mass.<\/p>\n\n<p>Am I doing this right?<\/p>\n","answers":[]}
{"QuestionId":42285937,"AnswerCount":1,"Tags":"<python><windows><tensorflow><pip>","CreationDate":"2017-02-16T22:31:19.363","AcceptedAnswerId":42286074.0,"OwnerUserId":1226676.0,"Title":"Unable to upgrade to tensorflow v1.0 on WIndows","Body":"<p>I tried to upgrade to TensorFlow v1.0 after the recent release, but pip is unable to install it for me.  I originally installed it with<\/p>\n\n<p><\/p>\n\n<p>and tried to run<\/p>\n\n<p><\/p>\n\n<p>But I was given the following output:<\/p>\n\n<pre><\/pre>\n\n<p>I've tried to manually uninstall tensorflow-gpu with pip, but I'm unable to uninstall it, and it looks like unless I can uninstall it, I can't upgrade to the latest version.<\/p>\n\n<p>Any thoughts on why pip can't uninstall tensorflow?<\/p>\n","answers":[{"AnswerId":"42286074","CreationDate":"2017-02-16T22:42:12.757","ParentId":null,"OwnerUserId":"1226676","Title":null,"Body":"<p>Looks like the issue was with Windows path lengths beyond 260 characters not resolving properly.<\/p>\n\n<p>To fix this, set the registry value <code>HKEY_LOCAL_MACHINE\\SYSTEM\\CurrentControlSet\\Control\\FileSystem\\LongPathsEnabled<\/code> to 1.<\/p>\n\n<p>I found the solution on the python 3.5 homepage:<\/p>\n\n<p><a href=\"https:\/\/docs.python.org\/3\/using\/windows.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/3\/using\/windows.html<\/a><\/p>\n"}]}
{"QuestionId":42286113,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-16T22:44:53.690","AcceptedAnswerId":42286944.0,"OwnerUserId":4416941.0,"Title":"Apart from image and sound recognition, what are real-world use cases for Tensorflow?","Body":"<p>When I look at tensorflow, I find lots of cool stuff I can geek out over, but I haven't been able to figure out anywhere that I would use it in the real world.  Google didn't spend a gajillion dollars on it without seeing real world applications.<\/p>\n\n<p>There are lots of cool tutorials on how to build cool stuff with Tensorflow, but they start with the assumption that you are already fluent in their dialect of \"greek without an R\" and can extrapolate business usecases from a demo of recognizing a handwritten letter e in a fixed size cell.<\/p>\n\n<p>I have built demonstration neural networks in pascal, C, C++, and Java at different times, so I have some grasp of the principles.  Is is possible to express this in a manner that an old-style Pascal guy who has dabbled i nthe underlying technology a bit can grasp? <\/p>\n","answers":[{"AnswerId":"42286944","CreationDate":"2017-02-16T23:53:13.217","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>You have to keep in mind that TensorFlow is, as far as the user is concerned, \"just\" a machine learning API. People may happen to use it for image classification - the 2017 Dev Summit showed medical use cases in <a href=\"https:\/\/www.youtube.com\/watch?v=toK1OSLep3s\" rel=\"nofollow noreferrer\">skin cancer detection<\/a> and <a href=\"https:\/\/www.youtube.com\/watch?v=oOeZ7IgEN4o\" rel=\"nofollow noreferrer\">retinal imaging<\/a> - but <em>all<\/em> the topics of supervised and unsupervised machine learning are candidates for TensorFlow, just like they are for any other ML library; regression of sales by advertisement budget, clustering of users in a social network and recommending books based on previous purchases via collaborative filtering, just to name a few.<\/p>\n\n<p>If you heard about the recent self-driving car projects, think about obtaining steering wheel and brake control commands from a live camera feed. NVIDIA had <a href=\"https:\/\/arxiv.org\/abs\/1604.07316\" rel=\"nofollow noreferrer\">a paper on it<\/a>, for example.<\/p>\n\n<p>One rather interesting use case are <a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\" rel=\"nofollow noreferrer\">sequence to sequence models<\/a> to transform one arbitrary sequence of inputs to another one; according to <a href=\"https:\/\/www.youtube.com\/watch?v=RIR_-Xlbp7s\" rel=\"nofollow noreferrer\">this video<\/a>, Google Translate might be taking advantage of it on the phone. If you're thinking of image and video retrieval, sequence labelling is another topic, where you train a network to describe, in human words, the content of a video. Or even natural language processing, where you try to determine the <em>concepts<\/em> within written text.<\/p>\n\n<p>There are also papers <a href=\"https:\/\/arxiv.org\/abs\/1610.09460\" rel=\"nofollow noreferrer\">like this<\/a> describing the usage of recurrent models like LSTMs for energy usage prediction (Note the paper isn't specific to TensorFlow, but LSTMs are part of the core library).  <a href=\"http:\/\/de.slideshare.net\/TaegyunJeon1\/electricity-price-forecasting-with-recurrent-neural-networks\" rel=\"nofollow noreferrer\">Here<\/a> are slides on electricity price forecasting with TensorFlow, if you're interested in it.<\/p>\n"}]}
{"QuestionId":42286426,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning><conv-neural-network>","CreationDate":"2017-02-16T23:06:32.420","AcceptedAnswerId":42289273.0,"OwnerUserId":195651.0,"Title":"What is the difference between xavier_initializer and xavier_initializer_conv2d?","Body":"<p>I noticed that TensorFlow 1.0 contains two Xavier initalization helpers in contrib,<\/p>\n\n<ul>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/xavier_initializer\" rel=\"noreferrer\">tf.contrib.layers.xavier_initializer()<\/a> and<\/li>\n<li><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/xavier_initializer\" rel=\"noreferrer\">tf.contrib.layers.xavier_initializer_conv2d()<\/a><\/li>\n<\/ul>\n\n<p>Both link to the same documentation page and have the same signature:<\/p>\n\n<pre><\/pre>\n\n<p>however the difference between them is not explained at all. I can guess by the name that the  version should be used for 2D convolutional layers, but would it have a noticeable impact if one were to use the regular version?<\/p>\n","answers":[{"AnswerId":"42289273","CreationDate":"2017-02-17T04:13:30.543","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>There is actually no difference. Because both the functions implement the weight initialization from:<\/p>\n\n<blockquote>\n  <p>Xavier Glorot and Yoshua Bengio (2010): <a href=\"http:\/\/www.jmlr.org\/proceedings\/papers\/v9\/glorot10a\/glorot10a.pdf?hc_location=ufi\" rel=\"noreferrer\">Understanding the difficulty of training deep feedforward neural networks.<\/a> International conference on artificial intelligence and statistics.<\/p>\n<\/blockquote>\n\n<p>Both the initializer is designed to keep the scale of the gradients roughly the same in all layers. In uniform distribution this ends up being the range: <code>x = sqrt(6. \/ (in + out)); [-x, x]<\/code> and for normal distribution a standard deviation of <code>sqrt(3. \/ (in + out))<\/code> is used.<\/p>\n"}]}
{"QuestionId":42286972,"AnswerCount":6,"Tags":"<python><pandas><tensorflow>","CreationDate":"2017-02-16T23:56:26.243","AcceptedAnswerId":null,"OwnerUserId":7508642.0,"Title":"Converting from Pandas dataframe to TensorFlow tensor object","Body":"<p>I'm still new to Python, Machine Learning and TensorFlow, but doing my best to jump right in head-first. I could use some help though.<\/p>\n\n<p>My data is currently in a Pandas dataframe. How can I convert this to TensorFlow object? I've tried<\/p>\n\n<pre><\/pre>\n\n<p>But, I get errors .<\/p>\n\n<p>I'm sure this is probably a straightforward question, but I could really use the help.<\/p>\n\n<p>Many thanks<\/p>\n\n<p>ps. I'm running tensorflow 0.12 with Anaconda Python 3.5 on Windows 10 <\/p>\n","answers":[{"AnswerId":"55825712","CreationDate":"2019-04-24T08:35:07.843","ParentId":null,"OwnerUserId":"11374349","Title":null,"Body":"<p>You can use <code>tf.estimator.inputs.pandas_input_fn<\/code> in your <code>make_input_fn(X, y, num_epochs)<\/code> function. I've not managed to get it to work with a multi-index, however. I fixed this issue by turning it into a standard integer index, using <code>df.reset_index(drop=True)<\/code><\/p>\n"},{"AnswerId":"42295872","CreationDate":"2017-02-17T10:44:45.670","ParentId":null,"OwnerUserId":"7241513","Title":null,"Body":"<p>The following works easily based on <code>numpy<\/code> array input data:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\na = np.array([1,2,3])\nwith tf.Session() as sess:\n    tf.global_variables_initializer().run()\n\n    dataVar = tf.constant(a)\n    print(dataVar.eval())\n\n-&gt; [1 2 3]\n<\/code><\/pre>\n\n<p>Don't forget to start the <code>session<\/code> and <code>run()<\/code> or <code>eval()<\/code> your tensor object to see its content; otherwise it will just give you its generic description.<\/p>\n\n<p>I suspect that since your data is in the DataFrame rather than a simply array, you need to experiment with the <code>shape<\/code> <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"nofollow noreferrer\">parameter<\/a> of <code>tf.constant()<\/code>, which you are currently not specifying, in order to help it understand the dimensionality of the DataFrame and deal with its indices, etc.? <\/p>\n"},{"AnswerId":"51105896","CreationDate":"2018-06-29T16:21:29.003","ParentId":null,"OwnerUserId":"8007589","Title":null,"Body":"<p>here is one solution i found that works on google colab , probably should work in a local machine too <\/p>\n\n<pre><code>import pandas as pd\nimport tensorflow as tf\n#Read the file to a pandas object\ndata=pd.read_csv('filedir')\n#convert the pandas object to a tensor\ndata=tf.convert_to_tensor(data)\ntype(data)\n<\/code><\/pre>\n\n<p>This must print something like<\/p>\n\n<pre><code>tensorflow.python.framework.ops.Tensor\n<\/code><\/pre>\n\n<p>Hope this helps :)<\/p>\n\n<p>`   <\/p>\n"},{"AnswerId":"54397699","CreationDate":"2019-01-28T08:01:48.957","ParentId":null,"OwnerUserId":"2056686","Title":null,"Body":"<p>hottbox.pdtools.utils (the Pandas integration tools of the HOTTBOX API) provides the functions<\/p>\n\n<pre><code>   pd_to_tensor(df[, keep_index])\n   tensor_to_pd(tensor[, col_name])\n<\/code><\/pre>\n\n<p>for conversion in both directions.<\/p>\n"},{"AnswerId":"57816111","CreationDate":"2019-09-06T05:17:44.513","ParentId":null,"OwnerUserId":"6241826","Title":null,"Body":"<p>You can convert a the dataframe column to a tensor object like so:<\/p>\n\n<pre><code>tf.constant((df['column_name']))\n<\/code><\/pre>\n\n<p>This should return you a tensor variable which looks something like this:<\/p>\n\n<pre><code>&lt;tf.Tensor: id=275634, shape=(48895,), dtype=float64, numpy=\narray([1, 2, ...])&gt;\n<\/code><\/pre>\n\n<p>Also, you can ad any number of dataframe columns as you want, like so:  <\/p>\n\n<pre><code>tf.constant(([cdf['column1'], cdf['column2']]))\n<\/code><\/pre>\n\n<p>Hope this helps.<\/p>\n"},{"AnswerId":"42302125","CreationDate":"2017-02-17T15:53:51.570","ParentId":null,"OwnerUserId":"7508642","Title":null,"Body":"<p>I think I've got it! :D<\/p>\n\n<p>I've converted my Pandas dataframe to a Numpy array using .as_matrix()<\/p>\n\n<p>Now, using <\/p>\n\n<pre><code>dataVar_tensor = tf.constant(dataVar, dtype = tf.float32, shape=[15780,9])\ndepth_tensor = tf.constant(depth, 'float32',shape=[15780,1])\n<\/code><\/pre>\n\n<p>seems to work. I can't say it does definitively because I have other hurdles to overcome to get my code working, but it's hopefully a step in the right direction. Thanks for all your help<\/p>\n\n<p>As an aside, my trials of getting the tutorial to work on my own data are continued in my next question <a href=\"https:\/\/stackoverflow.com\/questions\/42302498\/converting-tensorflow-tutorial-to-work-with-my-own-data\">Converting TensorFlow tutorial to work with my own data<\/a><\/p>\n"}]}
{"QuestionId":42287173,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-17T00:15:31.973","AcceptedAnswerId":null,"OwnerUserId":2382483.0,"Title":"Tensorflow viewing a uint8 tensor as a float32 tensor","Body":"<p>I have a bunch of training example files in a custom format id like to load into a tensorflow model, but I can't figure out how to efficiently parse the data. <\/p>\n\n<p>The files have a 20 byte header describing the number of features, number of labels, number of examples, etc. The header is followed by some number of examples. Each example is some number of float32s representing the features, followed by some number of uint8s representing the labels, so the rest of the file is an alternating pattern of floats and uints. <\/p>\n\n<p>I have been using  and  to get the file as a vector of uint8s, and now id like to slice the features from the labels, and view the features vector as a float32 vector in little endian order. Any chance this sort of thing is possible in tensorflow?<\/p>\n\n<p>I have been successfully parsing the files with  and converting the features and labels to s, but it seems like it would be way more memory efficient to load it straight into tensorflow.<\/p>\n","answers":[{"AnswerId":"42308268","CreationDate":"2017-02-17T22:11:43.597","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>Assuming it's all fixed-size data, you need to de-interleave the data using <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/array_ops\" rel=\"nofollow noreferrer\">Slicing and Joining<\/a> operators. Then, you can use <code>decode_raw<\/code> with <code>out_dtype=tf.float32<\/code> on the image data, and slice again to get your vectors individually.<\/p>\n"}]}
{"QuestionId":42287270,"AnswerCount":2,"Tags":"<tensorflow><backpropagation>","CreationDate":"2017-02-17T00:27:45.993","AcceptedAnswerId":42304848.0,"OwnerUserId":1058511.0,"Title":"use resilient propagation on tensorflow","Body":"<p>is there a way to use resilient propagation on Tensorflow? I know there are a number of back propagation strategies. Is there one that is close to rPorp? Can I insert my own implementation of resilient propagation?<\/p>\n","answers":[{"AnswerId":"42304848","CreationDate":"2017-02-17T18:17:08.650","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>There is no rprop implementation in tensorflow, although it would be fairly trivial to create one. You could create one by writing an op, or directly in python by combining ops.<\/p>\n\n<p>There is an <a href=\"https:\/\/www.github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/rmsprop.py#L50\" rel=\"nofollow noreferrer\">RMSProp<\/a> which is a different thing.<\/p>\n\n<p>Note that RProp doesn't work well with stochastic updates. The batch sizes would have to be very large for it to work.<\/p>\n"},{"AnswerId":"45960501","CreationDate":"2017-08-30T12:31:21.543","ParentId":null,"OwnerUserId":"8537848","Title":null,"Body":"<p>RProp and RMSProp are quite similar. They both compare the sign of the gradient of the current batch (or single sample) to the previous, to update a unique value for each parameter (usually multiplying by 1.2 when the signs agree and by 0.5 when they don't). But while RProp compares each batch gradient, RMSProp uses a discount-factor to keep a running average for comparing signs. RProp uses this unique value to take an absolute step in the direction of the gradient while RMSProp multiplies the value and the gradient.<\/p>\n\n<p>RProp works great for larger batches, but doesn't work well for stochastic updates, since the sign of the gradient will flicker causing the steps to approach minimum which stops learning. The running average of RMSProp solves this issue. But because RMSProp multiplies the value and gradient, it's more susceptible to saturation than RProp (at least for Sigmoid and Tanh - but you can of course use Relu or Leaky Relu to get around that).<\/p>\n"}]}
{"QuestionId":42287784,"AnswerCount":2,"Tags":"<python><machine-learning><keras>","CreationDate":"2017-02-17T01:25:13.137","AcceptedAnswerId":null,"OwnerUserId":5171213.0,"Title":"Keras: learning rate schedule","Body":"<p>I am implementing an MLP in Keras, and tweaking the hyperparameters. One object of experimentation is the learning rate. There are two schedules I'm trying to use, both outlined in <a href=\"http:\/\/machinelearningmastery.com\/using-learning-rate-schedules-deep-learning-models-python-keras\/\" rel=\"nofollow noreferrer\">this tutorial<\/a>. One is specifically defined using learning rate \/ epochs, and one uses a separately-defined step decay function. The necessary code is below. <\/p>\n\n<p>The error is 'The output of the \"schedule\" function should be float'. I specifically cast the learning rate as a float, so I'm not sure where I'm going wrong?<\/p>\n\n<p>EDIT: original code was not a MWE, I apologize. To reproduce this error, you can save the data snippets below and run this code.<\/p>\n\n<pre><\/pre>\n\n<p>Error message:<\/p>\n\n<pre><\/pre>\n\n<p>Data snippet (train_ex.csv):<\/p>\n\n<pre><\/pre>\n\n<p>Data snippet (test_ex.csv):<\/p>\n\n<pre><\/pre>\n\n<p>EDIT 2:<\/p>\n\n<p>Based on @sascha's comments, I've tried modifying a bit (this is the relevant section below). Same error.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42320685","CreationDate":"2017-02-18T21:05:28.740","ParentId":null,"OwnerUserId":"2320035","Title":null,"Body":"<p>First of all: <strong>i misinterpreted your code earlier and my comments are deprecated! Sorry!<\/strong>.<\/p>\n\n<p>The error-message leads us to the real problem here!<\/p>\n\n<p>You define your scheduler like this:<\/p>\n\n<pre><code>def step_decay(epoch):\n    drop_every = 10\n    decay_rate = (learning_rate_*np.power(0.5, np.floor((1+drop_every)\/drop_every))).astype('float32')\n    return decay_rate\n<\/code><\/pre>\n\n<p><strong>Check the type it returns!*<\/strong> It's <code>&lt;class 'numpy.float32'&gt;<\/code>. (Try it: with python's <code>type()<\/code> function)<\/p>\n\n<p>Somehow keras is not doing a very general check for these types and expects <code>&lt;class 'float'&gt;<\/code> (python's native float).<\/p>\n\n<p>Just convert your numpy-float to a native python-float:<\/p>\n\n<p>Replace: <code>decay_rate = (learning_rate_*np.power(0.5, np.floor((1+drop_every)\/drop_every))).astype('float32')<\/code><\/p>\n\n<p>with: <code>decay_rate = (learning_rate_*np.power(0.5, np.floor((1+drop_every)\/drop_every))).astype('float32').item()<\/code><\/p>\n\n<p><a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.ndarray.item.html\" rel=\"nofollow noreferrer\">Read the docs of numpy.ndarray.item<\/a> (especially the notes on the why of this behaviour)<\/p>\n\n<p>The blog-author does not have this problem as he is not using numpy within his scheduler and uses python's math-functions. This will result in a native-float!<\/p>\n"},{"AnswerId":"42963385","CreationDate":"2017-03-22T21:55:42.163","ParentId":null,"OwnerUserId":"448357","Title":null,"Body":"<p>You can also try to check out the <a href=\"https:\/\/keras.io\/callbacks\/#reducelronplateau\" rel=\"noreferrer\">ReduceLROnPlateau<\/a> callback to reduce the learning rate by a pre-defined factor, if a monitored value has not changed for a certain number of epochs, e.g. half the learning rate if the validation accuracy has not improved for five epochs looks like this:<\/p>\n\n<pre><code>learning_rate_reduction = ReduceLROnPlateau(monitor='val_acc', \n                                            patience=5, \n                                            verbose=1, \n                                            factor=0.5, \n                                            min_lr=0.0001)\nmodel.fit_generator(..., callbacks=[learning_rate_reduction], ...)\n<\/code><\/pre>\n"}]}
{"QuestionId":42287954,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-17T01:46:00.433","AcceptedAnswerId":null,"OwnerUserId":5538737.0,"Title":"tf.argmax() vs tf.arg_max() in TensorFlow","Body":"<p>It seems  and  are doing the same thing for outputting the indies of the biggest element in matrix.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/argmax\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/argmax<\/a>\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/arg_max\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/arg_max<\/a><\/p>\n\n<p>Why does TensorFlow have two APIs for this?<\/p>\n","answers":[{"AnswerId":"42289528","CreationDate":"2017-02-17T04:40:40.720","ParentId":null,"OwnerUserId":"3502070","Title":null,"Body":"<p>The following is the source code of <code>argmax<\/code> (from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/math_ops.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/ops\/math_ops.py<\/a>).<\/p>\n\n<pre><code># pylint: disable=redefined-builtin\n# TODO(aselle): deprecate arg_max\ndef argmax(input, axis=None, name=None, dimension=None):\n  if dimension is not None:\n    if axis is not None:\n      raise ValueError(\"Cannot specify both 'axis' and 'dimension'\")\n    axis = dimension\n  elif axis is None:\n    axis = 0\n  return gen_math_ops.arg_max(input, axis, name)\n<\/code><\/pre>\n\n<p>As you can see, <code>argmax<\/code> is using <code>arg_max<\/code> inside. Also from the code, I recommend using <code>argmax<\/code> because <code>arg_max<\/code> could be deprecated soon.<\/p>\n"},{"AnswerId":"56751665","CreationDate":"2019-06-25T10:12:35.243","ParentId":null,"OwnerUserId":"7677598","Title":null,"Body":"<p>With TensorFlow version <strong>1.14.0<\/strong> it's suggested to use <code>tf.math.argmax<\/code> function. And also <code>dimension<\/code> parameter is deprecated instead of it you can use <code>axis<\/code>.<\/p>\n\n<p>Sample<\/p>\n\n<pre><code>tf.math.argmax(input, axis=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":42288142,"AnswerCount":1,"Tags":"<python><numpy><linear-regression><keras><conv-neural-network>","CreationDate":"2017-02-17T02:04:09.367","AcceptedAnswerId":null,"OwnerUserId":1158977.0,"Title":"Binning Data Using Numpy To Simplify Linear Regression","Body":"<p>I have training data in the form of images taken by a  from an RaspberryPi RC car while I drive it in between two lane lines.<\/p>\n\n<p>Each image is labelled with Left and Right motor controls.  I've plotted them in the below graph.<\/p>\n\n<p>I am using  to do the regression with  and  as backend.<\/p>\n\n<p>It's clearly visible that regression can be much simpler if I am able to remove the training samples which are to the left of the visible regression.<\/p>\n\n<p>The code for loading the images and labels is very simple and is below:<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6qWy2.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/6qWy2.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Firstly is there a good approach for removing data which is out of bounds of the visible regression from the training samples?<\/p>\n\n<p>Also, I have a different approach in mind which is to create  data bins with edges of  to , then take 50 samples from each bin such that my data is balanced.<\/p>\n\n<p>Is there a  way to put the data into bins so that I don't need to do that with some custom function?<\/p>\n","answers":[{"AnswerId":"42288366","CreationDate":"2017-02-17T02:30:57.003","ParentId":null,"OwnerUserId":"4492932","Title":null,"Body":"<p>Answering the first question:<\/p>\n\n<ol>\n<li>Do the linear regression fit on the whole dataset<\/li>\n<li>Remove the points with the largest residuals<\/li>\n<li>Repeat 1 and 2 until all residuals are comfortably small<\/li>\n<\/ol>\n"}]}
{"QuestionId":42288290,"AnswerCount":1,"Tags":"<python><cuda><tensorflow><gpu><cudnn>","CreationDate":"2017-02-17T02:22:00.007","AcceptedAnswerId":42296976.0,"OwnerUserId":3768495.0,"Title":"cuda kernel doesn't match during installation of Tensorflow","Body":"<p>I am installing Tensorflow as instructed on the <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux\" rel=\"nofollow noreferrer\">official page<\/a> and at the step of \"Validate your installation\".<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0ueSe.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0ueSe.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<pre><\/pre>\n\n<p>This is what I got when I try to start a session. I don't quite understand what's going on. Please kindly help. Thank you!<\/p>\n","answers":[{"AnswerId":"42296976","CreationDate":"2017-02-17T11:38:42.820","ParentId":null,"OwnerUserId":"4262324","Title":null,"Body":"<p>I don't believe the error is related to Tensorflow. You should get the same error running <strong>nvidia-smi<\/strong>.<\/p>\n\n<p>Is it possible you have updated your NVIDIA GPU drivers after installing the CUDA toolkit? It looks like the toolkit expects drivers version 367.57, while you are running a more recent version, 375.26.<\/p>\n\n<p>For a quick check, try reverting to NVIDIA drivers version 367.57; you can do it from <strong>System Settings > Software and Updates > Additional Drivers<\/strong>.<\/p>\n\n<p>Once confirmed the mismatch in expected drivers version is the issue, you can either stay with drivers version 367.57, or un-install CUDA Toolkit and cuDNN, update the drivers to 375.26, and the re-install CUDA Toolkit and cuDNN.<\/p>\n"}]}
{"QuestionId":42288907,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-17T03:32:15.013","AcceptedAnswerId":null,"OwnerUserId":6315546.0,"Title":"Tensorflw MultiRNNCell's parameter in doc ,I think it is wrong","Body":"<p>I found in the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\" rel=\"nofollow noreferrer\">tensorflow doc<\/a>:<\/p>\n\n<p><\/p>\n\n<p>I need to use MultiRNNCell<\/p>\n\n<p>but, I write those lines<\/p>\n\n<p><\/p>\n\n<p>Its output is .<\/p>\n\n<p>Can  use the same object  as a list for parameter?<\/p>\n","answers":[{"AnswerId":"42305172","CreationDate":"2017-02-17T18:37:33.850","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>The <code>MultiRNNCell<\/code> takes a list of <a href=\"https:\/\/www.github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py#L668\" rel=\"nofollow noreferrer\">cells<\/a> as a parameter.<\/p>\n\n<p>The <code>BasicLSTMCell<\/code> is an <a href=\"https:\/\/www.github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py#L124\" rel=\"nofollow noreferrer\">RNN cell<\/a>.<\/p>\n\n<p>So yes, <code>MultiRNNCell<\/code> should be able to take a list of <code>BasicLSTMCell<\/code> objects as the <code>cell<\/code> parameter.<\/p>\n"}]}
{"QuestionId":42289274,"AnswerCount":1,"Tags":"<python><tensorflow><virtualenv>","CreationDate":"2017-02-17T04:13:40.717","AcceptedAnswerId":null,"OwnerUserId":4422065.0,"Title":"virtualenv install a package that exists globally?","Body":"<p>I set up a VirtualEnvironment. Using the  command, I am sure that python 3.5 and pip 9.0.1 which is from Python3.5 are in the local virtualenv. My global system has pip2.7 installed.<\/p>\n\n<p>I have a git clone that uses python2 and the Tensorflow library. I use pip3 to install tensorflow 1.0.0. Running on python2.7 I get no module named Tensorflow. Executing with python3.5 trips because the project was written in 2.<\/p>\n\n<p><strong>What am I trying to do<\/strong>: install pip2.7 in my virtualenv (instead of my global MacOS, which it is currently under), and then I can use to install Tensorflow (python 2.7) aka this:<\/p>\n\n<pre><\/pre>\n\n<p>If I install it with pip3.5 (virtualenv) I get: <\/p>\n\n<p>If I install it with pip2.7 (MacOS Global), one of the global dependencies, Numpy, throws an exception and it won't install Tensorflow.<\/p>\n\n<p><strong>All I need to do is get pip2.7 on virtualenv.<\/strong> Attempts like  while in the virtualenv fail as it says . As you know, virtualenv's  has been made the default so that's not the issue.<\/p>\n\n<p>I've tried some other answers, like --upgrade which produces <\/p>\n","answers":[{"AnswerId":"42289448","CreationDate":"2017-02-17T04:32:47.407","ParentId":null,"OwnerUserId":"4422065","Title":null,"Body":"<p>The solution was actually to make a new virtualenv, this time with Python2.7<\/p>\n\n<p><code>virtualenv -p python2.7 ~\/my\/env\/dir\/<\/code><\/p>\n\n<p>pip2.7 and python2.7 were installed and I could install Tensorflow beautifully!<\/p>\n"}]}
{"QuestionId":42289871,"AnswerCount":1,"Tags":"<machine-learning><lua><neural-network><conv-neural-network><torch>","CreationDate":"2017-02-17T05:10:46.267","AcceptedAnswerId":null,"OwnerUserId":2838606.0,"Title":"How to apply changes to elements of a tensor\/vector without changing its reference?","Body":"<p>I am sharing parameters of a network and want to apply some operations to change to the elements of the parameter vector. I cannot do those operations directly on the parameter tensor because they will certainly change the parameter vector reference and sharing breaks. So <strong>:clone()<\/strong> the shared parameters vector and apply the changes I want on the new vector and use the <strong>:copy()<\/strong> function to replace the elements in the original parameter vector. I thought the <strong>tensor:copy()<\/strong> function does not change the vector\/tensor reference but it seems that it does since the parameter sharing collapses soon after apply it. So I wonder, can anyone fix the code below and suggest a way to change the elements of the parameter vector without breaking the sharing?<\/p>\n\n<pre><\/pre>\n\n<p>One of the operations I'm interested in doing is clamping but the clamp() function also breaks the sharing if applied directly kon the parameter vector.<\/p>\n","answers":[{"AnswerId":"42344833","CreationDate":"2017-02-20T12:38:51.933","ParentId":null,"OwnerUserId":"1624600","Title":null,"Body":"<p>Basically, you need a pointer to the Tensor that will modify its raw data, preserve the reference and sharing. This link (Tensor doc) clearly explains it. \n<a href=\"https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#result-datatensor-asnumber\" rel=\"nofollow noreferrer\">https:\/\/github.com\/torch\/torch7\/blob\/master\/doc\/tensor.md#result-datatensor-asnumber<\/a><\/p>\n"}]}
{"QuestionId":42289880,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-02-17T05:11:17.007","AcceptedAnswerId":null,"OwnerUserId":3755376.0,"Title":"How can I generate data layer (HDF5) for training and testing in same prototxt?","Body":"<p>I have a data layer with HDF5 type. It contains both Train and Test phase as expected<\/p>\n\n<pre><\/pre>\n\n<p>I want to use python to generate it. This is my code <\/p>\n\n<pre><\/pre>\n\n<p>However, my output is only Test phase. How can I fix it? Thanks<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42291579","CreationDate":"2017-02-17T07:12:08.263","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>This is an open issue <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/4044\" rel=\"nofollow noreferrer\">in caffe<\/a> (you can find links to other relevant SO threads there).<\/p>\n\n<p>What you can do is write two prototxt one for train and one for test. solver.prototxt support definition of train net file name and test net file name.<\/p>\n"}]}
{"QuestionId":42290182,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-02-17T05:36:24.537","AcceptedAnswerId":null,"OwnerUserId":3010409.0,"Title":"How to give more than one labels with an image in tensorflow?","Body":"<p>I want to implememt multitask Neural Network in tensorflow, for which I need my input as:<\/p>\n\n<blockquote>\n  <p><strong>[image label1 label2]<\/strong><\/p>\n<\/blockquote>\n\n<p>which I can give to the neural network for training.<\/p>\n\n<p>My question is, how can I associate more than one label with image in TFRecord file?<\/p>\n\n<p>I currently was using <strong>build_image_data.py<\/strong> file of inception model for genertrating TFRecord file but in that cases there is just one label per image.<\/p>\n","answers":[{"AnswerId":"42292153","CreationDate":"2017-02-17T07:42:41.303","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>Do they have always two labels? If so try \"label1-label2\" as one label. \nOr simply build two networks, one for label 1 and the other for label 2.<\/p>\n\n<p>Are they hierarchical labels? Then, check out Hierarchical classifiers.<\/p>\n"},{"AnswerId":"42298590","CreationDate":"2017-02-17T12:58:00.043","ParentId":null,"OwnerUserId":"3010409","Title":null,"Body":"<ol>\n<li>I got this working. For any one looking for reference, you can modify Example proto of build_image_data.py file and associate it with two labels. :)<\/li>\n<\/ol>\n"}]}
{"QuestionId":42291135,"AnswerCount":1,"Tags":"<tensorflow><conv-neural-network>","CreationDate":"2017-02-17T06:46:34.620","AcceptedAnswerId":null,"OwnerUserId":860481.0,"Title":"Tensorflow - retrain inception for ethnicity recognition and hair color","Body":"<p>I'm new to tensorflow and the inception model. I found the following tutorial online (<a href=\"https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/image_retraining\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/how_tos\/image_retraining\/<\/a>) and wanted to test this on my own project. <\/p>\n\n<p>I'm trying to let the model recognize ethnicity based on people in a picture. I have made a training set of approx 850 images per category.<\/p>\n\n<p>For some reason I'm unable to get more than a 65% accuracy level. I tried increasing the training steps and amount of images as well. <\/p>\n\n<p>Maybe the inception model is not the correct model to use for this?<\/p>\n\n<p>Could someone point me in a good direction of what I can do to improve the results?<\/p>\n\n<p>Regards,\nP.<\/p>\n","answers":[{"AnswerId":"42292080","CreationDate":"2017-02-17T07:38:04.773","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Do you get 65% accuracy on the train or on the test set?\nIf it is on the train set, you are probably doing something wrong with your code. \nIf it is on the test set, you are indeed using the wrong model. The Inception model is a very very big model, and having only 850 images per category won't give you a good general model. It will simply \"remember\" those 850 images. (think of remembering the answer to each question on a test, instead of learning for a test)<\/p>\n\n<p>Maybe you can try building a simpler, smaller model first and see how well that model learns!<\/p>\n"}]}
{"QuestionId":42292995,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-17T08:32:18.543","AcceptedAnswerId":null,"OwnerUserId":3627108.0,"Title":"How to make tf.contrib.learn.estimator not to add summary?","Body":"<p>I'm using tf.contrib.learn.DNNClaisifer to train my models and meanwhile the estimator generate tensorflow summary in the model_dir which spent a lot of space. I failed to diable the summary since I couldn't find any arguments or interface to specify whether write the tensorflow summary. Could you tell me how to diable the summary writer in tf.contrib.learn.DNNClasifier?<\/p>\n","answers":[{"AnswerId":"42305278","CreationDate":"2017-02-17T18:44:52.577","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>There appears to be no such option, but you can disable it from the <a href=\"http:\/\/google3\/third_party\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn.py?l=61\" rel=\"nofollow noreferrer\">code<\/a>. Remove the contents of <code>contrib.learn.python.learn.estimators.dnn._add_hidden_layer_summary<\/code>.<\/p>\n"}]}
{"QuestionId":42293600,"AnswerCount":1,"Tags":"<sequence><keras><prediction><lstm>","CreationDate":"2017-02-17T09:02:44.610","AcceptedAnswerId":42303656.0,"OwnerUserId":2703401.0,"Title":"Accuracy\/Loss doesn't change","Body":"<p>It might be duplication of the previous posts but here is my code.\nMy inputs X are sequences of characters each of length 10 encoded as 1-26 numbers with added random noise.output is the next word in the sequence.<\/p>\n\n<pre><\/pre>\n\n<p>Tried changing different loss functions and optimizers. No luck.<\/p>\n","answers":[{"AnswerId":"42303656","CreationDate":"2017-02-17T17:10:19.820","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Encoding characters by number is not a good way. It will be interpreted as numbers so it's like saying that Y and Z are close together which doesn't make sense. This is why the Embedding() layers exist. Or you might consider one-hot encoding . Characters are then one-hot vectors of length 26. \n\"a\" would become [1 0 0 0 0 0 0 0 0 ... 0] for example.<\/p>\n\n<p>That being said, the reason it's not working is because you put a Softmax on a layer which has only one value... Softmax on one value will always give output 1, so your network can't learn since output is 1 whatever happens before.<\/p>\n\n<p>Softmax is used to make a probability density out of a tensor, if there is only one possible value, it will get probability 1. If you want that one neuron to be a probability (between 0 and 1) use only the sigmoid, not the softmax. <\/p>\n\n<p>I hope this helps :) <\/p>\n"}]}
{"QuestionId":42296045,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-02-17T10:52:29.540","AcceptedAnswerId":null,"OwnerUserId":826983.0,"Title":"Negative dimension size caused by subtracting 2 from 1 for 'MaxPool'","Body":"<p>I am trying to build a convolution layer as follows:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>But I am getting the error:<\/p>\n\n<pre><\/pre>\n\n<p>I don't quite understand why this is the case. Do I have to specify the input? I thought that this is not necessary except for the model inputs?<\/p>\n","answers":[{"AnswerId":"56967531","CreationDate":"2019-07-10T09:23:47.750","ParentId":null,"OwnerUserId":"11764019","Title":null,"Body":"<p>I have met the same problem before.Every time you use Maxpooling(pool_size=[2,2]),image_size will be reduced by 2 times.Maybe the parameter has been smaller than 2, so that you cannot use Maxpool again.\nHope that would be useful to you.^_^<\/p>\n"}]}
{"QuestionId":42296782,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-02-17T11:28:20.600","AcceptedAnswerId":42297021.0,"OwnerUserId":5512020.0,"Title":"ValueError when executing softmax_cross_entropy_with_logits","Body":"<p>I am following the <a href=\"https:\/\/pythonprogramming.net\/tensorflow-neural-network-session-machine-learning-tutorial\/\" rel=\"noreferrer\">tensorflow tutorial<\/a>. There has been recent tensor flow update in which the cost function  has been modified. Hence the code in the tutorial is giving the following error:<\/p>\n\n<p> <\/p>\n\n<p>What does it mean and how to rectify it?<\/p>\n\n<p>Here's the entire code till that point:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42297021","CreationDate":"2017-02-17T11:40:43.260","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>Change<\/p>\n\n<pre><code>tf.nn.softmax_cross_entropy_with_logits(prediction,y)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y)\n<\/code><\/pre>\n"},{"AnswerId":"57597357","CreationDate":"2019-08-21T18:28:02.940","ParentId":null,"OwnerUserId":"11675463","Title":null,"Body":"<p>the function <code>softmax_cross_entropy_with_logits<\/code> has been deprecated. The new function is <code>softmax_cross_entropy_with_logits_v2<\/code><\/p>\n\n<pre><code>cost=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(pred, y))\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">Documentation<\/a><\/p>\n"},{"AnswerId":"50579415","CreationDate":"2018-05-29T08:11:59.967","ParentId":null,"OwnerUserId":"8076093","Title":null,"Body":"<p>This is because the code donot support Tensorflow v 1.0\nIt works well when you make slight changes to the following lines of code:\nreplace <br><\/p>\n\n<pre><code>cost = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits(prediction,y) )\n<\/code><\/pre>\n\n<p>with <br><br>\n <code>cost = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y))<\/code><\/p>\n"}]}
{"QuestionId":42297359,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-17T11:56:32.457","AcceptedAnswerId":42300533.0,"OwnerUserId":826983.0,"Title":"TypeError: Output tensors to a Model must be Keras tensors","Body":"<p>I want to take an input image  (which also has negative values) and feed it into two activation layers. However, I want to make a simple transformation e.g. multiply the whole image with :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>If I do it this way I am getting:<\/p>\n\n<pre><\/pre>\n\n<p>and I am not sure how I can fix that. Is there a  side  method that I can use for such a thing? Or can I wrap the result of  somehow such that I can pass it on to ?<\/p>\n\n<p><strong>Please note:<\/strong> The negative values may be important. Thus transforming the image s.t. the minimum is simply  is not a solution here.<\/p>\n\n<hr>\n\n<p>I am getting the same error for<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<hr>\n\n<p>The same error for:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42300533","CreationDate":"2017-02-17T14:36:39.850","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Does creating a Lambda Layer to wrap your function work?<\/p>\n\n<p>See doc <a href=\"https:\/\/keras.io\/layers\/core\/#lambda\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<pre><code>from keras.layers import Lambda\nimport tensorflow as tf\n\ndef mul_minus_one(x):\n    return tf.mul(x,-1.0)\ndef mul_minus_one_output_shape(input_shape):\n    return input_shape\n\nmyCustomLayer = Lambda(mul_minus_one, output_shape=mul_minus_one_output_shape)\nright = myCustomLayer(img)\nright = Activation('relu')(right)\n<\/code><\/pre>\n"}]}
{"QuestionId":42297980,"AnswerCount":1,"Tags":"<python><gpu><caffe>","CreationDate":"2017-02-17T12:27:49.557","AcceptedAnswerId":null,"OwnerUserId":249226.0,"Title":"Does Caffe's PythonLayer support GPU?","Body":"<p>Is it possible for the forward and\/or backward pass of a Caffe  to access the GPU?<\/p>\n\n<p>This question was recently ask here:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/5286\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/issues\/5286<\/a><\/p>\n","answers":[{"AnswerId":"42297981","CreationDate":"2017-02-17T12:27:49.557","ParentId":null,"OwnerUserId":"249226","Title":null,"Body":"<h2>Update<\/h2>\n\n<p>This is possible with the merging of the <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/5904\" rel=\"nofollow noreferrer\">Pull Request 5904<\/a>.<\/p>\n\n<h2>Before PR 5904<\/h2>\n\n<p>From the <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/ufpGi3VviVA\" rel=\"nofollow noreferrer\">Caffe-Users mailing<\/a> and the referenced pull request <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2102\" rel=\"nofollow noreferrer\">2102<\/a>,\nit appears that (as of this writing), Caffe does not support passing GPU data between custom Python layers. Since <code>PythonLayer<\/code> is full Python, it can access the GPU. It would just require data to be copied between the CPU and GPU multiple times.<\/p>\n"}]}
{"QuestionId":42299378,"AnswerCount":0,"Tags":"<python-3.x><machine-learning><tensorflow><gmm>","CreationDate":"2017-02-17T13:37:37.290","AcceptedAnswerId":null,"OwnerUserId":5217799.0,"Title":"ValueError: Features are incompatible with given information. Given features: Tensor(\"input:0\", shape=(?, 198), dtype=float32)","Body":"<p>Hi I am trying to implement GMM using tensorflow. But I am getting the following error :-\nValueError: Features are incompatible with given information. Given features: Tensor(\"input:0\", shape=(?, 198), dtype=float32), required signatures: TensorSignature(dtype=tf.float64, shape=TensorShape([Dimension(None), Dimension(198)]), is_sparse=False).<\/p>\n\n<p>Following is my code:-<\/p>\n\n<pre><\/pre>\n\n<p>x is a numpy array of shape (2384, 198)<\/p>\n\n<p>Stack Trace:-<\/p>\n\n<pre><\/pre>\n\n<p>Update:-<\/p>\n\n<pre><\/pre>\n\n<p>If pass a float64 data then I get the following error:-<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42299394,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-02-17T13:38:15.920","AcceptedAnswerId":null,"OwnerUserId":7581009.0,"Title":"Tensorflow error [image not found]","Body":"<p>I installed tensorflow ver1.0.0 but it did not work!!\nI use Python3.6.0, Tensorflow-gpu-1.0.0. Please tell me how to fix it!<\/p>\n\n<pre><\/pre>\n\n<p>seems above, I was able to install tensorflow, and activate tensorflow-gpu 1.0.0....\nIn reference to other similar question in stackoverflow, I disable csrutil!\n    csrutil disable<\/p>\n\n<p>Wasn't tensorflow installed..??<\/p>\n","answers":[{"AnswerId":"42552655","CreationDate":"2017-03-02T10:18:54.903","ParentId":null,"OwnerUserId":"7608560","Title":null,"Body":"<p>I solved the error by adding<\/p>\n\n<pre><code>$export DYLD_LIBRARY_PATH=\/Developer\/NVIDIA\/CUDA-8.0\/lib\\\n                     ${DYLD_LIBRARY_PATH:+:${DYLD_LIBRARY_PATH}}\n<\/code><\/pre>\n\n<p>like in the CUDA Installation Guide.<\/p>\n"}]}
{"QuestionId":42299857,"AnswerCount":0,"Tags":"<debugging><tensorflow><keras>","CreationDate":"2017-02-17T14:03:43.497","AcceptedAnswerId":null,"OwnerUserId":4402325.0,"Title":"Showing or saving intermediate image tensor in keras","Body":"<p>I am trying to debug a custom loss function and I would like to visualize the images generated during the intermediate computation step in the objective function. A tf_summary_image or a simple imshow would be perfect, but the summary it is not working without calling a sess.run() with a proper feed_dict. For simplicity, let's say I have:<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestions?<\/p>\n","answers":[]}
{"QuestionId":42299928,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-17T14:07:56.517","AcceptedAnswerId":null,"OwnerUserId":7581206.0,"Title":"Any existing implementation of distributed matrix multiplication in tensorflow?","Body":"<p>From the github code, it seems the MatMul op doesn't support partitioned matrixes. So is there any tool in tensorflow that supports multiplication of two huge matrixes that are distributed across multiple nodes?<\/p>\n","answers":[{"AnswerId":"42300288","CreationDate":"2017-02-17T14:24:16.980","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Support for distributing computation across machines is built into TensorFlow. I would recommend reading distributed TensorFlow docs to figure out how to setup a TensorFlow cluster.<\/p>\n\n<p>Once cluster is setup, you can decide how to partition your problem and use <code>with tf.device<\/code> to allocate each worker to their partition of work.<\/p>\n\n<p>For instance, suppose you are multiplying <code>a*a'<\/code>, and you want to split intermediate multiplication evenly over 2 workers, and the aggregate results on the 3rd.<\/p>\n\n<p>You would do something like this:<\/p>\n\n<pre><code>with tf.device(worker0):\n  # load a1\n  b1 = tf.matmul(a1, tf.transpose(a1))\n\nwith tf.device(worker1):\n  # load a2 \n  b2 = tf.matmul(a2, tf.transpose(a2))\n\nwith tf.device(worker2):\n  result = b1+b2\n<\/code><\/pre>\n\n<p>The <code>load a1<\/code> part depends on how big is your matrix is stored. If it's huge, then perhaps <code>load a1<\/code> will read it from disk. If it fits in memory, you can use <code>a1=a[:n\/2,:]<\/code> to get a partition of it<\/p>\n"}]}
{"QuestionId":42300339,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter-notebook>","CreationDate":"2017-02-17T14:26:50.797","AcceptedAnswerId":42309125.0,"OwnerUserId":833530.0,"Title":"Tensorflow model not training when changing the input data","Body":"<p>I am trying to use tensorflow to train a neural network (LeNet) using the traffic sign images.  I want to check the effect of a preprocessing technique on the performance of the nn. So, I preprocessed the images and stored the results (trainingimages, validationimages,testimages, final testimages) as a tuple in a dict.<\/p>\n\n<p>I then tried to iterate over this dict and then use the training and validation operations of the tensorflow as follows<\/p>\n\n<pre><\/pre>\n\n<p>The evaluate function is as follows<\/p>\n\n<pre><\/pre>\n\n<p>Once I execute the program, it works good for the first iteration of dataset but from the second iteration, the network doesnt train and continues to do so for all the other iterations. <\/p>\n\n<pre><\/pre>\n\n<p>However, if I restart the kernel and use any datatype (any iteration), it works. I figured out that I must clear the graph or run multiple sessions for multiple datatypes but I am not yet clear on how to do that. I tried using  but seems like it does not have any effect. Can somebody point me in the right direction?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42309125","CreationDate":"2017-02-17T23:26:08.063","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>You might want to make try with data that is normalized to zero mean and unit variance before feeding it to the network, e.g. by scaling images to <code>-1..1<\/code> range; that said, <code>0..1<\/code> range mostly sounds sane as well. Depending on the activations used in the network, the value range can make all the difference: ReLUs, for example, die out for inputs below zero, sigmoids start to saturate when values are below <code>-4<\/code> or above <code>+4<\/code> and tanh activations miss out on half of their value range if no value is ever below <code>0<\/code> - if the value range is too big, gradients may explode as well, preventing training altogether. From <a href=\"http:\/\/yann.lecun.com\/exdb\/publis\/pdf\/sermanet-ijcnn-11.pdf\" rel=\"nofollow noreferrer\">this paper<\/a>, the authors seem to subtract the (batch) image mean instead of the value range mean.<\/p>\n\n<p>You can try to use a <em>smaller<\/em> learning rate as well (although personally, I usually start experimenting around <code>0.0001<\/code> for Adam).<\/p>\n\n<p>As for your multiple sessions part of the question: The way it is currently implemented in your code, you are basically cluttering the default graph. By calling<\/p>\n\n<pre><code>for key in finalInputdata.keys():\n    for procTypes in range(0,(len(finalInputdata[key]))):\n        if np.shape(finalInputdata[key][procTypes][0]) != ():\n\n    # ...\n\n    x = tf.placeholder(tf.float32, (None, 32, 32,np.shape(X_train)[-1]))\n    y = tf.placeholder(tf.int32, (None))\n    one_hot_y = tf.one_hot(y,43)\n\n    # Tensor Operations\n    logits = LeNet(x,np.shape(X_train)[-1])\n\n    # ... etc ...\n<\/code><\/pre>\n\n<p>you are creating <code>len(finalInputdata) * N<\/code> <em>different<\/em> instances of LeNet, all within the default graph. That might be an issue when variables are internally reused in the network. <\/p>\n\n<p>If you do want to reset your <em>default<\/em> graph in order to try different hyperparameters, try<\/p>\n\n<pre><code>for key in finalInputdata.keys():\n    for procTypes in range(0,(len(finalInputdata[key]))):\n\n        tf.reset_default_graph()\n        # define the graph\n\n        sess = tf.InteractiveSession()\n        # train\n<\/code><\/pre>\n\n<p>but it is probably better to explictly create Graphs and Sessions like so:<\/p>\n\n<pre><code>for key in finalInputdata.keys():\n    for procTypes in range(0,(len(finalInputdata[key]))):\n\n        with tf.Graph().as_default() as graph:\n            # define the graph\n\n        with tf.Session(graph=graph) as sess:\n            # train\n<\/code><\/pre>\n\n<p>Instead of calling <code>sess = tf.get_default_session()<\/code> you would then directly use the <code>sess<\/code> reference.<\/p>\n\n<p>I also found that Jupyter kernels and GPU enabled TensorFlow don't play together that well when iterating on networks, sometimes running into out of memory errors or downright crashing the browser tab.<\/p>\n"}]}
{"QuestionId":42300475,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-17T14:33:30.890","AcceptedAnswerId":null,"OwnerUserId":7581282.0,"Title":"Change device placement of an existing tensorflow variable","Body":"<p>How do I change the device placement of a tf.Variable() ?\nI tried two methods<\/p>\n\n<pre><\/pre>\n\n<p>this creates a new variable on the gpu and doesn't change device assignment for <strong>a<\/strong><\/p>\n\n<p>I tried enforcing variable reuse by using <\/p>\n\n<pre><\/pre>\n\n<p>Here is the code :<\/p>\n\n<pre><\/pre>\n\n<p>this creates a new variable on the gpu and doesn't change device assignment for <strong>a<\/strong><\/p>\n\n<p>This time, I get an error saying the variable <strong>a<\/strong> did not exist in the gpu.<\/p>\n\n<p>Any help on changing-device-placement or lazy-device-assignment would be appreciated. Thanks<\/p>\n","answers":[]}
{"QuestionId":42301218,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-17T15:09:14.513","AcceptedAnswerId":42301461.0,"OwnerUserId":2877091.0,"Title":"Simple model gets 0.0 accuracy","Body":"<p>I am training a simple model on a dataset containing labels always equal to 0, and am getting a 0.0 accuracy.<\/p>\n\n<p>The model is the following:<\/p>\n\n<pre><\/pre>\n\n<p>And here is the dataset:<\/p>\n\n<pre><\/pre>\n\n<p>As the model trains, y of the data shown above decreases, and reaches -1000 after 1000 iterations.<\/p>\n\n<p>What could be the cause of the failure to train the model ?<\/p>\n","answers":[{"AnswerId":"42301461","CreationDate":"2017-02-17T15:19:43.193","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Your accuracy checks if the predicted float is exactly equal to the value you expect. With the network you made this is a very difficult task (although you might have a chance as you are also overfitting your data). <\/p>\n\n<p>To get better results: \n- Define accuracy to be higher\/lower than a value (closer to 1 or closer to 0). \n- Normalise your input data, I don't know the range of your input, but 444444 is a rediculous value to use as input, and it is difficult to train weights that can handle these values. <\/p>\n\n<p>Also: try to add some sanity checks. For example: what is the output your model is predicting? (y.eval) And what is the cross entropy you have during training your network? (sess.run([accuracy,cross_entropy], feed_dict={x: data, y_: labels})<\/p>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":42301677,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-17T15:30:22.370","AcceptedAnswerId":null,"OwnerUserId":6306884.0,"Title":"Find compute bottleneck","Body":"<p>I have a Resnet based network and it really slow<br>\nIt use 2 minutes to run one image   <\/p>\n\n<p>Here is what I did that tried to find bottleneck    <\/p>\n\n<p>I tried to use <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1824#issuecomment-225754659\" rel=\"nofollow noreferrer\">timeline<\/a> to find compute bottleneck<br>\nBut  show nothing.(the only change is =>)    <\/p>\n\n<p>The timeline file has value, file size is 885.9MB<\/p>\n\n<pre><\/pre>\n\n<p>When I saving the timeline file, console show below message, does it important?<\/p>\n\n<pre><\/pre>\n\n<p>BTW, will this kind of message means process slow down? It show each iteraction, but only show  at first iteraction.<\/p>\n\n<pre><\/pre>\n\n<p>At first, I think PoolAllocator caused by limited GPU memory.\nAfter check  with , I found the GPU memory usage is  far away from not enough, right?<\/p>\n\n<p>Follow this <a href=\"https:\/\/stackoverflow.com\/questions\/39774449\/fully-convolutional-resnets-using-tf-slim-run-very-slow\">Question<\/a>, I add , doesn't work:(<\/p>\n","answers":[]}
{"QuestionId":42301781,"AnswerCount":0,"Tags":"<tensorflow><skflow>","CreationDate":"2017-02-17T15:35:07.280","AcceptedAnswerId":null,"OwnerUserId":7581542.0,"Title":"Weight\/bias initialization in Skflow","Body":"<p>I've been using TensorFlow to create feed-forward neural nets with customized initial weights and biases (cannot provide code for proprietary reasons, but another algorithm outputs initial matrices that are then plugged into tf for optimization). \nEverything has been working beautifully for classification, however when I started testing on regression datasets I was getting enormously large errors. I ran the skflow Boston housing data example (DNN regressor) and it gets much better results. Like others (<a href=\"https:\/\/stackoverflow.com\/questions\/36753304\/tensorflow-model-has-different-results-than-the-same-model-in-skflow-optimizer\">tensorflow model has different results than the same model in skflow (optimizer)<\/a>) I tried reproducing the skflow code in tf (abandoning my weight initialization idea) and am getting drastically worse performance in tf. <\/p>\n\n<p>So I'd like ideas on one of two things, either:\n1. How can you initialize with custom weight matrices in skflow? Digging in the source code has yet to reveal simple solutions...\nor\n2. What is the key difference between the skflow dnn regressor and the code written in tensorflow (what is the answer to the question I linked to)? <\/p>\n","answers":[]}
{"QuestionId":42302498,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-17T16:12:29.757","AcceptedAnswerId":42373487.0,"OwnerUserId":7508642.0,"Title":"Converting TensorFlow tutorial to work with my own data","Body":"<p>This is a follow on from my last question <a href=\"https:\/\/stackoverflow.com\/questions\/42286972\/converting-from-pandas-dataframe-to-tensorflow-tensor-obeject\/42302125#42302125\">Converting from Pandas dataframe to TensorFlow tensor object<\/a><\/p>\n\n<p>I'm now on the next step and need some more help. I'm trying to replace this line of code<\/p>\n\n<pre><\/pre>\n\n<p>with a replacement for my own data. I've found this answer on StackOverflow: <a href=\"https:\/\/stackoverflow.com\/questions\/40368697\/where-does-next-batch-in-the-tensorflow-tutorial-batch-xs-batch-ys-mnist-trai\">Where does next_batch in the TensorFlow tutorial batch_xs, batch_ys = mnist.train.next_batch(100) come from?<\/a> But I don't understand:<\/p>\n\n<p>1) Why the .next_batch() doesn't work on my tensor. Am I creating it incorrectly<\/p>\n\n<p>2) How to implement the pseudocode that was given in the answer to the question on .next_batch()<\/p>\n\n<p>I currently have two tensor objects, one with the parameters I wish to use to train the model (dataVar_tensor) and one with the correct result (depth_tensor). I obviously need to keep their relationship to keep the correct response with the correct parameters. <\/p>\n\n<p>Please can you take some time to help me understand what's going on and to replace this line of code?<\/p>\n\n<p>Many thanks<\/p>\n","answers":[{"AnswerId":"42373487","CreationDate":"2017-02-21T16:59:28.620","ParentId":null,"OwnerUserId":"7241513","Title":null,"Body":"<p>I stripped off the non-relevant stuff so as to preserve the formatting and indentation. Hopefully it should be clear now. The following code reads a CSV file in batches of N lines (N specified in a constant at the top). Each line contains a date (first cell), then a list of floats (480 cells) and a one-hot vector (3 cells). The code then simply prints the batches of these dates, floats, and one-hot vector as it reads them. The place where it prints them is normally where you'd actually run your model and feed these in place of the placeholder variables.<\/p>\n\n<p>Just keep in mind that here it reads each line as a String, and then converts the specific cells within that line into floats, simply because the first cell is easier to read as a string. If all your data is numeric, then simply set the defaults into a float\/int rather than an 'a' and get rid of the code that converts strings to floats. It's not needed otherwise!<\/p>\n\n<p>I put some comments to clarify what it's doing. Let me know if something is unclear.<\/p>\n\n<pre><code>import tensorflow as tf\n\nfileName = 'YOUR_FILE.csv'\n\ntry_epochs = 1\nbatch_size = 3\n\nTD = 1 # this is my date-label for each row, for internal pruposes\nTS = 480 # this is the list of features, 480 in this case\nTL = 3 # this is one-hot vector of 3 representing the label\n\n# set defaults to something (TF requires defaults for the number of cells you are going to read)\nrDefaults = [['a'] for row in range((TD+TS+TL))]\n\n# function that reads the input file, line-by-line\ndef read_from_csv(filename_queue):\n    reader = tf.TextLineReader(skip_header_lines=False) # i have no header file\n    _, csv_row = reader.read(filename_queue) # read one line\n    data = tf.decode_csv(csv_row, record_defaults=rDefaults) # use defaults for this line (in case of missing data)\n    dateLbl = tf.slice(data, [0], [TD]) # first cell is my 'date-label' for internal pruposes\n    features = tf.string_to_number(tf.slice(data, [TD], [TS]), tf.float32) # cells 2-480 is the list of features\n    label = tf.string_to_number(tf.slice(data, [TD+TS], [TL]), tf.float32) # the remainin 3 cells is the list for one-hot label\n    return dateLbl, features, label\n\n# function that packs each read line into batches of specified size\ndef input_pipeline(fName, batch_size, num_epochs=None):\n    filename_queue = tf.train.string_input_producer(\n        [fName],\n        num_epochs=num_epochs,\n        shuffle=True)  # this refers to multiple files, not line items within files\n    dateLbl, features, label = read_from_csv(filename_queue)\n    min_after_dequeue = 10000 # min of where to start loading into memory\n    capacity = min_after_dequeue + 3 * batch_size # max of how much to load into memory\n    # this packs the above lines into a batch of size you specify:\n    dateLbl_batch, feature_batch, label_batch = tf.train.shuffle_batch(\n        [dateLbl, features, label], \n        batch_size=batch_size,\n        capacity=capacity,\n        min_after_dequeue=min_after_dequeue)\n    return dateLbl_batch, feature_batch, label_batch\n\n# these are the date label, features, and label:\ndateLbl, features, labels = input_pipeline(fileName, batch_size, try_epochs)\n\nwith tf.Session() as sess:\n\n    gInit = tf.global_variables_initializer().run()\n    lInit = tf.local_variables_initializer().run()\n\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n\n    try:\n        while not coord.should_stop():\n            # load date-label, features, and label:\n            dateLbl_batch, feature_batch, label_batch = sess.run([dateLbl, features, labels])      \n\n            print(dateLbl_batch);\n            print(feature_batch);\n            print(label_batch);\n            print('----------');\n\n    except tf.errors.OutOfRangeError:\n        print(\"Done looping through the file\")\n\n    finally:\n        coord.request_stop()\n\n    coord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":42303299,"AnswerCount":1,"Tags":"<python><theano><keras>","CreationDate":"2017-02-17T16:50:05.230","AcceptedAnswerId":42303643.0,"OwnerUserId":1627234.0,"Title":"Keras error compiling node","Body":"<p>I am trying to train a small CNN using Keras with the  like so:<\/p>\n\n<pre><\/pre>\n\n<p>I am trying to solve a binary classification problem, but I'm getting the following error<\/p>\n\n<blockquote>\n  <p>Exception: ('The following error happened while compiling the node',\n  GpuElemwise{RoundHalfToEven,no_inplace}(GpuSoftmaxWithBias.0)<\/p>\n<\/blockquote>\n\n<p>followed by a large amount of cuda options. The line this is failing at is<\/p>\n\n<pre><\/pre>\n\n<p>I am completely lost as to what this could be, I've tried several different architectures for the cnn, I've also verified that thte ImageDataGenerator works fine. I have not been able to figure out what the problem could be.<\/p>\n\n<p>I'm using Python 3.6.0, Theano 0.8.2 and Keras 1.2.2<\/p>\n","answers":[{"AnswerId":"42303643","CreationDate":"2017-02-17T17:09:37.250","ParentId":null,"OwnerUserId":"1627234","Title":null,"Body":"<p>Doing some more digging I found a seemingly unrelated <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/5408\" rel=\"nofollow noreferrer\">issue<\/a>, but that had the same error message.<\/p>\n\n<p>Apparently, this is an issue with the particular version of Theano, and had already been fixed in the master branch. Running <\/p>\n\n<pre><code>pip install --upgrade git+https:\/\/github.com\/Theano\/Theano.git#egg=Theano\n<\/code><\/pre>\n\n<p>solved my problem.<\/p>\n"}]}
{"QuestionId":42303336,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-17T16:52:34.663","AcceptedAnswerId":42305995.0,"OwnerUserId":5359914.0,"Title":"Tensorflow math operation reduce_sum","Body":"<pre><\/pre>\n\n<p>The result is :<\/p>\n\n<pre><\/pre>\n\n<p>But it is not the answer what i want.<\/p>\n\n<p>I want to get the answer <\/p>\n\n<p>[5+8+9+8+5,1+4+9+16+25,1+4+9+16+25]\n                        =[35,55,55]<\/p>\n","answers":[{"AnswerId":"42305995","CreationDate":"2017-02-17T19:31:35.377","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As xxi mentioned in <a href=\"https:\/\/stackoverflow.com\/questions\/42303336\/tensorflow-math-operation-reduce-sum#comment71762288_42303336\">their comment<\/a>, the correct solution here is to use the optional <code>axis<\/code> argument when calling <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_sum\" rel=\"nofollow noreferrer\"><code>tf.reduce_sum()<\/code><\/a>. In your case, you want to reduce along the column axis, so the following code will work:<\/p>\n\n<pre><code>product = tf.multiply(a, b)\nproduct_sum = tf.reduce_sum(product, axis=1)\n<\/code><\/pre>\n\n<p>(Note also that in TensorFlow 1.0, <code>tf.mul()<\/code> is now <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/multiply\" rel=\"nofollow noreferrer\"><code>tf.multiply()<\/code><\/a>.)<\/p>\n"}]}
{"QuestionId":42303399,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-17T16:55:35.403","AcceptedAnswerId":42303930.0,"OwnerUserId":5171213.0,"Title":"Keras: masking zero-padded input for non-RNN","Body":"<p>A colleague of mine pointed out the very cool option to use  instead of a masking layer when you need to mask input to a non-RNN in Keras. <\/p>\n\n<p>In my case, I have 62 columns in the input, with the 63rd being the response. Over 97% of the nonzero entries in the 62 columns are contained in the first 30 columns. I'm trying to just get this working, so I'd like to weight the last 32 columns to be 0 in training, essentially creating a 'poor-man's mask'. <\/p>\n\n<p>This is an 8-class classification task, using an MLP. The response variable has been transformed using the  function in Keras.<\/p>\n\n<p>Here's the implementation:<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix my  to 'mask' the first 32 columns of the input?<\/p>\n","answers":[{"AnswerId":"42303930","CreationDate":"2017-02-17T17:24:00.863","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Sample weight isn't working like that:<\/p>\n\n<blockquote>\n  <p><code>sample_weight<\/code>: optional array of the same length as <code>x<\/code>, containing weights to apply to the model's loss for each sample. In the case of temporal data, you can pass a 2D array with shape <code>(samples, sequence_length)<\/code>, to apply a different weight to every timestep of every sample. In this case you should make sure to specify <code>sample_weight_mode=\"temporal\"<\/code> in <code>compile()<\/code>. <a href=\"https:\/\/keras.io\/models\/model\/#fit\" rel=\"nofollow noreferrer\">source<\/a><\/p>\n<\/blockquote>\n\n<p>In other words, this setting puts different weights on the <strong>samples<\/strong> of the training data, not on the features of each sample. This is used only at training step.\nI think you should use masking if you don't want the layer to use those features. Or just remove them from your dataset? Or, if it's not too complicated, let the network learn by itself which the useful features are.<\/p>\n\n<p>Does this help?<\/p>\n"}]}
{"QuestionId":42303577,"AnswerCount":1,"Tags":"<machine-learning><computer-vision><deep-learning><keras><conv-neural-network>","CreationDate":"2017-02-17T17:04:50.793","AcceptedAnswerId":null,"OwnerUserId":6522749.0,"Title":"How to apply mean\/average pooling over the batch size to get a single output for the whole batch in Keras?","Body":"<p>For eg.- the input with dimensions [10,1,224,224] is required to be reduced to [1,1,224,224] where [samples,channels,rows,columns] is the convention for the dimensions.<\/p>\n","answers":[{"AnswerId":"42303784","CreationDate":"2017-02-17T17:16:49.157","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Then your problem is badly formuled, consider using [10,1,224,224] as input_shape and make batches of such tensors. Then use Averagepooling3D, see doc <a href=\"https:\/\/keras.io\/layers\/pooling\/#averagepooling3d\" rel=\"nofollow noreferrer\">here<\/a>.\nYou won't be able to make operations on batches with the usual layers, except maybe if you build your own custom layer : see <a href=\"https:\/\/keras.io\/layers\/writing-your-own-keras-layers\/\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"}]}
{"QuestionId":42305812,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-17T19:19:38.313","AcceptedAnswerId":42308448.0,"OwnerUserId":4172685.0,"Title":"Using summary with tf slim or tf layers","Body":"<p>I cannot not find how to add a summary to visualize weights for networks which have been defined with  or .<\/p>\n\n<p>For instance, if I have:<\/p>\n\n<pre><\/pre>\n\n<p>How can I add the weights and bias into a summary ?<\/p>\n\n<p>Edit: I just saw that I could use . That probably works for what I want to do.<\/p>\n","answers":[{"AnswerId":"42308448","CreationDate":"2017-02-17T22:27:33.183","ParentId":null,"OwnerUserId":"4172685","Title":null,"Body":"<p>Here is the solution I found after deeper investigation, in case someone has the same problem than me.<\/p>\n\n<p>For <code>tf.contrib.slim<\/code>, (deprecated) the weights are added to the collection <code>tf.GraphKeys.TRAINABLE_VARIABLES<\/code> which can be visualized with:<\/p>\n\n<pre><code>tf.contrib.layers.summarize_collection(tf.GraphKeys.TRAINABLE_VARIABLES)\n<\/code><\/pre>\n\n<p>For <code>tf.contrib.layers<\/code>, there are some convenience methods<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>tf.contrib.layers.summarize_weights()  # tf.GraphKeys.WEIGHTS\ntf.contrib.layers.summarize_biases()  # tf.GraphKeys.BIASES\n# For tf &gt;= 1.3:\ntf.contrib.layers.summarize_tensors()\n<\/code><\/pre>\n\n<p>To have a better control on which variables add, it's possible by customizing the code of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/layers\/python\/layers\/summaries.py\" rel=\"noreferrer\">summarize_collection<\/a>.<\/p>\n"}]}
{"QuestionId":42306484,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-17T20:06:35.243","AcceptedAnswerId":45630180.0,"OwnerUserId":577666.0,"Title":"Configure input_map when importing a tensorflow model from metagraph file","Body":"\n\n<p>I've trained a DCGAN model and would now like to load it into a library that visualizes the drivers of neuron activation through image space optimization.<\/p>\n\n<p>The following code works, but forces me to work with (1, width, height, channels) images when doing subsequent image analysis, which is a pain (the library assumptions about the shape of network input).<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I'd like to change the input_map, After reading the source, I expected this code to work:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>But got an error:<\/p>\n\n<blockquote>\n  <p>ValueError: tf.import_graph_def() requires a non-empty  if  is used.<\/p>\n<\/blockquote>\n\n<p>When the stack gets down to  the name field is set to import_scope, so I tried the following:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Which netted me the following :<\/p>\n\n<blockquote>\n  <p>KeyError: \"The name 'gradients\/discriminator\/minibatch\/map\/while\/TensorArrayWrite\/TensorArrayWriteV3_grad\/TensorArrayReadV3\/RefEnter:0' refers to a Tensor which does not exist. The operation, 'gradients\/discriminator\/minibatch\/map\/while\/TensorArrayWrite\/TensorArrayWriteV3_grad\/TensorArrayReadV3\/RefEnter', does not exist in the graph.\"<\/p>\n<\/blockquote>\n\n<p>If I set 'import_scope', I get the same error whether or not I set 'input_map'. <\/p>\n\n<p>I'm not sure where to go from here.<\/p>\n","answers":[{"AnswerId":"45630180","CreationDate":"2017-08-11T08:14:25.107","ParentId":null,"OwnerUserId":"1896918","Title":null,"Body":"<p>In the newer version of tensorflow>=1.2.0, the following step works fine.<\/p>\n\n<pre><code>t_input = tf.placeholder(np.float32, shape=[None, width, height, channels], name='new_input') # define the input tensor\n\n# here you need to give the name of the original model input placeholder name\n# For example if the model has input as; input_original=  tf.placeholder(tf.float32, shape=(1, width, height, channels, name='original_placeholder_name'))\nnew_saver = tf.train.import_meta_graph(\/path\/to\/checkpoint_file.meta, input_map={'original_placeholder_name:0':  t_input})\nnew_saver.restore(sess, '\/path\/to\/checkpointfile')\n<\/code><\/pre>\n"},{"AnswerId":"47586663","CreationDate":"2017-12-01T04:31:04.360","ParentId":null,"OwnerUserId":"2153958","Title":null,"Body":"<p>So, the main issue is that you're not using the syntax right. Check the documentation for <code>tf.import_graph_def<\/code> for the use of <code>input_map<\/code> (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/import_graph_def\" rel=\"nofollow noreferrer\">link<\/a>).<\/p>\n\n<p>Let's breakdown this line:<\/p>\n\n<pre><code>new_saver = tf.train.import_meta_graph(model_fn, input_map={'images': t_input}, import_scope='import')\n<\/code><\/pre>\n\n<p>You didn't outline what <code>model_fn<\/code> is, but it needs to be a path to the file.\nFor the next part, in <code>input_map<\/code>, you're saying: <em>replace<\/em> the input in the <em>original graph<\/em> (DCGAN) whose <code>name<\/code> is <code>images<\/code> with my variable (in the current graph) called <code>t_input<\/code>. Problematically, <code>t_input<\/code> and <code>images<\/code> are referencing the same object in different ways as per this line:<\/p>\n\n<pre><code> t_input = tf.placeholder(np.float32, name='images')\n<\/code><\/pre>\n\n<p>In other words, <code>images<\/code> in <code>input_map<\/code> should actually be whatever the variable name is that you're trying to replace in the DCGAN graph. You'll have to import the graph in its base form (i.e., without the <code>input_map<\/code> line) and figure out what the name of the variable you want to link to is. It'll be in the list returned by <code>tf.get_collection('variables')<\/code> after you have imported the graph. Look for the dimensions (1, width, height, channels), but with the values in place of the variable names. If it's a placeholder, it'll look something like <code>scope\/Placeholder:0<\/code> where <code>scope<\/code> is replaced with whatever the variable's scope is.<\/p>\n\n<p>Word of caution:<\/p>\n\n<p>Tensorflow is very finicky about what it expects graphs to look like. So, if in the original graph specification the width, height, and channels are explicitly specified, then Tensorflow will complain (throw an error) when you try to connect a <code>placeholder<\/code> with a different set of dimensions. And, this makes sense. If the system was trained with some set of dimensions, then it only knows how to generate images with those dimensions.<\/p>\n\n<p>In theory, you can still stick all kinds of weird stuff on the front of that network. But, you will need to scale it down so it meets those dimensions first (and the Tensorflow documentation says it's better to do that with the CPU outside of the graph; i.e., before inputing it with <code>feed_dict<\/code>).<\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":42306632,"AnswerCount":1,"Tags":"<python><tensorflow><computer-vision><deep-learning><keras>","CreationDate":"2017-02-17T20:15:24.667","AcceptedAnswerId":46287502.0,"OwnerUserId":4674270.0,"Title":"Fitting a Keras model with large amount of data using a custom data generator","Body":"<p>I'm trying to fit my Keras model with quite large amount of data.<\/p>\n\n<p>To do this, I'm using custom data generators and  function.<\/p>\n\n<p>However, I can't seem to understand if I'm doing this correctly.<\/p>\n\n<p>Here's what I have:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>My dataset consists of images with names like:<\/p>\n\n<ul>\n<li><p>, i.e.: <\/p><\/li>\n<li><p>, i.e.: <\/p><\/li>\n<\/ul>\n\n<p>So, basically, I'm trying to train a model to perform a binary cat-dog classification.<\/p>\n\n<hr>\n\n<p>Is my  function correctly implemented for mini-batch optimization?<\/p>\n\n<p>How can I add the usage of  to my  function (from the  function)?<\/p>\n","answers":[{"AnswerId":"46287502","CreationDate":"2017-09-18T20:14:24.107","ParentId":null,"OwnerUserId":"4674270","Title":null,"Body":"<p>Okay, here's the github link to my final version of the project that I got working:<\/p>\n\n<p><a href=\"https:\/\/github.com\/yakovenkodenis\/dogs-vs-cats-kaggle\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yakovenkodenis\/dogs-vs-cats-kaggle<\/a><\/p>\n\n<p>Hope, it helps somebody<\/p>\n"}]}
{"QuestionId":42306848,"AnswerCount":0,"Tags":"<python><gpu><theano>","CreationDate":"2017-02-17T20:30:10.440","AcceptedAnswerId":null,"OwnerUserId":7581502.0,"Title":"python, theano, GPU, PicklingError: Can't pickle","Body":"<p>I'm new to python and theano. I want to bring a demo to work, which claims to parallelize an optimization of an objective function, compiled in theano, on a GPU.<\/p>\n\n<p>The demo can be found here <a href=\"https:\/\/github.com\/yaringal\/VSSGP\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yaringal\/VSSGP<\/a><\/p>\n\n<p>I had to modify some code in order to run it on windows. For the file  I changed it to:<\/p>\n\n<pre><\/pre>\n\n<p>So I have put the most code in a new function, STARTME(), so that it is not on modul stage.<\/p>\n\n<p>I also changed  to:<\/p>\n\n<pre><\/pre>\n\n<p>to get rid of the global variables. Now when I execute the function STARTME():<\/p>\n\n<pre><\/pre>\n\n<p>it runs fine until line 53 of , where I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>What can I fix, to bring it to work?\nI already tested with a test-file, whether my system works on a GPU on a theano example of their page, and it was true, so this can't be the error.<\/p>\n\n<p>Thanks for your help!!<\/p>\n\n<p>Romaxx<\/p>\n","answers":[]}
{"QuestionId":42307908,"AnswerCount":1,"Tags":"<python><hadoop><tensorflow>","CreationDate":"2017-02-17T21:42:59.153","AcceptedAnswerId":null,"OwnerUserId":7582975.0,"Title":"Running tensorflow with file on HDFS (cannot find libhdfs.so)","Body":"<p>I am running into an error \"libhdfs.so: cannot open shared object file: No such file or directory\" (stack trace below) while trying to run a python script invoking a Tensorflow reader on a file stored in HDFS. I am running the script  on a node on the cluster which has Tensorflow in a virtualenv, activated at the time of execution. I set the following environment variables before execution:<\/p>\n\n<ul>\n<li>export HADOOP_HDFS_HOME=$HADOOP_HDFS_HOME:\/opt\/cloudera\/parcels\/CDH<\/li>\n<li>export JAVA_HOME=\/usr\/lib\/jvm\/java-7-openjdk-amd64<\/li>\n<li>export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:\/opt\/cloudera\/parcels\/CDH\/lib\/libhdfs.so<\/li>\n<li>export\nLD_LIBRARY_PATH=$LD_LIBRARY_PATH:${JAVA_HOME}\/jre\/lib\/amd64\/server<\/li>\n<\/ul>\n\n<p>I execute the script as such:<\/p>\n\n<ul>\n<li>CLASSPATH=$($LD_LIBRARY_PATH} classpath --glob) python TEST.py<\/li>\n<\/ul>\n\n<p>This is the code in the script:<\/p>\n\n<pre><\/pre>\n\n<p>Below is the stack trace of the error. Any ideas on what is causing this error is appreciated (I have already checked that the LD_LIBRARY_PATH variable has an explicit pointer to the libhdfs.so file prior to execution, unable to figure out why it still cannot find the file).<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43490524","CreationDate":"2017-04-19T08:35:15.520","ParentId":null,"OwnerUserId":"4023592","Title":null,"Body":"<p>I also encountered this issue the solution for me was copying this file to:<\/p>\n\n<p>$HADOOP_HDFS_HOME\/lib\/native<\/p>\n\n<p>If you don't know the location of this file execute the following command to find it's location:<\/p>\n\n<pre><code>sudo updatedb\nlocate libhdfs.so\n<\/code><\/pre>\n\n<p>This will give you the location of the file. Next copy the file to $HADOOP_HDFS_HOME\/lib\/native:<\/p>\n\n<pre><code>cp locationOflibhdfs.so $HADOOP_HDFS_HOME\/lib\/native\n<\/code><\/pre>\n\n<p>Note: Replace locationOflibhdfs.so with the location of the libhdfs.so file.<\/p>\n"}]}
{"QuestionId":42307975,"AnswerCount":1,"Tags":"<memory><tensorflow><gpu><cpu>","CreationDate":"2017-02-17T21:47:31.197","AcceptedAnswerId":null,"OwnerUserId":582967.0,"Title":"Memory management when using GPU in TensorFlow","Body":"<p>I have some doubts about using GPU in Tensorflow. I was following convolutional neural network tutorial <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\" rel=\"nofollow noreferrer\">here<\/a> (tensorflow\/models\/image\/cifar10\/cifar10_train.py). As in the tutorial, all parameters (e.g., weights) are stored and updated in CPU memory and GPUs are only used to compute gradients or inference. <\/p>\n\n<p>Since the weights are stored in CPU, they should be synchronized every iteration and it seems that GPU is underutilized (about 60% according to ). In case of using multiple GPUs, I understand that weights should be stored in CPU memory to synchronize between the GPUs. However, why does this tutorial store all weights in CPU even in single GPU? Is there any way to store and update them in GPU memory?<\/p>\n\n<p>In case of inference, does the weights copied to GPU once and reuse them? or should they copied every time they are used?<\/p>\n\n<p>How about image data? It seems that those data reside in GPU (not sure). When does this data transferred to GPU? When they are loaded from disk? or when they are required in the GPU?<\/p>\n\n<ul>\n<li>If they are copied to GPU right after they are loaded from disk, what happens if size of the image data is too large to fit in the GPU memory? In such case, there is any way to copy data separately (something like prefetching)?<\/li>\n<li>If they are copied to GPU on demand, is there any way to prefetch them before they are actually used by GPU to avoid idle time? <\/li>\n<\/ul>\n\n<p><strong>EDIT:<\/strong> It would be helpful if there is any way to check where the send\/recv nodes are inserted between CPU and GPU (as in the <a href=\"https:\/\/static.googleusercontent.com\/media\/research.google.com\/en\/\/pubs\/archive\/45166.pdf\" rel=\"nofollow noreferrer\">white paper<\/a>).<\/p>\n","answers":[{"AnswerId":"42310498","CreationDate":"2017-02-18T02:44:24.583","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Those tutorials are meant to show off the API, so they don't optimize for performance. It's faster to keep variable on GPU for single tower model, and also faster for multi-tower model when you have p2p communication enabled between GPU. To pin variables to GPU, use the same <code>tf.device('\/gpu:0')<\/code> approach as for any other op.<\/p>\n\n<p>You can see all the memory copies between GPUs if you enable partition graphs, ie do something like this:<\/p>\n\n<pre><code>metadata = tf.RunMetadata()\nsess.run(x, options=tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE,\n                                  output_partition_graphs=True),\n         run_metadata=metadata)\n\ntimeline = Timeline(metadata.step_stats)\nwith open(\"dynamic_stitch_gpu_profile.json\", \"w\") as f:\n    f.write(timeline.generate_chrome_trace_format())\nwith open(\"dynamic_stitch_gpu_profile.pbtxt\", \"w\") as f:\n    f.write(str(metadata))\n<\/code><\/pre>\n\n<p>See this issue for an example of using this technique to track down copies:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7251#issuecomment-277385212\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7251#issuecomment-277385212<\/a><\/p>\n\n<p>For prefetching to GPU, see this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5722\" rel=\"nofollow noreferrer\">issue<\/a><\/p>\n\n<p>There are new <code>stage_op<\/code> ops that have been added that allow prefetching to GPU and are dramatically faster than using Python queue runner approach. They are in process of being documented.<\/p>\n"}]}
{"QuestionId":42308475,"AnswerCount":1,"Tags":"<boost><cmake><caffe>","CreationDate":"2017-02-17T22:29:54.427","AcceptedAnswerId":null,"OwnerUserId":1487170.0,"Title":"Undefined Boost Thread Reference when Compiling Caffe Code","Body":"<p>I have installed Boost 1.63.0 locally, as well as Caffe (version rc3). I am currently trying to compile some Caffe code using CMake, but I get the following linker error:<\/p>\n\n<pre><\/pre>\n\n<p>My first thought is that I'm somehow linking to Boost the wrong way in my CMake file. Right now, in CMakeLists.txt, I have:<\/p>\n\n<pre><\/pre>\n\n<p>My second thought was that this had something to do with pthreads, so I linked to pthreads in my CMakeLists.txt using this StackOverflow answer: <a href=\"https:\/\/stackoverflow.com\/questions\/1620918\/cmake-and-libpthread\">cmake and libpthread<\/a>. However, I still get the same error after doing this.<\/p>\n\n<p>Does anyone have other suggestions about what might be the problem here? Thank you.<\/p>\n","answers":[{"AnswerId":"43689040","CreationDate":"2017-04-28T21:20:30.060","ParentId":null,"OwnerUserId":"3399850","Title":null,"Body":"<p>When compiling <code>Caffe<\/code> using a locally installed <code>Boost 1.63<\/code>, source code build, with <code>CMake 3.7<\/code>, I've run into a similar issues.  The <code>Caffe 1.0<\/code> <code>CMake<\/code> folder contains <code>Dependencies.cmake<\/code> and <code>Caffe<\/code> is looking for <code>Boost 1.45<\/code> not <code>Boost 1.62<\/code> or <code>Boost 1.63<\/code> and <code>Python 3.5.3<\/code> <code>Ubuntu 17.04<\/code> launchpad distribution files.  The locally installed file <code>libboost_python.so<\/code> or <code>libboost_python.so.1.63.0<\/code> in folder <code>\/usr\/local\/lib<\/code> are differently named from the launchpad distribution files in folder <code>\/usr\/lib\/x86_64-linux-gnu<\/code> <code>libboost_python-py35.so<\/code>.<\/p>\n"}]}
{"QuestionId":42308481,"AnswerCount":1,"Tags":"<python><deep-learning><keras><keras-layer>","CreationDate":"2017-02-17T22:30:17.457","AcceptedAnswerId":null,"OwnerUserId":7526688.0,"Title":"Matrix Size Error with Keras Model","Body":"<p>Having some issues with keras.  I am just trying to build the model and get it running and then tune it in.  So that being said I am only using 99 images and 99 labels.  For reference I am using it to give me a continuous output not just a class label.  Below is the code I am using.  First I have a script that imports all the data.  99 images and 99 labels.  <\/p>\n\n<p>When I get to the fitting of the model part it tosses me an error.  \"ValueError: Error when checking model target: expected cropping2d_1 to have 4 dimensions, but got array with shape (99, 1)\". <\/p>\n\n<p>I read some other threads about similar errors and it seems it could be the order of the array I am sending keras.  I played around with it and got the following.  Currently the shape of the images array is (99,160,320,3).  I tried changing the order of the \"input_shape\" in keras to (3,160,320).  This gave me and error \"ValueError: Error when checking model input: expected cropping2d_input_1 to have shape (None, 3, 160, 320) but got array with shape (99, 160, 320, 3)\". I then reshaped the images_center array accordingly at got the same error as above. <\/p>\n\n<p>I've left out the import statements just to keep it short here.  <\/p>\n\n<p>Any thoughts on next steps?  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42319074","CreationDate":"2017-02-18T18:31:01.363","ParentId":null,"OwnerUserId":"7558284","Title":null,"Body":"<p>Your labels (i.e. the \"target\") are of shape (99, 1) so the network should produce an output in the same shape. Try adding a fully connected layer at the end, like <code>model.add(Dense(1))<\/code>.<\/p>\n"}]}
{"QuestionId":42308650,"AnswerCount":0,"Tags":"<python><tensorflow><lstm><recurrent-neural-network><mnist>","CreationDate":"2017-02-17T22:43:29.110","AcceptedAnswerId":null,"OwnerUserId":6659899.0,"Title":"TensorFlow LSTM to classify MNIST - understanding changing number of timesteps","Body":"<p><strong>Context<\/strong><\/p>\n\n<p>I am trying to understand RNNs, both implementation (in TensorFlow) and theory. As part of this I've written a simple LSTM using TensorFlow to classify MNIST handwritten digits.<\/p>\n\n<p>For this I use TensorFlow's , with an input shape of  (and ). <\/p>\n\n<p>When I feed the MNIST image into the model with 28 timesteps, and 28 pixels input at each timestep (for 784 total pixels), the model works well, training fast and achieving high accuracy (~1 min\/epoch, 98% accuracy with 128 hidden units).<\/p>\n\n<p>However, if I feed the image into the model pixel-by-pixel, so that there are 784 timesteps each with input size 1, the model performs extremely poorly (~30 mins\/epoch, max accuracy of 40%).<\/p>\n\n<p><br\/><\/p>\n\n<p><strong>Questions<\/strong>: <\/p>\n\n<ul>\n<li>What exactly is going on here? Why does feeding the images in pixel-by-pixel cause the model to perform so poorly? Is it to do with there not being enough context or there being too many timesteps or must there be something wrong with the model?<\/li>\n<li>Is there any way to change this so that the model will work well when fed the images pixel-by-pixel? I have heard of truncating the backpropagation at a certain number of timesteps, but the TensorFlow docs are unclear on how exactly to do that and I haven't found any decent guides yet.<\/li>\n<\/ul>\n\n<p><br\/><\/p>\n\n<p><strong>Code, If It's Helpful:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42308979,"AnswerCount":0,"Tags":"<python><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-17T23:12:10.163","AcceptedAnswerId":null,"OwnerUserId":288609.0,"Title":"datasets module in python for imagenet","Body":"<p>In one of the programs, author uses the following code to input imagenet<\/p>\n\n<pre><\/pre>\n\n<p>Due to some system constraints, I cannot have  module installed. Are there any walk-around so that I can get the  from imagenet?<\/p>\n","answers":[]}
{"QuestionId":42309202,"AnswerCount":1,"Tags":"<tensorflow><profiling><gpu>","CreationDate":"2017-02-17T23:33:09.133","AcceptedAnswerId":null,"OwnerUserId":582967.0,"Title":"Profiling TensorFlow using tfprof","Body":"<p>I am trying to profile computation\/memory usage of TensorFlow and found that <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/tfprof\" rel=\"noreferrer\">tfprof<\/a> is a right tool for my purpose. However, I was not able to get FLOPS of all operators. <\/p>\n\n<p>Here is what I did following the tfprof tutorial using cifar10 tutorial in TensorFlow repository (tensorflow\/models\/image\/cifar10\/cifar10_train.py):<\/p>\n\n<pre><\/pre>\n\n<p>And the result is<\/p>\n\n<pre><\/pre>\n\n<p>However, the result does not contain all of the ops such as max pooling, relu, gradient of conv layers. Maybe flops stats of those ops are not defined (RegisterStatistics('flops')). Therefore, to provide runtime information, as in the tfprof tutorial 11), I tried to create  (See code above).<\/p>\n\n<p>However, I am not sure how can I add op information (How can I get entry name of the ops?). Is there any way to add <em>ALL<\/em> ops it contains?<\/p>\n\n<p>Or any other tool rather than tfprof? Perhaps profiling tool from NVIDIA?<\/p>\n","answers":[{"AnswerId":"42310726","CreationDate":"2017-02-18T03:26:25.407","ParentId":null,"OwnerUserId":"7583767","Title":null,"Body":"<p>You are right that the other ops don't have flops before they don't have RegisterStatistics('flops'). You are welcome to contribute.<\/p>\n\n<p>I'm not sure if NVIDA has tools for it.<\/p>\n"}]}
{"QuestionId":42309421,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-17T23:59:17.260","AcceptedAnswerId":42309513.0,"OwnerUserId":395857.0,"Title":"What's the point in constraining the shape of a TensorFlow's placeholder?","Body":"<p><a href=\"https:\/\/web.archive.org\/web\/20170218000039\/https:\/\/github.com\/petewarden\/tensorflow_ios\/blob\/master\/tensorflow\/g3doc\/resources\/faq.md\" rel=\"nofollow noreferrer\">https:\/\/web.archive.org\/web\/20170218000039\/https:\/\/github.com\/petewarden\/tensorflow_ios\/blob\/master\/tensorflow\/g3doc\/resources\/faq.md<\/a> states that:<\/p>\n\n<blockquote>\n  <p>The  op allows you to define tensors that must be fed, and optionally allows you to constrain their shape as well.<\/p>\n<\/blockquote>\n\n<p>What's the point in constraining the shape of a placeholder? Is it just for code clarity, or does it have other advantages such as reducing the running time?<\/p>\n","answers":[{"AnswerId":"42309513","CreationDate":"2017-02-18T00:11:50.760","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The main advantage of constraining the shape of a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a> is that the (possibly partial) shape can be used to infer the shapes of tensors that are derived from it. Shape inference can then be used to detect errors at the point where you create an op, rather than at runtime. The shape inference is analogous to type inference in a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Gradual_typing\" rel=\"nofollow noreferrer\">gradually typed<\/a> language, like TypeScript, where if you provide more information about the shapes (types) of placeholders (arguments), the system can use that information to improve the robustness of your program.<\/p>\n\n<p>For example, if you had the following program:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32)\ny = tf.placeholder(tf.float32)\nz = tf.matmul(x, y)\n\n# ...\n\nsess.run(z, feed_dict={x: [[1., 2., 3.]], y: [[2.], [3.]]})\n<\/code><\/pre>\n\n<p>...you would get a runtime error from the <code>MatMul<\/code> kernel.<\/p>\n\n<p>By contrast, if you constrained the shapes in the <code>tf.placeholder()<\/code> calls as follows:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, shape=[1, 3])\ny = tf.placeholder(tf.float32, shape=[2, 1])\nz = tf.matmul(x, y)\n<\/code><\/pre>\n\n<p>...you'd get an error when you call <code>tf.matmul()<\/code>.<\/p>\n\n<p>Shape inference can also improve the performance at runtime. There are various cases where the result of <code>tf.shape()<\/code> can be constant folded, and other optimizations will be made, which reduce the number of operations that must run. Backprop code in particular has to be very conservative about the shapes of its arguments, and additional static information there can result in much simpler graphs.<\/p>\n"}]}
{"QuestionId":42309978,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-02-18T01:13:38.690","AcceptedAnswerId":42560924.0,"OwnerUserId":5896419.0,"Title":"Keras\/Tensorflow: Training successful on MBP 13\" with Theano, but throws ResourceExhaustedError on a powerful computer with TensorFlow","Body":"<p>I have successfully trained a Keras model on a 13\" Macbook Pro with Theano, albeit at a slow speed, but when I train the exact same model with the same data on a more powerful computer (32 GB RAM, 8 GB Nvidia Quadro GPU, 8 CPU cores) with TensorFlow on Ubuntu, the following error occurs:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/9J005.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9J005.png\" alt=\"ResourceExhaustedError\"><\/a> <\/p>\n\n<p>Here is the script that I use:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42560924","CreationDate":"2017-03-02T16:35:16.543","ParentId":null,"OwnerUserId":"5896419","Title":null,"Body":"<p>As of this writing, the problem seems to be a TensorFlow issue. For me, the solution was to switch the backend to Theano. To switch the Keras backend, perform the following steps:<\/p>\n\n<ol>\n<li><p>Find the Keras configuration file<\/p>\n\n<p>~\/.keras\/keras.json<\/p><\/li>\n<li><p>Change the value of the backend field and, optionally, the ordering field<\/p>\n\n<p>{\n    \"image_dim_ordering\": \"th\",\n    \"epsilon\": 1e-07,\n    \"floatx\": \"float32\",\n    \"backend\": \"theano\"\n}<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42310081,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-02-18T01:32:01.570","AcceptedAnswerId":null,"OwnerUserId":3755376.0,"Title":"How to store volumetric patch into HDF5?","Body":"<p>I have a volumetric data with size . Due to limit memory, I cannot use the whole data to directly feed to CAFFE. Hence, I will randomly choose  patches  which extract from the volumetric data and store them into HDF5. I was successful to randomly extract the patches from raw data and its label by   function. <\/p>\n\n<p>I want to store these patches into the HDF5 data. The bellow code performs the task. Could you look at and verify help me my implementation? I am wondering about initial matrix  (lines 2,3), order matrix storing in lines 8,9,10,11). Thank you for reading<\/p>\n\n<pre><\/pre>\n\n<p>This is my prototxt<\/p>\n\n<pre><\/pre>\n\n<p>Update: Script to test convolution<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42310274,"AnswerCount":1,"Tags":"<python><graph><tensorflow><construction>","CreationDate":"2017-02-18T02:05:18.750","AcceptedAnswerId":null,"OwnerUserId":7583479.0,"Title":"Diffrent outputs caused by a \" irrelevant \" statement in graph construction (symbolic style programs ) in tensorflow","Body":"<p>I have a basic issue about the symbolic style programs in tensorflow, below are codes ( tensorflow 0.12 ):<\/p>\n\n<pre><\/pre>\n\n<p>if the 'x = z' is commented, this is a typical directed acyclic graph (DAG), the result is:<\/p>\n\n<pre><\/pre>\n\n<p>if the 'x = z' is uncommented, the result is:<\/p>\n\n<pre><\/pre>\n\n<p>I think the result should be irrelevant of the statement \"x = z\", and the result should be 3. What caused the diffrence? Thank you! <\/p>\n","answers":[{"AnswerId":"42499208","CreationDate":"2017-02-28T02:40:09.877","ParentId":null,"OwnerUserId":"7583479","Title":null,"Body":"<p>when x=z is added, the x acually indicates the same graph node with z, when if x is feeding [1], so z is also [1]. Thanks Yaroslav.<\/p>\n"}]}
{"QuestionId":42310706,"AnswerCount":1,"Tags":"<deep-learning><keras><evaluation><mnist>","CreationDate":"2017-02-18T03:22:28.583","AcceptedAnswerId":null,"OwnerUserId":6471247.0,"Title":"Keras custom accuracy metric for list output","Body":"<p>I have a model that predicts digits from the SVHN dataset.  I think I may need a custom accuracy metric because some of the digits are correct even if the whole sequence is not correct. Here is the code and a sample output.  Does anyone know how I could go about creating a custom metric?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42357391","CreationDate":"2017-02-21T02:01:16.390","ParentId":null,"OwnerUserId":"6471247","Title":null,"Body":"<p>The following will calculate the accuracy of individual digits:<\/p>\n\n<pre><code>def new_accuracy(predictions, labels):\n  return (100.0 * np.sum(np.argmax(predictions, 2).T == labels) \/ predictions.shape[1] \n          \/ predictions.shape[0])\n<\/code><\/pre>\n"}]}
{"QuestionId":42310835,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-02-18T03:45:40.467","AcceptedAnswerId":42346487.0,"OwnerUserId":3763831.0,"Title":"Input Queue not responding TensorFlow program hanging","Body":"<p>I am currently trying to train a neural network. I have an array of file names and their corresponding labels. However I am having issues when trying to train the network.<\/p>\n\n<pre><\/pre>\n\n<p>At the line in which I am trying to print the image data, the program hangs. Even when this line is removed, the program is hanging on the last sess.run call in which I am feeding this data. I have initialized queue runners, coordinators, etc. However, I have a feeling that the filename_queue is having an issue. Is there anything I am missing in the tf.train.slice_input_producer line? Also is the program hanging or is it just taking a while to load. How much time would it usually take to load an epoch with a batch size of 100 and images of 80 by 70?<\/p>\n","answers":[{"AnswerId":"42346487","CreationDate":"2017-02-20T13:56:20.810","ParentId":null,"OwnerUserId":"5903959","Title":null,"Body":"<p>It looks like an issue I opened. While feeding data the input queue runners was hanging. This is because you have to start it. <\/p>\n\n<p>From the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5396\" rel=\"nofollow noreferrer\">issue<\/a>, we have:<\/p>\n\n<blockquote>\n  <p>Quoting: <em>RudrakshTuwani<\/em><br>\n  For anyone else struggling with this, please read the documentation as\n  mentioned by girving. For the lazy ones:<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\nsess.run(init)\nthreads = tf.train.start_queue_runners()\nprint(sess.run(name_of_output_tensor))\n<\/code><\/pre>\n<\/blockquote>\n\n<p>As well as:<\/p>\n\n<blockquote>\n  <p>From: <em>girving<\/em><br>\n  You probably need to start queue runners. Please see the documentation\n  at <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/how_tos\/threading_and_queues\/index.html\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.11\/how_tos\/threading_and_queues\/index.html<\/a><\/p>\n<\/blockquote>\n\n<p>Hope it helps!<br>\n<code>pltrdy<\/code><\/p>\n\n<hr>\n\n<p>Note that in my case I got confused because the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/ptb_word_lm.py#L352\" rel=\"nofollow noreferrer\">original code<\/a> was using:<\/p>\n\n<pre><code>sv = tf.train.Supervisor(logdir=FLAGS.save_path)\n    with sv.managed_session() as session:\n<\/code><\/pre>\n\n<p>instead of my (and your):<\/p>\n\n<pre><code>with tf.Session() as session:\n<\/code><\/pre>\n\n<p>The first one actually implicitely starts queue runners.<\/p>\n"}]}
{"QuestionId":42311007,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-18T04:20:31.073","AcceptedAnswerId":42311633.0,"OwnerUserId":5512020.0,"Title":"AttributeError: 'tensorflow.python.ops.rnn' has no attribute 'rnn'","Body":"<p>I am following <a href=\"https:\/\/pythonprogramming.net\/rnn-tensorflow-python-machine-learning-tutorial\/\" rel=\"noreferrer\">this tutorial<\/a> on Recurrent Neural Networks.<\/p>\n\n<p>This is the imports:<\/p>\n\n<pre><\/pre>\n\n<p>This is code for input processing:<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the following error for the :<\/p>\n\n<pre><\/pre>\n\n<p>TensorFlow was updated recently, so what should be the new code for the offending line<\/p>\n","answers":[{"AnswerId":"57677900","CreationDate":"2019-08-27T15:22:27.143","ParentId":null,"OwnerUserId":"1024979","Title":null,"Body":"<p>Thanks @suku<\/p>\n\n<p>I get the following error: <code>ImportError: No module named 'tensorflow.contrib.rnn.python.ops.core_rnn'<\/code><\/p>\n\n<p>To solve:<\/p>\n\n<pre><code>from tensorflow.contrib.rnn.python.ops import core_rnn\n<\/code><\/pre>\n\n<p>replaced by:<\/p>\n\n<pre><code>from tensorflow.python.ops import rnn, rnn_cell\n<\/code><\/pre>\n\n<hr>\n\n<p>and in my code I had used <code>core_rnn.static_rnn<\/code>:<\/p>\n\n<pre><code> outputs,_ = core_rnn.static_rnn(cell, input_list, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>I got the this error:<\/p>\n\n<p><code>NameError: name 'core_rnn' is not defined<\/code><\/p>\n\n<p>This is solved by replacing the line by:<\/p>\n\n<pre><code>outputs,_ = rnn.static_rnn(cell, input_list, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>python: 3.6 64bit\nrensorflow:1.10.0<\/p>\n"},{"AnswerId":"42311633","CreationDate":"2017-02-18T05:55:41.353","ParentId":null,"OwnerUserId":"5512020","Title":null,"Body":"<p>For people using the newer version of tensorflow, add this to the code:<\/p>\n\n<pre><code>from tensorflow.contrib import rnn \n\n\nlstm_cell = rnn.BasicLSTMCell(rnn_size) \noutputs, states = rnn.static_rnn(lstm_cell, x, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>instead of <\/p>\n\n<pre><code>from tensorflow.python.ops import rnn, rnn_cell \nlstm_cell = rnn_cell.BasicLSTMCell(rnn_size,state_is_tuple=True) \noutputs, states = rnn.rnn(lstm_cell, x, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>PS: @BrendanA suggested that <code>tf.nn.rnn_cell.LSTMCell<\/code> be used instead of <code>rnn_cell.BasicLSTMCell<\/code><\/p>\n"}]}
{"QuestionId":42311242,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-02-18T04:58:12.793","AcceptedAnswerId":null,"OwnerUserId":7583897.0,"Title":"Tensorflow: Unexpected EOF while parsing using CPU tf.device","Body":"<p>I'm kinda new in Python and using Tensorflow, but now I'm just trying to implement basic Word2Vec tutorial from <a href=\"https:\/\/www.tensorflow.org\/code\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>It has code like this:<\/p>\n\n<pre><\/pre>\n\n<p>When I run the  part it outputs this error:<\/p>\n\n<pre><\/pre>\n\n<p>And when I run the whole block, it outputs <\/p>\n\n<pre><\/pre>\n\n<p>Thank you before!<\/p>\n","answers":[]}
{"QuestionId":42311601,"AnswerCount":2,"Tags":"<python><tensorflow><ipython><anaconda>","CreationDate":"2017-02-18T05:51:54.410","AcceptedAnswerId":null,"OwnerUserId":7584020.0,"Title":"How to change Python version within IPython in anaconda","Body":"<p>since tensorflow only works in python 3.5.2, my default IPython env in Anaconda invoke Python 3.6 which can't import tensorflow. Does anyone know how to change that? Much thanks.<\/p>\n","answers":[{"AnswerId":"42312456","CreationDate":"2017-02-18T07:31:26.700","ParentId":null,"OwnerUserId":"6837980","Title":null,"Body":"<p>You can use Anaconda 4.2.0, which installs python3.5.\n<a href=\"https:\/\/repo.continuum.io\/archive\/index.html\" rel=\"nofollow noreferrer\">https:\/\/repo.continuum.io\/archive\/index.html<\/a><\/p>\n"},{"AnswerId":"42314730","CreationDate":"2017-02-18T11:47:44.773","ParentId":null,"OwnerUserId":"2449192","Title":null,"Body":"<p>You can create an environment that installs Python 3.5. See here: <a href=\"https:\/\/conda.io\/docs\/using\/envs.html\" rel=\"nofollow noreferrer\">https:\/\/conda.io\/docs\/using\/envs.html<\/a><\/p>\n\n<p>To summarize<\/p>\n\n<pre><code>conda create --name tenflow python=3.5 tensorflow ipython\n<\/code><\/pre>\n\n<p>Then, on Mac or Linux<\/p>\n\n<pre><code>source activate tenflow\n<\/code><\/pre>\n\n<p>or on Windows<\/p>\n\n<pre><code>activate tenflow\n<\/code><\/pre>\n"}]}
{"QuestionId":42311607,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-18T05:52:32.110","AcceptedAnswerId":42311787.0,"OwnerUserId":6815505.0,"Title":"how to normalize prediction values in tensorflow","Body":"<p>A simple feed forward DNN with relevant .csv files can be found here <a href=\"https:\/\/github.com\/jhsmith12345\/tensorflow\/blob\/normalize_prediction\/tf_from_csv.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/jhsmith12345\/tensorflow\/blob\/normalize_prediction\/tf_from_csv.py<\/a><\/p>\n\n<p>This piece of code<\/p>\n\n<pre><\/pre>\n\n<p>is outputting <\/p>\n\n<pre><\/pre>\n\n<p>I am expecting a prediction that conforms to the labels, which are 1 or 0. Something like<\/p>\n\n<pre><\/pre>\n\n<p>I believe that my predictive values are not getting normalized by a softmax, but have no idea how to proceed. Any help is appreciated! Also, I'm more than happy to post the full code here but didn't want to clutter the post. Thanks!<\/p>\n","answers":[{"AnswerId":"42311787","CreationDate":"2017-02-18T06:13:22.483","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>Let me clear, in your code<\/p>\n\n<pre><code>prediction = neural_network_model(x)\nprediction.eval(feed_dict={x: [[9,3]]})\n# output is [[ -12.2412138  -17.24327469 ]]\n<\/code><\/pre>\n\n<p>and you confuse why the range is not 0 ~ 1, right ?<br>\nbecause softmax doesn't apply on prediction<br>\nyou use <code>tf.nn.softmax_cross_entropy_with_logits<\/code><br>\nFor I know, this function apply softmax to <code>prediction<\/code> before compute cross entropy<br>\nBut it doesn't change the value of <code>prediction<\/code>   <\/p>\n\n<p>I think you can\ndo softmax then compute cross entropy, finally print <code>prediction<\/code> directly (it means can't use <code>tf.nn.softmax_cross_entropy_with_logits<\/code>)    <\/p>\n\n<p>or, change nothing, but do softmax before print <code>prediction<\/code><\/p>\n"}]}
{"QuestionId":42311733,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><keras>","CreationDate":"2017-02-18T06:07:44.880","AcceptedAnswerId":null,"OwnerUserId":7584059.0,"Title":"ValueError: expected convolution2d_input_8 to have shape (None, 3, 150, 150) but got array with shape (32, 150, 150, 3)","Body":"<p>I got the following error<\/p>\n\n<blockquote>\n  <p>ValueError: Error when checking model input: expected convolution2d_input_8 to have shape (None, 3, 150, 150) but got array with shape (32, 150, 150, 3)<\/p>\n<\/blockquote>\n\n<p>I think its is because of the dim_ordering. I am using tensorflow backend.\nBut if I remove it I get another error:<\/p>\n\n<blockquote>\n  <p>ValueError: Negative dimension size caused by subtracting 3 from 1 for 'Conv2D_17' (op: 'Conv2D') with input shapes: [?,1,74,16], [3,3,16,32].<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42311997","CreationDate":"2017-02-18T06:39:53.947","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<blockquote>\n  <p>Input shape<\/p>\n  \n  <p>4D tensor with shape: (samples, channels, rows, cols) if\n  dim_ordering='th' or 4D tensor with shape: (samples, rows, cols,\n  channels) if dim_ordering='tf'.<\/p>\n  \n  <p>Output shape<\/p>\n  \n  <p>4D tensor with shape: (samples, nb_filter, new_rows, new_cols) if\n  dim_ordering='th' or 4D tensor with shape: (samples, new_rows,\n  new_cols, nb_filter) if dim_ordering='tf'. rows and cols values might\n  have changed due to padding.\n  <a href=\"https:\/\/keras.io\/layers\/convolutional\/#convolution2d\" rel=\"nofollow noreferrer\">source<\/a><\/p>\n<\/blockquote>\n\n<p>Indeed it comes from dimension ordering. \nNot only the input must be different when using Theano or Tensorflow, the output will be too. <\/p>\n\n<p>If I understand correctly, you should use tf backend as usual, but then change your input shape to become (150,150,3). \nIf you really want to change to Theano, consider setting dim_ordering to theano for <strong>all<\/strong> of the convolutionnal layers, so that they can \"speak\" to each other regarding output and input shapes. <\/p>\n\n<p>Does this help?<\/p>\n"}]}
{"QuestionId":42311773,"AnswerCount":1,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-02-18T06:11:29.720","AcceptedAnswerId":42312602.0,"OwnerUserId":4809029.0,"Title":"Select subset of tensor using boolean tensor in tensorflow","Body":"<p>I have two rank-2 tensors  and  of shape  by . The tensor  is boolean; precisely one entry in each of its rows is . I want to extract a new rank-1 tensor  of length , where the th entry of  is equal to the entry from the th row of  corresponding to where the th row of  is equal to .<\/p>\n\n<p>In , I can do this as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Can I do something similar in ? I know that I could  my tensors and then use  functions, but that seems inefficient.<\/p>\n\n<p>This <a href=\"https:\/\/stackoverflow.com\/questions\/33769041\/tensorflow-indexing-with-boolean-tensor\">question<\/a> suggests the use of  and , but it does not deal with collapsing the dimensionality of the output like my question does.<\/p>\n","answers":[{"AnswerId":"42312602","CreationDate":"2017-02-18T07:48:20.630","ParentId":null,"OwnerUserId":"681870","Title":null,"Body":"<p>You could maybe use <code>tf.boolean_mask<\/code>?<\/p>\n\n<pre><code>from __future__ import print_function\n\nimport tensorflow as tf\n\nwith tf.Session() as sess:\n    arr1 = tf.constant([[1,2],\n                        [3,4]])\n    arr2 = tf.constant([[False, True],\n                        [True, False]])\n    print(sess.run(tf.boolean_mask(arr1, arr2)))\n<\/code><\/pre>\n\n<p>Should give: [2, 3]<\/p>\n"}]}
{"QuestionId":42312367,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-18T07:21:51.323","AcceptedAnswerId":42816011.0,"OwnerUserId":6837980.0,"Title":"tensorflow naming: capitalized or not?","Body":"<p>Why tensorflow mixes capitalized and non-capitalized naming? I don't think it make sense. Maybe it's due to some legacy code?<\/p>\n\n<p>Below are some examples\n- tf.constant(), tf.Variable()\n- tf.Session().run()<\/p>\n","answers":[{"AnswerId":"42816011","CreationDate":"2017-03-15T16:51:23.990","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>It uses pep8, functions are snake case, classes are camel case.<\/p>\n"}]}
{"QuestionId":42312938,"AnswerCount":1,"Tags":"<android><python><tensorflow>","CreationDate":"2017-02-18T08:32:06.813","AcceptedAnswerId":42314524.0,"OwnerUserId":4086817.0,"Title":"Tensorflow Android demo : No output","Body":"<p>I am trying to use retrained inception-v3 model in Tensroflow android demo app, no output shown.<\/p>\n\n<p><strong>What I have done<\/strong><\/p>\n\n<p>Trainded the model per description <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">retrain inception<\/a>. After training (only five classes) I have tested the graph using <\/p>\n\n<pre><\/pre>\n\n<p>and following are the outputs<\/p>\n\n<pre><\/pre>\n\n<p>Done the optimization for inference using <\/p>\n\n<pre><\/pre>\n\n<p>and tested the output graph again works fine.<\/p>\n\n<p>finaly ran the following strip_unsued.py<\/p>\n\n<pre><\/pre>\n\n<p>tested the graph again works fine.<\/p>\n\n<p><strong>Android app chnages in Classifier Activity<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>Build &amp; Ran the project.<\/p>\n\n<p><strong>Traceback<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>When I use the app to identify an object no output is shown. <a href=\"https:\/\/scontent-lhr3-1.xx.fbcdn.net\/v\/t34.0-12\/16830128_1858197347729123_830189794_n.png?oh=2faf1f9fd39a8d49c286b33b1b4ea298&amp;oe=58A9B455\" rel=\"nofollow noreferrer\">Output<\/a><\/p>\n\n<p>this one is also show in logs <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42314524","CreationDate":"2017-02-18T11:26:30.560","ParentId":null,"OwnerUserId":"4086817","Title":null,"Body":"<p>I have figured it out.There is typo mistake in ClassifierActivity.<\/p>\n\n<pre><code>private static final int INPUT_SIZE = 229;\n<\/code><\/pre>\n\n<p>should be <\/p>\n\n<pre><code>private static final int INPUT_SIZE = 299;\n<\/code><\/pre>\n"}]}
{"QuestionId":42312956,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-18T08:34:56.243","AcceptedAnswerId":null,"OwnerUserId":4993842.0,"Title":"Training tensorflow network to check if two numbers are same doesn't work?","Body":"<p>I'm new to python and tensorflow (and machine learning) and this is my first program. Once trained, it should accept two numbers and return [1, 0] if they're unequal and [0, 1] if they're equal. <\/p>\n\n<p>I've modified the beginner tensorflow tutorial at <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners<\/a> (which looks for numbers in an image array) to this purpose. I train it on 1000000 pairs with both equal and unequal pairs with the numbers in [0, 100] range.<\/p>\n\n<p><strong>Problems<\/strong> - Training brings up the accuracy to 0.5 but no higher. Instead of returning [0, 1] or [1, 0], it returns [x, -x] (where x is a floating point number). I'm not sure if the problem lies in the python code or tensorflow usage but I suspect it is a silly beginner's mistake. What am I doing wrong?<\/p>\n\n<p>Below is my code -<\/p>\n\n<pre><\/pre>\n\n<p>Then I test it with some test data -<\/p>\n\n<pre><\/pre>\n\n<p>Which always prints 0.5. <\/p>\n\n<p>Here are the outputs of more variables -<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42313046,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><keras>","CreationDate":"2017-02-18T08:45:35.143","AcceptedAnswerId":42410436.0,"OwnerUserId":4359305.0,"Title":"Keras - Predict on model after learning sentiment data throws errors","Body":"<p>My problem is in getting results through predict method in keras using tensorflow backend. But first a small introduction.<\/p>\n\n<p>I am using <\/p>\n\n<ul>\n<li>Python 2.7.12<\/li>\n<li>Keras==1.2.1<\/li>\n<li>numpy==1.12.0<\/li>\n<li>tensorflow==0.12.1<\/li>\n<\/ul>\n\n<p>I created a convolutional neural network like in these documentation: <a href=\"https:\/\/blog.keras.io\/using-pre-trained-word-embeddings-in-a-keras-model.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/using-pre-trained-word-embeddings-in-a-keras-model.html<\/a><\/p>\n\n<p>I trained the network with 11842 prepared twitter texts. The only individual change is that I have 3 possibilities for a result (0,1,2). I defined that in following code line:<\/p>\n\n<pre><\/pre>\n\n<p>So the method fit works without a problem and I am achieving between 88-92% acc.<\/p>\n\n<pre><\/pre>\n\n<p>After the learning process I saved the model in .h5 format (also works fine).<\/p>\n\n<p>Now I try to load the models and predict with them. First example (trained_model) is done via the same data I used to train ... because I wanted to compare them. The second example (trained_model_2) is done via new twitter texts (I collected earlier).<\/p>\n\n<pre><\/pre>\n\n<p>For comparing the training dataset with the \"live \/ new\" data set: <\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>Also both are from type <\/p>\n\n<p>Following code line raises the first error:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: indices[0,999] = 13608 is not in [0, 13480)\n       [[Node: Gather_1 = Gather[Tindices=DT_INT32, Tparams=DT_FLOAT, validate_indices=true, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](embedding_1_W_1\/read, _recv_input_1_1_0)]]<\/p>\n<\/blockquote>\n\n<p>Following code line raises the second error:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): indices[0,999] = 13608 is not in [0, 13480)\n       [[Node: Gather_1 = Gather[Tindices=DT_INT32, Tparams=DT_FLOAT, validate_indices=true, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](embedding_1_W_1\/read, _recv_input_1_1_0)]]<\/p>\n<\/blockquote>\n\n<p><strong>EDIT<\/strong>\nSource Code of the methods I created. Method \"trainModule\" is only used for training the network \/ save it. The method \"predict_sentiment\" is used for my predict tests. The first prediction_results works and returns a numpy array with following shape <strong>(11842, 3)<\/strong>\n<a href=\"http:\/\/pastebin.com\/vZ8bVpfW\" rel=\"nofollow noreferrer\">Code - pastbin<\/a><\/p>\n\n<p>The whole error output: <a href=\"http:\/\/pastebin.com\/uxEweWUn\" rel=\"nofollow noreferrer\">Error output - pastbin<\/a><\/p>\n\n<p>If some additional information is needed, I will update the question...<\/p>\n","answers":[{"AnswerId":"42410436","CreationDate":"2017-02-23T08:23:31.540","ParentId":null,"OwnerUserId":"4359305","Title":null,"Body":"<p>The problem was the trained model can't find the word in the embedding matrix. That means I used a different vocabulary for training and prediction. Because the of the fixed vocabulary I need the same vocabulary for train and new data.<\/p>\n\n<p>In general I only had to fix the tokenizer from:<\/p>\n\n<pre><code>tokenizer_predict = Tokenizer(nb_words=MAX_NB_WORDS)\ntokenizer_predict.fit_on_texts(texts_predict)\nsequence_predict = tokenizer_predict.texts_to_sequences(predict_data)\n<\/code><\/pre>\n\n<p>To:<\/p>\n\n<pre><code>tokenizer_predict = Tokenizer(nb_words=MAX_NB_WORDS)\ntokenizer_predict.fit_on_texts(texts_train)\nsequence_predict = tokenizer_predict.texts_to_sequences(predict_data)\n<\/code><\/pre>\n"},{"AnswerId":"42314890","CreationDate":"2017-02-18T12:03:31.510","ParentId":null,"OwnerUserId":"7584913","Title":null,"Body":"<p>Maybe it tried to access a list of [0, 13480) with the index 13608 (which obviously will not work). Someone else had also a similar issue: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2734\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2734<\/a>\nIt seems he tried to access a vocabulary [0, 10000) with the index of 10535.<\/p>\n"}]}
{"QuestionId":42313412,"AnswerCount":2,"Tags":"<python><tensorflow><keras><recurrent-neural-network><resnet>","CreationDate":"2017-02-18T09:28:56.903","AcceptedAnswerId":42358696.0,"OwnerUserId":1844390.0,"Title":"Keras pretrain CNN with TimeDistributed","Body":"<p>Here is my problem, I want to use one of the pretrain CNN network in a TimeDistributed layer. But I have some problem to implement it.<\/p>\n\n<p>Here is my model:<\/p>\n\n<pre><\/pre>\n\n<p>When compiling the model I have no error. But when I start training I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I checked and I do send float32 but for input1, input2 is the input present in the pretrain Resnet.<\/p>\n\n<p>Just to have an overview here is the model summary. (Note: it's strange that it doesn't show what happen inside Resnet but never mind)<\/p>\n\n<pre><\/pre>\n\n<p>I am guessing that I do not use the TimeDistributed correctly and I saw nobody trying to do this. I hope someone can guide me on this.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>The problem comes from the fact that  create its own input in the graph.<\/p>\n\n<p>So I guess I need to do something like  but I do not see how to couple it with .<\/p>\n\n<p>I tried<\/p>\n\n<pre><\/pre>\n\n<p>But it does not work.<\/p>\n","answers":[{"AnswerId":"57535293","CreationDate":"2019-08-17T10:08:03.253","ParentId":null,"OwnerUserId":"8523421","Title":null,"Body":"<p>Considering you are using a pre-trained network from keras, you can replace it with your own pre-trained network too.<\/p>\n\n<p>Here's a simple solution::<\/p>\n\n<pre><code>model_vgg=keras.applications.VGG16(input_shape=(256, 256, 3),\n                                           include_top=False,\n                                           weights='imagenet')\nmodel_vgg.trainable = False\nmodel_vgg.summary()\n<\/code><\/pre>\n\n<p>If you want to use any intermediate layers then, otherwise replace 'block2_pool' with last layer's name::<\/p>\n\n<pre><code>intermediate_model= Model(inputs=model_vgg.input, outputs=model_vgg.get_layer('block2_pool').output)\nintermediate_model.summary()\n<\/code><\/pre>\n\n<p>Finally wrap it in a TimeDistributed Layer<\/p>\n\n<pre><code>input_tensor = Input(shape=(time_steps,height, width, channels))\ntimeDistributed_layer = TimeDistributed( intermediate_model )(input_tensor)\n<\/code><\/pre>\n\n<p>Now you can simply do::<\/p>\n\n<pre><code>my_time_model = Model( inputs = input_tensor, outputs = timeDistributed_layer )\n<\/code><\/pre>\n"},{"AnswerId":"42358696","CreationDate":"2017-02-21T04:26:45.897","ParentId":null,"OwnerUserId":"1844390","Title":null,"Body":"<p>My quick solution is a little bit ugly.<\/p>\n\n<p>I just copied the code of ResNet and added TimeDistributed to all layers and then loaded the weights from a \"basic\" ResNet on my customized ResNet.<\/p>\n\n<p>Note:<\/p>\n\n<p>To be able to analyze sequence of images like this does take a huge amount of memory on the gpu.<\/p>\n"}]}
{"QuestionId":42313781,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-18T10:09:43.867","AcceptedAnswerId":null,"OwnerUserId":2039825.0,"Title":"repmat with interlace or Kronecker product in Tensorflow","Body":"<p>Suppose I have a tensor:    <\/p>\n\n<pre><\/pre>\n\n<p>Which is a matrix with 2 rows and 3 columns.<\/p>\n\n<p>I would like to replicate it, suppose twice, to get the following tensor:<\/p>\n\n<pre><\/pre>\n\n<p>Using  will clearly replicate it differently, so I tried the following code (which works):<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, it seems to be working very slow when the number of columns become large. Executing it in Matlab using Kronecker product with a vector of ones (Matlab's \"kron\") seems to be much faster.<\/p>\n\n<p>Can anyone help?<\/p>\n","answers":[]}
{"QuestionId":42313828,"AnswerCount":0,"Tags":"<image-processing><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2017-02-18T10:15:36.503","AcceptedAnswerId":null,"OwnerUserId":1803124.0,"Title":"Dynamic graphs in tensorflow","Body":"<p>I would like to implement a 2D LSTM as in <a href=\"https:\/\/arxiv.org\/pdf\/0705.2011.pdf\" rel=\"nofollow noreferrer\">this paper<\/a>, specifically I would like to do so dynamically, so using tf.while. In brief this network works as follows.<\/p>\n\n<ul>\n<li>order the pixels in an image so that pixel i, j -> i * width + j<\/li>\n<li>run a 2D-LSTM over this sequence\n\n<blockquote>\n  <p>The difference between a 2D and regular LSTM is we have a recurrent connection between the previous element in the sequence and the pixel directly above the current pixel, so at pixel i,j are connections to i - 1, j and i, j - 1.<\/p>\n<\/blockquote><\/li>\n<\/ul>\n\n<h1>What I have done<\/h1>\n\n<p>I have tried to do this using tf.while where in each iteration of the loop I accumulate the activations and cell states into a tensor whose shape I allow to vary. This is what the following block of code tries to do.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>This works, at least in the forward direction. That is to say if I look at  what is returned in a session then I get what I want which is a 3D tensor, call it T, of shape [height, width, units] where T[i,j,:] contains the activation of the LSTM cell at input i, j.<\/p>\n\n<p>I then would like to classify each pixel and for this purpose I conv2D across T then reshape the result into [height * width, num_labels] and construct the cross entropy loss.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<h1>The problem<\/h1>\n\n<p>However now when I try with an image which 28 x 28 and 32 units<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>I get the following error<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>I think this is a problem with calculating the gradients resulting from the tf.while loop but I am pretty lost at this point.<\/p>\n","answers":[]}
{"QuestionId":42313902,"AnswerCount":1,"Tags":"<machine-learning><keras>","CreationDate":"2017-02-18T10:25:24.260","AcceptedAnswerId":null,"OwnerUserId":1762870.0,"Title":"Can I remove layers in a pre-trained Keras model?","Body":"<p>I'm importing a pre-trained VGG model in Keras, with <\/p>\n\n<pre><\/pre>\n\n<p>I've noticed that the type of a standard model is , while a pre-trained model is . I usually add and remove layers with  and  for sequential models respectively, however, I cannot seem to use  with pre-trained models.<\/p>\n\n<p>Is there an alternative to  for these type of models?<\/p>\n","answers":[{"AnswerId":"48796569","CreationDate":"2018-02-14T21:30:37.447","ParentId":null,"OwnerUserId":"9357409","Title":null,"Body":"<p>Depends on what you're wanting to remove. If you want to remove the last softmax layer and use the model for transfer learning, you can pass the <code>include_top=False<\/code> kwarg into the model like so:<\/p>\n\n<pre><code>from keras.applications.vgg16 import VGG16\nIN_SHAPE = (256, 256, 3) # image dimensions and RGB channels\n\npretrained_model = VGG16(\n  include_top=False,\n  input_shape=IN_SHAPE,\n  weights='imagenet'\n)\n<\/code><\/pre>\n\n<p>I wrote a blog post on this use case recently that has some code examples and goes into a bit more detail: <a href=\"http:\/\/innolitics.com\/10x\/pretrained-models-with-keras\/\" rel=\"nofollow noreferrer\">http:\/\/innolitics.com\/10x\/pretrained-models-with-keras\/<\/a><\/p>\n\n<p>If you're wanting to modify the model architecture more than that, you can access the <code>pop()<\/code> method via <code>pretrained_model.layers.pop()<\/code>, as is explained in the link @indraforyou posted.<\/p>\n\n<p>Side note: When you're modifying layers in a pretrained model, it can be especially helpful to have a visualization of the structure and input\/output shapes. <code>pydot<\/code> and <code>graphviz<\/code> is particularly useful for this:<\/p>\n\n<pre><code>import pydot\npydot.find_graphviz = lambda: True\nfrom keras.utils import plot_model\nplot_model(model, show_shapes=True, to_file='..\/model_pdf\/{}.pdf'.format(model_name))\n<\/code><\/pre>\n"}]}
{"QuestionId":42314204,"AnswerCount":1,"Tags":"<tensorflow><tflearn>","CreationDate":"2017-02-18T10:54:37.827","AcceptedAnswerId":null,"OwnerUserId":5512020.0,"Title":"tflearn cannot read mnist data","Body":"<p>I am trying to do Convolutional Neural Network using tensorflow. I am using tflearn for this. I have downloaded the data from <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">this site<\/a> <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42329291","CreationDate":"2017-02-19T15:49:44.777","ParentId":null,"OwnerUserId":"6552952","Title":null,"Body":"<p>The MNIST files will work with the GPU-supported TensorFlow version that uses NVIDIA cuDNN, and here is a post about installing them <a href=\"https:\/\/stackoverflow.com\/questions\/31326015\/how-to-verify-cudnn-installation\">How to verify CuDNN installation?<\/a>.<\/p>\n"}]}
{"QuestionId":42314206,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-18T10:54:57.860","AcceptedAnswerId":42314263.0,"OwnerUserId":3102085.0,"Title":"How to understand shapes of a tensor","Body":"<p>For example here is the code snippet<\/p>\n\n<pre><\/pre>\n\n<p>I do not understand what does the shape of a tensor mean? How do I understand what values kernel has etc. <\/p>\n","answers":[{"AnswerId":"42314263","CreationDate":"2017-02-18T10:59:51.263","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p><code>kernel<\/code> is a tensor with <code>shape=[3,3,3,64]<\/code> that means:<\/p>\n\n<ul>\n<li><code>3, 3<\/code> = kernel width x kernel height<\/li>\n<li><code>3<\/code> = kernel depth<\/li>\n<li><code>64<\/code> = number of kernel <code>3 x 3 x 64<\/code><\/li>\n<\/ul>\n\n<p>This means that every single kernel is a volume <code>3x3x3<\/code> and you got 64 of this.<\/p>\n\n<p>The kernel depth is equal to the depth of the input the kernel is convolved with: so you're convolving it with a RGB image, probably.<\/p>\n\n<p>The output of the convolution of this 3 channel image with 64 kernels <code>3x3x3<\/code> is a volume with depth <code>64<\/code>: one feature map for every convolutional kernel.<\/p>\n"}]}
{"QuestionId":42314317,"AnswerCount":2,"Tags":"<c++><tensorflow><bazel>","CreationDate":"2017-02-18T11:04:53.560","AcceptedAnswerId":null,"OwnerUserId":7584738.0,"Title":"Eigen functor undeclared when building Tensorflow","Body":"<p>I'm trying to contribute to the Tensorflow code by adding non-inverse hyperbolic functions (sinh and cosh), as requested in [issue #7531]\n(<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7531\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7531<\/a>).<\/p>\n\n<p>When I test the build with the command<\/p>\n\n<pre><\/pre>\n\n<p>But I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>The full output can be seen <a href=\"http:\/\/pastebin.com\/uDzX3CH5\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>What I did was add these two templates in :<\/p>\n\n<pre><\/pre>\n\n<p>and made two new files  and  , which are just copies of the non-hyperbolic versions with references to sin or cos replaced with sinh and cosh, respectively. As far as I can see, the hyperbolic functions are now implemented exactly like the other math functions from the Eigen library. But as far as I can see the error it gives references to missing declarations in the Eigen source.<\/p>\n\n<p>This is my very first open source contribution and on top of that I'm fairly new to C++. What I did wrong might be obvious with more C++ experience. <\/p>\n","answers":[{"AnswerId":"43070955","CreationDate":"2017-03-28T13:28:48.350","ParentId":null,"OwnerUserId":"7778502","Title":null,"Body":"<p>(<em>This should be a comment on OP's question but I lack the required karma.<\/em>)<\/p>\n\n<p>The <code>--config opt<\/code> flag seems wrong, did you mean <code>-c opt<\/code> for optimized build? <code>--config opt<\/code> means use the \"opt\" configuration from the bazelrc file, which supposes you have lines in the bazelrc file like \"build:opt --flag1=value1\" and similar. If you don't, then Bazel should display a warning saying:<\/p>\n\n<pre><code>WARNING: Config values are not defined in any .rc file: \"opt\"\n<\/code><\/pre>\n\n<p>or something to this effect.<\/p>\n"},{"AnswerId":"42829498","CreationDate":"2017-03-16T09:05:47.057","ParentId":null,"OwnerUserId":"1009896","Title":null,"Body":"<p>Did you add your `cwise_op_sinh.cc' and 'cwise_op_cosh.cc' into the corresponding BUILD file? Quick search tells me it should be <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2d23960e29c3efb1b30c5cb6f8f3368635b97f29\/tensorflow\/compiler\/tf2xla\/kernels\/BUILD\" rel=\"nofollow noreferrer\">here<\/a> but I can be wrong. You definitely have to tell Bazel about your new files though. <\/p>\n"}]}
{"QuestionId":42315202,"AnswerCount":1,"Tags":"<tensorflow><histogram><tensorboard>","CreationDate":"2017-02-18T12:35:35.013","AcceptedAnswerId":42318280.0,"OwnerUserId":364772.0,"Title":"Understanding TensorBoard (weight) histograms","Body":"<p>It is really straightforward to see and understand the scalar values in TensorBoard. However, it's not clear how to understand histogram graphs. <\/p>\n\n<p>For example, they are the histograms of my network weights.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/IttNH.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RVOA5.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>(After fixing a bug thanks to sunside)\n<a href=\"https:\/\/i.stack.imgur.com\/IttNH.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IttNH.jpg\" alt=\"enter image description here\"><\/a>\nWhat is the best way to interpret these? Layer 1 weights look mostly flat, what does this mean?<\/p>\n\n<p>I added the network construction code here.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42318280","CreationDate":"2017-02-18T17:21:19.067","ParentId":null,"OwnerUserId":"195651","Title":null,"Body":"<p>It appears that the network hasn't learned anything in the layers one to three. The last layer does change, so that means that there either may be something wrong with the gradients (if you're tampering with them manually), you're constraining learning to the last layer by optimizing only its weights or the last layer really 'eats up' all error. It could also be that only biases are learned. The network appears to learn something though, but it might not be using its full potential. More context would be needed here, but playing around with the learning rate (e.g. using a smaller one) might be worth a shot.<\/p>\n\n<p>In general, histograms display the number of occurrences of a value relative to each other values. Simply speaking, if the possible values are in a range of <code>0..9<\/code> and you see a spike of amount <code>10<\/code> on the value <code>0<\/code>, this means that 10 inputs assume the value <code>0<\/code>; in contrast, if the histogram shows a plateau of <code>1<\/code> for all values of <code>0..9<\/code>, it means that for 10 inputs, each possible value <code>0..9<\/code> occurs <em>exactly<\/em> once.\nYou can also use histograms to visualize probability distributions when you normalize all histogram values by their total sum; if you do that, you'll intuitively obtain the likelihood with which a certain value (on the x axis) will appear (compared to other inputs).<\/p>\n\n<p>Now for <code>layer1\/weights<\/code>, the plateau means that:<\/p>\n\n<ul>\n<li>most of the weights are in the range of -0.15 to 0.15<\/li>\n<li>it is (mostly) equally likely for a weight to have any of these values, i.e. they are (almost) uniformly distributed<\/li>\n<\/ul>\n\n<p>Said differently, almost the same number of weights have the values <code>-0.15<\/code>, <code>0.0<\/code>, <code>0.15<\/code> and everything in between. There are some weights having slightly smaller or higher values.\nSo in short, this simply looks like the weights have been initialized using a uniform distribution with zero mean and value range <code>-0.15..0.15<\/code> ... give or take. If you do indeed use uniform initialization, then this is typical when the network has not been trained yet.<\/p>\n\n<p>In comparison, <code>layer1\/activations<\/code> forms a bell curve (gaussian)-like shape: The values are centered around a specific value, in this case <code>0<\/code>, but they may also be greater or smaller than that (equally likely so, since it's symmetric). Most values appear close around the mean of <code>0<\/code>, but values do range from <code>-0.8<\/code> to <code>0.8<\/code>.\nI assume that the <code>layer1\/activations<\/code> is taken as the distribution over all layer outputs in a batch. You can see that the values do change over time.<\/p>\n\n<p>The layer 4 histogram doesn't tell me anything specific. From the shape, it's just showing that some weight values around <code>-0.1<\/code>, <code>0.05<\/code> and <code>0.25<\/code> tend to be occur with a higher probability; a reason <em>could<\/em> be, that different parts of each neuron there actually pick up the same information and are basically redundant. This can mean that you could actually use a smaller network or that your network has the potential to learn more distinguishing features in order to prevent overfitting. These are just assumptions though.<\/p>\n\n<p>Also, as already stated in the comments below, do add bias units. By leaving them out, you are forcefully constraining your network to a possibly invalid solution.<\/p>\n"}]}
{"QuestionId":42315225,"AnswerCount":1,"Tags":"<machine-learning><neural-network><keras>","CreationDate":"2017-02-18T12:37:41.547","AcceptedAnswerId":42348324.0,"OwnerUserId":1213715.0,"Title":"Viewing layer activations with Keras","Body":"<p>In NVIDIA's paper \"<a href=\"https:\/\/arxiv.org\/abs\/1604.07316\" rel=\"nofollow noreferrer\">End to End Learning for Self-Driving Cars<\/a>\" there's an illustration showing the activation of first-layer feature maps:<\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/abs\/1604.07316\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Kq5gm.png\" alt=\"Image from End to End Learning for Self-Driving Cars\"><\/a><\/p>\n\n<p>I'm trying to recreate this with the <a href=\"https:\/\/github.com\/commaai\/research\" rel=\"nofollow noreferrer\">Comma AI model<\/a>, but the only visualisation tools I've found are <a href=\"https:\/\/blog.keras.io\/how-convolutional-neural-networks-see-the-world.html\" rel=\"nofollow noreferrer\">Keras' instructions for gradient ascent and descent<\/a>, rather than simply viewing activations. what should I be looking for?<\/p>\n\n<p>EDIT IN RESPONSE TO COMMENT<\/p>\n\n<p>I tried implementing the code in <a href=\"https:\/\/stackoverflow.com\/questions\/41711190\/keras-how-to-get-the-output-of-each-layer\/41712013#41712013\">this answer<\/a> using the below code:<\/p>\n\n<pre><\/pre>\n\n<p>This give the following output error:<\/p>\n\n<pre><\/pre>\n\n<p>I thought this might be a problem with th vs tf dimensions, so tried changing the test input to:<\/p>\n\n<pre><\/pre>\n\n<p>which gave the following error:<\/p>\n\n<pre><\/pre>\n\n<p>EDIT: Output of model.summary()<\/p>\n\n<pre><\/pre>\n\n<p>EDIT: DEBUGGING WITH SINGLE LAYER<\/p>\n\n<p>In order to debug the issue with input shapes, I rewrote the script for a single layer:<\/p>\n\n<pre><\/pre>\n\n<p>The output from this is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>However, attempting the same on the first convolutional layer (convolution2d_1) with an 80x40 image returns the same error:<\/p>\n\n<pre><\/pre>\n\n<p>EDIT: OUTPUT LAYER DATA AS IMAGE<\/p>\n\n<p>The following code replaces the random image with a loaded one, and takes the layer output and saves it as an image:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42348324","CreationDate":"2017-02-20T15:24:39.650","ParentId":null,"OwnerUserId":"1213715","Title":null,"Body":"<p>Here is the final code that does what I want it to do, still rough and in need of tidying up:<\/p>\n\n<pre><code>from keras import backend as K\nfrom PIL import Image\nfrom scipy.misc import imread\nfrom scipy.misc import imsave\nimport numpy as np\nimport json\nfrom keras.models import model_from_json\n\nwith open('outputs\/steering_model\/steering_angle.json', 'r') as jfile:\n    z = json.load(jfile)\n    model = model_from_json(z)\nprint(\"Loaded model\")\nmodel.load_weights('outputs\/steering_model\/steering_angle.keras')\nprint(\"Loaded weights\")\n\nlayer_name = 'lambda_1'\n#layer_name = 'convolution2d_1'\n#layer_name = 'elu_1'\n#layer_name = 'convolution2d_2'\n\nimpaths = ['track','road','mway']\n\nimg_width = 500\nimg_height = 375\n\ninp = model.input\nlayer_idx = [idx for idx, layer in enumerate(model.layers) if layer.name == layer_name][0]\noutput = model.layers[layer_idx].output\nfunctor = K.function([inp]+ [K.learning_phase()], output)  # evaluation function\n\nfor impath in impaths:\n    input_img_data = imread('testimages\/'+impath+'.png').astype(np.float32)\n    input_img_data = np.rollaxis(input_img_data,2,0)   # change to (channels,h,w)\n    test = np.expand_dims(input_img_data,axis=0)       # change to (dims,channels,h,w)\n\n    print(\"Test Shape: %s\" % (test.shape,))            # check shape\n\n    layer_out = functor([test, 1.])\n    print (\"Output Shape: %s\" % (layer_out.shape,))    # check output shape\n\n    # save multiple greyscale images\n    layer_out = np.rollaxis(layer_out,0,4)\n    print (\"Output Image Shape: %s\" % (layer_out.shape,))    # check output shape    \n    count = 1\n    for x in layer_out:\n        x = np.rollaxis(x,2,0)\n        print (\"Final Image Shape: %s\" % (x.shape,))    # check output shape\n        imsave('activationimages\/activ_%s_%s_%d.png' % (layer_name,impath,count),x[0,:,:])\n        count = count + 1\n<\/code><\/pre>\n\n<p>The main issue was wrangling the shapes of the various input and output layers - hence all the print commands in the above code, for debugging.<\/p>\n\n<p>A second confusion was that I was interpreting an array shape of (3,w,h) as a single RGB (3-channel) image, rather than one greyscale image.<\/p>\n\n<p>The version above tests an array of images at a time (hardcoded image path). The lambda_1 level outputs a single RGB image per test image, convolution2d_1 and elu_1 output sixteen smaller (25%) greyscale images - one for each filter. And, I hope, so on.<\/p>\n\n<p>I will add a Github link to a tidied gist with image stitching when I've done this. I've learned a lot.<\/p>\n"}]}
{"QuestionId":42315234,"AnswerCount":2,"Tags":"<tensorflow><tflearn>","CreationDate":"2017-02-18T12:38:51.027","AcceptedAnswerId":42316723.0,"OwnerUserId":5512020.0,"Title":"AttributeError: module 'tensorflow' has no attribute 'unpack'","Body":"<p>I am trying to run a  model using tfLearn and I get this error: <\/p>\n\n<pre><\/pre>\n\n<p>with the following line:<\/p>\n\n<pre><\/pre>\n\n<p>These are the lines of code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42354530","CreationDate":"2017-02-20T21:28:09.873","ParentId":null,"OwnerUserId":"5838046","Title":null,"Body":"<p>I had the same problem and installed the latest ('bleeding edge') version of TFLearn and I did not get the 'unpack' attribute error anymore with TensorFlow 1.0 .<\/p>\n\n<p>I used the following command in terminal to install <strong>TFLearn 0.3<\/strong>:<\/p>\n\n<pre><code>pip install git+https:\/\/github.com\/tflearn\/tflearn.git\n<\/code><\/pre>\n\n<p>This is according to instructions on the TFLearn <a href=\"https:\/\/github.com\/tflearn\/tflearn#installation%22Github%20Page%22\" rel=\"nofollow noreferrer\">GitHub page<\/a>.<\/p>\n"},{"AnswerId":"42316723","CreationDate":"2017-02-18T14:55:13.730","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>In tf 1.0, there's no <code>unpack<\/code>. You may want to use <code>unstack<\/code> instead.\nTo upgrade previous code, you can refer to <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/migration<\/a>.\nBut I don't know if there's a tool for updating an entire deep learning library such like tflearn = =<\/p>\n"}]}
{"QuestionId":42315289,"AnswerCount":3,"Tags":"<python><tensorflow><computer-vision><deep-learning><keras>","CreationDate":"2017-02-18T12:44:18.550","AcceptedAnswerId":42315889.0,"OwnerUserId":4674270.0,"Title":"'Resource exhausted' memory error when trying to train a Keras model","Body":"<p>I'm trying to train a VGG19 model for a binary image classification problem. My dataset doesn't fit into the memory, so I use batches and the  function of the .<\/p>\n\n<p>However, even when trying to train with batches, I get the following error:<\/p>\n\n<blockquote>\n  <p>W tensorflow\/core\/common_runtime\/bfc_allocator.cc:275] Ran out of\n  memory trying to allocate 392.00MiB.  See logs for memory state.<\/p>\n  \n  <p>W tensorflow\/core\/framework\/op_kernel.cc:975] Resource exhausted: OOM\n  when allocating tensor with shape<\/p>\n<\/blockquote>\n\n<p>Here's the console output about my GPU when starting the training script:<\/p>\n\n<pre><\/pre>\n\n<p>I don't know, but I think 1.5+ GB should be enough to train on small batches, right?<\/p>\n\n<p>The full output of the script is quite huge and I'll paste a piece of it to <a href=\"http:\/\/pastebin.com\/dtCLZSR0\" rel=\"nofollow noreferrer\">this pastebin<\/a>.<\/p>\n\n<p>Below is the code for my model:<\/p>\n\n<pre><\/pre>\n\n<p>I run the model the following way:<\/p>\n\n<pre><\/pre>\n\n<p>and my  and  folders consist of two directories each:  and , so that  function could separate my classes correctly.<\/p>\n\n<hr>\n\n<p>What am I doing wrong here? Is it just that VGG19 is too big for my machine or it's some problem with batch sizes?<\/p>\n\n<p>What can I do to train the model on my machine?<\/p>\n\n<hr>\n\n<p>PS: if I don't interrput the training script (even though it outputs lots of similar errors like one in the pastebin above), the last lines of the output are the following:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<h2>Update 1<\/h2>\n\n<p>Following @rmeertens's advice, I've made last Dense layers smaller:<\/p>\n\n<p>last block:<\/p>\n\n<pre><\/pre>\n\n<p>and the error changed a bit. It's still an OOM error though: <a href=\"http:\/\/%20pastebin.com\/SamkUbJA\" rel=\"nofollow noreferrer\">pastebin.com\/SamkUbJA<\/a><\/p>\n","answers":[{"AnswerId":"42315889","CreationDate":"2017-02-18T13:41:43.063","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>In this case the OOM error appears because your graph is too large. What is the shape of the tensor you tried to allocate when everything goes down? <\/p>\n\n<p>Anyway, a first thing you could try is allocating the model without having any of data in memory. Is something else still running (another jupyter notebook, some other model service in the background). <\/p>\n\n<p>Also, maybe you can save space in the last layers: <\/p>\n\n<pre><code>model.add(Dense(4096, activation='relu'))\nmodel.add(Dense(4096, activation='relu'))\nmodel.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>A 4096x4096 matrix is pretty big (and immediately going back to 1 is a bad idea anyway ;) )<\/p>\n"},{"AnswerId":"42331099","CreationDate":"2017-02-19T18:25:15.593","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>If you tell me that this model is working i'd be very surprised. <\/p>\n\n<p>Having a softmax activation on 1 output (your last layer) doesnt make sense. \nThe softmax normalizes the outputs of a layer so that they sum up to 1... if you only have one output it will be 1 all the time! So if you want a binary probability, either use sigmoid on 1 output or use softmax on 2 outputs! <\/p>\n"},{"AnswerId":"46032412","CreationDate":"2017-09-04T07:40:09.023","ParentId":null,"OwnerUserId":"7148586","Title":null,"Body":"<p>In my case, I solved the problem by following way:<br>\nchange<code>Convolution2D<\/code> -> <code>Conv2D<\/code><\/p>\n\n<pre><code>from keras.layers import Conv2D\n\nmodel.add(Convolution2D(64, 3, 3, activation='relu')) -&gt; model.add(Conv2D(64, 3, 3, activation='relu'))\n<\/code><\/pre>\n"}]}
{"QuestionId":42315881,"AnswerCount":1,"Tags":"<csv><tensorflow>","CreationDate":"2017-02-18T13:40:52.910","AcceptedAnswerId":null,"OwnerUserId":7585135.0,"Title":"Tensorflow: Use softmax in training ,got result W,b value always as zero?","Body":"<p>I tried to write a tensorflow code to train samples for the first time, but I seems the weight factor W and b are always zero after every step of training. <\/p>\n\n<p>The training data are very simple, that are 10000 samples (x,y) when 00.3, y=1. I imported these data from a csv file.\nTraing data sotred in csv file shown as follows(There are 10000 data in total):<\/p>\n\n<pre><\/pre>\n\n<p>But when I run this code and print W and b in each step, I found W, b is always zero, seems like they are not been trained. Training result:<\/p>\n\n<pre><\/pre>\n\n<p>I'm confused, could anyone help me to find what the problem is?Thank you very much!\nCode is attched here:<\/p>\n\n<pre><\/pre>\n\n<p>)<\/p>\n","answers":[{"AnswerId":"42315913","CreationDate":"2017-02-18T13:43:16.193","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Jep, you could expect this problem with this weight initialisation: <\/p>\n\n<pre><code>#Weight\nW = tf.Variable(tf.zeros([1, 2]))\nb = tf.Variable(tf.zeros([2]))\n<\/code><\/pre>\n\n<p>Your weights should be initialised randomly ;)<\/p>\n"}]}
{"QuestionId":42316928,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-18T15:16:01.337","AcceptedAnswerId":42320470.0,"OwnerUserId":4993842.0,"Title":"TensorFlow logical operation ((A == B) && (C == D)) results in \"Incompatible shapes: [2] vs. [3]\"","Body":"<p>I'm trying to build the following logical expression -<\/p>\n\n<pre><\/pre>\n\n<p>But it results in the below error -<\/p>\n\n<blockquote>\n  <p>Incompatible shapes: [2] vs. [3]<\/p>\n<\/blockquote>\n\n<p> and  work fine separately, the error occurs only with .  expects boolean tensors and  returns boolean tensors, so all arguments are in order so not sure why it fails.<\/p>\n\n<p>To give some context, the original code is below and I'm just trying to update  to include both 0 and 1 for <\/p>\n\n<p>UPDATE1 Start (this adds all of the variable declarations)<\/p>\n\n<pre><\/pre>\n\n<p>UPDATE1 End (this adds all of the variable declarations)<\/p>\n\n<p><strong>This is where the problem is located -<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>How can I debug this error?<\/p>\n","answers":[{"AnswerId":"42320470","CreationDate":"2017-02-18T20:41:08.897","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem arises because <code>tf.equal()<\/code> is an <em>elementwise<\/em> operation and it returns a tensor with the same shape as its arguments. The easiest way to fix your expression is to use <code>tf.reduce_all()<\/code> to aggregate the results of <code>tf.equal()<\/code> down to a scalar before computing the <code>tf.logical_and()<\/code>, as follows:<\/p>\n\n<pre><code>tf.logical_and(\n    tf.reduce_all(tf.equal(tf.argmax(y_conv, 0), tf.argmax(y_, 0))), \n    tf.reduce_all(tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))))\n<\/code><\/pre>\n"}]}
{"QuestionId":42317075,"AnswerCount":8,"Tags":"<python><windows><tensorflow><install>","CreationDate":"2017-02-18T15:28:43.803","AcceptedAnswerId":null,"OwnerUserId":7192050.0,"Title":"Tensorflow r1.0 : could not a find a version that satisfies the requirement tensorflow","Body":"<p>I want to install Tensorflow 1.o for python on windows.<\/p>\n\n<p><strong>This is information for my system.<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>But, when I execute below command,<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I don't understand what the problem is...<\/p>\n\n<p>And I tried another way...<\/p>\n\n<p><strong>This is case when I use Conda<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>What is the problem?<\/p>\n","answers":[{"AnswerId":"53018580","CreationDate":"2018-10-27T03:56:28.263","ParentId":null,"OwnerUserId":"1166437","Title":null,"Body":"<p>From your python version output, looks like that you are using Anaconda python, in that case, there is a simple way to install tensorflow.<\/p>\n\n<pre><code>conda install -c conda-forge tensorflow\n<\/code><\/pre>\n\n<p>This command will take care of all dependencies like upgrade\/downgrade etc.<\/p>\n"},{"AnswerId":"42872285","CreationDate":"2017-03-18T08:50:15.843","ParentId":null,"OwnerUserId":"7657597","Title":null,"Body":"<p>Tensorflow on windows needs python 3.5. You can follow following steps to activate python 3.5 in anaconda:<\/p>\n\n<ol>\n<li>See which version of python you have: <code>conda search python<\/code><\/li>\n<li>If you already have python 3.5 then go to step 3\notherwise use <code>conda create -n py35 python=3.5 anaconda<\/code> to create python 3.5<\/li>\n<li>Activate python 3.5 using <code>activate py35<\/code><\/li>\n<li>Now install tensorflow using <code>conda install tensorflow<\/code><\/li>\n<\/ol>\n\n<p>If step4 is not working i.e, something like \"tensorflow: no package found \" then follow <a href=\"https:\/\/github.com\/conda-forge\/tensorflow-feedstock\" rel=\"noreferrer\">this<\/a> tutorial to forge <code>conda-forge<\/code> channel and then try installing tensorflow using step4. It worked for me. <\/p>\n"},{"AnswerId":"44218561","CreationDate":"2017-05-27T15:22:25.033","ParentId":null,"OwnerUserId":"102040","Title":null,"Body":"<p>I was getting the same error<\/p>\n\n<ol>\n<li>Get Python 3.5<\/li>\n<li>Upgrade pip version to 9<\/li>\n<li>Install tensorflow<\/li>\n<\/ol>\n\n<p>It worked for me<\/p>\n"},{"AnswerId":"57692563","CreationDate":"2019-08-28T12:34:48.143","ParentId":null,"OwnerUserId":"6738841","Title":null,"Body":"<p>upgrading <code>pip<\/code> worked for me<\/p>\n\n<pre><code>python -m pip install --upgrade pip\n<\/code><\/pre>\n"},{"AnswerId":"42317281","CreationDate":"2017-02-18T15:50:31.697","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>On Microsoft Windows, TensorFlow needs Python <strong>3.5<\/strong> 64-bit. You seem to use Python 32-bit.<\/p>\n"},{"AnswerId":"45773532","CreationDate":"2017-08-19T15:54:49.387","ParentId":null,"OwnerUserId":"3370380","Title":null,"Body":"<p>I was in same problem. <\/p>\n\n<p>Below command solved my problem<\/p>\n\n<pre><code>pip3 install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-1.0.0-py3-none-any.whl\n<\/code><\/pre>\n\n<p>to find the list of all the urls based on the python version and CPU or GPU only refer to:\n<a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/pip<\/a><\/p>\n"},{"AnswerId":"47565739","CreationDate":"2017-11-30T04:11:25.463","ParentId":null,"OwnerUserId":"1759199","Title":null,"Body":"<p>I did it with: <\/p>\n\n<pre><code>python3 -m pip install --upgrade tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"44254885","CreationDate":"2017-05-30T06:21:07.543","ParentId":null,"OwnerUserId":"7665961","Title":null,"Body":"<p><strong>Try this<\/strong><\/p>\n\n<p><strong>Installing with Anaconda<\/strong><\/p>\n\n<pre><code>conda create --name tensorflow python=3.5\nactivate tensorflow\nconda install jupyter\nconda install scipy\npip install tensorflow\nor\npip install tensorflow-gpu\n<\/code><\/pre>\n\n<p>It is important to add python=3.5 at the end of the first line, because it will install Python 3.5.<\/p>\n"}]}
{"QuestionId":42317586,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><neural-network><tflearn>","CreationDate":"2017-02-18T16:18:46.093","AcceptedAnswerId":null,"OwnerUserId":7585646.0,"Title":"Error with shape in tflearn(neural network)","Body":"<p>I'm write a code in my data test, but have error in shape.\nmy matrix \"X\" contains 4 columns and 3824 lines.\nI want to predict the dose of radiation in a certain place with 4 inputs and 1 output\nIn the future I will increase the entries but I want to do a test with this database first<\/p>\n\n<p>using example: <a href=\"https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/examples\/images\/dnn.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tflearn\/tflearn\/blob\/master\/examples\/images\/dnn.py<\/a><\/p>\n\n<p>code:<\/p>\n\n<pre><\/pre>\n\n<p>messager error:<\/p>\n\n<pre><\/pre>\n\n<p>Please help?<\/p>\n","answers":[]}
{"QuestionId":42317636,"AnswerCount":1,"Tags":"<python><graph><tensorflow><deep-learning><distributed>","CreationDate":"2017-02-18T16:23:13.453","AcceptedAnswerId":null,"OwnerUserId":7583479.0,"Title":"How to make the generated data in remote worker span iterations in in-graph replica in distributed tensorflow?","Body":"<p>I use the in-graph replication of tensorflow to do distributed training. For reducing communicaiton cost purpose, i need hold some generated data (such as the cell states in LSTM) in some remote worker in one training iteration to next iteration, but i found that i can not achieve it.<\/p>\n\n<p>If i use the fetch operation of 'session.run' interface to retrieve the data generated in one remote worker, and feed the data to this remoter worker in the next training iteration, the unnecessary network costs are produced, as below codes show:<\/p>\n\n<pre><\/pre>\n\n<p>The 'final_state[i]' generated in iteration <em>k<\/em> need be assigned to the 'initial_state[i]' in iteration <em>k+1<\/em> for every remote worker, \nhow can we do the assignment in remote worker machine without fetching to the master(grpc:\/\/localhost:port) machine and feeding again to the remote workers ?   <\/p>\n","answers":[{"AnswerId":"42499145","CreationDate":"2017-02-28T02:33:37.343","ParentId":null,"OwnerUserId":"7583479","Title":null,"Body":"<p>Variable objects and persistent tensors can replace feed_dict, as Yaroslav proposed. Thanks Yaroslav.<\/p>\n"}]}
{"QuestionId":42317906,"AnswerCount":0,"Tags":"<keras><lstm>","CreationDate":"2017-02-18T16:49:36.813","AcceptedAnswerId":null,"OwnerUserId":2439854.0,"Title":"Keras LSTM input shape","Body":"<p>I am not sure whether I understand the shape of the input I should feed into the LSTM in keras. So I have the following setup:<\/p>\n\n<pre><\/pre>\n\n<p>The input_shape is (4,20,3) where 20 is the time step and 3 is the dimensionality of my feature vector. and 4 should be the batch size I think. Lets assume that I have a time series T and I select 4 random staring point P. so then my input to the network is constructed as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I increase each p in P by 20 after every train_on_batch call and reset the state based on a random variable. As I understand the input has now 4 different batch each having its own distinct state. It is like training the network 4 times with different starting points.<\/p>\n\n<p>The target labels are contained in a different time series L and it is constructed as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I am getting some weird results and I wanted to make sure I get the initial idea right. So could you help to re insure me that it works as I have just described or tell me where I got it wrong?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":42317953,"AnswerCount":1,"Tags":"<python><tensorflow><mnist>","CreationDate":"2017-02-18T16:54:20.133","AcceptedAnswerId":null,"OwnerUserId":7585792.0,"Title":"Select specific MNIST classes to train a neural network in TensorFlow","Body":"<p>currently i am looking for a way to filter specific classes out of my training dataset (MNIST) to train a neural network on different constellations, e.g. train a network only on classes 4,5,6 then train it on 0,1,2,3,4,5,6,7,8,9 to evaluate the results with the test dataset.<\/p>\n\n<p>I'd like to do it with an argument parser via console to chose which classes should be in my training dataset so i can split this into mini batches. I think i could do it with sorting out via labels but i am kinda stuck at this moment... would appreciate any tip!!!<\/p>\n\n<p>Greetings,\nAlex<\/p>\n","answers":[{"AnswerId":"47821847","CreationDate":"2017-12-14T20:51:54.267","ParentId":null,"OwnerUserId":"7585792","Title":null,"Body":"<p>Found the answer i guess... one hot=True transformed the scalar into a one hot vector :) thanks for your time anyway!<\/p>\n"}]}
{"QuestionId":42317984,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-18T16:57:06.840","AcceptedAnswerId":null,"OwnerUserId":3102085.0,"Title":"Weights in the network are not being updated during backpropogation","Body":"<p>Here is my feature extracting network.<\/p>\n\n<pre><\/pre>\n\n<p>It is a siames fashioned CNN. The problem is that the latter layers do not get updated. And the first ones are updated to Nan. Is there something wrong with sess.run()? The way I am calling the attributes.\nUsually one of the reasons for this is faulty input. And I tried printing the values img1 and img2 and the values contained in the matrices is 255 only<\/p>\n","answers":[]}
{"QuestionId":42318495,"AnswerCount":0,"Tags":"<scipy><tensorflow><tflearn>","CreationDate":"2017-02-18T17:40:13.437","AcceptedAnswerId":null,"OwnerUserId":4057320.0,"Title":"Scipy installed but Tflearn says Scipy not supported (win7)","Body":"<p>I am using Windows 7 and tflearn (version 0.2.2, tensorflow version 0.12.1, python 3.5)and I have scipy manually installed from <a href=\"http:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/\" rel=\"nofollow noreferrer\">http:\/\/www.lfd.uci.edu\/~gohlke\/pythonlibs\/<\/a> \nWhen I am running a tflearn code there is always the error message \u201cscipy not supported\u201d. <\/p>\nThe code is running fine so far, if I don\u2019t need scipy depending things.\nWhen I use tflearn.data_preprocessing.ImageAugementation  that needs scipy, it gives me  a NameError: name \u2018scipy\u2019not defined.\nI checked the script and it tries to import scipy.ndimage.<\/p>\nI have no idea, what is causing this error, scipy is installed, but tflearn can\u2019t find it somehow\u2026<\/p>\n","answers":[]}
{"QuestionId":42318945,"AnswerCount":1,"Tags":"<tensorflow><word2vec><named-entity-recognition><syntaxnet>","CreationDate":"2017-02-18T18:19:31.367","AcceptedAnswerId":null,"OwnerUserId":2862428.0,"Title":"TensorFlow RNNs for named entity recognition","Body":"<p>I'm trying to work out what's the best model to adapt for an open named entity recognition problem (biology\/chemistry, so no dictionary of entities exists but they have to be identified by context).<\/p>\n\n<p>Currently my best guess is to adapt Syntaxnet so that instead of tagging words as N, V, ADJ etc, it learns to tag as BEGINNING, INSIDE, OUT (IOB notation).<\/p>\n\n<p>However I am not sure which of these approaches is the best?<\/p>\n\n<ul>\n<li>Syntaxnet<\/li>\n<li>word2vec<\/li>\n<li>seq2seq (I think this is not the right one as I need it to learn on two aligned sequences, whereas seq2seq is designed for sequences of differing lengths as in translation)<\/li>\n<\/ul>\n\n<p>Would be grateful for a pointer to the right method! thanks!<\/p>\n","answers":[{"AnswerId":"42322361","CreationDate":"2017-02-19T00:43:51.273","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>Syntaxnet can be used to for named entity recognition, e.g.  see: <a href=\"https:\/\/stackoverflow.com\/a\/39833949\/395857\">Named Entity Recognition with Syntaxnet<\/a><\/p>\n\n<p>word2vec alone isn't very effective for named entity recognition. I don't think seq2seq is commonly used either for that task.<\/p>\n\n<p>As drpng mentions, you may want to look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/crf\" rel=\"noreferrer\">tensorflow\/tree\/master\/tensorflow\/contrib\/crf<\/a>. Adding an LSTM before the CRF layer would help a bit, which gives <a href=\"https:\/\/arxiv.org\/pdf\/1606.03475.pdf\" rel=\"noreferrer\">something like<\/a>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/RIV1j.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/RIV1j.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>LSTM+CRF code in TensorFlow: <a href=\"https:\/\/github.com\/Franck-Dernoncourt\/NeuroNER\" rel=\"noreferrer\">https:\/\/github.com\/Franck-Dernoncourt\/NeuroNER<\/a><\/p>\n"}]}
{"QuestionId":42319101,"AnswerCount":3,"Tags":"<python><tensorflow><dataset>","CreationDate":"2017-02-18T18:33:28.660","AcceptedAnswerId":null,"OwnerUserId":3791642.0,"Title":"ImportError: No module named datasets","Body":"<p>from datasets import dataset_utils ImportError: No module named datasets.\nwhen i am writing this in python sript.<\/p>\n\n<pre><\/pre>\n\n<p>But i am getting error.<\/p>\n\n<pre><\/pre>\n\n<p>I found this solution \n<a href=\"https:\/\/stackoverflow.com\/questions\/39811840\/how-can-jupyter-access-a-new-tensorflow-module-installed-in-the-right-path\">How can jupyter access a new tensorflow module installed in the right path?<\/a>\n I did the same and i have dataset packages at path anaconda\/lib\/python2.7\/site-packages\/. Still i am getting same error.<\/p>\n","answers":[{"AnswerId":"43405184","CreationDate":"2017-04-14T04:24:15.203","ParentId":null,"OwnerUserId":"1506477","Title":null,"Body":"<p>Datasets is present in <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\/datasets\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\/datasets<\/a>\nSince 'models' are not installable from pip (at the time of writing), they are not available in python load paths by default. So either we copy them or manually add to the path.\nHere is how I setup env before running the code:<\/p>\n\n<pre><code># git clone or wget\nwget https:\/\/github.com\/tensorflow\/models\/archive\/master.zip -O models.zip \nunzip models.zip\n# add it to Python PATH\nexport PYTHONPATH=$PYTHONPATH:$PWD\/models-master\/slim\n# now we are good to call `python mytensorflow.py`\n<\/code><\/pre>\n"},{"AnswerId":"43454772","CreationDate":"2017-04-17T15:49:16.333","ParentId":null,"OwnerUserId":"2448616","Title":null,"Body":"<p>You can find the folder address on your device and append it to system path.  <\/p>\n\n<pre><code>import sys  \nsys.path.append(r\"D:\\Python35\\models\\slim\\datasets\"); import dataset_utils  \n<\/code><\/pre>\n\n<p>You'll need to do the same with 'nets' and 'preprocessing'<\/p>\n\n<pre><code>sys.path.append(r\"D:\\Python35\\models\\slim\\nets\"); import vgg\nsys.path.append(r\"D:\\Python35\\models\\slim\\preprocessing\"); import vgg_preprocessing  \n<\/code><\/pre>\n"},{"AnswerId":"42334574","CreationDate":"2017-02-20T00:17:37.863","ParentId":null,"OwnerUserId":"3151464","Title":null,"Body":"<p>It's using the datasets package in the TF-slim image models library, which is in:<\/p>\n\n<p>git clone <a href=\"https:\/\/github.com\/tensorflow\/models\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/<\/a><\/p>\n\n<p>Having done that though, in order to import the module as shown in the example on the slim image page, empty <strong>init<\/strong>.py have to be added to the models and models\/slim directories. <\/p>\n"}]}
{"QuestionId":42319337,"AnswerCount":1,"Tags":"<python><tensorflow><artificial-intelligence>","CreationDate":"2017-02-18T18:54:01.310","AcceptedAnswerId":null,"OwnerUserId":4724265.0,"Title":"Tensorflow Type Error: Value passed to parameter 'shape' has DataType float32 not in list of allowed values: int32, int64","Body":"<p>I am trying to create a DCGAN and I'm running into this error when I think I'm trying to use a linear() method:<\/p>\n\n<pre><\/pre>\n\n<p>I believe the problem resides in one of these areas:<\/p>\n\n<pre><\/pre>\n\n<p>This is the generator method where I try to reshape a tensor. Before that a build method is called which sets up all the place holders and other variables:<\/p>\n\n<pre><\/pre>\n\n<p>Finally, this is the linear() method being called:<\/p>\n\n<pre><\/pre>\n\n<p>I tried the suggestion of changing self.z to int32. I receive this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43868596","CreationDate":"2017-05-09T11:28:40.233","ParentId":null,"OwnerUserId":"7606144","Title":null,"Body":"<p>The error raises in the following line:<\/p>\n\n<pre><code>self.h0 = tf.reshape(self.z,[-1, sample_H16, sample_W16, self.gen_dimension * 8])\n<\/code><\/pre>\n\n<p>you probably only have to cast to int the parameters that may not be:<\/p>\n\n<pre><code>self.h0 = tf.reshape(self.z,[-1, sample_H16, sample_W16, int(self.gen_dimension * 8)])\n<\/code><\/pre>\n"}]}
{"QuestionId":42319632,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-18T19:20:09.053","AcceptedAnswerId":42319950.0,"OwnerUserId":2814502.0,"Title":"TensorFlow XLA with multiple gpu does not use GPU at the same time","Body":"<p>I'm trying to use XLA on multi GPU machine. bat when I turn on XLA JIT, tensorflow does not use GPU at the same time. <\/p>\n\n<p>when XLA on, gpu0 and gpu1 be active alternately.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ExRWv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ExRWv.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8SOXz.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8SOXz.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>when XLA off, gpu0 and gpu1 are both active at same time. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/srDl0.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/srDl0.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>what is happen on my environment?<\/p>\n\n<h1>my code<\/h1>\n\n<pre><\/pre>\n\n<h1>Environment info<\/h1>\n\n<p>Operating System: Ubuntu 16.04\nGPU : GTX 1080 x2\nconfig option(gcc) : -march=native -O3\nconfig option(cuda capability) : 6.1<\/p>\n\n<p>Installed version of CUDA and cuDNN: <\/p>\n\n<pre><\/pre>\n\n<p>The tensorflow commit hash c56c873fbaf976d26d487ad57c8efbc87f05331c<\/p>\n\n<p>The output of <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42319950","CreationDate":"2017-02-18T19:51:07.887","ParentId":null,"OwnerUserId":"7119965","Title":null,"Body":"<p>At the moment, XLA is single GPU.<\/p>\n"},{"AnswerId":"51055857","CreationDate":"2018-06-27T06:23:59.317","ParentId":null,"OwnerUserId":"3558756","Title":null,"Body":"<p>Don't turn on JIT compilation at the session level. You should use jit_scope() to scope which part of the model should be included in XLA JIT compilation and this works at multiple GPU training. \nFor more in details, you can refer to this link:\n<a href=\"http:\/\/danny270degree.blogspot.com\/2018\/06\/xla-jit-how-to-turn-on-xla-jit.html\" rel=\"nofollow noreferrer\">http:\/\/danny270degree.blogspot.com\/2018\/06\/xla-jit-how-to-turn-on-xla-jit.html<\/a><\/p>\n"}]}
{"QuestionId":42319786,"AnswerCount":0,"Tags":"<tensorflow><gpu><lstm>","CreationDate":"2017-02-18T19:35:12.207","AcceptedAnswerId":null,"OwnerUserId":442896.0,"Title":"Low GPU Usage & Performance with Tensorflow + RNNs","Body":"<p>I have implemented a network that tries to predict a word from a sentence. The network is actually pretty complex, but here\u2019s a simple version of it:<\/p>\n\n<ol>\n<li>Take indices of words in a sentences and convert to embeddings<\/li>\n<li>Run each sentence through LSTM<\/li>\n<li>Give each word in the sentence a score with a linear multiplication of the LSTM output<\/li>\n<\/ol>\n\n<p>And here\u2019s the code: <\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/gist.github.com\/benb969\/e3217df41ac1737d3bcb2a58d01bf316\" rel=\"nofollow noreferrer\" title=\"here\">Reproducible code on gist<\/a><\/p>\n\n<p><strong>My problem is low GPU utilization, around 40%, using Titan X.<\/strong><\/p>\n\n<p>A few notes:<\/p>\n\n<ul>\n<li>I am using a batch size of 40. While I can make it bigger, say 1000, and get very good speeds per sample (remains almost the same time per batch! ~0.7s vs ~0.8s) - I want my batch size to remain at this size due to various reasons, among others to remain similar to an article I\u2019m trying to implement.<\/li>\n<li>In this example the given batch is also the entire data, so there are no issues of batching, queuing, etc, which might make it easier to understand the problem.<\/li>\n<\/ul>\n\n<p>What I suspect to be the problem, and as noted by answers to similar problems around forums, is repeated CPU to GPU memory transfers. This is how the tracing result looks:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/JewRy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/JewRy.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The green bars (pid 9 and 11) are all MEMCPYHtoD or MEMCPYDtoH. How can I avoid them?<\/p>\n\n<p>What can I do to make speed performance better without making the batch size bigger?<\/p>\n\n<p>The sample reproducible code can be found <a href=\"https:\/\/gist.github.com\/benb969\/e3217df41ac1737d3bcb2a58d01bf316\" rel=\"nofollow noreferrer\" title=\"here\">here<\/a>. Using Tensorflow 1.0, cudnn 5.1, cuda 8<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
{"QuestionId":42320222,"AnswerCount":0,"Tags":"<neural-network><keras><conv-neural-network>","CreationDate":"2017-02-18T20:17:19.313","AcceptedAnswerId":null,"OwnerUserId":7153052.0,"Title":"What is a convolutional ply?","Body":"<p>I want to recreate the result of this <a href=\"https:\/\/www.microsoft.com\/en-us\/research\/publication\/convolutional-neural-networks-for-speech-recognition-2\/\" rel=\"nofollow noreferrer\">paper<\/a>. They use the term convolutional ply for the neural network they apply on the audio spectogram.  I am not sure I understand what a convolutional ply is, and how it differs from an ordinary convolutional neural network (cnn).<\/p>\n\n<p>The paper states this as being the difference:<\/p>\n\n<blockquote>\n  <p>A convolution ply differs from a standard, fully connected hidden\n  layer in two important aspects, however. First, each convolutional\n  unit receives input only from a local area of the input. This means\n  that each unit represents some features of a local region of the \n  input. Second, the units of the convolution ply can themselves be\n  organized into a number of feature maps, where all units in the same\n  feature map share the same weights but receive input from different\n  locations of the lower layer<\/p>\n<\/blockquote>\n\n<p>Which to me sound like a ordinary cnn network. What is the difference?<\/p>\n","answers":[]}
{"QuestionId":42322176,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-19T00:14:45.193","AcceptedAnswerId":42322993.0,"OwnerUserId":4192671.0,"Title":"Tensorflow: Custom data reader using py_func","Body":"<p>I am trying to queue up data from an hdf5 file. Since Tensorflow does not support hdf5, I created a python function that reads examples from an hdf5 file and raises  when it reaches the end of the file. I then wrap this python function with  and use it as an enqueue op for my queue.<\/p>\n\n<p>This is my code:<\/p>\n\n<pre><\/pre>\n\n<p>I expected that the  would be handled by the QueueRunner, however, I get the following error and the program crashes. Is it possible to do this kind of reading from a py_func, and if it is, what am I doing wrong? If not, what approach should I use instead?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42322993","CreationDate":"2017-02-19T02:29:41.077","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Looks like exception handling in <code>py_func<\/code> is not supported.<\/p>\n\n<p>Consider this code in <code>py_func.cc<\/code><\/p>\n\n<pre><code>\/\/ Invokes the trampoline.\n  PyObject* result = PyEval_CallObject(trampoline, args);\n  Py_DECREF(args);\n  if (result == nullptr) {\n    if (PyErr_Occurred()) {\n      \/\/ TODO(zhifengc): Consider pretty-print error using LOG(STDERR).\n      PyErr_Print();\n    }\n    return errors::Internal(\"Failed to run py callback \", call-&gt;token,\n                            \": see error log.\");\n  }\n<\/code><\/pre>\n\n<p><code>PyErr_Occurred<\/code> is set when there's an exception generated, so this will cause execution to throw <code>Failed to run py callback<\/code>.<\/p>\n\n<p><code>py_func<\/code> is an odd creature because it runs in your Python client environment. Normally when op (like reader op) fails, it which propagates from TF runtime returns not ok status to Python client, which then converts it to Python exception in <code>raise_exception_on_not_ok_status<\/code> (in client.py:session.run). Since <code>py_func<\/code> body runs in Python client, TensorFlow would need to be modified to handle <code>PyErr_Occurred<\/code> to insert a bad status back into TensorFlow runtime.<\/p>\n"}]}
{"QuestionId":42322698,"AnswerCount":1,"Tags":"<multithreading><concurrency><tensorflow><keras><python-multithreading>","CreationDate":"2017-02-19T01:35:17.700","AcceptedAnswerId":45843766.0,"OwnerUserId":233293.0,"Title":"TensorFlow\/Keras multi-threaded model fitting","Body":"<p>I'm attempting to train multiple  models with different parameter values using multiple threads (and the  backend). I've seen a few examples of using the same model within multiple threads, but in this particular case, I run into various errors regarding conflicting graphs, etc. Here's a simple example of what I'd like to be able to do:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The resulting error is . I've also tried initializing the models within the threads which gives a similar failure. <\/p>\n\n<p>Any thoughts on the best way to go about this? I'm not at all attached to this exact structure, but I'd prefer to be able to use multiple threads rather than processes so all the models are trained within the same GPU memory allocation.<\/p>\n","answers":[{"AnswerId":"45843766","CreationDate":"2017-08-23T15:24:37.183","ParentId":null,"OwnerUserId":"1679257","Title":null,"Body":"<p>Tensorflow Graphs are not threadsafe (see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph<\/a>) and when you create a new Tensorflow Session, it by default uses the default graph.<\/p>\n\n<p>You can get around this by creating a new session with a new graph in your parallelized function and constructing your keras model there.<\/p>\n\n<p>Here is some code that creates and fits a model on each available gpu in parallel:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import concurrent.futures\nimport numpy as np\n\nimport keras.backend as K\nfrom keras.layers import Dense\nfrom keras.models import Sequential\n\nimport tensorflow as tf\nfrom tensorflow.python.client import device_lib\n\ndef get_available_gpus():\n    local_device_protos = device_lib.list_local_devices()\n    return [x.name for x in local_device_protos if x.device_type == 'GPU']\n\nxdata = np.random.randn(100, 8)\nytrue = np.random.randint(0, 2, 100)\n\ndef fit(gpu):\n    with tf.Session(graph=tf.Graph()) as sess:\n        K.set_session(sess)\n        with tf.device(gpu):\n            model = Sequential()\n            model.add(Dense(12, input_dim=8, activation='relu'))\n            model.add(Dense(8, activation='relu'))\n            model.add(Dense(1, activation='sigmoid'))\n\n            model.compile(loss='binary_crossentropy', optimizer='adam')\n            model.fit(xdata, ytrue, verbose=0)\n\n            return model.evaluate(xdata, ytrue, verbose=0)\n\ngpus = get_available_gpus()\nwith concurrent.futures.ThreadPoolExecutor(len(gpus)) as executor:\n    results = [x for x in executor.map(fit, gpus)]\nprint('results: ', results)\n<\/code><\/pre>\n"}]}
{"QuestionId":42322736,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network><sequential>","CreationDate":"2017-02-19T01:41:38.340","AcceptedAnswerId":53093000.0,"OwnerUserId":1191087.0,"Title":"Reading sequential data from TFRecords files within the TensorFlow graph?","Body":"<p>I'm working with video data, but I believe this question should apply to any sequential data. I want to pass my RNN 10 sequential examples (video frames) from a TFRecords file. When I first start reading the file, I need to grab 10 examples, and use this to create a sequence-example which is then pushed onto the queue for the RNN to take when it's ready. However, now that I have the 10 frames, next time I read from the TFRecords file, I only need to take 1 example and just shift the other 9 over. But when I hit the end of the first TFRecords file, I need to restart the process on the second TFRecords file. It's my understanding that the  op will process the ops required under each condition even if that condition is not the one that is to be used. This would be a problem when using a condition to check whether to read 10 examples or only 1. Is there anyway to resolve this problem to still have the desired result outlined above?<\/p>\n","answers":[{"AnswerId":"53093000","CreationDate":"2018-10-31T22:50:15.870","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can use the recently added <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.12\/api_docs\/python\/tf\/data\/Dataset#window\" rel=\"nofollow noreferrer\"><code>Dataset.window()<\/code><\/a> transformation in TensorFlow 1.12 to do this:<\/p>\n\n<pre><code>filenames = tf.data.Dataset.list_files(...)\n\n# Define a function that will be applied to each filename, and return the sequences in that\n# file.\ndef get_examples_from_file(filename):\n  # Read and parse the examples from the file using the appropriate logic.\n  examples = tf.data.TFRecordDataset(filename).map(...)\n\n  # Selects a sliding window of 10 examples, shifting along 1 example at a time.\n  sequences = examples.window(size=10, shift=1, drop_remainder=True)\n\n  # Each element of `sequences` is a nested dataset containing 10 consecutive examples.\n  # Use `Dataset.batch()` and get the resulting tensor to convert it to a tensor value\n  # (or values, if there are multiple features in an example).\n  return sequences.map(\n      lambda d: tf.data.experimental.get_single_element(d.batch(10)))\n\n# Alternatively, you can use `filenames.interleave()` to mix together sequences from\n# different files.\nsequences = filenames.flat_map(get_examples_from_file)\n<\/code><\/pre>\n"}]}
{"QuestionId":42323035,"AnswerCount":1,"Tags":"<keras><keras-layer>","CreationDate":"2017-02-19T02:37:01.560","AcceptedAnswerId":42343210.0,"OwnerUserId":996366.0,"Title":"Assertion Error in keras","Body":"<p>I am trying to write a GAN in keras but I am getting this assertion error while running it. After searching I found out the most likely cause of the problem is an old version theano. I updated theano to the latest github development version <strong>0.9.0beta1<\/strong> but still I am getting the same error.<\/p>\n\n<pre><\/pre>\n\n<p>The problem happens in this section <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42343210","CreationDate":"2017-02-20T11:20:48.830","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>I think it's a typo in your code... Please change the last layer of the Generator from:<\/p>\n\n<pre><code>G.add(Dense(1, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>G.add(Dense(100, activation='sigmoid'))\n<\/code><\/pre>\n\n<p>I guess you didn't want your generator to produce only 1 pixel since your discriminator takes 100 inputs.<\/p>\n\n<p>The error came from the fact that your first model outputs a tensor with shape <code>(batch_size, 1)<\/code> and your second model input takes a input shape of <code>(batch_size, 100)<\/code>. Hence the assert error. <\/p>\n\n<p>It's compiling on my laptop now.<\/p>\n"}]}
{"QuestionId":42323086,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-19T02:47:45.163","AcceptedAnswerId":42323377.0,"OwnerUserId":7587086.0,"Title":"new to tensorflow, test errors, fix is to upgrade nightly?","Body":"<p>I recently installed python 3.5.3 and followed the instructions on: <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a>\nto install tensorflow using pip3.\nI ran the small test code to see if everything works and got this:<\/p>\n\n<pre><\/pre>\n\n<p>After a bit of research i found this thread <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7500\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7500<\/a> that says I have to upgrade to the latest nightly...but I don't really know what that means or how to do it exactly. Also they say this error is not a big deal but I still want to fix.<\/p>\n\n<p>EDIT: I seem to have a new problem: <\/p>\n\n<pre><\/pre>\n\n<p>Do I have to recompile things and how?<\/p>\n","answers":[{"AnswerId":"42323377","CreationDate":"2017-02-19T03:38:48.773","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>To upgrade to the latest nightly build for CPU:<\/p>\n\n<pre><code>pip install --upgrade http:\/\/ci.tensorflow.org\/view\/Nightly\/job\/nightly-win\/DEVICE=cpu,OS=windows\/lastBuild\/artifact\/cmake_build\/tf_python\/dist\/tensorflow-1.0.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>You can go on <a href=\"http:\/\/ci.tensorflow.org\/view\/Nightly\/job\/nightly-win\/\" rel=\"nofollow noreferrer\">http:\/\/ci.tensorflow.org\/view\/Nightly\/job\/nightly-win\/<\/a> for more wheels.<\/p>\n"}]}
{"QuestionId":42323388,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><chatbot><dialogflow>","CreationDate":"2017-02-19T03:41:40.677","AcceptedAnswerId":null,"OwnerUserId":7587196.0,"Title":"When to use tensorflow or api.ai","Body":"<p>I am evaluating the chatbot development technologies. Tensorflow and API.ai are (both from google) the options I am considering right now. But I am not sure which one to choose. Can you provide pointers when Tensorflow is suitable and when to use API.ai?<\/p>\n","answers":[{"AnswerId":"42628652","CreationDate":"2017-03-06T14:53:05.490","ParentId":null,"OwnerUserId":"4762340","Title":null,"Body":"<p><strong>API.ai:<\/strong><\/p>\n\n<ul>\n<li>You can build a running bot in a few minutes.<\/li>\n<li>User friendly UI.<\/li>\n<li>Develop a bot with even wright single line of code.<\/li>\n<li>Verify fast integration with messaging platform like Facebook, telegram,     slack etc..<\/li>\n<li>It is also good for prototyping bot app.<\/li>\n<li>Already trained(ready-made) lot of general purpose conversation so that you can use directly.<\/li>\n<li>Difficult to build complex conversation. <\/li>\n<\/ul>\n\n<p><strong>Tensorflow:<\/strong><\/p>\n\n<ul>\n<li>Tensorflow is general purpose machine learning\/deep learing library.<\/li>\n<li>Good to build complex conversation.<\/li>\n<li>Time consume train conversation.  <\/li>\n<\/ul>\n"},{"AnswerId":"42325104","CreationDate":"2017-02-19T08:25:44.600","ParentId":null,"OwnerUserId":"364772","Title":null,"Body":"<p>API.ai basically provides an NLP engine which understands NLP and extracts intents entities, etc. for your use. For example, you can simply build a simple chat-bot using this. <\/p>\n\n<p>On the other hand, tensorflow is a general deep learning\/machine learning framework. You can build (almost) any deep learning\/ machine learning algorithms using TensorFlow.<\/p>\n"}]}
{"QuestionId":42323640,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-19T04:29:41.960","AcceptedAnswerId":null,"OwnerUserId":4993842.0,"Title":"Tensorflow: AttributeError: module 'tensorflow' has no attribute 'Supervisor'","Body":"<p>I'm trying to use the Supervisor class to create checkpoints that can be used to save\/load partial trainings, as mentioned in the TensorFlow documentation - <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/supervisor\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/supervisor<\/a> . <\/p>\n\n<p>But when I try to use it as mentioned in the docs - <\/p>\n\n<pre><\/pre>\n\n<p>It throws the below error -<\/p>\n\n<blockquote>\n  <p>Tensorflow: AttributeError: module 'tensorflow' has no attribute 'Supervisor'<\/p>\n<\/blockquote>\n\n<p>Whats am I missing?<\/p>\n","answers":[]}
{"QuestionId":42323762,"AnswerCount":1,"Tags":"<machine-learning><neural-network><computer-vision><deep-learning><caffe>","CreationDate":"2017-02-19T04:50:01.723","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"What is the purpose of dropout layer in deploy.prototxt?","Body":"<p>I read somewhere,  during testing there is no Dropout applied. What is the purpose of  layers in ? Is not  file for testing the model? \nThanks in advance<\/p>\n","answers":[{"AnswerId":"42324588","CreationDate":"2017-02-19T07:04:33.823","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>A <code>\"Dropout\"<\/code> layer is used as a type of regularization by \"switching off\" responses at random during training: The layer will randomly set to zero <code>dropout_ratio<\/code> out of the inputs.<br>\nDepending on the way dropout is implemented, you might need to replace the probabilistic drop of the training by a scaling at test time. The test time dropout does this scaling for you.<\/p>\n\n<p>Recent versions of caffe already \"folds\" this scaling into the training phase (see <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/layers\/dropout_layer.cpp#L41\" rel=\"nofollow noreferrer\"><code>dropout_layer.cpp<\/code> line 41 and 59<\/a>), thus making this layer redundant during TEST phase. <\/p>\n"}]}
{"QuestionId":42323995,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-02-19T05:32:31.640","AcceptedAnswerId":null,"OwnerUserId":3763831.0,"Title":"Invalid Argument error Expected begin[0] = 0","Body":"<p>I am currently developing a neural network, and I got all the data and I got the code to the point that an image is being fed to the CNN for training. However, in the training process, for the first image an error pops up with the following code.<\/p>\n\n<pre><\/pre>\n\n<p>The stack trace looked like this<\/p>\n\n<pre><\/pre>\n\n<p>This error seems to originate from the data causing some confliction with the softmax function. However I have absolutely no idea what is causing this problem.<\/p>\n","answers":[{"AnswerId":"44767671","CreationDate":"2017-06-26T19:54:12.317","ParentId":null,"OwnerUserId":"6679308","Title":null,"Body":"<p>I followed this tutorial: <a href=\"https:\/\/www.kaggle.com\/sentdex\/first-pass-through-data-w-3d-convnet\" rel=\"nofollow noreferrer\">Sentdex,\nFirst pass through Data w\/ 3D ConvNet\n<\/a> to build a 3D CNN and got the same error as yours here.<\/p>\n\n<p>This error occurs because the dimension of the label vector of my input data (for example, the location of the first label vector in Sentdex's train data is <code>train_data[0][1]<\/code>) should be the same number as <code>n_classes<\/code> which in the tutorial is 2.<\/p>\n\n<p>In my wrong try, I just use a binary value 0 or 1 to represent it, whose dimension is 1 where should be 2. So the <code>tf.nn.softmax_cross_entropy_with_logits()<\/code> function was confused by the wrong size of label vector.<\/p>\n\n<p>Try expand your label vectors' dimension to be equal to your <code>n_classes<\/code>.<\/p>\n"}]}
{"QuestionId":42324140,"AnswerCount":2,"Tags":"<python><github><neural-network><deep-learning><keras>","CreationDate":"2017-02-19T05:55:40.710","AcceptedAnswerId":null,"OwnerUserId":7452543.0,"Title":"How to change Keras optimizer code","Body":"<p>I am really new to Keras so forgive me if my query is a bit silly. I installed Keras in my system using the default methods and it works fine. I want to add a new optimizer to Keras so that I can easily mention \"optimizer = mynewone \" under the model.compile function. How do I go about changing the \" optimizer.py \" code in Keras  and ensuring that the change is reflected on my Keras environment. Here is what I tried: <\/p>\n\n<p>Suppose I change the optimizer name  from rmsprop to rmsprops in the code I get the  following error:<\/p>\n\n<pre><\/pre>\n\n<p>Then when I click on optimizers.py I get the code developed by Keras in my environment. After that in the code  I replaced all \"rmsprop\" keywords with \"rmsprops\" and saved the file. So I thought I must have the updated optimizers.py in my system. But when I go back to my original file and run model.compile it throws the same error. <\/p>\n\n<p>Any help would be really appreciated. Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42326554","CreationDate":"2017-02-19T11:16:04.520","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>I think your approach is complicated and it doesn't have to be. Let's say you implement your own optimizer by subclassing keras.optimizers.Optimizer:<\/p>\n\n<pre><code>class MyOptimizer(Optimizer):\n    optimizer functions here.\n<\/code><\/pre>\n\n<p>Then to instantiate it in your model you can do this:<\/p>\n\n<pre><code>myOpt = MyOptimizer()\nmodel.compile(loss='binary_crossentropy', optimizer=myOpt, metrics= ['accuracy'])\n<\/code><\/pre>\n\n<p>Just pass an instance of your optimizer as the optimizer parameter of model.compile and that's it, Keras will now use your optimizer.<\/p>\n"},{"AnswerId":"42326255","CreationDate":"2017-02-19T10:43:00.423","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Are you sure that it is a new optimizer that you want? Not a custom objective function? Objectives can be custom it's easy to define, optimizers are trickier. <\/p>\n\n<p>There is already a huge number of optimizers with a lot of parameters. However if you really want to go down that road I would advise you to go to tensorflow! \nThen you will be able to use <a href=\"https:\/\/keras.io\/optimizers\/#tfoptimizer\" rel=\"nofollow noreferrer\">this in Keras<\/a><\/p>\n\n<p>It's all I can do for you, but maybe there is another way that I don't know of. <\/p>\n"}]}
{"QuestionId":42324981,"AnswerCount":1,"Tags":"<deep-learning><caffe><matconvnet>","CreationDate":"2017-02-19T08:04:53.223","AcceptedAnswerId":null,"OwnerUserId":5658932.0,"Title":"How to convert .caffemodel into .mat format","Body":"<p>I wanted to use my fine-tuned caffe model in matconvnet for further processing. But the script provided by Andrea vedaldi at \n<a href=\"https:\/\/github.com\/vlfeat\/matconvnet\/blob\/master\/utils\/import-caffe.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/vlfeat\/matconvnet\/blob\/master\/utils\/import-caffe.py<\/a>\nleaves the following errors which I am not able to fix:<\/p>\n\n<p>google.protobuf.text_format.ParseError: 8:1 : Message type \"caffe.NetParameter\" has no field named \"layer\".<\/p>\n\n<p>When I include my caffe.proto file in the argument to import-caffe.py it leaves the following error<\/p>\n\n<p>google.protobuf.text_format.ParseError: 8:1 : Message type \"caffe.NetParameter\" has no field named \"layer\".<\/p>\n\n<p>I would be grateful if any researcher might have already solved this issue.<\/p>\n\n<p>Thanks in advance\nTharun<\/p>\n","answers":[{"AnswerId":"42333414","CreationDate":"2017-02-19T21:59:26.103","ParentId":null,"OwnerUserId":"7511408","Title":null,"Body":"<p>Basically the error message says that the field named layer doesn't exist in the prototxt file. It's probably a typo. Check the prototxt file at line 8 column 1. My guess is that the name of the field should be layers.<\/p>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":42325223,"AnswerCount":1,"Tags":"<c++><neural-network><deep-learning><caffe>","CreationDate":"2017-02-19T08:41:08.173","AcceptedAnswerId":42325732.0,"OwnerUserId":5745846.0,"Title":"Caffe C++ - Access data in other layers","Body":"<p>I am trying to access the various layers using Caffe C++. However, I am unable to get any data:<\/p>\n\n<pre><\/pre>\n\n<p>Above is the approach I used. Even the \"data\" layer seems to be empty, with a vector size of 0. What is going on here?<\/p>\n","answers":[{"AnswerId":"42325732","CreationDate":"2017-02-19T09:44:56.193","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your code inspect layer's blobs: these are the internal parameters of the net (trainable parameters). <code>\"Data\"<\/code> layer has no internal parameters, thus the <code>size<\/code> of the blob you got is zero.<\/p>\n\n<p>On the other hand, if you want to inspect the net's data blobs (i.e., inputs and responses) you should look at <code>net_.blob_names()<\/code> and <code>net_.blobs()<\/code>.<\/p>\n"}]}
{"QuestionId":42325648,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-19T09:35:49.540","AcceptedAnswerId":null,"OwnerUserId":1879547.0,"Title":"TensorFlow: How to skip broken data","Body":"<p>I am playing with TensorFlow 1.0. My input data are bulk of jpeg images. Some of them are broken for different reasons, and I just want to skip them at input.<\/p>\n\n<p>Image loading part of the Graph is the following:<\/p>\n\n<pre><\/pre>\n\n<p>Model running part as usual:<\/p>\n\n<pre><\/pre>\n\n<p>When running this code I see the following, and I could not figure out how to catch this exception correctly.<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately, an answer given in\n<a href=\"https:\/\/stackoverflow.com\/questions\/40690603\/skipping-nonexistent-or-corrupt-files-in-tensorflow\">Skipping nonexistent or corrupt files in Tensorflow<\/a>\ndoesn't work for me. It seems that in my case an exception is raised by  which is too late.<\/p>\n","answers":[{"AnswerId":"43150512","CreationDate":"2017-03-31T22:20:53.463","ParentId":null,"OwnerUserId":"5587428","Title":null,"Body":"<p>Sorry for the late response. The answer may be contained in your error message:<\/p>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors_impl.InvalidArgumentError: Invalid JPEG data, size 0\n           [[Node: DecodeJpeg = DecodeJpegacceptable_fraction=1, channels=0, dct_method=\"\", fancy_upscaling=true, ratio=1, try_recover_truncated=false, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n<\/blockquote>\n\n<p>For whatever reason a JPEG file may be corrupted. However, you've used the default settings for <code>tf.image_decode_jpeg<\/code> which requires perfect decoding. Instead you may want to allow some error by setting parameters <code>try_recover_truncated = True<\/code> and <code>acceptable_fraction=0.5<\/code> (or whatever). See <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/tf\/image\/decode_jpeg\" rel=\"nofollow noreferrer\">this link<\/a> for more.<\/p>\n"}]}
{"QuestionId":42326245,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning><recurrent-neural-network>","CreationDate":"2017-02-19T10:42:00.530","AcceptedAnswerId":42328643.0,"OwnerUserId":6317571.0,"Title":"Tensorflow error in import tf.nn.rnn_cell","Body":"<p>I am using Tensorflow 1.0.0 and Python 3.5. \nWhen I try to do: <\/p>\n\n<pre><\/pre>\n\n<p>I get the following error: <\/p>\n\n<p>AttributeError                           <\/p>\n\n<pre><\/pre>\n\n<p>Can someone help me?<\/p>\n","answers":[{"AnswerId":"42328643","CreationDate":"2017-02-19T14:48:27.787","ParentId":null,"OwnerUserId":"1601927","Title":null,"Body":"<p>TensorFlow changes a lot of APIs before 1.0.<\/p>\n\n<p>You'll need to replace <code>tf.nn.rnn_cell.BasicLSTMCell<\/code> by <code>tf.contrib.rnn.BasicLSTMCell<\/code><\/p>\n"}]}
{"QuestionId":42326466,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-02-19T11:07:07.330","AcceptedAnswerId":47781091.0,"OwnerUserId":562769.0,"Title":"Why is the accuracy of my CNN not reproducible?","Body":"<p>I want reproducible results for the CNNs I train. Hence I set the seed in <a href=\"https:\/\/github.com\/MartinThoma\/HASY\/blob\/master\/scripts\/tf_hasy.py\" rel=\"noreferrer\">my script<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>The docs of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/set_random_seed\" rel=\"noreferrer\"><\/a> and <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.random.seed.html\" rel=\"noreferrer\"><\/a> do not report any special behaviour for a seed of .<\/p>\n\n<p>When I run the same script twice on the same machine within a couple of minutes and without making updates, I expected to get the same results. However, this is not the case:<\/p>\n\n<p>Run 1:<\/p>\n\n<pre><\/pre>\n\n<p>Run 2:<\/p>\n\n<pre><\/pre>\n\n<p>How can I make the network produce reproducible results?<\/p>\n\n<h2>System<\/h2>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42327852","CreationDate":"2017-02-19T13:34:48.030","ParentId":null,"OwnerUserId":"7353970","Title":null,"Body":"<p>Might be a scope problem. Make sure to set the seed within your scope in which you're using the graph, e.g. after <\/p>\n\n<pre><code>with tf.Graph().as_default()\n tf.set_random_seed(0)\n<\/code><\/pre>\n\n<p>This also has to be done after calling <code>tf.reset_default_graph()<\/code>.<\/p>\n\n<p>For a full example, see <a href=\"https:\/\/stackoverflow.com\/questions\/36288235\/how-to-get-stable-results-with-tensorflow-setting-random-seed\">How to get stable results with TensorFlow, setting random seed<\/a><\/p>\n"},{"AnswerId":"47781091","CreationDate":"2017-12-12T20:33:28.333","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>While I didn't solve the problem, here are possible reasons why the results are not always the same (roughly ordered from most likely\/easiest to fix to most unlikely\/hardest to fix). I also try to give a solution after the problem.<\/p>\n\n<ol>\n<li><strong>Human error<\/strong> - you missread a number \/ made a typo when you copied a result from one shell to the paper: Logging. Create an <code>2017-12-31-23-54-experiment-result.log<\/code> for every single experiment you run. Not manually,\nbut the experiment creates it. Yes, the time stamp in the name for easier finding it again. All following should be logged to that file for each single experiment.<\/li>\n<li><strong>Code<\/strong> changed: Version control (e.g. git)<\/li>\n<li><strong>Configuration file<\/strong> changed: Version control<\/li>\n<li><strong>Pseudorandom number<\/strong> changed: set seed for random \/ tensorflow \/ numpy (yes, you might have to set more than one seed)<\/li>\n<li><strong>Data loading<\/strong> differently \/ in a different order: Version control + seed (is the preprocessing really the same?)<\/li>\n<li><strong>Environment variables<\/strong> changed: Docker<\/li>\n<li><strong>Software (version)<\/strong> changed: Docker<\/li>\n<li><strong>Driver (version)<\/strong> changed: Logging<\/li>\n<li><strong>Hardware<\/strong> changed: Logging<\/li>\n<li><strong>Hardware\/Software has some reproducibility problems<\/strong>. Such as the fact that <a href=\"https:\/\/en.wikipedia.org\/wiki\/Associative_property#Nonassociativity_of_floating_point_calculation\" rel=\"nofollow noreferrer\">floating point multiplication is not associative<\/a> and different cores on a GPU might finish computations at different times (I'm not sure about this)<\/li>\n<li><strong>Hardware has errors<\/strong><\/li>\n<\/ol>\n\n<p>In any case, running the \"same\" thing multiple times might help to get a gut feeling for how different things are.<\/p>\n\n<h2>Writing a paper<\/h2>\n\n<p>If you write a paper, I think the following would be the best practice for reproducibility:<\/p>\n\n<ol>\n<li>Add a link to a <strong>repository<\/strong> (e.g. git) where all code is<\/li>\n<li>The code has to be <strong>containerized<\/strong> (e.g. Docker)<\/li>\n<li>If there is Python code and a <code>requirements.txt<\/code> you have to give the <strong>exact software version<\/strong>, not something like <code>tensorflow&gt;=1.0.0<\/code> but <code>tensorflow==1.2.3<\/code><\/li>\n<li>Add the <strong>git hash<\/strong> of the version you used for the experiments. It might be different hashes if you changed something in between.<\/li>\n<li>Always log information about <strong>drivers<\/strong> (e.g. <a href=\"https:\/\/stackoverflow.com\/a\/47781255\/562769\">like this for nVidia<\/a>) and <strong>hardware<\/strong>. Add this to the appendix of your paper. So in case of later changes one can at least check if there was a change which might cause numbers being different.<\/li>\n<\/ol>\n\n<p>For logging the versions, you might want to use something like this:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n\n# core modules\nimport subprocess\n\n\ndef get_logstring():\n    \"\"\"\n    Get important environment information that might influence experiments.\n\n    Returns\n    -------\n    logstring : str\n    \"\"\"\n    logstring = []\n    with open('\/proc\/cpuinfo') as f:\n        cpuinfo = f.readlines()\n    for line in cpuinfo:\n        if \"model name\" in line:\n            logstring.append(\"CPU: {}\".format(line.strip()))\n            break\n\n    with open('\/proc\/driver\/nvidia\/version') as f:\n        version = f.read().strip()\n    logstring.append(\"GPU driver: {}\".format(version))\n    logstring.append(\"VGA: {}\".format(find_vga()))\n    return \"\\n\".join(logstring)\n\n\ndef find_vga():\n    vga = subprocess.check_output(\"lspci | grep -i 'vga\\|3d\\|2d'\",\n                                  shell=True,\n                                  executable='\/bin\/bash')\n    return vga\n\n\nprint(get_logstring())\n<\/code><\/pre>\n\n<p>which gives something like<\/p>\n\n<pre><code>CPU: model name    : Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz\nGPU driver: NVRM version: NVIDIA UNIX x86_64 Kernel Module  384.90  Tue Sep 19 19:17:35 PDT 2017\nGCC version:  gcc version 5.4.0 20160609 (Ubuntu 5.4.0-6ubuntu1~16.04.5)\nVGA: 00:02.0 VGA compatible controller: Intel Corporation Skylake Integrated Graphics (rev 06)\n02:00.0 3D controller: NVIDIA Corporation GM108M [GeForce 940MX] (rev a2)\n<\/code><\/pre>\n"}]}
{"QuestionId":42326748,"AnswerCount":5,"Tags":"<tensorflow>","CreationDate":"2017-02-19T11:36:37.143","AcceptedAnswerId":42652258.0,"OwnerUserId":1268771.0,"Title":"tensorflow on GPU: no known devices, despite cuda's deviceQuery returning a \"PASS\" result","Body":"<blockquote>\n  <p>Note : this question was initially <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7648#issuecomment-280866214\" rel=\"noreferrer\">asked on github<\/a>, but it was asked to be here instead<\/p>\n<\/blockquote>\n\n<p>I'm having trouble running tensorflow on gpu, and it does not seems to be the usual cuda's configuration problem, because everything seems to indicate cuda is properly setup.<\/p>\n\n<p>The main symptom: when running tensorflow, my gpu is not detected (<a href=\"https:\/\/gist.github.com\/oelmekki\/cafda411bf5c2ea695d984fa98e0995b\" rel=\"noreferrer\">the code being run<\/a>, and <a href=\"https:\/\/gist.github.com\/oelmekki\/77235c6b0dde99b3438f190eb557f40f\" rel=\"noreferrer\">its output<\/a>).<\/p>\n\n<p>What differs from usual issues is that cuda seems properly installed and running  from cuda samples is successful (<a href=\"https:\/\/gist.github.com\/oelmekki\/fe65a15daec45aa90ec33b10b51d3aae\" rel=\"noreferrer\">output<\/a>).<\/p>\n\n<p>I have two graphical cards:<\/p>\n\n<ul>\n<li>an old GTX 650 used for my monitors (I don't want to use that one with tensorflow)<\/li>\n<li>a GTX 1060 that I want to dedicate to tensorflow<\/li>\n<\/ul>\n\n<p>I use:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/pypi.python.org\/pypi\/tensorflow\" rel=\"noreferrer\">tensorflow-1.0.0<\/a><\/li>\n<li>cuda-8.0 (<a href=\"https:\/\/gist.github.com\/oelmekki\/6e5e9d7d1ea871e1d73efae307efe9ce\" rel=\"noreferrer\">ls -l \/usr\/local\/cuda\/lib64\/libcud*<\/a>)<\/li>\n<li>cudnn-5.1.10<\/li>\n<li>python-2.7.12<\/li>\n<li>nvidia-drivers-375.26 (this was installed by cuda and replaced my distro driver package)<\/li>\n<\/ul>\n\n<p>I've tried:<\/p>\n\n<ul>\n<li>adding  to <\/li>\n<li>forcing gpu placement in tensorflow script using  (and  when it failed, for good measure)<\/li>\n<li>whitelisting the gpu I wanted to use with , in case the presence of my old unsupported card did cause problems<\/li>\n<li>running the script with sudo (because why not)<\/li>\n<\/ul>\n\n<p>Here are the outputs of <a href=\"https:\/\/gist.github.com\/oelmekki\/7bdcb5cc2f791cea561a60f8b21e87b5\" rel=\"noreferrer\">nvidia-smi<\/a> and <a href=\"https:\/\/gist.github.com\/oelmekki\/b83a5a0a72e8924aeb44b70b3598f9b4\" rel=\"noreferrer\">nvidia-debugdump -l<\/a>, in case it's useful.<\/p>\n\n<p>At this point, I feel like I have followed all the breadcrumbs and have no idea what I could try else. I'm not even sure if I'm contemplating a bug or a configuration problem. Any advice about how to debug this would be greatly appreciated. Thanks!<\/p>\n\n<p><strong>Update<\/strong>: with the help of Yaroslav on github, I gathered more debugging info by raising log level, but it doesn't seem to say much about the device selection : <a href=\"https:\/\/gist.github.com\/oelmekki\/760a37ca50bf58d4f03f46d104b798bb\" rel=\"noreferrer\">https:\/\/gist.github.com\/oelmekki\/760a37ca50bf58d4f03f46d104b798bb<\/a><\/p>\n\n<p><strong>Update 2<\/strong>: Using theano detects gpu correctly, but interestingly it complains about cuDNN being too recent, then fallback to cpu (<a href=\"https:\/\/gist.github.com\/oelmekki\/34b6e41a0ff2b17ff9f39bcf56d0635a\" rel=\"noreferrer\">code ran<\/a>, <a href=\"https:\/\/gist.github.com\/oelmekki\/11626d6b34058337dae64f1915e5a9fe\" rel=\"noreferrer\">output<\/a>). Maybe that could be the problem with tensorflow as well?<\/p>\n","answers":[{"AnswerId":"50957900","CreationDate":"2018-06-20T22:24:55.303","ParentId":null,"OwnerUserId":"4400634","Title":null,"Body":"<p>Might seem dumb but a <code>sudo reboot<\/code> has fixed the exact same problem for me and a couple others.<\/p>\n"},{"AnswerId":"45783412","CreationDate":"2017-08-20T14:50:59.990","ParentId":null,"OwnerUserId":"1831452","Title":null,"Body":"<p>In my case:<\/p>\n\n<pre><code>pip3 uninstall tensorflow\n<\/code><\/pre>\n\n<p>is not enough. Because when reinstall with:<\/p>\n\n<pre><code>pip3 install tensorflow-gpu\n<\/code><\/pre>\n\n<p>It's still reinstall tensorflow with cpu not gpu.\nSo, before install tensorfow-pgu, I tried to remove all related tensor folders in site-packages uninstall protobuf, and it works!<\/p>\n\n<p>For conclusion:<\/p>\n\n<pre><code>pip3 uninstall tensorflow\n<\/code><\/pre>\n\n<p>Remove all tensor folders in ~\\Python35\\Lib\\site-packages<\/p>\n\n<pre><code>pip3 uninstall protobuf\npip3 install tensorflow-gpu\n<\/code><\/pre>\n"},{"AnswerId":"57868606","CreationDate":"2019-09-10T10:14:00.230","ParentId":null,"OwnerUserId":"12045866","Title":null,"Body":"<p>This error may be caused by your GPU's compute capability, CUDA officially supports GPU's compute capability within 3.5 ~ 5.0, you can check here: <a href=\"https:\/\/en.wikipedia.org\/wiki\/CUDA\" rel=\"nofollow noreferrer\">https:\/\/en.wikipedia.org\/wiki\/CUDA<\/a><\/p>\n\n<p>In my case, the error was like this:<\/p>\n\n<blockquote>\n  <p>Ignoring visible gpu device (device: 0, name: GeForce GT 640M, pci bus id: 0000:01:00.0, compute capability: 3.0) with Cuda compute capability 3.0. The minimum required Cuda capability is 3.5.<\/p>\n<\/blockquote>\n\n<p>For now we can only compile from source code on Linux (or mac OS) to break the '3.5~5.0' limit.<\/p>\n"},{"AnswerId":"42652258","CreationDate":"2017-03-07T15:34:20.790","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>From the log output, it looks like you are running the CPU version of TensorFlow (PyPI: <a href=\"https:\/\/pypi.python.org\/pypi\/tensorflow\" rel=\"noreferrer\"><code>tensorflow<\/code><\/a>), and not the GPU version (PyPI: <a href=\"https:\/\/pypi.python.org\/pypi\/tensorflow-gpu\" rel=\"noreferrer\"><code>tensorflow-gpu<\/code><\/a>). Running the GPU version would either log information about the CUDA libraries, or an error if it failed to load them or open the driver.<\/p>\n\n<p>If you run the following commands, you should be able to use the GPU in  subsequent runs:<\/p>\n\n<pre><code>$ pip uninstall tensorflow\n$ pip install tensorflow-gpu\n<\/code><\/pre>\n"},{"AnswerId":"48079860","CreationDate":"2018-01-03T14:57:21.557","ParentId":null,"OwnerUserId":"4134062","Title":null,"Body":"<p>None of the other answers here worked for me. After a bit of tinkering I found that this fixed my issues when dealing with Tensorflow built from binary:<\/p>\n\n<hr>\n\n<p><strong>Step 0: Uninstall protobuf<\/strong><\/p>\n\n<pre><code>pip uninstall protobuf\n<\/code><\/pre>\n\n<p><strong>Step 1: Uninstall tensorflow<\/strong><\/p>\n\n<pre><code>pip uninstall tensorflow\npip uninstall tensorflow-gpu\n<\/code><\/pre>\n\n<p><strong>Step 2: Force reinstall Tensorflow with GPU support<\/strong><\/p>\n\n<pre><code>pip install --upgrade --force-reinstall tensorflow-gpu\n<\/code><\/pre>\n\n<p><strong>Step 3: If you haven't already, set CUDA_VISIBLE_DEVICES<\/strong><\/p>\n\n<p>So for me with 2 GPUs it would be<\/p>\n\n<pre><code>export CUDA_VISIBLE_DEVICES=0,1\n<\/code><\/pre>\n"}]}
{"QuestionId":42327006,"AnswerCount":2,"Tags":"<deep-learning><keras>","CreationDate":"2017-02-19T12:05:25.487","AcceptedAnswerId":null,"OwnerUserId":6222724.0,"Title":"How to calculate top5 accuracy in keras?","Body":"<p>I want to calculate top5 in imagenet2012 dataset, but i don't know how to do it in keras.\nfit function just can calculate top 1 accuracy.<\/p>\n","answers":[{"AnswerId":"51844925","CreationDate":"2018-08-14T15:27:37.133","ParentId":null,"OwnerUserId":"1159513","Title":null,"Body":"<p>Frank Wilson's answer is probably the more official answer, but just you can also calculate it like this.<\/p>\n\n<pre><code>top1 = 0.0\ntop5 = 0.0    \nclass_probs = model.predict(x)\nfor i, l in enumerate(labels):\n    class_prob = class_probs[i]\n    top_values = (-class_prob).argsort()[:5]\n    if top_values[0] == l:\n        top1 += 1.0\n    if np.isin(np.array([l]), top_values):\n        top5 += 1.0\n\nprint(\"top1 acc\", top1\/len(labels))\nprint(\"top1 acc\", top5\/len(labels))\n<\/code><\/pre>\n"},{"AnswerId":"42328087","CreationDate":"2017-02-19T13:58:29.007","ParentId":null,"OwnerUserId":"1900169","Title":null,"Body":"<p>If you are just after the topK you could always call tensorflow directly (you don't say which backend you are using).<\/p>\n\n<pre><code>from keras import backend as K\nimport tensorflow as tf\n\ntop_values, top_indices = K.get_session().run(tf.nn.top_k(_pred_test, k=5))\n<\/code><\/pre>\n\n<p>If you want an accuracy metric you can add it to your model <a href=\"https:\/\/keras.io\/metrics\/#top_k_categorical_accuracy\" rel=\"noreferrer\">'top_k_categorical_accuracy'<\/a>.<\/p>\n\n<pre><code>model.compile('adam', 'categorical_crossentropy', ['accuracy', 'top_k_categorical_accuracy'])\n\nhistory = model.fit(X_train, y_train, nb_epoch=3, validation_split=0.2)\n\nTrain on 31367 samples, validate on 7842 samples\nEpoch 1\/3\n31367\/31367 [==============================] - 6s - loss: 0.0818 - acc: 0.9765 - top_k_categorical_accuracy: 0.9996 - \n...\n<\/code><\/pre>\n\n<p>The default <code>k<\/code> for this metric is 5 but if you wanted to change that to say 3 you would set up your model like this:<\/p>\n\n<pre><code>top3_acc = functools.partial(keras.metrics.top_k_categorical_accuracy, k=3)\n\ntop3_acc.__name__ = 'top3_acc'\n\nmodel.compile('adam', 'categorical_crossentropy', ['accuracy', top3_acc])\n<\/code><\/pre>\n"}]}
{"QuestionId":42327497,"AnswerCount":1,"Tags":"<python><machine-learning><scikit-learn><neural-network><keras>","CreationDate":"2017-02-19T13:00:02.197","AcceptedAnswerId":null,"OwnerUserId":6715147.0,"Title":"Keras neural net assertion error","Body":"<p>I am having trouble with a Keras assertion Error and would like to ask if anyone can please help:<\/p>\n\n<p>I have ran Keras nn before with 2D convolution and never seen this error.<\/p>\n\n<pre><\/pre>\n\n<p>)<\/p>\n\n<p>I get an error saying:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you very much! <\/p>\n","answers":[{"AnswerId":"42344054","CreationDate":"2017-02-20T12:00:51.370","ParentId":null,"OwnerUserId":"6715147","Title":null,"Body":"<p>Problem seems to have gone away when I upgraded from Keras 1.1.1 to 1.2.0. May be a version problem.<\/p>\n"}]}
{"QuestionId":42327543,"AnswerCount":2,"Tags":"<tensorflow><neural-network><deep-learning><conv-neural-network>","CreationDate":"2017-02-19T13:05:45.443","AcceptedAnswerId":42420014.0,"OwnerUserId":195651.0,"Title":"Adam optimizer goes haywire after 200k batches, training loss grows","Body":"<p>I've been seeing a very strange behavior when training a network, where after a couple of 100k iterations (8 to 10 hours) of learning fine, everything breaks and the training loss <em>grows<\/em>:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/47T4x.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/47T4x.jpg\" alt=\"Loss explodes\"><\/a><\/p>\n\n<p>The training data itself is randomized and spread across many  files containing  examples each, then shuffled again in the input stage and batched to  examples.<\/p>\n\n<h3>The background<\/h3>\n\n<p>I am designing a network that performs four different regression tasks at the same time, e.g. determining the likelihood of an object appearing in the image and simultanously determining its orientation. The network starts with a couple of convolutional layers, some with residual connections, and then branches into the four fully-connected segments.<\/p>\n\n<p>Since the first regression results in a probability, I'm using cross entropy for the loss, whereas the others use classical L2 distance. However, due to their nature, the probability loss is around the order of , while the orientation losses can be much larger, say . I already normalized both input and output values and use clipping<\/p>\n\n<pre><\/pre>\n\n<p>in cases where things can get really bad.<\/p>\n\n<p>I've been (successfully) using the Adam optimizer to optimize on the tensor containing all distinct losses (rather than ing them), like so:<\/p>\n\n<pre><\/pre>\n\n<p>In order to display the results in TensorBoard, I then actually do<\/p>\n\n<pre><\/pre>\n\n<p>for a scalar summary.<\/p>\n\n<p>Adam is set to learning rate  and epsilon  (I see the same behavior with the default value for epislon and it breaks even faster when I keep the learning rate on ). Regularization also has no influence on this one, it does this sort-of consistently at some point.<\/p>\n\n<p>I should also add that stopping the training and restarting from the last checkpoint - implying that the training input files are shuffled again as well - results in the same behavior. The training always seems to behave similarly at that point.<\/p>\n","answers":[{"AnswerId":"42420014","CreationDate":"2017-02-23T15:34:10.663","ParentId":null,"OwnerUserId":"3990607","Title":null,"Body":"<p>Yes. This is a known problem of Adam. <\/p>\n\n<p>The equations for Adam are<\/p>\n\n<pre><code>t &lt;- t + 1\nlr_t &lt;- learning_rate * sqrt(1 - beta2^t) \/ (1 - beta1^t)\n\nm_t &lt;- beta1 * m_{t-1} + (1 - beta1) * g\nv_t &lt;- beta2 * v_{t-1} + (1 - beta2) * g * g\nvariable &lt;- variable - lr_t * m_t \/ (sqrt(v_t) + epsilon)\n<\/code><\/pre>\n\n<p>where <code>m<\/code> is an exponential moving average of the mean gradient and <code>v<\/code> is an exponential moving average of the squares of the gradients. The problem is that when you have been training for a long time, and are close to the optimal, then <code>v<\/code> can become very small. If then all of a sudden the gradients starts increasing again it will be divided by a very small number and explode. <\/p>\n\n<p>By default <code>beta1=0.9<\/code> and <code>beta2=0.999<\/code>. So <code>m<\/code> changes much more quickly than <code>v<\/code>. So <code>m<\/code> can start being big again while <code>v<\/code> is still small and cannot catch up. <\/p>\n\n<p>To remedy to this problem you can increase <code>epsilon<\/code> which is <code>10-8<\/code> by default. Thus stopping the problem of dividing almost by 0. \nDepending on your network, a value of <code>epsilon<\/code> in <code>0.1<\/code>, <code>0.01<\/code>, or <code>0.001<\/code> might be good. <\/p>\n"},{"AnswerId":"44898328","CreationDate":"2017-07-04T06:23:13.900","ParentId":null,"OwnerUserId":"1774237","Title":null,"Body":"<p>Yes this could be some sort of super complicated unstable numbers\/equations case, but most certainty your training rate is just simply to high as your loss quickly decreases until 25K and then oscillates a lot in the same level. Try to decrease it by factor of 0.1 and see what happens. You should be able to reach even lower loss value. <\/p>\n\n<p>Keep exploring! :)<\/p>\n"}]}
{"QuestionId":42327891,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-19T13:38:12.803","AcceptedAnswerId":null,"OwnerUserId":6467953.0,"Title":"When I run 'tensorflow-mnist-tutorial' and I got this: Expected int32, got list containing Tensors of type '_Message' instead","Body":"<p>when I running the Tensorflow's official sample code: '<a href=\"https:\/\/github.com\/martin-gorner\/tensorflow-mnist-tutorial\" rel=\"nofollow noreferrer\">tensorflow-mnist-tutorial<\/a>':<\/p>\n\n<pre><\/pre>\n\n<p>and something wrong:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42327956,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-19T13:44:17.853","AcceptedAnswerId":null,"OwnerUserId":2476373.0,"Title":"How to set random seed on tensorflow slim models?","Body":"<p>I am wondering how can I set the random seed for building a graph with the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\" rel=\"nofollow noreferrer\">tensorflow slim classifier models<\/a>. I understand I should use . However it is not clear where to add it, because the tf graph is not explicitly declared before calling the .<\/p>\n\n<p>I.e. there's no <\/p>\n\n<pre><\/pre>\n\n<p>to put it under<\/p>\n","answers":[]}
{"QuestionId":42328034,"AnswerCount":1,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-02-19T13:51:27.227","AcceptedAnswerId":42338474.0,"OwnerUserId":6115356.0,"Title":"Keras load_model not working after training","Body":"<p>I have used the sample code here: <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2295\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/2295<\/a><\/p>\n\n<p>I am unable to load my model after training it in keras. I get the following error: <\/p>\n\n<p><\/p>\n\n<p>I tried using HDFView and deleting the optimiser weights and reloading. But then I get the error:     <\/p>\n\n<pre><\/pre>\n\n<p>I also tried saving the architecture to JSON and saving weights separately and then loading but even that fails. <\/p>\n","answers":[{"AnswerId":"42338474","CreationDate":"2017-02-20T07:12:09.070","ParentId":null,"OwnerUserId":"6115356","Title":null,"Body":"<p>I was able to devise a workaround. I was able to load the trained model using the steps below:<\/p>\n\n<ol>\n<li><p>Create the Sequential model.\nExample:\n<code>\nmodel = Sequential()\nmodel.add(...)\nmodel.add(...)\nmodel.compile(...)\nmodel.fit(...)\n<\/code><\/p><\/li>\n<li><p>After training save only model weights using model.save_weights()\nExample:\n<code>model.save_weights(SaveLocation)<\/code><\/p><\/li>\n<li><p>To load the model weights, create the model programmatically just as in step 1 but do not use the model.compile function.\nExample:\n<code>\nmodel = Sequential()\nmodel.add(...)\nmodel.add(...)\nmodel.load_weights(weightFile)\n<\/code><\/p><\/li>\n<li><p>The weights are now loaded successfully.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42328062,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-19T13:55:26.427","AcceptedAnswerId":42342351.0,"OwnerUserId":4787681.0,"Title":"How to define loss that can't be expressed by tensors in tensorflow","Body":"<p>For example, <strong>x<\/strong> is fed into network <strong>A<\/strong> and output <strong>y<\/strong>. Then <strong>y<\/strong> is used to train a new network <strong>B<\/strong>, after <strong>B<\/strong> is trained up to a specific iteration, evaluate <strong>B<\/strong> on <strong>x<\/strong> and output a loss <strong>c<\/strong>, the loss <strong>c<\/strong> is used as the loss for training network <strong>A<\/strong>. In short, I will define a loss that is a trainable network based on the output of previous network. Is there any way to define such loss in tensorflow? Thanks.<\/p>\n","answers":[{"AnswerId":"42342351","CreationDate":"2017-02-20T10:40:18.083","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>Not sure if this works, but you could try the following: once you computed the loss <code>c<\/code>, you could feed this value back into network <code>A<\/code> using a placeholder. Then assign this placeholder to a non-trainable Variable inside model <code>A<\/code>. This assign operation will then be the \"training operation\" which you can feed into an optimizer and fetch in the optimization step. If both networks are loaded in the same graph, you probably won't even need the workaround through the placeholder.<\/p>\n"}]}
{"QuestionId":42328949,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-19T15:18:03.270","AcceptedAnswerId":null,"OwnerUserId":3604746.0,"Title":"Maybe a bug: Different eval step between CPU and GPU?","Body":"<p><strong>Environment info<\/strong><\/p>\n\n<p>ubuntu 14.04\nInstalled version of CUDA and cuDNN:\n(please attach the output of ls -l \/path\/to\/cuda\/lib\/libcud*):<\/p>\n\n<blockquote>\n  <p>-rw-r--r-- 1 root root 322936 Aug 16 2015 \/usr\/local\/cuda\/lib64\/libcudadevrt.a lrwxrwxrwx 1 root root 16 Aug 15\n  2015 \/usr\/local\/cuda\/lib64\/libcudart.so -> libcudart.so.7.5 lrwxrwxrwx\n  1 root root 19 Aug 15 2015 \/usr\/local\/cuda\/lib64\/libcudart.so.7.5 ->\n  libcudart.so.7.5.18 -rwxr-xr-x 1 root root 383336 Aug 15 2015\n  \/usr\/local\/cuda\/lib64\/libcudart.so.7.5.18 -rw-r--r-- 1 root root\n  720192 Aug 16 2015 \/usr\/local\/cuda\/lib64\/libcudart_static.a lrwxrwxrwx\n  1 root root 13 Oct 20 14:34 \/usr\/local\/cuda\/lib64\/libcudnn.so ->\n  libcudnn.so.5 lrwxrwxrwx 1 root root 17 Oct 20 14:33\n  \/usr\/local\/cuda\/lib64\/libcudnn.so.5 -> libcudnn.so.5.1.3 -rwxr-xr-x 1\n  root root 60696704 Oct 20 14:28\n  \/usr\/local\/cuda\/lib64\/libcudnn.so.5.1.3 -rw-r--r-- 1 root root\n  59715990 Oct 20 14:28 \/usr\/local\/cuda\/lib64\/libcudnn_static.a<\/p>\n<\/blockquote>\n\n<p><strong>If installed from source, provide<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I'm sorry that I can't reproduce it in minimal (mnist) example(it seems to be all right), but in my code, the global_step would pulse agin in the eval stage:<\/p>\n\n<pre><\/pre>\n\n<p>Just focus on the block after if step % FLAGS.eval_steps == 1: It is all right\uff08the step shouldn't add one\uff09 to run on CPU(add line with tf.device(\"\/cpu:0\") afterwith tf.Session() as sess: ) , But as show the below image, it will add one when we use GPU:\nOn CPU:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vsDn0.png\" rel=\"nofollow noreferrer\">image-cpu<\/a><\/p>\n\n<p>On GPU:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/sWcxB.png\" rel=\"nofollow noreferrer\">image-gpu<\/a><\/p>\n\n<p>It's very strange for me! And I try to reproduce it on MNIST but i seem to be all right both On CPU and GPU.<\/p>\n\n<p>Anyone could help me with it ?<\/p>\n","answers":[]}
{"QuestionId":42329059,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-19T15:28:09.270","AcceptedAnswerId":null,"OwnerUserId":195651.0,"Title":"How is tf.summary.tensor_summary meant to be used?","Body":"<p>TensorFlow provides a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/tensor_summary\" rel=\"noreferrer\"><\/a> function that appears to be a multidimensional variant of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/scalar\" rel=\"noreferrer\"><\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I thought it could be useful for summarizing inferred probabilities per class ... somewhat like<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However it appears that TensorBoard doesn't provide a way to display these summaries at all. How are they meant to be used?<\/p>\n","answers":[{"AnswerId":"42859256","CreationDate":"2017-03-17T13:49:34.380","ParentId":null,"OwnerUserId":"5131071","Title":null,"Body":"<p>I cannot get it to work either. It seems like that feature is still under development. See this video from the TensorFlow Dev Summit that states that the tensor_summary is still under development (starting at 9:17): <a href=\"https:\/\/youtu.be\/eBbEDRsCmv4?t=9m17s\" rel=\"nofollow noreferrer\">https:\/\/youtu.be\/eBbEDRsCmv4?t=9m17s<\/a>. It will probably be better defined and examples should be provided in the future.<\/p>\n"}]}
{"QuestionId":42329317,"AnswerCount":0,"Tags":"<visual-c++><deep-learning><caffe>","CreationDate":"2017-02-19T15:51:53.647","AcceptedAnswerId":null,"OwnerUserId":1074258.0,"Title":"when i add caffe Prebuilt binaries to my project a lot of error occurred","Body":"<p>i download caffe prebuilt binaries from : <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/windows\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/tree\/windows<\/a><\/p>\n\n<p>create project property sheet for add caffe in my project :<\/p>\n\n<pre><\/pre>\n\n<p>and write simple code to test caffe like this :<\/p>\n\n<pre><\/pre>\n\n<p>but when i want build project a lot of error occur:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pydkk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pydkk.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>how can i add caffe to my project?<\/p>\n","answers":[]}
{"QuestionId":42329394,"AnswerCount":2,"Tags":"<python><matplotlib><tensorflow><tensorboard>","CreationDate":"2017-02-19T15:58:39.533","AcceptedAnswerId":49510312.0,"OwnerUserId":386590.0,"Title":"Displaying matplotlib figure in tensorboard instead of dumping to file","Body":"<p>I am trying to display images that are generated by Tensorflow directly in Tensorboard. I tried to use this solution <a href=\"https:\/\/stackoverflow.com\/questions\/38543850\/tensorflow-how-to-display-custom-images-in-tensorboard-e-g-matplotlib-plots\">Tensorflow: How to Display Custom Images in Tensorboard (e.g. Matplotlib Plots)<\/a> but I don't get how you could link this to images generated during training as the summary is defined before creating the Tensorflow graph:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"49510312","CreationDate":"2018-03-27T10:20:11.007","ParentId":null,"OwnerUserId":"242594","Title":null,"Body":"<p>With <a href=\"http:\/\/github.com\/cheind\/tf-matplotlib\" rel=\"nofollow noreferrer\">tf-matplotlib<\/a> a simple scatter plot boils down to:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nimport tfmpl\n\n@tfmpl.figure_tensor\ndef draw_scatter(scaled, colors): \n    '''Draw scatter plots. One for each color.'''  \n    figs = tfmpl.create_figures(len(colors), figsize=(4,4))\n    for idx, f in enumerate(figs):\n        ax = f.add_subplot(111)\n        ax.axis('off')\n        ax.scatter(scaled[:, 0], scaled[:, 1], c=colors[idx])\n        f.tight_layout()\n\n    return figs\n\nwith tf.Session(graph=tf.Graph()) as sess:\n\n    # A point cloud that can be scaled by the user\n    points = tf.constant(\n        np.random.normal(loc=0.0, scale=1.0, size=(100, 2)).astype(np.float32)\n    )\n    scale = tf.placeholder(tf.float32)        \n    scaled = points*scale\n\n    # Note, `scaled` above is a tensor. Its being passed `draw_scatter` below. \n    # However, when `draw_scatter` is invoked, the tensor will be evaluated and a\n    # numpy array representing its content is provided.   \n    image_tensor = draw_scatter(scaled, ['r', 'g'])\n    image_summary = tf.summary.image('scatter', image_tensor)      \n    all_summaries = tf.summary.merge_all() \n\n    writer = tf.summary.FileWriter('log', sess.graph)\n    summary = sess.run(all_summaries, feed_dict={scale: 2.})\n    writer.add_summary(summary, global_step=0)\n<\/code><\/pre>\n\n<p>When executed, this results in the following plot inside Tensorboard\n<img src=\"https:\/\/github.com\/cheind\/tf-matplotlib\/raw\/master\/etc\/scatter.png\" alt=\"\"><\/p>\n"},{"AnswerId":"42329802","CreationDate":"2017-02-19T16:31:50.643","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>I think you are confusing two things: Tensorboard summaries (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/image\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/image<\/a>) and saving an image. <\/p>\n\n<p>If you want to display your samples in tensorboard, you have to use tf.summary.image in your code and feed it your G_sample tensor. If you look at <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/summary#Generation_of_Summaries\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/summary#Generation_of_Summaries<\/a> you get more information about how you can use Filewriters etc. <\/p>\n\n<p>Let me know if this helps, or you get stuck somewhere else!<\/p>\n"}]}
{"QuestionId":42329598,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-19T16:13:56.340","AcceptedAnswerId":null,"OwnerUserId":2234380.0,"Title":"Running distributed training with tf.learn","Body":"<p>I've been trying to set up a distributed cluster running the Boston Housing example mentioned in the TensorFlow tutorial but so far I'm a bit lost. Googling or searching in the tutorials was no help. <\/p>\n\n<pre><\/pre>\n\n<p>I'm not sure if I've set up TF_CONFIG correctly here. I used a cluster of 4 machines - two PSs and two workers but I didn't set 'environment' in cluster nor 'master' machines. I first started two PSs running, and then when I ran two workers, it was stuck right after \"INFO:tensorflow:Create CheckpointSaverHook.\" Did I do anything wrong here?<\/p>\n\n<p>I appreciate your help.<\/p>\n","answers":[{"AnswerId":"42652774","CreationDate":"2017-03-07T15:59:10.727","ParentId":null,"OwnerUserId":"5925734","Title":null,"Body":"<p>I had the exact same problem. The issue is that the grpc server never actually gets started. I made the same assumption you did - that tf.learn starts the grpc server - but it does not. You can start a server from inside your python script. Then, depending on if the process is running a 'ps' or 'worker' task, you either call <code>server.join()<\/code> or run the rest of your model's code:<\/p>\n\n<pre><code>job = sys.argv[1]\ntask = int(sys.argv[2])\n\ncluster = {'worker': ['localhost:2223'],\n           'ps': ['localhost:2222']}\n\nos.environ['TF_CONFIG'] = json.dumps({'cluster': cluster,\n                                      'task': {'type': job, 'index': task}})\n\n# Create the server\nserver = tf.train.Server(cluster,\n                         job_name=job,\n                         task_index=task)\n\nif job == \"ps\":\n    server.join()\nelif job == \"worker\":\n    # Load input\n    # estimator.fit()\n<\/code><\/pre>\n\n<p>For more information, checkout:\n<a href=\"https:\/\/stackoverflow.com\/questions\/37712509\/how-to-run-tensorflow-distributed-mnist-example\">how to run tensorflow distributed mnist example<\/a><\/p>\n\n<p>And <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/deploy\/distributed#putting-it-all-together-example-trainer-program\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/deploy\/distributed#putting-it-all-together-example-trainer-program<\/a><\/p>\n"}]}
{"QuestionId":42329874,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><conv-neural-network>","CreationDate":"2017-02-19T16:38:34.413","AcceptedAnswerId":42330406.0,"OwnerUserId":4969370.0,"Title":"What gives the logits this unexpected shape?","Body":"<p>I am currently developing an audio classifier with the Python API of TensorFlow, using the UrbanSound8K dataset, collecting exactly 176400 data points from each file and trying to distinguish between 10 mutually exclusive classes.<\/p>\n\n<p>I have adapted this example code for a convolutional neural net:\n<a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/pros\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/pros<\/a><\/p>\n\n<p>Unfortunately, I am getting the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>Here is a slightly edited version of the code:<\/p>\n\n<pre><\/pre>\n\n<p>When inspecting the Tensor shapes before calling sess.run(...) then everything is as expected.<\/p>\n\n<p>So why do the logits have the shape [7000, n_labels] instead of [batch_size, n_labels]?<\/p>\n","answers":[{"AnswerId":"42330406","CreationDate":"2017-02-19T17:26:22.387","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>Your network has incorrect structure, the crucial problem is here<\/p>\n\n<pre><code>with tf.name_scope(\"dens\"):\n    n_edges = widths[2] * channels[2]\n    wf1 = weights_x([n_edges, channels[3]], \"wf1\")\n    bf1 = biases(2, \"bf1\")\n    pf1 = tf.reshape(p2, [-1, n_edges], name=\"pf1\")\n    f1 = tf.nn.relu(tf.matmul(pf1, wf1) + bf1, name=\"f1\")\n<\/code><\/pre>\n\n<p>p2 has a shape [10, 1, 4900, 64] and n_edges is not equal to 4900 * 64 = 313600 but rather it is 448 (way too small layer!), if you make n_edges = 313600 everything is fine, however it is up to you whether this is the architecture you had in mind. It looks like you merged two non-compatible things, you used shape of the convolution kernel to compute how big is the layer to flatten it. However this is not how convolution works - shape of the layer depends on the size of the input and kernel and padding. Thus in general it is <strong>way bigger<\/strong>, and as  in this example - fully connected layer actually should have over 300k input neurons, and not as in your code - just 448. The crucial distinction here is that this fully connected layer works on the <strong>output<\/strong> of the convolution, not on the <strong>parameters<\/strong>.<\/p>\n\n<p>This 7000 is just a result of operation batch_size  * (4900 * 64) \/ (n_edges) = 10 * 313600 \/ 448 = 7000 (reshaping of pf1).<\/p>\n\n<p>The more generic fix is to do<\/p>\n\n<pre><code>p2s = p2.get_shape()\nn_edges = int(p2s[1] * p2s[2] * p2s[3])\n<\/code><\/pre>\n\n<p>since at this point all the shapes of p2 (besides 0th) are known thus can be read and used for the construction of the reminder of the network.<\/p>\n"}]}
{"QuestionId":42329959,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network><regression><deep-learning>","CreationDate":"2017-02-19T16:46:50.023","AcceptedAnswerId":47662628.0,"OwnerUserId":5729466.0,"Title":"My DNN return the same prediction for all test data (tensorflow)","Body":"<p>I am trying to run this code on my data for regression. It seems the the network can predict the first test data but all other predictions are the same as the first one. The first function generates random weights for initialization. The number of predictors is 54 and the number of outputs is 4. Here is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47662628","CreationDate":"2017-12-05T21:04:27.920","ParentId":null,"OwnerUserId":"5729466","Title":null,"Body":"<p>Normalization of the label vector was the solution.<\/p>\n"},{"AnswerId":"42335236","CreationDate":"2017-02-20T01:48:38.453","ParentId":null,"OwnerUserId":"1536848","Title":null,"Body":"<p>check the data of test_X, maybe all records of text_X is same.<\/p>\n"}]}
{"QuestionId":42330232,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-19T17:12:14.370","AcceptedAnswerId":null,"OwnerUserId":4538696.0,"Title":"Tensorflow regression model doesn't learn from negative values","Body":"<p>I use tensorflow for nontrivial regression problem. \ninput of my network (array with 192 binary (0 or 1) elements) representing poker hand, poker bets and action and output (array with single element) representing gain from this action.\nNetwork has 3 hidden layers with 100 neurons<\/p>\n\n<pre><\/pre>\n\n<p>I try to minimize square error with learning rate 0.001 and GradientDescentOptimizer:<\/p>\n\n<pre><\/pre>\n\n<p>As a result of training (100 iterations) I got for this target:<\/p>\n\n<pre><\/pre>\n\n<p>this output  :<\/p>\n\n<pre><\/pre>\n\n<p>I also didn't see any negative outputs in log files during training. So my network don't learn from negative values. Any ideas?<\/p>\n","answers":[{"AnswerId":"42330641","CreationDate":"2017-02-19T17:47:31.293","ParentId":null,"OwnerUserId":"4538696","Title":null,"Body":"<p>Simple answer:\nActivation function ReLU is: <code>max(0,x)<\/code> where <code>x<\/code> is weighted sum of inputs. So the function out always positive. You should use other activate function (like <code>tanh<\/code>) <\/p>\n"}]}
{"QuestionId":42330470,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-19T17:31:58.893","AcceptedAnswerId":null,"OwnerUserId":6770703.0,"Title":"Why are some nodes ignored by `print_model_analysis ` when `run_meta` is provided?","Body":"<p>I want to compute the number of variables and the number of floating point operations of models. However, it seems that  ignores the first node when   is provided.<\/p>\n\n<p>For example, (test with tensorflow 1.0.0)<\/p>\n\n<pre><\/pre>\n\n<p>It only contains the number of floating point operations for . How can I analyze the model correctly using ?<\/p>\n","answers":[]}
{"QuestionId":42330522,"AnswerCount":1,"Tags":"<tensorflow><nan>","CreationDate":"2017-02-19T17:36:37.137","AcceptedAnswerId":null,"OwnerUserId":7520802.0,"Title":"Tensorflow training\/validation loss nan questions","Body":"<p>I've read a few other posts here on what to do when getting nan on training\/validation loss. I'm assuming my issue is not enough decay in my learning rate though I'm wondering if someone could just take a look and agree\/disagree.<\/p>\n\n<p>I'm following the awesome blog post <a href=\"http:\/\/danielnouri.org\/notes\/2014\/12\/17\/using-convolutional-neural-nets-to-detect-facial-keypoints-tutorial\/#dropout\" rel=\"nofollow noreferrer\">here<\/a> but implementing in tensorflow. It's fairly easy to convert the model over, but the momentum and learning rate is a bit trickier and I think that's the issue. I'm having problems where I can only go so many epochs before the loss jumps up and to nan. The model I'm using should be equivalent to net4\/5 in the blog tutorial.<\/p>\n\n<pre><\/pre>\n\n<p>So I've seen this and it seems like it's just a case of needing to lower the learning rate at that point. It doesn't match up well with the tutorials numbers though which is worrying.<\/p>\n\n<p>The next step in the blog post is adding dropout. I've already implemented it into the model, where I just pass a tensor bool to tell it if its training or not. So with dropout enabled I'm getting nans in under 150 epochs and I'm not sure what the problem is. Since it's supposed to be regularizing the system I wasn't expecting this to happen.<\/p>\n\n<pre><\/pre>\n\n<p>Any thoughts on what could be the issues with having dropout enabled? I've built the exact same model afaik though even without the nan issues my losses aren't as good.<\/p>\n\n<p>My code: <a href=\"https:\/\/github.com\/sdeck51\/CNNTutorials\/blob\/master\/7.%20FacialFeatureDetection_Tutorial\/FaceDetector.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/sdeck51\/CNNTutorials\/blob\/master\/7.%20FacialFeatureDetection_Tutorial\/FaceDetector.ipynb<\/a><\/p>\n\n<p>EDIT:<\/p>\n\n<p>So I have my convolution layers set up incorrectly. I've gone over the tutorial which has this.<\/p>\n\n<pre><\/pre>\n\n<p>and I've just updated mine so I think it's now the same.<\/p>\n\n<pre><\/pre>\n\n<p>Still not working though. With dropout enabled on the convolution layers and the first fully connected layer the model lasts for under 50 epochs and then the error goes through the roof. Even with very small learning rates the problem still occurs.<\/p>\n\n<pre><\/pre>\n\n<p>And it looks like the non dropout method is broken and doing the same thing >_>...<\/p>\n\n<p>EDIT: I think I've figured out the issue. I'm using a momentum optimization algorithm that increases momentum over time. I think the small increase from that was causing it to overshoot. Currently running without dropout, but I'm getting better results than before by having a constant momentum. After I run 1000 epochs I'm going to check out it with dropout<\/p>\n\n<p>Running with dropout now and it's not blowing up so I think I've fixed the issue.<\/p>\n","answers":[{"AnswerId":"42349146","CreationDate":"2017-02-20T16:02:07.300","ParentId":null,"OwnerUserId":"7520802","Title":null,"Body":"<p>The problem was indeed the optimizer. I'm using Momentum optimizer and I have it initially set to .9 and it should become .999 towards the end of its epoch cycles. For some reason the extra momentum is causing the loss to sky rocket. Leaving it as .9 fixes the issue.<\/p>\n"}]}
{"QuestionId":42330830,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><classification>","CreationDate":"2017-02-19T18:02:08.387","AcceptedAnswerId":42331403.0,"OwnerUserId":4969370.0,"Title":"How should audio be pre-processed for classification?","Body":"<p>I am currently developing an audio classifier with the Python API of TensorFlow, using the UrbanSound8K dataset and trying to distinguish between 10 mutually exclusive classes.<\/p>\n\n<p>The audio files are 4 seconds long and contain 176400 data points which results in serious memory issues. How should the audio be pre-processed to reduce memory usage?<\/p>\n\n<p>And how can more useful features be extracted from the audio (using convolution and pooling)?<\/p>\n","answers":[{"AnswerId":"42331403","CreationDate":"2017-02-19T18:53:43.887","ParentId":null,"OwnerUserId":"7421526","Title":null,"Body":"<p>I personally prefer spectrograms as input for neural nets when it comes to sound classification. This way, raw audio data is transformed into an image representation and you can treat it like a basic image classification task.<\/p>\n\n<p>There are a number of ways to choose from, here is what I usually do using scipy, <a href=\"https:\/\/github.com\/jameslyons\/python_speech_features\" rel=\"nofollow noreferrer\">python_speech_features<\/a> and <a href=\"https:\/\/github.com\/jiaaro\/pydub\" rel=\"nofollow noreferrer\">pydub<\/a>:<\/p>\n\n<pre><code>import numpy as np\nimport scipy.io.wavfile as wave\nimport python_speech_features as psf\nfrom pydub import AudioSegment\n\n#your sound file\nfilepath = 'my-sound.wav'\n\ndef convert(path):\n\n    #open file (supports all ffmpeg supported filetypes) \n    audio = AudioSegment.from_file(path, path.split('.')[-1].lower())\n\n    #set to mono\n    audio = audio.set_channels(1)\n\n    #set to 44.1 KHz\n    audio = audio.set_frame_rate(44100)\n\n    #save as wav\n    audio.export(path, format=\"wav\")\n\ndef getSpectrogram(path, winlen=0.025, winstep=0.01, NFFT=512):\n\n    #open wav file\n    (rate,sig) = wave.read(path)\n\n    #get frames\n    winfunc=lambda x:np.ones((x,))\n    frames = psf.sigproc.framesig(sig, winlen*rate, winstep*rate, winfunc)\n\n    #Magnitude Spectrogram\n    magspec = np.rot90(psf.sigproc.magspec(frames, NFFT))\n\n    #noise reduction (mean substract)\n    magspec -= magspec.mean(axis=0)\n\n    #normalize values between 0 and 1\n    magspec -= magspec.min(axis=0)\n    magspec \/= magspec.max(axis=0)\n\n    #show spec dimensions\n    print magspec.shape    \n\n    return magspec\n\n#convert file if you need to\nconvert(filepath)\n\n#get spectrogram\nspec = getSpectrogram(filepath)\n<\/code><\/pre>\n\n<p>First, you need to standardize your audio files in terms of sample rate and channels. You can do that (and more) with the excellent <strong>pydub<\/strong> package.<\/p>\n\n<p>After that, you need to transform your audio signal into an image with FFT. You can do that with <strong>scipy.io.wavefile<\/strong> and the sigproc modul of <strong>python_speech_features<\/strong>. I like the magnitude spectrogram, rotate it 90 degrees, normalize it and use the resulting NumPy array as input for my convnets. You can change the spatial dimensions of the spectrogram by adjusting the values of <strong><em>winstep<\/em><\/strong> and <strong><em>NFFT<\/em><\/strong> to fit your input size.<\/p>\n\n<p>There might be easier ways to do all that; I achieved good overall classification results using the code above.<\/p>\n"}]}
{"QuestionId":42330839,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-19T18:02:56.000","AcceptedAnswerId":42333817.0,"OwnerUserId":5351157.0,"Title":"Batch normalization layer in Tensorflow is not updating its moving mean and moving variance","Body":"<p>Batch Normalization is not saving its moving mean and moving variance<\/p>\n\n<p>When I train I get perfect overfitting on my training data (as expected). With batch normalization the training is much faster, also as expected. However, when, immediately after a training step, I run the <strong>same model<\/strong> on the <strong>same data<\/strong> with \"is_training\" = False it gives a vastly inferior result. Furthermore, every time I look at moving_mean and moving_variance they are their default values. They never update.  <\/p>\n\n<pre><\/pre>\n\n<p>I have the update_ops code in place, but it doesn't seem to be doing the trick. update_collections = None makes it function, but I've been told that's a sub-optimal solution for performance reasons.<\/p>\n\n<pre><\/pre>\n\n<p>My code is below<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42333817","CreationDate":"2017-02-19T22:44:15.157","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<p>Batch normalization creates operations that you must run in order to update the values. That said, it also adds them to particular collections, and if you use the <code>tf.contrib.layers.optimize_loss<\/code> function, it collects these for you and runs them whenever this op is run.<\/p>\n\n<p>So to resolve, replace:<\/p>\n\n<pre><code>    train_step = AdamOptimizer().minimize(cost)\n<\/code><\/pre>\n\n<p>with<\/p>\n\n<pre><code>    train_step = optimize_loss(loss, step, learning_rate, optimizer='ADAM')\n<\/code><\/pre>\n"}]}
{"QuestionId":42331396,"AnswerCount":1,"Tags":"<python><tensorflow><keras><lstm><recurrent-neural-network>","CreationDate":"2017-02-19T18:53:05.927","AcceptedAnswerId":42336555.0,"OwnerUserId":1870832.0,"Title":"Input Shape Error in Second-layer (but not first) of Keras LSTM","Body":"<p>EDITED for conciseness.<\/p>\n\n<p>I am trying to build an LSTM model, working off the documentation example at <\/p>\n\n<blockquote>\n  <p><a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/recurrent\/<\/a><\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>The following three lines of code (plus comment) are taken directly from the documentation link above:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer lstm_2: expected\n  ndim=3, found ndim=2<\/p>\n<\/blockquote>\n\n<p>I get that error above after executing the second model.add() statement, but before exposing the model to my data, or even compiling it.<\/p>\n\n<p>What am I doing wrong here? Any help is much appreciated.\nFYI I'm using Keras 1.2.1. EDIT: Just upgraded to current 1.2.2, still having same issue.<\/p>\n","answers":[{"AnswerId":"42336555","CreationDate":"2017-02-20T04:30:40.990","ParentId":null,"OwnerUserId":"1870832","Title":null,"Body":"<p>Thanks to patyork for answering this on <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/5447\" rel=\"nofollow noreferrer\">github<\/a>:<\/p>\n\n<p>\"the second LSTM layer is not getting a 3D input that it expects (with a shape of (batch_size, timesteps, features). This is because the first LSTM layer has (by fortune of default values) return_sequences=False, meaning it only output the last feature set at time t-1 which is of shape (batch_size, 32), or 2 dimensions that doesn't include time.\"<\/p>\n\n<p>So to offer a code example of how to use a stacked LSTM to achieve many-to-one (return_sequences=False) sequence classification, just make sure to use return_sequences=True on the intermediate layers like this: <\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(32, input_dim=64, input_length=10, return_sequences=True))\nmodel.add(LSTM(24, return_sequences=True))\nmodel.add(LSTM(16, return_sequences=True))\nmodel.add(LSTM(1,  return_sequences=False))\n\nmodel.compile(optimizer = 'RMSprop', loss = 'categorical_crossentropy')\n<\/code><\/pre>\n\n<p>(no errors)<\/p>\n"}]}
{"QuestionId":42331830,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-19T19:29:40.307","AcceptedAnswerId":null,"OwnerUserId":2713740.0,"Title":"combine layers from different neural networks","Body":"<p>I am using tensorflow to train two instances of the same neural network with two different datasets. the network itself is quite simple with an input and output layer and 6 hidden layers (each layer is a 20 meurons followed by a non-linear activation function).<\/p>\n\n<p>I can train the network with two different datasets and that is fine. Now, what i want to do is basically create a new network which is a combination of these two trained networks. In particular, I want the input and the first 3 layers to be from one of the trained network and the last 3 layers and the output layer to be from the other network. I am very new to tensorflow and have not found a way to do this. Can someone point me to the API or some way to do this sort of hybrid networks?<\/p>\n","answers":[{"AnswerId":"42657707","CreationDate":"2017-03-07T20:28:49.607","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Constructing your network with Keras will make this easy; see <a href=\"https:\/\/keras.io\/\" rel=\"nofollow noreferrer\">the keras documentation<\/a> for how to reuse layers across networks.<\/p>\n"},{"AnswerId":"48858069","CreationDate":"2018-02-19T00:20:20.613","ParentId":null,"OwnerUserId":"4454404","Title":null,"Body":"<p>You might be asking about multitask learning aspect,well it can be simplified by seperating the weight matrix of each individual variables trained with different datasets and sum there weight layers individually to a sharable_weight_layer variable after a, b trained networks and finally evaluate your model as summed network in multitasking method. <\/p>\n"}]}
{"QuestionId":42333101,"AnswerCount":0,"Tags":"<python><tensorflow><lstm>","CreationDate":"2017-02-19T21:29:21.493","AcceptedAnswerId":null,"OwnerUserId":6347839.0,"Title":"Predicting Next Word of LSTM Model from Tensorflow Example","Body":"<p>My buddy and I are trying to utilize the trained model from the LSTM tensorflow example <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\" rel=\"nofollow noreferrer\">here<\/a>. We've been able to train our model, save the model, and then import the model. We've just used tensorflow's Supervisor. It was in the tutorial, but you can read more about it <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard6\/tf.train.Supervisor.md\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n\n<p>It's weird because there's not a whole lot of clear documentation for this. I understand that tensorflow is an API that's going through a lot of changes and adaptations right now, but it's hard to find clear answers. For example, we want to use , but we aren't sure if there is anything comparable to 's . <\/p>\n\n<p>More to the point, however, we <em>just want to use our model<\/em>. We want to be able to map a string using . We're not sure how to do this. We pass in a string, and we want to do a simple prediction in terms of like predicting the next word in a string. So, something similar to this:<\/p>\n\n<pre><\/pre>\n\n<p>But again, my buddy and I are pretty new to this, so we're not sure about where to go. I know this is probably too broad of a question for stack, but we've been pouring over the documentation and haven't been able to make much progress. <strong>ANY<\/strong> help would be appreciated so much! <\/p>\n\n<p>We've already found these other Stack links. Check them out <a href=\"https:\/\/stackoverflow.com\/questions\/36286594\/predicting-the-next-word-using-the-lstm-ptb-model-tensorflow-example\">here<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/33773661\/predicting-next-word-using-the-language-model-tensorflow-example\">here<\/a>.<\/p>\n\n<p>We are not sure how to associate the  probability list with any meaningful words.<\/p>\n","answers":[]}
{"QuestionId":42333163,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><keras><keras-layer>","CreationDate":"2017-02-19T21:34:00.640","AcceptedAnswerId":42339083.0,"OwnerUserId":5896419.0,"Title":"Keras: NaN Training Loss After Introducing Batch Normalization","Body":"<p>The goal of the model is to categorically classify video sequences, where each input is a grayscale, 45-frame, 100x150 video sequence (1, 45, 100, 150), which corresponds to a one-hot encoded categorical output for one of 3 classes e.g. [0, 0, 1]. Here is the script used to train the model:<\/p>\n\n<pre><\/pre>\n\n<p>Introducing batch normalization seems to have caused the training loss to be NaN very quickly:<a href=\"https:\/\/i.stack.imgur.com\/UtZ5j.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/UtZ5j.png\" alt=\"enter image description here\"><\/a> <\/p>\n","answers":[{"AnswerId":"42339083","CreationDate":"2017-02-20T07:50:04.040","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Seems like you introduce nan inputs into your network. For nan input you will get nan output. Do you by any chance have an input that is entirely 1 number? Then dividing through the variance is dividing through 0 -> nan inputs.<\/p>\n\n<p>Edit:<\/p>\n\n<p>You are using dim_ordering th. Therefore you want to batch normalize the axis 1. This has to be specified for the batch normalization layer. The default argument only works for tf dim_ordering.<\/p>\n"}]}
{"QuestionId":42333971,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><tensorlayer>","CreationDate":"2017-02-19T23:01:55.723","AcceptedAnswerId":null,"OwnerUserId":6379348.0,"Title":"how to do hyperparameters gridsearch in tensorlayer?","Body":"<p>I just started to learn tensorlayer, an excellent high-level wrapper over tensorflow framework for deep learning.  Does anyone know how to do parameter gridsearch to tune DNN? Is there anyway that I can use GridSearchCV from sklearn to tune network parameters from tensorlayer? Thanks a lot.   <\/p>\n","answers":[{"AnswerId":"50088147","CreationDate":"2018-04-29T15:18:26.380","ParentId":null,"OwnerUserId":"6804142","Title":null,"Body":"<p>you can do something like this.<\/p>\n\n<pre><code>for .... (different hyper-parameter):\n    with tf.Graph().as_default() as graph:  # clear all variables of TF\n       tl.layers.clear_layers_name()         # clear all layer name of TL\n       sess = tf.InteractiveSession()\n       # define and train a new model here\n<\/code><\/pre>\n\n<p>ref: <a href=\"https:\/\/github.com\/tensorlayer\/tensorlayer\/issues\/214\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorlayer\/tensorlayer\/issues\/214<\/a><\/p>\n"}]}
{"QuestionId":42334274,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><computer-vision>","CreationDate":"2017-02-19T23:36:41.463","AcceptedAnswerId":null,"OwnerUserId":288609.0,"Title":"get_shape() and ndims for an image read through opencv","Body":"<p>I read an image as follows<\/p>\n\n<pre><\/pre>\n\n<p>The shape of the read image is .<\/p>\n\n<p>When this read array is passed to the following code  , an error message will appear <\/p>\n\n<p>It seems to me that  and  are expecting a Tensorflow tensor. However, I read image using cv2 as the above. How to transform an image read by cv2 to a tensorflow tensor.<\/p>\n\n<p>Originally, the image is read as follows<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"55796868","CreationDate":"2019-04-22T15:31:51.597","ParentId":null,"OwnerUserId":"4674445","Title":null,"Body":"<p>Try converting it to a tensor first:<\/p>\n\n<pre><code>image_cv = cv2.imread('\/data\/cat.jpg', cv2.IMREAD_UNCHANGED)\n\nimage_tensor = tf.convert_to_tensor(image_float, dtype=np.int32)\nimage_float = tf.to_float(image_tensor, name='ToFloat')\n<\/code><\/pre>\n\n<p>BTW why do you need the <code>IMREAD_UNCHANGED<\/code> flag if it's a JPEG (with no support of alpha channel)?<\/p>\n\n<p>You can also just ask TF to load the jpeg from that file:<\/p>\n\n<pre><code>image = tf.image.decode_jpeg(tf.read_file('\/data\/cat.jpg'), channels=3)\n<\/code><\/pre>\n"}]}
{"QuestionId":42334335,"AnswerCount":1,"Tags":"<python><neural-network><nlp><keras><lstm>","CreationDate":"2017-02-19T23:45:37.037","AcceptedAnswerId":42338223.0,"OwnerUserId":5514476.0,"Title":"How to structure an LSTM neural network for classification","Body":"<p>I have data that has various conversations between two people. Each sentence has some type of classification. I am attempting to use an NLP net to classify each sentence of the conversation. I tried a convolution net and get decent results (not ground breaking tho). I figured that since this a back and forth conversation, and LSTM net may produce better results, because what was previously said may have a large impact on what follows.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lpl08.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lpl08.png\" alt=\"Type of RNN nets\"><\/a><\/p>\n\n<p>If I follow the structure above, I would assume that I am doing a many-to-many. My data looks like.<\/p>\n\n<pre><\/pre>\n\n<p>Data has been processed using word2vec. I then design my network as follows..<\/p>\n\n<pre><\/pre>\n\n<p>I assume that this setup will feed one batch of sentences in at a time. However, if in model.fit, shuffle is not equal to false its receiving shuffled batches, so why is an LSTM net even useful in this case? From research on the subject, to achieve a many-to-many structure one would need to change the LSTM layer too<\/p>\n\n<pre><\/pre>\n\n<p>and the output layer would need to be...<\/p>\n\n<pre><\/pre>\n\n<p>When switching to this structure I get an error on the input size. I'm unsure of how to reformat the data to meet this requirement, and also how to edit the embedding layer to receive the new data format.<\/p>\n\n<p>Any input would be greatly appreciated. Or if you have any suggestions on a better method, I am more than happy to hear them!<\/p>\n","answers":[{"AnswerId":"42338223","CreationDate":"2017-02-20T06:55:22.217","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Your first attempt was good. The shuffling takes place between sentences, the only shuffle the training samples between them so that they don't always come in in the same order. The words inside sentences are not shuffled. <\/p>\n\n<p>Or maybe I didn't understand the question correctly?<\/p>\n\n<p><strong>EDIT<\/strong> :<\/p>\n\n<p>After a better understanding of the question, here is my proposition.<\/p>\n\n<p><strong><em>Data preparation :<\/em><\/strong>  You slice your corpus in blocks of <code>n<\/code> sentences (they can overlap).\nYou should then have a shape like <code>(number_blocks_of_sentences, n, number_of_words_per_sentence)<\/code> so basically a list of 2D arrays which contain blocks of <code>n<\/code> sentences. <code>n<\/code> shouldn't be too big because LSTM can't handle huge number of elements in the sequence when training (vanishing gradient).\nYour targets should be an array of shape <code>(number_blocks_of_sentences, n, 1)<\/code> so also a list of 2D arrays containing the class of each sentence in your block of sentences.<\/p>\n\n<p><strong><em>Model :<\/em><\/strong> <\/p>\n\n<pre><code>n_sentences = X_train.shape[1]  # number of sentences in a sample (n)\nn_words = X_train.shape[2]      # number of words in a sentence\n\nmodel = Sequential()\n# Reshape the input because Embedding only accepts shape (batch_size, input_length) so we just transform list of sentences in huge list of words\nmodel.add(Reshape((n_sentences * n_words,),input_shape = (n_sentences, n_words)))\n# Embedding layer - output shape will be (batch_size, n_sentences * n_words, embedding_dim) so each sample in the batch is a big 2D array of words embedded \nmodel.add(Embedding(len(vocabaulary), embedding_dim, input_length = n_sentences * n_words ))\n# Recreate the sentence shaped array\nmodel.add(Reshape((n_sentences, n_words, embedding_dim))) \n# Encode each sentence - output shape is (batch_size, n_sentences, 88)\nmodel.add(TimeDistributed(LSTM(88)))\n# Go over lines and output hidden layer which contains info about previous sentences - output shape is (batch_size, n_sentences, hidden_dim)\nmodel.add(LSTM(hidden_dim, return_sequence=True))\n# Predict output binary class - output shape is (batch_size, n_sentences, 1)\nmodel.add(TimeDistributed(Dense(1,activation='sigmoid')))\n...\n<\/code><\/pre>\n\n<p>This should be a good start.<\/p>\n\n<p>I hope this helps<\/p>\n"}]}
{"QuestionId":42334646,"AnswerCount":5,"Tags":"<tensorflow>","CreationDate":"2017-02-20T00:26:37.667","AcceptedAnswerId":48535322.0,"OwnerUserId":1191087.0,"Title":"TensorFlow - Pad unknown size tensor to a specific size?","Body":"<p>Is there a way to pad a tensor of variable size to a given shape with a specific pad value? For example given the tensors:<\/p>\n\n<pre><\/pre>\n\n<p>and <\/p>\n\n<pre><\/pre>\n\n<p>Is there a way to have a generic operation which would take either and pad them with a value (say, to shape  with value ) to result in:<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>respectively? My reasoning (in case there is a better solution) is that I have examples from a TFRecords file, part of which has a variable length. For processing, a static length makes them easier to work with.<\/p>\n","answers":[{"AnswerId":"48535322","CreationDate":"2018-01-31T05:51:36.343","ParentId":null,"OwnerUserId":"4082104","Title":null,"Body":"<p>Yes. There is. Provided you do not need to change the rank of the tensor, it's very simple.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/pad\" rel=\"noreferrer\"><code>tf.pad()<\/code><\/a> accepts regular python lists with tensors. The format of the padding is a list of pairs of how much to pad on each side of that dimension.<\/p>\n\n<p>e.g.<\/p>\n\n<pre><code>t = tf.constant([[1, 2], [3, 4]])\npaddings = [[0, 0], [0, 4-tf.shape(t)[0]]]\nout = tf.pad(t, paddings, 'CONSTANT', constant_values=-1)\nsess.run(out)\n# gives: \n# array([[ 1,  2, -1, -1],\n#       [ 3,  4, -1, -1]], dtype=int32)\n<\/code><\/pre>\n\n<hr>\n\n<p>If you want to generalise this to a useful function, you could do something like:<\/p>\n\n<pre><code>def pad_up_to(t, max_in_dims, constant_values):\n    s = tf.shape(t)\n    paddings = [[0, m-s[i]] for (i,m) in enumerate(max_in_dims)]\n    return tf.pad(t, paddings, 'CONSTANT', constant_values=constant_values)\n<\/code><\/pre>\n\n<p>where <code>max_in_dims<\/code> is essentially the desired shape of the output. <em>Note:<\/em> this function will fail if you provide a shape that is strictly smaller than <code>t<\/code> in any dimension.<\/p>\n\n<p>You can use it like:<\/p>\n\n<pre><code>t = tf.constant([[1, 2], [3, 4]]) # shape = [2, 2]\nt_padded = pad_up_to(t, [2, 4], -1) # shape = [2, 4], padded with -1s\n<\/code><\/pre>\n\n<p>or<\/p>\n\n<pre><code>t = tf.placeholder(tf.float32, [None, None]) # shape = [?, ?]\nt_padded = pad_up_to(t, [5,5], -1) # shape = [5, 5], padded with -1s\nt_np = np.random.uniform(0, 1, [3,4]) # shape = [3,4], no padding\nt_padded_out = sess.run(t_padded, {t: t_np})\nt_np2 = np.random.uniform(0, 1, [2,1]) # shape = [2,1], no padding\nt_padded_out2 = sess.run(t_padded, {t: t_np2})\n<\/code><\/pre>\n\n<p>Although the dimension sizes are calculated dynamically, the number of dimensions is not, so make sure that <code>max_in_dims<\/code> has the same number of elements as t.shape.<\/p>\n"},{"AnswerId":"42335002","CreationDate":"2017-02-20T01:14:10.720","ParentId":null,"OwnerUserId":"7565103","Title":null,"Body":"<p>There is no way in TensorFlow to do this.You will have to preprocess the data and add a set of -1 to get your desired length. For example you get the length and loop through <code>list[0].append(-1)<\/code> and <code>list[1].append(-1)<\/code> the desired amount of times - the length of the list. Hope this helps!<\/p>\n"},{"AnswerId":"51936821","CreationDate":"2018-08-20T18:47:13.477","ParentId":null,"OwnerUserId":"4055338","Title":null,"Body":"<p>An extension of <a href=\"https:\/\/stackoverflow.com\/a\/48535322\/4055338\">Multihunter's solution<\/a> so that padding is only performed when necessary and does not yield an error for longer inputs:<\/p>\n\n<p>Suppose we have a sequential input called <code>inp_seq<\/code>, which is a tensor of rank 4 and should be padded in order to have a minimum length of <code>filter_size<\/code> in dimension 1.<\/p>\n\n<pre><code>def dynamic_padding(inp, min_size):\n\n    pad_size = min_size - tf.shape(inp)[1]\n    paddings = [[0, 0], [0, pad_size], [0, 0], [0, 0]] # assign here, during graph execution\n    return tf.pad(inp, paddings)\n\n# Pad only if necessary\npadded = tf.cond(tf.less(tf.shape(inp_seq)[1], filter_size), true_fn=lambda: dynamic_padding(inp_seq, filter_size), false_fn=lambda: inp_seq)  \n<\/code><\/pre>\n"},{"AnswerId":"45721424","CreationDate":"2017-08-16T19:21:17.660","ParentId":null,"OwnerUserId":"414707","Title":null,"Body":"<p>I ran into something similar.  Not fully general, but you can do something like:<\/p>\n\n<pre><code>test_a = tf.constant([[1, 2], [3, 4]])\ntest_b = tf.constant([[1, 2, 3], [4, 5, 6]])\n\ndef pad_second_dim(input, desired_size):\n    padding = tf.tile([[0]], tf.stack([tf.shape(input)[0], desired_size - tf.shape(input)[1]], 0))\n    return tf.concat([input, padding], 1)\n\nwith tf.Session() as sess:\n    print sess.run(pad_second_dim(test_a, 4))\n    # [[1 2 0 0] [3 4 0 0]]\n    print sess.run(pad_second_dim(test_b, 4))\n    # [[1 2 3 0] [4 5 6 0]]\n<\/code><\/pre>\n"},{"AnswerId":"52085537","CreationDate":"2018-08-29T20:23:26.060","ParentId":null,"OwnerUserId":"4068846","Title":null,"Body":"<p><a href=\"https:\/\/stackoverflow.com\/a\/48535322\/4068846\">The accepted answer<\/a> didn't work for me either and I am reluctant to do <a href=\"https:\/\/stackoverflow.com\/a\/45721424\/4068846\">assignments in the graph<\/a>. Here, I adjusted Multihunter's answer in such a way that it should work with variable sizes. A variation on this worked for me. Specifically, I am consuming data with <code>tf.data.TFREcordDataset<\/code> and wanted to apply padding on load instead of writing out the data pre-padded.<\/p>\n\n<pre><code>MIN_SIZE = 100\n\n# v shape is undefined on the second dimension. \nv = tf.get_variable(shape=(2, None), dtype=tf.int32)\n\n# Note: this will blow up if `padding_len` &lt; 0\npadding_len = MIN_SIZE - tf.shape(v)[-1]\n\n# paddings = [ [0, 0], [0, padding_len] ]\npaddings = tf.concat( ([[0, 0]], [[0, padding_len ]]), axis=0)\n\n# padded.shape = (2, 100)\npadded = tf.pad(t, paddings, 'CONSTANT', constant_values=-1)\n<\/code><\/pre>\n"}]}
{"QuestionId":42334674,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-02-20T00:30:12.290","AcceptedAnswerId":42345095.0,"OwnerUserId":3763831.0,"Title":"How to train CNN with an RGB Image","Body":"<p>I am currently building a CNN to differentiate between a rotten apple and a normal apple. I feel that it would be of great benefit if I could feed the CNN with rgb images. However, what exactly do I need to change to the following network?<\/p>\n\n<pre><\/pre>\n\n<p>I have tried to change certain values however I continuously get one error after another. This network is currently meant to take 28 by 28 channel 1 grayscale images. <\/p>\n","answers":[{"AnswerId":"42345095","CreationDate":"2017-02-20T12:52:29.720","ParentId":null,"OwnerUserId":"2444909","Title":null,"Body":"<p>The only difference between a gray scale and an RGB image are the number of bands, respectively 1 and 3.<\/p>\n\n<p>So, your CNN has to take 3 bands as an input, instead of 1. The rest will be handled.<\/p>\n\n<p>Without making your code run, you need at least to change : <\/p>\n\n<pre><code>weights = {'W_conv1':tf.Variable(tf.random_normal([5,5,3,32]))\nx = tf.reshape(x, shape=[-1, 28, 28, 3])\n<\/code><\/pre>\n"}]}
{"QuestionId":42334714,"AnswerCount":0,"Tags":"<python><python-3.x><windows-7><tensorflow><pip>","CreationDate":"2017-02-20T00:34:08.100","AcceptedAnswerId":null,"OwnerUserId":3006691.0,"Title":"Tensor Flow import error on Python 3.5.2 - 64bit Windows 7","Body":"<p>I installed tensor flow using pip as follows\"<\/p>\n\n<p><\/p>\n\n<p>But when I try to test tensor flow installation I get the following error:<\/p>\n\n<p>                                                                                                                                                                                                                                                                                                   <\/p>\n\n<p>Any suggestion on how to get tensor flow working based on the above errors ?<\/p>\n","answers":[]}
{"QuestionId":42334777,"AnswerCount":3,"Tags":"<python><tensorflow><pip><virtualenv><sse>","CreationDate":"2017-02-20T00:42:53.577","AcceptedAnswerId":null,"OwnerUserId":7590434.0,"Title":"Tensorflow installation using SSE instructions with pip","Body":"<p>I successfully installed cpu only tensorflow on ubuntu 16.04 using the default instructions provided <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#installing_with_virtualenv\" rel=\"noreferrer\">here<\/a>. The instructions recommended using virtualenv and pip so I did not build from source. I had no problems installing with these instructions.<\/p>\n\n<p>I validated my installation using the instructions provided <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#validate_your_installation\" rel=\"noreferrer\">further down<\/a> on the same page, and while the program ran successfully, it output the following warnings.<\/p>\n\n<pre><\/pre>\n\n<p>all the stackoverflow pages i've looked at only address people who have built from source using bazel, but don't seem to apply to people who used pip. <\/p>\n\n<p>How do I recompile or update my installation so as to the SSE instructions?<\/p>\n","answers":[{"AnswerId":"42976143","CreationDate":"2017-03-23T12:31:33.253","ParentId":null,"OwnerUserId":"5825372","Title":null,"Body":"<pre><code>bazel build --linkopt='-lrt' -c opt --copt=-mavx --copt=-msse4.2 --copt=-msse4.1 --copt=-msse3-k \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>If above solved your problem ,vote up <code>Christian Frei's<\/code> answer on <a href=\"https:\/\/stackoverflow.com\/questions\/42270739\/how-do-i-resolve-these-tensorflow-warnings\/42975902#42975902\">How do I resolve these tensorflow warnings?<\/a><\/p>\n"},{"AnswerId":"42335112","CreationDate":"2017-02-20T01:27:21.623","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You must use <code>bazel<\/code> with <code>--config=opt<\/code> option to build a wheel customized for your architecture, and then use <code>pip<\/code> to install the resulting wheel.<\/p>\n\n<p>I plan to occasionally upload 2014 MacBook + Xeon V3 optimized builds to  <a href=\"https:\/\/github.com\/yaroslavvb\/tensorflow-community-wheels\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yaroslavvb\/tensorflow-community-wheels<\/a><\/p>\n\n<p>Once you install Bazel, making a wheel for your <code>pip<\/code> command is roughly as follows<\/p>\n\n<pre><code>.\/configure\n\nexport LD_LIBRARY_PATH=\"\/usr\/local\/cuda\/lib64:\/usr\/local\/cuda\/extras\/CUPTI\/lib64:$LD_LIBRARY_PATH\"\nexport flags=\"--config=opt --config=cuda -k\"\n\nbazel build $flags -k \/\/tensorflow\/tools\/pip_package:build_pip_package\nbazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n<\/code><\/pre>\n\n<p>Then wheel is dropped into <code>\/tmp\/tensorflow_pkg<\/code><\/p>\n"},{"AnswerId":"44603206","CreationDate":"2017-06-17T09:58:57.350","ParentId":null,"OwnerUserId":"5255006","Title":null,"Body":"<p>Those are just warnings. As per this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7778\" rel=\"nofollow noreferrer\">link<\/a>, add this lines before importing TensorFlow,<\/p>\n\n<pre><code>import os\nos.environ['TF_CPP_MIN_LOG_LEVEL']='2'\nimport tensorflow as tf\n<\/code><\/pre>\n\n<p>I also have the same problem, but the suggested link solved my problem. <\/p>\n\n<p>Also look at this <a href=\"https:\/\/stackoverflow.com\/a\/43135194\/5255006\">answer<\/a>, it will help to know more about SSE information and also the different types of log indicators.<\/p>\n"}]}
{"QuestionId":42334855,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-20T00:54:46.673","AcceptedAnswerId":null,"OwnerUserId":7590388.0,"Title":"state output from tf.nn.dynamic_rnn operation","Body":"<p>For this code snippet:<\/p>\n\n<pre><\/pre>\n\n<p>I was expected the last time index from rnn_out to be equal to state. Or, perhaps the tanh of the state. But this isn't what I am seeing - they don't match. In the context of this RNN recurrence relation, what value does state contain?<\/p>\n\n<p>h(t) = tanh[b + W<em>h(t-1) + U<\/em>x(t)]<\/p>\n\n<p>The answer here, implies the last time index of rnn_out and state should be equal (but they are not):<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/40384791\/for-the-tf-nn-rnn-cell-basicrnn-whats-the-difference-between-the-state-and-outp\">for the tf.nn.rnn_cell.BasicRNN,what&#39;s the difference between the state and output<\/a><\/p>\n\n<p>The TF documentation isn't clear to me on this point.<\/p>\n","answers":[]}
{"QuestionId":42335408,"AnswerCount":1,"Tags":"<python><caffe><ubuntu-16.04>","CreationDate":"2017-02-20T02:10:09.920","AcceptedAnswerId":42338788.0,"OwnerUserId":4096101.0,"Title":"Caffe compilation error on Ubuntu 16.04","Body":"<p>I think I pretty much followed all the instructions in:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-16.04-or-15.10-Installation-Guide\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Ubuntu-16.04-or-15.10-Installation-Guide<\/a>.<\/p>\n\n<p>I still get the error:<\/p>\n\n<pre><\/pre>\n\n<p>Has anyone come across this?<\/p>\n","answers":[{"AnswerId":"42338788","CreationDate":"2017-02-20T07:31:33.177","ParentId":null,"OwnerUserId":"7511408","Title":null,"Body":"<p>Did you install OpenBLAS? Try the following command and repeat the compillation process:<\/p>\n\n<pre><code>sudo apt-get install -y libopenblas-dev\n<\/code><\/pre>\n\n<p>If it fails, you could try compilling caffe with cmake.<\/p>\n\n<p>From the build folder, do:<\/p>\n\n<pre><code>cmake -DBLAS=open ..\nmake all -j $(($(nproc) + 1))\nmake install\nmake runtest\n<\/code><\/pre>\n\n<p>You should provide the config file to better understand the links.<\/p>\n"}]}
{"QuestionId":42335630,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><mnist><kaggle>","CreationDate":"2017-02-20T02:40:32.637","AcceptedAnswerId":null,"OwnerUserId":7419836.0,"Title":"Same code on mnist from two different data sources, but very different results","Body":"<p>I followed tensorflow tutorial and the my code yield 90%ish prediction rate as expected. <\/p>\n\n<p>However, I tried the same code with mnist data set from Kaggle in the link below,\n<a href=\"https:\/\/www.kaggle.com\/c\/digit-recognizer\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/digit-recognizer<\/a>\nbut the prediction rate is completely screwed.<\/p>\n\n<p>I carefully checked the training data and labels format, but could not resolve where the difference come from. Please help me with the problem.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42335856,"AnswerCount":1,"Tags":"<python><linux><tensorflow><keras><keras-layer>","CreationDate":"2017-02-20T03:10:21.193","AcceptedAnswerId":42341893.0,"OwnerUserId":835269.0,"Title":"Keras: ValueError: Input 0 is incompatible layer issues","Body":"<p>I am using Keras with Tensorflow as backends and get incompatible errors:<\/p>\n\n<pre><\/pre>\n\n<p>The following error shows:<\/p>\n\n<pre><\/pre>\n\n<p>How can I fix this problem?<\/p>\n\n<p>Keras version: 1.2.2\nTensorflow version: 0.12<\/p>\n","answers":[{"AnswerId":"42341893","CreationDate":"2017-02-20T10:19:41.920","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The <code>LSTM<\/code> layer is accepting input in shape of <code>(len_of_sequences, nb_of_features)<\/code>. The input shape you provided is only <code>1-dim<\/code> so this where error comes from. The exact form of error message comes from the fact that the actual shape of data includes the <code>batch_size<\/code>. So the actual shape of data fed to the layer is <code>(batch_size, len_of_sequences, nb_of_features)<\/code>. Your shape is <code>(batch_size, 1)<\/code> and this is the reason behind <code>3d<\/code> vs <code>2d<\/code> inputs.<\/p>\n\n<p>Moreover - you might have a similiar problem with a second layer. In order to make your <code>LSTM<\/code> layer to return a sequences you should change its definition to:<\/p>\n\n<pre><code>model.add(LSTM(64, input_shape = (len_of_seq, nb_of_features), return_sequences=True)\n<\/code><\/pre>\n\n<p>or:<\/p>\n\n<pre><code>model.add(LSTM(64, input_dim = nb_of_features, input_len = len_of_sequence, return_sequences=True)\n<\/code><\/pre>\n"}]}
{"QuestionId":42337210,"AnswerCount":0,"Tags":"<gpu><cpu><caffe>","CreationDate":"2017-02-20T05:37:44.020","AcceptedAnswerId":null,"OwnerUserId":5869383.0,"Title":"Why GPU takes longer time to process first frame of video than CPU in Caffe?","Body":"<p>I tried to run the code of <a href=\"https:\/\/github.com\/alexgkendall\/SegNet-Tutorial\/blob\/master\/Scripts\/webcam_demo.py\" rel=\"nofollow noreferrer\">SegNet<\/a> web demo. When I use<\/p>\n\n<pre><\/pre>\n\n<p>the average frame rate is about 9 second per frame. <\/p>\n\n<p>However, when I use <\/p>\n\n<pre><\/pre>\n\n<p>it takes about <strong>107s<\/strong> to process the first frame and about <strong>5 frames per second<\/strong> starting from the second frame. <\/p>\n\n<p>Why it takes longer time using GPU than using CPU when it processes first frame? Is there any way to improve the frame rate while processing the first frame?<\/p>\n\n<p>Thank you!!!! Really need help. <\/p>\n","answers":[]}
{"QuestionId":42337497,"AnswerCount":0,"Tags":"<macos><installation><tensorflow>","CreationDate":"2017-02-20T06:04:18.797","AcceptedAnswerId":null,"OwnerUserId":1099571.0,"Title":"What is the minimum Mac OS version for Tensorflow 1.0.0?","Body":"<p>I tried to install tensorflow on my Mac OS X 10.10.5 (Yosemite) system, with Python 3.6. This failed with the message:<\/p>\n\n<p><\/p>\n\n<p>There is of course a <a href=\"https:\/\/pypi.python.org\/pypi\/tensorflow\/1.0.0\" rel=\"nofollow noreferrer\"> package on pypi<\/a>.  There is no requirement for macOS version number listed on that page, but the installable \"whale\" files have names like . <\/p>\n\n<p>I suspect that the \"\" notation means that one must have Mac OS X 10.11 \"El Capitan\" or higher to install, but I can't find a statement to that effect.  <\/p>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/install\/install_mac\" rel=\"nofollow noreferrer\">Tensorflow Mac installation instructions<\/a> are silent on macOS version requirements.  They also direct us to ask installation questions here on SO. The optional <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow noreferrer\">nVidia CUDA<\/a> library seems to require 10.11 or higher.<\/p>\n\n<p>Can anyone give a definitive answer?<\/p>\n","answers":[]}
{"QuestionId":42337549,"AnswerCount":1,"Tags":"<neural-network><protocol-buffers><deep-learning><caffe><hdf5>","CreationDate":"2017-02-20T06:08:10.750","AcceptedAnswerId":42337834.0,"OwnerUserId":2467772.0,"Title":"How to load trained caffe model in h5 format to c++ caffe net?","Body":"<p>Normally trained caffe model are in  extension and actually they are in  format. <\/p>\n\n<p>Any idea how to load a caffe model in  format to caffe net in c++?<\/p>\n\n<p>I have a model trained with python caffe in  format.<\/p>\n\n<p>My application is in c++ using caffe c++ version and I prefer to use c++ than python.<\/p>\n\n<p>How to read the model in caffe trained model in hdf5 format to c++ caffe net?<\/p>\n\n<p>I know that caffe has hdf5data layer inside.\nIs there a sample program for that?<\/p>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I used CopyTrainedLayersFromHDF5() api and got following runtime errors.<\/p>\n\n<pre><\/pre>\n\n<p><strong>EDIT 1:<\/strong> <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42337834","CreationDate":"2017-02-20T06:28:53.633","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Have you considered the <code>net<\/code> object method <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/include\/caffe\/net.hpp#L116\" rel=\"nofollow noreferrer\"><code>void CopyTrainedLayersFromHDF5(const string trained_filename);<\/code><\/a>? It seems like it does what you are looking for.<\/p>\n\n<p>As for <code>\"HDF5Data\"<\/code> layer: you are confusing two things here. The hdf5 file you have stores the <strong>trained parameters<\/strong> of the net. In contrast, <code>\"HDF5Data\"<\/code> layer stores the <strong>training examples<\/strong> used for training the net. <\/p>\n"}]}
{"QuestionId":42337863,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-20T06:30:35.460","AcceptedAnswerId":null,"OwnerUserId":1629860.0,"Title":"Tensorflow can't work on Zeppelin","Body":"<p>I want to run tensorflow on Zeppelin:System Ubuntu 16.04,python2.7. \nFirst,I directly pip install tensorflow on system  and download zeppelin\nbinary from the website. When I test the tensorflow from the system commandline it's ok. But when I start Zeppelin and use the Zeppelin from\nanother pc get errors:<\/p>\n\n<pre><\/pre>\n\n<p>I search the web and think this error may caused by the Pip install. So I download the source and compile the tensorflow. It's still the same error.<\/p>\n\n<p>But when I use the jupyter notebook. It's OK!<br>\nSo I ask for help\nthank you<\/p>\n","answers":[{"AnswerId":"42337922","CreationDate":"2017-02-20T06:34:45.533","ParentId":null,"OwnerUserId":"1950503","Title":null,"Body":"<p>You're running the script in the installation directory. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3217\" rel=\"nofollow noreferrer\">This is a common issue<\/a> with Tensorflow.<\/p>\n"}]}
{"QuestionId":42338277,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-20T06:58:48.103","AcceptedAnswerId":null,"OwnerUserId":7591344.0,"Title":"Is it Ok that creating TensorFlow device multiple times","Body":"<p>I've run a image processing script using tensorflow API. It turns out that the processing time decreased quickly when I set the for-loop outside the session running procedure. Could anyone tell me why? Is there any side-effects?<\/p>\n\n<p>The original code:<\/p>\n\n<pre><\/pre>\n\n<p>The code modified: <\/p>\n\n<pre><\/pre>\n\n<p>The time cost in the original code would keep increasing from 200ms to even 20000ms. While after modified, the the logs messages indicate that there are more than one graph and tensorflow devices were created during running, why is that?<\/p>\n\n<blockquote>\n  <p>python random_crop_images_hongyuan.py  I\n  tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA\n  library libcublas.so.8.0 locally I\n  tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA\n  library libcudnn.so.5 locally I\n  tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA\n  library libcufft.so.8.0 locally I\n  tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA\n  library libcuda.so.1 locally I\n  tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA\n  library libcurand.so.8.0 locally W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use SSE3 instructions, but these are\n  available on your machine and could speed up CPU computations. W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use SSE4.1 instructions, but these are\n  available on your machine and could speed up CPU computations. W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use SSE4.2 instructions, but these are\n  available on your machine and could speed up CPU computations. W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use AVX instructions, but these are\n  available on your machine and could speed up CPU computations. W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use AVX2 instructions, but these are\n  available on your machine and could speed up CPU computations. W\n  tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use FMA instructions, but these are\n  available on your machine and could speed up CPU computations. I\n  tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:910] successful\n  NUMA node read from SysFS had negative value (-1), but there must be\n  at least one NUMA node, so returning NUMA node zero I\n  tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:885] Found device 0\n  with properties:  name: GeForce GT 730M major: 3 minor: 5\n  memoryClockRate (GHz) 0.758 pciBusID 0000:01:00.0 Total memory:\n  982.88MiB Free memory: 592.44MiB I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:906] DMA: 0  I\n  tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:916] 0:   Y  I\n  tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:975] Creating\n  TensorFlow device (\/gpu:0) -> (device: 0, name: GeForce GT 730M, pci\n  bus id: 0000:01:00.0) Done image 3000 th in 317 ms <\/p>\n  \n  <p>I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:975] Creating\n  TensorFlow device (\/gpu:0) -> (device: 0, name: GeForce GT 730M, pci\n  bus id: 0000:01:00.0) Done image 3001 th in 325 ms <\/p>\n  \n  <p>I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:975] Creating\n  TensorFlow device (\/gpu:0) -> (device: 0, name: GeForce GT 730M, pci\n  bus id: 0000:01:00.0) Done image 3002 th in 312 ms <\/p>\n  \n  <p>I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:975] Creating\n  TensorFlow device (\/gpu:0) -> (device: 0, name: GeForce GT 730M, pci\n  bus id: 0000:01:00.0) Done image 3003 th in 147 ms <\/p>\n  \n  <p>I tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:975] Creating\n  TensorFlow device (\/gpu:0) -> (device: 0, name: GeForce GT 730M, pci\n  bus id: 0000:01:00.0) Done image 3004 th in 447 ms<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42359918","CreationDate":"2017-02-21T06:10:56.957","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It's not possible to be 100% certain without seeing all of your code, but I would guess that the <code>crop_image()<\/code> function is calling various TensorFlow op functions to build a graph.<\/p>\n\n<p>It is almost never a good idea to build a graph inside a <code>for<\/code> loop. <a href=\"https:\/\/stackoverflow.com\/a\/35705890\/3574081\">This answer<\/a> explains why: some operations (such as the first <code>Session.run()<\/code> call to a new operation) take time that is linear in the number of operations in the graph. If you add more operations in each iteration, iteration <code>i<\/code> will do work that is linear in <code>i<\/code>, and so the overall execution time will be quadratic.<\/p>\n\n<p>The modified version of your code (with a <code>with tf.Graph().as_default():<\/code> block inside the loop) will be faster because it creates a new, empty <code>tf.Graph<\/code> in each iteration, and therefore each iteration does a constant amount of work.<\/p>\n\n<p>An even more efficient solution would be to build the graph and session once, using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a> tensors to represent the filename and bbox arguments to <code>crop_image<\/code>, and feeding different values to these placeholders in each iteration.<\/p>\n"},{"AnswerId":"42342152","CreationDate":"2017-02-20T10:30:35.800","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>My guess is that this happens because creating the session is an expensive operation. May be it could also happen that the session is not properly cleaned when the with-statement is left, so each new allocation on the device will have less resources available. In short, I would not recommend doing it this way, rather initialize just one session and try to reuse it.<\/p>\n\n<p>EDIT:\nIn answer to your comment: The session is closed automatically as soon as the with-block is exited. I've read in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/1578\" rel=\"nofollow noreferrer\">this github issue<\/a> that the memory on the GPU is only really released when the whole program exits. But I guess that when you allocate a new session after you closed the last one, Tensorflow will internally just re-use the previously allocated resources. So, in retrospective my answer is probably not very insightful. Sorry if I caused confusion.<\/p>\n"}]}
{"QuestionId":42338981,"AnswerCount":1,"Tags":"<python><tensorflow><dimension>","CreationDate":"2017-02-20T07:44:08.197","AcceptedAnswerId":42339182.0,"OwnerUserId":5538922.0,"Title":"It seems inconsistent the ways tensorflow allows me to specify variable length dimension","Body":"<p>I'm a novice to tensorflow. I was practicing coding with this <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/recurrent_network.py\" rel=\"nofollow noreferrer\">tutorial code<\/a>. Most of all the code made sense to me but at some points I got stuck.<\/p>\n\n<pre><\/pre>\n\n<p>With  function I had to specify variable length dimesion with . But with  I had to use , not . In documentation for the two functions, both of the pertaining arguments have the name . So I am feeling lost here. Do they really have different meanings? Or is it just a small design mistake of the tensorflow developers?<\/p>\n","answers":[{"AnswerId":"42339182","CreationDate":"2017-02-20T07:55:34.333","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>You can understand it this way: in a placeholder the value \"None\" indicates: \"could be any value\". Like in your case: you have a batch size that can be anything. \nIn a reshape function, -1 indicates \"whatever value is remaining to make this shape work\". In you case, your x gets the shape (batch*n_steps), as this is the shape x needs to be to fit the same data in a matrix. <\/p>\n\n<p>Interesting note: you CAN use multiple None values in a placeholder (to indicate any batch size, any width and height of an image)... But you can't use multiple -1 values in a reshape function!<\/p>\n"}]}
{"QuestionId":42339779,"AnswerCount":0,"Tags":"<keras><conv-neural-network><convergence>","CreationDate":"2017-02-20T08:36:50.607","AcceptedAnswerId":null,"OwnerUserId":7505030.0,"Title":"Rerun of Neural Network does not converge","Body":"<p>I have a Convolutional Neural Network for semantic segmentation, implemented in Keras with Tensorflow backend. The architecture is of type encoder-decoder, using convolutional layers and max-pooling in the encoding part and deconvolutional layers and upsampling in the decoding part. Two classes to separate: objects of interest and background. Last layer has a sigmoid activation, the rest have Relu. There is one batch normalization layer in the \"middle\" of the model. Loss function is an Intersection over Union score. Learning algorithm : stochastic gradient descent, with learning rate 0.001 and momentum 0.9 .<\/p>\n\n<p>I have trained the model and it achieved convergence with a quite good performance on the segmentation task. The trouble here, being a newbie in Neural Networks, I did not save any data regarding the learning progress : graphs with loss values, accuracy, etc. which I need to include in my report. Therefore, I tried to retrain the same network (nothing changed), but it does not converge (at all). <\/p>\n\n<p>What to do in this case ? I did not choose any specific initializations for the weights ( as I understood and later checked, the default in Keras is glorot_uniform ), so naturally each time I run it, they will be different. Is there something I can do ? Or am I at the mercy of chance to reproduce the initial results ? <\/p>\n\n<p>Any thoughts\/suggestions\/advices are most appreciated. Thanks !<\/p>\n","answers":[]}
{"QuestionId":42340668,"AnswerCount":4,"Tags":"<python-3.x><ubuntu><tensorflow>","CreationDate":"2017-02-20T09:23:11.143","AcceptedAnswerId":null,"OwnerUserId":7172549.0,"Title":"AttributeError: module 'tensorflow' has no attribute 'Graph'","Body":"<p>I am writing the following code in Jupyter notebook.<\/p>\n\n<pre><\/pre>\n\n<p>And I encounter the following error while executing this simple code.<\/p>\n\n<blockquote>\n  <p>AttributeError: module 'tensorflow' has no attribute 'Graph'<\/p>\n<\/blockquote>\n\n<p>If I execute the same from the console. It works. Any idea what's happening here.<\/p>\n\n<p><strong>More information:<\/strong>(test is my virtualenv)<\/p>\n\n<p>ipython, jupyter, python version:<\/p>\n\n<pre><\/pre>\n\n<p>Execute from Console:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"48600145","CreationDate":"2018-02-03T17:41:03.060","ParentId":null,"OwnerUserId":"6907129","Title":null,"Body":"<p>A simple solution worked for me.\nFrom <code>anaconda navigator<\/code> first go to environments and create a tensorflow environment like this image<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GLcHf.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/GLcHf.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I named it <code>tensorflow<\/code>. Then from the right select all from the drop down menu and search for tensorflow.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/wnpSD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wnpSD.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Select it and hit apply. It will download some additional packages along with tensorflow. When you are done, select tensorflow from home and you should be good to go. Here it works.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/prk2u.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/prk2u.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"55991769","CreationDate":"2019-05-05T12:08:15.437","ParentId":null,"OwnerUserId":"2876639","Title":null,"Body":"<p>Check if TensorFlow is installed in Anaconda\/Environments. If not, just search and install. Then try to run g=tf.Graph() again. It worked for me. <\/p>\n"},{"AnswerId":"42343607","CreationDate":"2017-02-20T11:39:41.137","ParentId":null,"OwnerUserId":"7172549","Title":null,"Body":"<p>I resolved this issue by restarting Jupyter Notebook kernel. It's weird.<\/p>\n"},{"AnswerId":"50491265","CreationDate":"2018-05-23T14:40:56.060","ParentId":null,"OwnerUserId":"2207137","Title":null,"Body":"<p>Make sure you haven't called any of your python files as 'tensorflow.py'<\/p>\n"}]}
{"QuestionId":42342022,"AnswerCount":1,"Tags":"<python-3.x><numpy><tensorflow>","CreationDate":"2017-02-20T10:25:11.307","AcceptedAnswerId":42344327.0,"OwnerUserId":5217799.0,"Title":"Get random gamma distribution in tensorflow like numpy.random.gamma","Body":"<p>Hi I am new to tensorflow and I am trying to generate random gamma distribution in tensorflow just like numpy.random.gamma<\/p>\n\n<p>My numpy code is :-<\/p>\n\n<pre><\/pre>\n\n<p>where  and <\/p>\n\n<p>My tensorflow code is :-<\/p>\n\n<pre><\/pre>\n\n<p>Is it a correct implementation? I believe I failed to understand the parameters of  became .<\/p>\n","answers":[{"AnswerId":"42344327","CreationDate":"2017-02-20T12:13:58.210","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>You are setting different shape parameters in your distribution, so it is expected that they differ.<\/p>\n\n<p>One thing to watch out for is that numpy has a \"scale\" parameter while TF has an \"inverse scale\" parameter. So one has to be inverted to get the same distribution.<\/p>\n\n<p>Jupyter notebook example with matching distributions:<\/p>\n\n<pre><code>%matplotlib inline\nimport tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nsize = (50000,)\nshape_parameter = 1.5\nscale_parameter = 0.5\nbins = np.linspace(-1, 5, 30)\n\nnp_res = np.random.gamma(shape=shape_parameter, scale=scale_parameter, size=size)\n\n# Note the 1\/scale_parameter here\n\ntf_op = tf.random_gamma(shape=size, alpha=shape_parameter, beta=1\/scale_parameter)\nwith tf.Session() as sess:\n    tf_res = sess.run(tf_op)\n\nplt.hist(tf_res, bins=bins, alpha=0.5);\nplt.hist(np_res, bins=bins, alpha=0.5);\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/u3v2e.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/u3v2e.png\" alt=\"Histogram plot of results\"><\/a><\/p>\n"}]}
{"QuestionId":42342235,"AnswerCount":0,"Tags":"<python><tensorflow><dataset><size>","CreationDate":"2017-02-20T10:34:16.773","AcceptedAnswerId":null,"OwnerUserId":7499201.0,"Title":"Size of Tensorflows translation model vs size of dataset","Body":"<p>I am working with the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\" rel=\"nofollow noreferrer\">translation model<\/a> created by tensorflow. My GPU has 12Gb memory, which restricts my model to ,  and . As my dataset is from <a href=\"http:\/\/cs.mcgill.ca\/~jpineau\/datasets\/ubuntu-corpus-1.0\/\" rel=\"nofollow noreferrer\">Ubuntu Dialoge Corpus<\/a>, I got plenty of data. The restrictions, however, makes it impossible to expand the model further and by training on all of the data, the model will probably never converge (underfitting).<\/p>\n\n<p>Because of the underfitting, I want to limit the training data by slicing away some of it. The question is; how much? Is there a guide line of how much training data a model of a certain size should be fed, in order to converge but avoid overfitting? How many pairs of x_data and y_data can a translation model possible learn from?<\/p>\n","answers":[]}
{"QuestionId":42343301,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2017-02-20T11:25:54.150","AcceptedAnswerId":null,"OwnerUserId":1045995.0,"Title":"Tensorflow: attention decoder","Body":"<p>In Tensorflow 1.0, the seq2seq API was largely changed, and is no longer compatible with previous seq2seq examples. In particular, I find attention decoders quite a bit more challenging to build: the old  function has been removed, instead the new API expects the user to provide the  a couple of different attention functions during training and predicting, which in turn rely on a  function.<\/p>\n\n<p>Has anybody got an example of how to build an attention decoder, providing only the inputs and the final encoder state?<\/p>\n","answers":[{"AnswerId":"50405079","CreationDate":"2018-05-18T06:29:59.413","ParentId":null,"OwnerUserId":"7208627","Title":null,"Body":"<p>This seq2seq with attention model can translate number pronunciations into digits, e.g. \"one hundred and twenty seven\" -> \"127\". See if it helps.<\/p>\n\n<p><a href=\"https:\/\/acv.aixon.co\/attention_word_to_number.html\" rel=\"nofollow noreferrer\">https:\/\/acv.aixon.co\/attention_word_to_number.html<\/a><\/p>\n"}]}
{"QuestionId":42343922,"AnswerCount":1,"Tags":"<deep-learning><caffe><pycaffe><deeplearning4j>","CreationDate":"2017-02-20T11:55:42.167","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"How can I improve the learning of FCN32 for semantic segmentation?","Body":"<p>I have trained <a href=\"https:\/\/github.com\/shelhamer\/fcn.berkeleyvision.org\" rel=\"nofollow noreferrer\">FCN32<\/a> for semantic segmentation from scratch for my data, and I got the following output:\n<a href=\"https:\/\/github.com\/shelhamer\/fcn.berkeleyvision.org\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NHvKi.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>As it can be seen, this is not a good learning curve showing an improper training on data. \n is as follows: <\/p>\n\n<pre><\/pre>\n\n<p>after changing the learning rate to 0.001, it became worse:\n<a href=\"https:\/\/i.stack.imgur.com\/wllM3.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/wllM3.png\" alt=\"enter image description here\"><\/a>\nI am wondering what can I do for improving the training? Thanks <\/p>\n","answers":[{"AnswerId":"42350562","CreationDate":"2017-02-20T17:13:23.930","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>You can try varying the learning rate. Good values are normally something between 0.1 and 0.0001.<\/p>\n"}]}
{"QuestionId":42344012,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><gradient-descent>","CreationDate":"2017-02-20T11:58:42.843","AcceptedAnswerId":42346265.0,"OwnerUserId":2736559.0,"Title":"How can I have multiple losses in a network in Caffe?","Body":"<p>If I define multiple loss layers in a network, will there be multiple back propagation happening from those ends to the beginning of the network? I mean, do they even work that way?<\/p>\n\n<p>Suppose I have something like this:<\/p>\n\n<pre><\/pre>\n\n<p>So basically suppose instead of having one classification\/loss per each network, we have several in between as well.<\/p>\n\n<p>And in case they work, how could I add all the losses together and make my final loss the summation of all previous losses? <\/p>\n","answers":[{"AnswerId":"42346265","CreationDate":"2017-02-20T13:47:07.883","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Caffe does this for you.<br>\nFurthermore, for each loss layer you have <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L322-L325\" rel=\"nofollow noreferrer\"><code>loss_weight<\/code><\/a> parameter that allows you to decide how influential this specific loss is with respect to all other losses in the net.<br>\nOverall, the loss minimized by caffe is the <em>weighted<\/em> sum of all loss layers in the model. <\/p>\n"}]}
{"QuestionId":42344090,"AnswerCount":4,"Tags":"<python><python-3.x><numpy><tensorflow><tensor>","CreationDate":"2017-02-20T12:02:56.803","AcceptedAnswerId":42344214.0,"OwnerUserId":5217799.0,"Title":"What is the alternative of numpy.newaxis in tensorflow?","Body":"<p>Hi I am new to tensorflow. I want to implement the following python code in tensorflow.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42344214","CreationDate":"2017-02-20T12:08:50.267","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>I think that would be <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/expand_dims\" rel=\"noreferrer\"><code>tf.expand_dims<\/code><\/a> -<\/p>\n\n<pre><code>tf.expand_dims(a, 1) # Or tf.expand_dims(a, -1)\n<\/code><\/pre>\n\n<p>Basically, we list the axis ID where this new axis is to be inserted and the trailing axes\/dims are <em>pushed-back<\/em>.<\/p>\n\n<p>From the linked docs, here's few examples of expanding dimensions -<\/p>\n\n<pre><code># 't' is a tensor of shape [2]\nshape(expand_dims(t, 0)) ==&gt; [1, 2]\nshape(expand_dims(t, 1)) ==&gt; [2, 1]\nshape(expand_dims(t, -1)) ==&gt; [2, 1]\n\n# 't2' is a tensor of shape [2, 3, 5]\nshape(expand_dims(t2, 0)) ==&gt; [1, 2, 3, 5]\nshape(expand_dims(t2, 2)) ==&gt; [2, 3, 1, 5]\nshape(expand_dims(t2, 3)) ==&gt; [2, 3, 5, 1]\n<\/code><\/pre>\n"},{"AnswerId":"47269417","CreationDate":"2017-11-13T16:46:05.573","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>If you're interested in exactly the same type(i.e. <code>None<\/code>) as in NumPy, then <code>tf.newaxis<\/code> is the exact alternative to <code>np.newaxis<\/code>.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code>In [71]: a1 = tf.constant([2,2], name=\"a1\")\n\nIn [72]: a1\nOut[72]: &lt;tf.Tensor 'a1_5:0' shape=(2,) dtype=int32&gt;\n\n# add a new dimension\nIn [73]: a1_new = a1[tf.newaxis, :]\n\nIn [74]: a1_new\nOut[74]: &lt;tf.Tensor 'strided_slice_5:0' shape=(1, 2) dtype=int32&gt;\n\n# add one more dimension\nIn [75]: a1_new = a1[tf.newaxis, :, tf.newaxis]\n\nIn [76]: a1_new\nOut[76]: &lt;tf.Tensor 'strided_slice_6:0' shape=(1, 2, 1) dtype=int32&gt;\n<\/code><\/pre>\n\n<p>This is exactly the same kind of operations that you do in NumPy. Just use it exactly at the same dimension where you want it to be increased.<\/p>\n"},{"AnswerId":"44787333","CreationDate":"2017-06-27T18:23:44.697","ParentId":null,"OwnerUserId":"1735003","Title":null,"Body":"<p>The corresponding command is <code>tf.newaxis<\/code> (or <code>None<\/code>, as in numpy). It does not have an entry on its own in tensorflow's documentation, but is briefly mentioned on the doc page of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/strided_slice\" rel=\"noreferrer\"><code>tf.stride_slice<\/code><\/a>.<\/p>\n\n<pre><code>x = tf.ones((10,10,10))\ny = x[:, tf.newaxis] # or y = x [:, None]\nprint(y.shape)\n# prints (10, 1, 10, 10)\n<\/code><\/pre>\n\n<p>Using <code>tf.expand_dims<\/code> is fine too but, as stated in the link above,<\/p>\n\n<blockquote>\n  <p>Those interfaces are much more friendly, and highly recommended.<\/p>\n<\/blockquote>\n"},{"AnswerId":"55669785","CreationDate":"2019-04-13T20:54:09.203","ParentId":null,"OwnerUserId":"4960855","Title":null,"Body":"<p>Consider <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/keras\/layers\/Reshape\" rel=\"nofollow noreferrer\">tf.keras.layers.Reshape<\/a>:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># as first layer in a Sequential model\nmodel = Sequential()\nmodel.add(Reshape((3, 4), input_shape=(12,)))\n# now: model.output_shape == (None, 3, 4)\n# note: `None` is the batch dimension\n\n# as intermediate layer in a Sequential model\nmodel.add(Reshape((6, 2)))\n# now: model.output_shape == (None, 6, 2)\n\n# also supports shape inference using `-1` as dimension\nmodel.add(Reshape((-1, 2, 2)))\n# now: model.output_shape == (None, 3, 2, 2)\n<\/code><\/pre>\n"}]}
{"QuestionId":42344457,"AnswerCount":1,"Tags":"<tensorflow><convolution>","CreationDate":"2017-02-20T12:20:15.910","AcceptedAnswerId":42371656.0,"OwnerUserId":4126652.0,"Title":"Tensorflow Row Wise 1D Convolution","Body":"<p>I was going through this paper: <a href=\"http:\/\/www.aclweb.org\/anthology\/P14-1062\" rel=\"nofollow noreferrer\">CNN for Sentence Modelling<\/a><\/p>\n\n<p>The paper seems to describe a convolution along the following lines:<\/p>\n\n<p>With a matrix input <strong>I<\/strong> of size [d,m] and a weight matrix <strong>W<\/strong> of size [d,s]<\/p>\n\n<p>The convolution is done like each row of Input matrix is convoluted with each row of the weight matrix, (basically a row-wise 1D convolution).\nThe output will be of size [d,m] with appropriate padding or [d,m-s+1] without padding.<\/p>\n\n<p>My understanding is that, The only difference between this and simply doing a 1D convolution is that each row of the input matrix will have it's own set of weights to convolve with.<\/p>\n\n<p>Is it possible to achieve such a convolution layer in Tensorflow?<\/p>\n","answers":[{"AnswerId":"42371656","CreationDate":"2017-02-21T15:37:08.043","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>You can do this.<\/p>\n\n<pre><code>I = tf.reshape(I, [1,d,1,m])\nW = tf.reshape(W, [1,d,s,1)\nstrides = [1,stride,1,1]\noutput = tf.nn.depthwise_conv2d(I, W, strides, padding='SAME')\n<\/code><\/pre>\n\n<p>I think this is what you're looking for.  It will apply each filter row wise across the input and return back the convolved result.<\/p>\n"}]}
{"QuestionId":42344795,"AnswerCount":1,"Tags":"<python><neural-network><gpu><keras><autoencoder>","CreationDate":"2017-02-20T12:36:37.630","AcceptedAnswerId":42358827.0,"OwnerUserId":7561641.0,"Title":"I am trying to run autoencoder_layers.py using keras on gpu but i get this error","Body":"<p>autoencoder_layers.py <a href=\"https:\/\/github.com\/nanopony\/keras-convautoencoder\/blob\/master\/autoencoder_layers.py\" rel=\"nofollow noreferrer\">github code<\/a><\/p>\n\n<pre><\/pre>\n\n<p>but I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>how can I fix it?<\/p>\n","answers":[{"AnswerId":"42358827","CreationDate":"2017-02-21T04:40:50.930","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Comment the line <code>from keras.backend.theano_backend import _on_gpu<\/code> and define <code>_on_gpu<\/code> as: <\/p>\n\n<pre><code>def _on_gpu():\n    return theano.config.device[:3] == 'gpu'\n<\/code><\/pre>\n\n<p>It will fix your error.<\/p>\n"}]}
{"QuestionId":42344936,"AnswerCount":1,"Tags":"<python-3.x><neural-network><deep-learning><caffe><pycaffe>","CreationDate":"2017-02-20T12:44:24.427","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"How to solve this issue in Python (creating weights for Infogain Loss layer)?","Body":"<p>I am working on semantic segmentation using CNNs. I have an imbalance number of pixels for each class. \nBased on this <a href=\"https:\/\/stackoverflow.com\/questions\/30486033\/tackling-class-imbalance-scaling-contribution-to-loss-and-sgd\/30497907#30497907\">link<\/a>, I am trying to create weight matrix  in order to define  for my imbalance class members.\nMy data has five classes. I wrote the following code in python:\nReads a sample image:<\/p>\n\n<pre><\/pre>\n\n<p>Counts the number of pixels of each class<\/p>\n\n<pre><\/pre>\n\n<p>output:\n 39817   13751     1091   10460  417<\/p>\n\n<pre><\/pre>\n\n<p>The output, after removing some unimportant lines, is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>As it can be seen, it is giving an error. My question can be folded into two parts:<\/p>\n\n<ol>\n<li><p>How to solve this error?\nI replaced  with  as follows:<\/p>\n\n<p><\/p><\/li>\n<\/ol>\n\n<p>Now, it is not giving error and last line shows this:<\/p>\n\n<pre><\/pre>\n\n<p>It created the file , Is this correct?<\/p>\n\n<ol start=\"2\">\n<li>Is this matrix H should be constant for the all images in database?<\/li>\n<\/ol>\n\n<p>I really appreciate any help. <\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42345358","CreationDate":"2017-02-20T13:04:26.973","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<ol>\n<li><p>You have a simple \"copy-paste\" bug. You copied your code from <a href=\"https:\/\/stackoverflow.com\/a\/27645934\/1714410\">this answer<\/a> where <code>L<\/code> was an integer representing the number of classes. In your code, on the other hand, <code>L<\/code> is a <em>list<\/em> with the class weights. replacing <code>L<\/code> with 5 in your code does indeed solves the problem.<\/p><\/li>\n<li><p>Should <code>H<\/code> be constant? This is really up to you to decide.<\/p><\/li>\n<\/ol>\n\n<p>BTW, AFAIK, current caffe version does not support pixel-wise infogain loss, you might need to use the code in <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/3855\" rel=\"nofollow noreferrer\">PR #3855<\/a>.<\/p>\n"}]}
{"QuestionId":42345015,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-02-20T12:48:24.730","AcceptedAnswerId":42534570.0,"OwnerUserId":7592380.0,"Title":"How can I make the inception-v3 model pre-trained from Imagenet (classify_image.py) in the Tensorflow tutorial importable as a module?","Body":"<p>I wonder how I can modify classify_image.py (from <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_recognition#usage_with_python_api\" rel=\"nofollow noreferrer\">this tutorial<\/a> so that I can import it from another python script. I would basically like it to have the same functionality it already has, but instead of providing the image path and getting the response printed out in the terminal, I would like to give a function the image path and get the function to return the top 5 results with their probabilities.<\/p>\n\n<p>I haven\u2019t found a direct solution to this problem yet, but I realize my problem solving and search for previous answers are limited since I unfortunately haven\u2019t learned the basics of Tensorflow yet. <\/p>\n\n<p>Of course, if there is another pre-trained Tensorflow model that is just as good and meets my demands I would happily use that instead.<\/p>\n\n<p>Regards,\nPontus<\/p>\n\n<p><strong>UPDATE<\/strong> Maybe I should clearify a little:<\/p>\n\n<p>I don't want to train a model, just use a pre-trained one for image recognition, and in this case have an image recognition script that I could import as a module in another python application.<\/p>\n\n<p>I have also tried with code from <a href=\"https:\/\/stackoverflow.com\/questions\/39582703\/using-pre-trained-inception-resnet-v2-with-tensorflow\">this tutorial<\/a> but I got stuck there too, and in that case it includes a lot of manual installation where I might have failed in some step. The good thing with the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_recognition#usage_with_python_api\" rel=\"nofollow noreferrer\">classify_image.py example<\/a> is that I got it to work as intended to in the tutorial so I thought that the step from that to using it as a pluggable module shouldn't be that big.<\/p>\n\n<p>What I have tried (with classify_image.py) is moving the lines beneath  to  in order for them to get executed when I call them from another script but I am still having problems. I am mainly having problems with the  function, that wants me to pass it an argument, and from searching around I figured  seems to be some kind of placeholder used when getting input from the cli. All the FLAGS stuff seems to be cli related too, which is what I want to move away from. I am also unsure whether the model weights etc get saved correctly for me to be able to use it from another script. Again, at this point I just want to play around with the image classifier and further on hopefully learn more about the machine learning behind it. Sorry for my lack of knowledge in the basics of this!<\/p>\n\n<p>classify_image.py:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42347450","CreationDate":"2017-02-20T14:43:21.870","ParentId":null,"OwnerUserId":"3994228","Title":null,"Body":"<p>1) First question is about how to return the predicted values. \nFollowing code snippet does predition on the given image: <\/p>\n\n<pre><code>    top_k = predictions.argsort()[-FLAGS.num_top_predictions:][::-1]\n    for node_id in top_k:\n      human_string = node_lookup.id_to_string(node_id)\n      score = predictions[node_id]\n      print('%s (score = %.5f)' % (human_string, score))\n<\/code><\/pre>\n\n<p>Instead of print you can save the result in some data structure and return. By default, 5 top predictions will be returned, if you would like to change this behavior set proper value to <code>--num_top_predictions<\/code>.<\/p>\n\n<p>2) <strong>Regarding Model<\/strong>:\nThere are two parts to it- <\/p>\n\n<ol>\n<li>You need to have quality dataset as Imagenet is.<\/li>\n<li>Assuming if you have such quality dataset, the infrastructure to train inception would require very powerful GPUs. Lot of time too.<\/li>\n<\/ol>\n\n<p>but if you still want to train your system with your own dataset, I would say initially train with imagenet and later train the final layer(tensor name is '<strong>final_result<\/strong>') with your own dataset. Please find this <a href=\"https:\/\/petewarden.com\/2016\/09\/27\/tensorflow-for-mobile-poets\/\" rel=\"nofollow noreferrer\">tutorial<\/a>.  <\/p>\n"},{"AnswerId":"42534570","CreationDate":"2017-03-01T14:39:52.580","ParentId":null,"OwnerUserId":"7592380","Title":null,"Body":"<p>In the end I managed to use the code from the SO article reffered to in the update in the original question. I modified the code with the additional <code>im = 2*(im\/255.0)-1.0<\/code> from the answer of said SO question, some line to fix PIL on my computer plus a function to convert classes to human readable labels (found on github), link to that file below. I made it a callable function that takes a list of images as input and outputs a list of labels and predict values. If you'd like to use it, this is what you have to to:<\/p>\n\n<ol>\n<li>Install the latest Tensorflow version (1.0 at the moment, which is needed).<\/li>\n<li><code>git clone https:\/\/github.com\/tensorflow\/models\/<\/code>where you want the models.<\/li>\n<li>Put <a href=\"http:\/\/download.tensorflow.org\/models\/inception_resnet_v2_2016_08_30.tar.gz\" rel=\"nofollow noreferrer\">this checkpoint file<\/a> from the SO question I referred to earlier (needs to be extracted, of course) in the directory of your project.<\/li>\n<li>Put <a href=\"https:\/\/gist.github.com\/yrevar\/942d3a0ac09ec9e5eb3a\" rel=\"nofollow noreferrer\">this text file<\/a> (the human readable labels) in the directory of your project.<\/li>\n<li><p>Use this code from the SO question with some modifications from my side, put it in a .py file in your project:<\/p>\n\n<pre><code>import tensorflow as tf\nslim = tf.contrib.slim\nimport PIL as pillow\nfrom PIL import Image\n#import Image\nfrom inception_resnet_v2 import *\nimport numpy as np\n\nwith open('imagenet1000_clsid_to_human.txt','r') as inf:\n    imagenet_classes = eval(inf.read())\n\ndef get_human_readable(id):\n    id = id - 1\n    label = imagenet_classes[id]\n\n    return label\n\ncheckpoint_file = '.\/inception_resnet_v2_2016_08_30.ckpt'\n\n#Load the model\nsess = tf.Session()\narg_scope = inception_resnet_v2_arg_scope()\ninput_tensor = tf.placeholder(tf.float32, [None, 299, 299, 3])  \nwith slim.arg_scope(arg_scope):\n    logits, end_points = inception_resnet_v2(input_tensor, is_training=False)\nsaver = tf.train.Saver()\nsaver.restore(sess, checkpoint_file)\n\ndef classify_image(sample_images):\n    classifications = []\n    for image in sample_images:\n        im = Image.open(image).resize((299,299))\n        im = np.array(im)\n        im = im.reshape(-1,299,299,3)\n        im = 2*(im\/255.0)-1.0\n        predict_values, logit_values = sess.run([end_points['Predictions'], logits], feed_dict={input_tensor: im})\n        #print (np.max(predict_values), np.max(logit_values))\n        #print (np.argmax(predict_values), np.argmax(logit_values))\n        label = get_human_readable(np.argmax(predict_values))\n        predict_value = np.max(predict_values)\n        classifications.append({\"label\":label, \"predict_value\":predict_value})\n\n    return classifications\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"47775918","CreationDate":"2017-12-12T15:16:49.353","ParentId":null,"OwnerUserId":"6320806","Title":null,"Body":"<p>In my case just replace <code>[-FLAGS.num_top_predictions:]<\/code> with <code>[-5:]<\/code><\/p>\n\n<p>Then replace other FLAG with directory and file the image sit on.<\/p>\n"}]}
{"QuestionId":42345959,"AnswerCount":0,"Tags":"<python><torch>","CreationDate":"2017-02-20T13:34:07.253","AcceptedAnswerId":null,"OwnerUserId":7302262.0,"Title":"access model using dir","Body":"<p>I have used <a href=\"https:\/\/github.com\/bshillingford\/python-torchfile\" rel=\"nofollow noreferrer\">This link<\/a> to load <a href=\"https:\/\/github.com\/satoshiiizuka\/siggraph2016_colorization\" rel=\"nofollow noreferrer\">Torch Model<\/a> and access it to get it's layers trained weights, i tried many times but i can't use <strong>dir<\/strong> command to access it's <strong>layers weights<\/strong>, so what am i missing ? is that the wrong way to access them ? can any one help me, and this is the code and it's output.<\/p>\n\n<pre><\/pre>\n\n<p><strong>output<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42346075,"AnswerCount":0,"Tags":"<python><neural-network><theano><conv-neural-network><theano.scan>","CreationDate":"2017-02-20T13:39:13.227","AcceptedAnswerId":null,"OwnerUserId":2444293.0,"Title":"Nonlinear convolution in Theano","Body":"<p>I'm trying to make a \"custom\" convolutional layer in , where instead of linear filters convoluted with an image, I'm applying sup- or inf- convolutions (essentially dilations and erosions). How would I go in writing such a convolution in an efficient way?<\/p>\n\n<p>Given a tensor  intended to contain a set of vectors as inputs, its dilation with a filter  can be written as<\/p>\n\n<pre><\/pre>\n\n<p>The problem is that I don't know how to properly apply this operation to subregions of an image taking into account padding, tensor slicing etc. for filters of arbitrary size. I also read in the documentation that using  to implement convolutions is pretty inefficient. Any idea on how to do this?<\/p>\n","answers":[]}
{"QuestionId":42346516,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-02-20T13:57:43.373","AcceptedAnswerId":null,"OwnerUserId":3218279.0,"Title":"Error in LSTM during testing","Body":"<p>My data is of 68871 x 43, where the features are in the column no. 1-42 and label in column no. 43<\/p>\n\n<p>My keras LSTM code for classification of the data is <\/p>\n\n<pre><\/pre>\n\n<p>I am getting this error during testing time <a href=\"https:\/\/i.stack.imgur.com\/VjSUG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/VjSUG.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Please help resolve this, many thanks in advance<\/p>\n","answers":[{"AnswerId":"42346782","CreationDate":"2017-02-20T14:09:20.870","ParentId":null,"OwnerUserId":"5903959","Title":null,"Body":"<p>I think that your problem is that <code>model.evaluate(testX, testY)<\/code> is only returning one value.<\/p>\n\n<p>Your error message tells you that <code>numpy.float64<\/code> isn't iterable. What it means it that <code>model.evaluate(testX, testY)<\/code> returns a <code>float64<\/code> and thus, you cannot put it return value into two variables <code>score, acc<\/code>.<\/p>\n\n<p>It would be like doing:<\/p>\n\n<pre><code>def single_return():\n    return np.float64(10)\na, b = single_return()\n<\/code><\/pre>\n\n<p>(Note that this code will raise the exact same error).<\/p>\n\n<p>I would then suggest, both to fix it now, but also as a quite good practice for the future to always return into a single variable, then to split. It make error message more clear, as only line having problem will be the affectation, not the <code>evaluation<\/code> one.<\/p>\n\n<p>Hope it helps.<br>\n<code>pltrdy<\/code><\/p>\n"}]}
{"QuestionId":42346898,"AnswerCount":2,"Tags":"<python><numpy><keras>","CreationDate":"2017-02-20T14:14:36.283","AcceptedAnswerId":42361960.0,"OwnerUserId":2182857.0,"Title":"Loading one dimensional data into a Dense layer in a sequential Keras model","Body":"<p>I have the results of a trained model, ending in a Flatten layer in numpy output files.\nI try to load them and use them as inputs of a Dense layer.<\/p>\n\n<pre><\/pre>\n\n<p>However I get the following error message:<\/p>\n\n<pre><\/pre>\n\n<p>My input dimensions are (16,1536) - 16 images for this limited trail run, 1536 features. <\/p>\n\n<pre><\/pre>\n\n<p>The dense layer should expect a one dimensional 1536 long array.<\/p>\n\n<pre><\/pre>\n\n<p>What should I do?\nMany thanks!<\/p>\n","answers":[{"AnswerId":"42361960","CreationDate":"2017-02-21T08:17:43.707","ParentId":null,"OwnerUserId":"2182857","Title":null,"Body":"<p>I have found my problem - I did not define the lables correctly. I have switched the model compilation to a sparse categorical crossentropy mode.<\/p>\n\n<p>my current code is <\/p>\n\n<pre><code>def train_top_model():\n    train_data = np.load(open('bottleneck_flat_features_train.npy'))\n    train_labels = np.array([0] * (nb_train_samples \/ 2) + [1] * (nb_train_samples \/ 2))\n#\n    validation_data = np.load(open('bottleneck_flat_features_validation.npy'))\n    validation_labels = np.array([0] * (nb_validation_samples \/ 2) + [1] * (nb_validation_samples \/ 2))\n#\n    top_m  = Sequential()\n    top_m.add(Dense(2,input_shape=train_data.shape[1:], activation='softmax', name='top_dense1'))\n    top_m.compile(optimizer='rmsprop', loss='sparse_categorical_crossentropy', metrics=['accuracy'])\n#\n    top_m.fit(train_data, train_labels,\n    nb_epoch=nb_epoch, batch_size=my_batch_size,\n    validation_data=(validation_data, validation_labels))\n<\/code><\/pre>\n\n<p>Now it works and converges. <\/p>\n"},{"AnswerId":"42352827","CreationDate":"2017-02-20T19:31:36.033","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Don't you want to compile and fit top_m instead of model?<\/p>\n"}]}
{"QuestionId":42347678,"AnswerCount":1,"Tags":"<nlp><deep-learning><keras><sentiment-analysis>","CreationDate":"2017-02-20T14:55:05.737","AcceptedAnswerId":42658120.0,"OwnerUserId":4590360.0,"Title":"Error in implementing Aspect-based sentiment analysis deep learning model","Body":"<p>I have a task of  where I have to first predict aspect for each sentence. Aspects are pre-defined &amp; they are in total 19.     <\/p>\n\n<p>I have to implement a  (for the above task) where the first layer is fully connected &amp; second layer outputs a softmax distribution.<br>\nEach sentence is represented by an average of the word vectors. Word vector model which is used is  [ <a href=\"https:\/\/drive.google.com\/file\/d\/0B7XkCwpI5KDYNlNUTTlSS21pQmM\/edit?usp=sharing\" rel=\"nofollow noreferrer\">https:\/\/drive.google.com\/file\/d\/0B7XkCwpI5KDYNlNUTTlSS21pQmM\/edit?usp=sharing<\/a> ].<br>\nSo for each sentence I have a 300 dimensional word vector (average of all the 300-dim word vectors of all the words in the sentence).<br>\nThe output  is defined as  when the sentence has aspect i &amp; total of k aspects<br>\notherwise <br>\n(Note that - A sentence can have multiple aspects)    <\/p>\n\n<p>To implement above model -<br>\nI have in total 1739 sentences &amp; for each of them I have a  word vector. So I have a   of shape . Similarly  is  of shape .       <\/p>\n\n<p>My code using keras is -      <\/p>\n\n<pre><\/pre>\n\n<p>This code throws me error -<br>\n          <\/p>\n\n<p>I am new to deep learning &amp; keras, so if anybody can throw some light on what I am doing wrong?<\/p>\n","answers":[{"AnswerId":"42658120","CreationDate":"2017-03-07T20:53:06.797","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The problem lies in a wrong dimensionality of your data. It should have shape <code>(1739, 300)<\/code>. Recheck your data preparation as an error probably occured there.<\/p>\n"}]}
{"QuestionId":42347725,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-20T14:57:02.767","AcceptedAnswerId":null,"OwnerUserId":4724265.0,"Title":"Tensorflow: TypeError: run() got an unexpected keyword argument 'feedis_dict'","Body":"<p>I get this error when trying to add the  to a  command:<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestions as to why  is unexpected ?<\/p>\n\n<p>I believe im using tensorflow gpu 1.0.0<\/p>\n","answers":[]}
{"QuestionId":42349903,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-20T16:41:52.490","AcceptedAnswerId":42493744.0,"OwnerUserId":7302262.0,"Title":"Combine layer chrominance with image luminance","Body":"<p>I have read this <a href=\"http:\/\/hi.cs.waseda.ac.jp\/~iizuka\/projects\/colorization\/data\/colorization_sig2016.pdf\" rel=\"nofollow noreferrer\">Colorization paper<\/a> and it said:<\/p>\n\n<blockquote>\n  <p>The output layer of the colorization network consists of a\n  convolutional layer with a Sigmoid transfer function that outputs the\n  chrominance of the input grayscale image.<\/p>\n<\/blockquote>\n\n<p>and in order to get the colored image they said:<\/p>\n\n<blockquote>\n  <p>the computed chrominance is combined with the input intensity image to\n  produce the resulting color image.<\/p>\n<\/blockquote>\n\n<p>So I have implemented it and get the output layer with depth two, but how can I get the color image? How can I combine the greyscale image luminance values with the output layer of depth 2 (a*b colors) to get the final image?<\/p>\n\n<p>I use tensorflow and python.<\/p>\n","answers":[{"AnswerId":"42493744","CreationDate":"2017-02-27T19:02:35.997","ParentId":null,"OwnerUserId":"7302262","Title":null,"Body":"<p>Ok, I tried to implement it by using <strong>Skimage<\/strong> library to make a tensor of image chrominance values and compine it with the luminance by the same method.<\/p>\n"}]}
{"QuestionId":42349931,"AnswerCount":1,"Tags":"<python><tensorflow><artificial-intelligence>","CreationDate":"2017-02-20T16:43:17.903","AcceptedAnswerId":42351403.0,"OwnerUserId":4724265.0,"Title":"Tensorflow ValueError:Cannot feed value of shape (40, 24, 24, 4) for Tensor u'real_images:0', which has shape '(40, 24, 24, 3)'","Body":"<p>When trying to implement a DCGAN i get this error message when trying to use my training function:<\/p>\n\n<pre><\/pre>\n\n<p>This error occurs when trying to use the line:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas as to why the tensors are different ? Below i have attached the code where i believe the error lies:<\/p>\n\n<p>real_images placeholder (self.colour_dim is 3):<\/p>\n\n<pre><\/pre>\n\n<p>Segment where the error occurs:<\/p>\n\n<pre><\/pre>\n\n<p>The getImage and imread functions used:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you all!<\/p>\n","answers":[{"AnswerId":"42351403","CreationDate":"2017-02-20T18:00:07.793","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>It looks like your image loading function is giving you RGBA images, while the network expects RGB images. Replacing <code>batch_images<\/code> with <code>batch_images[:,:,:,:3]<\/code> in the feed dict should be an easy hotfix, you can also look if your loading function supports giving you an RGB image directly.<\/p>\n"}]}
{"QuestionId":42350009,"AnswerCount":4,"Tags":"<java><windows><tensorflow>","CreationDate":"2017-02-20T16:46:41.047","AcceptedAnswerId":null,"OwnerUserId":1295822.0,"Title":"Native library for TensorFlow java on windows","Body":"<p>Now that there is experimental support for Java on TensorFlow, I wanted to get started on Windows, following the steps in the readme<\/p>\n\n<p><strong><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/README.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/README.md<\/a><\/strong><\/p>\n\n<p>I got the expected error:<\/p>\n\n<pre><\/pre>\n\n<p>The Python version is running fine, but that is not Java....<\/p>\n\n<p>The readme points to native libraries for Mac and Linux, can someone confirm that I just have to wait a bit ?<\/p>\n","answers":[{"AnswerId":"45527856","CreationDate":"2017-08-06T01:41:10.367","ParentId":null,"OwnerUserId":"3491759","Title":null,"Body":"<p>This works for me when I added the native i.e tensorflow_jni.dll on PC or libtensorflow_jni.dylib on Mac(this should be the same for Linux distributions, not tested though) to my project root folder as seen here <a href=\"https:\/\/stackoverflow.com\/questions\/43711054\/adding-tensorflow-to-java-project-netbeans\">Adding Tensorflow to Java Project Netbeans<\/a>. Though I'm using Intellij IDEA<\/p>\n"},{"AnswerId":"44821058","CreationDate":"2017-06-29T09:19:21.073","ParentId":null,"OwnerUserId":"2289729","Title":null,"Body":"<p>Add the below code in Maven pom.xml<\/p>\n\n<pre><code> &lt;dependencies&gt;\n   &lt;dependency&gt;\n     &lt;groupId&gt;org.tensorflow&lt;\/groupId&gt;\n     &lt;artifactId&gt;tensorflow&lt;\/artifactId&gt;\n     &lt;version&gt;1.2.0&lt;\/version&gt;\n   &lt;\/dependency&gt;\n &lt;\/dependencies&gt;\n<\/code><\/pre>\n\n<p>You can find the installation of tensorflow for java detailed description in below link\n<a href=\"https:\/\/www.tensorflow.org\/install\/install_java\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_java<\/a><\/p>\n"},{"AnswerId":"50622531","CreationDate":"2018-05-31T10:48:16.440","ParentId":null,"OwnerUserId":"2413870","Title":null,"Body":"<p><em>java.lang.UnsatisfiedLinkError: tensorflow_native_libraries tensorflow_jni.dll: Can't find depedent libraries<\/em> happens because of some Missing DLLs in Windows which tensorflow_jni.dll cannot seem to find.<\/p>\n\n<ul>\n<li><p>Install Microsoft Visual C++ 2015 Resdistributable on Windows\nmachine <a href=\"https:\/\/www.microsoft.com\/en-us\/download\/details.aspx?id=52685\" rel=\"nofollow noreferrer\">Download here<\/a><\/p><\/li>\n<li><p>Try to close existing command prompts and then re-run the Tensorflow\nJava Program.<\/p><\/li>\n<\/ul>\n\n<p>See below link for more details<br>\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/14456\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/14456<\/a><\/p>\n"},{"AnswerId":"48991371","CreationDate":"2018-02-26T15:00:45.750","ParentId":null,"OwnerUserId":"1833961","Title":null,"Body":"<p>Not really an answer, but nevertheless I would like to warn you in advance.<\/p>\n\n<p>The support for Tensor Flow is very experimental indeed.\n(February 2018) - You are still very dependent of Python.<\/p>\n\n<p>Most importantly, there is <strong>no support for training networks in Java<\/strong>.\nYou can only apply\/run networks.<\/p>\n\n<p>(Not even to mention the very limited number of available classes).\nIf you do want to go ahead, I recommend you to <strong>take a look on Github<\/strong> because there are some <strong>utility classes<\/strong> there which will be more than useful for you. At least that will give you more comfort.<\/p>\n"}]}
{"QuestionId":42350299,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><keras>","CreationDate":"2017-02-20T17:00:26.757","AcceptedAnswerId":42350605.0,"OwnerUserId":288609.0,"Title":"regarding sparse tensor representation","Body":"<p>In the context of sparse tensor representation of Tensorflow, the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/sparse_ops\/sparse_tensor_representation\" rel=\"nofollow noreferrer\">document<\/a> gives the following definition and corresponding example. I can see why , but I am not very clear why  and . Thanks.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/D9Tv7.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/D9Tv7.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8J9Uo.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8J9Uo.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42350605","CreationDate":"2017-02-20T17:15:43.423","ParentId":null,"OwnerUserId":"5903959","Title":null,"Body":"<p>We usually use sparse vector\/matrix\/tensor when one value is way (way) more represented than the other one. Usually, this value is zero, as you see in the example.<\/p>\n\n<p>Now, the trick is to consider that tensor's values are 0 by default, then to specify which elements are not 0. <\/p>\n\n<p>You can consider <code>indices<\/code> as coordinates in an <code>ndims<\/code> space, and <code>values<\/code> the corresponding values.<\/p>\n\n<ul>\n<li><p><strong>Dense Tensor Philosophy<\/strong>: \"create a tensor like: <code>[[1, 0,0,0],[0,0,2,0],[0,0,0,0]]<\/code>\"<\/p><\/li>\n<li><p><strong>Sparse Tensor Philosophy<\/strong>: <\/p>\n\n<ul>\n<li>\"create a zero valued tensor having a [3,4] shape<\/li>\n<li>put <code>1<\/code> (=<code>values[0]<\/code>) at <code>[0,0]<\/code> (=<code>indices[0]<\/code>)<\/li>\n<li>then put <code>2<\/code> (=<code>values[1]<\/code>) at <code>[1,2]<\/code> (=<code>indices[1]<\/code>).<\/li>\n<\/ul><\/li>\n<\/ul>\n\n<p>For example, if you want to add a 3 a the bottom-right corner, you must set <code>indices=[[0,0],[1,2],[2, 4]]<\/code>, <code>values=[1,2,3]<\/code>.<\/p>\n\n<p>Got it?\nHope it helps<\/p>\n"}]}
{"QuestionId":42351184,"AnswerCount":5,"Tags":"<python><tensorflow><regression>","CreationDate":"2017-02-20T17:48:07.010","AcceptedAnswerId":42351397.0,"OwnerUserId":6112394.0,"Title":"How to Calculate R^2 in Tensorflow","Body":"<p>I am trying to do regression in Tensorflow. I'm not positive I am calculating R^2 correctly as Tensorflow gives me a different answer than  Can someone please look at my below code and let me know if I implemented the pictured equation correctly. Thanks<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/3vEI6.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/3vEI6.png\" alt=\"The formula I am attempting to create in TF\"><\/a><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"53796652","CreationDate":"2018-12-15T19:43:32.787","ParentId":null,"OwnerUserId":"805030","Title":null,"Body":"<p>All the other solutions wouldn't produce the right R squared score for multidimensional y. The right way to calculate R2 (variance weighted) in TensorFlow is:<\/p>\n\n<pre>\nunexplained_error = tf.reduce_sum(tf.square(labels - predictions))\ntotal_error = tf.reduce_sum(tf.square(labels - tf.reduce_mean(labels, axis=0)))\nR2 = 1. - tf.div(unexplained_error, total_error)\n<\/pre>\n\n<p>The result from this TF snippet matches exactly the result from sklearn's:<\/p>\n\n<pre>\nfrom sklearn.metrics import r2_score\nR2 = r2_score(labels, predictions, multioutput='variance_weighted')\n<\/pre>\n"},{"AnswerId":"51420297","CreationDate":"2018-07-19T10:28:24.300","ParentId":null,"OwnerUserId":"4306852","Title":null,"Body":"<p>The function is given <a href=\"https:\/\/www.kaggle.com\/rohumca\/linear-regression-in-tensorflow\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre><code>def R_squared(y, y_pred):\n  residual = tf.reduce_sum(tf.square(tf.subtract(y, y_pred)))\n  total = tf.reduce_sum(tf.square(tf.subtract(y, tf.reduce_mean(y))))\n  r2 = tf.subtract(1.0, tf.div(residual, total))\n  return r2\n<\/code><\/pre>\n\n<p>The concept is explained <a href=\"http:\/\/blog.minitab.com\/blog\/adventures-in-statistics-2\/regression-analysis-how-do-i-interpret-r-squared-and-assess-the-goodness-of-fit\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"42351397","CreationDate":"2017-02-20T17:59:55.003","ParentId":null,"OwnerUserId":"224671","Title":null,"Body":"<p>What you are computing the \"R^2\" is <\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/FgWsa.png\" alt=\"R^2_{\\text{wrong}} = \\operatorname{mean}_i \\left( \\frac{(y_i-\\hat y_i)^2}{(y_i-\\mu)^2} - 1\\right)[1]\"><\/p>\n\n<p>compared to the given expression, you are computing the mean at the wrong place. You should take the mean when computing the errors, before doing the division.<\/p>\n\n<pre><code>total_error = tf.reduce_sum(tf.square(tf.sub(y, tf.reduce_mean(y))))\nunexplained_error = tf.reduce_sum(tf.square(tf.sub(y, prediction)))\nR_squared = tf.sub(1, tf.div(unexplained_error, total_error))\n<\/code><\/pre>\n"},{"AnswerId":"46304041","CreationDate":"2017-09-19T15:13:43.027","ParentId":null,"OwnerUserId":"8634830","Title":null,"Body":"<p>It should actually be the opposite on the rhs. Unexplained variance divided by total variance<\/p>\n"},{"AnswerId":"47320697","CreationDate":"2017-11-16T02:43:56.147","ParentId":null,"OwnerUserId":"5836315","Title":null,"Body":"<p>I think it should be like this:<\/p>\n\n<pre><code>total_error = tf.reduce_sum(tf.square(tf.sub(y, tf.reduce_mean(y))))\nunexplained_error = tf.reduce_sum(tf.square(tf.sub(y, prediction)))\nR_squared = tf.sub(1, tf.div(unexplained_error, total_error))\n<\/code><\/pre>\n"}]}
{"QuestionId":42351638,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-20T18:14:47.497","AcceptedAnswerId":null,"OwnerUserId":7520802.0,"Title":"Tensorflow loss change when adding noise to input images","Body":"<p>I was wondering if it was typical to see huge spikes in the loss over time graph with models that are being fed images that get noise applied to them.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/PHK6a.png\" rel=\"nofollow noreferrer\">Image<\/a><\/p>\n\n<p>I was wondering if this is normal(look at the lines with DA which means data augmented), and if not what I should do to keep the loss from being so noisy. I'm guessing lowering the amount of noise I apply to the image, provided it keeps the same performance.<\/p>\n\n<p>Thanks<\/p>\n\n<p>edit: Whoops Just noticed I have train and valid losses flipped. Higher ones should be validation.<\/p>\n","answers":[{"AnswerId":"46936173","CreationDate":"2017-10-25T15:18:00.630","ParentId":null,"OwnerUserId":"786559","Title":null,"Body":"<p>Well, of course it is ! Especially if you do it 'online', so basically it never sees the same image twice.<\/p>\n\n<p>However, I am wondering why you regularly get those big spikes in the top blue line, which isn't DA . It is difficult to say without more insight into your data and problem<\/p>\n"}]}
{"QuestionId":42351645,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><deep-learning>","CreationDate":"2017-02-20T18:15:25.240","AcceptedAnswerId":42353659.0,"OwnerUserId":4014075.0,"Title":"TensorFlow + Batch Gradient Descent: 0% accuracy on each batch, but network converges?","Body":"<p>I'm currently very much a beginner with TensorFlow and Deep Learning in general, and I was trying to make a pretty simple 2-layer neural network with the  activation function for the hidden layer, and  for the output layer. In particular, I was training on the well-known notMNIST dataset, which has the exact same shape as MNIST but with more difficult examples. This was how I solved it (using TensorFlow ):<\/p>\n\n<pre><\/pre>\n\n<p>With a simple runner in this fashion:<\/p>\n\n<pre><\/pre>\n\n<p>I get these results:<\/p>\n\n<pre><\/pre>\n\n<p>As can be seen, the minibatch accuracy is <em>always<\/em> , but the minibatch loss is going down and the validation accuracy is going up. It seems the model \"works\", but I think that something else is going on that's indicative of a larger problem. The sudden jump after 500 epochs is also suspicious. Since I don't have much intuition for this, I tried various superficial things like altering the learning rate and batch size, but they didn't do anything to that perpetual 0% accuracy.<\/p>\n\n<p>It would be really appreciated if someone more experienced in TensorFlow can tell me what might be causing this, so I can learn to avoid it in the future.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42353659","CreationDate":"2017-02-20T20:28:58.900","ParentId":null,"OwnerUserId":"5951780","Title":null,"Body":"<p>Try <\/p>\n\n<pre><code>train_prediction = tf.nn.softmax(output)\n<\/code><\/pre>\n\n<p>instead of <\/p>\n\n<pre><code>train_prediction = tf.nn.softmax(hidden_layer_output)\n<\/code><\/pre>\n\n<p>and it should work.<\/p>\n\n<p>By the way: I wouldn't call logit what you called logits. Your outputs would be called logits, but that's just a naming issue...<\/p>\n"}]}
{"QuestionId":42351728,"AnswerCount":5,"Tags":"<windows><python-3.x><cmd><tensorflow>","CreationDate":"2017-02-20T18:21:18.903","AcceptedAnswerId":42463158.0,"OwnerUserId":4564238.0,"Title":"Tensorflow installation on Windows 10, error 'Not a supported wheel on this platform'","Body":"<p>This question is for a Windows 10 laptop. I'm currently trying to install tensorflow, however, when I run:<\/p>\n\n<p><\/p>\n\n<p>I get the following error:<\/p>\n\n<p><\/p>\n\n<p>I am trying to install the cpu-version only of tensorflow in an Anaconda 4.3.0 version. I had python 3.6.0 and then I downgraded to 3.5.0, none of them worked. <\/p>\n","answers":[{"AnswerId":"42395104","CreationDate":"2017-02-22T15:03:52.430","ParentId":null,"OwnerUserId":"3111214","Title":null,"Body":"<p>Inside your Anaconda environment, try running this:<\/p>\n\n<pre><code>pip install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>This will do the job. The issue was discussed <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7552\" rel=\"nofollow noreferrer\">here<\/a> also.<\/p>\n\n<p>Here is the screenshot of how this helped me:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lwyvK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lwyvK.png\" alt=\"pip install --upgrade tensorflow\"><\/a><\/p>\n"},{"AnswerId":"42352213","CreationDate":"2017-02-20T18:51:28.460","ParentId":null,"OwnerUserId":"4638103","Title":null,"Body":"<p>So are you sure you correctly downgraded your python? Run this command on command line <code>pip -V<\/code>. This should print the pip version and the python version.<\/p>\n"},{"AnswerId":"42463158","CreationDate":"2017-02-26T00:04:40.680","ParentId":null,"OwnerUserId":"4469065","Title":null,"Body":"<p>I also had same problem when I installed anaconda 4.3 version<\/p>\n\n<p>Here is my solution.<\/p>\n\n<ul>\n<li>Instead of using Anaconda3 4.3, install Anaconda3 4.2(Anaconda3-4.2.0-Windows-x86_64.exe)<\/li>\n<li><p>Type on command line(If you are using GPU version)<\/p>\n\n<pre><code>pip install -U --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre><\/li>\n<li><p>Typeon command line(If you are using CPU only)<\/p>\n\n<pre><code>pip install -U --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre><\/li>\n<\/ul>\n\n<p>I'm working on win10 with python version=3.5.2, 64 bit<\/p>\n"},{"AnswerId":"42565642","CreationDate":"2017-03-02T20:48:25.993","ParentId":null,"OwnerUserId":"3632894","Title":null,"Body":"<p>You can use same version of anaconda and execute this command<\/p>\n\n<pre><code>conda create -n tensorflow python=3.5\nactivate tensorflow\npip install tensorflow-gpu\n<\/code><\/pre>\n\n<p>It worked for <strong>conda 4.0.8<\/strong><\/p>\n"},{"AnswerId":"45117546","CreationDate":"2017-07-15T11:17:40.457","ParentId":null,"OwnerUserId":"8246394","Title":null,"Body":"<p>If you have python3 on Windows insatalled, you can use the following command(non GPU):<\/p>\n\n<p>pip install --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.10.0-py3-none-any.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/mac\/cpu\/tensorflow-0.10.0-py3-none-any.whl<\/a><\/p>\n\n<p>Worked for me.<\/p>\n"}]}
{"QuestionId":42351740,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-02-20T18:22:17.153","AcceptedAnswerId":46248975.0,"OwnerUserId":5436892.0,"Title":"Tensorflow, feeding Estimator.fit(batch)","Body":"<p>Could you provide an example of using the high-level API Estimators with placeholders and feeding batches  like for a basic use:<\/p>\n\n<pre><\/pre>\n\n<p>How to do the same with Estimator API?\nEstimator takes batch_size, steps, input_fuc or feed_fun as an argument of the fit function (see doc <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/contrib.learn\/estimators\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/contrib.learn\/estimators<\/a>)  but it is not clear for me how to implement a function which will load data as a batch from e.g. disk?<\/p>\n","answers":[{"AnswerId":"46248975","CreationDate":"2017-09-16T00:31:03.873","ParentId":null,"OwnerUserId":"2912797","Title":null,"Body":"<p>I don't think whether estimators are really meant to be used with placeholders. They use the concept of <code>input_fn<\/code> which is properly described <a href=\"https:\/\/www.tensorflow.org\/get_started\/input_fn\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>If you realy need to use a placeholder you might use a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/FeedFnHook\" rel=\"nofollow noreferrer\"><code>FeedFnHook<\/code><\/a>:<\/p>\n\n<pre><code>def input_fn():  # empty input_fn, returns features and labels\n    return {}, {}\n\nfeed_dict = {x:batch_of_inputs,y:batch_of_targets}\ndef feed_fn():  # feed_fn with hardcoded feed_dict\n    return feed_dict\n\nhooks = [tf.train.FeedFnHook(feed_fn=feed_fn)]\nestimator.train(input_fn=input_fn, hooks=hooks, steps=1)\n<\/code><\/pre>\n"}]}
{"QuestionId":42352593,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-20T19:15:35.443","AcceptedAnswerId":42354235.0,"OwnerUserId":3614410.0,"Title":"Tensorflow r1.0 control_flow_ops.with_dependencies() missing","Body":"<p>control_flow_ops.with_dependencies() seems to be missing in the 1.0 release, can't find anything in the changelog about it. Should we not use the function anymore or was it replaced by something else?<\/p>\n","answers":[{"AnswerId":"42354235","CreationDate":"2017-02-20T21:09:03.630","ParentId":null,"OwnerUserId":"5219154","Title":null,"Body":"<p>I think you are looking for <code>tf.control_dependencies()<\/code>: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/control_dependencies\" rel=\"nofollow noreferrer\">TF docs<\/a><\/p>\n"}]}
{"QuestionId":42352930,"AnswerCount":1,"Tags":"<python><tensorflow><nan>","CreationDate":"2017-02-20T19:38:01.883","AcceptedAnswerId":43103318.0,"OwnerUserId":2294865.0,"Title":"NaN result when summing two tensors which individually are both real numbers","Body":"<p>So I've been playing with this interesting application of tensorflow: <a href=\"http:\/\/kevinhughes.ca\/blog\/tensor-kart\" rel=\"nofollow noreferrer\">TensorKart<\/a><\/p>\n\n<p>The full source code is here: <a href=\"https:\/\/github.com\/kevinhughes27\/TensorKart\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kevinhughes27\/TensorKart<\/a><\/p>\n\n<p>I'm getting a NaN result in my loss calculation and it's totally stumped me. The problem lies in this line of code:<\/p>\n\n<pre><\/pre>\n\n<p>When I evaluate the first part:<\/p>\n\n<pre><\/pre>\n\n<p>I get values around 0.8\nWhen I evaluate the second part:<\/p>\n\n<pre><\/pre>\n\n<p>I get values around 3.1<\/p>\n\n<p>BUT when I sum the two parts I get NaN?!?!<\/p>\n\n<p>I've checked the shape and types and they both come back as: Tensor(\"Shape:0\", shape=(0,), dtype=int32)<\/p>\n\n<p>Things that I have tried: using tf.reduce_sum instead of tf.add_n, casting them to tf.float32, summing them with tf.add....<\/p>\n\n<p>Does anyone have any insights?<\/p>\n","answers":[{"AnswerId":"43103318","CreationDate":"2017-03-29T20:16:21.723","ParentId":null,"OwnerUserId":"2294865","Title":null,"Body":"<p>So after having gone through every operation leading up to the calculation of the loss value, it turns out there were NaN values in my input data. Converting them to zero resolved the issue of getting a NaN result in the final loss value calculation.<\/p>\n\n<p>What I haven't figured out till now is why I was only getting an error at the point I sum the two parts of the loss value but not while calculating the parts individually.<\/p>\n\n<p>Utterly bizarre.<\/p>\n"}]}
{"QuestionId":42353056,"AnswerCount":2,"Tags":"<python><numpy><neural-network><keras><recurrent-neural-network>","CreationDate":"2017-02-20T19:46:46.427","AcceptedAnswerId":43266817.0,"OwnerUserId":2751997.0,"Title":"Keras Masking for RNN with Varying Time Steps","Body":"<p>I'm trying to fit an RNN in Keras using sequences that have varying time lengths. My data is in a Numpy array with format  where  is determined at run-time as the number of time steps available for the sample with the most time stamps. I've padded the beginning of each time series with , except for the longest one, obviously.<\/p>\n\n<p>I've initially defined my model like so...<\/p>\n\n<pre><\/pre>\n\n<p>For completeness, here's the code for the loss function:<\/p>\n\n<pre><\/pre>\n\n<p>And here's the code for the custom activation function:<\/p>\n\n<pre><\/pre>\n\n<p>When I fit the model and make some test predictions, <em>every sample in the test set gets exactly the same prediction<\/em>, which seems fishy.<\/p>\n\n<p>Things get better if I remove the masking layer, which makes me think there's something wrong with the masking layer, but as far as I can tell, I've followed the documentation exactly.<\/p>\n\n<p>Is there something mis-specified with the masking layer? Am I missing something else?<\/p>\n","answers":[{"AnswerId":"43266817","CreationDate":"2017-04-06T22:34:59.927","ParentId":null,"OwnerUserId":"3110805","Title":null,"Body":"<p>I could not validate without actual data, but I had a similar experience with an RNN. In my case normalization solved the issue. Add a normalization layer to your model.<\/p>\n"},{"AnswerId":"43268654","CreationDate":"2017-04-07T02:22:57.620","ParentId":null,"OwnerUserId":"5896419","Title":null,"Body":"<p>The way you implemented masking should be correct. If you have data with the shape <strong>(samples, timesteps, features)<\/strong>, and you want to mask timesteps lacking data with a zero mask of the same size as the features argument, then you add <code>Masking(mask_value=0., input_shape=(timesteps, features))<\/code>. See here: <a href=\"https:\/\/keras.io\/layers\/core\/#masking\" rel=\"nofollow noreferrer\">keras.io\/layers\/core\/#masking<\/a><\/p>\n\n<p>Your model could potentially be too simple, and\/or your number of epochs could be insufficient for the model to differentiate between all of your classes. Try this model:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Masking(mask_value=0., input_shape=(max_time, 24)))\nmodel.add(LSTM(256, input_dim=24))\nmodel.add(Dense(1024))\nmodel.add(Dense(2))\nmodel.add(Activation(activate))\nmodel.compile(loss=weibull_loglik_discrete, optimizer=RMSprop(lr=.01))\nmodel.fit(train_x, train_y, nb_epoch=100, batch_size=1000, verbose=2, validation_data=(test_x, test_y)) \n<\/code><\/pre>\n\n<p>If that does not work, try doubling the epochs a few times (e.g. 200, 400) and see if that improves the results.<\/p>\n"}]}
{"QuestionId":42353188,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><theano><keras>","CreationDate":"2017-02-20T19:56:28.903","AcceptedAnswerId":42353316.0,"OwnerUserId":288609.0,"Title":"regarding transforming an ndarray(image input via cv2 or skimage) to a tensor","Body":"<p>I have read an image as follows using <\/p>\n\n<pre><\/pre>\n\n<p>This read image cause the error message when it was called by <\/p>\n\n<p>The error message is as <\/p>\n\n<p>Are there any mechanisms that can transform an image represented as ndarray to a Tensorflow tensor. Thanks.<\/p>\n","answers":[{"AnswerId":"42353316","CreationDate":"2017-02-20T20:04:06.363","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>You have to read up on the sess.run function. In the array you have as argument of your function you specify what you want to get OUT of your run command. In your case, you probably only want your pred and logits. <\/p>\n\n<p>If you want to put something IN the network you have to specify a tf.placeholder in your graph, and feed your image like this: <\/p>\n\n<p><code>np_pred,np_logits = sess.run([pred, logits],feed_dict={image_placeholder: image})<\/code><\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":42353400,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-02-20T20:10:12.663","AcceptedAnswerId":42377086.0,"OwnerUserId":6211276.0,"Title":"Visualizing the inputs to convolution neural networks using keras and tensorflow as backend","Body":"<p>There is an in-detailed post on <a href=\"https:\/\/blog.keras.io\/how-convolutional-neural-networks-see-the-world.html\" rel=\"nofollow noreferrer\">keras blog<\/a>.<br>\nBut when compiling the code I get the error as follows: <br><\/p>\n\n<pre><\/pre>\n\n<p>This error goes when I set .<br>\nBut as I am using tensorflow backend so dimension ordering should be .<br>\nEven after setting dim_ordering as 'th', I get error while  from  as follows : <br><\/p>\n\n<pre><\/pre>\n\n<p>As detailed in <a href=\"https:\/\/stackoverflow.com\/questions\/39547279\/loading-weights-in-th-format-when-keras-is-set-to-tf-format\">this post<\/a> about 'th' and 'tf'.<br> The above error implies layer weights are in 'tf' (but I set it to 'th' to avoid first error) and provided weight shape in 'th' ordering.<br>\n?<\/p>\n","answers":[{"AnswerId":"42377086","CreationDate":"2017-02-21T20:15:19.843","ParentId":null,"OwnerUserId":"6211276","Title":null,"Body":"<p>Answer to this question was pretty simple. As, I was using <code>tensorflow<\/code> as backend. So, to convert I inserted the line <br>\n<code>if K.backend()=='tensorflow':\n    K.set_image_dim_ordering(\"th\")<\/code><br>\nafter <code>from keras import backend as K<\/code>.<br>\nThis is because the <code>vgg16_weights.h5<\/code> has <code>th<\/code> format and also <code>cifar10.load_data()<\/code>.<\/p>\n"}]}
{"QuestionId":42353581,"AnswerCount":0,"Tags":"<tensorflow><syntaxnet><parsey-mcparseface>","CreationDate":"2017-02-20T20:23:22.140","AcceptedAnswerId":null,"OwnerUserId":804409.0,"Title":"syntaxnet POS tagger models","Body":"<p>I'm trying to make the (romanian) syntaxnet model run on a part of my corpus. I have the unzipped files mapped in a docker container. When i run the script for tokenizing:\n\nI'm getting a core dumped error related to a missing file, specifically char-map.\nLooking inside the unzipped folder, i notice that indeed, there is no char-map file there (only char-ngram-map). \nWhat am i missing here?\nIf the archive is indeed missing some files, then how did it get uploaded there ?<\/p>\n\n<p>Thanks<\/p>\n","answers":[]}
{"QuestionId":42353676,"AnswerCount":4,"Tags":"<python><numpy><matplotlib><tensorflow><mnist>","CreationDate":"2017-02-20T20:30:12.463","AcceptedAnswerId":42353867.0,"OwnerUserId":4333347.0,"Title":"Display MNIST image using matplotlib","Body":"<p>I am using tensorflow to import some MNIST input data. I followed this tutorial...<a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners<\/a><\/p>\n\n<p>I am importing them as so...<\/p>\n\n<pre><\/pre>\n\n<p>I want to be able to display any of the images from the training set. I know the location of the images is , so I try to access the first images and display it like so...<\/p>\n\n<pre><\/pre>\n\n<p>I a attempt to convert the image to a 28 by 28 numpy array because I know that each image is 28 by 28 pixels.<\/p>\n\n<p>However, when I run the code all I get is the following...<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/uhHFc.jpg\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/uhHFc.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Clearly I am doing something wrong. When I print out the matrix, everything seems to look good, but I think I am incorrectly reshaping it.<\/p>\n","answers":[{"AnswerId":"47343074","CreationDate":"2017-11-17T03:59:39.243","ParentId":null,"OwnerUserId":"4520029","Title":null,"Body":"<p>Here is the complete code for showing image using matplotlib<\/p>\n\n<pre><code>first_image = mnist.test.images[0]\nfirst_image = np.array(first_image, dtype='float')\npixels = first_image.reshape((28, 28))\nplt.imshow(pixels, cmap='gray')\nplt.show()\n<\/code><\/pre>\n"},{"AnswerId":"42353867","CreationDate":"2017-02-20T20:44:12.597","ParentId":null,"OwnerUserId":"893159","Title":null,"Body":"<p>You are casting an array of floats (<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/tutorials\/mnist\/beginners\/\" rel=\"nofollow noreferrer\">as described in the docs<\/a>) to <code>uint8<\/code>, which truncates them to 0, if they are not <code>1.0<\/code>. You should either round them or use them as floats or multiply with 255.<\/p>\n\n<p>I am not sure, why you don't see the white background, but i would suggest to use a well defined gray scale anyway.<\/p>\n"},{"AnswerId":"50200235","CreationDate":"2018-05-06T13:28:41.323","ParentId":null,"OwnerUserId":"9525493","Title":null,"Body":"<p>For those of you who want to do it with PIL.Image: <\/p>\n\n<pre><code>import numpy as np\nimport PIL.Image as pil\nfrom tensorflow.examples.tutorials.mnist import input_data\n\nmnist = input_data.read_data_sets('mnist')\n\ntestImage = (np.array(mnist.test.images[0], dtype='float')).reshape(28,28)\n\nimg = pil.fromarray(np.uint8(testImage * 255) , 'L')\nimg.show()\n<\/code><\/pre>\n"},{"AnswerId":"45565062","CreationDate":"2017-08-08T09:50:14.827","ParentId":null,"OwnerUserId":"8433267","Title":null,"Body":"<p>The following code shows example images displayed from the MNIST digit database used for training neural networks. It uses a variety of pieces of code from around stackflow and avoids pil.<\/p>\n\n<pre><code># Tested with Python 3.5.2 with tensorflow and matplotlib installed.\nfrom matplotlib import pyplot as plt\nimport numpy as np\nfrom tensorflow.examples.tutorials.mnist import input_data\nmnist = input_data.read_data_sets('MNIST_data', one_hot = True)\ndef gen_image(arr):\n    two_d = (np.reshape(arr, (28, 28)) * 255).astype(np.uint8)\n    plt.imshow(two_d, interpolation='nearest')\n    return plt\n\n# Get a batch of two random images and show in a pop-up window.\nbatch_xs, batch_ys = mnist.test.next_batch(2)\ngen_image(batch_xs[0]).show()\ngen_image(batch_xs[1]).show()\n<\/code><\/pre>\n\n<p>The definition of mnist is at: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py<\/a> <\/p>\n\n<p>The tensorflow neural network that led me to the need to display the MNINST images is at: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/examples\/tutorials\/mnist\/mnist_deep.py\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.2\/tensorflow\/examples\/tutorials\/mnist\/mnist_deep.py<\/a> <\/p>\n\n<p>Since I have only been programming Python for two hours, I might have made some newby errors. Please feel free to correct.<\/p>\n"}]}
{"QuestionId":42354269,"AnswerCount":1,"Tags":"<machine-learning><statistics><tensorflow>","CreationDate":"2017-02-20T21:10:38.363","AcceptedAnswerId":42354636.0,"OwnerUserId":3970726.0,"Title":"Tensorflow: sigmoid_cross_entropy_with_logits","Body":"<p>I have an image multilabel classification problem that I would like to solve with tensorflow.<\/p>\n\n<p>I'm trying to construct proper loss function and a \"proper\" final layer for CNN network.<\/p>\n\n<p>What kind of arguments the  <\/p>\n\n<pre><\/pre>\n\n<p>function expects?<\/p>\n\n<p>Am I safe to assume that:<\/p>\n\n<ul>\n<li>labels are vectors with binary values {0,1}<\/li>\n<li>logits are vectors with same dimmension as labels with values from whole ]-\u221e, \u221e[<\/li>\n<\/ul>\n\n<p>Therefore I should skip ReLU in the last layer (to ensure final output can be negative).<\/p>\n\n<p>Or maybe logits are bounded and represent probabilty?<\/p>\n\n<p>I'm not 100% sure on this.<\/p>\n","answers":[{"AnswerId":"42354636","CreationDate":"2017-02-20T21:34:45.697","ParentId":null,"OwnerUserId":"5951780","Title":null,"Body":"<p>You are right. Your label though can be any real number between 0 and 1, even though it would be either or usually. But in theory (and practice occasionally) labels here can be any real number in the range [0,1].<\/p>\n\n<p>For what regards the logits: No activation. You are right. <\/p>\n\n<p>The sigmoid_cross_entropy_with_logits uses the logistic loss as in:\nz * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))\nwith x being your logit. You basically already have the activation in that sigmoid-part.<\/p>\n"}]}
{"QuestionId":42354526,"AnswerCount":1,"Tags":"<lua><deep-learning><torch><resnet>","CreationDate":"2017-02-20T21:27:56.137","AcceptedAnswerId":null,"OwnerUserId":3867261.0,"Title":"How to save feature values of all batch data from pretrained torch networks?","Body":"<p>Now I'm using fb torch library from github <a href=\"https:\/\/github.com\/facebook\/fb.resnet.torch\" rel=\"nofollow noreferrer\">fb torch resnet<\/a><\/p>\n\n<p>It's my first time to use torch and lua, so Im encountering some problems.<\/p>\n\n<p>My goal is to save the feature vector of specific layer (last avg pooling of resnet) into a one file with the class of the input image. All input images are from cifar-10 db.<\/p>\n\n<p>The file format that i want to get is like belows<\/p>\n\n<pre><\/pre>\n\n<p>Now I have seen some sample code of that github <a href=\"https:\/\/github.com\/facebook\/fb.resnet.torch\/blob\/master\/pretrained\/extract-features.lua\" rel=\"nofollow noreferrer\">extract-features.lua<\/a><\/p>\n\n<p>Because it's my first time for lua, I feel so hard to understand this code and to modify to the way i want. And i don't want my data to save into t7 file format.<\/p>\n\n<ol>\n<li>How can i access only one specific layer from network in torch via lua? (last average pooling)<\/li>\n<li>How can i access values of the layer and classification result index?<\/li>\n<li>How can read all each images from cifar-10 db file(t7 batch)?<\/li>\n<\/ol>\n\n<p>Sorry for too many questions. But im feeling hard using torch because of pool amouns of community threads and posting of torch.. please understand me. <\/p>\n","answers":[{"AnswerId":"42362794","CreationDate":"2017-02-21T09:00:29.993","ParentId":null,"OwnerUserId":"6076729","Title":null,"Body":"<blockquote>\n  <p>How can i access only one specific layer from network in torch via lua? (last average pooling)<\/p>\n<\/blockquote>\n\n<p>To access each layer you just have to load the model and get it using an integer number. If you do <code>print model<\/code> you will be able to see in which position the last average pooling is.<\/p>\n\n<pre><code>model = torch.load(path_to_model):cuda()\navg_pooling_layer = model:get(position_of_the_avg_pooling_layer)\n<\/code><\/pre>\n\n<blockquote>\n  <p>How can i access values of the layer and classification result index?<\/p>\n<\/blockquote>\n\n<p>I do not quite understand what you mean by this. If you want to see the output or the weights from a specific layer. (following the code above) You need to get these elements from the <em>layer<\/em> table. Again, to see which ones are the possible elements to get use <code>print avg_pooling_layer<\/code><\/p>\n\n<pre><code>weights = avg_pooling_layer.weight -- get the weights of the layer\noutput = avg_pooling_layer.output  -- get the output of the layer\n<\/code><\/pre>\n\n<blockquote>\n  <p>How can read all each images from cifar-10 db file(t7 batch)?<\/p>\n<\/blockquote>\n\n<p>To read the images from a t7 file use the torch function <code>torch.load<\/code>. (used before to load the model).<\/p>\n\n<pre><code>cifar_10 = torch.load(\"path_to_cifar-10.t7\")\n<\/code><\/pre>\n\n<p>Once loaded you could have the training and test set in subtables or functions. Again, print the table and visualize which values are the ones you need to get.<\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":42354607,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-02-20T21:33:14.177","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"Tensorflow: Create a batch from a list of image tensors","Body":"<p>I have a function  that performs preprocessing on my input images. I gather all images that belong to the same batch in a list like this:<\/p>\n\n<pre><\/pre>\n\n<p>Now I want to convert this list into one single tensor with the first dimension being the batch size dimension, such that I could feed it to the input placeholder of my network.<\/p>\n\n<pre><\/pre>\n\n<p>Any idea how I could do that?<\/p>\n\n<pre><\/pre>\n\n<p>When I try to concatenate all tensors, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>So maybe it would be enough to convert the tensor back into a numpy array before feeding it to the network?<\/p>\n","answers":[{"AnswerId":"52229703","CreationDate":"2018-09-07T21:14:35.927","ParentId":null,"OwnerUserId":"9415153","Title":null,"Body":"<p>The issue here is using a tensor as a value in feed_dict. Instead of feeding <code>batch<\/code> as the value for input_placeholder, why don't you use <code>batch<\/code> instead of <code>input_placeholder<\/code>, assuming <code>batch<\/code> is your batched tensor? <\/p>\n\n<p>So, instead of:<\/p>\n\n<pre><code>input_placeholder = tf.Placeholder(tf.int32)\nloss = some_function(input_placeholder)\nsess.run(loss, feed_dict={input_placeholder: batch})\n<\/code><\/pre>\n\n<p>Do:<\/p>\n\n<pre><code>loss = some_function(batch)\nsess.run(batch)\n<\/code><\/pre>\n"},{"AnswerId":"43852695","CreationDate":"2017-05-08T16:25:44.857","ParentId":null,"OwnerUserId":"449886","Title":null,"Body":"<p>In the TF r1.1 pack has been replaced with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/stack\" rel=\"nofollow noreferrer\">tf.stack<\/a><\/p>\n"},{"AnswerId":"42499686","CreationDate":"2017-02-28T03:31:14.407","ParentId":null,"OwnerUserId":"7633030","Title":null,"Body":"<p>You can use <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/array_ops\/slicing_and_joining#pack\" rel=\"nofollow noreferrer\">tf.pack<\/a> to pack a list of tensors into a batch.<\/p>\n\n<pre><code>image_list = [get_image(file_path) for file_path in batch_files]\nimage_batch = tf.pack(image_list)\n<\/code><\/pre>\n\n<p>You can also use tf.concat to concatenate the list along the first dimension and reshape it.<\/p>\n"}]}
{"QuestionId":42354634,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-20T21:34:39.233","AcceptedAnswerId":null,"OwnerUserId":1951176.0,"Title":"Range of size of another tensor's dimension","Body":"<p>I would like to calculate a subset of a tensor in a following matter:\nI have a tensor  and a tensor of indices . I would like to get a tensor  such that:<\/p>\n\n<pre><\/pre>\n\n<p>To do that I thought I'll do:<\/p>\n\n<pre><\/pre>\n\n<p>But I have a problem in the second line:<\/p>\n\n<pre><\/pre>\n\n<p>The shape of  is dependent on the size of some placeholder. Is there a different way to solve my problem or to get a range of size of another tensor's dimension, which is not know at TF-graph compile time?<\/p>\n","answers":[{"AnswerId":"42355637","CreationDate":"2017-02-20T22:51:48.907","ParentId":null,"OwnerUserId":"6791223","Title":null,"Body":"<p>Use <code>tf.shape<\/code> instead of <code>get_shape<\/code>.<\/p>\n"}]}
{"QuestionId":42354675,"AnswerCount":2,"Tags":"<macos><tensorflow><bazel>","CreationDate":"2017-02-20T21:37:17.730","AcceptedAnswerId":null,"OwnerUserId":7595292.0,"Title":"Where does Bazel store TenserFlow build?","Body":"<p>I'm trying to build TensorFlow from sources following this guide: <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"nofollow noreferrer\">Installing TensorFlow from Sources<\/a>. The build seems to have worked fine, but then there's the last step:<\/p>\n\n<blockquote>\n  <p>Invoke pip install to install that pip package. The filename of the\n  .whlfile depends on your platform. For example, the following command\n  will install the pip package for TensorFlow 1.0.0 on Linux: <\/p>\n\n<pre><\/pre>\n<\/blockquote>\n\n<p>I suppose that's great if you run Linux, but I would have appreciated the location on Mac OS X as well.<\/p>\n\n<p>Where is the package stored on Mac? I can't find it int , nor . And since search is broken on Mac, I'm out of luck.<\/p>\n","answers":[{"AnswerId":"42358003","CreationDate":"2017-02-21T03:17:04.550","ParentId":null,"OwnerUserId":"5020937","Title":null,"Body":"<p>It sounds like you may have skipped a step. Bazel does not create this file. The program that Bazel builds does.<\/p>\n\n<p>The prior step on <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources<\/a> to the one that you mention is to run<\/p>\n\n<pre><code>$ bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n<\/code><\/pre>\n\n<p>The second argument specified where to put the wheel file. Furthermore, that program logs its output directory:<\/p>\n\n<pre><code>$ bazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\nMon Feb 20 22:08:08 EST 2017 : === Using tmpdir: \/var\/folders\/yt\/9r8s598x603bx58zq85yrx680096yv\/T\/tmp.XXXXXXXXXX.gp5oIM0Z\nMon Feb 20 22:08:13 EST 2017 : === Building wheel\nMon Feb 20 22:08:45 EST 2017 : === Output wheel file is in: \/tmp\/tensorflow_pkg\n$ ls \/tmp\/tensorflow_pkg\/\ntensorflow-1.0.0-cp27-cp27m-macosx_10_12_intel.whl\n<\/code><\/pre>\n"},{"AnswerId":"42355152","CreationDate":"2017-02-20T22:12:37.203","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Here's the procedure I use for building on MacOS<\/p>\n\n<p>Pre-requisites<\/p>\n\n<ul>\n<li>Install Bazel<\/li>\n<li>Install CUDA 8.0<\/li>\n<li>Clone tensorflow into ~\/git\/tensorflow<\/li>\n<\/ul>\n\n<p>Then<\/p>\n\n<pre><code>cd ~\/git\/tensorflow\ngit fetch --all\nbrew upgrade bazel\n\ngit stash\ngit rebase tf\/master\ngit stash pop\n\n# to resolve conflict from previous .\/configure run\ngit checkout --theirs -- tensorflow\/core\/platform\/default\/build_config.bzl\n\n# do the actual build in TMUX session to avoid losing state when closing terminal\ntmux new-session -s bazel -n 0\ncd ~\/git\/tensorflow\n\n.\/configure\n\nexport CUDA_VISIBLE_DEVICES=\nexport CUDA_HOME=\/usr\/local\/cuda\nexport DYLD_LIBRARY_PATH=\/usr\/local\/cuda\/lib:\/usr\/local\/cuda\/extras\/CUPTI\/lib\nexport LD_LIBRARY_PATH=$DYLD_LIBRARY_PATH\nexport PATH=$DYLD_LIBRARY_PATH:$PATH\n\nexport date=feb11\nexport tag=head\nexport flags=\"--config=cuda --config=opt -k --action_env PATH --action_env DYLD_LIBRARY_PATH --action_env LD_LIBRARY_PATH\"\n\nbazel build $flags -k \/\/tensorflow\/tools\/pip_package:build_pip_package\nrm -Rf \/tmp\/tensorflow_pkg\nmkdir \/tmp\/tensorflow_pkg\nbazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg\n\nmkdir -p ~\/g\/tfbins\/$date.$tag\ncp `find \/tmp\/tensorflow_pkg -type f ` ~\/tfbins\/$date.$tag\n<\/code><\/pre>\n\n<p>This will drop your wheel file into <code>~\/tfbins\/feb11.head<\/code>\nYou can automate <code>.\/configure<\/code> part by using something like\n<a href=\"https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/configure_tf.sh\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/configure_tf.sh<\/a><\/p>\n\n<p>Note, CUDA building on MacOS is broken with latest Bazel release right now, you need to downgrade to 0.4.3 or build Bazel from head if you want CUDA -- <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7227\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7227<\/a><\/p>\n\n<p>You can then\u00a0share the wheel publicly by doing this (setup gsutil first and choose different $bucket)<\/p>\n\n<pre><code>export wheel=`find ~\/tfbins\/$date.$tag -type f`\nexport basename=`find ~\/tfbins\/$date.$tag -type f -printf \"%f\\n\"`\ncd ~\/tfbins\/$date.$tag\nfullname=$date.$tag.$basename\nln -s $basename $fullname\nexport bucket=tensorflow-community-wheels\ngsutil cp $fullname gs:\/\/$bucket\ngsutil acl set public-read gs:\/\/$bucket\/$fullname\n\necho https:\/\/storage.cloud.google.com\/$bucket\/$fullname\n<\/code><\/pre>\n\n<p>Then share the link printed from last line to <a href=\"https:\/\/github.com\/yaroslavvb\/tensorflow-community-wheels\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yaroslavvb\/tensorflow-community-wheels<\/a> :)<\/p>\n"}]}
{"QuestionId":42355005,"AnswerCount":1,"Tags":"<python><tensorflow><softmax>","CreationDate":"2017-02-20T22:01:21.960","AcceptedAnswerId":42355510.0,"OwnerUserId":4333347.0,"Title":"Cannot print correct predictions using tensor flow","Body":"<p>I've implemented a logistic regression that works fine. It prints out the accuracy correctly. I display the accuracy as so...<\/p>\n\n<pre><\/pre>\n\n<p>As I said, this works fine. However, after reading a tutorial, I know that  should be an array of booleans that tells us whether our predictions are true. I want to print this boolean value but I run into issues. I tried the following...<\/p>\n\n<pre><\/pre>\n\n<p>and then I tried...<\/p>\n\n<pre><\/pre>\n\n<p>I am quite new to TensorFlow. How can I print out this variable as an array of predictions?<\/p>\n","answers":[{"AnswerId":"42355510","CreationDate":"2017-02-20T22:40:55.487","ParentId":null,"OwnerUserId":"5951780","Title":null,"Body":"<p>You still need to feed in the data. Try: <\/p>\n\n<pre><code>print(correct_prediction.eval({x: mnist.test.images, y: mnist.test.labels}))\n<\/code><\/pre>\n"}]}
{"QuestionId":42355122,"AnswerCount":4,"Tags":"<csv><tensorflow>","CreationDate":"2017-02-20T22:10:33.823","AcceptedAnswerId":42358524.0,"OwnerUserId":562769.0,"Title":"Can I export a tensorflow summary to CSV?","Body":"<p>Is there a way to extract scalar summaries to CSV (preferably from within tensorboard) from tfevents files?<\/p>\n\n<h2>Example code<\/h2>\n\n<p>The following code generates tfevent files in a  within the same directory. Suppose you let it run and you find something interesting. You want to get the raw data for further investigation. How would you do that?<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"52095336","CreationDate":"2018-08-30T10:57:34.047","ParentId":null,"OwnerUserId":"2230045","Title":null,"Body":"<p>While the answer here is as requested within tensorboard it only allows to download a csv for a single run of a single tag.\nIf you have for example 10 tags and 20 runs (what is not at all much) you would need to do the above step 200 times (that alone will probably take you more than a hour).\nIf now you for some reason would like to actually do something with the data for all runs for a single tag you would need to write some weird CSV accumulation script or copy everything by hand (what will probably cost you more than a day).<\/p>\n\n<p>Therefore I would like to add a solution that extracts a CSV file for every tag with all runs contained. Column headers are the run path names and row indices are the run step numbers.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import os\nimport numpy as np\nimport pandas as pd\n\nfrom collections import defaultdict\nfrom tensorboard.backend.event_processing.event_accumulator import EventAccumulator\n\n\ndef tabulate_events(dpath):\n    summary_iterators = [EventAccumulator(os.path.join(dpath, dname)).Reload() for dname in os.listdir(dpath)]\n\n    tags = summary_iterators[0].Tags()['scalars']\n\n    for it in summary_iterators:\n        assert it.Tags()['scalars'] == tags\n\n    out = defaultdict(list)\n    steps = []\n\n    for tag in tags:\n        steps = [e.step for e in summary_iterators[0].Scalars(tag)]\n\n        for events in zip(*[acc.Scalars(tag) for acc in summary_iterators]):\n            assert len(set(e.step for e in events)) == 1\n\n            out[tag].append([e.value for e in events])\n\n    return out, steps\n\n\ndef to_csv(dpath):\n    dirs = os.listdir(dpath)\n\n    d, steps = tabulate_events(dpath)\n    tags, values = zip(*d.items())\n    np_values = np.array(values)\n\n    for index, tag in enumerate(tags):\n        df = pd.DataFrame(np_values[index], index=steps, columns=dirs)\n        df.to_csv(get_file_path(dpath, tag))\n\n\ndef get_file_path(dpath, tag):\n    file_name = tag.replace(\"\/\", \"_\") + '.csv'\n    folder_path = os.path.join(dpath, 'csv')\n    if not os.path.exists(folder_path):\n        os.makedirs(folder_path)\n    return os.path.join(folder_path, file_name)\n\n\nif __name__ == '__main__':\n    path = \"path_to_your_summaries\"\n    to_csv(path)\n<\/code><\/pre>\n\n<p>My solution builds upon: <a href=\"https:\/\/stackoverflow.com\/a\/48774926\/2230045\">https:\/\/stackoverflow.com\/a\/48774926\/2230045<\/a><\/p>\n\n<hr>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>I created a more sophisticated version and released it on GitHub: <a href=\"https:\/\/github.com\/Spenhouet\/tensorboard-aggregator\" rel=\"noreferrer\">https:\/\/github.com\/Spenhouet\/tensorboard-aggregator<\/a><\/p>\n\n<p>This version aggregates multiple tensorboard runs and is able to save the aggregates to a new tensorboard summary or as a <code>.csv<\/code> file.<\/p>\n"},{"AnswerId":"56402388","CreationDate":"2019-05-31T22:34:13.707","ParentId":null,"OwnerUserId":"3995368","Title":null,"Body":"<p>Just to add to @Spen <\/p>\n\n<p>in case you want to export the data when you have varying numbers of steps.\nThis will make one large csv file.\nMight need to change around the keys for it to work for you.<\/p>\n\n<pre><code>import os\nimport numpy as np\nimport pandas as pd\n\nfrom collections import defaultdict\nfrom tensorboard.backend.event_processing.event_accumulator import EventAccumulator\nimport glob\nimport pandas as pd\nlistOutput = (glob.glob(\"*\/\"))\n\nlistDF = []\n\nfor tb_output_folder in listOutput:\n print(tb_output_folder)\n x = EventAccumulator(path=tb_output_folder)\n x.Reload()\n x.FirstEventTimestamp()\n keys = ['loss', 'mean_absolute_error', 'val_loss', 'val_mean_absolute_error'] \n\n listValues = {}\n\n steps = [e.step for e in x.Scalars(keys[0])]\n wall_time = [e.wall_time for e in x.Scalars(keys[0])]\n index = [e.index for e in x.Scalars(keys[0])]\n count = [e.count for e in x.Scalars(keys[0])]\n n_steps = len(steps)\n listRun = [tb_output_folder] * n_steps\n printOutDict = {}\n\n data = np.zeros((n_steps, len(keys)))\n for i in range(len(keys)):\n     data[:,i] = [e.value for e in x.Scalars(keys[i])]\n\n printOutDict = {keys[0]: data[:,0], keys[1]: data[:,1],keys[2]: data[:,2],keys[3]: data[:,3]}\n\n printOutDict['Name'] = listRun\n\n DF = pd.DataFrame(data=printOutDict)\n\n listDF.append(DF)\n\ndf = pd.concat(listDF)\ndf.to_csv('Output.csv')   \n\n<\/code><\/pre>\n"},{"AnswerId":"57411105","CreationDate":"2019-08-08T11:04:05.590","ParentId":null,"OwnerUserId":"4293113","Title":null,"Body":"<p>Here is my solution which bases on the previous solutions but can scale up.<\/p>\n\n<pre><code>import os\nimport numpy as np\nimport pandas as pd\n\nfrom collections import defaultdict\nfrom tensorboard.backend.event_processing.event_accumulator import EventAccumulator\n\n\ndef tabulate_events(dpath):\n\n    final_out = {}\n    for dname in os.listdir(dpath):\n        print(f\"Converting run {dname}\",end=\"\")\n        ea = EventAccumulator(os.path.join(dpath, dname)).Reload()\n        tags = ea.Tags()['scalars']\n\n        out = {}\n\n        for tag in tags:\n            tag_values=[]\n            wall_time=[]\n            steps=[]\n\n            for event in ea.Scalars(tag):\n                tag_values.append(event.value)\n                wall_time.append(event.wall_time)\n                steps.append(event.step)\n\n            out[tag]=pd.DataFrame(data=dict(zip(steps,np.array([tag_values,wall_time]).transpose())), columns=steps,index=['value','wall_time'])\n\n        if len(tags)&gt;0:      \n            df= pd.concat(out.values(),keys=out.keys())\n            df.to_csv(f'{dname}.csv')\n            print(\"- Done\")\n        else:\n            print('- Not scalers to write')\n\n        final_out[dname] = df\n\n\n    return final_out\nif __name__ == '__main__':\n    path = \"youre\/path\/here\"\n    steps = tabulate_events(path)\n    pd.concat(steps.values(),keys=steps.keys()).to_csv('all_result.csv')\n<\/code><\/pre>\n"},{"AnswerId":"42358524","CreationDate":"2017-02-21T04:09:44.953","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>Just check the \"Data download links\" option on the upper-left in TensorBoard, and then click on the \"CSV\" button that will appear under your scalar summary.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/12F2d.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/12F2d.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":42355723,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-02-20T22:58:34.570","AcceptedAnswerId":null,"OwnerUserId":6063977.0,"Title":"Is it possible to train on multiple images sizes in keras?","Body":"<p>Keras takes numpy arrays as input for training data, however it is possible to create models that can take variable input sizes. I'm wondering if there is a way to incorporate images of various dimensions in the training data for a model.<\/p>\n","answers":[{"AnswerId":"42356212","CreationDate":"2017-02-20T23:47:55.367","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>You cannot give variable size images to train a model in Keras. According to Keras API, the <code>Input<\/code> layer function looks as follows.<\/p>\n\n<pre><code>Input(shape=(3,None,None))\n<\/code><\/pre>\n\n<p>Where, <code>3<\/code> is presenting number of channels for RGB images. But you have to clearly tell Keras, what is the width and height of the training images. So, there is no way Keras can handle variable length images.<\/p>\n\n<p>So, you have to transform the images to a specific size first and then train model using Keras.<\/p>\n"}]}
{"QuestionId":42356027,"AnswerCount":0,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2017-02-20T23:29:54.400","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"What is the upside of using `tf.nn.rnn` instead of `tf.nn.dynamic_rnn` in TensorFlow?","Body":"<p>What is the upside of using <a href=\"https:\/\/web.archive.org\/web\/20170220232722\/https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn\/recurrent_neural_networks#rnn\" rel=\"nofollow noreferrer\"><\/a> instead of <a href=\"https:\/\/web.archive.org\/web\/20170220232722\/https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn\/recurrent_neural_networks#dynamic_rnn\" rel=\"nofollow noreferrer\"><\/a>?<\/p>\n\n<p>The <a href=\"https:\/\/web.archive.org\/web\/20170220232722\/https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn\/recurrent_neural_networks#dynamic_rnn\" rel=\"nofollow noreferrer\">documentation<\/a>  says :<\/p>\n\n<blockquote>\n  <p>[] is functionally identical to the function  above, but performs fully dynamic unrolling of inputs.<\/p>\n<\/blockquote>\n\n<p>Is there any case where one might prefer to use  instead of ?<\/p>\n","answers":[]}
{"QuestionId":42356207,"AnswerCount":1,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2017-02-20T23:47:32.460","AcceptedAnswerId":null,"OwnerUserId":5496253.0,"Title":"How to pass a TensorFlow tensor to a TFLearn model","Body":"<p>I am working on a project in TensorFlow that performs operations on already-trained machine learning models. Following the tutorial <a href=\"http:\/\/tflearn.org\/tutorials\/quickstart.html\" rel=\"nofollow noreferrer\">TFLearn Quickstart<\/a>, I built a deep neural network that predicts survival from the <a href=\"https:\/\/www.kaggle.com\/c\/titanic\" rel=\"nofollow noreferrer\">Titanic Dataset<\/a>. I would like to use the TFLearn model in the same way that I would use a TensorFlow model.<\/p>\n\n<p>The TFLearn docs homepage says<\/p>\n\n<blockquote>\n  <p>Full transparency over Tensorflow. All functions are built over tensors and can be used independently of TFLearn<\/p>\n<\/blockquote>\n\n<p>This makes me think that I would be able to pass tensors as inputs, etc. to the the TFLearn model.<\/p>\n\n<pre><\/pre>\n\n<p>At present, when I pass a tensor into the model I am greeted with <em>ValueError: setting an array element with a sequence.<\/em><\/p>\n\n<p>Specifically, how can you pass tensors into a TFLearn model?\nGenerally, what limits are placed on how I can use tensors on a TFLearn model?<\/p>\n","answers":[{"AnswerId":"45647957","CreationDate":"2017-08-12T08:00:33.910","ParentId":null,"OwnerUserId":"8198272","Title":null,"Body":"<p>I don't know if you're still looking for an answer to your problem, but I think the issue is on your very last line:<\/p>\n\n<pre><code>print(model.predict([testTF]))\n<\/code><\/pre>\n\n<p>Try this instead:<\/p>\n\n<pre><code>print(model.predict(testTF))\n<\/code><\/pre>\n\n<p>I think that you nested a list inside another list.  This isn't a TFlearn issue <em>per se<\/em>.  Hope that helps.<\/p>\n"}]}
{"QuestionId":42356240,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-20T23:50:40.937","AcceptedAnswerId":null,"OwnerUserId":6405712.0,"Title":"In TensorFlow, what is the difference between a constant and an initializer?","Body":"<p>Consider the following in TensorFlow:<\/p>\n\n<pre><\/pre>\n\n<p>What is the difference between the two? When do you use an initializer vs. a constant?<\/p>\n","answers":[]}
{"QuestionId":42356396,"AnswerCount":3,"Tags":"<python><matplotlib><tensorflow><importerror><conda>","CreationDate":"2017-02-21T00:07:20.473","AcceptedAnswerId":42356600.0,"OwnerUserId":6917400.0,"Title":"ImportError: No module named 'matplotlib' -- Using Anaconda tensorflow environment","Body":"<p>I'm just trying to learn Tensorflow, but am totally new to Python, so I'm using Anaconda\nI created a conda environment: <\/p>\n\n<pre><\/pre>\n\n<p>Of course I activated my conda environment<\/p>\n\n<pre><\/pre>\n\n<p>Then I played a bit around in Spyder, tried to plot a MNIST-digit (copy-paste code from my tutor which is tested several times), it includes of course <\/p>\n\n<pre><\/pre>\n\n<p>but executing the Python file with bash gives me: <\/p>\n\n<pre><\/pre>\n\n<p>I'm quite confused right now, as the  in the bash obviously denotes that my conda tensorflow environment works (at least from my understanding). Also, from what I understood, conda should have matplotlib built in, right? And it should also load this in my conda tensorflow environment, right? This is what my tutor's slide said <\/p>\n\n<blockquote>\n  <p>There is no need to install further packages like numpy or matplotlib, since Anaconda contains current versions of them already.'<\/p>\n<\/blockquote>\n\n<p>and also what I was able to take from everything I Googled and StackOverflowed. Neither Googling nor StackOverflowing gave me any good answer (might also just be because I don't understand enough yet). <\/p>\n\n<p>My best guess would be that I still have to include matplotlib into my tensorflow conda environment, but this would be contradicting both my tutor &amp; Google, while I also would not know how to do this.<\/p>\n\n<p>edit:  gave me that matplotlib was not in my tensorflowenvironment, so I went<\/p>\n\n<pre><\/pre>\n\n<p>I'm still afraid something is wrong with my conda tensorflow environment, shouldn't the matplotlib have been in there by default? It also told me: <\/p>\n\n<pre><\/pre>\n\n<p>Which tells me also numpy was missing? Can someone confirm this to be correct now, or is there something fishy with my conda? <\/p>\n","answers":[{"AnswerId":"42356600","CreationDate":"2017-02-21T00:29:15.773","ParentId":null,"OwnerUserId":"1634191","Title":null,"Body":"<p>You just created a conda environment named tensorflow and switched into it.  You haven't installed the tensorflow package or any of the default anaconda packages.<\/p>\n\n<p>To do that, do<\/p>\n\n<pre><code>conda create -n tensorflow python=3.5 anaconda # install anaconda3 default packages\nsource activate tensorflow # switch into it\nconda install -c conda-forge tensorflow # install tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"57303251","CreationDate":"2019-08-01T06:49:35.580","ParentId":null,"OwnerUserId":"1518748","Title":null,"Body":"<p>I faced the same problem on my mac.\nSo I ran conda list to see whether the matplotlib is installed or not.\nOnce i found it missing, i went ahead and ran the command <code>conda install matplotlib<\/code>.\nAfter this step to verify it's properly installed; Do the following.<\/p>\n\n<pre><code>conda activate tf\n<\/code><\/pre>\n\n<p>This activates the tensorflow environment on anaconda.\nAfter this start interactive python shell on the same terminal.<\/p>\n\n<pre><code>import matplotlib\n<\/code><\/pre>\n\n<p>If its installed properly this should not throw any error now.<\/p>\n"},{"AnswerId":"55404841","CreationDate":"2019-03-28T18:46:24.447","ParentId":null,"OwnerUserId":"11274048","Title":null,"Body":"<p>I ran into the same problem using these instructions:\n<a href=\"https:\/\/www.anaconda.com\/tensorflow-in-anaconda\/\" rel=\"nofollow noreferrer\">https:\/\/www.anaconda.com\/tensorflow-in-anaconda\/<\/a>\nfor tensorflow-gpu.<\/p>\n\n<p>Running<\/p>\n\n<pre><code>conda create -n tensorflow_gpuenv tensorflow-gpu\nconda activate tensorflow_gpuenv\n<\/code><\/pre>\n\n<p>should ensure that \"TensorFlow is now installed and ready for use.\"\nBut it doesn't.  Running 'conda list' shows matplotlib was not installed.  So you'll need to install that as well:<\/p>\n\n<pre><code>conda install -c conda-forge matplotlib\n<\/code><\/pre>\n"}]}
{"QuestionId":42356596,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-21T00:28:58.173","AcceptedAnswerId":42356937.0,"OwnerUserId":4333347.0,"Title":"How to get the trained weights created by a model","Body":"<p>I implemented a simple logistic regression. Before running the training algorithm, I created a placeholder for my weights where I initialized all the weights to 0...<\/p>\n\n<pre><\/pre>\n\n<p>After initializing all my variables correctly, the logistic regression is implemented (which I've tested and runs correctly)...<\/p>\n\n<pre><\/pre>\n\n<p>My issue is, I need to extract the weights used in the model. I used the following for my model...<\/p>\n\n<pre><\/pre>\n\n<p>I tried extracting the following way...<\/p>\n\n<pre><\/pre>\n\n<p>I thought that the trained weights would be located in , however when I run the  function, I get an array of zeroes.<\/p>\n\n<p>What I want, is the all sets of weights. But for some reason I am getting arrays of zeroes instead of the actual weights of the classifier.<\/p>\n","answers":[{"AnswerId":"42356933","CreationDate":"2017-02-21T01:05:48.307","ParentId":null,"OwnerUserId":"1318734","Title":null,"Body":"<p>The variable <code>W<\/code> should refer to the trained weights. Please try simply doing: <code>sess.run(W)<\/code><\/p>\n"},{"AnswerId":"42356937","CreationDate":"2017-02-21T01:06:16.680","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Is much easier, just evaluate the weights with the run function and you will get back the numpy array with the values:<\/p>\n\n<pre><code>sess.run([x, W, b])\n<\/code><\/pre>\n"}]}
{"QuestionId":42356826,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><keras>","CreationDate":"2017-02-21T00:55:22.157","AcceptedAnswerId":42357692.0,"OwnerUserId":288609.0,"Title":"Regarding defining input_placeholder to predict over a given image","Body":"<p>In the following function, author aims to run prediction over a downloaded image based on vgg model.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>I have been trying to predict over an existing image read via opencv, the only modification I made is to read image via cv2, add input_placeholder, and modify sess.run correspondingly. However, I got the following error message: <\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>Would you like to let me know which modification I made is wrong?<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42357692","CreationDate":"2017-02-21T02:36:44.850","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>For reference first take a look to the official documentation: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run<\/a><\/p>\n\n<p>For each graph element passed to the <code>fetches<\/code> parameter of <code>sess.run()<\/code> you get one value returned. In your case you are passing the following list as fetches: <code>[pred,logits]<\/code> so <code>sess.run([pred,logits], ...)<\/code> will return 2 values: the result of running the <code>pred<\/code> op and the <code>logits<\/code> op.<\/p>\n\n<p>Citing the documentation<\/p>\n\n<blockquote>\n  <p>The value returned by run() has the same shape as the fetches\n  argument, where the leaves are replaced by the corresponding values\n  returned by TensorFlow.<\/p>\n<\/blockquote>\n\n<p>However in this line <\/p>\n\n<p><code>segmentation, np_image, np_logits = sess.run([pred,logits],feed_dict={input_placeholder:image})<\/code><\/p>\n\n<p>you are trying to assign these 2 values to 3 different python variables (<code>segmentation<\/code>, <code>np_image<\/code>, <code>np_logits<\/code>) hence you get the <code>ValueError<\/code>.<\/p>\n\n<p>If you look at the original example you provided, the final line is:<\/p>\n\n<p><code>segmentation, np_image, np_logits = sess.run([pred, image, logits])<\/code><\/p>\n\n<p>To mimic the original example try removing the <code>np_image<\/code> declaration from your code like so:<\/p>\n\n<p><code>segmentation, np_logits = sess.run([pred,logits],feed_dict={input_placeholder:image})<\/code><\/p>\n"},{"AnswerId":"42356854","CreationDate":"2017-02-21T00:58:08.040","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>is a python error:<\/p>\n\n<pre><code>segmentation, np_image, np_logits = sess.run([pred,logits],feed_dict={input_placeholder:image})\n<\/code><\/pre>\n\n<p>should be<\/p>\n\n<pre><code>np_image, np_logits = sess.run([pred,logits],feed_dict={input_placeholder:image})\n<\/code><\/pre>\n\n<p>you are only passing two fetches to the run function<\/p>\n"}]}
{"QuestionId":42357268,"AnswerCount":0,"Tags":"<tensorflow><gmm>","CreationDate":"2017-02-21T01:47:18.503","AcceptedAnswerId":null,"OwnerUserId":7595963.0,"Title":"Looking for a method to replace sklearn.mixture.GaussianMixture.predict_proba(X) in Tensorflow","Body":"<p>In , I can use  if I want to Predict posterior probability of data per each component.<\/p>\n\n<p>However, I cannot find a similar method in 'tensorflow gmm_ops'<\/p>\n\n<p>Did anyone find that before?<\/p>\n","answers":[]}
{"QuestionId":42357609,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-21T02:27:35.333","AcceptedAnswerId":42359652.0,"OwnerUserId":6302651.0,"Title":"difference in checkpoint of Tensorflow version 0.11 and version 0.12?","Body":"<p>I developed a CNN algorithm using Python and Tensorflow library.\nI run the program on two Tensorflow versions as 0.11 and 0.12.\nIn the generated checkpoint folder, there is a difference between two versions:<\/p>\n\n<ol>\n<li>Version 0.11\n\n<ul>\n<li>checkpoint100.ckp<\/li>\n<li>checkpoint100.ckp.meta<\/li>\n<\/ul><\/li>\n<li>Version 0.12\n\n<ul>\n<li>checkpoint100.ckp.index<\/li>\n<li>checkpoint100.ckp.meta<\/li>\n<li>checkpoint100.ckp.data-00000-of-00001<\/li>\n<\/ul><\/li>\n<\/ol>\n\n<p>So could anybody explain what is actual difference in checkpoint between 2 tensorflow versions? Thank you!<\/p>\n","answers":[{"AnswerId":"42359652","CreationDate":"2017-02-21T05:49:48.040","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In TensorFlow 0.12, we introduced the more efficient \"V2\" checkpoint format, with two main differences:<\/p>\n\n<ul>\n<li>A V2 checkpoint includes a separate <code>.index<\/code> file, which provides  a compact index of the particular data file (and location within that file) where a particular saved tensor can be found.<\/li>\n<li>The V2 checkpoint format supports larger tensors than its predecessor (which used a Protocol Buffer to store tensor values, with a limit of at most 2GB per tensor).<\/li>\n<\/ul>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver#__init__\" rel=\"nofollow noreferrer\"><code>tf.train.Saver<\/code> initializer<\/a> accepts an optional <code>write_version<\/code> argument, which controls the format that is used when writing a checkpoint, and defaults to the V2 format. Since TensorFlow 0.12, the <code>tf.train.Saver<\/code> has supported reading and writing both checkpoint formats.<\/p>\n"}]}
{"QuestionId":42357850,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-21T02:59:20.993","AcceptedAnswerId":null,"OwnerUserId":2638485.0,"Title":"Tensorboard Scalars Count","Body":"<p>What is the count shown on the right side of a graph on the tensorboard scalars tab (see 1 on the right side of the attached image).<\/p>\n\n<p>There are no runs shown for this measure.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/aDFFL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aDFFL.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42358408","CreationDate":"2017-02-21T03:59:50.777","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>It's the amount of summaries grouped under that category when checking the \"Split on underscores\" option on the upper-left part:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/D2Dpk.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/D2Dpk.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>For example, I have two scalar summaries, one called <code>losses_per-epoch<\/code> and the other called <code>losses_per-batch<\/code>. When checking the \"Split on underscores\" option both are grouped under the \"losses\" category, and the count becomes 2:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xImhG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xImhG.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":42357895,"AnswerCount":1,"Tags":"<python><neural-network><caffe><conv-neural-network><pycaffe>","CreationDate":"2017-02-21T03:04:31.677","AcceptedAnswerId":null,"OwnerUserId":7003063.0,"Title":"Caffe module installation failed in Windows 10","Body":"<p>I am very new to caffe. I was following this link:\n<a href=\"http:\/\/adilmoujahid.com\/posts\/2016\/06\/introduction-deep-learning-python-caffe\/\" rel=\"nofollow noreferrer\">http:\/\/adilmoujahid.com\/posts\/2016\/06\/introduction-deep-learning-python-caffe\/<\/a><\/p>\n\n<p>When I ran this command: <strong>python create_lmdb.py<\/strong>,  I got the following error:<\/p>\n\n<p><strong>File \"create_lmdb.py\", line 19, in  ,     import caffe ;\nImportError: No module named 'caffe'<\/strong><\/p>\n\n<p>I tried to install caffe using this link : <a href=\"https:\/\/github.com\/BVLC\/caffe\/tree\/windows\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/tree\/windows<\/a>, specially the following codes:<\/p>\n\n<pre><\/pre>\n\n<p>I have got the folllowing error:<\/p>\n\n<pre><\/pre>\n\n<p>I downloaded Ninja, but not sure what to do with it. So I tried to change WITH_NINJA=1 to WITH_NINJA=0, then I have got the following error<\/p>\n\n<pre><\/pre>\n\n<p>Please let me know what should be my next step! Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42424735","CreationDate":"2017-02-23T19:41:26.273","ParentId":null,"OwnerUserId":"7613191","Title":null,"Body":"<p>Step #1<\/p>\n\n<p>Open build_win.cmd in notepad++ and update file line as below (to diable Ninja build)<\/p>\n\n<pre><code>:: Change to 1 to use Ninja generator (builds much faster)\nif NOT DEFINED WITH_NINJA set WITH_NINJA=0\n<\/code><\/pre>\n\n<p>I am using Microsoft Visual Studio 2013<\/p>\n\n<pre><code>:: Change MSVC_VERSION to 12 to use VS 2013\nif NOT DEFINED MSVC_VERSION set MSVC_VERSION=12\n<\/code><\/pre>\n\n<p>Step # 2 <\/p>\n\n<p>Open build folder which is inside the caffe folder and delete all previously generated files.<\/p>\n\n<p>Step # 3<\/p>\n\n<p>Python 3.6 is not supported officially and python 3.5 has some challenging issues. For a beginner i would recommend to install python 2.7 straight forward.<\/p>\n\n<pre><code>Install anaconda 2.7 [Anaconda  2.7][1]\n<\/code><\/pre>\n\n<p>Step # 4 <\/p>\n\n<pre><code>Execute scripts\/build_win.cmd\n[1]: https:\/\/www.continuum.io\/downloads\n<\/code><\/pre>\n\n<p>Reformat the code for easy to read<\/p>\n"}]}
{"QuestionId":42358358,"AnswerCount":0,"Tags":"<python><macos><tensorflow><pip><python-module>","CreationDate":"2017-02-21T03:54:56.557","AcceptedAnswerId":null,"OwnerUserId":1930337.0,"Title":"How to get rid of python pip permission warning?","Body":"<p>When I try to install TensorFlow through pip, I get the following warning which I  want to get rid of. <\/p>\n\n<pre><\/pre>\n\n<p>The warning message<\/p>\n\n<pre><\/pre>\n\n<p>should I give all permissions to the Library\/Caches\/pip directory?<\/p>\n","answers":[]}
{"QuestionId":42358685,"AnswerCount":0,"Tags":"<python-3.x><anaconda><theano><theano-cuda>","CreationDate":"2017-02-21T04:25:42.347","AcceptedAnswerId":null,"OwnerUserId":2652482.0,"Title":"Failed to compile cuda_ndarray.cu theano on windows 7","Body":"<p>Hi I am trying to run theano on gpu. <\/p>\n\n<p>My configuration is <\/p>\n\n<p>Windows 7<\/p>\n\n<p>CUDA toolkit 8.0<\/p>\n\n<p>Anaconda python 3.5<\/p>\n\n<p>Visual studio 14.0<\/p>\n\n<pre><\/pre>\n\n<p>Any idea ?<\/p>\n","answers":[]}
{"QuestionId":42359848,"AnswerCount":1,"Tags":"<c++><opencl><caffe>","CreationDate":"2017-02-21T06:05:46.633","AcceptedAnswerId":44980474.0,"OwnerUserId":6863445.0,"Title":"OpenCL: conversion from ' float* ' to ' cl_mem '","Body":"<p>In caffe source code (branch opencl),  sometimes I can see some like these(eg. in      src\/caffe\/layers\/base_conv_layer.cpp line 400):<\/p>\n\n<pre><\/pre>\n\n<p>I think it is really strange, so I googled it, but got nothing about it.<\/p>\n\n<p>Then I wrote a simple test code to figure out what happened, but I found this conversion is meaningless:<\/p>\n\n<pre><\/pre>\n\n<p>Their value is same !<\/p>\n\n<p>So my question is why this conversion raises no problem in caffe opencl ?\nWhat happened in this conversion ?<\/p>\n","answers":[{"AnswerId":"44980474","CreationDate":"2017-07-07T22:22:18.877","ParentId":null,"OwnerUserId":"8273142","Title":null,"Body":"<p>If you look into class <code>SyncedMemory<\/code> (in <code>include\/caffe\/syncedmem.hpp<\/code> and <code>src\/caffe\/syncedmem.cpp<\/code>), you will find that a memory buffer is first created by <code>clCreateBuffer<\/code> and the handle is stored as a <code>cl_mem<\/code> object. The handle is then cast to type <code>DType*<\/code> when <code>gpu_data()<\/code> or <code>mutable_gpu_data()<\/code> is called. It can therefore be cast back to type <code>cl_mem<\/code> because it is a <code>cl_mem<\/code> object in the first place.<\/p>\n"}]}
{"QuestionId":42359860,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><theano><keras>","CreationDate":"2017-02-21T06:06:39.623","AcceptedAnswerId":null,"OwnerUserId":6719992.0,"Title":"add Batch Normalization immediately before non-linearity or after in Keras?","Body":"<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>When I use official <a href=\"https:\/\/github.com\/fchollet\/deep-learning-models\/blob\/master\/inception_v3.py\" rel=\"nofollow noreferrer\">inception_v3<\/a> model in keras, I find that they use BatchNormalization after 'relu' nonlinearity as above code script.<\/p>\n\n<p>But in the Batch Normalization paper, the authors said <\/p>\n\n<blockquote>\n  <p>we add the BN transform immediately before the nonlinearity, by\n  normalizing x=Wu+b.<\/p>\n<\/blockquote>\n\n<p>Then I view the implementation of inception in tensorflow which add BN immediately before the nonlinearity as they said. For more details in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/slim\/ops.py\" rel=\"nofollow noreferrer\">inception ops.py<\/a><\/p>\n\n<p>I'm confused. Why do people use above style in Keras other than the following? <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>In the Dense case:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42949857","CreationDate":"2017-03-22T11:06:26.140","ParentId":null,"OwnerUserId":"7456923","Title":null,"Body":"<p>I also use it before the activation, which is indeed how it was designed, and so do other libraries, such as lasagne's batch_norm <a href=\"http:\/\/lasagne.readthedocs.io\/en\/latest\/modules\/layers\/normalization.html#lasagne.layers.batch_norm\" rel=\"nofollow noreferrer\">http:\/\/lasagne.readthedocs.io\/en\/latest\/modules\/layers\/normalization.html#lasagne.layers.batch_norm<\/a> .<\/p>\n\n<p>However it seems that in practice placing it after the activation works a bit better:<\/p>\n\n<p><a href=\"https:\/\/github.com\/ducha-aiki\/caffenet-benchmark\/blob\/master\/batchnorm.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ducha-aiki\/caffenet-benchmark\/blob\/master\/batchnorm.md<\/a>\n(this is just one benchmark though)<\/p>\n"}]}
{"QuestionId":42360110,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-02-21T06:24:57.547","AcceptedAnswerId":null,"OwnerUserId":541106.0,"Title":"keras CNN with low and constant accuracies","Body":"<p>I am dealing with the Street View House Number recognition problem. I am trying to train a CNN with Keras. <\/p>\n\n<p>Here is how I prepared the input<\/p>\n\n<pre><\/pre>\n\n<p>And here is how I created and trained the model<\/p>\n\n<pre><\/pre>\n\n<p>The problem is I get very low accuracies that remain with the same value at each epoch<\/p>\n\n<p>Train on 13400 samples, validate on 6600 samples\nEpoch 1\/10\n13400\/13400 [==============================] - 78s - loss: 34.7407 - dense_740_loss: 0.1161 - dense_741_loss: 0.6879 - dense_742_loss: 4.7988 - dense_743_loss: 14.7893 - dense_744_loss: 14.3486 - dense_740_acc: 0.9902 - dense_741_acc: 0.9542 - dense_742_acc: 0.7001 - dense_743_acc: 0.0810 - dense_744_acc: 0.1055 - val_loss: 34.7760 - val_dense_740_loss: 0.0049 - val_dense_741_loss: 0.7131 - val_dense_742_loss: 4.8721 - val_dense_743_loss: 14.8091 - val_dense_744_loss: 14.3769 - val_dense_740_acc: 0.9997 - val_dense_741_acc: 0.9558 - val_dense_742_acc: 0.6977 - val_dense_743_acc: 0.0812 - val_dense_744_acc: 0.1080\nEpoch 2\/10\n13400\/13400 [==============================] - 70s - loss: 34.7032 - dense_740_loss: 0.0036 - dense_741_loss: 0.6760 - dense_742_loss: 4.7861 - dense_743_loss: 14.8118 - dense_744_loss: 14.4257 - dense_740_acc: 0.9998 - dense_741_acc: 0.9581 - dense_742_acc: 0.7031 - dense_743_acc: 0.0810 - dense_744_acc: 0.1050 - val_loss: 34.7760 - val_dense_740_loss: 0.0049 - val_dense_741_loss: 0.7131 - val_dense_742_loss: 4.8721 - val_dense_743_loss: 14.8091 - val_dense_744_loss: 14.3769 - val_dense_740_acc: 0.9997 - val_dense_741_acc: 0.9558 - val_dense_742_acc: 0.6977 - val_dense_743_acc: 0.0812 - val_dense_744_acc: 0.1080<\/p>\n","answers":[{"AnswerId":"42361266","CreationDate":"2017-02-21T07:38:07.230","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Set the learning rate of adam according to your model. Instead of using the default adam you can instantiate the optimizer (<a href=\"https:\/\/keras.io\/optimizers\/#adam\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/optimizers\/#adam<\/a>) yourself using:<\/p>\n\n<pre><code>from keras.optimizers import Adam\nadam = Adam(lr=0.1)\n<\/code><\/pre>\n\n<p>This seems promising, because your loss is actually decreasing, just only by a very low margin.<\/p>\n"},{"AnswerId":"42363290","CreationDate":"2017-02-21T09:22:06.073","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>I don't see what the problem is, your loss is decreasing. The accuracy can only start changing very late in the process, your optimization happens on the loss. \nIt might be also that your data set is biased, do you have the same number of samples in every classes ?<\/p>\n"}]}
{"QuestionId":42360167,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-21T06:28:34.973","AcceptedAnswerId":null,"OwnerUserId":6405712.0,"Title":"In what order does TensorFlow evaluate nodes in a computation graph?","Body":"<p>I am having a strange bug in TensorFlow. Consider the following code, part of a simple feed-forward neural network:<\/p>\n\n<pre><\/pre>\n\n<p>(Notice that prob is not used to define the loss function. This is because sigmoid_cross_entropy applies sigmoid internally in its definition)\nI later run the optimizer in the following line:<\/p>\n\n<pre><\/pre>\n\n<p>The above works just fine. However, if I instead run the following line to run the code, the network seems to perform terribly, even though there shouldn't be any difference!<\/p>\n\n<pre><\/pre>\n\n<p>I have a feeling it has something to do with the order in which TF computes the nodes in the graph during a session, but I'm not sure. What could the issue be?<\/p>\n","answers":[{"AnswerId":"42361435","CreationDate":"2017-02-21T07:46:28.790","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>It's not an issue with the graph, it's just that you are looking at different things.<\/p>\n\n<p>In the first example you provide:<\/p>\n\n<p><code>result,step_loss,_ = sess.run(fetches = [output,loss,optimizer],feed_dict = {x_ : np.array([[x,y,x*x,y*y,x*y]]), y_ : [[1,0]]})<\/code><\/p>\n\n<p>you are saving the result of running the <code>output<\/code> op in the <code>result<\/code> python variable.<\/p>\n\n<p>In the second one:<\/p>\n\n<p><code>result,step_loss,_ = sess.run(fetches = [prob,loss,optimizer],feed_dict = {x_ : np.array([[x,y,x*x,y*y,x*y]]), y_ : [[1,0]]})<\/code><\/p>\n\n<p>you are saving the result of the <code>prob<\/code> op in the <code>result<\/code> python variable.<\/p>\n\n<p>Since both ops are different it is to be expected that the values returned by them would be different.<\/p>\n\n<p>You could run <\/p>\n\n<p><code>logits, activation, step_loss, _ = sess.run(fetches = [output, prob, loss, optimizer], ...)<\/code><\/p>\n\n<p>to check your results.<\/p>\n"}]}
{"QuestionId":42360210,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-02-21T06:31:22.823","AcceptedAnswerId":null,"OwnerUserId":4439730.0,"Title":"the weights remain the same, even after training","Body":"<p>I am a beginner in machine learning. I am trying to train a 2 layered fully connected neural network. But after I complete the training, the weights and the biases are pretty much the same as before training. I am initialling them with random numbers.<\/p>\n\n<p>I also noticed that my cost function is returning the same value every time.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42360614,"AnswerCount":3,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-02-21T06:57:07.420","AcceptedAnswerId":42361083.0,"OwnerUserId":3608555.0,"Title":"What needs to be saved to reuse a model in TensorFlow","Body":"<p>I have started exploring TensorFlow library and trying out Image classification <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"nofollow noreferrer\">example<\/a> that uses MNIST data. I want the model to be stored in a file after training phase is over, so that I can use it as and when required. I have checked <a href=\"https:\/\/stackoverflow.com\/questions\/38000180\/save-tensorflow-model-to-file\">this link<\/a> that tells about how to save values from TensorFlow to any file and read that as well. So far, I am able to save some variables from the script to file using pickle as suggested in the link. But, I am not able grasp on what needs to saved in file to store current state of the model for its later use. \nPlease, someone could explain that part with example on storing a model and loading that model.<\/p>\n","answers":[{"AnswerId":"42381923","CreationDate":"2017-02-22T03:02:45.933","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>Only <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"nofollow noreferrer\">Variables<\/a> can be saved and restored. When you need to reuse the saved variables you need to first build the graph by creating the neural network and setting the NN's parameters like layer numbers, learning rate and dropout and etc. The only values restored from a checkpoint are the Variables defined in the training process. You can take a look at any example, for example <a href=\"https:\/\/github.com\/nicolas-ivanov\/tf_seq2seq_chatbot\/blob\/a4827727bbac5be87a2ba7980492f5c88b0e6c2f\/tf_seq2seq_chatbot\/lib\/seq2seq_model_utils.py\" rel=\"nofollow noreferrer\">this one<\/a>. <\/p>\n\n<p>To sum up, only Variables can and need to be saved and restored, neural network configures and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"nofollow noreferrer\">placeholders<\/a> can not. <\/p>\n"},{"AnswerId":"42360975","CreationDate":"2017-02-21T07:20:53.643","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>First, you should check out this <a href=\"https:\/\/stackoverflow.com\/questions\/33759623\/tensorflow-how-to-restore-a-previously-saved-model-python\">other question<\/a>.<\/p>\n\n<p>TensorFlow has methods implemented for managing both saving and restoring <em>checkpoints<\/em>, specifically, the <code>tf.train.saver<\/code> class. Check out the official documentation <a href=\"http:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"nofollow noreferrer\">here<\/a>. Checkpoints basically store the values of your tensors (among other things) in disk.<\/p>\n\n<p>Citing the documentation:<\/p>\n\n<blockquote>\n  <p>Checkpoints are binary files in a proprietary format which map variable names to tensor values. The best way to examine the contents of a checkpoint is to load it using a <code>Saver<\/code>.<\/p>\n<\/blockquote>\n"},{"AnswerId":"42361083","CreationDate":"2017-02-21T07:26:22.700","ParentId":null,"OwnerUserId":"7045779","Title":null,"Body":"<p>To save and restore variables in Tensorflow, the following things are needed.<\/p>\n\n<p>1) A list of variables to be saved and restored\n2) tf.train.Saver<\/p>\n\n<p>Generally, 1) is achieved by<\/p>\n\n<pre><code># To save and restore whole tf variables\nall_vars = tf.global_variables()\n<\/code><\/pre>\n\n<p>or,<\/p>\n\n<pre><code># To save and restore the specific tf variables using scope\nall_vars = tf.global_variables()\nmodel_vars = [k for k in all_vars if k.name.startswith(\"xxx\")]\n# \"xxx\" is the expected scope\n<\/code><\/pre>\n\n<p>Then, 2) is achieved by<\/p>\n\n<pre><code>saver = tf.train.Saver(vars_list)\n# vars_list is list of variables from above\n<\/code><\/pre>\n\n<p>Finally, to save variables, (with tf.Session() running with named 'sess')<\/p>\n\n<pre><code>saver.save(sess, '\/directory\/to\/chechpoint\/file.ckpt')\n<\/code><\/pre>\n\n<p>and to restore them,<\/p>\n\n<pre><code>saver.restore(sess, '\/directory\/to\/chechpoint\/file.ckpt')\n<\/code><\/pre>\n"}]}
{"QuestionId":42360774,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-21T07:06:55.753","AcceptedAnswerId":null,"OwnerUserId":6215726.0,"Title":"can tensorflow build spp-net model for different input image size?","Body":"<p>my input img height is fixed but the width  range from 10 to 100000(mean is 2000).\nI wanted use spp-net but tf.nn.conv2d dosen't support different size.\nIs there any way to handle this problem? <\/p>\n","answers":[{"AnswerId":"42361466","CreationDate":"2017-02-21T07:48:27.853","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>It depends on what you want to do, but conv2d DOES support different sizes. If your placeholder has shape [None, None, None, 3] for example, you can pass this as an argument to a conv2d operation (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d<\/a>). <\/p>\n\n<p>For your application, where you apply SPP after the convolutional layers, this should work!<\/p>\n"}]}
{"QuestionId":42361275,"AnswerCount":1,"Tags":"<neural-network><computer-vision><deep-learning><caffe>","CreationDate":"2017-02-21T07:38:26.800","AcceptedAnswerId":42361677.0,"OwnerUserId":2467772.0,"Title":"Feature extraction using caffe model","Body":"<p>I'd like to discuss feature extraction using Caffe model called GoggleNet.\nI am referring to this paper <a href=\"https:\/\/arxiv.org\/abs\/1506.04878\" rel=\"nofollow noreferrer\">\"End to end people detection in crowded scenes\"<\/a>. For those who are familiar with caffe, should be able to cope with my queries.<\/p>\n\n<p>The paper has its own <a href=\"https:\/\/github.com\/Russell91\/ReInspect\" rel=\"nofollow noreferrer\">library<\/a> using Python, I also run through the library but can't cope with some points mentioned in the paper.<\/p>\n\n<p>The input image is passed through with  layer. <\/p>\n\n<p>Then output is formed as multidimensional array in 15x20x1024. So each 1024 vector represents a bounding box in the center of 64x64 region. Since it is 50% overlapping, there are 15x20 matrix for 640x480 image and each cell has third dimension of 1024 vector in length.<\/p>\n\n<p>My query is <\/p>\n\n<p>(1)how this 15x20x1024 array output can be obtained?<\/p>\n\n<p>(2)how this 1x1x1024 data can represent 64x64 region in the image?\nThere is a description in the source code as<\/p>\n\n<pre><\/pre>\n\n<p>That conversion is implemented using the function in Python as<\/p>\n\n<pre><\/pre>\n\n<p>I can't cope that portion as how each 1x1x1024 represents that size of bounding box rectangle.<\/p>\n","answers":[{"AnswerId":"42361677","CreationDate":"2017-02-21T08:00:11.650","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Since you are looking at a 1x1 cell very deep in the net, it's effective <a href=\"https:\/\/www.quora.com\/What-is-a-receptive-field-in-a-convolutional-neural-network\" rel=\"nofollow noreferrer\">recptive field<\/a> is quite large and can be (and probably is) 64x64 pixels in the original image.<br>\nThat is, each feature in <code>\"inception_5b\/output\"<\/code> is affected by 64x64 pixels in the input image.<\/p>\n"}]}
{"QuestionId":42361500,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><pycaffe><matconvnet>","CreationDate":"2017-02-21T07:50:24.890","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"How to design a CNN (the inputs and outputs of the layers)?","Body":"<p>I am a beginner and learning deep learning with baby steps. I have a question about designing the nets. I see in the papers, there are layers with different inputs\/outputs and I do not know how to calculate\/design before implementation. \nFor instance, in this <a href=\"https:\/\/people.eecs.berkeley.edu\/~jonlong\/long_shelhamer_fcn.pdf\" rel=\"nofollow noreferrer\">paper<\/a>, there are some numbers beside the schematic layers output (see the following figure). How these filter size and other parameters are being specified for a network with a specific image size as input. \n<a href=\"https:\/\/people.eecs.berkeley.edu\/~jonlong\/long_shelhamer_fcn.pdf\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/V6L42.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>or in another paper, they have the following design:\n<a href=\"https:\/\/i.stack.imgur.com\/qFjQI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qFjQI.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>and they have mentioned, For a  input image,\nthe total sub-sampling factor of the network is , resulting in\na , where L is the number of class labels. How this  size is being obtained?<\/p>\n\n<p>How can I learn to design the net and calculates inputs\/outputs of layers?<\/p>\n\n<p>Thank you for any help<\/p>\n","answers":[{"AnswerId":"42361797","CreationDate":"2017-02-21T08:07:40.937","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<ol>\n<li><p>If you pool twice with <code>stride=2<\/code> means you reduce by a factor 2 the image size twice, resulting with a total of x4 reduction (sub sampling) of image size. Hence, if you start with an image of size 256: 256\/4=64.<\/p><\/li>\n<li><p>How do choose kernel size, number of output of each layer, strides and other design parameters? There's actually no single answer for that and basically many papers\/works approach the same tasks with different settings. AFAIK there's no clear guidelines or obvious choice of parameters that suits any specific task.<br>\nThat being said, you can find <a href=\"https:\/\/arxiv.org\/abs\/1611.00847\" rel=\"nofollow noreferrer\">this work<\/a> surveying some emerging deep nets design patterns.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42361513,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-21T07:51:40.253","AcceptedAnswerId":42386311.0,"OwnerUserId":5703903.0,"Title":"What does :0 behind names in TensorFlow mean?","Body":"<p>With reference to the tutorial <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variable_scope\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>What does the  mean?<\/p>\n\n<p>As I see it, it only appears on variables but not on ops. Is it possible for this to be :1 or even :2?<\/p>\n","answers":[{"AnswerId":"42386311","CreationDate":"2017-02-22T08:31:28.377","ParentId":null,"OwnerUserId":"2595935","Title":null,"Body":"<p>If you search for <code>:%d<\/code> in the codebase, you will find not only the variable has <code>:0<\/code> but also ops: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/ops.py#L317\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/framework\/ops.py#L317<\/a><\/p>\n\n<p>The <code>:0<\/code> simply means the first output of that node. It is possible to have <code>:1<\/code> or <code>:2<\/code> in some cases.<\/p>\n"}]}
{"QuestionId":42361668,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-21T08:00:01.523","AcceptedAnswerId":null,"OwnerUserId":2309451.0,"Title":"Tensorflow bidirectional_dynamic_rnn() FailedPreconditionError: Attempting to use uninitialized value BiRNN\/FW\/LSTMCell\/B","Body":"<p>I'm getting the above error when attempting to call . I've called . At first I thought it's because I didn't pass in  to . However, even after I did, it's still shows the same error.<\/p>\n\n<p>Any idea?<\/p>\n\n<p>Stacktrace:<\/p>\n\n<pre><\/pre>\n\n<p>Code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42383465","CreationDate":"2017-02-22T05:35:21.313","ParentId":null,"OwnerUserId":"2309451","Title":null,"Body":"<p>Never mind. It turns out I didn't initialize the variables properly..<\/p>\n\n<p>I moved this line: <\/p>\n\n<p><code>pred = BidirectionalLSTMEncoder().add_prediction_op(input_placeholder, 6)<\/code><\/p>\n\n<p>above this line:<\/p>\n\n<p><code>with tf.Session() as session:<\/code><\/p>\n\n<p>Then everything worked fine. The functions containing variable initialization need to be called before <code>tf.global_variables_initializer()<\/code><\/p>\n"}]}
{"QuestionId":42361700,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2017-02-21T08:02:02.493","AcceptedAnswerId":null,"OwnerUserId":1735308.0,"Title":"caffe segnet deep observation wrong batch layer parameters","Body":"<p>I created a caffe model on top of segnet\/deepobservationEO\nI compiled caffe successfully. I used the model proposed here: \n<a href=\"https:\/\/github.com\/nshaud\/DeepNetsForEO\/blob\/master\/models\/segnet_isprs_vaihingen_irrg.prototxt\" rel=\"nofollow noreferrer\">https:\/\/github.com\/nshaud\/DeepNetsForEO\/blob\/master\/models\/segnet_isprs_vaihingen_irrg.prototxt<\/a><\/p>\n\n<p>I get that issue:<\/p>\n\n<pre><\/pre>\n\n<p>any idea about what to do, where to check?\nthx\ncolin<\/p>\n","answers":[{"AnswerId":"42363792","CreationDate":"2017-02-21T09:45:05.357","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like you have <code>lr_mult: 0.5<\/code> for layer <code>conv1_1_bn<\/code> of type <code>\"BatchNorm\"<\/code>.<br>\nCaffe only supports <code>lr_mult: 0<\/code> for layer <code>\"BatchNorm\"<\/code>.<\/p>\n"}]}
{"QuestionId":42362532,"AnswerCount":4,"Tags":"<numpy><tensorflow>","CreationDate":"2017-02-21T08:47:46.473","AcceptedAnswerId":null,"OwnerUserId":7597359.0,"Title":"'numpy.ndarray' object has no attribute 'train'","Body":"<p>how can i solve this ? this is my first time for Tensortflow. I try to copy Train and Evaluate the Model from tensortflow tutorial but it seem not work. Can someone help me to solve my problem? Thanks!<\/p>\n\n<p><a href=\"http:\/\/pastebin.com\/NCQKNyKy\" rel=\"nofollow noreferrer\">http:\/\/pastebin.com\/NCQKNyKy<\/a><\/p>\n\n<pre><\/pre>\n\n<p>This is result: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"45858828","CreationDate":"2017-08-24T10:08:37.747","ParentId":null,"OwnerUserId":"8484697","Title":null,"Body":"<p>You may try to use numpy.reshape to turn your data from 2 dimension into 3 dimension. \nFor example if you had 20 samples and 100 features, so a (20,100) data matrix and used a minibatch size of 5. Then you could reshape using np.reshape(data,[10,5,-1]) to get a (10,5,40) matrix. <\/p>\n\n<p>*The \"-1\" meaning that you leave numpy to count the array for your, the total number of array is 20,000. \nThus, in this example: 10*5*40 = 20,000.<\/p>\n"},{"AnswerId":"42367339","CreationDate":"2017-02-21T12:27:03.703","ParentId":null,"OwnerUserId":"4729700","Title":null,"Body":"<p>It is not quite clear what you are trying to do. The problem occurs because data is a numpy array generated in this line<\/p>\n\n<pre><code>data = genfromtxt('circle_deeplearn_data_small.txt',delimiter=',')\n<\/code><\/pre>\n\n<p>The error occurs when you try to use the method train of data, which does not exist, in the following line<\/p>\n\n<pre><code>batch = data.train.next_batch(50)\n<\/code><\/pre>\n\n<p>Instead you need to feed data to tensorflow.<\/p>\n"},{"AnswerId":"53054873","CreationDate":"2018-10-29T22:34:22.230","ParentId":null,"OwnerUserId":"2256217","Title":null,"Body":"<p>I have faced same problem. Actually, it's not a problem. Literally, I didn't know the structure of the data that's why I have faced this problem. Te datasets comes from <code>tensorflow<\/code> lib are compressed in a single file and separated in a file as <code>train, test<\/code>, and <code>validation<\/code> set. That's why when we call <code>dataset.train.next_batch()<\/code> it does work. You own datatset is not compressed in the same way that's why it doesn't work. You have to configure your dataset on the own way so do the batch system and looping.<\/p>\n"},{"AnswerId":"45791967","CreationDate":"2017-08-21T07:59:36.990","ParentId":null,"OwnerUserId":"8484697","Title":null,"Body":"<p>here <code>data<\/code>is just a numpy array. You may need to write ur own train data iterator <\/p>\n"}]}
{"QuestionId":42362542,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-21T08:48:18.713","AcceptedAnswerId":null,"OwnerUserId":5037646.0,"Title":"How to monitor dead relus","Body":"<p>I think I might have a problem with dead Relus, but I don't really know how to check it with tensorboard or any other way. Your help would be really appreciated.<\/p>\n","answers":[{"AnswerId":"47851863","CreationDate":"2017-12-17T03:03:17.003","ParentId":null,"OwnerUserId":"2122215","Title":null,"Body":"<p>I initially had this same question myself and couldn't find an answer, so here's how I'm doing it with Tensorboard (This assumes some familiarity with Tensorboard).<\/p>\n\n<pre><code>activation = tf.nn.relu(layer) \nactive = tf.count_nonzero(tf.count_nonzero(activation, axis=0))\ntf.summary.scalar('pct-active-neurons', active \/ layer.shape[1])\n<\/code><\/pre>\n\n<p>In this snip, <code>activation<\/code> is my post-ReLU activation for this particular layer. The first call to <code>tf.count_nonzero(out, axis=0)<\/code> is counting how many activations each neuron has seen across all training examples for the current step of training. The second call <code>tf.count_nonzero( ... )<\/code> that wraps the first call counts how many neurons in the layer had at least one activation for the batch of training examples for this step. Finally, I convert it to a percentage by dividing the number of neurons that had at least one activation in the training step by the total number of neurons for the layer.<\/p>\n\n<p>More information on setting up Tensorboard can be found <a href=\"https:\/\/www.tensorflow.org\/get_started\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"48782899","CreationDate":"2018-02-14T08:44:03.267","ParentId":null,"OwnerUserId":"5681083","Title":null,"Body":"<p>How about adding the summary \n<code>\ntf.summary.scalar('%s\/fraction_of_zero_values' % tag, tf.nn.zero_fraction(value))\n<\/code><\/p>\n"}]}
{"QuestionId":42363749,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-21T09:42:56.207","AcceptedAnswerId":null,"OwnerUserId":7261750.0,"Title":"In tensorflow, when I use DNNLinearClassifier [wide_n_deep_tutorial.py], how to restore the history model trained out and train base on it","Body":"<p>I was learning the Google <a href=\"https:\/\/www.tensorflow.org\/tutorials\/wide_and_deep\" rel=\"nofollow noreferrer\">wide and deep model<\/a> in tensorflow.<\/p>\n\n<p>My code was modified based on the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/wide_n_deep_tutorial.py\" rel=\"nofollow noreferrer\">tf example<\/a>.<\/p>\n\n<p>My question is: <\/p>\n\n<ol>\n<li><p>How to load history model, and train new sample based on it. I had search many on the web, but most methods are like the following code.<\/p>\n\n<p>Examples of saver<\/p>\n\n<pre><\/pre>\n\n<p>But in  train, it define a method \"build_estimator\" which returns:<\/p>\n\n<p>. <\/p>\n\n<p>When I use the method above to define a saver, it gave error no Varibles to save. And the \"m\" returned have no save, restore method.<\/p>\n\n<p>I have also try  in , but it also doesn't work.<\/p><\/li>\n<li><p>Second is related with above, how to save model. Or how to use model which the fit method saved in \/tmp\/. (checkpoint)<\/p><\/li>\n<\/ol>\n\n<p>My question can be summarized as:<\/p>\n\n<p>How to save model and restore history model when I use the   in tensorflow?<\/p>\n","answers":[{"AnswerId":"42380542","CreationDate":"2017-02-22T00:27:47.473","ParentId":null,"OwnerUserId":"2595935","Title":null,"Body":"<p>There is an argument called <code>model_dir<\/code> in <code>DNNLinearCombinedClassifier<\/code>: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn_linear_combined.py#L476\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/dnn_linear_combined.py#L476<\/a>. Just point it to the directory where you store your model checkpoint.<\/p>\n"}]}
{"QuestionId":42364283,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-21T10:05:08.347","AcceptedAnswerId":null,"OwnerUserId":1660762.0,"Title":"Tensorflow: calculate gradient for tf.multiply","Body":"<p>I'm building a neural network that has the following two layers<\/p>\n\n<pre><\/pre>\n\n<p>I then want to multiply them using  (which, unlike  multiplies corresponding indices, i.e. c_ij = a_ij * b_ij)<\/p>\n\n<pre><\/pre>\n\n<p>My goal is to learn . So I run<\/p>\n\n<pre><\/pre>\n\n<p>But it doesn't work. The network doesn't change at all.<\/p>\n\n<p>Looking at tensorboard, I could see that 'input' has no gradient, so I'm assuming that's the problem. Any ideas how to solve this?<\/p>\n\n<p>From reading tensorflow docs it seems like I might have to write a gradient op for tf.multiply, but I find it hard to believe no one needed to do this before.<\/p>\n","answers":[{"AnswerId":"42368769","CreationDate":"2017-02-21T13:34:12.470","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>I thought the <code>pseudo_inputs<\/code> should be set as Placeholders in the first line. <\/p>\n\n<p>And in this line: <\/p>\n\n<pre><code>train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss, var_list=[weights])\n<\/code><\/pre>\n\n<p>Since weights are to be trained in the graph by minimizing loss then it should not passed as a parameter here.<\/p>\n\n<pre><code>train = tf.train.AdamOptimizer(learn_rate).minimize(loss)\n<\/code><\/pre>\n\n<p>Then you should first run the train using the samples(<strong>you don't have labels<\/strong>) you have. <\/p>\n\n<pre><code>for x_train, y_train in samples:\n    sess.run(train, {pseudo_inputs:x_train, y:y_train})\n<\/code><\/pre>\n\n<p>And after that you can get weights by:<\/p>\n\n<pre><code>W_c, loss_c  = sess.run([W, loss], {pseudo_inputs=x_train, y:y_train})\n<\/code><\/pre>\n"}]}
{"QuestionId":42364513,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><deep-learning><image-recognition>","CreationDate":"2017-02-21T10:16:27.750","AcceptedAnswerId":null,"OwnerUserId":3425890.0,"Title":"How to locate multiple objects in the same image?","Body":"<p>I am a newbie in TensorFlow.<\/p>\n\n<p>Currently, I am testing some classification's examples \"Convolutional Neural Network\" in the TensorFlow website, and it explains how to classify input images into pre-defined classes, but the problem is: I can't figure out how to locate multiple objects in the same image. For example, I had an input image with a cat and dog and I want my graph to display in the output that there are both of them \"a cat and a dog\" in the image.<\/p>\n","answers":[{"AnswerId":"45159411","CreationDate":"2017-07-18T06:45:11.203","ParentId":null,"OwnerUserId":"6323882","Title":null,"Body":"<p>A naive approach for what you are trying to do would be to classify parts of the image independently.<\/p>\n\n<p>But there are some better techniques for object detection. Actually, there is the  <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/object_detection%22Object%20Detection%20API%22\" rel=\"nofollow noreferrer\">TensorFlow Object Detection API<\/a>, which gives you access to the most common object detection methods like Faster R-CNN or SSD.<\/p>\n"},{"AnswerId":"45159189","CreationDate":"2017-07-18T06:34:03.593","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Great question. Detecting multiple objects in the same image boils is essentially a \"segmentation problem\". Two nice and popular algorithms are \nYOLO (You Only Look Once), and SSD(Single Shot Multibox Detector). I included links to them at the bottom. <\/p>\n\n<p>I would watch a few videos on how YOLO works, and see if you grasp the idea. Then read the paper on SSD, and see if you get why this algorithm is even faster and more precise. <\/p>\n\n<p>Both algorithms are single-pass: they only look at the image \"once\" and predict bounding boxes for the categories they spot. There are more precise algorithms, but they are slower (they first pick many spots they want to look, and then run a classifier on only that spot. The result is that they run this classifier many times per image, which is slow). <\/p>\n\n<p>As you stated you are a newbie to Tensorflow, you can try this code other people made: <a href=\"https:\/\/github.com\/thtrieu\/darkflow\" rel=\"noreferrer\">https:\/\/github.com\/thtrieu\/darkflow<\/a> . The very extensive readme shows you how to get started on your own dataset. <\/p>\n\n<p>Good luck, and let us know if you have other questions, or if these algorithms do not fit your use-case. <\/p>\n\n<ul>\n<li>YOLO 9000 (<a href=\"https:\/\/pjreddie.com\/darknet\/yolo\/\" rel=\"noreferrer\">https:\/\/pjreddie.com\/darknet\/yolo\/<\/a>)<\/li>\n<li>SSD (Single shot multibox detector) (<a href=\"https:\/\/arxiv.org\/abs\/1512.02325\" rel=\"noreferrer\">https:\/\/arxiv.org\/abs\/1512.02325<\/a>)<\/li>\n<\/ul>\n"}]}
{"QuestionId":42364629,"AnswerCount":4,"Tags":"<python><python-2.7><tensorflow>","CreationDate":"2017-02-21T10:21:44.120","AcceptedAnswerId":57644594.0,"OwnerUserId":2847505.0,"Title":"Tensorflow - casting from int to float strange behavior","Body":"<p>I am working on tensorflow 0.12 and am having problem with casting.\nThe following snippet of code does a strange thing:<\/p>\n\n<pre><\/pre>\n\n<p>I get a value:\n6.86574233e-36<\/p>\n\n<p>I also tried using  and . Both gave the same result.<\/p>\n\n<p>Please help.<\/p>\n","answers":[{"AnswerId":"42558206","CreationDate":"2017-03-02T14:33:40.757","ParentId":null,"OwnerUserId":"2427065","Title":null,"Body":"<p>No error in your program.<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.InteractiveSession()\na = tf.constant(1)\nb = tf.cast(a, tf.float32)\nprint b.eval()\n<\/code><\/pre>\n\n<p>This is an online environment for TF <a href=\"https:\/\/codeenv.com\/env\/run\/gXGpnR\/\" rel=\"nofollow noreferrer\">https:\/\/codeenv.com\/env\/run\/gXGpnR\/<\/a><br>\nTest your code there to run, use <\/p>\n\n<ul>\n<li>click on test_tf.py<\/li>\n<li>add your code<\/li>\n<li>in left side CLI, type ipython test_tf.py<\/li>\n<\/ul>\n"},{"AnswerId":"42527737","CreationDate":"2017-03-01T09:18:56.930","ParentId":null,"OwnerUserId":"3805227","Title":null,"Body":"<p>I checked the code in python3 and python2 for the same tensorflow version as well the code seems to be working correctly as in both the cases I got the following output for python2<\/p>\n\n<pre><code>print b.eval()\n1.0\n<\/code><\/pre>\n\n<p>I would suggest checking the tensorflow installation or the virtualenv.<\/p>\n"},{"AnswerId":"42364757","CreationDate":"2017-02-21T10:27:58.690","ParentId":null,"OwnerUserId":"325241","Title":null,"Body":"<pre><code>sess = tf.InteractiveSession()\na = tf.constant(1, tf.int64)  &lt;--------\nb = tf.cast(a, tf.float32)\nprint b.eval()  # 1.0\n<\/code><\/pre>\n\n<p>You need to declare the <code>dtype<\/code> for your <code>tf.constant<\/code>: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant<\/a><\/p>\n"},{"AnswerId":"57644594","CreationDate":"2019-08-25T08:52:26.613","ParentId":null,"OwnerUserId":"2847505","Title":null,"Body":"<p>Since I see that this is still getting some attention, I should mention that the newer versions of tensorflow do not show this behavior, I suggest working with tensorflow version 1.13 or higher<\/p>\n"}]}
{"QuestionId":42364938,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-21T10:36:23.510","AcceptedAnswerId":null,"OwnerUserId":2768132.0,"Title":"Draw a graph using Tensorflow","Body":"<p>I am very new to tensorflow. I have installed Tensorflow CPU version and also installed python 3.5.2. I have also created a simple program using tensorflow like some addition of two numbers etc. Now i want to create a graphical program using python and tensorflow. Is it possible? any simple program will be helpful(like draw a line, draw a graph etc..). Kindly mention the steps. i have tried lots of examples but its not working.<\/p>\n\n<p>Thanks in advance!!<\/p>\n","answers":[{"AnswerId":"42365418","CreationDate":"2017-02-21T10:57:18.560","ParentId":null,"OwnerUserId":"5903959","Title":null,"Body":"<p>TensorFlow is not here to draw anything. Like, really not. It is built to compute stuff, over CPU and GPU and to give a Machine Learning library.<\/p>\n\n<p>Now, you may have been confused because TensorFlow often mention \"graphs\". Those graphs are computational graphs, which is how each operations are linked one to another. In fact, \"graphs\" refers to a set of <a href=\"https:\/\/en.wikipedia.org\/wiki\/Graph_theory\" rel=\"nofollow noreferrer\">vertice and edges<\/a>, which can be drawn, but is not a \"graphical\" object in itself. Still, you can see this graph, generated by TensorFlow, thanks to <a href=\"https:\/\/www.tensorflow.org\/get_started\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">Tensorboard<\/a>.<\/p>\n\n<p>Now, what is totally possible is, in one hand to run intensive computations thanks to TensorFlow AND, in the other hand, another librarie to draw, plot etc...<\/p>\n\n<p>The most famous\/documented\/used\/best by far is <a href=\"http:\/\/matplotlib.org\/api\/pyplot_api.html\" rel=\"nofollow noreferrer\">matplotlib, aka pyplot<\/a>. This is the lib to use, the lib to learn, as you will encounter it everywhere. In fact, I don't know any other lib. <\/p>\n\n<p>Tell me if my answer is somehow off topic, or if you would like me to precise something.<br>\nHave fun!<\/p>\n"}]}
{"QuestionId":42365158,"AnswerCount":0,"Tags":"<numpy><multidimensional-array><tensorflow><keras>","CreationDate":"2017-02-21T10:45:50.603","AcceptedAnswerId":null,"OwnerUserId":7544784.0,"Title":"How to transform tensors to numpy.ndarray in Keras?","Body":"<p>I have got a Tensorflow Tensor within Keras (&lt; class 'tensorflow.python.framework.ops.Tensor'>) which needs to be transformed into a numpy.ndarray. \nIs there an easy solution to this task? <\/p>\n\n<p>I did time-comsuming research via Google and trial\/error. \nIt seems that I have to use tensorflow sessions, but I don't wan't to use any tensorflow sessions, because using them I will receive errors like:<\/p>\n\n<pre><\/pre>\n\n<p>I need this placeholder in line 113 - it doesn't make troubles when using Keras without any tensorflow session. But as soon as I am trying to use a tensorflow session (in order to use .eval() for transforming the tensor into numpy.ndarray) my code explodes...<\/p>\n\n<p>I appreciate your help!<\/p>\n","answers":[]}
{"QuestionId":42365342,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-21T10:54:00.230","AcceptedAnswerId":null,"OwnerUserId":1346739.0,"Title":"Tensorflow - hello world program, prepends b to output","Body":"<p>After installing tensorflow and running the short program following instructions at <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#run_a_short_tensorflow_program\" rel=\"nofollow noreferrer\">tensorflow site<\/a><\/p>\n\n<pre><\/pre>\n\n<p>The expected result is \n\nBut I get <\/p>\n\n<p>Note that the result differ in that  is prepended and ouput itself is in single qoutes.<\/p>\n\n<p>Is there something that I am missing? I had earlier done the same example on Windows (Anaconda) and got same result. I installed only the CPU one.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[]}
{"QuestionId":42365367,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-21T10:55:11.833","AcceptedAnswerId":null,"OwnerUserId":4895074.0,"Title":"Custom RMSPE loss function in keras","Body":"<p>I am trying to define my own loss function in keras which is Root Mean Squared Percentage Error.  RMSPE is defined as :  <br><img src=\"https:\/\/latex.codecogs.com\/gif.latex?%5Csqrt%7B%5Cfrac%7B1%7D%7Bn%7D%5Csum_%7Bi%3D1%7D%5E%7Bn%7D%5Cleft%20%28%20%5Cfrac%7Byi-%5Chat%7By%7D%7D%7Byi%7D%20%5Cright%20%29%5E%7B2%7D%7D\" alt=\"equation\">  <br>I have defined my loss function as:<br><br> But after a few iterations it is giving me loss value as nan. Can someone point out what am i doing wrong. Thanks<\/p>\n","answers":[{"AnswerId":"45777419","CreationDate":"2017-08-20T00:25:45.933","ParentId":null,"OwnerUserId":"8489353","Title":null,"Body":"<p>When your denominator is 0, that will be Nan. Consider plus a small number on your denominator, like 0.0000001.<\/p>\n"}]}
{"QuestionId":42365547,"AnswerCount":0,"Tags":"<python><numpy><asynchronous><parallel-processing><tensorflow>","CreationDate":"2017-02-21T11:03:23.003","AcceptedAnswerId":null,"OwnerUserId":199556.0,"Title":"Pre load huge numpy arrays in parallel when train or predict with tensorflow","Body":"<p>First of all I'm using numpy arrays as initial data and I'm not going to change that.<\/p>\n\n<p>So each file contains separated example(let's say it's bigger than 100Mb). I do my logic on it in batches(kind of sliding windows). Then I need to save result for that sample.<\/p>\n\n<p>I've noticed that reading that file from disk adds noticeable delay(especially when you predict). <\/p>\n\n<p>My idea is to load next file in parallel with the training\/prediction process so that delay becomes negligible. \nProbably I should use FIFOQueue, but can't put it together.<\/p>\n\n<p>Schematically the whole pipeline can be represented as <\/p>\n\n<p>Read numpy array => Split it in chunks => Process each chunk(batch) => Store that slice in a new numpy array => Save predicted result for the whole file<\/p>\n","answers":[]}
{"QuestionId":42367281,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2017-02-21T12:24:37.323","AcceptedAnswerId":null,"OwnerUserId":4086817.0,"Title":"Tensorflow Android : retrained Inception v3 take too much time","Body":"<p>I have retrained the Inception-V3 final layer with my own 20 categories. When I am using retrained model in android demo app it takes 6 to 8 seconds to predict.<\/p>\n\n<p>Running on<\/p>\n\n<ul>\n<li>LG G4 Stylus  ->  6-8 sec<\/li>\n<li>S6, ->  3-4.5 sec<\/li>\n<\/ul>\n\n<p>I have done  it takes 6-9 sec and  it takes 7-11 sec. Is there any way to improve it?<\/p>\n\n<p>Output on LG G4 Stylus:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lKryY.png\" rel=\"nofollow noreferrer\">Output<\/a><\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I have followed <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">this<\/a><\/p>\n","answers":[{"AnswerId":"42389688","CreationDate":"2017-02-22T11:14:16.130","ParentId":null,"OwnerUserId":"7604321","Title":null,"Body":"<p>Comparing your debug output to the normal TF Classify app on my phone I see that you have a much larger node count which would suggest that, for some reason, your graph is a lot larger than it should be. I'm not too familiar with the quantize method but it looks like you have more conv2D layers than normal as well.<\/p>\n\n<p>Without any further information it's hard to say but I think that you should rebuild the graph and check that you've added the final layer properly.<\/p>\n"}]}
{"QuestionId":42367314,"AnswerCount":1,"Tags":"<image><rotation><caffe><detection>","CreationDate":"2017-02-21T12:26:07.370","AcceptedAnswerId":42395178.0,"OwnerUserId":6129428.0,"Title":"How to mark detected objects in a new image?","Body":"<p>I'm trying to detect multiple vehicles in satellite and aerial images. I have two main questions:<\/p>\n\n<p>1- After training the convolution network and getting the caffe model, how could I test it on a new image and mark the detected vehicles with sth like bounding boxes? Should I change the size of data blob to be able to use commands like this?<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XH5UX.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XH5UX.jpg\" alt=\"satellite image\"><\/a><\/p>\n\n<p>2- As you know the vehicles on the streets have different angles. Are deep learning techniques already rotation invariant? If not what can I do to deal with object's angles which can vary to 360 degrees?<\/p>\n\n<p>I would appreciate if anyone guide me through this.<\/p>\n","answers":[{"AnswerId":"42395178","CreationDate":"2017-02-22T15:06:44.187","ParentId":null,"OwnerUserId":"7134831","Title":null,"Body":"<p>You can use <a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\" rel=\"nofollow noreferrer\">faster R-CNN<\/a> base on caffe to train a vehicle detection model.<\/p>\n\n<ol>\n<li><p>Different image sizes can be input in the faster R-CNN framework, and there is <a href=\"https:\/\/github.com\/rbgirshick\/py-faster-rcnn\/blob\/master\/tools\/demo.py\" rel=\"nofollow noreferrer\">demo code<\/a> for you to reference.<\/p><\/li>\n<li><p>Because there are different angles vehicle in your training data, the trained model have the capacity to recognize these in new images.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42367538,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><neural-network><deep-learning><recurrent-neural-network>","CreationDate":"2017-02-21T12:36:01.643","AcceptedAnswerId":42369335.0,"OwnerUserId":6317571.0,"Title":"Tensorflow. Switching from BasicRNNCell to LSTMCell","Body":"<p>I have built a RNN with BasicRNN now I want to use the LSTMCell but the passage does not seem trivial. What should I change?<\/p>\n\n<p>First i define all the placeholders and variables: <\/p>\n\n<pre><\/pre>\n\n<p>Then I unstack the labels: <\/p>\n\n<pre><\/pre>\n\n<p>Then i define my RNN: <\/p>\n\n<pre><\/pre>\n\n<p>The error that I get is: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"50050475","CreationDate":"2018-04-26T19:18:52.980","ParentId":null,"OwnerUserId":"9706545","Title":null,"Body":"<p>The problem seems to be with the <code>init_state<\/code> variable.\nBasic RNN cells only have one state variable while LSTM has a visible and a hidden state. Specify the options <code>state_is_tuple=False<\/code> will concat the two state variables into one, therefore double the size of what you have specified in the init_state declaration. \nTo avoid this, one can use the built-in <code>zero_state<\/code> method for an LSTMCell to initialize the state in the correct way without worrying about size differences.\nSo it would simply be:<\/p>\n\n<pre><code>init_state = cell.zero_state(batch_size, dtype)\n<\/code><\/pre>\n\n<p>Of course will will have to be placed after the line where the cell is built.<\/p>\n"},{"AnswerId":"42369335","CreationDate":"2017-02-21T13:57:37.387","ParentId":null,"OwnerUserId":"5903959","Title":null,"Body":"<p>You should consider giving the error trace. Otherwise it is hard (or impossible) to help.<\/p>\n\n<p>I reproduced the situation and found that the issue was coming from state unpacking, i.e. line <code>c, h = state<\/code>.<\/p>\n\n<p>Try to set <code>state_is_tuple<\/code> to false i.e.<\/p>\n\n<pre><code>cell = tf.contrib.rnn.BasicLSTMCell(state_size, state_is_tuple=False)\n<\/code><\/pre>\n\n<p>I'm not sure why this is happening. Are you loading a previous model? What is your tensorflow version?<\/p>\n\n<hr>\n\n<p><strong>More information on TensorFlow RNN Cells:<\/strong><\/p>\n\n<p>I would suggest you to take a look at: <a href=\"http:\/\/www.wildml.com\/2016\/08\/rnns-in-tensorflow-a-practical-guide-and-undocumented-features\/\" rel=\"nofollow noreferrer\">WildML post<\/a>, section \"<em>RNN CELLS, WRAPPERS AND MULTI-LAYER RNNS<\/em>\".<\/p>\n\n<p>It states that:<\/p>\n\n<blockquote>\n  <ul>\n  <li>BasicRNNCell \u2013 A vanilla RNN cell. <\/li>\n  <li>GRUCell \u2013 A Gated Recurrent Unit    cell. <\/li>\n  <li>BasicLSTMCell \u2013 An LSTM cell based on Recurrent Neural Network    Regularization. No peephole connection or cell clipping. <\/li>\n  <li>LSTMCell \u2013 A more complex LSTM cell that allows for optional peephole connections and cell clipping. <\/li>\n  <li>MultiRNNCell \u2013 A wrapper to combine multiple cells into a multi-layer cell. <\/li>\n  <li>DropoutWrapper \u2013 A wrapper to add dropout to    input and\/or output connections of a cell.<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>Given this, I would suggest you to switch from <code>BasicRNNCell<\/code> to <code>BasicLSTMCell<\/code>. Where <code>Basic<\/code> here means \"use it unless you know what you are doing\". If you want to <em>try<\/em> LSTMs without going into details, thats the way to go. It may be straightforward, just replace with it and voil\u00e0!<\/p>\n\n<p>If not, share some of your code + error.<\/p>\n\n<p>Hope it helps<\/p>\n"}]}
{"QuestionId":42367757,"AnswerCount":2,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-21T12:47:27.790","AcceptedAnswerId":42369657.0,"OwnerUserId":2768132.0,"Title":"Graph using tensorboard","Body":"<p>How can see the graph, generated by TensorFlow using Tensorboard? Can anyone provide a simple example? I am very new to tensorflow. I am using python as well.<\/p>\n\n<p>Thanks in advance!!<\/p>\n","answers":[{"AnswerId":"46802742","CreationDate":"2017-10-18T04:18:57.770","ParentId":null,"OwnerUserId":"8216409","Title":null,"Body":"<p>First of all you need to create your computational graph<\/p>\n\n<pre><code>import tensorflow as tf\n\na = tf.constant(3)\nb = tf.constant(4)\n\nc = a+b\n\nwith tf.Session() as sess:\n    File_Writer = tf.summary.FileWriter('\/home\/theredcap\/Documents\/CS\/Personal\/Projects\/Test\/tensorflow\/tensorboard\/' , sess.graph )\n    print(sess.run(c))\n<\/code><\/pre>\n\n<p>Here the <code>tf.summary.FileWriter(' path to store your logs')<\/code>writes the graph to the Tensorflow log<\/p>\n\n<p>Then all you have to do is access the tensorboard with the mentioned log\/graph that you just wrote with <\/p>\n\n<pre><code>tensorboard --logdir \"path\/to\/logs\"\n<\/code><\/pre>\n\n<p>Note: When I did <\/p>\n\n<pre><code>tensorboard --logdir = \"the above mentioned path\"\n<\/code><\/pre>\n\n<p>Nothing was displayed on my tensorboard<\/p>\n\n<p>For more info<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorboard\/blob\/master\/README.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorboard\/blob\/master\/README.md<\/a><\/p>\n"},{"AnswerId":"42369657","CreationDate":"2017-02-21T14:11:05.350","ParentId":null,"OwnerUserId":"2911687","Title":null,"Body":"<p>You need to create a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/FileWriter\" rel=\"nofollow noreferrer\"><code>tf.summary.FileWriter<\/code><\/a> that writes your graph to a TensorBoard log. For instance:<\/p>\n\n<pre><code>summary_writer = tf.summary.FileWriter('logs', graph=tf.get_default_graph())\nsummary_writer.flush()\n<\/code><\/pre>\n\n<p>And then launch <code>tensorboard<\/code> with the appropriate <code>--logdir<\/code> parameter.<\/p>\n"}]}
{"QuestionId":42367908,"AnswerCount":1,"Tags":"<neural-network><keras><backpropagation>","CreationDate":"2017-02-21T12:54:27.330","AcceptedAnswerId":42377374.0,"OwnerUserId":2439854.0,"Title":"Does the place of the sigmoid function matter in neural network?","Body":"<p>I am trying to build a neural net in python using Keras with a custom loss and I was wandering whether having a sigmoid function as an activation function in the last layer and having a sigmoid in the beginning of the custom loss is the same or not. So here is what I mean by that:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/YmCbD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YmCbD.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>I have a feeling that in the second model the loss is calculated but it is not back propagated through the sigmoid meanwhile in the first model it is. Is that right?   <\/p>\n","answers":[{"AnswerId":"42377374","CreationDate":"2017-02-21T20:31:36.730","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Indeed, in the second case the backpropagation doesn't go through the sigmoid. It is a really bad thing to alter data inside of the loss function.<\/p>\n\n<p>The reason this is a bad thing to do is because then, you will backpropagate an error on the output which is not the real error that the network is making.<\/p>\n\n<p>Explaining myself with a simple case:\nyou have labels in a binary form say a tensor <code>[0, 0, 1, 0]<\/code>\nIf your sigmoid is inside your custom loss function, you might have outputs that look like this <code>[-100, 0, 20, 100],<\/code> the sigmoid in your loss will transform this into something looking approximately like tihs :<code>[0, 0.5, 1, 1]<\/code>\nThe error that will be backpropagated will then be <code>[0, -0.5, 0, -1]<\/code>. The backpropagation will <strong>not<\/strong> take into account the sigmoid and you will apply this error directly to the output. You can see that the magnitude of the error doesn't reflect at all the magnitude of the output's error: the last value is 100 and should be in negative territory, but the model will backpropagate a small error of -1 on that layer. <\/p>\n\n<p>To summarize, the sigmoid must be in the network so that the backpropagation takes it into account when backpropagating the error.<\/p>\n"}]}
{"QuestionId":42369041,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-ml>","CreationDate":"2017-02-21T13:44:58.390","AcceptedAnswerId":42371163.0,"OwnerUserId":3064462.0,"Title":"Field: version.deployment_uri Error: The total size of files in gs:\/\/my-bucket\/ml\/ is x bytes, which exceeds the allowed maximum of 1073741824 bytes","Body":"<p>When trying to create a new version in the google cloud console, I get an error like,<\/p>\n\n<p>Field: version.deployment_uri Error: The total size of files in gs:\/\/my-bucket\/ml\/ is 2150116163 bytes, which exceeds the allowed maximum of 1073741824 bytes.<\/p>\n\n<p>My model is an RNN model.  I believe the embed sequence, vocab size, is likely the cause of the large model.<\/p>\n\n<p>Is there a quota setting that can be adjusted for larger models?<\/p>\n","answers":[{"AnswerId":"42371163","CreationDate":"2017-02-21T15:15:40.383","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>Unfortunately, that limit is not adjustable at this time, although it may be in the future.<\/p>\n\n<p>Are you comfortable sharing how large your model is? That information is valuable for us for planning purposes.<\/p>\n\n<p>In the meantime, you will need to adjust the vocab and embedding sizes or otherwise reduce the size of the model.<\/p>\n"}]}
{"QuestionId":42369718,"AnswerCount":2,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-02-21T14:13:35.937","AcceptedAnswerId":null,"OwnerUserId":1190882.0,"Title":"Tensor flow: Passing bool variable while creating and evaluating graph","Body":"<p>I want to pass a Boolean value which can change while I am evaluating the graph. I am posting my code below:<\/p>\n\n<pre><\/pre>\n\n<p>I am getting the errors as shown below:  For <\/p>\n\n<pre><\/pre>\n\n<p>For :<\/p>\n\n<pre><\/pre>\n\n<p>Please can somebody explain me how should I overcome this problem. And if possible, can anybody tell me where I am going wrong in all of my approaches.<\/p>\n\n<p>Basically what I am doing is I have two models. First I train one of the models, then using the learned values of first model, I have to train the second model freezing the values of first model. <\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42373606","CreationDate":"2017-02-21T17:04:21.693","ParentId":null,"OwnerUserId":"2283175","Title":null,"Body":"<p>You can't just use <code>True<\/code> and <code>False<\/code> inside \"the graph\" - you need to pass them from <code>python<\/code> to <code>tensorflow<\/code>, one way to do this is with <code>tf.constant()<\/code>. You can try<\/p>\n\n<pre><code>def my_model(input, is_freeze):\n    trainable = tf.cond(is_freeze, lambda: tf.constant(False), lambda: tf.constant(True))\n<\/code><\/pre>\n\n<p>Otherwise looks fine to me! Please provide a minimal working example if you want more detailed help!<\/p>\n"},{"AnswerId":"42376804","CreationDate":"2017-02-21T19:59:43.030","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Seems that you are trying to embed the switch from inference \/ logic inside your tensorflow graph. Is not advisable to do so. Check out the new <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f821ce046df71f5784ed4ce7fb6b87f77d96b031\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py#L984\" rel=\"nofollow noreferrer\">Estimator<\/a> interface to see how you can achieve cleanly the same goal.<\/p>\n"}]}
{"QuestionId":42370388,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-21T14:41:29.740","AcceptedAnswerId":43734651.0,"OwnerUserId":799188.0,"Title":"Inception v3 Graph Definition File Download","Body":"<p>I am trying to follow the transfer learning examples <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\" rel=\"nofollow noreferrer\">1<\/a> and <a href=\"https:\/\/www.kernix.com\/blog\/image-classification-with-a-pre-trained-deep-neural-network_p11\" rel=\"nofollow noreferrer\">2<\/a>, both use a pretrained Inception v3 model. Both refer to a graph definition file, .<\/p>\n\n<p>Where can I download the original Inception v3  file?<\/p>\n\n<p>Any help would be appreciated.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"43734651","CreationDate":"2017-05-02T09:39:54.320","ParentId":null,"OwnerUserId":"3958070","Title":null,"Body":"<p>I seems you should follow the instructions from here:\n<a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/image_retraining<\/a><\/p>\n\n<p>to get sustainable results - the script employed within the lab loads the pre-trained Inception v3 model, removes the old top layer, and trains a new one on the images you've applied.<\/p>\n\n<p>There is more recent version of Inception-v3 model available:\n<a href=\"http:\/\/download.tensorflow.org\/models\/inception_v3_2016_08_28.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/inception_v3_2016_08_28.tar.gz<\/a><\/p>\n"},{"AnswerId":"42376621","CreationDate":"2017-02-21T19:49:38.027","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">From the official Github page<\/a><\/p>\n\n<pre><code>curl -O http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-v3-2016-03-01.tar.gz\n<\/code><\/pre>\n"}]}
{"QuestionId":42370740,"AnswerCount":2,"Tags":"<python><numpy><neural-network><keras><recurrent-neural-network>","CreationDate":"2017-02-21T14:57:03.230","AcceptedAnswerId":null,"OwnerUserId":7583561.0,"Title":"Preparing Time-Series Data for Keras LSTM - Network Trains with Extremely High Loss","Body":"<p>I am running into issues preparing my data for use in Keras's LSTM layer. The data is a 1,600,000 item  time-series csv consisting of a date and three features:<\/p>\n\n<p>\n My goal is to predict the value of F1 <em>prediction_period<\/em> timesteps in the future. Understanding that Keras's LSTM layer takes import data in the format (samples,timesteps,dimensions) I wrote the following function to convert my data into a 3D numpy array in this format (Using 2016-03-03 as an example):<\/p>\n\n<p>[[[.284, .278, .687], [.276, .305, .691], [.252, .316, .690]],...other samples...]<\/p>\n\n<p>This function creates the array by stacking copies of the data, with each copy shifted one step further back in time. Lookback is the number of \"layers\" in the stack and trainpercent is train\/test split:<\/p>\n\n<pre><\/pre>\n\n<p>While the network accepted the data and did train, the loss values were exceptionally high, which was very surprising considering that the data has a definite periodic trend. To try and isolate the problem, I replaced my dataset and network structure with a sin-wave dataset and structure from this example:\n<a href=\"http:\/\/www.jakob-aungiers.com\/articles\/a\/LSTM-Neural-Network-for-Time-Series-Prediction\" rel=\"nofollow noreferrer\">http:\/\/www.jakob-aungiers.com\/articles\/a\/LSTM-Neural-Network-for-Time-Series-Prediction<\/a>.<\/p>\n\n<p>Even with the sin wave dataset, the loss was still orders of magnitude higher that were produced by the example function. I went through the function piece by piece, using a one column sequential dataset and compared expected values with the actual values. I didn't find any errors.<\/p>\n\n<p>Am I structuring my input data incorrectly for Keras's LSTM layer? If so, what is the proper way to do this? If not, what would you expect to cause these symptoms (extremely high loss which does not decrease over time, even with 40+ epochs) in my function or otherwise.<\/p>\n\n<p>Thanks in advance for any advice you can provide!<\/p>\n","answers":[{"AnswerId":"42374204","CreationDate":"2017-02-21T17:37:33.910","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>A \"high loss\" is a very subjective thing. We can not assess this without seeing your model.<\/p>\n\n<p>It can come from multiple reasons: <\/p>\n\n<ul>\n<li>training loss can be influenced by regularization techniques. For example, the whole point of L2 regularization is to add the weights of the model in the loss.<\/li>\n<li>the loss is defined by an objective function, so it depends on what objective you are using.<\/li>\n<li>the optmizer you are using for that objective function might not be adapted. Some optimizers do not garantee convergence of the loss.<\/li>\n<li>your time serie might not be predictable (but apparently this is not your case).<\/li>\n<li>your model might not be adequate for the task you are trying to achieve.<\/li>\n<li>your training data is not correctly prepared (but you have investigated this)<\/li>\n<\/ul>\n\n<p>You see that there are plenty of possibilities. A high loss doesn't mean anything in itself. You can have a really small loss and just do + 1000 and your loss will be high eventhough the problem is solved<\/p>\n"},{"AnswerId":"45252148","CreationDate":"2017-07-22T07:57:54.260","ParentId":null,"OwnerUserId":"6599583","Title":null,"Body":"<p>Here are some things you can do to improve your predictions:<\/p>\n\n<ol>\n<li><p>First make sure you input data is centered i.e. apply some\nstandardization or normalization. You can either use the\nMinMaxScaler or StandardScaler from sklearn library or implement\nsome custom scaling based on your data.<\/p><\/li>\n<li><p>Make sure your network(LSTM\/GRU\/RNN) is big enough to capture the\ncomplexity in your data. <\/p><\/li>\n<li><p>Use the tensorboard callback in Keras to monitor your weight\nmatrices and loss functions.<\/p><\/li>\n<li><p>Use an adaptive optimizer instead of setting custom learning\nparameters. Maybe'adam' or 'adagrad' .<\/p><\/li>\n<\/ol>\n\n<p>Using these will at least make sure that your network is training. You should see gradual decrease of losses over time. After you've solved this problem you are free to experiment with your initial hyper-parameters and implementing different regularization techniques<\/p>\n\n<p>Good Luck !<\/p>\n"}]}
{"QuestionId":42370847,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-21T15:01:29.237","AcceptedAnswerId":null,"OwnerUserId":6215726.0,"Title":"How to build spp-net in Tensorflow?","Body":"<p>since tf.max_pool dosen't support variable size, the spp layer does not work. \nI feed the length to the graph, got a error.<\/p>\n\n<pre><\/pre>\n\n<p>TypeError: Expected int32, got 1.0 of type 'float' instead.<\/p>\n\n<p>Do you have any idea, how to deal with it?<\/p>\n","answers":[{"AnswerId":"42371386","CreationDate":"2017-02-21T15:25:12.283","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>In your strides parameter, change the 3rd value to integer.\ni.e.<\/p>\n\n<p><code>np.floor(shape[2] * 1. \/ l + 1)<\/code> to <code>np.floor(shape[2] * 1. \/ l + 1).astype(np.int32)<\/code><\/p>\n\n<p>Complete <code>stride<\/code> option:<\/p>\n\n<pre><code>strides=[1, np.floor(shape[1] * 1. \/ l + 1).astype(np.int32), \n           np.floor(shape[2] * 1. \/ l + 1).astype(np.int32), 1],\n<\/code><\/pre>\n\n<p>Hope this helps you.<\/p>\n"}]}
{"QuestionId":42370946,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-21T15:05:54.657","AcceptedAnswerId":null,"OwnerUserId":4520592.0,"Title":"Tensorflow: how to compute auc correctly","Body":"<p>Here is my function that computes accuracy and auc:<\/p>\n\n<pre><\/pre>\n\n<p>Here results produced by logging this function's output: <br><\/p>\n\n<pre><\/pre>\n\n<p>Auc is very high - too high - in my opinion. I've checked tensorflow docs for a hundred of times but could understand what could be wrong. <br><\/p>\n\n<p>Could anyone please help me find an error? <br><\/p>\n","answers":[]}
{"QuestionId":42371079,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-02-21T15:12:05.267","AcceptedAnswerId":42372005.0,"OwnerUserId":3425890.0,"Title":"How to specify the architecture of deep neural network in Tensorflow?","Body":"<p>I am newbie in Tensorflow<\/p>\n\n<p>Actually, I am testing some example in Tensorflow web-site, and I start to understand some features of the framwork, but what I don't understand is how I can design my architecture, I mean number of layers, type of Layer \"conv, pool...\", and if it is necessery to do that, because there are many predifined architectures like AmexNet, <\/p>\n\n<p>Thanks, <\/p>\n","answers":[{"AnswerId":"42372005","CreationDate":"2017-02-21T15:53:35.190","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>I would strongly recommend working through their hands on tutorial, depending on if you have previous ML experience (<a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/pros\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/pros<\/a>) or not (<a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners<\/a>). The questions you are asking are answered in there. <\/p>\n\n<p>The question on using predefined architectures or self defined depends on your use case. If you want to do something easy like classifying if there is only a car in the scene or not a more shallow architecture might work better, because it is faster and a more deep one is overkill. However most architectures are similar to the ones already defined in literature. <\/p>\n\n<p>Another question that arises naturally, while talking about pre defined architecture is about transfer learning \/ fine tuning. Often pre defined architectures are already learned on some big dataset (mostly ImageNet) and already perform really well out of the box for many tasks. With little training data it makes a lot of sense to use this. With lots of training data it can hinder your progress though. <\/p>\n"}]}
{"QuestionId":42371282,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-21T15:21:11.933","AcceptedAnswerId":null,"OwnerUserId":7251162.0,"Title":"feedback loop on tensorflow","Body":"<p>I am trying to implement a network that finds key points in an image.\nTo do so I tried to use an initial guess and then to reuse it as an input to the network (concatenate it to the initial input) and let it calculate the result again. How can I do this feedback? do I need to build a loss that uses those two output or just the output in the end. \nis  relevent to this task?\nThank you. Gilad<\/p>\n","answers":[]}
{"QuestionId":42374230,"AnswerCount":2,"Tags":"<python><tensorflow><google-cloud-platform>","CreationDate":"2017-02-21T17:38:29.453","AcceptedAnswerId":42379007.0,"OwnerUserId":7600246.0,"Title":"tensorflow: how to put my python code on google cloud as a service","Body":"<p>I'm a begginer in TensorFlow and I have some doubts about how to deploy my code in google cloud.<\/p>\n\n<h2>I ran de retrain.py in my local machine and I trainned the Inception with my own pictures. I have the graph (.pb and .meta). After that I develop another python script to test my model. Something like that:<\/h2>\n\n<pre><\/pre>\n\n<hr>\n\n<p>In my local machine I execute de script with one parameter where I put the picture's path that I want to predict. My script works fine and I have the response in my final print. But I dont know how to execute this script in google cloud as a service. I already saw some foruns and I can't find a solution.\nFor now, I turn this code in a webservice and I'm deploying in my own server. But I need to use the google cloud.<\/p>\n\n<p>Any suggestion?<\/p>\n","answers":[{"AnswerId":"42411893","CreationDate":"2017-02-23T09:38:28.740","ParentId":null,"OwnerUserId":"7219743","Title":null,"Body":"<p>As Google cloud do you mean <a href=\"https:\/\/cloud.google.com\/ml\/\" rel=\"nofollow noreferrer\">Cloud Machine Learning<\/a>?\nThis platform enables you to build machine learning models and is optimized for Tensorflow. <\/p>\n\n<p>To deploy your model on this platform you need to save you graph in two files named <a href=\"https:\/\/cloud.google.com\/ml\/docs\/how-tos\/preparing-models\" rel=\"nofollow noreferrer\">export and export.meta<\/a> (visit the link to more details).<\/p>\n\n<p>In my view the only problem is how pass the image to predict to the service, because cloud ml don't accept jpg file, and you have to convert your image in a text format.<\/p>\n\n<p>I think that you could find useful suggestions <a href=\"https:\/\/stackoverflow.com\/questions\/40849195\/how-convert-a-jpeg-image-into-json-file-in-google-machine-learning\/40856329#40856329\">here<\/a> and <a href=\"https:\/\/cloud.google.com\/blog\/big-data\/2016\/12\/how-to-train-and-classify-images-using-google-cloud-machine-learning-and-cloud-dataflow\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"42379007","CreationDate":"2017-02-21T22:14:54.983","ParentId":null,"OwnerUserId":"2050","Title":null,"Body":"<p>The TensorFlow project provides <a href=\"https:\/\/github.com\/tensorflow\/serving\" rel=\"nofollow noreferrer\">Serving<\/a> to turn models into services. The only protocol supported at this time is protocol buffers (gRPC) at this time.<\/p>\n\n<p>The TensorFlow website and the Serving page provide sample code and  tutorials.<\/p>\n"}]}
{"QuestionId":42374401,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-02-21T17:48:19.673","AcceptedAnswerId":42497461.0,"OwnerUserId":886724.0,"Title":"Keras model\/layer regularization property deprecated","Body":"<p>I am using Keras to define a model, then I try to calibrate it using Distributed Tensorflow, as is done <a href=\"https:\/\/gist.github.com\/fchollet\/2c9b029f505d94e6b8cd7f8a5e244a4e\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>I used to deal with regularizer loss as done in the link<\/p>\n\n<pre><\/pre>\n\n<p>But now the regularizers property is depricated and a warning suggests to use  property of  so I tried :<\/p>\n\n<pre><\/pre>\n\n<p>But this causes a crash. Any help please<\/p>\n","answers":[{"AnswerId":"42497461","CreationDate":"2017-02-27T23:09:33.980","ParentId":null,"OwnerUserId":"886724","Title":null,"Body":"<p>The correct thing to do is simply :<\/p>\n\n<pre><code>loss = tf.reduce_mean(keras.objectives.mean_squared_error(targets, preds))\ntotal_loss = loss * 1.  # copy tensor\nfor reg_loss in model.losses:\ntotal_loss = total8loss + reg_loss\n<\/code><\/pre>\n"}]}
{"QuestionId":42374524,"AnswerCount":0,"Tags":"<r><tensorflow>","CreationDate":"2017-02-21T17:55:18.440","AcceptedAnswerId":null,"OwnerUserId":6785054.0,"Title":"How to read .tfrecord files into R?","Body":"<p>Is it possible to read TensorFlow's .tfrecord files into R?<\/p>\n\n<p>I suppose it should look something like this:<\/p>\n\n<pre><\/pre>\n\n<p>But I am not sure what the arguments  should be and whether I am even close.<\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>I managed to get this far:<\/p>\n\n<pre><\/pre>\n\n<p> is a list of the correct size I expected but I am not familiar with the class of the elements in the list:<\/p>\n\n<pre><\/pre>\n\n<p>[1] \"python.builtin.str\"        \"python.builtin.basestring\" \"python.builtin.object\"<\/p>\n\n<p>I might have to make use of the functions in the reticulate R package. Any suggestions would be very much appreciated.<\/p>\n","answers":[]}
{"QuestionId":42374769,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-21T18:07:45.400","AcceptedAnswerId":null,"OwnerUserId":7466373.0,"Title":"tensorflow sequence to sequence models error","Body":"<p>I am getting the following errors after executing translate.py from sequence to sequence models tutorial in tensorflow<a href=\"https:\/\/i.stack.imgur.com\/ydzNB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ydzNB.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Can anyone please help me in resolving this error\nThanks in advance<\/p>\n","answers":[]}
{"QuestionId":42374849,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-02-21T18:11:40.303","AcceptedAnswerId":null,"OwnerUserId":3218279.0,"Title":"Implementation of Stateful LSTM","Body":"<p>I want to create a stateful LSTM<\/p>\n\n<p>My data is of 68871 x 43, where the features are in the column no. 1-42 and label in column no. 43<\/p>\n\n<p>My keras LSTM code for classification of the data is<\/p>\n\n<pre><\/pre>\n\n<p>Please suggest what changes to be made in order to make the LSTM stateful\nThanks so much in advance !!!!!<\/p>\n","answers":[{"AnswerId":"42376795","CreationDate":"2017-02-21T19:59:08.420","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<pre><code>model.add(LSTM(10, input_shape=(5,43), stateful=True))\n<\/code><\/pre>\n\n<p>is what you are looking for I think. See <a href=\"https:\/\/keras.io\/layers\/recurrent\/#recurrent\" rel=\"nofollow noreferrer\">here for documentation<\/a>.<\/p>\n\n<p>Also, if you want to use stateful lstm it doesn't make sense to shuffle the training data (which is the default). Therefore modify this line too: <\/p>\n\n<pre><code>model.fit(trainX, trainY, nb_epoch=10, batch_size=5, shuffle=False)\n<\/code><\/pre>\n"}]}
{"QuestionId":42374972,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-21T18:18:13.043","AcceptedAnswerId":null,"OwnerUserId":1130295.0,"Title":"System blows up during install","Body":"<p>I downloaded a clean copy of python 35.3<\/p>\n\n<p>Tried installing tensorflow per the instructions.  It did not work.  The method is described as <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a>.  Here is the result.<\/p>\n\n<pre><\/pre>\n\n<p>I also tried pip and eliminating the upgrade term.  All syntax out on install.<\/p>\n\n<p>What am I missing?  <\/p>\n","answers":[{"AnswerId":"42375246","CreationDate":"2017-02-21T18:33:04.887","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It looks like you entered the <code>pip3 install \u2026<\/code> command at the interactive Python prompt. <code>pip3<\/code> is the name of a program to run, and not a command in the Python language itself.<\/p>\n\n<p>Instead you should enter this command at a Windows command prompt (e.g. the prompt you get by running <code>cmd.exe<\/code> or \"Command Prompt\" in the Start menu).<\/p>\n"}]}
{"QuestionId":42374988,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2017-02-21T18:18:43.823","AcceptedAnswerId":42449975.0,"OwnerUserId":1637894.0,"Title":"Theano scan acts strangely after 127 steps","Body":"<p>I'm confused about how Theano scan works. I would expect the following function to monotonically increase, but this doesn't happen:<\/p>\n\n<pre><\/pre>\n\n<p>The output:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42449975","CreationDate":"2017-02-24T23:28:14.330","ParentId":null,"OwnerUserId":"1637894","Title":null,"Body":"<p>Well, I'm dumb ... the type was set to <code>int8<\/code>!<\/p>\n"}]}
{"QuestionId":42375276,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><autoencoder>","CreationDate":"2017-02-21T18:34:47.123","AcceptedAnswerId":null,"OwnerUserId":4533188.0,"Title":"Tensorflow autoencoder: How to get representative output?","Body":"<h1>Setup<\/h1>\n\n<p>I build an autoencoder with Tensorflow for images. My images are around 30 Pixels in length and width. I am using 5 layers:<\/p>\n\n<ol>\n<li>Input layer<\/li>\n<li>Encoder layer with 256 neurons with linear functions. (This layer is supposed to function sort of as a preprocessing PCA.)<\/li>\n<li>Encoder layer with 128 neurons with sigmoid functions.<\/li>\n<li>Decoder layer with 256 neurons with sigmoid functions.<\/li>\n<li>Decoder\/output layer with as many neurons as the input with linear function.<\/li>\n<\/ol>\n\n<p>All layer use biases and are defined like this<\/p>\n\n<pre><\/pre>\n\n<p>My cost is defined by<\/p>\n\n<pre><\/pre>\n\n<p>and I used  with start learning rate of .<\/p>\n\n<p>I trained the autoencoder with around 250000 images. In the following I show the results for test data, with which I did not train. The upper row always shows the input images, the lower row always shows the output images.<\/p>\n\n<p>I used the autoencoder satisfactory on the MNIST data (array cells are values from 0-1):\n<a href=\"https:\/\/i.stack.imgur.com\/tjsAC.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tjsAC.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I used the autoencoder satisfactory on the photos I took myself (array cells are values from 0-100):\n<a href=\"https:\/\/i.stack.imgur.com\/KyHmK.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/KyHmK.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Where the root mean square errors<\/p>\n\n<pre><\/pre>\n\n<p>are as one would expect - higher for the first three photos (darkened, whitened, put a cross in), and lower for the unchanged photos:<\/p>\n\n<pre><\/pre>\n\n<p>But the autoencoder does not work satisfactory for texture images (array cells are from 0-1). First I used  and got<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/F4fJM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/F4fJM.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Since my cost where pretty high during training and did not change I followed the advice from <a href=\"https:\/\/stackoverflow.com\/a\/40956761\/4533188\">https:\/\/stackoverflow.com\/a\/40956761\/4533188<\/a> and used . Indeed I got better results with<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/LuIMv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/LuIMv.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>While the costs are lower, they are still constant during the epochs:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/8Nuak.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/8Nuak.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Also the output images are pretty dark. I believe this might also the reason why my s are is not as I would expect:<\/p>\n\n<pre><\/pre>\n\n<p>Please note, that I would expect that the s is high for the first image due to the dot in the middle and the second image since it is darkened. I guess the reason the s are not as hoped, is because the output images are so dark.<\/p>\n\n<h1>Questions<\/h1>\n\n<ol>\n<li>Why are the costs not decreasing over the epochs and what can I do to get this going?<\/li>\n<li>Why are the output images so dark and what can I do to get them to be better representations of the input images?<\/li>\n<li>How do I get the s to be as I expect?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"42375969","CreationDate":"2017-02-21T19:12:37.677","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Your network is clearly not training<\/p>\n\n<ul>\n<li>Your Autoenconder Layers are not symmetric <\/li>\n<li>Don't use sigmoid, use ReLU<\/li>\n<li>Use a better initialization technique (the specific distribution depends on the activation function)<\/li>\n<li>Share the weight of the Encoder and Decoder layers <\/li>\n<\/ul>\n"}]}
{"QuestionId":42375680,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><computer-vision><artificial-intelligence><deep-learning>","CreationDate":"2017-02-21T18:55:33.063","AcceptedAnswerId":null,"OwnerUserId":1085736.0,"Title":"Tensorflow Count Objects in Image","Body":"<p>New to machine learning so looking for some direction how to get started. The end goal is to be able to train a model to count the number of objects in an image using Tensorflow. My initial focus will be to train the model to count one specific type of object. So lets say I take coins. I will only train the model to count coins. Not worried about creating a generic counter for all different types of objects. I've only done Google's example of image classification of flowers and I understand the basics of that. So looking for clues how to get started. Is this an image classification problem and I can use the same logic as the flowers...etc etc?<\/p>\n","answers":[{"AnswerId":"42395245","CreationDate":"2017-02-22T15:10:09.990","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Probably the best performing solution for the coin problem would be to use a regression to solve this. Annotate 5k images with the amount of objects in the scene and run your model on it. Then your model just outputs the correct number. (Hopefully)<\/p>\n\n<p>Another way is to classify if an image shows a coin and use a sliding window approach like this one: <a href=\"https:\/\/arxiv.org\/pdf\/1312.6229.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1312.6229.pdf<\/a> to classify for each window if it shows a coin. Then you count the found regions. This one is easier to annotate and learn and better extensible. But you have the problem of choosing good windows and using the result of those windows in a concise way. <\/p>\n"}]}
{"QuestionId":42375988,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-02-21T19:13:32.300","AcceptedAnswerId":null,"OwnerUserId":3365859.0,"Title":"CentOS machine problems with tensorflow","Body":"<p>I have a GPU CentOS machine with OS version 7.2<\/p>\n\n<p>I installed the tensorflow version 1.0 using pip (I did not compile)<\/p>\n\n<p>When I run a problem using updated keras, I get<\/p>\n\n<blockquote>\n  <p>W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use SSE3 instructions, but these are\n  available on your machine and could speed up CPU computations. <\/p>\n  \n  <p>W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use SSE4.1 instructions, but these are\n  available on your machine and could speed up CPU computations. <\/p>\n  \n  <p>W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use SSE4.2 instructions, but these are\n  available on your machine and could speed up CPU computations. <\/p>\n  \n  <p>W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use AVX instructions, but these are\n  available on your machine and could speed up CPU computations. <\/p>\n  \n  <p>W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use AVX2 instructions, but these are\n  available on your machine and could speed up CPU computations. <\/p>\n  \n  <p>W tensorflow\/core\/platform\/cpu_feature_guard.cc:45] The TensorFlow\n  library wasn't compiled to use FMA instructions, but these are\n  available on your machine and could speed up CPU computations.<\/p>\n<\/blockquote>\n\n<p>Any thoughts on how to fix this? <\/p>\n","answers":[{"AnswerId":"42376286","CreationDate":"2017-02-21T19:30:52.243","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>These warnings are harmless, but indicate that\u2014if you wanted to\u2014you could compile a more optimized version of TensorFlow for your local machine. The published binaries for TensorFlow are somewhat conservative in what platform-specific optimizations they apply, so that they can run on a wide range of machines. Note that these optimizations only affect CPU performance, so for GPU-accelerated models they will not have as much of an effect.<\/p>\n\n<p>Building TensorFlow from source with the options suggested in <a href=\"https:\/\/stackoverflow.com\/a\/41584791\/3574081\">this answer<\/a> will give you a version that does not emit these warnings.<\/p>\n"}]}
{"QuestionId":42376790,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-21T19:59:00.997","AcceptedAnswerId":42378111.0,"OwnerUserId":667648.0,"Title":"Segmentation fault running tensorflow-gpu","Body":"<p>If I do:<\/p>\n\n<pre><\/pre>\n\n<p>(note: this is <em>all<\/em> the output I get)<\/p>\n\n<p>Currently I have installed  using . I have followed advice <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5141\" rel=\"nofollow noreferrer\">here<\/a> to prevent segmentation faults by putting the following in my  file.<\/p>\n\n<pre><\/pre>\n\n<p>Lastly, I have disable SIP. None of this seems to work though. Any thoughts on how to make it run?<\/p>\n","answers":[{"AnswerId":"42378111","CreationDate":"2017-02-21T21:14:20.350","ParentId":null,"OwnerUserId":"667648","Title":null,"Body":"<p>So I looked at yaroslavvb comment <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6729\" rel=\"nofollow noreferrer\">here<\/a> and changed my <code>.zshrc<\/code> file to this:<\/p>\n\n<pre><code>export CUDA_HOME=\/usr\/local\/cuda\nexport DYLD_LIBRARY_PATH=\/usr\/local\/cuda\/lib:\/usr\/local\/cuda\/extras\/CUPTI\/lib\nexport LD_LIBRARY_PATH=$DYLD_LIBRARY_PATH\nexport PATH=$DYLD_LIBRARY_PATH:$PATH\n<\/code><\/pre>\n\n<p>Got rid of the link (I still have SIP disabled, maybe I should set it back?)<\/p>\n\n<p>This gives:<\/p>\n\n<pre><code>\u279c  ~ python3 -c \"import numpy; import tensorflow;\"\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcublas.8.0.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:126] Couldn't open CUDA library libcudnn.5.dylib. LD_LIBRARY_PATH: \/usr\/local\/cuda\/lib:\/usr\/local\/cuda\/extras\/CUPTI\/lib\nI tensorflow\/stream_executor\/cuda\/cuda_dnn.cc:3517] Unable to load cuDNN DSO\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcufft.8.0.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcuda.1.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcurand.8.0.dylib locally\n<\/code><\/pre>\n\n<p>I then realized I still needed to install <code>cudnn<\/code> which I did by downloading it from Nvidia and <a href=\"https:\/\/christopher5106.github.io\/big\/data\/2015\/07\/16\/deep-learning-install-caffe-cudnn-cuda-for-digits-python-on-mac-osx.html\" rel=\"nofollow noreferrer\">following here.<\/a><\/p>\n\n<p>This then gave:<\/p>\n\n<pre><code>\u279c  ~ python3 -c \"import tensorflow;\" \nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcublas.8.0.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcudnn.5.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcufft.8.0.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcuda.1.dylib locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcurand.8.0.dylib locally\n<\/code><\/pre>\n"}]}
{"QuestionId":42377107,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2017-02-21T20:16:04.543","AcceptedAnswerId":null,"OwnerUserId":2145297.0,"Title":"In Keras what's the best way to save a model checkpoint and reload it for training on different data set?","Body":"<p>So let's say I have a trained Keras neural network called model. I want to save the model before further training so I can go back to that checkpoint as I train it on different datasets.<\/p>\n\n<pre><\/pre>\n\n<p>Now here's where my question comes in. I want to free up the GPU memory so I can reload checkpoint_1 before further training of the model. What I'm currently doing is ending the current tensorflow session and starting a new one.<\/p>\n\n<pre><\/pre>\n\n<p>I then load checkpoint_1 and continue training.<\/p>\n\n<pre><\/pre>\n\n<p>Is there a better way of doing this? Stopping and starting the tensorflow sessions takes a lot of time.<\/p>\n","answers":[]}
{"QuestionId":42378223,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><computer-vision><large-data>","CreationDate":"2017-02-21T21:20:58.790","AcceptedAnswerId":42380150.0,"OwnerUserId":3367721.0,"Title":"Loading\/ feed_dicting large dataset into Tensorflow session","Body":"<p>I am trying to consume 50k image dataset for a convNet in ratio tr-60%, test-20%, validate-20%. \nSo far I have created  a placeholder and feed_dicting it @ sess.run(), as follows:-<\/p>\n\n<pre><\/pre>\n\n<p>but according to official TF performance guide it is a poor way to implement, as follows:-\n<a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide\" rel=\"nofollow noreferrer\">link to TF guide<\/a><\/p>\n\n<blockquote>\n  <p>Unless for a special circumstance or for example code, do not feed\n  data into the session from Python variables, e.g. dictionary.<\/p>\n\n<pre><\/pre>\n<\/blockquote>\n\n<p>Can you please help in implementing queues for reading data in TF? <\/p>\n\n<p>One way I found is :-<\/p>\n\n<blockquote>\n  <p>Create an op which is loading your data in stream fashion<\/p>\n<\/blockquote>\n\n<p>But I am not sure \n<strong>1)<\/strong> if it is best way, \n<strong>2)<\/strong> I couldn't implement above suggestion, can you help in creating this op pseudo code?\nThanks alot.<\/p>\n","answers":[{"AnswerId":"42380150","CreationDate":"2017-02-21T23:47:43.240","ParentId":null,"OwnerUserId":"2595935","Title":null,"Body":"<p>It is generally a bad idea to feed data using feed_dict but you don't always have to write ops to process your data. You can convert your image data to the format that tensorflow can recognize: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/convert_to_records.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/convert_to_records.py<\/a>. This process can be made in parallel and you can output a list of files since tensorflow can consume file list as well.<\/p>\n\n<p>Then follow the tutorial on this page to create queues and feed data in python: <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/reading_data<\/a><\/p>\n"}]}
{"QuestionId":42378963,"AnswerCount":2,"Tags":"<python-2.7><tensorflow><deep-learning>","CreationDate":"2017-02-21T22:10:43.593","AcceptedAnswerId":42379657.0,"OwnerUserId":2033910.0,"Title":"Is it possible to point tensorflow to a specific version of CUDA after tensorflow has been installed?","Body":"<p>I am trying to run tensorflow on a cuda machine over the cloud where I do not have sudo privileges. <\/p>\n\n<p>I installed tensorflow in a virtualenv but I am not able to point it towards a specific cuda version. Can this be only done when installing from source?<\/p>\n","answers":[{"AnswerId":"42379101","CreationDate":"2017-02-21T22:22:17.967","ParentId":null,"OwnerUserId":"2050","Title":null,"Body":"<p>Usually <code>CUDA_HOME<\/code> allows to solve such kind of issues. As for TensorFlow up to 0.12, the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup\" rel=\"nofollow noreferrer\">homepage<\/a> gives detailed information.<\/p>\n\n<p>Version 1.0 is available now, and the documentation has been updated (not re-read at this point), so there might be extra information there if you switched.<\/p>\n"},{"AnswerId":"42379657","CreationDate":"2017-02-21T23:02:32.783","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>On my Ubuntu systems I use (to be executed in the same terminal as the one in which you run TensorFlow):<\/p>\n\n<pre><code># CUDA 7.5:\nexport PATH=\/usr\/local\/cuda-7.5\/bin:$PATH\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-7.5\/lib64:$LD_LIBRARY_PATH\n\n# CUDA 8.0:\nexport PATH=\/usr\/local\/cuda-8.0\/bin:$PATH\nexport LD_LIBRARY_PATH=\/usr\/local\/cuda-8.0\/lib64:$LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>You may also want to point to CuDNN and CNMeM:<\/p>\n\n<pre><code># CuDNN:\nexport LD_LIBRARY_PATH=\/scratch1\/francky\/cudnn\/cuda\/lib64:$LD_LIBRARY_PATH\nexport CPATH=\/scratch1\/francky\/cudnn\/cuda\/include:$CPATH\nexport LIBRARY_PATH=\/scratch1\/francky\/cudnn\/cuda\/lib64:$LD_LIBRARY_PATH\n\n# CNMeM:\nexport LD_LIBRARY_PATH=\/scratch1\/francky\/cudnn\/cnmem\/build:$LD_LIBRARY_PATH\nexport CPATH=\/scratch1\/francky\/cudnn\/cnmem\/include:$CPATH\nexport LIBRARY_PATH=\/scratch1\/francky\/cudnn\/cnmem\/build:$LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>FYI: <a href=\"https:\/\/stackoverflow.com\/q\/42396992\/395857\">Is there any point in using CNMeM when running TensorFlow?<\/a><\/p>\n"}]}
{"QuestionId":42379138,"AnswerCount":4,"Tags":"<machine-learning><neural-network><keras>","CreationDate":"2017-02-21T22:24:16.003","AcceptedAnswerId":42381526.0,"OwnerUserId":2383721.0,"Title":"Keras: ValueError: No data provided for \"input_1\". Need data for each key","Body":"<p>I am using the keras functional API with input images of dimension (224, 224, 3). I have the following model using the functional API, although a similar problem seems to arise with sequential models:<\/p>\n\n<pre><\/pre>\n\n<p>I am calling  where my generator has <\/p>\n\n<pre><\/pre>\n\n<p> is the input (224, 224, 3) image and  is in {-1,1}.<\/p>\n\n<p>On fitting the model, I get an error <\/p>\n\n<pre><\/pre>\n\n<p>One strange thing is that if I switch the  target of the dict to , the error switches to missing an input for , but goes back to missing  if both keys are in the data generator.<\/p>\n\n<p>This happens whether I call  or get batches from the generator and call .<\/p>\n\n<p>Does anyone know what's going on? From what I can tell, this should be the same as given in <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"noreferrer\">the documentation<\/a> although with a different input size.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"50666476","CreationDate":"2018-06-03T12:35:06.373","ParentId":null,"OwnerUserId":"5134333","Title":null,"Body":"<p>I encountered this error on 3 cases (In R):<\/p>\n\n<ol>\n<li>The input data does not have the same dimension as was declared in the first layer<\/li>\n<li>The input data includes missing values<\/li>\n<li>The input data is not a matrix (for example, a data frame)<\/li>\n<\/ol>\n\n<p>Please check all of the above.<\/p>\n\n<p>Maybe this code in R can help:<\/p>\n\n<pre><code>library(keras)\n\n#The network should identify the rule that a row sum greater than 1.5 should yield an output of 1\n\nmy_x=matrix(data=runif(30000), nrow=10000, ncol=3)\nmy_y=ifelse(rowSums(my_x)&gt;1.5,1,0)\nmy_y=to_categorical(my_y, 2)\n\nmodel = keras_model_sequential()\nlayer_dense(model,units = 2000, activation = \"relu\", input_shape = c(3))\nlayer_dropout(model,rate = 0.4)\nlayer_dense(model,units = 50, activation = \"relu\")\nlayer_dropout(model,rate = 0.3)\nlayer_dense(model,units = 2, activation = \"softmax\")\n\ncompile(model,loss = \"categorical_crossentropy\",optimizer = optimizer_rmsprop(),metrics = c(\"accuracy\"))\n\nhistory &lt;- fit(model,  my_x, my_y, epochs = 5, batch_size = 128, validation_split = 0.2)\n\nevaluate(model,my_x, my_y,verbose = 0)\n\npredict_classes(model,my_x)\n<\/code><\/pre>\n"},{"AnswerId":"42381526","CreationDate":"2017-02-22T02:19:28.280","ParentId":null,"OwnerUserId":"2383721","Title":null,"Body":"<p>This was due to me misunderstanding how the keras outputs work. The layer specified by the <code>output<\/code> argument to <code>Model<\/code> requires the output from the data. I misunderstood that the <code>output<\/code> key in the data dictionary automatically goes to the layer specified by the <code>output<\/code> argument.<\/p>\n"},{"AnswerId":"53262686","CreationDate":"2018-11-12T12:58:07.247","ParentId":null,"OwnerUserId":"7804867","Title":null,"Body":"<pre><code>yield ({'input_1': image}, {'output': classification}) \n<\/code><\/pre>\n\n<p>Replace <code>output<\/code> with <code>dense_1<\/code>.\nIt will work.<\/p>\n"},{"AnswerId":"50966358","CreationDate":"2018-06-21T10:33:45.683","ParentId":null,"OwnerUserId":"6183919","Title":null,"Body":"<p>I have encountered this issue as well and none of the above mentioned answers worked.  According to the keras <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">documentation<\/a> you can pass the arguments either as a dictionary like that:<\/p>\n\n<pre><code>model.fit({'main_input': headline_data, 'aux_input': additional_data},\n      {'main_output': labels, 'aux_output': labels},\n      epochs=50, batch_size=32)\n<\/code><\/pre>\n\n<p>or as a list like that:<\/p>\n\n<pre><code>model.fit([headline_data, additional_data], [labels, labels],\n      epochs=50, batch_size=32)\n<\/code><\/pre>\n\n<p>The dictionary version didn't work for me with keras version 2.0.9. I have used the list version as a workaround for now. <\/p>\n"}]}
{"QuestionId":42379808,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-21T23:15:22.957","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"How can I view the default configuration options of a TensorFlow session?","Body":"<p>How can I view the default configuration options of a <a href=\"https:\/\/web.archive.org\/web\/20170221231502\/https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session\" rel=\"noreferrer\">TensorFlow session<\/a>?<\/p>\n\n<p>I am aware that I can pass configuration options through the  parameter when initializing a session, e.g.: <\/p>\n\n<pre><\/pre>\n\n<p>But I don't know how to view the default configuration options.  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/protobuf\/config.proto\" rel=\"noreferrer\"><\/a> doesn't look like the default options.<\/p>\n","answers":[{"AnswerId":"42379943","CreationDate":"2017-02-21T23:27:10.763","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Tensorflow 1.0.0 from <\/p>\n\n<pre><code>pip install tensorflow-gpu\n<\/code><\/pre>\n\n<p>I ran:<\/p>\n\n<pre><code>import tensorflow as tf\nimport pprint\np = tf.ConfigProto()\npprint.pprint({ k:str(getattr(p,k)) for k in sorted(dir(p)) if not k.startswith('_')})\n<\/code><\/pre>\n\n<p>generates:<\/p>\n\n<pre><code>{'ALLOW_SOFT_PLACEMENT_FIELD_NUMBER': '7',\n 'ByteSize': '&lt;bound method ConfigProto.ByteSize of &gt;',\n 'Clear': '&lt;bound method ConfigProto._Clear of &gt;',\n 'ClearExtension': '&lt;bound method ConfigProto.ClearExtension of &gt;',\n 'ClearField': '&lt;bound method ConfigProto.ClearField of &gt;',\n 'CopyFrom': '&lt;bound method ConfigProto.CopyFrom of &gt;',\n 'DESCRIPTOR': '&lt;google.protobuf.descriptor.Descriptor object at 0x113dbf4d0&gt;',\n 'DEVICE_COUNT_FIELD_NUMBER': '1',\n 'DEVICE_FILTERS_FIELD_NUMBER': '4',\n 'DeviceCountEntry': \"&lt;class 'tensorflow.core.protobuf.config_pb2.DeviceCountEntry'&gt;\",\n 'DiscardUnknownFields': '&lt;bound method ConfigProto._DiscardUnknownFields of &gt;',\n 'FindInitializationErrors': '&lt;bound method ConfigProto.FindInitializationErrors of &gt;',\n 'FromString': '&lt;function FromString at 0x113de2578&gt;',\n 'GPU_OPTIONS_FIELD_NUMBER': '6',\n 'GRAPH_OPTIONS_FIELD_NUMBER': '10',\n 'HasExtension': '&lt;bound method ConfigProto.HasExtension of &gt;',\n 'HasField': '&lt;bound method ConfigProto.HasField of &gt;',\n 'INTER_OP_PARALLELISM_THREADS_FIELD_NUMBER': '5',\n 'INTRA_OP_PARALLELISM_THREADS_FIELD_NUMBER': '2',\n 'IsInitialized': '&lt;bound method ConfigProto.IsInitialized of &gt;',\n 'LOG_DEVICE_PLACEMENT_FIELD_NUMBER': '8',\n 'ListFields': '&lt;bound method ConfigProto.ListFields of &gt;',\n 'MergeFrom': '&lt;bound method ConfigProto.MergeFrom of &gt;',\n 'MergeFromString': '&lt;bound method ConfigProto.MergeFromString of &gt;',\n 'OPERATION_TIMEOUT_IN_MS_FIELD_NUMBER': '11',\n 'PLACEMENT_PERIOD_FIELD_NUMBER': '3',\n 'ParseFromString': '&lt;bound method ConfigProto.ParseFromString of &gt;',\n 'RPC_OPTIONS_FIELD_NUMBER': '13',\n 'RegisterExtension': '&lt;function RegisterExtension at 0x113de2500&gt;',\n 'SESSION_INTER_OP_THREAD_POOL_FIELD_NUMBER': '12',\n 'SerializePartialToString': '&lt;bound method ConfigProto.SerializePartialToString of &gt;',\n 'SerializeToString': '&lt;bound method ConfigProto.SerializeToString of &gt;',\n 'SetInParent': '&lt;bound method ConfigProto.Modified of &gt;',\n 'USE_PER_SESSION_THREADS_FIELD_NUMBER': '9',\n 'WhichOneof': '&lt;bound method ConfigProto.WhichOneof of &gt;',\n 'allow_soft_placement': 'False',\n 'device_count': '{}',\n 'device_filters': '[]',\n 'gpu_options': '',\n 'graph_options': '',\n 'inter_op_parallelism_threads': '0',\n 'intra_op_parallelism_threads': '0',\n 'log_device_placement': 'False',\n 'operation_timeout_in_ms': '0',\n 'placement_period': '0',\n 'rpc_options': '',\n 'session_inter_op_thread_pool': '[]',\n 'use_per_session_threads': 'False'}\n<\/code><\/pre>\n"},{"AnswerId":"42380023","CreationDate":"2017-02-21T23:35:57.447","ParentId":null,"OwnerUserId":"2595935","Title":null,"Body":"<p>It is equivalent to the question of default values of protobufs: <a href=\"https:\/\/developers.google.com\/protocol-buffers\/docs\/proto#optional\" rel=\"nofollow noreferrer\">https:\/\/developers.google.com\/protocol-buffers\/docs\/proto#optional<\/a><\/p>\n\n<p>The default values are type specific. Usually <code>0<\/code> for numerical fields, <code>False<\/code> for boolean and empty for repeated fields, etc.<\/p>\n"}]}
{"QuestionId":42380317,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T00:05:22.890","AcceptedAnswerId":null,"OwnerUserId":6665889.0,"Title":"Tensorflow FIFOQueue '_4_batch_join\/fifo_queue' is closed and has insufficient elements","Body":"<p>tensorflow version : 1.0.0<\/p>\n\n<pre><\/pre>\n\n<p>OutOfRangeError (see above for traceback): FIFOQueue '_4_batch_join\/fifo_queue' is closed and has insufficient elements (requested 32, current size 17)<\/p>\n","answers":[{"AnswerId":"42382116","CreationDate":"2017-02-22T03:24:28.830","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>This probably happens when preparing the last batch. Add <code>allow_smaller_final_batch=True<\/code> to your <code>train.batch_join<\/code> invocation:\nfrom the documentation:<\/p>\n\n<blockquote>\n  <p>If allow_smaller_final_batch is True, a smaller batch value than\n  batch_size is returned when the queue is closed and there are not\n  enough elements to fill the batch, otherwise the pending elements are\n  discarded. In addition, all output tensors' static shapes, as accessed\n  via the get_shape method will have a first Dimension value of None,\n  and operations that depend on fixed batch_size would fail.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42380322,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-22T00:05:41.660","AcceptedAnswerId":43763655.0,"OwnerUserId":5956459.0,"Title":"Show multiple steps per image in tensorboard","Body":"<p>Is it possible to view the images for all steps in the Image viewer of tensorboard?\nOnly the images corresponding to the last step are shown in the Images tab, though they are accessible changing the index of the the url of the image:<\/p>\n\n<p>(Changing the 0 for the required step in the url: localhost:6006\/data\/individualImage?index=<strong>0<\/strong>&amp;tag=777_mask_output_output_1%2Fimage%2F0&amp;run=tensorboard_logs&amp;ts=1487721119.455751)<\/p>\n","answers":[{"AnswerId":"43763655","CreationDate":"2017-05-03T14:59:52.773","ParentId":null,"OwnerUserId":"2494561","Title":null,"Body":"<p>As I understood, an additional slider was added at TensorFlow v. 1.1.0:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/i.stack.imgur.com\/YDEVo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YDEVo.png\" alt=\"enter image description here\"><\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":42380719,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-22T00:45:03.907","AcceptedAnswerId":null,"OwnerUserId":7601838.0,"Title":"how to add operation to tensorflow(tutorial)","Body":"<p>I want to add my own operation to tensorflow.\nso I read <a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op#use_the_op_in_python\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/extend\/adding_an_op#use_the_op_in_python<\/a>.\nAnd then I tried to add new operation but it doesn't work.\nerror massage: tensorflow.python.framework.errors_impl.NotFoundError: dlopen(zero_out.so, 6): image not found<\/p>\n\n<p>I put this source code to tensorflow\/tensorflow\/core\/user_ops naming zero_out.cc.<\/p>\n\n<pre><\/pre>\n\n<p>I just write same thing with tutorial.\nand then, I made BUILD file with same path.(tensorflow\/tensorflow\/core\/user_ops)<\/p>\n\n<pre><\/pre>\n\n<p>then, in terminal(path ~\/tensorflow) I typped bazel build command.<\/p>\n\n<pre><\/pre>\n\n<p>And I tried to run my code.<\/p>\n\n<pre><\/pre>\n\n<p>But only have error massage.<\/p>\n\n<pre><\/pre>\n\n<p>I want to know what's wrong.<\/p>\n\n<p>Thanks for reading.<\/p>\n","answers":[]}
{"QuestionId":42381074,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-22T01:28:49.127","AcceptedAnswerId":null,"OwnerUserId":3084800.0,"Title":"GPU trained models' keys not found on cpu","Body":"<p>Im finding that when I try and run a model that I have trained on a gpu now on a tensorflow cpu only docker image, Im getting the following <a href=\"https:\/\/gist.github.com\/IoDmitri\/43f6283cb048cd22196a7bf0ba22f8fc\" rel=\"nofollow noreferrer\">error<\/a> <\/p>\n\n<p>Here is my code for saving my model:<\/p>\n\n<pre><\/pre>\n\n<p>Here is now the code I use to restore a model<\/p>\n\n<pre><\/pre>\n\n<p>If I use the  script here is its output:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42381537,"AnswerCount":1,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2017-02-22T02:20:36.637","AcceptedAnswerId":null,"OwnerUserId":6312225.0,"Title":"Using a function to generate TF Learn DNN objects","Body":"<p>I'm working in a Jupyter notebook on trying to generate a long list of TF Learn DNN objects for some brute force trial and error testing (I know this isn't the most efficient method, just trying to show an example).  The data follows the Titanic quickstart tutorial.<\/p>\n\n<p>I have a function that, given a bunch of parameters, should return a tflearn.DNN() objects: <\/p>\n\n<pre><\/pre>\n\n<p>Then I use that function to generate a specific model:<\/p>\n\n<pre><\/pre>\n\n<p>But I get back a lovely error message that takes me into TF Learn code, and I get lost quickly:<\/p>\n\n<pre><\/pre>\n\n<p>Is returning models from functions beyond the scope of TF Learn?  Or is there some other hurdle in the way?<\/p>\n","answers":[{"AnswerId":"50167093","CreationDate":"2018-05-04T04:18:50.833","ParentId":null,"OwnerUserId":"7596400","Title":null,"Body":"<p>Try restarting the kernel and clear output and to run it again . I too faced this issue and this solution worked for me . It is because you have run model multiple times and it had crashed .<\/p>\n"}]}
{"QuestionId":42381548,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T02:21:48.777","AcceptedAnswerId":null,"OwnerUserId":986040.0,"Title":"Tensor Flow: Simple Network Surgery (dropping first layers)","Body":"<p>Assume you have a trained network that is comprised by five layers represented as <strong>L1->L2->L3->L4->L5<\/strong>. \nHere, L1, is the <em>input<\/em> layer, comprised by a tf.placeholder. <\/p>\n\n<p>How can you <em>fix<\/em> the output of a middle layer, like L3, to a value specified by the user and run a forward pass to see the output values of L5? In other words in this scenario we would like to treat L3 as the starting input layer and ignore L1 and L2 altogether.<\/p>\n\n<p>Finally, assume that there is <strong>no<\/strong> need for a backward pass: i.e., we only want to evaluate and not train further the model.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"52629821","CreationDate":"2018-10-03T14:50:49.533","ParentId":null,"OwnerUserId":"10191188","Title":null,"Body":"<p>We need to make L1 and L2 as identity operations and pass the input directly to L3.<\/p>\n\n<ol>\n<li><p>We need to load the model using the checkpoints.<\/p>\n\n<pre><code> graph = tf.Graph()\n with graph.as_default():\n    saver = tf.train.import_meta_graph(checkpoint_path + '.meta')\n    session = tf.Session(graph=graph)\n    saver.restore(session, checkpoint_path)\n<\/code><\/pre><\/li>\n<li><p>Lets say the name of the operations in L1 and L2 are 'l1' and 'l2'.We need to take the node and change the op to identity for L1 and L2 and for L3, change the input.<\/p>\n\n<pre><code>      graph_def = graph.as_graph_def()\n      for node in graph_def.node:\n          if node.name == \"L3\":\n          # Change its input from the output of L2 to the original input tensor op name.\n              node.input[1] = \"input\"\n          if node.name == \"L1\" or node.name == \"L2\":\n              # For L1 and L2, change its op to Identity\n              node.op = \"Identity\"\n<\/code><\/pre>\n\n<p>The detailed description can be found in\n<a href=\"https:\/\/github.com\/TheMoskowitz\/tensorflow-surgery\/blob\/master\/Tensorflow_Surgery.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/TheMoskowitz\/tensorflow-surgery\/blob\/master\/Tensorflow_Surgery.ipynb<\/a><\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42381661,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network>","CreationDate":"2017-02-22T02:33:30.383","AcceptedAnswerId":null,"OwnerUserId":7602147.0,"Title":"Trouble implementing super-resolution CNN","Body":"<p>So I'm currently trying to implement a fast super-resolution CNN (<a href=\"https:\/\/arxiv.org\/pdf\/1608.00367.pdf\" rel=\"nofollow noreferrer\">this paper<\/a>) by modifying <a href=\"https:\/\/github.com\/tegg89\/SRCNN-Tensorflow\" rel=\"nofollow noreferrer\">this repository<\/a> (a tensorflow implementation of the original super-resolution CNN).<\/p>\n\n<p>The problem is that the network instantly reaches a high loss after a few epochs and then stops learning immediately, no matter how many times I reset the network it always converges to the exact same high loss. If I try to feed-forward an image the result ends up being a shade of gray.<\/p>\n\n<p>On the other hand though if I hook up the first convolution layer directly to the final deconvolution layer the network actually trains and feed-forwarding creates a new up-scaled image. Of course this network is too shallow to actually learn any real features though.<\/p>\n\n<p>So what I'm wondering is what's going wrong between my first convolution layer conv1 and my last layer conv8?<\/p>\n\n<p>These are the network layers:<\/p>\n\n<pre><\/pre>\n\n<p>With their respective weights and biases:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"42383358","CreationDate":"2017-02-22T05:26:40.470","ParentId":null,"OwnerUserId":"7602147","Title":null,"Body":"<p>Found the problem, I was initializing my weights according to a gaussian distribution with standard deviation 0.001. This worked for the original SRCNN paper because it had less layers but in my deeper network it was causing the gradient to vanish. The initialization scheme I ended up using comes from <a href=\"https:\/\/arxiv.org\/pdf\/1502.01852.pdf\" rel=\"nofollow noreferrer\">this paper<\/a> on PreLU optimization.<\/p>\n"}]}
{"QuestionId":42382535,"AnswerCount":1,"Tags":"<csv><tensorflow><tensorboard>","CreationDate":"2017-02-22T04:07:31.093","AcceptedAnswerId":42385577.0,"OwnerUserId":7602394.0,"Title":"How can I feed batch CSV data for tensorflow learning","Body":"<p>Thanks for read my question<\/p>\n\n<p>I have data like these.\n19 data for input and 1 for label<\/p>\n\n<p>I already had tried mnist, tensorboard example and csv batch load example.\nNow I'm trying mixed it all.<\/p>\n\n<p>load csv data and make it batch.\nlearn just 1 layer and check cost.\nthat is all I want to do<\/p>\n\n<p>this is my code<\/p>\n\n<pre><\/pre>\n\n<p>I checked batch data correctly feed to the cross_entropy.\nbut If i use these<\/p>\n\n<pre><\/pre>\n\n<p>I had Got a error message like this<\/p>\n\n<pre><\/pre>\n\n<p>what's wrong in my code?<\/p>\n","answers":[{"AnswerId":"42385577","CreationDate":"2017-02-22T07:50:23.940","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>One thing that is wrong with your code is the initialization of your weights... You should give them random weights (otherwise your network will not learn). The summary talks about there being a NaN (Not a number) in your weights. Try running W.eval() to see what exactly is in your weights matrix!<\/p>\n"}]}
{"QuestionId":42382712,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T04:27:02.583","AcceptedAnswerId":null,"OwnerUserId":7572130.0,"Title":"How to delete existed tensorflow variable?","Body":"<p>I am using windows GPU tensorflow 1.0. I want to delete useless or wrong tf variable.     <\/p>\n\n<pre><\/pre>\n\n<p>Could you tell me how to delete old variable W1 and add new W1?<\/p>\n\n<p>(the newline edit function is useless)<\/p>\n","answers":[{"AnswerId":"42385497","CreationDate":"2017-02-22T07:46:01.737","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>First of all, I don't know exactly what you are trying to achieve by changing the shape of your variable. If you would have asked the question in such a way we knew your intent, maybe we could help you better. <\/p>\n\n<p>Now there are multiple options to solve your problem: \n- ignore the name of your variable (what bugs does it cause??)\n- remove all variables: <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework\/utility_functions#reset_default_graph\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework\/utility_functions#reset_default_graph<\/a>\n- Change the shape of the variable instead of defining a new one. Check the answer of mrry here: <a href=\"https:\/\/stackoverflow.com\/questions\/33654754\/how-can-i-change-the-shape-of-a-variable-in-tensorflow\">How can I change the shape of a variable in TensorFlow?<\/a><\/p>\n"}]}
{"QuestionId":42383490,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-ml><google-apis-explorer>","CreationDate":"2017-02-22T05:37:00.400","AcceptedAnswerId":null,"OwnerUserId":3064462.0,"Title":"{\"error\": \"Error loading the model\"} when using \/ml\/v1beta1\/ml.projects.predict","Body":"<p>Using the following API Explorer and body,I get the error \n{\"error\": \"Error loading the model\"}.  I was going to start using <a href=\"https:\/\/developers.google.com\/resources\/api-libraries\/documentation\/ml\/v1beta1\/python\/latest\/ml_v1beta1.projects.html#predict\" rel=\"nofollow noreferrer\">https:\/\/developers.google.com\/resources\/api-libraries\/documentation\/ml\/v1beta1\/python\/latest\/ml_v1beta1.projects.html#predict<\/a>, but would like to verify everything is okay first.  <\/p>\n\n<p>Is there a way to see the actual error? <\/p>\n\n<hr>\n\n<p><a href=\"https:\/\/developers.google.com\/apis-explorer\/?authuser=1#p\/ml\/v1beta1\/ml.projects.predict\" rel=\"nofollow noreferrer\">https:\/\/developers.google.com\/apis-explorer\/?authuser=1#p\/ml\/v1beta1\/ml.projects.predict<\/a>?<\/p>\n\n<p>POST <a href=\"https:\/\/ml.googleapis.com\/v1beta1\/projects\/\" rel=\"nofollow noreferrer\">https:\/\/ml.googleapis.com\/v1beta1\/projects\/<\/a>{project}\/models\/{model_name}\/versions\/v1:predict?key={YOUR_API_KEY}<\/p>\n\n<p>{\n \"httpBody\": {\n  \"data\": \"[{\\\"placeholder_name\\\": [44, 158, 178, 156, 111, 101, 110, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], \\\"key\\\": 1}]\"\n }\n}<\/p>\n\n<p>200<\/p>\n\n<p>cache-control:  private\ncontent-encoding:  gzip\ncontent-length:  53\ncontent-type:  text\/html; charset=utf-8\ndate:  Wed, 22 Feb 2017 05:25:14 GMT\nserver:  ESF\nvary:  Origin, X-Origin, Referer<\/p>\n\n<p>{\"error\": \"Error loading the model\"}\n`<\/p>\n","answers":[{"AnswerId":"42418687","CreationDate":"2017-02-23T14:35:14.630","ParentId":null,"OwnerUserId":"3064462","Title":null,"Body":"<p>I added export.meta, export.index and a renamed export.data-00000-of-00001 -> export too cloud storage.<\/p>\n\n<p>It was an incorrect guess based on the documentation <a href=\"https:\/\/cloud.google.com\/ml\/docs\/concepts\/deployment-overview#deployment_location\" rel=\"nofollow noreferrer\">https:\/\/cloud.google.com\/ml\/docs\/concepts\/deployment-overview#deployment_location<\/a>.<\/p>\n\n<p>So it looks like you just need the files created by <code>saver.save(sess, os.path.join(FLAGS.model_dir, 'export'))<\/code><\/p>\n"}]}
{"QuestionId":42384412,"AnswerCount":0,"Tags":"<caffe><imagenet>","CreationDate":"2017-02-22T06:42:06.347","AcceptedAnswerId":null,"OwnerUserId":1279459.0,"Title":"Cannot find mean file while training on Imagenet","Body":"<p>I am trying to train and validate a network on Imagenet. The validation process works without any problems (with the pretrained weights). However, when I try to perform the training, there appears an error that the  file is not found; the very same file that has worked for the valiudation process. What is wrong?<\/p>\n\n<pre><\/pre>\n\n<p>Here is the prototxt I am using:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42384602,"AnswerCount":1,"Tags":"<keras><image-segmentation><resnet>","CreationDate":"2017-02-22T06:54:33.997","AcceptedAnswerId":42391339.0,"OwnerUserId":6691357.0,"Title":"Implementing skip connections in keras","Body":"<p>I am implementing ApesNet in keras. It has an ApesBlock that has skip connections. How do I add this to a sequential model in keras? The ApesBlock has two parallel layers that merge at the end by element-wise addition.<img src=\"https:\/\/i.stack.imgur.com\/UrFP8.png\" alt=\"enter image description here\"><\/p>\n","answers":[{"AnswerId":"42391339","CreationDate":"2017-02-22T12:20:40.817","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The easy answer is don't use a sequential model for this, use the functional API instead, implementing skip connections (also called residual connections) are then very easy, as shown in this example from the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"noreferrer\">functional API guide<\/a>:<\/p>\n\n<pre><code>from keras.layers import merge, Convolution2D, Input\n\n# input tensor for a 3-channel 256x256 image\nx = Input(shape=(3, 256, 256))\n# 3x3 conv with 3 output channels (same as input channels)\ny = Convolution2D(3, 3, 3, border_mode='same')(x)\n# this returns x + y.\nz = merge([x, y], mode='sum')\n<\/code><\/pre>\n"}]}
{"QuestionId":42385106,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-02-22T07:23:27.303","AcceptedAnswerId":42392218.0,"OwnerUserId":2143228.0,"Title":"keras model fit_generator ValueError: Error when checking model target: expected cropping2d_4 to have 4 dimensions, but got array with shape (32, 1)","Body":"<p>I'm trying to use keras model.fit_generator() to fit a model, below is my definition of the generator:<\/p>\n\n<pre><\/pre>\n\n<p>Here the output shape is:\nX_train shape:  (32, 160, 320, 3)  y_train shape: (32,)<\/p>\n\n<p>The model fit code is:<\/p>\n\n<pre><\/pre>\n\n<p>Then I get the error message: <\/p>\n\n<p>ValueError: Error when checking model target: expected cropping2d_6 to have 4 dimensions, but got array with shape (32, 1)<\/p>\n\n<p>Could someone help let me know what's the issue?<\/p>\n","answers":[{"AnswerId":"42392218","CreationDate":"2017-02-22T13:00:07.747","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>The big question here is : do you know what you are trying to do ? <\/p>\n\n<p>1) If you read <a href=\"https:\/\/keras.io\/layers\/convolutional\/#cropping2d\" rel=\"nofollow noreferrer\">here<\/a>, the input is a 4D tensor and the output is ALSO a 4D tensor. Your target is a 2D tensor of shape (batch_size,1). So of course, when keras tries to compute the error between the output which has 3D (without batch dimension) and the target which has 1D (without batch dimension), it can not make sense out of that. Outputs and targets must have the same dimensions.<\/p>\n\n<p>2) Do you know what cropping2D is actually doing ? It is cropping your images... So removing values at the beginning and end of your cropping dimensions. In your case you are outputing images of shape (90, 218, 3). This is not a prediction, there is no weight to train on this layer so no reason to fit the \"model\". Your model is just cropping images. No training needed for that.<\/p>\n"}]}
{"QuestionId":42385125,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><distributed>","CreationDate":"2017-02-22T07:24:23.217","AcceptedAnswerId":null,"OwnerUserId":7603030.0,"Title":"Run distributed tensorflow example with error","Body":"<p>I have three node to run a distributed tensorflow, which is two worker(one has GPU,one not)and one ps(without GPU).The code is below:<\/p>\n\n<pre><\/pre>\n\n<p>I run the above code on my three node with instruction below in terminal:<\/p>\n\n<pre><\/pre>\n\n<p>However, after the Variables initialized, I met a question that the terminal of worker always print :<\/p>\n\n<pre><\/pre>\n\n<p>and the terminal of ps don't proceed.\nThe IP of ps is 192.168.1.102, and the IP of the worker is 192.168.1.103,192.168.1.104,just like the code above.\nAnyone can help me?<\/p>\n","answers":[{"AnswerId":"44660190","CreationDate":"2017-06-20T18:09:10.443","ParentId":null,"OwnerUserId":"5251457","Title":null,"Body":"<p>I guess filtering out device should help here. Could you please try adding  device_filter to your session ?<\/p>\n\n<pre><code>        config = tf.ConfigProto(\n        allow_soft_placement=True,\n        log_device_placement=False,\n        device_filters=[\"\/job:ps\", \"\/job:worker\/task:%d\" % FLAGS.task_index])\n\nwith sv.prepare_or_wait_for_session(server.target, config=\\config) as sess:\n<\/code><\/pre>\n\n<p>This should fix the issue.<\/p>\n"}]}
{"QuestionId":42385199,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T07:28:20.280","AcceptedAnswerId":null,"OwnerUserId":7603240.0,"Title":"Why I am not getting the Hello Tensorflow output?","Body":"<p>I couldn't get the output Hello Tesorflow? <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/qbeDn.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/qbeDn.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42385403","CreationDate":"2017-02-22T07:40:30.337","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Well, you do get your output... See the last line...<\/p>\n\n<p>Maybe it is a good idea to check your installation\/reinstall Tensorflow. <\/p>\n"}]}
{"QuestionId":42385720,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T08:00:00.747","AcceptedAnswerId":null,"OwnerUserId":6334174.0,"Title":"Cannot set feed_previous to tf.bool in embedding_attention_seq2seq","Body":"<p>I am using the  function and I try to set the  argument to , however, this does not work. Error messages:<\/p>\n\n<pre><\/pre>\n\n<p>Code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44465503","CreationDate":"2017-06-09T19:36:17.373","ParentId":null,"OwnerUserId":"7845325","Title":null,"Body":"<p>It looks like that the copy.deepcopy(cell) method does not support placeholders. Since this method is used to copy the RNN model specified as the cell parameter so that both the encoder and decoder can use the same architecture, and it does not give you any other alternatives to specify them separately, the only possibility, in case you do need to use placeholders, you may have to reproduce this seq2seq module by yourself (by copying the source code and make slight changes). In your own implementation, you just create two separate models for the encoder and decoder.<\/p>\n"}]}
{"QuestionId":42385769,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-22T08:02:13.673","AcceptedAnswerId":null,"OwnerUserId":5107084.0,"Title":"TensorFlow: How does tf.set_shape change the channel information of grayscale images?","Body":"<p>If I have a grayscale image that undergoes a preprocessing that has its shape shaped as , where does the additional channel information come from? Does the channel information actually replicate? I am referring to the following preprocessing that could turn a grayscale image into an image with 3 channels:<\/p>\n\n<pre><\/pre>\n\n<p>if you look at the code line , this is what that changed my grayscale image into 3 channels. Where does the additional information come from, and if I were to fine-tune a pre-trained model on grayscale images, should I let this happen? Because if the image is 3 channels, I can reuse the variables from the input layer scope for fine-tuning.<\/p>\n\n<p>Another source of error may come from the  decoder found in this line of the code: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/data\/tfexample_decoder.py#L269\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/slim\/python\/slim\/data\/tfexample_decoder.py#L269<\/a><\/p>\n\n<p>If I set the number of channels to 3, the channels really ended up with 3 channels although there was only 1 channel initially. What happens to the information here?<\/p>\n","answers":[]}
{"QuestionId":42386463,"AnswerCount":1,"Tags":"<python><tensorflow><keras><jupyter-notebook>","CreationDate":"2017-02-22T08:39:57.347","AcceptedAnswerId":42390626.0,"OwnerUserId":6487788.0,"Title":"Keras jupyter notebook outputs blocks during training","Body":"<p>When using Keras in a Jupyter Notebook environment, during training my Notebook crashes. While running the fit function, apart from the normal training progress, a lot of squares appear. In fact, there are so many squares the browser crashes...<\/p>\n\n<p>When running the same code in a normal terminal these squares do not appear. <\/p>\n\n<p>Here is an example of the output: <\/p>\n\n<p>Train on 33164 samples, validate on 8292 samples\nEpoch 1\/3\n  960\/33164 [..............................] - ETA: 46s - loss: 0.1997 - acc: 0.9394 <\/p>\n\n<p>Hope anybody can help!<\/p>\n","answers":[{"AnswerId":"42390626","CreationDate":"2017-02-22T11:52:22.800","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>see this <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/4676\" rel=\"noreferrer\">issue on github<\/a><\/p>\n\n<p>in short 4 options :<\/p>\n\n<ul>\n<li>Use command lines instead of ipython notebooks<\/li>\n<li>Use <code>verbose=0<\/code> or <code>verbose=2<\/code> in <code>fit()<\/code>, the first will not output anything, the second will only output one log line per epoch.<\/li>\n<li>If you use tensorflow as backend you can monitor everything on tensorboard (slow but better than nothing)<\/li>\n<li>You can use the <a href=\"https:\/\/keras.io\/callbacks\/#remotemonitor\" rel=\"noreferrer\">RemoteMonitor callback<\/a> but this seems to require a bit of work.<\/li>\n<\/ul>\n\n<p>I hope this helps.<\/p>\n"}]}
{"QuestionId":42386493,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-22T08:41:22.703","AcceptedAnswerId":null,"OwnerUserId":5497286.0,"Title":"Tensorflow: concatenating 2 tensors of shapes containing None","Body":"<p>My problem is the following: I have a tensor  of shape  () and a tensor . At runtime it is <strong>guaranteed<\/strong> that  and  will have the same shape. I would like to concatenate these two tensors along  axis.<\/p>\n\n<p>PS. Note that I simplified my original problem - so having a tensor of shape  in the first place is not an option.<\/p>\n","answers":[{"AnswerId":"42386754","CreationDate":"2017-02-22T08:55:17.300","ParentId":null,"OwnerUserId":"5497286","Title":null,"Body":"<p>Before asking a question I should probably try to run the code :) Using <code>tf.concat(values=[A, B], concat_dim=3)<\/code> seems to be working.<\/p>\n"}]}
{"QuestionId":42386650,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2017-02-22T08:50:06.940","AcceptedAnswerId":42851839.0,"OwnerUserId":5244666.0,"Title":"Why am I getting error: Initializing Tensorflow?","Body":"<p>Im trying to build .apk file by this tutorial - <a href=\"https:\/\/www.oreilly.com\/learning\/tensorflow-on-android\" rel=\"nofollow noreferrer\">https:\/\/www.oreilly.com\/learning\/tensorflow-on-android<\/a><\/p>\n\n<p>I want to use my own retrained_graph.pb<\/p>\n\n<p>I change  (in tutorial its a , but I didnt find this file).\nI replaced:<\/p>\n\n<pre><\/pre>\n\n<p>But when I run apk on the device - it crashes with .<\/p>\n\n<pre><\/pre>\n\n<p>I found that this code handles exception:<\/p>\n\n<pre><\/pre>\n\n<p>in <\/p>\n\n<p>I spent already 2 days, guys help me please! What can be reason of the error?<\/p>\n","answers":[{"AnswerId":"42851839","CreationDate":"2017-03-17T07:40:22.643","ParentId":null,"OwnerUserId":"3378821","Title":null,"Body":"<p>I solved the problem, the comment in the java code is not updated, now just change:<\/p>\n\n<pre><code>private static final String INPUT_NAME = \"Mul:0\";\nprivate static final String OUTPUT_NAME = \"final_result:0\";\n<\/code><\/pre>\n\n<p>To:<\/p>\n\n<pre><code>private static final String INPUT_NAME = \"Mul\";\nprivate static final String OUTPUT_NAME = \"final_result\";\n<\/code><\/pre>\n"}]}
{"QuestionId":42386797,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-02-22T08:57:20.770","AcceptedAnswerId":42635977.0,"OwnerUserId":6311081.0,"Title":"TFLearn import error","Body":"<p>I have installed tensorflow with docker in my windows machine. I am opening the tensorflow with powershell <\/p>\n\n<blockquote>\n  <p>docker run -it -p 8888:8888 tensorflow\/tensorflow<\/p>\n<\/blockquote>\n\n<p>It's opening on a ipython notebook. But when I want to import<\/p>\n\n<blockquote>\n  <p>import tflearn\n  ImportError: No module named tflearn.<\/p>\n<\/blockquote>\n\n<p>After that I have installed tflearn with<\/p>\n\n<blockquote>\n  <p>pip install tflearn<\/p>\n<\/blockquote>\n\n<p>But getting still same error. What is the problem anyone know that?<\/p>\n","answers":[{"AnswerId":"42635977","CreationDate":"2017-03-06T21:33:43.750","ParentId":null,"OwnerUserId":"1869715","Title":null,"Body":"<p>I see that you are having trouble installing TFLearn inside tensorflow docker container. In order to do that you can simply launch a new terminal from the Ipython Notebook interface and install your package there using pip.\n<a href=\"https:\/\/i.stack.imgur.com\/lG92u.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lG92u.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>To install TFLearn, use the same command that you used <code>pip install tflearn<\/code>.<br>\nIf pip is not installed by default, you can do that by <code>sudo apt-get install python-pip<\/code>.<\/p>\n"}]}
{"QuestionId":42387156,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T09:13:18.030","AcceptedAnswerId":42401657.0,"OwnerUserId":7345996.0,"Title":"what is the meaning of shadow variable in tensorflow and what is it used for?","Body":"<p>I am reading the tensroflow document. In <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/train\/moving_averages\" rel=\"nofollow noreferrer\">this<\/a> page, I have encounted a term shadow variable,i couldn't figure out what is the meaning of it and what is it used for?<\/p>\n","answers":[{"AnswerId":"42401657","CreationDate":"2017-02-22T20:30:56.717","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>It says it in the document you linked: <\/p>\n\n<blockquote>\n  <p>The apply() method adds shadow copies of trained variables and add ops\n  that maintain a moving average of the trained variables in their\n  shadow copies<\/p>\n<\/blockquote>\n\n<p>Shadow Variables are <em>copies<\/em> of the main train variables that are used to mantain the moving average. The idea is that they <em>follow<\/em> the main Variable like a <em>Shadow<\/em>  ^_^<\/p>\n"}]}
{"QuestionId":42387584,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-02-22T09:30:00.460","AcceptedAnswerId":42659009.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Is it possible to restore checkpoint models for multi-gpu training?","Body":"<p>I am currently using a supervisor and constructed just one graph to perform transfer learning using the pre-trained weights from TF-slim. I am wondering if there is a way to restore checkpoint models to multiple inference models at the outset? My primary concern is that firstly, the name scopes that are defined as in a <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/image\/cifar10\/cifar10_multi_gpu_train.py\" rel=\"nofollow noreferrer\">reference code<\/a> on the TF repository may cause the pre-trained variables to be unable to be restored due to a name mismatch. Also, given that I have to use a supervisor with an  that takes in only one saver that restores the variables, how could I have multiple savers to restore the same variables to multiple GPUs (If I even need to have multiple savers at all).<\/p>\n\n<p>One idea I have is that perhaps I could just restore the variables to one graph, and let the other GPUs use the same graph for training. However, would the training for the next GPU take place only after the first GPU has completed? But this way, I won't be able to restore the weights according to the original checkpoint model variable names too, unless I edit the names of the checkpoint weights.<\/p>\n","answers":[{"AnswerId":"42659009","CreationDate":"2017-03-07T21:50:06.290","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variables\" rel=\"nofollow noreferrer\">tensorflow documentation on saving and restoring variables<\/a> points you to the saver object allowing you to specify what saved variables get restored as what model variables, by <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver#__init__\" rel=\"nofollow noreferrer\">passing a dictionary from saved name to variable object when constructing the saver<\/a>.<\/p>\n"}]}
{"QuestionId":42387757,"AnswerCount":1,"Tags":"<python><numpy><machine-learning><tensorflow><neural-network>","CreationDate":"2017-02-22T09:37:00.387","AcceptedAnswerId":null,"OwnerUserId":4504245.0,"Title":"Feeding timeseries data into Tensorflow for LSTM classifier training","Body":"<p>I have a dataframe of shape  with timestamp as index:<\/p>\n\n<p>I'm trying to implement a LSTM classifier but I'm struggling to feed it into the DataFlow<\/p>\n\n<p>The final arrays I'm trying to feed are of shape '(X_train = (38307, 25), y_train = (38307, 2))' <\/p>\n\n<p>I have added the code in case <\/p>\n\n<pre><\/pre>\n\n<p>EDIT the RNN function:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42388213","CreationDate":"2017-02-22T09:57:03.160","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Unfortunately, the most important part of your code, is hidden in the RNN function. <\/p>\n\n<p>Some tips to help you out: I guess you are trying to build a dynamic RNN... (is that correct? ) In that case, a common mistake I see is that people confuse the time major and batch major setting of these RNNs. In other words, is you input data [batch,time,variables], or [time,batch,variables]. \nMore about this can be found here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard8\/tf.nn.dynamic_rnn.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard8\/tf.nn.dynamic_rnn.md<\/a><\/p>\n"}]}
{"QuestionId":42387999,"AnswerCount":0,"Tags":"<android><tensorflow><bazel>","CreationDate":"2017-02-22T09:47:08.733","AcceptedAnswerId":null,"OwnerUserId":7591349.0,"Title":"Bazel build error in tensorflow","Body":"<p>I am using TensorFlow Android Camera Demo. I follow step given in README file.I do following things:<\/p>\n\n<pre><\/pre>\n\n<p>While I run bazel build command :\n   bazel build -c opt \/\/tensorflow\/examples\/android:tensorflow_demo<\/p>\n\n<p>I getting error that <strong>Couldn't find java at '$(ls -d C:\/Program\\ Files\/Java\/jdk* | sort | tail -n 1)\/bin\/java.exe'<\/strong>.<\/p>\n\n<p>My JDK is installed at  . I also set environment variable \n and  <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42388299,"AnswerCount":1,"Tags":"<function><theano><pymc3>","CreationDate":"2017-02-22T10:00:29.307","AcceptedAnswerId":null,"OwnerUserId":7603665.0,"Title":"Theano custom Op for PyMC3","Body":"<p>I tried to implement a custom Op to which I get the \"Bad input argument to theano function\" error. Here is the code. The problem as I understand is: how to convert the PyMC3 variables into theano understandable type?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43679607","CreationDate":"2017-04-28T11:54:29.210","ParentId":null,"OwnerUserId":"7344908","Title":null,"Body":"<p>I think I can answer your question.<\/p>\n\n<p><strong>Firstly<\/strong>, you should not use <code>f = theano.function( [x1,x2,x3], trial_Op()(x1,x2,x3) )<\/code>. Once it defined, <code>f<\/code> takes numeric values as arguments. However in the pymc3 model, <code>x1<\/code> and <code>x2<\/code> defined as <code>Normal<\/code> are not numeric but symbolic. So it will throw the error you just faced. If you are familiar with <em><code>@as_op<\/code><\/em> method documented in the tutorial, the solution is simple: change <code>sum_of_x1_x2_x3 = f(x1,x2,x3)<\/code> to <code>sum_of_x1_x2_x3 = Trial_Op(x1,x2,x3)<\/code> ;<\/p>\n\n<p><strong>secondly<\/strong>, in your code, it seems <code>dmatrix<\/code> type is not necessary. so I modify the code following:        <\/p>\n\n<pre><code>N = 20 #data array length\n\nclass Trial_Op(theano.Op):\n    __props__ = ()\n    itypes = [t.dscalar, t.dscalar]\n    otypes = [t.dvector] #if the data has multiple values i.e. data array\n\n    def perform(self,node,inputs,output_storage):\n        x0 = inputs[0]\n        x1 = inputs[1]\n\n        f = np.add(x0,x1)\n        out = np.empty(N)\n        out[:] = f\n        z = output_storage[0]\n        z[0] = out\n\n\ntrial_Op = Trial_Op()\n\n\nimport pymc3 as pm\ntrue_Data = np.random.normal(2,1,N)\n\nwith pm.Model() as model:\n    x1 = pm.Normal('x1', mu = 0, sd = 0.1)\n    x2 = pm.Normal('x2', mu = 3, sd = 0.5)    \n\n    mu = trial_Op(x1, x2)\n    z = pm.Normal('z', mu = mu, sd = 1., observed = true_Data)\n\n    step = pm.Metropolis()\n    trace = pm.sample(10000, step)\n\npm.traceplot(trace)\n<\/code><\/pre>\n\n<p>Note in the custom function, the <code>otypes<\/code> is <code>dvector<\/code> to satisfy the <code>mu<\/code> in the  <code>z = pm.Normal('z', mu = mu, sd = 1., observed = true_Data)<\/code>. And the sample number expands to 10000. The result image shows:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/VMTkz.png\" rel=\"nofollow noreferrer\">result<\/a><\/p>\n\n<p><strong>However<\/strong>, I don't know how to define the <em><code>grad()<\/code><\/em> method in the custom theano function. Maybe later someone or me can solve it to enable NUTS sampling method in the model :).<\/p>\n"}]}
{"QuestionId":42388563,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><word-embedding>","CreationDate":"2017-02-22T10:10:50.497","AcceptedAnswerId":null,"OwnerUserId":6059430.0,"Title":"Tensorflow: \"GraphDef cannot be larger than 2GB.\" error when saving model after assigning variables","Body":"<p>I want to use a pretrained model to warmly start another model with a little difference. Simply, I create a new model, and assign the variables with same name with pretrained model weights. But, when saving the model, error occurred.<\/p>\n\n<p><\/p>\n\n<p>The example code is as follow:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42401742","CreationDate":"2017-02-22T20:35:55.803","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<blockquote>\n  <p>There is a hard limit of 2GB for serializing individual tensors\n  because of the 32bit signed size in protobuf.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4291\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4291<\/a><\/p>\n"},{"AnswerId":"42906762","CreationDate":"2017-03-20T14:51:18.247","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Fabrizio <a href=\"https:\/\/stackoverflow.com\/a\/42401742\/3574081\">correctly points out<\/a> that there's a hard 2GB limit on the size of protocol buffers, but you might be wondering <em>why<\/em> your program hits that limit. The problem stems from these lines: <\/p>\n\n<pre><code>for v in tf.trainable_variables():\n  embedding = np.random.uniform(-1, 1, (400000, 1024))\n  sess.run(v.assign(embedding))\n<\/code><\/pre>\n\n<p>When the execution hits <code>v.assign(embedding)<\/code>, new nodes are added to the TensorFlow graph. In particular, each <code>embedding<\/code> array is converted to a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"nofollow noreferrer\"><code>tf.constant()<\/code><\/a> tensor, which will be quite large (approximately 328MB by my estimate).<\/p>\n\n<p>The best way to avoid this is to load the variables from the previous model directly into your new model using a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"nofollow noreferrer\"><code>tf.train.Saver<\/code><\/a>. Since the models might have a different structure, you might need to specify a mapping from the names of variables in the old model to the <code>tf.Variable<\/code> objects in your new model.<\/p>\n\n<hr>\n\n<p>An alternative way to solve your problem would be to pre-create a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"nofollow noreferrer\"><code>tf.placeholder()<\/code><\/a> op for assigning a value to each variable. This might require more restructuring of your actual code, but the following worked for me:<\/p>\n\n<pre><code>v1 = tf.get_variable(\"L_enc\", [400000, 1024])\nv2 = tf.get_variable(\"L_dec\", [400000, 1024])\n\n# Define a separate placeholder and assign op for each variable, so\n# that we can feed the initial value without adding it to the graph.\nvars = [v1, v2]\nplaceholders = [tf.placeholder(tf.float32, shape=[400000, 1024]) for v in vars]\nassign_ops = [v.assign(p) for (v, p) in zip(vars, placeholders)]\n\ninit_op = tf.global_variables_initializer()\n\nsaver = tf.train.Saver(tf.all_variables())\n\nwith tf.Session() as sess:\n  sess.run(init_op)\n  for p, assign_op in zip(placeholders, assign_ops):\n    embedding = np.random.uniform(-1, 1, (400000, 1024))\n    sess.run(assign_op, {p: embedding})\n\n  # Save the variables to disk.\n  save_path = saver.save(sess, \".\/model.ckpt\")\n  print(\"Model saved in file: %s\" % save_path)\n<\/code><\/pre>\n"}]}
{"QuestionId":42388935,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-22T10:42:56.067","AcceptedAnswerId":null,"OwnerUserId":5841587.0,"Title":"Batching for tf.learn input_fn in tensorflow 1.0","Body":"<p>When using an input function to feed training data to tf.learn what is the best method in tf 1.0?<\/p>\n\n<p>I was using learn.dataframe.queues.feeding_functions.enqueue_data in tf0.12 but having explored tf.train.batch, tf.train.slice_input_producer and tf.FIFOQueue I am not sure how to rewrite this for tf1.0.<\/p>\n\n<p>The code I was using is<\/p>\n\n<pre><\/pre>\n\n<p>from <a href=\"https:\/\/medium.com\/@ilblackdragon\/tensorflow-tutorial-part-4-958c29c717a0#.4vwo2ato7\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@ilblackdragon\/tensorflow-tutorial-part-4-958c29c717a0#.4vwo2ato7<\/a><\/p>\n","answers":[]}
{"QuestionId":42389754,"AnswerCount":2,"Tags":"<tensorflow><graph-visualization><tensorboard>","CreationDate":"2017-02-22T11:16:55.410","AcceptedAnswerId":42401817.0,"OwnerUserId":5648649.0,"Title":"Tensor board Visualization","Body":"<p>is there anything wrong with this code, cause this is generating an output folder with tf file inside, but when I am trying to run it on tensorboard its showing- No Graph definition file found.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42396585","CreationDate":"2017-02-22T16:06:20.350","ParentId":null,"OwnerUserId":"7604321","Title":null,"Body":"<p>Two things. Firstly, do you get any error running the code because there's no indentation in the with: block.<\/p>\n\n<p>Secondly, make sure that you are pointing Tensorboard to the correct directory in the correct location. Is anything showing up at all in Tensorboard?<\/p>\n"},{"AnswerId":"42401817","CreationDate":"2017-02-22T20:39:51.017","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<ul>\n<li>make sure there is a <code>output<\/code> directory in the same directory where you are running that script<\/li>\n<li>make sure some files are inside <code>.\/output<\/code><\/li>\n<li>launch tensorboard: <code>tensorboard \u2014logdir=.\/output<\/code> from the same directory<\/li>\n<\/ul>\n"}]}
{"QuestionId":42389822,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning><bigdata>","CreationDate":"2017-02-22T11:20:05.123","AcceptedAnswerId":42390693.0,"OwnerUserId":6751201.0,"Title":"What should I do if I want to use large datasets that can't load into the memory with TensorFlow?","Body":"<p>I want to use a large dataset that cannot load into the memory once to train a model with TensorFlow. But I don't know what exacty I should do.<\/p>\n\n<p>I have read some great posts about  file format and the official document. Bus I still can't figure it out. <\/p>\n\n<p>Is there a complete solution plan with TensorFlow?<\/p>\n","answers":[{"AnswerId":"42390268","CreationDate":"2017-02-22T11:37:22.823","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Normally you use a batch wise training anyways so you can load the data on the fly. For example for images:<\/p>\n\n<pre><code>for bid in nrBatches:\n     batch_x, batch_y = load_data_from_hd(bid)\n     train_step.run(feed_dict={x: batch_x, y_: batch_y})\n<\/code><\/pre>\n\n<p>So you load every batch on the fly and only load the data which you need to load at any given moment. Naturally your training time will increase while using the harddisk instead of memory to load data.<\/p>\n"},{"AnswerId":"42390693","CreationDate":"2017-02-22T11:55:13.190","ParentId":null,"OwnerUserId":"281089","Title":null,"Body":"<p>Consider using <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/io_ops\" rel=\"nofollow noreferrer\"><code>tf.TextLineReader<\/code><\/a> which in conjunction with <code>tf.train.string_input_producer<\/code> allows you to load data from multiple files on disk (if your dataset is large enough that it needs to be spread out into multiple files).<\/p>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data#reading_from_files\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/reading_data#reading_from_files<\/a><\/p>\n\n<p>Code snippet from the link above:<\/p>\n\n<pre><code>filename_queue = tf.train.string_input_producer([\"file0.csv\", \"file1.csv\"])\n\nreader = tf.TextLineReader()\nkey, value = reader.read(filename_queue)\n\n# Default values, in case of empty columns. Also specifies the type of the\n# decoded result.\nrecord_defaults = [[1], [1], [1], [1], [1]]\ncol1, col2, col3, col4, col5 = tf.decode_csv(\n    value, record_defaults=record_defaults)\nfeatures = tf.stack([col1, col2, col3, col4])\n\nwith tf.Session() as sess:\n  # Start populating the filename queue.\n  coord = tf.train.Coordinator()\n  threads = tf.train.start_queue_runners(coord=coord)\n\n  for     filename_queue = tf.train.string_input_producer([\"file0.csv\", \"file1.csv\"])\n\nreader = tf.TextLineReader()\nkey, value = reader.read(filename_queue)\n\n# Default values, in case of empty columns. Also specifies the type of the\n# decoded result.\nrecord_defaults = [[1], [1], [1], [1], [1]]\ncol1, col2, col3, col4, col5 = tf.decode_csv(\n    value, record_defaults=record_defaults)\nfeatures = tf.stack([col1, col2, col3, col4])\n\nwith tf.Session() as sess:\n  # Start populating the filename queue.\n  coord = tf.train.Coordinator()\n  threads = tf.train.start_queue_runners(coord=coord)\n\n  for i in range(1200):\n    # Retrieve a single instance:\n    example, label = sess.run([features, col5])\n\n  coord.request_stop()\n  coord.join(threads)i in range(1200):\n    # Retrieve a single instance:\n    example, label = sess.run([features, col5])\n\n  coord.request_stop()\n  coord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":42390543,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-22T11:48:47.963","AcceptedAnswerId":42401238.0,"OwnerUserId":281089.0,"Title":"Enable XLA when using tf.contrib.learn.Estimator","Body":"<p>Seeing the <a href=\"https:\/\/youtu.be\/kAOanJczHA0\" rel=\"nofollow noreferrer\">TF Dev Summit talk<\/a> on the potential (and highly experimental) benefits of turning on XLA on TensorFlow graphs, I thought I'd experiment with it a bit.<\/p>\n\n<p><strong>Question<\/strong>: When using <a href=\"https:\/\/www.tensorflow.org\/get_started\/tflearn\" rel=\"nofollow noreferrer\"><\/a>, how do I enable <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/experimental\/xla\/jit#turning_on_jit_compilation\" rel=\"nofollow noreferrer\">JIT XLA<\/a>?<\/p>\n\n<p>I can flag certain ops for JIT XLA by <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>However, the docs warn that this is meant just for testing. I'd like to be able to do this using the recommended way<\/p>\n\n<pre><\/pre>\n\n<p>but I can't figure out a way of setting the  from outside the .<\/p>\n\n<p>This pseudo-code may clarify the problem better:<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard4\/tf.contrib.learn.RunConfig.md\" rel=\"nofollow noreferrer\"><\/a> seemed like a good candidate but it doesn't have something for the session (which I guess makes sense, why would the RunConfig have a pointer to the existing session?)<\/p>\n\n<p>I could be overthinking this and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_default_session\" rel=\"nofollow noreferrer\"><\/a> could be all I need but can I modify the config of session after it's created?<\/p>\n","answers":[{"AnswerId":"53472482","CreationDate":"2018-11-25T22:06:47.437","ParentId":null,"OwnerUserId":"4042398","Title":null,"Body":"<p>Now it's possible via <code>session_config<\/code> param of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/RunConfig#session_config\" rel=\"nofollow noreferrer\">RunConfig<\/a>.\nHere's an example:<\/p>\n\n<pre><code>session_config = tf.ConfigProto()                                               \noptimizer_options = session_config.graph_options.optimizer_options               \nif xla:                                                     \n    optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1           \n\nrun_config = tf.estimator.RunConfig(                                        \n    session_config=session_config,  # use this session config\n    log_device_placement=True                                           \n)                                                                           \n\nbase_classifier = tf.estimator.Estimator(                                   \n    model_fn=model_fn,                                                      \n    model_dir=model_dir,                                                    \n    config=run_config,                                                      \n    params=model_params                                                                \n)                          \n<\/code><\/pre>\n"},{"AnswerId":"42401238","CreationDate":"2017-02-22T20:06:52.927","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Looks like not yet. The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/f821ce046df71f5784ed4ce7fb6b87f77d96b031\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/estimator.py#L962\" rel=\"nofollow noreferrer\">MonitoredSession is being instantiated directly<\/a> and the options are passed locally. \nYour only option would be to subclass and override the train method with a monitored session that uses the XLA configuration.<\/p>\n"}]}
{"QuestionId":42391598,"AnswerCount":1,"Tags":"<vector><tensorflow><expand>","CreationDate":"2017-02-22T12:31:56.320","AcceptedAnswerId":42392811.0,"OwnerUserId":6716760.0,"Title":"Expand Vector in Tensorflow and space elements with zeros","Body":"<p>I would like to space vector elements from each others and fill it with zeros:<\/p>\n\n<pre><\/pre>\n\n<p>Space elements of a with two zeros:<\/p>\n\n<pre><\/pre>\n\n<p>The number of zeros that I space the elements with should be flexible.<\/p>\n\n<p>What's the best way to do this on the Graph in Tensorflow?<\/p>\n","answers":[{"AnswerId":"42392811","CreationDate":"2017-02-22T13:27:04.627","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>Could you do it as explained <a href=\"https:\/\/stackoverflow.com\/questions\/37061808\/how-does-tensorflow-indexing-work\">in this post<\/a> (second example of the accepted answer)?<\/p>\n\n<p>Basically I would first create <code>b<\/code> as a a vector of zeros, then compute the  indices which point into <code>b<\/code> for all values in <code>a<\/code> and then use the code from the linked post to assign the values of <code>a<\/code> to these indices.<\/p>\n\n<p>Something like this maybe:<\/p>\n\n<pre><code>a = tf.constant(np.array([1, 3, 5, 7]))\ndim = a.get_shape()[0].value\nn = 2\nb = tf.fill(dims=[dim*(n+1)], value=0.0)\nnew_indices = np.array([i + i*n for i in range(0, dim)])\n# now call referenced code with new_indices to update `b`\n<\/code><\/pre>\n\n<p>I'm not sure that this is the best way per se, but it would get the job done.<\/p>\n"}]}
{"QuestionId":42392441,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras>","CreationDate":"2017-02-22T13:10:48.073","AcceptedAnswerId":42401875.0,"OwnerUserId":7596902.0,"Title":"How to record val_loss and loss pre batch in keras","Body":"<p>I'm using the callback function in keras to record the  and  per epoch, But I would like to a do the same but per batch. I found a callback function called , but I not sure how to use it.<\/p>\n","answers":[{"AnswerId":"42401875","CreationDate":"2017-02-22T20:43:00.420","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Following and modyfing an example from <a href=\"https:\/\/keras.io\/callbacks\/\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre><code>class LossHistory(keras.callbacks.Callback):\n    def on_train_begin(self, logs={}):\n        self.losses = []\n        self.val_losses = []\n\n\ndef on_batch_end(self, batch, logs={}):\n    self.losses.append(logs.get('loss'))\n    self.val_losses.append(logs.get('val_loss'))\n\nmodel = Sequential()\nmodel.add(Dense(10, input_dim=784, init='uniform'))\nmodel.add(Activation('softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='rmsprop')\n\nhistory = LossHistory()\nmodel.fit(X_train, Y_train, batch_size=128, nb_epoch=20, verbose=0, validation_split=0.1,\n          callbacks=[history])\n\nprint history.losses\n# outputs\n'''\n[0.66047596406559383, 0.3547245744908703, ..., 0.25953155204159617, 0.25901699725311789]\n'''\nprint history.val_losses\n<\/code><\/pre>\n"}]}
{"QuestionId":42392701,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-02-22T13:22:11.530","AcceptedAnswerId":null,"OwnerUserId":5611409.0,"Title":"DataType string for attr 'TI' not in list of allowed values: uint8, int32, int64","Body":"<p>I have been using a CNN for text classification and have used tensorflow's contrib learn.<\/p>\n\n<p>However, when I try to execute the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I am running in the following error:<\/p>\n\n<blockquote>\n  <p>ERROR:DataType string for attr 'TI' not in list of allowed values:\n  uint8, int32, int64 on line 'classifier.fit'<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42401418","CreationDate":"2017-02-22T20:17:13.673","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>You need to convert your inputs <code>y_train<\/code> to the given type. \n<code>print(type(y_train))<\/code> most likely is a float instead of an integer. <\/p>\n"}]}
{"QuestionId":42392798,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><keras><conv-neural-network>","CreationDate":"2017-02-22T13:26:27.633","AcceptedAnswerId":null,"OwnerUserId":5653822.0,"Title":"Error during playing with keras example pretrained_word_embeddings","Body":"<p>I'am trying to modify the keras example pretrained_word_embeddings that is available <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/pretrained_word_embeddings.py\" rel=\"nofollow noreferrer\">here<\/a> and I faced the following issue : if I decreased  varibae for example to  value I would get the following error:<\/p>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"C:\\Program\n  Files\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\common_shapes.py\",\n  line 670, in _call_cpp_shape_fn_impl\n      status)   File \"C:\\Program Files\\Anaconda3\\lib\\contextlib.py\", line 66, in <strong>exit<\/strong>\n      next(self.gen)   File \"C:\\Program Files\\Anaconda3\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\",\n  line 469, in raise_exception_on_not_ok_status\n      pywrap_tensorflow.TF_GetCode(status)) tensorflow.python.framework.errors_impl.InvalidArgumentError: Negative\n  dimension size caused by subtracting 5 from 2 for 'Conv2D_2' (op:\n  'Conv2D') with input shapes: [?,2,1,128], [5,1,128,128].<\/p>\n<\/blockquote>\n\n<p>I need to change it in case when I for example need to work with small messages like tweets. I use a Tensorflow backend.<\/p>\n\n<p>Please, help me to clarify \n1) What is the problem with ?\n2) What is the reason that  is in trace instead of  that I'am using in the model.<\/p>\n","answers":[{"AnswerId":"42400099","CreationDate":"2017-02-22T19:02:40.987","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Let's go through network definition and analyse the shapes of layer outputs when <code>MAX_SEQUENCE_LENGTH=95<\/code>:<\/p>\n\n<pre><code>sequence_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')\nembedded_sequences = embedding_layer(sequence_input)\n# Output shape: (95, EMBEDDING_DIM)\nx = Conv1D(128, 5, activation='relu')(embedded_sequences)\n# Output shape: (91, 128) (because of valid border mode) \nx = MaxPooling1D(5)(x)\n# Output shape: (18, 128)\nx = Conv1D(128, 5, activation='relu')(x)\n# Output shape: (14, 128)\nx = MaxPooling1D(5)(x)\n# Output shape: (2, 128)\nx = Conv1D(128, 5, activation='relu')(x)\n# Output shape: (2 - 4??, 128) - PROBLEM!!\nx = MaxPooling1D(35)(x) # In the easiest way - change 35 to 2.\nx = Flatten()(x)\nx = Dense(128, activation='relu')(x)\npreds = Dense(100, activation='softmax')(x)\n<\/code><\/pre>\n\n<p>As you may see the problem lies in the last <code>Conv1D<\/code> layer where there is not enough dimension to apply convolution with a <code>valid<\/code> border mode. There are many ways how you could solve that. The easiest one is to crop the last <code>Conv-MaxPool<\/code> duo and change the network definition to:<\/p>\n\n<pre><code>sequence_input = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='int32')\nembedded_sequences = embedding_layer(sequence_input)\n# Output shape: (95, EMBEDDING_DIM)\nx = Conv1D(128, 5, activation='relu')(embedded_sequences)\n# Output shape: (91, 128) (because of valid border mode) \nx = MaxPooling1D(5)(x)\n# Output shape: (18, 128)\nx = Conv1D(128, 5, activation='relu')(x)\n# Output shape: (14, 128)\nx = MaxPooling1D(5)(x)\n# Output shape: (2, 128)\nx = Flatten()(x) # Here - everything is ok.\nx = Dense(128, activation='relu')(x)\npreds = Dense(100, activation='softmax')(x)\n<\/code><\/pre>\n\n<p>Of course - there are more way to do this (like playing with pooling size, etc.). <\/p>\n\n<p>The reason behind using <code>Conv2D<\/code> comes from the fact that in case of <code>TensorFlow<\/code> backend a <code>Conv1D<\/code> are implemented by a two dimensional convolutions with one dimension squeezed to have a size of <code>1<\/code>.<\/p>\n"}]}
{"QuestionId":42393407,"AnswerCount":1,"Tags":"<python><python-2.7><neural-network><deep-learning><caffe>","CreationDate":"2017-02-22T13:52:37.513","AcceptedAnswerId":null,"OwnerUserId":3919542.0,"Title":"How to load different version of caffe in a single python file","Body":"<p>I am running two neural networks for my project, both of which use different versions of caffe. I am trying to add the path of the caffe module using the sys module. The first neural network runs without any difficulty, however the second neural network has problem regarding the path to its caffe module. <\/p>\n\n<p>I have tried removing the caffe path location that i added earlier and then tried adding the caffe path for the second neural network. Then also it is giving me error.<\/p>\n\n<p>The only solution i found is to run my second neural network in a subprocess. However it is not efficient as i have to save all the data in a file and then read and do the computation in the main process later.<\/p>\n\n<p>I was suggested by someone to use environment variables but i don't know how to go along this concept.<\/p>\n\n<p>Any help as how can i load multiple caffe version would be great. <\/p>\n\n<p><strong>EDIT<\/strong><\/p>\n\n<p>I have tried adding different versions of the modules using the logic given by <strong>Shai<\/strong> ex: Adding different versions of numpy. However the same logic doesn't work in the case of caffe.<\/p>\n\n<pre><\/pre>\n\n<p><strong>error displayed<\/strong> : WARNING: Logging before InitGoogleLogging() is written to STDERR\nF0223 08:21:57.810040 17753 layer_factory.hpp:69] Check failed: registry.count(type) == 0 (1 vs. 0) Layer type Convolution already registered.\n<strong>* Check failure stack trace: *<\/strong>\nAborted (core dumped)<\/p>\n\n<p><strong>Another approach tried and failed:<\/strong><\/p>\n\n<p>I tried to use <strong><em>threading<\/em><\/strong>, where i tried to load different caffe version in each independent thread. Now i don't get any error However only one version of caffe is getting loaded. Actually the  variables is globally shared between all the threads and thus when the second  statement is executed, python just sees that the modules are already present and does not load the new caffe modules. <\/p>\n\n<p>Are there any ways using which we can change the behaviour of the threads?<\/p>\n","answers":[{"AnswerId":"42408731","CreationDate":"2017-02-23T06:40:40.323","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>I'm not exactly an expert in python, but I'll give it a try...<\/p>\n\n<p>I assume you have two folders <code>\/path\/to\/caffe-a<\/code> and <code>\/path\/to\/caffe-b<\/code> each of them is a clone of a different branch of caffe. Thus you have<\/p>\n\n<pre><code>\/path\/to\/caffe-a\/python\/caffe\/_caffe.so\n\/path\/to\/caffe-b\/python\/caffe\/_caffe.so\n<\/code><\/pre>\n\n<p>And so on...<\/p>\n\n<p>In order to distinguish between the two versions I suggest you create soft links:<\/p>\n\n<pre><code>~$ ln -s \/path\/to\/caffe-a\/python\/caffe \/path\/to\/caffe-a\/python\/caffe_a\n~$ ln -s \/path\/to\/caffe-b\/python\/caffe \/path\/to\/caffe-b\/python\/caffe_b\n<\/code><\/pre>\n\n<p>In python, add both versions of caffe to path<\/p>\n\n<pre><code>import sys\nsys.path.insert(0, '\/path\/to\/caffe-a\/python')\nsys.path.insert(0, '\/path\/to\/caffe-b\/python')\n<\/code><\/pre>\n\n<p>Now import them with different names<\/p>\n\n<pre><code>import caffe_a as caffe_a, caffe_b as caffe_b\n<\/code><\/pre>\n\n<p>Build a net using <code>'a'<\/code> version<\/p>\n\n<pre><code>net_a = caffe_a.Net('\/path\/to\/deploy.prototxt', caffe.TEST)\nnet_a.forward()\n<\/code><\/pre>\n\n<p>Build a net using <code>'b'<\/code> version:<\/p>\n\n<pre><code>net_b = caffe_b.Net('\/path\/to\/other\/deploy.prototxt', caffe.TEST)\nnet_b.forward()\n<\/code><\/pre>\n\n<p>Good luck...<\/p>\n"}]}
{"QuestionId":42393470,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><mnist>","CreationDate":"2017-02-22T13:55:51.830","AcceptedAnswerId":null,"OwnerUserId":7177838.0,"Title":"Tensorflow: Reading MNIST Data in Python 3.5","Body":"<p>im just getting started with ML and Tensorflow. After having (i think) successfully installed tensorflow on my ubuntu system. I wanted to implement a CNN with tensorflow for handwritten didgits. If have read through a couple of questions in stackoverflow and fooled around 2 days without any solution.<\/p>\n\n<pre><\/pre>\n\n<p>What i get out is:<\/p>\n\n<pre><\/pre>\n\n<p>Im almost certain that the 4 gz files are at the wrong place, but I put them almost everywhere. So according to this code where exactly do I have to put the data? Or is the code wrong? What do I miss?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42394506","CreationDate":"2017-02-22T14:37:58.200","ParentId":null,"OwnerUserId":"3271168","Title":null,"Body":"<p>The problem from a stacktrace is not in with a path to mnist dataset. You've installed <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#determine_which_tensorflow_to_install\" rel=\"nofollow noreferrer\">tensorflow with GPU<\/a> support, but it cannot link to required cudnn library. If you really want to use tensorflow with gpu, make sure you have all <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#nvidia_requirements_to_run_tensorflow_with_gpu_support\" rel=\"nofollow noreferrer\">GPU requirements<\/a>. Otherwise you just need to reinstall <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#installing_with_anaconda\" rel=\"nofollow noreferrer\">tensorflow with CPU<\/a> (assuming you use anaconda).<\/p>\n"}]}
{"QuestionId":42393660,"AnswerCount":1,"Tags":"<tensorflow><neural-network><activation-function>","CreationDate":"2017-02-22T14:02:26.363","AcceptedAnswerId":null,"OwnerUserId":3552975.0,"Title":"What activation function is used in the nce_loss?","Body":"<p>I am stuck with the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py#L164\" rel=\"nofollow noreferrer\">nce_loss<\/a> activation function in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"nofollow noreferrer\">word2vec<\/a> model. I want to figure out what activation function it uses among all these listed <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/nn#nce_loss\" rel=\"nofollow noreferrer\">here<\/a>: <\/p>\n\n<blockquote>\n  <p>These include smooth nonlinearities (sigmoid, tanh, elu, softplus,\n  and softsign), continuous but not everywhere differentiable functions\n  (relu, relu6, crelu and relu_x), and random regularization (dropout).<\/p>\n<\/blockquote>\n\n<p>I have searched for it in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/131c3a67a7b8d27fd918e0bc5bddb3cb086de57e\/tensorflow\/python\/ops\/nn_impl.py#L872\" rel=\"nofollow noreferrer\">this function<\/a> and somewhere else but failed to get any ideas. \nI suppose it is the relu* series. Any hints please?<\/p>\n","answers":[{"AnswerId":"42401539","CreationDate":"2017-02-22T20:23:45.910","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>None of those. It uses <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/131c3a67a7b8d27fd918e0bc5bddb3cb086de57e\/tensorflow\/python\/ops\/nn_impl.py#L1117\" rel=\"nofollow noreferrer\">CrossEntropy<\/a>.<\/p>\n"}]}
{"QuestionId":42393760,"AnswerCount":0,"Tags":"<python><windows><tensorflow><anaconda>","CreationDate":"2017-02-22T14:06:23.853","AcceptedAnswerId":null,"OwnerUserId":7605018.0,"Title":"If I change the folder, why is the \"ModuleNotFouldError\" happened?","Body":"<p>My environment is Windows10, and I installed python3.6, tensorflow, and anaconda.\nThere is no error in the installation.<\/p>\n\n<p>If I run the \"command prompt\" (not use anaconda), I can success to run tensorflow.\nHowever, if I use anaconda, a problem is coming.<\/p>\n\n<p>1st: If I run anaconda prompt and import tensorflow by python, no problem.\n2nd: If I change the folder and import tensorflow, the error message \"ModuleNotFouldError\" is coming.<\/p>\n\n<p>Why I change the path, I cannot import a module?<\/p>\n\n<p>History is below.<\/p>\n\n<hr>\n\n<pre>\n(c:\\Users\\**Username**\\Anaconda3) c:\\Users\\**Username**\\>python\nPython 3.6.0 |Anaconda 4.3.0 (64-bit)| (default, Dec 23 2016, 11:57:41) [MSC v.1900 64 bit (AMD64)] on win32\nType \"help\", \"copyright\", \"credits\", or \"license\" for more information\n>>>import tensorflow\n>>>exit()\n\n(c:\\Users\\**Username**\\Anaconda3) c:\\Users\\**Username**\\>cd workspace\n(c:\\Users\\**Username**\\Anaconda3) c:\\Users\\**Username**\\workspace>python\nPython 3.6.0 |Anaconda 4.3.0 (64-bit)| (default, Dec 23 2016, 11:57:41) [MSC 1900 64 bit (AMD64)] on win32\nType \"help\", \"copyright\", \"credits\", or \"license\" for more information\n>>>import tensorflow\nTrackback (most recent call last):\nFile \"\", line 1, in \nModuleNotFoundError: No module named 'tensorflow'\n<\/pre>\n","answers":[]}
{"QuestionId":42394046,"AnswerCount":1,"Tags":"<tensorflow><nlp><chatbot>","CreationDate":"2017-02-22T14:18:12.407","AcceptedAnswerId":null,"OwnerUserId":7497449.0,"Title":"Handle context in a chatbot","Body":"<p>I have made a chatbot, using Tensorflow's <a href=\"https:\/\/www.tensorflow.org\/tutorials\/seq2seq\" rel=\"nofollow noreferrer\">translation model<\/a> (with some modifications), by feeding it with message-response pairs from the Ubuntu Dialogue Corpus. <strong>I was wondering if anyone has any ideas about how I can handle context in a conversation?<\/strong> I.e. I don't want the chatbot to forget what I have previously said, after I enter a new sentence.<\/p>\n\n<p>I can only think about one strategy, which is to handle the context in the pre-processing. Lets say that I have this conversation:<\/p>\n\n<pre>M1: Hi, how are you?\nR1: Hey, good! I just finished work at the restaurant. How are you?\nM2: Good. How was it?\nR2: Exhausting... \nM3: Many customers?\nR3: Yes, and they didn't tip well either! \n<\/pre>\n\n<p>Then I could put them in pairs like this: (M1-R1), (R1M2-R2), (R2M3-R3) etc... Another option would be to save the context from M1 in each pair e.g. (M1-R1), (M1R1M2-R2), (M1R1M2R2M3-R3) - but then the length of the training sentences will increase (a lot) - leading to more memory allocating during training and I would probably need to decrease my network (fewer neurons in each layer). <\/p>\n\n<p>They did something similar in <a href=\"https:\/\/arxiv.org\/abs\/1506.06714\" rel=\"nofollow noreferrer\">this paper<\/a>, but I don't understand how their model is built and how it will handle this. <\/p>\n","answers":[{"AnswerId":"42896644","CreationDate":"2017-03-20T05:51:18.290","ParentId":null,"OwnerUserId":"1993393","Title":null,"Body":"<p><a href=\"https:\/\/www.npmjs.com\/package\/bot-context\" rel=\"nofollow noreferrer\">Bot-context<\/a> is a cool library which provides an easy way to maintain context. The related <a href=\"https:\/\/medium.com\/@ashishshubham\/maintaining-context-in-chatbots-2016b6a5b7c6#.4b10epgv6\" rel=\"nofollow noreferrer\">blog post<\/a>.<\/p>\n"}]}
{"QuestionId":42394554,"AnswerCount":0,"Tags":"<python><macos><tensorflow><spyder>","CreationDate":"2017-02-22T14:40:25.253","AcceptedAnswerId":null,"OwnerUserId":3659855.0,"Title":"Tensorflow 1.0 GPU installation issue on OSX","Body":"<p>I have tried to upgrade tensorflow 0.12 (cpu) to 1.0 (gpu) on a MacBook. I already had installed tensorflow (cpu) but I think I did that on an anaconda environment (did't create a new one, instead I installed it on the existing root environment). <\/p>\n\n<p>To install the new version I followed these <a href=\"https:\/\/www.tensorflow.org\/install\/install_mac\" rel=\"nofollow noreferrer\">steps<\/a>:<\/p>\n\n<ol>\n<li>sudo pip uninstall tensorflow<\/li>\n<li>pip3 install tensorflow-gpu (succesful, plus I have already installed the latest CUDA and cuDNN)<\/li>\n<\/ol>\n\n<p>When I try to run tensorflow on spyder (iPython) or via the terminal on python initially it looks that the library is loaded but without 'extracting' the CUDA and cuDNN drivers (when you get a list with drivers or components after import tensorflow as tf). And in general I receive weird messages when I try to execute the Hello TensorFlow, <\/p>\n\n<p>I executed: <\/p>\n\n<p>I received:<\/p>\n\n<pre><\/pre>\n\n<p>So, version 1.0 GPU is installed but still I couldn't get tensorflow working. And if I had the CPU version 0.12 uninstalled where did python (and spyder) get their package from?\nI looked around and I found out that I can check the location of the imported packages by using on the terminal the command python -v:<\/p>\n\n<p>According to python when I import tensorflow I receive this message:<\/p>\n\n<pre><\/pre>\n\n<p>That's a different location from what the pip3 show command shows. Actually, I followed the path to that folder and is empty. So, I deleted the folder from anaconda and now when I execute import tensorflow on python-v I get:<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone know how can I fix that? So spyder and python can see the location of tensorflow-gpu?<\/p>\n","answers":[]}
{"QuestionId":42394585,"AnswerCount":2,"Tags":"<tensorflow><tfrecord>","CreationDate":"2017-02-22T14:41:48.763","AcceptedAnswerId":42402484.0,"OwnerUserId":1501285.0,"Title":"How to inspect a Tensorflow .tfrecord file?","Body":"<p>I have a  but I don't know how it is structured. How can I inspect the schema to understand what the  file contains?<\/p>\n\n<p>All Stackoverflow answers or documentation seem to assume I know the structure of the file.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42402484","CreationDate":"2017-02-22T21:20:02.103","ParentId":null,"OwnerUserId":"1501285","Title":null,"Body":"<p>Found it!<\/p>\n\n<pre><code>import tensorflow as tf\n\nfor example in tf.python_io.tf_record_iterator(\"data\/foobar.tfrecord\"):\n    print(tf.train.Example.FromString(example))\n<\/code><\/pre>\n\n<p>You can also add:<\/p>\n\n<pre><code>from google.protobuf.json_format import MessageToJson\n...\njsonMessage = MessageToJson(tf.train.Example.FromString(example))\n<\/code><\/pre>\n"},{"AnswerId":"42398972","CreationDate":"2017-02-22T17:58:46.600","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Use TensorFlow <code>tf.TFRecordReader<\/code> with the <code>tf.parse_single_example<\/code> decoder as specified in <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/reading_data<\/a><\/p>\n\n<p>PS, tfrecord contains 'Example' records defined in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/example\/example.proto<\/a><\/p>\n\n<p>Once you extract the record into a string, parsing it is something like this<\/p>\n\n<pre><code>a=tf.train.Example()\nresult = a.ParseFromString(binary_string_with_example_record)\n<\/code><\/pre>\n\n<p>However, I'm not sure where's the raw support for extracting individual records from a file, you can track it down in <code>TFRecordReader<\/code><\/p>\n"}]}
{"QuestionId":42394845,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T14:52:39.870","AcceptedAnswerId":42395062.0,"OwnerUserId":7601767.0,"Title":"Keep getting a SyntaxError message in Tensorflow","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/YeRAP.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/YeRAP.png\" alt=\"My code\"><\/a><\/p>\n\n<p>What is the problem in this code?<\/p>\n\n<p>I keep getting a  on line 33, right after .<\/p>\n","answers":[{"AnswerId":"42395062","CreationDate":"2017-02-22T15:01:59.047","ParentId":null,"OwnerUserId":"2058281","Title":null,"Body":"<p>That print statement will give you an error in Python 3.X because it is no longer a statement, it is a function call.<\/p>\n\n<p><code>print('{0} {1} {2} {3}'.format(step, sess.run(cost), sess.run(W), sess.run(b)))<\/code><\/p>\n\n<p>See the python documentation or <a href=\"http:\/\/sebastianraschka.com\/Articles\/2014_python_2_3_key_diff.html#the-print-function\" rel=\"nofollow noreferrer\">here<\/a> for more information.<\/p>\n\n<p><strong>EDIT<\/strong>: As per late comments above, try changing <code>xrange()<\/code> to <code>range()<\/code>, since <code>xrange<\/code> does not exist in Python 3.<\/p>\n"}]}
{"QuestionId":42395109,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-22T15:04:15.597","AcceptedAnswerId":null,"OwnerUserId":7587072.0,"Title":"How to construct a new distributions object from a tensor in tensorflow","Body":"<p>I want compute the KL divergence between Normal distribution and tensor output by CNN.<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/contrib.distributions\/kullback_leibler_divergence\" rel=\"nofollow noreferrer\">KL function<\/a> <\/p>\n\n<p>the argu must be instance of distributions.Distribution.\nSo, I don't know how to constucting an instance of Distribution from my tensor which is output of CNN.<\/p>\n","answers":[]}
{"QuestionId":42395827,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-02-22T15:34:26.177","AcceptedAnswerId":null,"OwnerUserId":772466.0,"Title":"How to create a custom connected neural network using tensorflow?","Body":"<p>I want to create a network that has specific fixed connections between layers.\nFor example,\n<a href=\"https:\/\/i.stack.imgur.com\/tFGBJ.png\" rel=\"nofollow noreferrer\">Sparsely connected neural network<\/a><\/p>\n\n<p>I tried looking into functions in Tensorflow, but I only found dense networks with regularizers, which doesn't function as I want.<\/p>\n\n<p>If it's not possible in tensorflow, then please suggest some other library that can be used. Thanks!<\/p>\n","answers":[{"AnswerId":"42403523","CreationDate":"2017-02-22T22:30:43.467","ParentId":null,"OwnerUserId":"2595935","Title":null,"Body":"<p>You can always find a workaround. Let's say a layer does <code>y = xW<\/code> (<code>Wx<\/code> is also correct) but you want some of the entries in <code>W<\/code> always be zeros. You can do it column-wise:<\/p>\n\n<p>For column <code>i<\/code> (or element <code>i<\/code> since <code>y<\/code> is a vector) of the output, <code>y_i = x * D_i * W_i<\/code>. The matrix <code>D_i<\/code> is a constant diagonal matrix (<code>tf.constant<\/code>, <code>tf.diag<\/code>) that controls what element would be zeros.<\/p>\n\n<p>Then you can use <code>tf.concat<\/code> to combine all <code>y_i<\/code> to matrix <code>Y<\/code>.<\/p>\n\n<p>You can abstract this into a function whose signature may look like <code>def sparse_layer(input_layer, gates_matrix, activation_f, ...)<\/code> which returns the output layer.<\/p>\n"}]}
{"QuestionId":42396286,"AnswerCount":2,"Tags":"<c++><segmentation-fault><tensorflow><include><eigen3>","CreationDate":"2017-02-22T15:53:49.553","AcceptedAnswerId":null,"OwnerUserId":2312653.0,"Title":"Segmentation fault when using tensorflow with another library, both linked to eigen3","Body":"<p>I am currently trying to use Tensorflow's shared library with another stereo vision library inside my program but I get a segmentation fault when I use both at the same time.<\/p>\n\n<p>I created a shared library for tensorflow, \"libtensorflow_cc.so\", so that I can use TensorFlow in other programs using CMake. I can compile a simple tensorflow program using it and run it: the neural network is working fine, I reproduced results I got with Python. TensorFlow (r0.12) is compiled with Protobuf (3.1.0) and Eigen (v3). <\/p>\n\n<p>I can compile my stereo vision program with the dedicated stereo-vision shared library and run it without TensorFlow.<\/p>\n\n<p>However, when I include Tenserflow like this:<\/p>\n\n<pre><\/pre>\n\n<p>in the previous stereo-vision program, without using anything else from TensorFlow, I can recompile but I get a segmentation error at run time.<\/p>\n\n<p>The stereo-vision dedicated shared library includes eigen3. TensorFlow has its own version of eigen3 that uses some functions of the original eigen3.<\/p>\n\n<p>Therefore, I think that the definitions included by TensorFlow replace the original definitions used when including eigen3. When the dedicated stereo-vision library uses the eigen3 definitions from TensorFlow, it crashes. I ran Valgrind on the program to see this:<\/p>\n\n<pre><\/pre>\n\n<p>The function  in the stereo-vision library () is using definitions from the eigen of TensorFlow () while it should use definitions from the eigen installed on the computer ().<\/p>\n\n<p>How can I fix this ? Thank you very much in advance for your help<\/p>\n","answers":[{"AnswerId":"42398857","CreationDate":"2017-02-22T17:52:59.030","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>It sounds like your <code>cmake<\/code> procedure is exporting Eigen symbols, whereas it shouldn't.<\/p>\n\n<p>Building with bazel makes sure those symbols are not exported using <code>exported_symbols_list<\/code> directive. I'm not sure what the equivalent command is in cmake. See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7380\" rel=\"nofollow noreferrer\">Limit the exported symbols from _pywrap_tensorflow.so<\/a> for more info<\/p>\n"},{"AnswerId":"44131405","CreationDate":"2017-05-23T09:50:25.233","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>Tensorflow uses its own copy of eigen3 which may be different than the one you need in your project.<\/p>\n"}]}
{"QuestionId":42396656,"AnswerCount":1,"Tags":"<parallel-processing><tensorflow><neural-network><theano><keras>","CreationDate":"2017-02-22T16:09:55.997","AcceptedAnswerId":42396841.0,"OwnerUserId":4338302.0,"Title":"Training neural networks in parallel, theoretical framework","Body":"<p>I'm working with neural networks and currently exploring other ways than my cpu to accelerate the training process. Using Keras with Tensorflow on a GPU machine in AWS I was able to really speed up the process. <\/p>\n\n<p><strong>How does this parallelization work in theory?<\/strong> <em>(I am not considering cluster parallelization)<\/em><\/p>\n\n<p>I struggled to find a reference so I ask you for an insight into how it works. <\/p>\n\n<p>Thank you in advance<\/p>\n","answers":[{"AnswerId":"42396841","CreationDate":"2017-02-22T16:18:15.050","ParentId":null,"OwnerUserId":"1755715","Title":null,"Body":"<p>While I'm not familiar with the relevant implementations, training a neural network means optimizing that neural network's parameters, e.g. the weights of neural connections.<\/p>\n\n<p>Common optimization methods include <a href=\"https:\/\/en.wikipedia.org\/wiki\/Quasi-Newton_method#Search_for_extrema:_optimization\" rel=\"nofollow noreferrer\">quasi-Newton methods<\/a>, which are rate-limited by matrix math operations like matrix inversion.  GPU's can greatly improve calculation speed here.<\/p>\n\n<p>References:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/developer.nvidia.com\/sites\/default\/files\/akamai\/cuda\/files\/Misc\/mygpu.pdf\" rel=\"nofollow noreferrer\">\"Matrix computations on the GPU\"<\/a>, Nvidia (2013-08), discusses how their GPU's can perform large matrix operations in parallel.<\/p><\/li>\n<li><p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/using_gpu\" rel=\"nofollow noreferrer\">\"Using GPUs\"<\/a>, TensorFlow, talks about how TensorFlow can be configured to use GPU's.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":42396992,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T16:24:48.953","AcceptedAnswerId":42400521.0,"OwnerUserId":395857.0,"Title":"Is there any point in using CNMeM when running TensorFlow?","Body":"<p>The <a href=\"https:\/\/github.com\/NVIDIA\/cnmem\" rel=\"nofollow noreferrer\">CNMeM library<\/a> is a \"simple library to help the Deep Learning frameworks manage CUDA memory.\"<\/p>\n\n<p><a href=\"https:\/\/datascience.stackexchange.com\/q\/13381\/843\">CNMeM has been reported to give some interesting speed improvements<\/a>, and is supported by Theano, Torch, and Caffe. However, TensorFlow  preallocates GPU memory when starting a session, unlike Theano, Torch, and Caffe.<\/p>\n\n<p>Does using CNMeM when running a TensorFlow-based program help (e.g., reduce the running time)?<\/p>\n","answers":[{"AnswerId":"42400521","CreationDate":"2017-02-22T19:26:01.607","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>No. Tensorflow has its own GPU memory management. Indeed it takes upfront the whole GPU memory regardless of the size of your problem.<\/p>\n"}]}
{"QuestionId":42397370,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T16:42:12.573","AcceptedAnswerId":42397517.0,"OwnerUserId":886724.0,"Title":"Distributed Tensorflow save fails no device","Body":"<p>I am following the example <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/how_tos\/distributed\/\" rel=\"nofollow noreferrer\">here<\/a> to create a distributed tensorflow model with a parameter server and n workers. I do not have any GPU, all work is distributed on CPU<\/p>\n\n<p>In the chief worker, I want to save my variables every some steps, but invoking the saver results in the following exception :<\/p>\n\n<pre><\/pre>\n\n<p>I tried :<\/p>\n\n<pre><\/pre>\n\n<p>I am using a supervisor :<\/p>\n\n<pre><\/pre>\n\n<p>and creating my sesion as follows :<\/p>\n\n<pre><\/pre>\n\n<p>but I am still having the exact same error<\/p>\n","answers":[{"AnswerId":"42397517","CreationDate":"2017-02-22T16:48:45.057","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This line in the error message:<\/p>\n\n<pre><code>available devices: \/job:localhost\/replica:0\/task:0\/cpu:0\n<\/code><\/pre>\n\n<p>...suggests that your <code>tf.Session<\/code> is not connected to the <code>tf.train.Server<\/code> you created. In particular, it seems to be a <em>local<\/em> (or \"direct\") session that can only access devices in the local process.<\/p>\n\n<p>To fix this problem, when you create your session, pass <code>server.target<\/code> to the initializer. For example, depending on which API you are using to create the session, you might want to use one of the following:<\/p>\n\n<pre><code># Creating a session explicitly.\nwith tf.Session(server.target) as sess:\n  # ...\n\n# Using a `tf.train.Supervisor` called `sv`.\nwith sv.managed_session(server.target):\n  # ...\n\n# Using a `tf.train.MonitoredTrainingSession`.\nwith tf.train.MonitoredTrainingSession(server.target):\n  # ...\n<\/code><\/pre>\n"}]}
{"QuestionId":42397490,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-22T16:47:36.067","AcceptedAnswerId":42405247.0,"OwnerUserId":6061661.0,"Title":"Update Tensorflow Checkpoint files to 1.0","Body":"<p>I have a model that was trained in Tensorflow r0.12, that created checkpoint files using .  My model was an RNN making use of  and  from .  Since the change to 1.0, this package has moved to  in  according to <a href=\"https:\/\/stackoverflow.com\/a\/41384322\/6061661\">this answer<\/a><\/p>\n\n<p>I ran the  file from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/compatibility\" rel=\"nofollow noreferrer\">here<\/a> to update my files to the new revision.  However, ever since updating, my old checkpoint files do not work.  It seems that some of the variables required by the new  implementation were not present or had a different name.<\/p>\n\n<p>Example errors (there are 132 such errors):<\/p>\n\n<pre><\/pre>\n\n<p>Saving\/Loading worked perfectly until the update.  What can I do to update my old checkpoint files to r1.0?<\/p>\n\n<p>If it matters, I am using python2.7 and the same error occurs when using either CPU-only tensorflow or tensorflow with CUDA.  <\/p>\n","answers":[{"AnswerId":"42405247","CreationDate":"2017-02-23T01:06:17.280","ParentId":null,"OwnerUserId":"1067799","Title":null,"Body":"<p>There is no easy way to do this... one approach would be to use get_variable_to_shape_map()<\/p>\n\n<pre><code>  ckpt_reader = tf.train.NewCheckpointReader(filepath)\n  ckpt_vars = ckpt_reader.get_variable_to_shape_map()\n<\/code><\/pre>\n\n<p>that will  give you a list of variable names to shapes in your saved checkpoint. then... create a dict that maps from old names to new names i.e.<\/p>\n\n<pre><code>old_to_new={}\nold_to_new[old_name] = new_name\n<\/code><\/pre>\n\n<p>then  instantitate a saver and restore just those vars<\/p>\n\n<pre><code>saver = tf.Saver(old_to_new)\nsaver.restore(filepath)\n<\/code><\/pre>\n\n<p>Good luck, hope this helps.<\/p>\n"}]}
{"QuestionId":42397591,"AnswerCount":1,"Tags":"<tensorflow><neural-network><argmax>","CreationDate":"2017-02-22T16:51:46.673","AcceptedAnswerId":42403113.0,"OwnerUserId":2470337.0,"Title":"Tensorflow Argmax: What's the difference between \"axis\" and \"dimension\" parameter?","Body":"<p>The current version of Tensorflow Argmax doesn't specify what's the difference between the \"axis\" and \"dimension\" arguments. Here is the only information given in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard8\/tf.argmax.md\" rel=\"nofollow noreferrer\">official manual<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone clarify? Which one is actually the dimension being reduced?<\/p>\n","answers":[{"AnswerId":"42403113","CreationDate":"2017-02-22T22:01:37.897","ParentId":null,"OwnerUserId":"2595935","Title":null,"Body":"<p>TensorFlow is transitioning to use <code>axis<\/code> rather than <code>dimension<\/code> which is going to be deprecated soon: <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/migration<\/a><\/p>\n"}]}
{"QuestionId":42397670,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-22T16:55:04.840","AcceptedAnswerId":null,"OwnerUserId":7605635.0,"Title":"Out of memory with conv3d (TensorFlow)","Body":"<p>Please consider this very simple minimal example:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I have a batch  of tensors with shape  and a kernel  with shape . By performing convolution with strides of , the shape of the resulting tensor will be . I'm feeding data using placeholders to simulate a real-world use case where data is fed from the outside.<\/p>\n\n<p>When I run the code above on a GTX960 with 2GB of memory, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>How is it possible that TensorFlow cannot allocate such a little amount of memory? Can this be a bug in TensorFlow? The error disappears if I use a smaller , but I am wondering what is going on here, since in principle the whole problem should fit entirely in GPU memory (indeed the problem size is only 640.00MiB).<\/p>\n\n<p>The log says:<\/p>\n\n<pre><\/pre>\n\n<p>I have identified the different chunks as follows:<\/p>\n\n<ul>\n<li>the chunk sized  is the kernel : 5 * 8 * 8 * 4 = 1280, where 4 is the size of a <\/li>\n<li>the chunk sized  (2.00MiB) is the convolution output: 8 * 1 * 256 * 256 * 4 = 2097152, where 8 is the  and 4 is again the size of a <\/li>\n<li>the chunk sized  (640.00MiB) is the input : 8 * 5 * 2048 * 2048 * 4 =  671088640<\/li>\n<li>the chunk sized  (more than 1GiB) is the remainder amount of GPU memory, that TensorFlow allocates by default. As a matter of fact, if I change the problem size (i.e. ) the size of this chunk varies accordingly so that all chunks always sum up to 1.71GiB on my system.<\/li>\n<\/ul>\n\n<p>If I understand things correctly, there should be plenty of memory still available. So I can't figure out why I get the out of memory error. Am I missing something, or can this be a bug? I'd appreciate if anyone could shed some light on this. Thanks!<\/p>\n\n<p>P.S. Here is the output of . There are no processes taking up GPU memory.<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42397950,"AnswerCount":0,"Tags":"<tensorflow><gpu><sparse-matrix>","CreationDate":"2017-02-22T17:07:34.523","AcceptedAnswerId":null,"OwnerUserId":6716760.0,"Title":"Tensorflow: Gradient Calculation with sparse tensors on GPU","Body":"<p>I built up a tensorflow model similar to the GPU Implementation of CIFAR10. I have a basic model that is executed on every GPU while the variables for the network are on the CPU. Everything works fine as long as I don't use sparse tensors as weight matrices in the layers.<\/p>\n\n<p>My sparse weight matrices are constructed with the function <strong>tf.sparse_to_dense()<\/strong> or <strong>tf.diag()<\/strong>. When I run it on the CPU everything works fine, but when I run it on the GPU I get the message that there is <strong>no GPU implementation of these functions<\/strong>.<\/p>\n\n<p>Is there an alternative? How do I construct sparse weight matrices that are located on a CPU but trained on a GPU?<\/p>\n","answers":[]}
{"QuestionId":42398645,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><keras><keras-layer>","CreationDate":"2017-02-22T17:42:00.283","AcceptedAnswerId":42399348.0,"OwnerUserId":5352399.0,"Title":"TimeDistributed vs. TimeDistributedDense Keras","Body":"<p>I have gone through the <a href=\"https:\/\/keras.io\/layers\/wrappers\/\" rel=\"noreferrer\">official documentation<\/a> but still can't understand what actually  does as a layer in Keras model? <\/p>\n\n<p>I couldn't understand the difference between  and ? When will someone use ? Is it only to reduce training data set? Does it have other benefit?<\/p>\n\n<p>Can anyone explain with a precise example that what these two type of layer wrappers does?<\/p>\n","answers":[{"AnswerId":"42399348","CreationDate":"2017-02-22T18:21:05.593","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So - basically the <code>TimeDistributedDense<\/code> was introduced first in early versions of Keras in order to apply a <code>Dense<\/code> layer stepwise to sequences. <code>TimeDistributed<\/code> is a Keras wrapper which makes possible to get any static (non-sequential) layer and apply it in a sequential manner. So if e.g. your layer accepts as an input something of shape <code>(d1, .., dn)<\/code> thanks to <code>TimeDistributed<\/code> wrapper your layer could accept an input with a shape of <code>(sequence_len, d1, ..., dn)<\/code> by applying a layer provided to <code>X[0,:,:,..,:]<\/code>, <code>X[1,:,...,:]<\/code>, <code>...<\/code>, <code>X[len_of_sequence,:,...,:]<\/code>. <\/p>\n\n<p>An example of such usage might be using a e.g. pretrained convolutional layer to a short video clip by applying <code>TimeDistributed(conv_layer)<\/code> where <code>conv_layer<\/code> is applied to each frame of a clip. It produces the sequence of outputs which might be then consumed by next recurrent or <code>TimeDistributed<\/code> layer.<\/p>\n\n<p>It's good to know that usage of <code>TimeDistributedDense<\/code> is depreciated and it's better to use <code>TimeDistributed(Dense)<\/code>.<\/p>\n"},{"AnswerId":"42580885","CreationDate":"2017-03-03T14:00:37.017","ParentId":null,"OwnerUserId":"7605750","Title":null,"Body":"<p><code>TimeDistributedDense<\/code> is the same than <code>TimeDistributed<\/code> with the only difference that <code>TimeDistributed<\/code> can be used with different types of layer and not just Dense layer. <\/p>\n\n<p>Keras documentation says that about <code>TimeDistributed<\/code> : <\/p>\n\n<p>\"Note this is strictly equivalent to using <code>layers.core.TimeDistributedDense<\/code>. However what is different about  <code>TimeDistributed<\/code> is that it can be used with arbitrary layers, not just Dense, for instance with a <code>Convolution2D<\/code> layer\"<\/p>\n"}]}
{"QuestionId":42398672,"AnswerCount":3,"Tags":"<python><tensorflow><pycharm><remote-debugging>","CreationDate":"2017-02-22T17:43:04.717","AcceptedAnswerId":null,"OwnerUserId":3585696.0,"Title":"PyCharm remote interpreter and Tensorflow -> can not import Cudart.so","Body":"<p>I'm using PyCharm for remote debugging, client Win10, server Ubuntu 16. On the Ubuntu machine I have Cuda 8.0 correctly installed, along with python 2.7 and tensorflow 1.0.0. Tensorflow works perfectly from the server (ssh + invoking python, import tensorflow etc..).<\/p>\n\n<p>In PyCharm I set up a remote debugger to my server machine, but now importing tensorflow results in an error:<\/p>\n\n<pre><\/pre>\n\n<p>I tried:<\/p>\n\n<p>adding the correct paths to the sys.path variable<\/p>\n\n<p>adding LD_LIBRARY_PATH to os.environ and\/or to the PyCharm environment variables<\/p>\n\n<p>No success!<\/p>\n\n<p>I recall time ago I was able to do debugging without any problem, but I can't figure out what's wrong. Any suggestion?<\/p>\n\n<p><strong>UPDATE 1<\/strong>\nAfter rebooting, the error changed to this<\/p>\n\n<pre><\/pre>\n\n<p><strong>UPDATE 2 \/ SOLVED<\/strong>\nAs I tried before, the solution was to add on pycharm the LD_LIBRARY_PATH to the environment variables on the configuration panel. I think before it didn't work for a blank space after the name (-> 'LD_LIBRARY_PATH '). Here the solution:\n<a href=\"https:\/\/i.stack.imgur.com\/gBPer.png\" rel=\"noreferrer\">solution<\/a><\/p>\n","answers":[{"AnswerId":"42398798","CreationDate":"2017-02-22T17:50:24.107","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>The error message you get is \"please exit the tensorflow source tree, and relaunch your python interpreter from there.\". <\/p>\n\n<p>Normally you get this error message if you download the tensorflow source, compile it manually etc. install it, and then start python and import tensorflow while still in the source folder. <\/p>\n\n<p>Things you can try to stop this error:\n- Check if the tensorflow source is in the folder you are running your remote application in (did you follow something like this to get it working? <a href=\"http:\/\/www.pinchofintelligence.com\/faster-writing-and-testing-machine-learning-applications\/\" rel=\"nofollow noreferrer\">http:\/\/www.pinchofintelligence.com\/faster-writing-and-testing-machine-learning-applications\/<\/a> )\n- Try to find and remove the tensorflow source folder\n- Restart your pc (this clears the \/tmp folder you are probably running your stuff in)<\/p>\n\n<p>Hopefully one of these things work, let me know!<\/p>\n"},{"AnswerId":"42405029","CreationDate":"2017-02-23T00:41:35.817","ParentId":null,"OwnerUserId":"3585696","Title":null,"Body":"<p>Solution is to add the correct path to the Pycharm environment panel:<\/p>\n\n<pre><code>LD_LIBRARY_PATH='\/usr\/local\/cuda\/lib64'\n<\/code><\/pre>\n\n<p>or the one that suits your configuration.<\/p>\n"},{"AnswerId":"52092053","CreationDate":"2018-08-30T08:04:33.167","ParentId":null,"OwnerUserId":"10293859","Title":null,"Body":"<p><strong>Remote nvidia-docker case<\/strong><\/p>\n\n<p>I spended a couple of hours to solve similar problem. I actually read stackoverflow several times and add all this env variables without any effort. <\/p>\n\n<p>So if you have remote nvidia-docker with oppened ssh then read this thread:\n<a href=\"https:\/\/github.com\/bethgelab\/docker\/issues\/5\" rel=\"nofollow noreferrer\">https:\/\/github.com\/bethgelab\/docker\/issues\/5<\/a><\/p>\n\n<p>nvidia-docker starts incorrectly terminal when you connect with ssh.<\/p>\n\n<p>Pycharm actually usses ssh to connect for remote debugging, so it also can't start tensorflow-gpu<\/p>\n"}]}
{"QuestionId":42399012,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-22T18:01:21.837","AcceptedAnswerId":null,"OwnerUserId":4533188.0,"Title":"Tensorflow: Transforming manually build layers to tf.contrib.layers","Body":"<p>I have these four layers defined:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to transform this code into code based on . So far I got<\/p>\n\n<pre><\/pre>\n\n<p>by reading up on <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/layers\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/layers\/<\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/fully_connected\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/fully_connected<\/a>. I read in <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.layers#Higher_level_ops_for_building_neural_network_layers\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.layers#Higher_level_ops_for_building_neural_network_layers<\/a> that  is an alternative for the linear layer.<\/p>\n\n<p>But my output is more different compared to what I got earlier, then that this could be by chance. What did I do wrong in the configuration of the layers?<\/p>\n","answers":[{"AnswerId":"42402400","CreationDate":"2017-02-22T21:13:36.363","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>One difference between your code and the <code>tf.contrib.layers<\/code> version is that  the default initializers are different:<\/p>\n\n<ul>\n<li>The initializer for the weights defaults to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/xavier_initializer\" rel=\"nofollow noreferrer\"><code>tf.contrib.layers.xavier_initializer()<\/code><\/a>.<\/li>\n<li>The initializer for the biases defaults to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/zeros_initializer\" rel=\"nofollow noreferrer\"><code>tf.zeros_initializer()<\/code><\/a>.<\/li>\n<\/ul>\n\n<p>These are generally considered to be good defaults for a fully connected layer, but you can override them with a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/random_normal_initializer\" rel=\"nofollow noreferrer\"><code>tf.random_normal_initializer<\/code><\/a> as follows:<\/p>\n\n<pre><code>layer_1 = tf.contrib.layers.fully_connected(\n    inputs=input,\n    num_outputs=n_hidden_1,\n    activation_fn=None,\n    weights_initializer=tf.random_normal_initializer(),\n    biases_initializer=tf.random_normal_initializer())\n# ...\n<\/code><\/pre>\n"}]}
{"QuestionId":42399401,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T18:24:04.563","AcceptedAnswerId":null,"OwnerUserId":2118130.0,"Title":"Use of grads_ys parameter in tf.gradients - TensorFlow","Body":"<p>I want to understand the  paramter in . I've seen it used like a multiplyer of the true gradient but its not crear in the definition. Mathematically how would the whole expression look like?<\/p>\n","answers":[{"AnswerId":"42400196","CreationDate":"2017-02-22T19:07:14.243","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Edit: better clarification of notation is <a href=\"https:\/\/medium.com\/@yaroslavvb\/backprop-and-systolic-arrays-24e925d2050\" rel=\"noreferrer\">here<\/a><\/p>\n\n<p><code>ys<\/code> are summed up to make a single scalar <code>y<\/code>, and then <code>tf.gradients<\/code> computes <code>dy\/dx<\/code> where <code>x<\/code> represents variables from <code>xs<\/code><\/p>\n\n<p><code>grad_ys<\/code> represent the \"starting\" backprop value. They are 1 by default, but a different value can be when you want to chain several <code>tf.gradients<\/code> calls together -- you can pass in the output of previous <code>tf.gradients<\/code> call into <code>grad_ys<\/code> to continue the backprop flow.<\/p>\n\n<p>For formal definition, look at the chained expression in Reverse Accumulation here: <a href=\"https:\/\/en.wikipedia.org\/wiki\/Automatic_differentiation#Reverse_accumulation\" rel=\"noreferrer\">https:\/\/en.wikipedia.org\/wiki\/Automatic_differentiation#Reverse_accumulation<\/a><\/p>\n\n<p>The term corresponding to <code>dy\/dw3 * dw3\/dw2<\/code> in TensorFlow is a vector of 1's (think of it as if TensorFlow wraps cost with a dummy identity op). When you specify <code>grad_ys<\/code> this term is replaced with <code>grad_ys<\/code> instead of vector of <code>1<\/code>s<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Ej7E4.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Ej7E4.png\" alt=\"enter image description here\"><\/a><\/p>\n"}]}
{"QuestionId":42399866,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-22T18:50:25.460","AcceptedAnswerId":42417109.0,"OwnerUserId":7531531.0,"Title":"Triplet, Siamese and Softmax in Tensorflow","Body":"<p>I'd like to compare the performance of following types of CNNs for two different large image data sets. The goal is to measure the similarity between two images, which both have not been seen during training. I have access to 2 GPUs and 16 CPU cores.<\/p>\n\n<ol>\n<li>Triplet CNN (Input: Three images, Label: encoded in position)<\/li>\n<li>Siamese CNN (Input: Two images, Label: one binary label)<\/li>\n<li>Softmax CNN for Feature Learning (Input: One image, Label: one integer label)<\/li>\n<\/ol>\n\n<p>For Softmax I can store the data in a binary format (Sequentially store label and image). Then read it with a TensorFlow reader.<\/p>\n\n<p>To use the same method for Triplet and Siamese Networks, I'd have to generate the combinations in advance and store them to disk. That would result in a big overhead in both the time it takes to create the file and in disk space. How can it be done on the fly?<\/p>\n\n<p>Another easy way would be to use feed_dict, but this would be slow. Therefore the problem would be solved if it would be possible to run the same function which I'd use for feed_dict in parallel and convert the result to a TensorFlow tensor as a last step. But as far as I know such a conversion does not exist so one has to read the files with a TensorFlow reader in the first place and do the whole process with TensorFlow methods. Is this correct?<\/p>\n","answers":[{"AnswerId":"42417109","CreationDate":"2017-02-23T13:27:14.243","ParentId":null,"OwnerUserId":"4844184","Title":null,"Body":"<p>Short answer do the pair\/triplet creation online with numpy no need to convert it to a tensor as the feed_dict arguments accepts numpy arrays already. <\/p>\n\n<p>The best would be to use <code>tf.nn.embedding_lookup()<\/code> from already existing batches in combination with <code>itertools<\/code> to create the indices of the pairs but for a na\u00efve non-optimal solution you can look at the gen_batches_siamese.py script in <a href=\"https:\/\/github.com\/jeandut\/tensorflow-models\/tree\/master\/siamese-on-mnist\" rel=\"nofollow noreferrer\">my github repository<\/a>. Where I reimplemented the caffe siamese example.\nObviously it will be less efficient than using tensorflow queues but my advice would be to try this baseline first before going to the pure tensorflow solution.<\/p>\n"}]}
{"QuestionId":42400235,"AnswerCount":0,"Tags":"<tensorflow><bazel>","CreationDate":"2017-02-22T19:09:31.830","AcceptedAnswerId":null,"OwnerUserId":4172141.0,"Title":"I have problems like configuring basel on windows","Body":"<p>Good afternoon I just setup and install basel for windows to be able to put the tensorflow in android studio but I mark error in the gradle by default comes the route of linux that is this<\/p>\n\n<p>Def bazel_location = '\/ usr \/ local \/ bin \/ bazel' but at the moment I install it appears in the cosola the path of basel that is being this The install of bazel was successful.\nSoftware installed to 'C: \\ ProgramData \\ chocolatey \\ lib \\ bazel' on windows<\/p>\n\n<p>Here I have my gradle code:<\/p>\n\n<p>\/\/def bazel_location = '\/usr\/local\/bin\/bazel'\ndef bazel_location = 'C:\/ProgramData\/chocolatey\/lib\/bazel'<\/p>\n\n<p>android {\n    compileSdkVersion 24\n    buildToolsVersion \"24.0.1\"<\/p>\n\n<pre><\/pre>\n\n<p>task buildNative(type: Exec) {<\/p>\n\n<p>workingDir '..\/..\/..'\n    commandLine bazel_location, 'build', '-c', 'opt',  \\\n       'tensorflow\/examples\/android:tensorflow_native_libs',  \\\n        '--crosstool_top=\/\/external:android\/crosstool',  \\\n        '--cpu=' + cpuType,  \\\n        '--host_crosstool_top=@bazel_tools\/\/tools\/cpp:toolchain'<\/p>\n\n<p>}<\/p>\n\n<p>I did not put the complete code of the gradle but here comes the path and the error that marks me\nBut I mark this error when wanting to install the apk from android studio:<\/p>\n\n<p>:buildNative FAILED<\/p>\n\n<p>FAILURE: Build failed with an exception.<\/p>\n\n<ul>\n<li><p>What went wrong:\nExecution failed for task ':buildNative'.<\/p>\n\n<blockquote>\n  <p>A problem occurred starting process 'command 'C:\/ProgramData\/chocolatey\/lib\/bazel''<\/p>\n<\/blockquote><\/li>\n<li><p>Try:\nRun with --stacktrace option to get the stack trace. Run with --info or --debug option to get more log output.<\/p><\/li>\n<\/ul>\n\n<p>BUILD FAILED<\/p>\n\n<p>Total time: 0.951 sec<\/p>\n","answers":[]}
{"QuestionId":42401341,"AnswerCount":2,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2017-02-22T20:13:26.983","AcceptedAnswerId":null,"OwnerUserId":1538856.0,"Title":"Trouble understanding datasets in the Tensorflow RNN Tutorial","Body":"<p>Ok, I admit it - this is a bad question.  You should probably downvote it ;) However, as I really NEED some insight, if you could kindly wait until some kind soul answers it first....<\/p>\n\n<p>I have the RNN tutorial installed and running successfully.  I am a NOOB to this - though I have a decent feel for the code.  Unfortunately, <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/tutorials\/recurrent\/\" rel=\"nofollow noreferrer\">the TensorFlow RNN tutorial<\/a> doesn't define (at least to me) a few critical items needed to understand:<\/p>\n\n<ol>\n<li><p>What is the logical difference between the training and validation sets?  Specifically, in terms of the file contents in 'ptb.train.txt' and 'ptb.valid.txt'<\/p>\n\n<p>For instance, ptb.train.txt contains:<\/p>\n\n<pre><\/pre>\n\n<p>However, neither  nor  appears in the , thus it is unclear what is being trained and what is being validated.<\/p><\/li>\n<li><p>My 'Test Perplexity' is reported as 121.236.  What does this mean?  Again, I would really like to know in terms of the 'ptb.test.txt' file - what is that being measured against?<\/p>\n\n<p>Another typical example:  contains:<\/p>\n\n<pre><\/pre>\n\n<p>Neither this 'sequence' nor a close subset appears in either the  or the  files.  Maybe I am completely off-base, but I thought the intent was to match such sequences.  Is this incorrect?  How does the reported perplexity relate to that?<\/p><\/li>\n<li><p>Finally (again, NOOB here), it takes quite a while to train even this simple example.  Is there a way to save the 'trained' system so that I can run it against different test datasets?<\/p>\n\n<p><strong>Update<\/strong>\nI believe question #3 is <a href=\"https:\/\/stackoverflow.com\/questions\/40786771\/how-to-use-tensorflows-ptb-model-example\">answered in this SO question<\/a><\/p><\/li>\n<\/ol>\n","answers":[{"AnswerId":"42403831","CreationDate":"2017-02-22T22:53:14.943","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<ol>\n<li><p>There are a few elements that you need to learn to understand the procedure. Machine Learning algorithms are being trained on a given dataset. These algorithms could potentially \u201cmemorize\u201d your whole dataset so they will give you the right answer for your dataset all the time. The <em>right answer<\/em> is the expected result of your problem. If the model memorizes all your train data it will not respond very well to new input data that has not seen, and so his predictive power would not be very good. To train a model to have a better predictive power, you split the dataset in <em>train<\/em> and <em>validation<\/em>, where the validation subset is not used for training but is used to control that the model is not memorizing the <em>train<\/em> dataset. <\/p><\/li>\n<li><p>The perplexity is a specific metric to assess how good is the model <\/p><\/li>\n<li><p>Yes you can Consult the documentation: <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variables\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/variables<\/a><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"47840781","CreationDate":"2017-12-15T22:39:30.770","ParentId":null,"OwnerUserId":"6451548","Title":null,"Body":"<ul>\n<li>Training sets are to <strong>optimize<\/strong> the model.  <\/li>\n<li>Validation datasets are to check if the model is <strong>under fitting<\/strong> or <strong>over-fitting<\/strong> the training data.<\/li>\n<li>After you've trained the model and feel like it is preforming well you use the <strong>test dataset<\/strong> to check if <em>you<\/em> tuned the model too finely based on the results from your validation datasets and caused your model to over-fit real world data.<\/li>\n<\/ul>\n\n<p>Also, these questions belong in <a href=\"https:\/\/stats.stackexchange.com\/\">https:\/\/stats.stackexchange.com\/<\/a> :)<\/p>\n"}]}
{"QuestionId":42401420,"AnswerCount":1,"Tags":"<python><theano>","CreationDate":"2017-02-22T20:17:21.220","AcceptedAnswerId":42414385.0,"OwnerUserId":5352399.0,"Title":"How the function dimshuffle works in Theano","Body":"<p>I am having tough time understanding what and how <a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/tensor\/basic.html#theano.tensor._tensor_py_operators.dimshuffle\" rel=\"noreferrer\">dimshuffle()<\/a> works implemented in Theano? I got the following set of examples in the official documentation but couldn't understand their meaning.<\/p>\n\n<p>Can anyone explain what each examples mean in the following?<\/p>\n\n<pre><\/pre>\n\n<p>Please note, I know about <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.10.0\/user\/basics.broadcasting.html\" rel=\"noreferrer\">broadcasting concept in numpy<\/a> python.<\/p>\n","answers":[{"AnswerId":"42414385","CreationDate":"2017-02-23T11:23:46.647","ParentId":null,"OwnerUserId":"1950580","Title":null,"Body":"<p>Without <code>'x'<\/code>, <code>dimshuffle<\/code> is same as <code>transpose<\/code><\/p>\n\n<p>For explanatory purpose, let's fake numpy has a <code>dimshuffle<\/code> function<\/p>\n\n<pre><code>x = np.arange(60).reshape((3,4,5))\nx.dimshuffle(0, 1, 2).shape # gives (3, 4, 5)\nx.dimshuffle(2, 1, 0).shape # gives (5, 4, 3)\n<\/code><\/pre>\n\n<p>Since we have:<\/p>\n\n<pre><code>shp = (3,4,5)\n(shp[2], shp[1], shp[0]) == (5, 4, 3)\n<\/code><\/pre>\n\n<p>The arguments <code>2, 1, 0<\/code> to dimshuffle just means the <strong>permutation<\/strong> to the shape tuple.<\/p>\n\n<p>Whenever there is <code>'x'<\/code> present, it adds 1 sized dimension into the array:<\/p>\n\n<pre><code>x = np.arange(60).reshape((3,4,5))\nx.dimshuffle(2, 1, 0, 'x').shape # (5, 4, 3, 1)\nx.dimshuffle(2, 1, 'x', 0).shape # (5, 4, 1, 3)\nx.dimshuffle(2, 'x', 1, 0).shape # (5, 1, 4, 3)\n<\/code><\/pre>\n\n<p>Whenever the permutation is missing an index (or multiple), these indices are removed from shape tuple, provided that they are 1 (which is broadcastable)<\/p>\n\n<pre><code>x = np.arange(1337).reshape(2,1337,1)\ny = x.dimshuffle(1,0) # this works since shape[2] is 1\ny.shape # (1337, 2)\nz = y.dimshuffle(1) # ERROR\n<\/code><\/pre>\n\n<p>Note theano has no way to determine the shape of symbolic tensor, so <code>dimshuffle<\/code> with dimensional removal must refer to <code>broadcastable<\/code> attribute. (This is different to tensorflow as you can specify shape at compile-time)<\/p>\n\n<pre><code>&gt;&gt;&gt; x = T.vector()\n&gt;&gt;&gt; x.broadcastable\n(False,)\n&gt;&gt;&gt; y = x.dimshuffle('x', 0, 'x')\n&gt;&gt;&gt; y.broadcastable # the new dims are broadcastable because we added via 'x'\n(True, False, True)\n<\/code><\/pre>\n\n<p>With <code>dimshuffle<\/code>, you can save multiple calls to <code>transpose<\/code> and <code>expand_dims<\/code> (note Theano has no <code>expand_dims<\/code>)<\/p>\n"}]}
{"QuestionId":42402876,"AnswerCount":0,"Tags":"<performance><opencv><machine-learning><tensorflow><neural-network>","CreationDate":"2017-02-22T21:44:44.560","AcceptedAnswerId":null,"OwnerUserId":3763831.0,"Title":"Poor performance when using TensorFlow input functions in a loop","Body":"<p>I am currently training a set of around 10,000 images with ten epochs. My question is regarding the following code:<\/p>\n\n<pre><\/pre>\n\n<p>Here, I am feeding images in mini-batches of 100 to the neural network. The code is initially getting a string, file_contents. Then that string is being decoded into a jpg. However, when I use TensorFlow's functions, such as tf.decode_jpeg and tf.reshape(), etc., there is a difference in speed. When using TensorFlow's functions to accomplish the same task of decoding images, it starts out fast in the training process, and then becomes incredibly slow after the first epoch. <\/p>\n\n<p>To put this in perspective, using OpenCV, I could train this entire model with 10 epochs in around 1hr and 30 minutes. However, using TensorFlow's functions, it took over 12 hours to get past the first epoch and midway through the second, I stopped the training process after I saw the progress it was making. <\/p>\n\n<p>I am not sure if this as anything to do with the concept of network slowdown, as seen <a href=\"https:\/\/www.quora.com\/Why-does-neural-network-learning-slow-down-as-the-error-gets-lower-Does-this-need-to-be-the-case\" rel=\"nofollow noreferrer\">here<\/a>. I am simply replacing OpenCV functions with TensorFlow functions to decode an image and read a file. Why is there such as dramatic speed difference between OpenCV and TensorFlow? Why exactly does TensorFlow's functions slow down as the code progresses? Will this have any effect on the accuracy of model?  <\/p>\n\n<p>Note: The only functions I changed were at the top. I didn't use tf.train.batch in either versions. The only thing that was changed was the lines of code from file_contents to data.flatten(), to corresponding tensorflow functions.<\/p>\n","answers":[]}
{"QuestionId":42403501,"AnswerCount":2,"Tags":"<tensorflow><gpu><nvidia><cudnn>","CreationDate":"2017-02-22T22:29:06.460","AcceptedAnswerId":null,"OwnerUserId":1601580.0,"Title":"How does one have TensorFlow not run the script unless the GPU was loaded successfully?","Body":"<p>I have been trying to run some TensorFlow training on some machine with GPUs however, whenever I try to do so I get some type of error that seems to say it wasn't able to use the GPU for some reason (usually memory issue, or cuda issue or cudnn etc). However, since the thing TensorFlow does automatically is to just run in CPU if it can't use the GPU its been hard to tell for me if it was actually able to leverage the GPU or not. Thus, I wanted to have my script just fail\/halt unless the GPU is being used. How do I do that?<\/p>\n\n<hr>\n\n<p>For the sake of an example, currently I have the message:<\/p>\n\n<pre><\/pre>\n\n<p>where it seems to load all the cuda fine but then at the end complains. The complaining lines are:<\/p>\n\n<pre><\/pre>\n\n<p>we could try to debug these specific bug but for the moment it proceeds to train however, I have no idea if its using cpu or gpu. Can we just have it not proceed training if any weird cuda\/cudnn or whatever gpu bug comes up?<\/p>\n","answers":[{"AnswerId":"43016258","CreationDate":"2017-03-25T12:21:32.150","ParentId":null,"OwnerUserId":"212731","Title":null,"Body":"<p>Use <code>with tf.device('\/gpu:0'):<\/code>. This will kill your program if <code>\/gpu:0<\/code> doesnt exist.<\/p>\n\n<p>eg see <a href=\"https:\/\/github.com\/hughperkins\/tensorflow-cl\/blob\/tensorflow-cl\/tensorflow\/stream_executor\/cl\/test\/test_binary_ops.py#L52\" rel=\"nofollow noreferrer\">https:\/\/github.com\/hughperkins\/tensorflow-cl\/blob\/tensorflow-cl\/tensorflow\/stream_executor\/cl\/test\/test_binary_ops.py#L52<\/a><\/p>\n\n<pre><code>with tf.Graph().as_default():\n    with tf.Session(config=tf.ConfigProto(log_device_placement=False)) as sess:\n        with tf.device('\/gpu:0'):\n            tf_a = tf.placeholder(tf_dtype, [None, None], 'a')\n            tf_b = tf.placeholder(tf_dtype, [None, None], 'b')\n            tf_c = tf.__dict__[tf_func](tf_a, tf_b, name=\"c\")\n<\/code><\/pre>\n"},{"AnswerId":"42404103","CreationDate":"2017-02-22T23:15:18.417","ParentId":null,"OwnerUserId":"2595935","Title":null,"Body":"<p>You can list all available devices in tensorflow: <a href=\"https:\/\/stackoverflow.com\/questions\/38559755\/how-to-get-current-available-gpus-in-tensorflow\">How to get current available GPUs in tensorflow?<\/a>. If GPU is not in the list, you can make the program throw exceptions.<\/p>\n"}]}
{"QuestionId":42404040,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><autoencoder>","CreationDate":"2017-02-22T23:10:02.207","AcceptedAnswerId":null,"OwnerUserId":1226676.0,"Title":"sparse autoencoder cost function in tensorflow","Body":"<p>I've been going through a variety of TensorFlow tutorials to try to familiarize myself with how it works; and I've become interested in utilizing autoencoders.<\/p>\n\n<p>I started by using the model autoencoder in Tensorflow's models repository:<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/autoencoder\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/autoencoder<\/a><\/p>\n\n<p>I got it working, and while visualizing the weights, expected to see something like this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/F33rU.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/F33rU.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>however, my autoencoder gives me garbage-looking weights (despite accurately recreating the input image).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/tkOvj.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tkOvj.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Further reading suggests that what I'm missing is that my autoencoder is not sparse, so I need to enforce a sparsity cost to the weights.<\/p>\n\n<p>I've tried to add a sparsity cost to the original code (based off of this example<a href=\"https:\/\/github.com\/georgeiswang\/Sparse-Autoencoder-Tensorflow\/blob\/master\/sparse_autoencoder.py\/06_autoencoder.py\" rel=\"noreferrer\">3<\/a>), but it doesn't seem to change the weights to looking like the model ones.<\/p>\n\n<p>How can I properly change the cost to get features that look like the ones that a typically found in the autoencoded MNIST dataset?  My modified model is here:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"44447362","CreationDate":"2017-06-09T00:06:24.590","ParentId":null,"OwnerUserId":"2464597","Title":null,"Body":"<p>I don't know that this will work for you, but I have seen it promote some sparsity in my own networks. I would recommend modifying your loss to use a combination of softmax cross entropy (or KL divergence if you wish) and l2 regularization loss on the weights. I calculate the l2 loss with:<\/p>\n\n<pre><code>l2 = sum(tf.nn.l2_loss(var) for var in tf.trainable_variables() if not 'biases' in var.name)\n<\/code><\/pre>\n\n<p>This makes me regularize only on the weights, not biases, assuming that you have \"biases\" in the name of your bias tensors (lots of the tf.contrib.rnn library names bias tensors so that this works). The overall cost function I use is then:<\/p>\n\n<pre><code>cost = tf.nn.softmax_or_kl_divergence_or_whatever(labels=labels, logits=logits)\ncost = tf.reduce_mean(cost)\ncost = cost + beta * l2\n<\/code><\/pre>\n\n<p>where <code>beta<\/code> is a hyperparameter of the network that I then vary when exploring my hyperparameter space.<\/p>\n\n<p>Another option, very similar to this, is to use l1 regularization instead. <a href=\"http:\/\/cs231n.github.io\/neural-networks-2\/#reg\" rel=\"nofollow noreferrer\">This is supposed to promote sparsity more than l2 regularization<\/a>. In my own examples I was not explicitly trying to promote sparsity, but saw it as a consequence of l2 regularization, but maybe l1 will give you more luck. You can implement l1 regularization with something like:<\/p>\n\n<pre><code>l1 = sum(tf.reduce_sum(tf.abs(var)) for var in tf.trainable_variables() if not 'biases' in var.name)\n<\/code><\/pre>\n\n<p>followed by the cost definition above, substituting <code>l1<\/code> for <code>l2<\/code>.<\/p>\n"}]}
{"QuestionId":42404564,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-02-22T23:56:00.270","AcceptedAnswerId":42415441.0,"OwnerUserId":775755.0,"Title":"What does tf.train.get_global_step() do in TensorFlow?","Body":"<p>What is the use of the function  in TensorFlow? \nIn machine learning concepts what is it equivalent to?<\/p>\n","answers":[{"AnswerId":"50959195","CreationDate":"2018-06-21T01:40:29.027","ParentId":null,"OwnerUserId":"8037585","Title":null,"Body":"<p><code>tf.train.get_global_step()<\/code> return global step(variable, tensor from variable node or None) through <code>get_collection(tf.GraphKeys.GLOBAL_STEP)<\/code> or <code>get_tensor_by_name('global_step:0')<\/code><\/p>\n\n<p>global step is widely used in learn rate decay(like <code>tf.train.exponential_decay<\/code>, see <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/train#Decaying_the_learning_rate\" rel=\"nofollow noreferrer\">Decaying the learning rate<\/a> for more information). <\/p>\n\n<p>You can pass global step to optimzer apply_gradients or minimize method to  increment by one.<\/p>\n"},{"AnswerId":"42405877","CreationDate":"2017-02-23T02:20:16.527","ParentId":null,"OwnerUserId":"4922660","Title":null,"Body":"<p>while you defined the global step operator, you can get value of it by sess.run(global_step_op)<\/p>\n"},{"AnswerId":"42415441","CreationDate":"2017-02-23T12:12:22.623","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>You could use it to restart training exactly where you left off when the training procedure has been stopped for some reason. Of course you can always restart training without knowing the <code>global_step<\/code> (if you save checkpoints regularly in your code, that is), but unless you somehow keep track of how many iterations you already performed, you will not know how many iterations are left after the restart. Sometimes you really want your model to be trained exactly <code>n<\/code> iterations and not <code>n<\/code> plus <code>unknown amount before crash<\/code>. So in my opinion, this is more of a practicality than a theoretical machine learning concept.<\/p>\n"}]}
{"QuestionId":42404567,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-22T23:56:04.997","AcceptedAnswerId":42443853.0,"OwnerUserId":6620662.0,"Title":"Read batch of images is very slow?","Body":"<p>I have the following code, which is reading a batch of 10 images from files on my local disk.<\/p>\n\n<p>The problem is that the code seems to run very slow. It takes about 5-6 minutes to run to completion. The directory containing the images contains approx. 25.000 images.<\/p>\n\n<p>Is the code correct or have I done something stupid?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42443853","CreationDate":"2017-02-24T16:38:39.387","ParentId":null,"OwnerUserId":"2505209","Title":null,"Body":"<p>Reduce min_after_dequeue to 1000 and try once. Look at the following timelines for different min_after_dequeue values.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/HTCEw.png\" rel=\"nofollow noreferrer\">min_after_dequeue = 2000 => 2.1 sec to finish<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/6M0Pc.png\" rel=\"nofollow noreferrer\">min_after_dequeue = 100  => 0.13 sec to finish<\/a><\/p>\n\n<p>Do following to get timelines<\/p>\n\n<pre><code>from tensorflow.python.client import timeline\n\nrun_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)\nrun_metadata = tf.RunMetadata()\n\nb1 = sess.run(new_batch,options=run_options,run_metadata=run_metadata)\n# Create the Timeline object, and write it to a json\ntl = timeline.Timeline(run_metadata.step_stats)\nctf = tl.generate_chrome_trace_format()\nwith open('timelinestack1.json', 'w') as f:\n    f.write(ctf)\n<\/code><\/pre>\n\n<p>Also, make sure all of your images have same size as you mentioned. Otherwise, use below line before set_shape(). <\/p>\n\n<pre><code>image = tf.image.resize_images(imaged, [224, 224])\n<\/code><\/pre>\n\n<p>I hope I gave a reasonable answer.<\/p>\n"}]}
{"QuestionId":42404697,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-02-23T00:09:55.087","AcceptedAnswerId":42404903.0,"OwnerUserId":5516665.0,"Title":"CNN for behavior classification","Body":"<p>Based on this tutorial : <a href=\"https:\/\/medium.com\/@harvitronix\/continuous-online-video-classification-with-tensorflow-inception-and-a-raspberry-pi-785c8b1e13e1#.n7mlkeevl\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@harvitronix\/continuous-online-video-classification-with-tensorflow-inception-and-a-raspberry-pi-785c8b1e13e1#.n7mlkeevl<\/a><\/p>\n\n<p>They simply constructed a CRNN network by combining CNN and RNN to make a video classifier that differentiate between football game an ad on TV.<\/p>\n\n<p>My problem is classifying or simply decide if this input video contains a cheating behavior (exam cheating) or not , so according to this tutorial , I should train the CNN by set of cheating and non-cheating images , then I would pass a sequence of frames through this network to output a sequence of (cheater \/ non-cheater) flags, and finally train or test the RNN with that output.<\/p>\n\n<p>So is the CNN would performs well in that case , because I would use myself and my team to act as cheaters and non-cheaters to get a large dataset to train the CNN , that means the same person would act as cheater and non-cheater , which might result in confusion!<\/p>\n\n<p>My final question is can I use the approach declared in the tutorial for behavior classification? .. or can the CNN successfully differentiate between cheating and non-cheating frames? <\/p>\n","answers":[{"AnswerId":"42404903","CreationDate":"2017-02-23T00:30:28.507","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>Actually, using the same people for cheater and non-cheater clips will <em>improve<\/em> the training.  The NN will have to learn to differentiate by <em>action<\/em>, rather than the physical characteristics of the cheater.<\/p>\n\n<p>I expect that this will work reasonably well for you.  However, depending on the particular behaviour, you may need to use video sequences, rather than individual frames, to describe certain behaviours.  Sometimes, cheating consists of a sequence of request and response, rather than data contained within a single frame.<\/p>\n\n<p>Does that help?<\/p>\n"}]}
{"QuestionId":42404724,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2017-02-23T00:11:52.680","AcceptedAnswerId":43672308.0,"OwnerUserId":7338784.0,"Title":"contrib.learn.estimator() for tensorflow0.12","Body":"<p>I am using contrib.learn.estimator to predict in tensorflow0.12 environment.<\/p>\n\n<pre><\/pre>\n\n<p>and the LSTM model is <\/p>\n\n<pre><\/pre>\n\n<p>The error reports <\/p>\n\n<pre><\/pre>\n\n<p>if I modify the model line as<\/p>\n\n<pre><\/pre>\n\n<p>It still reports error:<\/p>\n\n<pre><\/pre>\n\n<p>I am confused with the contrib.learn.estimator() for tensorflow 0.12. Could anyone help me correct the code and teach me on Estimate? Thank you anyway.<\/p>\n","answers":[{"AnswerId":"43672308","CreationDate":"2017-04-28T05:02:59.080","ParentId":null,"OwnerUserId":"6215244","Title":null,"Body":"<p>Your model_fn argument shall not be the return of ModelFnOps(). A mamual model_fn name is required:<\/p>\n\n<pre><code>def model_fn(features, targets, mode):\n    '''Implement of model_fn API.\n\n    # Logic to do the following:\n    # 1. Configure the model via TensorFlow operations\n    # 2. Define the loss function for training\/evaluation\n    # 3. Define the training operation\/optimizer\n    # 4. Generate predictions\n    # 5. Return predictions\/loss\/train_op\/eval_metric_ops in ModelFnOps object\n\n    Args:\n        features: Inputs data for the model.\n        targets: Expected outputs of the model namely labels.\n        mode: learn.ModeKeys.TRAIN\/EVAL\/INFER\n\n    Returns:\n        ModelFnOps object.\n    '''\n<\/code><\/pre>\n\n<p>And then when you define an Estimator:<\/p>\n\n<pre><code>model = SKCompat(learn.Estimator(\n    model_fn=model_fn,\n    model_dir='path\/to\/your\/model'\n))\n<\/code><\/pre>\n"}]}
{"QuestionId":42405415,"AnswerCount":0,"Tags":"<tensorflow><keras>","CreationDate":"2017-02-23T01:26:23.580","AcceptedAnswerId":null,"OwnerUserId":3813428.0,"Title":"How to implement tensorflow Estimator with multiple models for GAN?","Body":"<p>I used to work with tensorflow and then found Keras. It's easy to train multiple models in Keras using API like <\/p>\n\n<pre><\/pre>\n\n<p>Now I need to turn back to tensorflow for its high flexibility. The problem is there is no model in tensorflow, and it seems  is something like model in Keras. However,  API only provides a  and a parameter arguments, which seems restricted us using only one . <\/p>\n\n<p>I'm wondering if we can implement  such that we can use some syntax like:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42406421,"AnswerCount":1,"Tags":"<tensorflow><gpu><nvidia><dma>","CreationDate":"2017-02-23T03:24:57.717","AcceptedAnswerId":null,"OwnerUserId":2455061.0,"Title":"What does the print of DMA and tensorflow mean? And is it possible to set them?","Body":"<p><a href=\"https:\/\/i.stack.imgur.com\/K5vRy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/K5vRy.png\" alt=\"tensorflow DMA\"><\/a><\/p>\n\n<p>What does the  or  mean?<\/p>\n\n<p>How to set the  and  by myself?<\/p>\n\n<p>By the way, the machine shutdown when it print (on the error machine)<\/p>\n\n<pre><\/pre>\n\n<p>And it run successfully on two GPU when print (on another machine)<\/p>\n\n<pre><\/pre>\n\n<p>So I wonder if it is the problem.<\/p>\n\n<p>EDIT:<\/p>\n\n<p>I use the program below could make it shutdown.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"58839351","CreationDate":"2019-11-13T14:29:08.107","ParentId":null,"OwnerUserId":"8668595","Title":null,"Body":"<p>It is Device interconnection. That shows how fast the data can be transferred between devices during multi-GPU training.<\/p>\n\n<p>It is not a problem. The \"NY\" configuration depends on your hardware configuration. You cannot set it manually. <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/55365280\/8668595\">This<\/a> post by @McAngus has a full answer.<\/p>\n"}]}
{"QuestionId":42406491,"AnswerCount":1,"Tags":"<numpy><theano><nvidia><keras><tesla>","CreationDate":"2017-02-23T03:31:39.663","AcceptedAnswerId":null,"OwnerUserId":3101193.0,"Title":"How to obtain the module\/object code of a Theano numpy program","Body":"<p>In my University we have a Cluster having Tesla GPUs. However the resource is shared by several departments and the supercomputing department requires users to provide uniquely the module\/code object of the program one needs to run in the cluster. In such a situation, I searched for some information about this. The supercomputer has a queue system (which is usual in supercomputers to be shared). As I understand, the supercomputing department requires one to follow procedures like <a href=\"http:\/\/www.cem.msu.edu\/supercomputer\/pbs_local.html\" rel=\"nofollow noreferrer\">this<\/a>. So, how to obtain the object code of a Keras-Theano model compiled for GPU? Just like the produced by  -->  which is what i need. <\/p>\n\n<p>Any other idea is very appreciated.<\/p>\n","answers":[{"AnswerId":"42413624","CreationDate":"2017-02-23T10:50:38.317","ParentId":null,"OwnerUserId":"1950580","Title":null,"Body":"<p>The easiest solution should be pickling theano function, however this only saves optimized graph but not generated code. I'm not sure if this works for your case.<\/p>\n\n<p>You can use command <code>theano-cache list<\/code> to find the directories for generated Op code, which is typically under <code>\/home\/user\/.theano\/<\/code>. However hand-compiling this into module may be complicated.<\/p>\n\n<p>There's also a <a href=\"https:\/\/github.com\/Theano\/Theano\/pull\/3350\" rel=\"nofollow noreferrer\">PR<\/a> for shared library generation however it's not merged yet.<\/p>\n"}]}
{"QuestionId":42406843,"AnswerCount":1,"Tags":"<tensorflow><tensorboard><nvidia-titan>","CreationDate":"2017-02-23T04:07:24.897","AcceptedAnswerId":null,"OwnerUserId":4093741.0,"Title":"Training TensorFlow model with summary operations is much slower than without summary operations","Body":"<p>I am training an Inception-like model using TensorFlow r1.0 with GPU Nvidia Titan X.<\/p>\n\n<p>I added some summary operations to visualize the training procedure, using code as follows:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>When I run these operations, the time cost of training one epoch is about 400 seconds. But when I turn off these operations, the time cost of training one epoch is just 90 seconds.<\/p>\n\n<p>How to optimize the graph to minimize the summary operations time cost?<\/p>\n","answers":[{"AnswerId":"54536701","CreationDate":"2019-02-05T14:33:24.233","ParentId":null,"OwnerUserId":"5224881","Title":null,"Body":"<p>Summaries of course slow down the training process, because you do more operations and you need to write them to disc. Also, histogram summaries slow the training even more, because for histograms you need more data to be copied from GPU to CPU than for scalar values.\nSo I would try to use histogram logging less often than the rest, that could make some difference. <\/p>\n\n<p>The usual solution is to compute summaries only every X batches. Since you compute summaries only one per epoch and not every batch, it might be worth trying even less summaries logging.<\/p>\n\n<p>Depends on how many batches you have in your dataset, but usually you don't lose much information by gathering a bit less logs.<\/p>\n"}]}
{"QuestionId":42407176,"AnswerCount":0,"Tags":"<tensorflow><google-cloud-ml>","CreationDate":"2017-02-23T04:37:55.947","AcceptedAnswerId":null,"OwnerUserId":1275011.0,"Title":"Exporting DNNRegressor model to Cloud ML","Body":"<p>I'm trying to use the operation:<\/p>\n\n<blockquote>\n  <p>tf.contrib.learn.DNNRegressor.export(*args, **kwargs)<\/p>\n<\/blockquote>\n\n<p>to create the export.meta file required for Google Cloud ML uploading.\nI've searched through the documentation (and Google) but didn't find a single practical example.<\/p>\n\n<p>I wrote a very simple regressor that approximates the sum between two numbers (smaller than one):<\/p>\n\n<pre><\/pre>\n\n<p>Everything works fine, but I was unable to build the input and output mappers required by the export function.\nWhen uploading to Cloud ML, I got errors like:<\/p>\n\n<blockquote>\n  <p>Model metagraph does not have outputs collection.<\/p>\n<\/blockquote>\n\n<p>Thanks!<\/p>\n","answers":[]}
{"QuestionId":42407277,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-02-23T04:49:38.553","AcceptedAnswerId":null,"OwnerUserId":3291312.0,"Title":"ValueError: all the input arrays must have same number of dimensions -for Tensorflow Manual MSINT","Body":"<p>I'm trying to take a stochastic gradient descent code I made in tensor flow and train 25112 images similar to the MINST dataset (the files look exactly like it). I apologize if this is a simple problem but I'm unsure of how to proceed. Thank you!<\/p>\n\n<p>I'm running into this error:<\/p>\n\n<p>\"ValueError: all the input arrays must have same number of dimensions\"<\/p>\n\n<p>On this line of code:<\/p>\n\n<p>x = np.c_[ np.ones(n), image_tensor2] #line75<\/p>\n\n<p>And I can't determine why this isn't working - I think it has something to do with how I've read in the image files but I can't be sure. Here is my code<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42407927","CreationDate":"2017-02-23T05:39:24.143","ParentId":null,"OwnerUserId":"4356204","Title":null,"Body":"<p>My intuition is that image_tensor is 3-D array while np.ones(n) is creating a 1-D array. <\/p>\n\n<p>But your goal is to insert a bias column (my guess) a quick way to do that would be:<\/p>\n\n<pre><code>b = np.ones((n + n+1))\nx = b[:, 1:] = image_tensor2\n<\/code><\/pre>\n\n<p><strong>example<\/strong><\/p>\n\n<pre><code>c = \narray([[7, 5, 6, 8, 7],\n       [5, 8, 7, 9, 7],\n       [9, 5, 6, 5, 8]])\n\nb = np.ones((3, 6))\nd = b[:, 1:] =c\n\nd =\narray([[ 1.,  7.,  5.,  6.,  8.,  7.],\n       [ 1.,  5.,  8.,  7.,  9.,  7.],\n       [ 1.,  9.,  5.,  6.,  5.,  8.]])\n<\/code><\/pre>\n"}]}
{"QuestionId":42407483,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-23T05:06:11.703","AcceptedAnswerId":42409162.0,"OwnerUserId":4993842.0,"Title":"Tensorflow supervisor for both training and evaluating operations?","Body":"<p>I've been using tensorflow supervisor (<a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/supervisor\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/supervisor<\/a>) for loading the model from the saved checkpoints for both training and running a network. But I noticed that the checkpoint files get updated even when running\/evaluating the model (timestamps of graph.pbtxt, model.ckpt.data files updated and new events.out created). <\/p>\n\n<p>This makes me wonder if using a supervisor for running\/evaluating the model resets\/alters the trained state as well? Is it advisable to use supervisor for anything other than training?<\/p>\n\n<p>Train -<\/p>\n\n<pre><\/pre>\n\n<p>Run\/Evaluate only. We don't want the below to modify the trained state of the model<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42409162","CreationDate":"2017-02-23T07:08:45.177","ParentId":null,"OwnerUserId":"2476373","Title":null,"Body":"<p>Your model is usually saved to a file names 'model.ckpt-NUM'. As long as the evaluation does not update that file (and it shouldn't), then you're safe.<\/p>\n\n<p>If you are worried about overwriting logging\/summaries. You should be careful in choosing the summaries names. <\/p>\n\n<p>E.g. for evaluation, choose a summary name <code>'eval\/' +metric_name<\/code>, and for training <code>'train\/' +metric_name<\/code>, for example, see <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/eval_image_classifier.py#L162\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/train_image_classifier.py#L320\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n\n<p>EDIT:\nYou can also choose a different directory (\"<code>logdir<\/code>\") for storing the evaluations results, as in the of API shown <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/summary\/generation_of_summaries_#FileWriter.__init__\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"}]}
{"QuestionId":42407611,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-23T05:17:08.937","AcceptedAnswerId":null,"OwnerUserId":7345996.0,"Title":"How tensorflow handle the computional graph when executing sess.run?","Body":"<p>when we using the code bellow:<\/p>\n\n<pre><\/pre>\n\n<p>does the tensorflow compute  twice, one for computing loss, one for computing the gradient, or compute  just once, using the computed  to compute  and ?\nThanks!<\/p>\n\n<p>edit.\nwhen tensorflow run the following code:<\/p>\n\n<pre><\/pre>\n\n<p>In this case, does tensorflow detect the overlapped subgraph between the subgraph of the  and  and then compute  first?<\/p>\n","answers":[{"AnswerId":"42408368","CreationDate":"2017-02-23T06:16:29.223","ParentId":null,"OwnerUserId":"2595935","Title":null,"Body":"<p>The <code>output<\/code> would only be computed once and then kept in memory until the the backward pass consumes it.<\/p>\n\n<p>There is some optimization that can be made in TensorFlow that if a tensor is fast to compute but large and would sit in memory for a long time, we could probably discard the output and recompute it when we need it next time.<\/p>\n"}]}
{"QuestionId":42407628,"AnswerCount":1,"Tags":"<python><python-2.7><machine-learning><tensorflow>","CreationDate":"2017-02-23T05:18:31.200","AcceptedAnswerId":null,"OwnerUserId":4370236.0,"Title":"ValueError: Dimensions must be equal, but are 784 and 500 for 'Mul' (op: 'Mul') with input shapes: [?,784], [784,500]","Body":"<p>I am trying to learn Tensor Flow and so I followed this tutorial on Neural Networks by <a href=\"https:\/\/pythonprogramming.net\/tensorflow-neural-network-session-machine-learning-tutorial\/\" rel=\"noreferrer\">https:\/\/pythonprogramming.net\/tensorflow-neural-network-session-machine-learning-tutorial\/<\/a><\/p>\n\n<p>I am trying to run the code, but keep getting the same dimension Error even when my dimensions seem correct.<\/p>\n\n<p>I am new to Tensor Flow, so I am not exactly sure what am I doing wrong.<\/p>\n\n<p>I'll post the code and the error.<\/p>\n\n<pre><\/pre>\n\n<p>The error I am getting is the following-<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42417595","CreationDate":"2017-02-23T13:47:49.920","ParentId":null,"OwnerUserId":"7096238","Title":null,"Body":"<p>The error comes because you use \"multiply\"<\/p>\n\n<p>In all the lines where you use <\/p>\n\n<pre><code>tf.add(tf.multiply(.....))\n<\/code><\/pre>\n\n<p>Use:<\/p>\n\n<pre><code>tf.add(tf.matmul(......))\n<\/code><\/pre>\n\n<p>Because this is the matrix multiplication.<\/p>\n"}]}
{"QuestionId":42407736,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><autoencoder>","CreationDate":"2017-02-23T05:26:19.127","AcceptedAnswerId":null,"OwnerUserId":4699762.0,"Title":"Training yields very low error but very incorrect (almost all ones) output","Body":"<p>I am trying to train an autoencoder on some simulated data where an input is basically a vector with Gaussian noise applied. The code is almost exactly the same as in this example: <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/autoencoder.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/autoencoder.py<\/a> <\/p>\n\n<p>The only differences are I changed the network parameters and the cost function:<\/p>\n\n<pre><\/pre>\n\n<p>During training, the error steadily decreases down to 0.00015, but the predicted and true values are very different, e.g.\n<a href=\"https:\/\/i.stack.imgur.com\/F1CNd.png\" rel=\"nofollow noreferrer\">as shown in this image<\/a>. In fact, the predicted y vector is almost all ones.<\/p>\n\n<p>How is it possible to get decreasing error with very wrong predictions? Is it possible that my network is just trying to move the weights closer to log(1) so as to minimize the cross entropy cost? If so, how do I combat this?<\/p>\n","answers":[{"AnswerId":"42415146","CreationDate":"2017-02-23T11:58:05.923","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>Yes, the network simply learns to to predict <code>1<\/code> which reduces the loss. The cross-entropy loss you are using is categorical which is used when <code>y_true<\/code> is one-hot code (example: [0,0,1,0]) and final layer is <code>softmax<\/code> (ensures sum of all output is 1). So when y_true[idx] is 0, the loss don't care while when the y_true[idx] is 1 and y_pred[idx] is 0 there is infinite(high) loss but if its 1 then loss is again 0. <\/p>\n\n<p>Now categorical cross-entropy loss is not suitable for autoencoders. For real valued inputs and hence outputs its mean-squared-error, which is what is used in example that you cited. But there the final activation layer is <code>sigmoid<\/code>, implicitly saying that each element of <code>x<\/code> is 0\/1. So either you need to convert your data to support the same or have the last layer of decoder linear. <\/p>\n\n<p>If you do want to use cross-entropy loss you can use binary cross-entropy \nFor inputs with 0,1 binary cross-entropy:  <code>tf.reduce_mean(y_true * tf.log(y_pred) + (1-y_true) * tf.log(1-y_pred))<\/code>. If you work it out in both misprediction case 0-1, 1-0, the network gets infinite loss. Note again here the final layer should be <code>softmax<\/code> and elements of <code>x<\/code> should be between 0 and 1<\/p>\n"}]}
{"QuestionId":42408293,"AnswerCount":1,"Tags":"<python><docker><tensorflow><deep-learning><image-recognition>","CreationDate":"2017-02-23T06:10:00.310","AcceptedAnswerId":43016216.0,"OwnerUserId":7124519.0,"Title":"Docker Error getting events from daemon: EOF","Body":"<hr>\n\n<h2>BUG REPORT INFORMATION<\/h2>\n\n<p><strong>Description<\/strong><\/p>\n\n<p>Hello everyone, after following the google codelabs, <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/?utm_campaign=chrome_series_machinelearning_063016&amp;utm_source=gdev&amp;utm_medium=yt-desc#3\" rel=\"nofollow noreferrer\">Codelabs<\/a> I have received an error    after  <\/p>\n\n<p>Update:\nI reran it and this shows up\n<\/p>\n\n<p><strong>Steps to reproduce the issue:<\/strong>\n1.\n```\n python tensorflow\/examples\/image_retraining\/retrain.py \\<\/p>\n\n<blockquote>\n  <p>--bottleneck_dir=\/tf_files\/bottlenecks \\\n  --how_many_training_steps 500 \\\n  --model_dir=\/tf_files\/inception \\\n  --output_graph=\/tf_files\/retrained_graph.pb \\\n  --output_labels=\/tf_files\/retrained_labels.txt \\\n  --image_dir \/tf_files\/flower_photos<\/p>\n<\/blockquote>\n\n<p>```<\/p>\n\n<p><strong>Describe the results you received:<\/strong>\n<\/p>\n\n<p><strong>Describe the results you expected:<\/strong>\n<\/p>\n\n<p><strong>Output of :<\/strong><\/p>\n\n<p><\/p>\n\n<p><strong>Output of :<\/strong><\/p>\n\n<p><\/p>\n\n<p><strong>Additional environment details (AWS, VirtualBox, physical, etc.):<\/strong>\nOS X with python 2.7, \nand this shows up\n<\/p>\n","answers":[{"AnswerId":"43016216","CreationDate":"2017-03-25T12:17:46.667","ParentId":null,"OwnerUserId":"7124519","Title":null,"Body":"<p>The solution is to increase the CPU size and Ram in Docker preference.<\/p>\n"}]}
{"QuestionId":42409884,"AnswerCount":2,"Tags":"<python><tensorflow><keras><python-multiprocessing>","CreationDate":"2017-02-23T07:49:08.733","AcceptedAnswerId":44300201.0,"OwnerUserId":3431836.0,"Title":"Keras + Tensorflow: Prediction on multiple gpus","Body":"<p>I'm using Keras with tensorflow as backend.\nI have one compiled\/trained model.<\/p>\n\n<p>My prediction loop is slow so I would like to find a way to parallelize the  calls to speed things up.\nI would like to take a list of batches (of data) and then per available gpu, run  over a subset of those batches.<br>\nEssentially:<\/p>\n\n<pre><\/pre>\n\n<p>I know that it's possible in pure Tensorflow to assign ops to a given gpu (<a href=\"https:\/\/www.tensorflow.org\/tutorials\/using_gpu\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/using_gpu<\/a>).  However, I don't know how this translates to my situation since I've built\/compiled\/trained my model using Keras' api.<\/p>\n\n<p>I had thought that maybe I just needed to use python's multiprocessing module and start a process per gpu that would run .  I know this is theoretically possible given another SO post of mine: <a href=\"https:\/\/stackoverflow.com\/questions\/42504669\/keras-tensorflow-and-multiprocessing-in-python\">Keras + Tensorflow and Multiprocessing in Python<\/a>.  However, this still leaves me with the dilemma of not knowing how to actually \"choose\" a gpu to operate the process on.<\/p>\n\n<p>My question boils down to: how does one parallelize prediction for one model in Keras across multiple gpus when using Tensorflow as Keras' backend?<\/p>\n\n<p>Additionally I am curious if similar parallelization for prediction is possible with only one gpu.   <\/p>\n\n<p>A high level description or code example would be greatly appreciated!<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"44300201","CreationDate":"2017-06-01T06:38:42.373","ParentId":null,"OwnerUserId":"1314209","Title":null,"Body":"<p>I created one simple example to show how to run keras model across multiple gpus. Basically, multiple processes are created and each of process owns a gpu. To specify the gpu id in process, setting env variable CUDA_VISIBLE_DEVICES is a very straightforward way (os.environ[\"CUDA_VISIBLE_DEVICES\"]).  Hope this git repo can help you.<\/p>\n\n<p><a href=\"https:\/\/github.com\/yuanyuanli85\/Keras-Multiple-Process-Prediction\" rel=\"noreferrer\">https:\/\/github.com\/yuanyuanli85\/Keras-Multiple-Process-Prediction<\/a><\/p>\n"},{"AnswerId":"44873088","CreationDate":"2017-07-02T16:47:30.383","ParentId":null,"OwnerUserId":"1351629","Title":null,"Body":"<p>You can use this function to parallelize a Keras model (credits to <a href=\"https:\/\/github.com\/kuza55\" rel=\"nofollow noreferrer\">kuza55<\/a>).<br>\n<a href=\"https:\/\/github.com\/kuza55\/keras-extras\/blob\/master\/utils\/multi_gpu.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kuza55\/keras-extras\/blob\/master\/utils\/multi_gpu.py<\/a><br>\n.<\/p>\n\n<pre><code>from keras.layers import merge\nfrom keras.layers.core import Lambda\nfrom keras.models import Model\n\nimport tensorflow as tf\n\ndef make_parallel(model, gpu_count):\n    def get_slice(data, idx, parts):\n        shape = tf.shape(data)\n        size = tf.concat([ shape[:1] \/\/ parts, shape[1:] ],axis=0)\n        stride = tf.concat([ shape[:1] \/\/ parts, shape[1:]*0 ],axis=0)\n        start = stride * idx\n        return tf.slice(data, start, size)\n\n    outputs_all = []\n    for i in range(len(model.outputs)):\n        outputs_all.append([])\n\n    #Place a copy of the model on each GPU, each getting a slice of the batch\n    for i in range(gpu_count):\n        with tf.device('\/gpu:%d' % i):\n            with tf.name_scope('tower_%d' % i) as scope:\n\n                inputs = []\n                #Slice each input into a piece for processing on this GPU\n                for x in model.inputs:\n                    input_shape = tuple(x.get_shape().as_list())[1:]\n                    slice_n = Lambda(get_slice, output_shape=input_shape, arguments={'idx':i,'parts':gpu_count})(x)\n                    inputs.append(slice_n)                \n\n                outputs = model(inputs)\n\n                if not isinstance(outputs, list):\n                    outputs = [outputs]\n\n                #Save all the outputs for merging back together later\n                for l in range(len(outputs)):\n                    outputs_all[l].append(outputs[l])\n\n    # merge outputs on CPU\n    with tf.device('\/cpu:0'):\n        merged = []\n        for outputs in outputs_all:\n            merged.append(merge(outputs, mode='concat', concat_axis=0))\n\n        return Model(input=model.inputs, output=merged)\n<\/code><\/pre>\n"}]}
{"QuestionId":42410237,"AnswerCount":1,"Tags":"<tensorflow><softmax>","CreationDate":"2017-02-23T08:11:18.693","AcceptedAnswerId":null,"OwnerUserId":775755.0,"Title":"tensor flow character recognition with softmax results in accuracy 1 due to [NaN...NaN] prediction","Body":"<p>I am trying to use the softmax regression method discussed in <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners<\/a> to recognize characters. <\/p>\n\n<p>My code is as follows. <\/p>\n\n<pre><\/pre>\n\n<p>However, I am getting an accuracy of 1, which shouldn't be the case. \nThe reason why I am getting it so is because my y tensor results with an array like<\/p>\n\n<pre><\/pre>\n\n<p>Can anyone explain to me what is wrong in my code?<\/p>\n\n<p>I converted each character to a one-hot encoding using the method below<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42410823","CreationDate":"2017-02-23T08:45:36.817","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Here is the problem you are having: <\/p>\n\n<ul>\n<li>You set your initial weights and biases to 0 (this is wrong, as your\nnetwork does not learn). <\/li>\n<li>The result is that y consists of all zeros<\/li>\n<li>You take the log of y.. and a log of 0 is not defined... Hence the NaN. <\/li>\n<\/ul>\n\n<p>Good luck!<\/p>\n\n<p>Edit to tell you how to fix it: look for an example on classifying MNIST characters and see what they do. You probably want to initialise your weights to be random normals ;)<\/p>\n"}]}
{"QuestionId":42410493,"AnswerCount":1,"Tags":"<caffe><pycaffe>","CreationDate":"2017-02-23T08:27:15.733","AcceptedAnswerId":null,"OwnerUserId":3198732.0,"Title":"How to use multi-GPU training with Python using Caffe (pycaffe)?","Body":"<p>I saw that recently a <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/4563\" rel=\"nofollow noreferrer\">new commit<\/a> was inserted to Caffe, that should enable multi-GPU training with Python interface\/Python custom layers.\nI can't seem to train a network using multiple-GPU from my Python code.\nI cloned the updated repository and built Caffe with USE_NCCL on.\nThe commands I tried to use in my Python script:<\/p>\n\n<pre><\/pre>\n\n<p>The error I get:<\/p>\n\n<blockquote>\n  <p>F0223 10:27:41.249380 23468 python_layer.hpp:25] PythonLayer does not\n  support CLI Multi-GPU, use train.py<\/p>\n<\/blockquote>\n\n<p>When running the script from the command line, it simply uses 1 GPU only.<\/p>\n","answers":[{"AnswerId":"45876002","CreationDate":"2017-08-25T07:00:52.113","ParentId":null,"OwnerUserId":"8344845","Title":null,"Body":"<p>I also recently tried the multi GPU support in Python Caffe. You might want to give this a read:\n<a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/docs\/multigpu.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/blob\/master\/docs\/multigpu.md<\/a><\/p>\n\n<p>Caffe only supports multi-GPU from command line and only during TRAIN i.e you have to use the train.py file (.\/build\/tools\/caffe train) and give the GPU's you want to use as arguments to this script. \nIt is pretty well explained in the file I pointed you above. <\/p>\n\n<p>I tried to use the caffe.set_solver.count(2) in my python script as well. I didn't get any error however my model trains only on one GPU. <\/p>\n\n<p>-Ayushi<\/p>\n"}]}
{"QuestionId":42411891,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2017-02-23T09:38:27.413","AcceptedAnswerId":42412124.0,"OwnerUserId":4338302.0,"Title":"How to extract bias weights in Keras sequential model?","Body":"<p>I'm running a simple feed-forward network using <em>Keras<\/em> . \nHaving just one hidden layer I would like to make some inference regarding the relevance of each input to each output and I would like to extract the weights. <\/p>\n\n<p>This is the model: <\/p>\n\n<pre><\/pre>\n\n<p>To extract the weight I wrote: <\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately I don't get the biases columns in the matrices, which I know Keras automatically puts in it. <\/p>\n\n<p><strong>Do you know how to retrieve the biases weights?<\/strong><\/p>\n\n<p>Thank you in advance for your help !<\/p>\n","answers":[{"AnswerId":"42412124","CreationDate":"2017-02-23T09:48:13.537","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>get_weights() for a Dense layer returns a list of two elements, the first element contains the weights, and the second element contains the biases. So you can simply do:<\/p>\n\n<pre><code>weights = model.layers[0].get_weights()[0]\nbiases = model.layers[0].get_weights()[1]\n<\/code><\/pre>\n\n<p>Note that weights and biases are already numpy arrays.<\/p>\n"}]}
{"QuestionId":42412103,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-23T09:47:11.500","AcceptedAnswerId":42414255.0,"OwnerUserId":6591278.0,"Title":"Tensorflow: Get correct accuracy of NN","Body":"<p>I'm stuck too long and would need some help (very new to tensorflow etc). I modified an MNIST example to my own data but keep getting 100% accuracy, even after 2 epochs.<br>\nMy X is (similar to MNIST) a [18, 1]-vector and y a float32.<br>\n<strong>variables:<\/strong>  <\/p>\n\n<pre><\/pre>\n\n<p><strong>my model:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Session:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The question is primarely why I always get 100% accuracy, which is obviously false. Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42414255","CreationDate":"2017-02-23T11:18:01.477","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>MNIST usually has one-hot coded output. In that case <code>correct_prediction = tf.equal(tf.argmax(prediction, 1), tf.argmax(y,1))<\/code> makes sense as <code>tf.argmax<\/code> converts the one-hot code to the actual class which is then compared by <code>tf.equal<\/code>. <\/p>\n\n<p>But in your case since the output dimension is <code>1<\/code> <code>tf.argmax<\/code> outputs <code>0<\/code> which is the only valid index for both and hence they are equal and hence 100% accuracy. <\/p>\n\n<p>You need to redefine the accuracy appropriate to your case. Assuming the output is a binary value (as accuracy as a metrics only makes sense for binary case) you can use the following:<\/p>\n\n<pre><code>  correct_prediction = tf.equal(tf.round(prediction), y)\n<\/code><\/pre>\n\n<p>Here you are rounding off the <code>prediction<\/code> and then comparing to <code>y<\/code>. For this to work you need to have sigmoid as the final layer activation <code>output = tf.nn.sigmoid(output)<\/code> or you need to clip the predictions:<\/p>\n\n<pre><code>  correct_prediction = tf.equal(tf.round(tf.clip_by_value(prediction,0,1)), y)\n<\/code><\/pre>\n\n<p>Other option for you will be to convert <code>y_vals_train<\/code>, <code>y_vals_test<\/code> to one-hot code and have the network output 2 dimension. <\/p>\n"}]}
{"QuestionId":42412660,"AnswerCount":1,"Tags":"<tensorflow><non-deterministic>","CreationDate":"2017-02-23T10:10:56.247","AcceptedAnswerId":null,"OwnerUserId":783758.0,"Title":"Non-deterministic Gradient Computation","Body":"<p>I realized that my models end up being different every time I train them, even though I keep the TensorFlow random seed the same.<\/p>\n\n<p>I verified that:<\/p>\n\n<ul>\n<li>Initialization is deterministic; the weights are identical before the first update.<\/li>\n<li>Inputs are deterministic. In fact, various forward computations, including the loss, are identical for the very first batch.<\/li>\n<li>The gradients for the first batch are different. Concretely, I'm comparing the outputs of . While  and  have identical values, the gradients are <em>sometimes<\/em> different for some of the Variables. The differences are quite significant (sometimes the sum-of-absolute-differences for a single variable's gradient is greater than 1).<\/li>\n<\/ul>\n\n<p>I conclude that it's the gradient computation that causes the non-determinism.\nI had a look at <a href=\"https:\/\/stackoverflow.com\/questions\/39938307\/determinism-in-tensorflow-gradient-updates\">this question<\/a> and the problem persists when running on a CPU with  and .<\/p>\n\n<p>How can the backward pass be non-deterministic when the forward pass isn't? How could I debug this further?<\/p>\n","answers":[{"AnswerId":"42414877","CreationDate":"2017-02-23T11:44:53.710","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>This answer might seem a little obvious, but do you use some kind of non-deterministic regularization such as dropout? Given that dropout \"drops\" some connections randomly when training, it may be causing that difference on the gradients.<\/p>\n\n<p><strong>Edit:<\/strong> Similar questions:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/36288235\/how-to-get-stable-results-with-tensorflow-setting-random-seed\/36289575#36289575\">How to get stable results with TensorFlow, setting random seed<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/36984291\/tensorflow-not-being-deterministic-where-it-should\">Tensorflow not being deterministic, where it should<\/a><\/li>\n<\/ul>\n\n<p><strong>Edit 2:<\/strong> This seems to be an issue with TensorFlow's implementation. See the following open issues in GitHub:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16889\" rel=\"nofollow noreferrer\">Problems Getting TensorFlow to behave Deterministically<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/openai\/baselines\/issues\/805\" rel=\"nofollow noreferrer\">Non-deterministic behaviour when ran on GPU<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":42414299,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2017-02-23T11:19:44.113","AcceptedAnswerId":42429595.0,"OwnerUserId":4692795.0,"Title":"Change number of GPUs when deploy","Body":"<p>I have a fine-tuned Inception v3 on a 2-GPU machine.\nNow I am trying to run the trained model on another machine with 1 GPU, but I got an error like this:<\/p>\n\n<blockquote>\n  <p>Cannot assign a device to node 'tower_1\/gradients\/tower_1\/conv0\/BatchNorm\/moments\/sufficient_statistics\/Sub_grad\/BroadcastGradientArgs': Could not satisfy explicit device specification '\/device:GPU:1' because no devices matching that specification are registered in this process; available devices: \/job:localhost\/replica:0\/task:0\/cpu:0, \/job:localhost\/replica:0\/task:0\/gpu:0<\/p>\n<\/blockquote>\n\n<p>It seems that the model wants a 2-GPU environment like the one it was trained on. Can I convert this model so that it uses only 1 GPU?<\/p>\n","answers":[{"AnswerId":"42429595","CreationDate":"2017-02-24T02:14:40.497","ParentId":null,"OwnerUserId":"4692795","Title":null,"Body":"<p>I changed two things and it worked.<\/p>\n\n<ol>\n<li><p>Turn on <code>allow_soft_placement<\/code> option in Session:<\/p>\n\n<pre><code>config = tf.ConfigProto(allow_soft_placement=True)\nsess = tf.Session(config=config)\n<\/code><\/pre><\/li>\n<li><p>Rename model file name from <code>model.ckpt-50000.data-00000-of-00001<\/code> to <code>model.ckpt-50000<\/code><\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42414359,"AnswerCount":0,"Tags":"<customization><keras><layer><convolution>","CreationDate":"2017-02-23T11:22:23.937","AcceptedAnswerId":null,"OwnerUserId":7158655.0,"Title":"Custom Layers for recursive interation bitween layers","Body":"<p>I have a question. I'm trying to make a CNN, the return of model feeds first Dence layer. Can I help?<br>\ne.g:<\/p>\n\n<pre><\/pre>\n\n<p>I Should have the output value feeds layer Dense (512)<\/p>\n\n<ol>\n<li>Dense (512 + 10, name='d1') output  d2 with input here.<\/li>\n<li>Activation()  <\/li>\n<li>Dropout(522)  <\/li>\n<li>Dense (10, name='d2')<\/li>\n<\/ol>\n","answers":[]}
{"QuestionId":42415076,"AnswerCount":2,"Tags":"<machine-learning><scikit-learn><pipeline><keras><hyperparameters>","CreationDate":"2017-02-23T11:54:30.357","AcceptedAnswerId":null,"OwnerUserId":1684118.0,"Title":"How to insert Keras model into scikit-learn pipeline?","Body":"<p>I'm using a Scikit-Learn custom pipeline () in conjunction with  for hyper-parameter optimization. This works great.<\/p>\n\n<p>Now I would like to insert a Keras model as a first step into the pipeline. Parameters of the model should be optimized. The computed (fitted) Keras model should then be used later on in the pipeline by other steps, so I think I have to store the model as a global variable so that the other pipeline steps can use it. Is this right?<\/p>\n\n<p>I know that Keras offers some wrappers for the Scikit-Learn API but the problem is that these wrappers already do classification \/ regression but I only want to compute the Keras model and nothing else.<\/p>\n\n<p>How can this be done?<\/p>\n\n<p>For example I have a method which returns the model:<\/p>\n\n<pre><\/pre>\n\n<p>The method needs some fixed parameters like a file path etc. but X and y is not needed (or can be ignored). The parameters of the model should be optimized (number of layers etc.).<\/p>\n","answers":[{"AnswerId":"47516072","CreationDate":"2017-11-27T16:54:56.987","ParentId":null,"OwnerUserId":"7447273","Title":null,"Body":"<p>This is a modification of the RBM example in sklearn documentation (<a href=\"http:\/\/scikit-learn.org\/stable\/auto_examples\/neural_networks\/plot_rbm_logistic_classification.html#sphx-glr-auto-examples-neural-networks-plot-rbm-logistic-classification-py\" rel=\"nofollow noreferrer\">http:\/\/scikit-learn.org\/stable\/auto_examples\/neural_networks\/plot_rbm_logistic_classification.html#sphx-glr-auto-examples-neural-networks-plot-rbm-logistic-classification-py<\/a>)<\/p>\n\n<p>but the neural network implemented in keras with tensorflow backend<\/p>\n\n<pre><code>    # -*- coding: utf-8 -*-\n    \"\"\"\n    Created on Mon Nov 27 17:11:21 2017\n\n    @author: ZED\n    \"\"\"\n\n    from __future__ import print_function\n\n    print(__doc__)\n\n    # Authors: Yann N. Dauphin, Vlad Niculae, Gabriel Synnaeve\n    # License: BSD\n\n    import numpy as np\n    import matplotlib.pyplot as plt\n\n    from scipy.ndimage import convolve\n\n    from keras.models import Sequential\n    from keras.layers.core import Dense,Activation\n    from keras.wrappers.scikit_learn import KerasClassifier\n    from keras.utils import np_utils\n\n    from sklearn import  datasets, metrics\n    from sklearn.model_selection import train_test_split\n    from sklearn.neural_network import BernoulliRBM\n    from sklearn.pipeline import Pipeline\n\n\n    #%%\n    # Setting up\n\n    def nudge_dataset(X, Y):\n        \"\"\"\n        This produces a dataset 5 times bigger than the original one,\n        by moving the 8x8 images in X around by 1px to left, right, down, up\n        \"\"\"\n        direction_vectors = [\n            [[0, 1, 0],\n             [0, 0, 0],\n             [0, 0, 0]],\n\n            [[0, 0, 0],\n             [1, 0, 0],\n             [0, 0, 0]],\n\n            [[0, 0, 0],\n             [0, 0, 1],\n             [0, 0, 0]],\n\n            [[0, 0, 0],\n             [0, 0, 0],\n             [0, 1, 0]]]\n\n        shift = lambda x, w: convolve(x.reshape((8, 8)), mode='constant',\n                                      weights=w).ravel()\n        X = np.concatenate([X] +\n                           [np.apply_along_axis(shift, 1, X, vector)\n                            for vector in direction_vectors])\n        Y = np.concatenate([Y for _ in range(5)], axis=0)\n        return X, Y\n\n    # Load Data\n    digits = datasets.load_digits()\n    X = np.asarray(digits.data, 'float32')\n    X, Y = nudge_dataset(X, digits.target)\n    X = (X - np.min(X, 0)) \/ (np.max(X, 0) + 0.0001)  # 0-1 scaling\n\n    X_train, X_test, Y_train, Y_test = train_test_split(X, Y,\n                                                        test_size=0.2,\n                                                        random_state=0)\n\n    #%%\n    def create_model():\n\n        model = Sequential()\n        model.add(Dense(100, input_dim=64))\n        model.add(Activation('tanh'))\n\n        \"\"\"\n        #other layer\n        model.add(Dense(500))\n        model.add(Activation('tanh'))\n        \"\"\"\n\n        model.add(Dense(10))\n        model.add(Activation('softmax'))\n        # Compile model\n        model.compile(loss = 'binary_crossentropy', optimizer = 'adadelta', metrics=['accuracy'])\n        return model\n\n    rbm = BernoulliRBM(random_state=0, verbose=True)\n\n    #This is the model you want. it is in sklearn format\n    clf = KerasClassifier(build_fn=create_model, verbose=0)\n\n    classifier = Pipeline(steps=[('rbm', rbm), ('VNN', clf)])\n\n    #%%\n    # Training\n\n    # Hyper-parameters. These were set by cross-validation,\n    # using a GridSearchCV. Here we are not performing cross-validation to\n    # save time.\n    rbm.learning_rate = 0.06\n    rbm.n_iter = 20\n    # More components tend to give better prediction performance, but larger\n    # fitting time\n    rbm.n_components = 64\n\n    #adapt targets to hot matrix\n    yTrain = np_utils.to_categorical(Y_train, 10)\n    # Training RBM-Logistic Pipeline\n    classifier.fit(X_train, yTrain)\n\n    #%%\n    # Evaluation\n\n    print()\n    print(\"NN using RBM features:\\n%s\\n\" % (\n        metrics.classification_report(\n            Y_test,\n            classifier.predict(X_test))))\n\n    #%%\n    # Plotting\n\n    plt.figure(figsize=(4.2, 4))\n    for i, comp in enumerate(rbm.components_):\n        plt.subplot(10, 10, i + 1)\n        plt.imshow(comp.reshape((8, 8)), cmap=plt.cm.gray_r,\n                   interpolation='nearest')\n        plt.xticks(())\n        plt.yticks(())\n    plt.suptitle('64 components extracted by RBM', fontsize=16)\n    plt.subplots_adjust(0.08, 0.02, 0.92, 0.85, 0.08, 0.23)\n\n    plt.show()\n<\/code><\/pre>\n"},{"AnswerId":"47520976","CreationDate":"2017-11-27T22:33:26.560","ParentId":null,"OwnerUserId":"436721","Title":null,"Body":"<p>You need to <strong>wrap your Keras model as a Scikit learn model<\/strong> first, and then just proceed as normal. <\/p>\n\n<p>Here's a quick example (I've omitted the imports for brevity)<\/p>\n\n<blockquote>\n  <p>Here is a full blog post with this one and many other examples: <a href=\"http:\/\/queirozf.com\/entries\/scikit-learn-pipeline-examples\" rel=\"noreferrer\">Scikit-learn Pipeline Examples<\/a><\/p>\n<\/blockquote>\n\n<pre><code># create a function that returns a model, taking as parameters things you\n# want to verify using cross-valdiation and model selection\ndef create_model(optimizer='adagrad',\n                 kernel_initializer='glorot_uniform', \n                 dropout=0.2):\n    model = Sequential()\n    model.add(Dense(64,activation='relu',kernel_initializer=kernel_initializer))\n    model.add(Dropout(dropout))\n    model.add(Dense(1,activation='sigmoid',kernel_initializer=kernel_initializer))\n\n    model.compile(loss='binary_crossentropy',optimizer=optimizer, metrics=['accuracy'])\n\n    return model\n\n# wrap the model using the function you created\nclf = KerasRegressor(build_fn=create_model,verbose=0)\n\n# just create the pipeline\npipeline = Pipeline([\n    ('clf',clf)\n])\n\npipeline.fit(X_train, y_train)\n<\/code><\/pre>\n"}]}
{"QuestionId":42415125,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-02-23T11:56:58.677","AcceptedAnswerId":null,"OwnerUserId":6888252.0,"Title":"How to implement a neural network in tensorflow with multiple right solutions","Body":"<p>I am currently trying to code a CNN with tensorflow in order to play poker.\nThe state of the game is the input a 9x17x17 tensor, the output should be 10 values in between 0 and 1 and the highest should then be determined as my move.\nTo train this network I prepared samples with the state, but I don't have the kind of labels which I used so far as for example in the MNIST dataset. Instead I got basically a list with 1 entry for each output containing the normalized win\/loss scenario. 0-0.5 means loss, where 0 means all chips were lost and 0.5-1 means win with 1 meaning that all chips were won.\nIn order to test if my network is working I created game states against one all-in bot.\nThe output of my network might be something like this:\n[0.322, 0.5553, 0.003, 0.773, 0.123, 0.444, 0.122, 0.888, 0.921, 0.232]\nEach entry represents one respective move.\nThe actual game results (vs the all-in player) are something like this:\n[0.475 1 1 1 1 1 1 1 1 1]<\/p>\n\n<p>My idea for the cost function was to simply use the maximum (or rather the index of the maximum) out of the predicted moves and use the respective result which I then want to maximize.\nI tried to implement this in the following way:<\/p>\n\n<pre><\/pre>\n\n<p>Unfortunately this does not work and I get the following error:\nValueError: No gradients provided for any variable<\/p>\n\n<p>How can I achieve a somewhat similar cost function?<\/p>\n","answers":[]}
{"QuestionId":42415515,"AnswerCount":1,"Tags":"<c++><caffe><ubuntu-16.04><openblas>","CreationDate":"2017-02-23T12:15:37.937","AcceptedAnswerId":null,"OwnerUserId":429476.0,"Title":"Error while compiling Caffe","Body":"<p>I am trying to compile Caffe with OpenBLAS in Ubuntu 16.04. I had downloaded and compile OpenBLAS with <strong>make NO_AFFINITY=1 USE_OPENMP=1<\/strong>\nOpenBLAS compilation is fine and all its tests are also run<\/p>\n\n<pre><\/pre>\n\n<p>However linking with OpenBLAS libs in Caffe install is giving the following error<\/p>\n\n<blockquote>\n  <p>Linking CXX executable upgrade_net_proto_binary\n  \/opt\/OpenBLAS\/lib\/libopenblas.so: undefined reference to `GOMP_parallel@GOMP_4.0'\n  collect2: error: ld returned 1 exit status\n  tools\/CMakeFiles\/upgrade_net_proto_binary.dir\/build.make:129: recipe for target 'tools\/upgrade_net_proto_binary' failed\n  make[2]: <strong>* [tools\/upgrade_net_proto_binary] Error 1\n  CMakeFiles\/Makefile2:403: recipe for target 'tools\/CMakeFiles\/upgrade_net_proto_binary.dir\/all' failed\n  make[1]: *<\/strong> [tools\/CMakeFiles\/upgrade_net_proto_binary.dir\/all] Error 2<\/p>\n<\/blockquote>\n","answers":[{"AnswerId":"42416019","CreationDate":"2017-02-23T12:39:13.347","ParentId":null,"OwnerUserId":"1317944","Title":null,"Body":"<p>Most likely the reason is that you're linking to the compiled OpenBLAS instead of installing it and linking to the installation. After you run <code>make<\/code>, you should run <code>make install PREFIX=\/path\/to\/install\/OpenBLAS<\/code>, and then you should link to the installed version.<\/p>\n"}]}
{"QuestionId":42415909,"AnswerCount":3,"Tags":"<tensorflow><neural-network><deep-learning><keras><lstm>","CreationDate":"2017-02-23T12:34:36.880","AcceptedAnswerId":42420389.0,"OwnerUserId":3022072.0,"Title":"Initializing LSTM hidden state Tensorflow\/Keras","Body":"<p>Can someone explain how can I initialize hidden state of LSTM in tensorflow? I am trying to build LSTM recurrent auto-encoder, so after i have that model trained i want to transfer learned hidden state of unsupervised model to hidden state of supervised model.\nIs that even possible with current API?\nThis is paper I am trying to recreate:<\/p>\n\n<p><a href=\"http:\/\/papers.nips.cc\/paper\/5949-semi-supervised-sequence-learning.pdf\" rel=\"noreferrer\">http:\/\/papers.nips.cc\/paper\/5949-semi-supervised-sequence-learning.pdf<\/a><\/p>\n","answers":[{"AnswerId":"42420389","CreationDate":"2017-02-23T15:51:08.720","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Yes - this is possible but truly cumbersome. Let's go through an example.<\/p>\n\n<ol>\n<li><p>Defining a model:<\/p>\n\n<pre><code>from keras.layers import LSTM, Input\nfrom keras.models import Model\n\ninput = Input(batch_shape=(32, 10, 1))\nlstm_layer = LSTM(10, stateful=True)(input)\n\nmodel = Model(input, lstm_layer)\nmodel.compile(optimizer=\"adam\", loss=\"mse\")\n<\/code><\/pre>\n\n<p>It's important to build and compile model first as in compilation the initial states are reset. Moreover - you need to specify a <code>batch_shape<\/code> where <code>batch_size<\/code> is specified as in this scenario our network should be <code>stateful<\/code> (which is done by setting a <code>stateful=True<\/code> mode.<\/p><\/li>\n<li><p>Now we could set the values of initial states:<\/p>\n\n<pre><code>import numpy\nimport keras.backend as K\n\nhidden_states = K.variable(value=numpy.random.normal(size=(32, 10)))\ncell_states = K.variable(value=numpy.random.normal(size=(32, 10)))\n\nmodel.layers[1].states[0] = hidden_states\nmodel.layers[1].states[1] = cell_states \n<\/code><\/pre>\n\n<p>Note that you need to provide states as a <code>keras<\/code> variables. <code>states[0]<\/code> holds hidden states and <code>states[1]<\/code> holds cell states.<\/p><\/li>\n<\/ol>\n\n<p>Hope that helps.<\/p>\n"},{"AnswerId":"57549285","CreationDate":"2019-08-19T00:23:06.797","ParentId":null,"OwnerUserId":"11944236","Title":null,"Body":"<p>I used this approach, totally worked out for me:<\/p>\n\n<p>lstm_cell = LSTM(cell_num, return_state=True) <\/p>\n\n<p>output, h, c = lstm_cell(input, initial_state=[h_prev, c_prev])<\/p>\n"},{"AnswerId":"53466346","CreationDate":"2018-11-25T09:49:57.970","ParentId":null,"OwnerUserId":"6646029","Title":null,"Body":"<p>Assuming an RNN is in layer 1 and hidden\/cell states are numpy arrays.  You can do this:<\/p>\n\n<pre><code>from keras import backend as K\n\nK.set_value(model.layers[1].states[0], hidden_states)\nK.set_value(model.layers[1].states[1], cell_states)\n<\/code><\/pre>\n\n<p>States can also be set using<\/p>\n\n<pre><code>model.layers[1].states[0] = hidden_states\nmodel.layers[1].states[1] = cell_states\n<\/code><\/pre>\n\n<p>but when I did it this way my state values stayed constant even after stepping the RNN.<\/p>\n"}]}
{"QuestionId":42416000,"AnswerCount":1,"Tags":"<python><while-loop><tensorflow><lstm>","CreationDate":"2017-02-23T12:38:23.667","AcceptedAnswerId":42420309.0,"OwnerUserId":7523983.0,"Title":"Gradients error using TensorArray Tensorflow","Body":"<p>i am trying to implement multidimentional lstm in tensorflow, I am using TensorArray to remember previous states, i am using a complicated way to get two neigbours state(above and from left). tf.cond want that both posible condition to exist and to have the same number of inputs. this is why i added one more cell.zero_state to the (last index +1) of the states. then i using a function to get the correct indexes to the states. when i am trying to use an optimizer in order to minimize a cost, i getting that error:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): TensorArray\n  MultiDimentionalLSTMCell-l1-multi-l1\/state_ta_262@gradients: Could not\n  read from TensorArray index 809 because it has not yet been written\n  to.<\/p>\n<\/blockquote>\n\n<p>Can someone tell how to fix it?<\/p>\n\n<p>Ps: without optimizer it works!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42420309","CreationDate":"2017-02-23T15:47:47.637","ParentId":null,"OwnerUserId":"4335224","Title":null,"Body":"<p>You should add parameter <code>parallel_iterations=1<\/code> to your while loop call.\nSuch as:<\/p>\n\n<pre><code>result, outputs_ta, states_ta = tf.while_loop(\n    condition, body, [time,outputs_ta,states_ta], parallel_iterations=1)\n<\/code><\/pre>\n\n<p>This is required because inside body you perform read and write operations on the same tensor array (<code>states_ta<\/code>). And in case of parallel loop execution(parallel_iterations > 1) some thread may try to read info from tensorArray, that was not written to it by another one.<\/p>\n\n<p>I've test your code snippet with parallel_iterations=1 on tensorflow 0.12.1 and it works as expected.<\/p>\n"}]}
{"QuestionId":42416890,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><computer-vision><conv-neural-network>","CreationDate":"2017-02-23T13:17:57.837","AcceptedAnswerId":null,"OwnerUserId":5119954.0,"Title":"CIFAR10 use only 2 feature instead 10 Tensorflow","Body":"<p>I would like to train a CNN in tensorflow with CIFAR10 dataset that contain 10 categories of images:<\/p>\n\n<pre><\/pre>\n\n<p>My question is : can i train my CNN with only 2 categories of this dataset? \nIf it's possible, i would like an explanation. \nThanks in advace<\/p>\n","answers":[]}
{"QuestionId":42418029,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2017-02-23T14:06:31.147","AcceptedAnswerId":42619727.0,"OwnerUserId":4274789.0,"Title":"Unable to use summary.merge in tensorboard for separate training and evaluation summaries","Body":"<p>I am trying to use tensorboard to watch the learning of a convolutional neural net. I am doing good with the tf.summary.merge_all function to create a merged summary. However, I would like to have tracking on accuracy and loss both for training and test data. This post is useful:<a href=\"https:\/\/stackoverflow.com\/questions\/34471563\/logging-training-and-validation-loss-in-tensorboard\">Logging training and validation loss in tensorboard<\/a>.<\/p>\n\n<p>To make things easier to handle, I would like to merge my summaries into two merged summaries, one for training and one for validation.(I will add more stuff eventually, like images weights etc.) I tried to follow the description from tensorboard <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/merge\" rel=\"noreferrer\">tf.summary.merge<\/a>. I can't make it work and I am unable to find any working examples to help me understand where I am going wrong. <\/p>\n\n<pre><\/pre>\n\n<p>This results in the following error:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): Could not parse one of the summary inputs\n           [[Node: Merge\/MergeSummary = MergeSummary[N=2, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](Merge\/MergeSummary\/inputs_0, Merge\/MergeSummary\/inputs_1)]]<\/p>\n<\/blockquote>\n\n<p>I would like to know why this doesn't work, and how I can find a solution, any working examples are appreciated.<\/p>\n","answers":[{"AnswerId":"42619727","CreationDate":"2017-03-06T07:21:23.583","ParentId":null,"OwnerUserId":"4274789","Title":null,"Body":"<p>I figured it out. I need to give the summaries names before merging. The code below solves the problem:<\/p>\n\n<pre><code>with tf.name_scope('Cost'):\n    cross_entropy = tf.reduce_mean(\n            tf.nn.softmax_cross_entropy_with_logits(logits=y_logits, labels=y))\n    opt = tf.train.AdamOptimizer(learning_rate=0.000003)\n    optimizer = opt.minimize(cross_entropy)\n    grads = opt.compute_gradients(cross_entropy, [b_fc_loc2])\n    cost_sum = tf.summary.scalar('val_cost', cross_entropy)\n    training_cost_sum = tf.summary.scalar('train_cost', cross_entropy)\n\n\nwith tf.name_scope('accuracy'):\n    correct_prediction = tf.equal(tf.argmax(y_logits, 1), tf.argmax(y, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))\n    train_accuracy = accuracy\n    accuracy_sum = tf.summary.scalar('val_accuracy', accuracy)\n    training_accuracy_sum = tf.summary.scalar('train_accuracy', accuracy)\n\n\nwith tf.Session() as sess:\n    writer = tf.summary.FileWriter('.\/logs\/{}\/{}'.format(session_name, run_num), sess.graph)\n    sess.run(tf.global_variables_initializer())\n    train_merged = tf.summary.merge([training_accuracy_sum, training_cost_sum])\n<\/code><\/pre>\n"}]}
{"QuestionId":42418322,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-23T14:18:20.157","AcceptedAnswerId":null,"OwnerUserId":4455881.0,"Title":"How to restore model(Graph and weights) and fine tune some variables?","Body":"<p>I want to <strong>pre-train<\/strong> a model on dataset_A with tensorflow, save the model with tf.train.Saver into checkpoint files. And then restore the model next time and <strong>fine tune some variables<\/strong> in the graph(i.e. not all the trainable_variables) with dataset_B.<\/p>\n\n<p>In the pre-training phase(pretrain.py), the variable is defined as follow:<\/p>\n\n<pre><\/pre>\n\n<p>In the fine-tuning phase(finetune.py), I have to get the variable for the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/AdamOptimizer\" rel=\"nofollow noreferrer\">optimizer<\/a>'s <strong>var_list<\/strong>. The codes bellow raise ValueError:<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>ValueError: Variable finetune\/full_connect_W:0 does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope?\n  enter code here<\/p>\n<\/blockquote>\n\n<p>Even if I remove the suffix \":0\" in the variable name, same error raised again.<\/p>\n\n<blockquote>\n  <p>ValueError: Variable finetune\/full_connect_W does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope?<\/p>\n<\/blockquote>\n\n<p>I managed to solve this problem in an ugly way. I restore the checkpoint file and find that <strong>finetune\/full_connect_W:0<\/strong> is the first item in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/trainable_variables\" rel=\"nofollow noreferrer\">tf.trainable_variables<\/a>. So I get the variable like this:<\/p>\n\n<pre><\/pre>\n\n<ul>\n<li>So the first question is: Is there any methods like <strong>get_variable<\/strong> or <strong>get_tensor_by_xxx<\/strong> to get the variable I want to fine tune?<\/li>\n<\/ul>\n\n<hr>\n\n<p>Another problem is, when I try to train the model with a new optimizer defined in the finetune phase (finetune.py), some error occurs. It seems the optimizer needs to be initialized. <\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>tensorflow.python.framework.errors.FailedPreconditionError: Attempting to use uninitialized value beta1_power_2<\/p>\n  \n  <p>[[Node: beta1_power_2\/read = IdentityT=DT_FLOAT, _class=[\"loc:@finetune\/full_connect_W\"], _device=\"\/job:localhost\/replica:0\/task:0\/gpu:0\"]]\n  Caused by op u'beta1_power_2\/read', defined at:<\/p>\n  \n  <p>File \"finetune_lstm_videotitle_test.py\", line 72, in \n      optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost, var_list=finetune_varlist)<\/p>\n<\/blockquote>\n\n<p>But if I initialize with tf.initialize_all_variables(), all the weights(variables) trained in the pre-trainning phrase with be reset.<\/p>\n\n<ul>\n<li>So the second question is: How can I fine tune specific variables and fixed other variables?<\/li>\n<\/ul>\n","answers":[]}
{"QuestionId":42418470,"AnswerCount":2,"Tags":"<android><xamarin><torch><flashlight><vision>","CreationDate":"2017-02-23T14:25:42.923","AcceptedAnswerId":null,"OwnerUserId":1792386.0,"Title":"How to turn on the torch\/flashlight with GooglePlay Services Vision API Xamarin Android","Body":"<p>I have been trying to implement the flashlight\/torch feature of the camera using the GooglePlay Services Vision API (using Nuget from Visual Studio) for the past few days without success. I have noticed that there is a GitHub implementation of this API which has such functionality but that is only available to Java users.\nI was wondering if there is anything related to C# Xamarin users.<\/p>\n\n<p>The Camera object is not made available on this API therefore I am not able to alter the Camera parameters needed to activate the flashlight.<\/p>\n\n<p>I would like to be sure if that functionality is not available so I don't waste more time over this. It just might be the case that the Xamarin developers have not attended to this functionality and they might in a near future.<\/p>\n\n<hr>\n\n<p>UPDATE<\/p>\n\n<p><a href=\"https:\/\/github.com\/googlesamples\/android-vision\/blob\/master\/visionSamples\/barcode-reader\/app\/src\/main\/java\/com\/google\/android\/gms\/samples\/vision\/barcodereader\/BarcodeCaptureActivity.java\" rel=\"nofollow noreferrer\">https:\/\/github.com\/googlesamples\/android-vision\/blob\/master\/visionSamples\/barcode-reader\/app\/src\/main\/java\/com\/google\/android\/gms\/samples\/vision\/barcodereader\/BarcodeCaptureActivity.java<\/a><\/p>\n\n<p>In there you can see that on line 214 we have such method call:<\/p>\n\n<pre><\/pre>\n\n<p>SetFlashMode is not a method of the CameraSource in Nuget, but it is on the GitHub (open source version).<\/p>\n","answers":[{"AnswerId":"42418576","CreationDate":"2017-02-23T14:30:40.103","ParentId":null,"OwnerUserId":"6625726","Title":null,"Body":"<p>Basically, anything you can do with Android can be done with Xamarin.Android. All the underlying APIs area available.<\/p>\n\n<p>Since you have existing Java code, you can create a binding project that enables you to call the code from your Xamarin.Android project. Here's a good article on how to get started: <a href=\"https:\/\/developer.xamarin.com\/guides\/android\/advanced_topics\/binding-a-java-library\/\" rel=\"nofollow noreferrer\">Binding a Java Library<\/a><\/p>\n\n<p>On the other hand, I don't think you need a library to do what you want to. If you only want torch\/flashlight functionality, you just need to adapt the Java code from <a href=\"https:\/\/stackoverflow.com\/a\/25097863\/6625726\">this answer<\/a> to work in Xamarin.Android with C#.<\/p>\n"},{"AnswerId":"45206306","CreationDate":"2017-07-20T05:58:45.583","ParentId":null,"OwnerUserId":"3993202","Title":null,"Body":"<p>Xamarin Vision Library Didn't expose the method to set Flash Mode.<\/p>\n\n<p><strong>WorkAround.<\/strong>\nUsing Reflection. You can get the Camera Object from CameraSouce and add the flash parameter then set the updated parameters to the camera.<\/p>\n\n<p><strong>This should be called after surfaceview has been created<\/strong><\/p>\n\n<p><strong>Code<\/strong><\/p>\n\n<pre><code>  public Camera getCameraObject (CameraSource _camSource)\n        {\n            Field [] cFields = _camSource.Class.GetDeclaredFields ();\n            Camera _cam = null;\n            try {\n                foreach (Field item in cFields) {\n                    if (item.Name.Equals (\"zzbNN\")) {\n                        Console.WriteLine (\"Camera\");\n                        item.Accessible = true;\n                        try {\n                            _cam = (Camera)item.Get (_camSource);\n                        } catch (Exception e) {\n                            Logger.LogException (this, e);\n                        }\n                    }\n                }\n            } catch (Exception e) {\n                Logger.LogException (this, e);\n            }\n            return _cam;\n        }\n\n        public void setFlash (bool isEnable)\n        {\n            try {\n                isTorch = !isEnable;\n                var _cam = getCameraObject (mCameraSource);\n                if (_cam == null) return;\n                var _pareMeters = _cam.GetParameters ();\n                var _listOfSuppo = _cam.GetParameters ().SupportedFlashModes;\n                _pareMeters.FlashMode = isTorch ? _listOfSuppo [0] : _listOfSuppo [3];\n                _cam.SetParameters (_pareMeters);\n            } catch (Exception e) {\n                Logger.LogException (this, e);\n            }\n        }\n<\/code><\/pre>\n"}]}
{"QuestionId":42419837,"AnswerCount":1,"Tags":"<python><matrix><tensorflow>","CreationDate":"2017-02-23T15:26:32.003","AcceptedAnswerId":42420093.0,"OwnerUserId":7611906.0,"Title":"Tensorflow tf.matmul example is incorrect?","Body":"<p>I read the official document for <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul\" rel=\"nofollow noreferrer\">tf.matmul<\/a> \nand I understand the first example.\nIt is a simple [2,3] x [3,2] operation:<\/p>\n\n<pre><\/pre>\n\n<p>However, the second example seems very strange :<\/p>\n\n<pre><\/pre>\n\n<p>Why the matrix with shape [2,2,3] is allowed to multiply with [2,3,2] ?<\/p>\n","answers":[{"AnswerId":"42420093","CreationDate":"2017-02-23T15:37:47.323","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>From the same page (<a href=\"https:\/\/web.archive.org\/web\/20170223153510\/https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul\" rel=\"nofollow noreferrer\">https:\/\/web.archive.org\/web\/20170223153510\/https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul<\/a>):<\/p>\n\n<blockquote>\n  <p>Returns:\n      A <code>Tensor<\/code> of the same type as <code>a<\/code> and <code>b<\/code> where each inner-most matrix is\n      the product of the corresponding matrices in <code>a<\/code> and <code>b<\/code>, e.g. if all\n      transpose or adjoint attributes are <code>False<\/code>:<BR\/>\n      <code>output<\/code>[..., i, j] = sum_k (<code>a<\/code>[..., i, k] * <code>b<\/code>[..., k, j]),\n      for all indices i, j.<\/p>\n<\/blockquote>\n\n<p>Hence a matrix with shape [2,2,3] is allowed to multiply with [2,3,2].<\/p>\n"}]}
{"QuestionId":42420219,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-23T15:43:49.557","AcceptedAnswerId":42421097.0,"OwnerUserId":3620605.0,"Title":"Sum dot products and constant in tensorflow","Body":"<p>I'm trying to do the following calculation in tensorflow<\/p>\n\n<pre><\/pre>\n\n<p>X's and W's have same shape, X*W is dot product.\nAfter playing around for a while I found that the following piece does the job<\/p>\n\n<pre><\/pre>\n\n<p>What is a concise way to perform this operation in tensorflow?<\/p>\n","answers":[{"AnswerId":"42421097","CreationDate":"2017-02-23T16:22:15.737","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>What about defining the dotproduct function?<\/p>\n\n<pre><code>def dotprod(X,W):\n  return tf.reduce_sum(tf.multiply(X,W))\nY = dotprod(X1,W1) + dotprod(X2,W2) + dotprod(X3,W3) + b\n<\/code><\/pre>\n\n<p>And I believe that you can use the + to add tensors. <\/p>\n\n<p>Let me know if this works!<\/p>\n"}]}
{"QuestionId":42420392,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-02-23T15:51:19.370","AcceptedAnswerId":null,"OwnerUserId":7545335.0,"Title":"how to implement L2 Regularization in caffe or DIGITS?","Body":"<p>I am using Caffe and also NVIDIA DIGITS. I want to use AlexNet pretrained on ImageNet and wanna fine tune it on my medical data. I have nearly 1000 images and using 80% for training, I generated 40,000 images by data augmentation (using cropping and rotation). However I face a severe overfitting. I tried to overcome this by adding multiple dropout layers. and the result change from :<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/BpbDM.png\" alt=\"results 1\"><\/p>\n\n<p>to:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/bIXXw.png\" alt=\"results 2\"><\/p>\n\n<p>but my accuracy does not improve.<\/p>\n\n<p>my network specifications:<\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p><\/p>\n\n<p>Now I want to add L2 regularization. I did not find such layer in Caffe and I should maybe make it myself. <\/p>\n\n<p> Do you have any solution for my problem? ( I have tried other ways like changing stepsize, changing learning rate from  to  and I found  is better, weigh decay changes, adding various dropout layer (which helped as you see))<\/p>\n\n<p> can you please help me how I can implement L2 regularization?? <\/p>\n","answers":[{"AnswerId":"42466370","CreationDate":"2017-02-26T08:31:55.857","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>You have <code>L2<\/code> regularization by default in caffe.<br>\nSee <a href=\"https:\/\/stackoverflow.com\/q\/32177764\/1714410\">this thread<\/a> for more information.<\/p>\n"}]}
{"QuestionId":42420709,"AnswerCount":0,"Tags":"<install><deep-learning><nvidia><pytorch>","CreationDate":"2017-02-23T16:04:54.793","AcceptedAnswerId":null,"OwnerUserId":5861307.0,"Title":"Can I install pytorch on nvidia jetson tx1?","Body":"<p>Because Tensorflow does not support ARM chips, I would like to install pytorch on Jetson TX1 to do some experiments. But I have not found any related resources about building pytorch into Jetson TX, is this possible?<\/p>\n","answers":[]}
{"QuestionId":42421380,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-23T16:34:54.510","AcceptedAnswerId":null,"OwnerUserId":2903666.0,"Title":"How to debug a segmentation fault 11 on TensorFlow?","Body":"<p>I installed cuda 8 and the new tensorflow 1.0. \nWhen I run \"import tensorflow as tf\" I get the following: <\/p>\n\n<pre><\/pre>\n\n<p>Knowing that nvcc -V gives the following: <\/p>\n\n<pre><\/pre>\n\n<p>Any idea how to fix this segmentation fault?<\/p>\n","answers":[{"AnswerId":"42588690","CreationDate":"2017-03-03T21:24:54.483","ParentId":null,"OwnerUserId":"954643","Title":null,"Body":"<p>You might be missing a library in your local cuda installation. E.g., <code>\/usr\/local\/cuda\/lib\/libcuda.dylib<\/code> was missing for me after trying to install <code>CUDA Toolkit 8.0<\/code> locally (possibly because I installed the drivers first before the toolkit, as this ancient thread suggests: <a href=\"https:\/\/render.otoy.com\/forum\/viewtopic.php?f=25&amp;t=1859\" rel=\"nofollow noreferrer\">https:\/\/render.otoy.com\/forum\/viewtopic.php?f=25&amp;t=1859<\/a>). Re-running the installer for just the driver installed it properly, and also symlinked it to another name (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3263#issuecomment-232184358\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3263#issuecomment-232184358<\/a>). <\/p>\n\n<p>Lastly, double check your environment variable paths, see if <code>echo $DYLD_LIBRARY_PATH<\/code> looks right.<\/p>\n\n<p>As an aside, I still saw some warnings when testing the install, e.g. <code>The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.<\/code>. These just are suggesting to build from source (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8037\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8037<\/a>), rather than using <code>pip install --upgrade tensorflow-gpu<\/code>. <\/p>\n"}]}
{"QuestionId":42422042,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-23T17:06:44.590","AcceptedAnswerId":42457923.0,"OwnerUserId":395857.0,"Title":"Forcing TensorFlow-GPU to use the CPU from the command line","Body":"<p>I have TensorFlow-GPU 1.0.0 installed on a server running Ubuntu 14.04.4 LTS x64.<\/p>\n\n<p>I know I can use <a href=\"https:\/\/stackoverflow.com\/a\/42382106\/395857\"><\/a> to hide one or several GPUs. Sometimes, I would like to hide all GPUs so that the TensorFlow-based program only uses the CPU. As a result, I tried<\/p>\n\n<pre><\/pre>\n\n<p>but this gives me the error message:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the  I use:<\/p>\n\n<pre><\/pre>\n\n<p>I know I could use  to prevent the TensorFlow-based program from using the GPU, but I wonder whether this can be achieved from the command line when launching the program (without changing the ). <\/p>\n\n<p>The option , according to the documentation, is supposed to have TensorFlow  automatically choose an existing and supported device to run the operations in case the specified one doesn't exist.<\/p>\n\n<p>My first reaction when I saw the message was that CUDA needs at least one GPU to successfully load but I <a href=\"https:\/\/stackoverflow.com\/a\/42287754\/395857\">read<\/a> that one can install GPU drivers and use TensorFlow-GPU on a machine even if the machine doesn't have GPU.<\/p>\n\n<hr>\n\n<p>Here is the   script I use for the test:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42457923","CreationDate":"2017-02-25T15:30:10.210","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>I think that \"E\" should've been a \"W\" or \"I\", that's just an\u00a0informational message and should not affect functioning of your program<\/p>\n"}]}
{"QuestionId":42422065,"AnswerCount":4,"Tags":"<linux><amazon-web-services><cuda><tensorflow><nvidia>","CreationDate":"2017-02-23T17:07:39.407","AcceptedAnswerId":null,"OwnerUserId":7479868.0,"Title":"How to install CUDA 8.0 in the latest version of Tensorflow (1.0) in AWS p2.xlarge instance, AMI ami-edb11e8d and nvidia drivers up to date (375.39)","Body":"<p>I have upgraded to Tensorflow version 1.0 and installed CUDA 8.0 with the cudnn 5.1 version and the nvidia drivers up to date 375.39. My NVIDIA hardware is the one that is on Amazon Web Services using the p2.xlarge instance, a Tesla K-80. My OS is Linux 64-bit.<\/p>\n\n<p>I get the next error message every time I use the command:  tf.Session()<\/p>\n\n<pre><\/pre>\n\n<p>I'm completely clueless about how to fix this. \nI have tried different versions of Nvidia drivers and CUDA but still it does not work.<\/p>\n\n<p>Any hints will be appreciated.<\/p>\n","answers":[{"AnswerId":"53794461","CreationDate":"2018-12-15T15:10:22.790","ParentId":null,"OwnerUserId":"7128719","Title":null,"Body":"<p>The AWS Deep Learning AMI has CUDA 8, 9 and 10 pre-installed and so you should not have to do this installation now. <\/p>\n\n<p>Reference: <a href=\"https:\/\/docs.aws.amazon.com\/dlami\/latest\/devguide\/overview-cuda.html\" rel=\"nofollow noreferrer\">https:\/\/docs.aws.amazon.com\/dlami\/latest\/devguide\/overview-cuda.html<\/a><\/p>\n"},{"AnswerId":"46991528","CreationDate":"2017-10-28T15:45:44.857","ParentId":null,"OwnerUserId":"5622761","Title":null,"Body":"<p>You can also try \"NVIDIA Volta Deep Learning AMI\" with p3 (v100 GPU) instance.<\/p>\n\n<p>Sign up on <a href=\"https:\/\/www.nvidia.com\/en-us\/gpu-cloud\/?ncid=van-gpu-cloud\" rel=\"nofollow noreferrer\">https:\/\/www.nvidia.com\/en-us\/gpu-cloud\/?ncid=van-gpu-cloud<\/a> and get your \"API Key\" to use the AMI free of charge.<\/p>\n\n<p>EC2\/GPU config info: <a href=\"https:\/\/aws.amazon.com\/blogs\/aws\/new-amazon-ec2-instances-with-up-to-8-nvidia-tesla-v100-gpus-p3\/\" rel=\"nofollow noreferrer\">https:\/\/aws.amazon.com\/blogs\/aws\/new-amazon-ec2-instances-with-up-to-8-nvidia-tesla-v100-gpus-p3\/<\/a><\/p>\n"},{"AnswerId":"42424313","CreationDate":"2017-02-23T19:18:00.790","ParentId":null,"OwnerUserId":"7601936","Title":null,"Body":"<p>Uninstall drivers &amp; cuda, then follow the <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-linux\/#axzz4ZXNjlCfN\" rel=\"nofollow noreferrer\">official guide<\/a> to reinstall. <\/p>\n\n<p>Run deviceQuery to check that the device is installed properly.<\/p>\n"},{"AnswerId":"42954674","CreationDate":"2017-03-22T14:34:05.350","ParentId":null,"OwnerUserId":"3868326","Title":null,"Body":"<p>You need to install a NVIDIA Driver and run the CUDA 8.0 installer.<\/p>\n\n<pre><code># Requirements\n# - NVIDIA Driver - NVIDIA-Linux-x86_64-375.39.run - http:\/\/www.nvidia.fr\/Download\/index.aspx\n# - CUDA runfile (local) - cuda_8.0.61_375.26_linux.run - https:\/\/developer.nvidia.com\/cuda-downloads\n# - cudnn-8.0-linux-x64-v5.0-ga.tgz\n\nsudo apt update -y &amp;&amp; sudo apt upgrade -y\nsudo apt install build-essential linux-image-extra-`uname -r` -y\n\nchmod +x NVIDIA-Linux-x86_64-375.39.run\nsudo .\/NVIDIA-Linux-x86_64-375.39.run\n\nchmod +x cuda_8.0.61_375.26_linux.run\n.\/cuda_8.0.61_375.26_linux.run --extract=`pwd`\/extracts\nsudo .\/extracts\/cuda-linux64-rel-8.0.61-21551265.run\n\necho -e \"export CUDA_HOME=\/usr\/local\/cuda\\nexport PATH=\\$PATH:\\$CUDA_HOME\/bin\\nexport LD_LIBRARY_PATH=\\$LD_LINKER_PATH:\\$CUDA_HOME\/lib64\" &gt;&gt; ~\/.bashrc\nsource .bashrc\n\ntar xf cudnn-8.0-linux-x64-v5.0-ga.tgz\ncd cuda\nsudo cp lib64\/* \/usr\/local\/cuda\/lib64\/\nsudo cp include\/cudnn.h \/usr\/local\/cuda\/include\/\n<\/code><\/pre>\n"}]}
{"QuestionId":42422594,"AnswerCount":1,"Tags":"<python><tensorflow><regression>","CreationDate":"2017-02-23T17:35:40.830","AcceptedAnswerId":42422818.0,"OwnerUserId":7508642.0,"Title":"Debugging Regression with TensorFlow Algorithm - Python","Body":"<p>Morning, I'm making some progress with TensorFlow, but it's still a struggle. I'm finding that a lot of the online examples don't run.<\/p>\n\n<p>Anyway, I've tried to write some code to apply a multi-layer neural network to solve a regression problem. However, I only get zeros out. Is anyone able to help me find where I'm going wrong in my code and my understanding?<\/p>\n\n<p>I'm running Tensorflow 0.12 with Python 3.5 in Windows 10<\/p>\n\n<p>Many thanks<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42422818","CreationDate":"2017-02-23T17:50:00.877","ParentId":null,"OwnerUserId":"1190882","Title":null,"Body":"<p>I am not sure by regression, did you mean linear regression or logistic regression. But looking into your code, I believe you mean linear regression.<\/p>\n\n<p>In case it is linear regression, your problem is in this line of code:<\/p>\n\n<p><code>prediction = tf.argmax(sess.run(pred, feed_dict={x:dataVar_scaled}),1).eval()<\/code><\/p>\n\n<p>You are taking the location of maximum (using <code>tf.argmax()<\/code>) value present in your code. Since <code>n_classes = 1<\/code>, there is only one class present in your output and it is always first which is why you are getting zero value.\nSo change that line to<\/p>\n\n<p><code>prediction = sess.run(pred, feed_dict={x:dataVar_scaled})<\/code>.<\/p>\n\n<p>Now you will get the value of the output and not the location of maximum value.<\/p>\n\n<p>In case it is logistic regression, you are clearly having only one class. How can you do classification when you have outputs only as one class.<\/p>\n\n<p>Hope this helps you.<\/p>\n"}]}
{"QuestionId":42422646,"AnswerCount":2,"Tags":"<tensorflow><machine-learning><keras><generative-adversarial-network>","CreationDate":"2017-02-23T17:39:08.783","AcceptedAnswerId":42470757.0,"OwnerUserId":2381889.0,"Title":"Keras train partial model issue (about GAN model)","Body":"<p>I came across a strange issue when using keras to implement GAN model.<\/p>\n\n<p>With GAN we need to build up G and D first, and then add a new Sequential model (GAN) and add(G), add(D) sequentially afterwards.<\/p>\n\n<p>Keras seems to backprop back to G (via GAN model) when I do , and I got an .<\/p>\n\n<p>If I remove the  (the last stacked G then D sequential model), it computes  correctly.<\/p>\n\n<p>My environment is:<\/p>\n\n<ul>\n<li>Ubuntu 16.04<\/li>\n<li>keras 1.2.2<\/li>\n<li>tensorflow-gpu 1.0.0<\/li>\n<li>keras config: <\/li>\n<\/ul>\n\n<p>I know that quite many people have succeeded in implementing GAN with keras, so I am wondering where I got wrong.<\/p>\n\n<pre><\/pre>\n\n<p>Error Message:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42470757","CreationDate":"2017-02-26T16:00:28.137","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>First, I would advise you to switch to the Functional API models. These kinds of mixed models are more easily handled by Functional models.<\/p>\n\n<p>I have no idea why your solution didn't work to be honnest, it seems like when you link the D model to a new input, it gets kind of \"corrupted\" and gets linked to it.\nThe way I have found around that problem, is to define the layers and use them for both the Discriminator and the GAN models. Here is the code :<\/p>\n\n<pre><code>import numpy as np\nfrom keras.layers import *\nimport keras.models as km\nimport keras.optimizers as ko\nfrom keras.datasets import mnist\n\nbatch_size = 16\nlr = 0.0001\n\ndef noise_gen(batch_size, z_dim):\n    noise = np.zeros((batch_size, z_dim), dtype=np.float32)\n    for i in range(batch_size):\n        noise[i, :] = np.random.uniform(-1, 1, z_dim)\n    return noise\n\n# Changes the traiable argument for all the layers of model\n# to the boolean argument \"trainable\"\ndef make_trainable(model, trainable):\n    model.trainable = trainable\n    for l in model.layers:\n        l.trainable = trainable\n\n# --------------------Generator Model--------------------\n\ng_input = Input(shape=(100,))\n\ng_hidden = Dense(1024, activation='relu')(g_input)\ng_hidden = Dense(7*7*128, activation='relu')(g_hidden)\ng_hidden = BatchNormalization()(g_hidden)\ng_hidden = Reshape((7,7,128))(g_hidden)\n\ng_hidden = Deconvolution2D(64,5,5, (None, 14, 14, 64), subsample=(2,2),\n        border_mode='same', activation='relu')(g_hidden)\ng_hidden = BatchNormalization()(g_hidden)\ng_output = Deconvolution2D(1,5,5, (None, 28, 28, 1), subsample=(2,2),\n        border_mode='same')(g_hidden)\n\nG = km.Model(input=g_input,output=g_output)\nG.compile(loss='binary_crossentropy', optimizer=ko.SGD(lr=lr, momentum=0.9, nesterov=True))\nG.summary()\n\n# --------------------Discriminator Model--------------------\n\nd_input = Input(shape=(28,28,1))\n\nd_l1 = Convolution2D(64,5,5, subsample=(2,2))\nd_hidden_1 = d_l1(d_input)\nd_l2 = LeakyReLU(alpha=0.2)\nd_hidden_2 = d_l2(d_hidden_1)\n\nd_l3 = Convolution2D(128,5,5, subsample=(2,2))\nd_hidden_3 = d_l3(d_hidden_2)\nd_l4 = BatchNormalization()\nd_hidden_4 = d_l4(d_hidden_3)\nd_l5 = LeakyReLU(alpha=0.2)\nd_hidden_5 = d_l5(d_hidden_4)\n\nd_l6 = Flatten()\nd_hidden_6 = d_l6(d_hidden_5)\nd_l7 = Dense(1, activation='sigmoid')\nd_output = d_l7(d_hidden_6)\n\nD = km.Model(input=d_input,output=d_output)\nD.compile(loss='binary_crossentropy',optimizer=ko.SGD(lr=lr,momentum=0.9, nesterov=True))\nD.summary()\n\n# --------------------GAN Model--------------------\nmake_trainable(D,False)\n\ngan_input = Input(shape=(100,))\ngan_hidden = G(gan_input)\ngan_hidden = d_l1(gan_hidden)\ngan_hidden = d_l2(gan_hidden)\ngan_hidden = d_l3(gan_hidden)\ngan_hidden = d_l4(gan_hidden)\ngan_hidden = d_l5(gan_hidden)\ngan_hidden = d_l6(gan_hidden)\ngan_output = d_l7(gan_hidden)\n\nGAN = km.Model(input=gan_input,output=gan_output)\nGAN.compile(loss='binary_crossentropy',optimizer=ko.SGD(lr=lr, momentum=0.9, nesterov=True))\nGAN.summary()\n\n# --------------------Main Code--------------------\n(X, _), _ = mnist.load_data()\nX = X \/ 255.\nX = X[:, :, :, np.newaxis]\n\nX_batch = X[0:batch_size, :]\nZ1_batch = noise_gen(batch_size, 100)\nZ2_batch = noise_gen(batch_size, 100)\n\nprint(type(X_batch),X_batch.shape)\nprint(type(Z1_batch),Z1_batch.shape)\n\nfake_batch = G.predict(Z1_batch)\nreal_batch = X_batch\nprint('--------------------Fake Image Generated!--------------------')\n\ncombined_X_batch = np.concatenate((real_batch, fake_batch))\ncombined_y_batch = np.concatenate((np.ones((batch_size, 1)), np.zeros((batch_size, 1))))\nprint('real_batch={}, fake_batch={}'.format(real_batch.shape, fake_batch.shape))\nprint(type(combined_X_batch),combined_X_batch.dtype,combined_X_batch.shape)\nprint(type(combined_y_batch),combined_y_batch.dtype,combined_y_batch.shape)\nmake_trainable(D,True)\nd_loss = D.train_on_batch(combined_X_batch, combined_y_batch)\nprint('--------------------Discriminator trained!--------------------')\nprint(d_loss)\n\nmake_trainable(D,False)\ng_loss = GAN.train_on_batch(Z2_batch, np.ones((batch_size, 1)))\nprint('--------------------GAN trained!--------------------')\nprint(g_loss)\n<\/code><\/pre>\n\n<p>Does that help? <\/p>\n"},{"AnswerId":"42501573","CreationDate":"2017-02-28T06:19:03.083","ParentId":null,"OwnerUserId":"2381889","Title":null,"Body":"<p>After strived for quite a long time, I finally get it that it's the Discriminator's BatchNormalization layer that caused the problem.<\/p>\n\n<p>If you just comment out the <code>model.add(kl.BatchNormalization())<\/code> in the Discriminator. It'll work fine.<\/p>\n\n<p>However, as @NassimBen shown, the functional API does not cause any problems.<\/p>\n\n<pre><code>import numpy as np\nimport keras.layers as kl\nimport keras.models as km\nimport keras.optimizers as ko\nfrom keras.datasets import mnist\n\nbatch_size = 16\nlr = 0.0001\n\ndef noise_gen(batch_size, z_dim):\n    noise = np.zeros((batch_size, z_dim), dtype=np.float32)\n    for i in range(batch_size):\n        noise[i, :] = np.random.uniform(-1, 1, z_dim)\n    return noise\n\n# --------------------Generator Model--------------------\n\nmodel = km.Sequential()\n\nmodel.add(kl.Dense(input_dim=100, output_dim=1024))\nmodel.add(kl.Activation('relu'))\n\nmodel.add(kl.Dense(7*7*128))\nmodel.add(kl.BatchNormalization())\nmodel.add(kl.Activation('relu'))\nmodel.add(kl.Reshape((7, 7, 128), input_shape=(7*7*128,)))\n\nmodel.add(kl.Deconvolution2D(64, 5, 5, (None, 14, 14, 64), subsample=(2, 2),\n                             input_shape=(7, 7, 128), border_mode='same'))\nmodel.add(kl.BatchNormalization())\nmodel.add(kl.Activation('relu'))\n\nmodel.add(kl.Deconvolution2D(1, 5, 5, (None, 28, 28, 1), subsample=(2, 2),\n                             input_shape=(14, 14, 64), border_mode='same'))\n\nG = model\nG.compile(  loss='binary_crossentropy', optimizer=ko.SGD(lr=lr, momentum=0.9, nesterov=True))\n\n# --------------------Discriminator Model--------------------\n\nmodel = km.Sequential()\n\nmodel.add(kl.Convolution2D( 64, 5, 5, subsample=(2, 2), input_shape=(28, 28, 1)))\nmodel.add(kl.LeakyReLU(alpha=0.2))\n\nmodel.add(kl.Convolution2D(128, 5, 5, subsample=(2, 2)))\n# model.add(kl.BatchNormalization())\nmodel.add(kl.LeakyReLU(alpha=0.2))\n\nmodel.add(kl.Flatten())\nmodel.add(kl.Dense(1))\nmodel.add(kl.Activation('sigmoid'))\n\nD = model\nD.compile(  loss='binary_crossentropy', optimizer=ko.SGD(lr=lr, momentum=0.9, nesterov=True))\n\n# --------------------GAN Model--------------------\n\nmodel = km.Sequential()\nmodel.add(G)\nD.trainable = False  # Is this necessary?\nmodel.add(D)\nGAN = model\nGAN.compile(loss='binary_crossentropy', optimizer=ko.SGD(lr=lr, momentum=0.9, nesterov=True))\n\n# --------------------Main Code--------------------\n(X, _), _ = mnist.load_data()\nX = X \/ 255.\nX = X[:, :, :, np.newaxis]\n\nX_batch = X[0:batch_size, :]\nZ1_batch = noise_gen(batch_size, 100)\nZ2_batch = noise_gen(batch_size, 100)\n\nfake_batch = G.predict(Z1_batch)\nreal_batch = X_batch\nprint('--------------------Fake Image Generated!--------------------')\n\ncombined_X_batch = np.concatenate((real_batch, fake_batch))\ncombined_y_batch = np.concatenate((np.ones((batch_size, 1)), np.zeros((batch_size, 1))))\nprint('real_batch={}, fake_batch={}'.format(real_batch.shape, fake_batch.shape))\n\nD.trainable = True\nd_loss = D.train_on_batch(combined_X_batch, combined_y_batch)\nprint('--------------------Discriminator trained!--------------------')\nprint(d_loss)\n\nD.trainable = False\ng_loss = GAN.train_on_batch(Z2_batch, np.ones((batch_size, 1)))\nprint('--------------------GAN trained!--------------------')\nprint(g_loss)\n<\/code><\/pre>\n"}]}
{"QuestionId":42423589,"AnswerCount":1,"Tags":"<c++><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2017-02-23T18:34:54.370","AcceptedAnswerId":42431495.0,"OwnerUserId":214488.0,"Title":"How to feed in and retrieve state of LSTM in tensorflow C\/ C++","Body":"<p>I'd like to build and train a multi-layer LSTM model (stateIsTuple=True) in python, and then load and use it in C++. But I'm having a hard time figuring out how to feed and fetch states in C++, mainly because I don't have string names which I can reference. <\/p>\n\n<p>E.g. I put the initial state in a named scope such as <\/p>\n\n<pre><\/pre>\n\n<p>and this appears in the graph as below, but how can I feed to these in C++?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/On4a9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/On4a9.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Also, how can I fetch the current state in C++? I tried the graph construction code below in python but I'm not sure if it's the right thing to do, because last_state should be a tuple of tensors, not a single tensor (though I can see that the last_state node in tensorboard is 2x2x50x128, which sounds like it just concatenated the states as I have 2 layers, 128 rnn size, 50 mini batch size, and lstm cell - with 2 state vectors).<\/p>\n\n<pre><\/pre>\n\n<p>and this is what it looks like in tensorboard<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lBXJT.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lBXJT.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Should I concat and split the state tensors so there is only ever one state tensor going in and out? Or is there a better way?<\/p>\n\n<p>P.S. Ideally the solution won't involve hard-coding the number of layers (or rnn size). So I can just have four strings input_node_name, output_node_name, input_state_name, output_state_name, and the rest is derived from there.<\/p>\n","answers":[{"AnswerId":"42431495","CreationDate":"2017-02-24T05:36:03.723","ParentId":null,"OwnerUserId":"214488","Title":null,"Body":"<p>I managed to do this by manually concatenating the state into a single tensor. I'm not sure if this is wise, since this is how tensorflow <em>used<\/em> to handle states, but is now deprecating that and switching to tuple states. Instead of setting state_is_tuple=False and risking my code being obsolete soon, I've added extra ops to manually stack and unstack the states to and from a single tensor. Saying that, it works fine both in python and C++.<\/p>\n\n<p>The key code is:<\/p>\n\n<pre><code># setting up\nzero_state = cell.zero_state(batch_size, tf.float32)\nstate_in = tf.identity(zero_state, name='state_in')         \n\n# based on https:\/\/medium.com\/@erikhallstrm\/using-the-tensorflow-multilayered-lstm-api-f6e7da7bbe40#.zhg4zwteg\nstate_per_layer_list = tf.unstack(state_in, axis=0)\nstate_in_tuple = tuple(\n    # TODO make this not hard-coded to LSTM\n    [tf.contrib.rnn.LSTMStateTuple(state_per_layer_list[idx][0], state_per_layer_list[idx][1])\n    for idx in range(num_layers)]\n)\n\noutputs, state_out_tuple = legacy_seq2seq.rnn_decoder(inputs, state_in_tuple, cell, loop_function=loop if infer else None)\nstate_out = tf.identity(state_out_tuple, name='state_out')\n\n# running (training or inference)\nstate = sess.run('state_in:0') # zero state\n\nloop:\n    feed = {'data_in:0': x, 'state_in:0': state}\n    [y, state] = sess.run(['data_out:0', 'state_out:0'], feed)\n<\/code><\/pre>\n\n<p>Here is the full code if anyone needs it\n<a href=\"https:\/\/github.com\/memo\/char-rnn-tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/memo\/char-rnn-tensorflow<\/a><\/p>\n"}]}
{"QuestionId":42424335,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-23T19:19:19.030","AcceptedAnswerId":42426372.0,"OwnerUserId":395857.0,"Title":"Why did a TensorFlow session fail to launch due to out of memory error on GPU despite specifying `device_count={'CPU': 1, 'GPU': 0}`?","Body":"<p>I tried running on a server whose GPUs' vRAM were >96% used:<\/p>\n\n<pre><\/pre>\n\n<p>It gave me  error that stopped the execution of the program:<\/p>\n\n<pre><\/pre>\n\n<p>Why can the level of vRAM usage interfer with my program given that I specified   when creating the TensorFlow session?<\/p>\n","answers":[{"AnswerId":"42426372","CreationDate":"2017-02-23T21:18:37.143","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>I'm not sure if <code>device_count={'GPU': 0}<\/code> works to prevent GPU memory allocation, I've not seen it used before. There's a chance it doesn't work because GPU allocator is a process-level concept since it's shared between sessions. So you are trying to configure a process-level setting through a session-level config. The most sure way is to make GPU invisible on a process level by setting env var --  <code>export CUDA_VISIBLE_DEVICES=<\/code><\/p>\n"}]}
{"QuestionId":42425262,"AnswerCount":1,"Tags":"<tensorflow><bazel>","CreationDate":"2017-02-23T20:11:12.390","AcceptedAnswerId":null,"OwnerUserId":7466373.0,"Title":"tensorflow install from sources bazel error","Body":"<p>As I am having an error \"AttributeError: 'module' object has no attribute 'GRUCell'\" when tried to run transate.py i removed tensorflow from my system and started installing it again from sources. I installed bazel and configured the installation. After \"Build the pip package\", I am getting the following error.<\/p>\n\n<p>Can anyone please help me in resolving these errors<\/p>\n\n<p>Thanks in advance\n<a href=\"https:\/\/i.stack.imgur.com\/TKKMv.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/TKKMv.png\" alt=\"enter image description here\"><\/a>  <\/p>\n","answers":[{"AnswerId":"43070599","CreationDate":"2017-03-28T13:13:48.030","ParentId":null,"OwnerUserId":"7778502","Title":null,"Body":"<p>The error says \"cannot stat 'bazel-bin\/...\/build_pip_package.runfiles\/tensorflow'\" meaning your runfiles tree doesn't exist or is unreadable. This indicates something went wrong during the build, particularly the runfiles tree creation step. The \"runfiles tree\" is a directory tree that contains the files that the binary you built needs at runtime, and the root of it is the \"bazel-bin\/...\/build_pip_package.runfiles\" directory.<\/p>\n\n<p>This suspicion of mine is also backed up by the hanging runfiles tree creation actions in your output (on the top of the image), which is supposed to create this directory structure, and should <em>definitely<\/em> not take 11 hours, it should complete within a few seconds to a few minutes. (If any action hangs for more than a few minutes that's already a sign something's likely wrong.)<\/p>\n\n<p>Some questions:<\/p>\n\n<ul>\n<li>If you rerun the build, does it hang? In other words, is this problem deterministic or did it occur just once?<\/li>\n<li>What if you run the build like <code>bazel --output_user_root=\/usr\/local\/foo build &lt;whatever&gt;<\/code>? You can use any other path that's on the local filesystem. What i'm poking at is, whether your output root was on NFS or a filesystem that got unmounted mid-build.<\/li>\n<\/ul>\n"}]}
{"QuestionId":42425750,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-23T20:38:47.983","AcceptedAnswerId":42452314.0,"OwnerUserId":7613421.0,"Title":"Shape must have rank 2 but it is rank 1","Body":"<p>I am trying to make a chatbot using tensorflow in python. But I am getting this error when I am training the dataset.<\/p>\n\n<p><strong>raise ValueError(err.message)\nValueError: Shape must be rank 2 but is rank 1 for 'model_with_buckets\/sequence_\nloss\/sequence_loss_by_example\/sampled_softmax_loss\/LogUniformCandidateSampler' (\nop: 'LogUniformCandidateSampler') with input shapes: [?].<\/strong><\/p>\n\n<p>Any help appreciated.\nThanks<\/p>\n","answers":[{"AnswerId":"42452314","CreationDate":"2017-02-25T05:26:32.213","ParentId":null,"OwnerUserId":"3377385","Title":null,"Body":"<p>I was getting similar error when i tried to run tensorflow \"seq2seq tutorial code\" compatible with the older version on tensorflow 1.0, it mainly due to change in the order of arguments to a function(<strong>tf.nn.sampled_softmax<\/strong>). It is expecting <strong>inputs(rank 2)<\/strong> but you are passing l<strong>abels(rank 1)<\/strong> So change the order of arguments according to definition of the function.<\/p>\n\n<p><strong>In tf 0.12.0:<\/strong>\nDefinition : tf.nn.sampled_softmax_loss(weights, biases, <strong>inputs, labels,<\/strong> num_sampled, num_classes, num_true=1, sampled_values=None, remove_accidental_hits=True, partition_strategy=\"mod\", name=\"sampled_softmax_loss\")<\/p>\n\n<p><strong>In tf 1.0:<\/strong>\ntf.nn.sampled_softmax_loss(weights, biases, <strong>labels, inputs,<\/strong> num_sampled, num_classes, num_true=1, sampled_values=None, remove_accidental_hits=True, partition_strategy='mod', name='sampled_softmax_loss')<\/p>\n"}]}
{"QuestionId":42425858,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2017-02-23T20:45:13.420","AcceptedAnswerId":null,"OwnerUserId":419780.0,"Title":"TensorBoard Distributions and Histograms with Keras and fit_generator","Body":"<p>I'm using Keras to train a CNN using the fit_generator function.<\/p>\n\n<p>It seems to be a <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/3358\" rel=\"noreferrer\">known issue<\/a> that TensorBoard doesn't show histograms and distributions in this setup.<\/p>\n\n<p>Did anybody figure out a way to make it work anyway?<\/p>\n","answers":[{"AnswerId":"42477664","CreationDate":"2017-02-27T04:06:40.190","ParentId":null,"OwnerUserId":"4259486","Title":null,"Body":"<p>There is no easy way to just plug it in with one line of code, you have to write your summaries by hand.<\/p>\n\n<p>Good news is that it's not difficult and you can use the <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks\/tensorboard_v1.py\" rel=\"nofollow noreferrer\">TensorBoard callback code<\/a> in Keras as a reference. \n(There is also a <a href=\"https:\/\/github.com\/keras-team\/keras\/blob\/master\/keras\/callbacks\/tensorboard_v2.py\" rel=\"nofollow noreferrer\">version 2<\/a> in preparation.)<\/p>\n\n<p>Basically, write a function e.g. <code>write_summaries(model)<\/code> and call it whenever you want to write your summaries (e.g. just after your <code>fit_generator()<\/code>)<\/p>\n\n<p>Inside your <code>write_summaries(model)<\/code> function use <code>tf.summary<\/code>, <code>histogram_summary<\/code> and other summary functions to log data you want to see on tensorboard.<\/p>\n\n<p>If you don't know exactly how, check <a href=\"https:\/\/www.tensorflow.org\/get_started\/summaries_and_tensorboard\" rel=\"nofollow noreferrer\">official tutorial<\/a>: \nand this great example of <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_with_summaries.py\" rel=\"nofollow noreferrer\">MNIST with summaries<\/a>.<\/p>\n"},{"AnswerId":"52674174","CreationDate":"2018-10-05T22:51:19.510","ParentId":null,"OwnerUserId":"177449","Title":null,"Body":"<p>I believe bartgras's explanation is superseded in more recent versions of Keras (I'm using Keras 2.2.2).  To get histograms in Tensorboard all I did was the following, (where <code>bg<\/code> is a data wrangling class which exposes a generator for <code>gb.training_batch()<\/code>; <code>gb.validation_batch()<\/code> however is NOT a generator):<\/p>\n\n<pre><code>NAME = \"Foo_{}\".format(datetime.now().isoformat(timespec='seconds')).replace(':', '-')\n\ntensorboard = keras.callbacks.TensorBoard(\n    log_dir=\"logs\/{}\".format(NAME),\n    histogram_freq=1,\n    write_images=True)\n\ncallbacks = [\n    tensorboard\n]\n\nhistory = model.fit_generator(\n    bg.training_batch(),\n    validation_data=bg.validation_batch(),\n    epochs=EPOCHS,\n    steps_per_epoch=bg.steps_per_epoch,\n    validation_steps=bg.validation_steps,\n    verbose=1,\n    shuffle=False,\n    callbacks=callbacks)\n<\/code><\/pre>\n"}]}
{"QuestionId":42426577,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-23T21:32:40.553","AcceptedAnswerId":42426767.0,"OwnerUserId":6926393.0,"Title":"Why are there different return types in TensorFlow for the same operation?","Body":"<p>In TensorFlow, I have two constant nodes:<\/p>\n\n<pre><\/pre>\n\n<p>When I run this graph in Jupyter Notebook, the return types are different for following two session runs:<\/p>\n\n<pre><\/pre>\n\n<p>For this operation, there is no output cell:<\/p>\n\n<pre><\/pre>\n\n<p>But when I run the session as:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone please explain the return types of these?\n<a href=\"https:\/\/i.stack.imgur.com\/bXazR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bXazR.png\" alt=\"Jupyter\"><\/a><\/p>\n","answers":[{"AnswerId":"42426767","CreationDate":"2017-02-23T21:45:18.663","ParentId":null,"OwnerUserId":"1405065","Title":null,"Body":"<p>Your second line has two <code>print<\/code> statements separated by commas. The <code>print<\/code> function returns <code>None<\/code>, and a single <code>None<\/code> is ignored by the Python REPL (read-eval-print loop). However, the comma creates a tuple, which is not ignored even when it contains only <code>None<\/code> values.<\/p>\n\n<p>If you want to run multiple statements on a single line, use <code>;<\/code> instead of <code>,<\/code>:<\/p>\n\n<pre><code>print(\"foo\"); print(\"bar\")\n<\/code><\/pre>\n"},{"AnswerId":"42426762","CreationDate":"2017-02-23T21:44:37.293","ParentId":null,"OwnerUserId":"366553","Title":null,"Body":"<h2>You're creating a tuple<\/h2>\n\n<p>print is a function just like any other. It doesn't return a value, so it returns None, which is not generally shown.<\/p>\n\n<p>However, the comma operator creates a tuple, and jupyter shows the returned value.<\/p>\n\n<p>Writing 4,4 would result in (4,4).<\/p>\n\n<p>Writing None,None would result in what you're seeing.<\/p>\n\n<p>If you simply want to execute two operations, put them each in their own line - the comma is not for that.<\/p>\n"}]}
{"QuestionId":42426960,"AnswerCount":4,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-02-23T21:57:13.193","AcceptedAnswerId":42647980.0,"OwnerUserId":1601580.0,"Title":"How does one train multiple models in a single script in TensorFlow when there are GPUs present?","Body":"<p>Say I have access to a number of GPUs in a single machine (for the sake of argument assume 8GPUs each with max memory of 8GB each in one single machine with some amount of RAM and disk). I wanted to run in <strong>one single script<\/strong> and in one single machine a program that evaluates multiple models (say 50 or 200) in TensorFlow, each with a different hyper parameter setting (say, step-size, decay rate, batch size, epochs\/iterations, etc). At the end of training assume we just record its accuracy and get rid of the model (if you want assume the model is being check pointed every so often, so its fine to just throw away the model and start training from scratch. You may also assume some other data may be recorded like the specific hyper params, train, validation, train errors are recorded as we train etc).<\/p>\n\n<p>Currently I have a (pseudo-)script that looks as follow:<\/p>\n\n<pre><\/pre>\n\n<p>essentially it tries lots of models in one single run but it builds each model in a separate graph and runs each one in a separate session.<\/p>\n\n<p>I guess my main worry is that its unclear to me how tensorflow under the hood allocates resources for the GPUs to be used. For example, does it load the (part of the) data set only when a session is ran? When I create a graph and a model, is it brought in the GPU immediately or when is it inserted in the GPU? Do I need to clear\/free the GPU each time it tries a new model? I don't actually care too much if the models are ran in parallel in multiple GPU (which can be a nice addition), but I want it to first run everything serially without crashing. Is there anything special I need to do for this to work?<\/p>\n\n<hr>\n\n<p>Currently I am getting an error that starts as follow:<\/p>\n\n<pre><\/pre>\n\n<p>and further down the line it says:<\/p>\n\n<pre><\/pre>\n\n<p>however further down the output file (where it prints) it seems to print fine the errors\/messages that should show as training proceeds. Does this mean that it didn't run out of resources? Or was it actually able to use the GPU? If it was able to use the CPU instead of the CPU, when why is this an error only happening when GPU are about to be used?<\/p>\n\n<p>The weird thing is that the data set is really not that big (all 60K points are 24.5M) and when I run a single model locally in my own computer it seems that the process uses less than 5GB. The GPUs have at least 8GB and the computer with them has plenty of RAM and disk (at least 16GB). Thus, the errors that tensorflow is throwing at me are quite puzzling. What is it trying to do and why are they occurring? Any ideas?<\/p>\n\n<hr>\n\n<p>After reading the answer that suggests to use the multiprocessing library I came up with the following script:<\/p>\n\n<pre><\/pre>\n\n<p>honestly I am not sure why his answer suggests to use pool, or why there are weird tuple brackets but this is what would make sense for me. Would the resources for tensorflow be re-allocated every time a new process is created in the above loop?<\/p>\n","answers":[{"AnswerId":"42647980","CreationDate":"2017-03-07T12:11:36.953","ParentId":null,"OwnerUserId":"2476373","Title":null,"Body":"<p>I think that running all models in one single script can be bad practice in the long term (see my suggestion below for a better alternative). However, if you would like to do it, here is a solution: You can encapsulate your TF session into a process with the <code>multiprocessing<\/code> module, this will make sure TF releases the session memory once the process is done. Here is a code snippet:<\/p>\n\n<pre><code>from multiprocessing import Pool\nimport contextlib\ndef my_model((param1, param2, param3)): # Note the extra (), required by the pool syntax\n    &lt; your code &gt;\n\nnum_pool_worker=1 # can be bigger than 1, to enable parallel execution \nwith contextlib.closing(Pool(num_pool_workers)) as po: # This ensures that the processes get closed once they are done\n     pool_results = po.map_async(my_model,\n                                    ((param1, param2, param3)\n                                     for param1, param2, param3 in params_list))\n     results_list = pool_results.get()\n<\/code><\/pre>\n\n<p>Note from OP: The random number generator seed does not reset automatically with the multi-processing library if you choose to use it. Details here: <a href=\"https:\/\/stackoverflow.com\/questions\/9209078\/using-python-multiprocessing-with-different-random-seed-for-each-process\">Using python multiprocessing with different random seed for each process<\/a><\/p>\n\n<p>About TF resource allocation: Usually TF allocates much more resources than it needs. Many times you can restrict each process to use a fraction of the total GPU memory, and discover through trial and error the fraction your script requires.<\/p>\n\n<p>You can do it with the following snippet<\/p>\n\n<pre><code>gpu_memory_fraction = 0.3 # Choose this number through trial and error\ngpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=gpu_memory_fraction,)\nsession_config = tf.ConfigProto(gpu_options=gpu_options)\nsess = tf.Session(config=session_config, graph=graph)\n<\/code><\/pre>\n\n<p>Note that sometimes TF increases the memory usage in order to accelerate the execution. Therefore, reducing the memory usage might make your model run slower.<\/p>\n\n<p><strong>Answers to the new questions in your edit\/comments:<\/strong><\/p>\n\n<ol>\n<li><p>Yes, Tensorflow will be re-allocated every time a new process is created, and cleared once a process ends.<\/p><\/li>\n<li><p>The for-loop in your edit should also do the job. I suggest to use Pool instead, because it will enable you to run several models concurrently on a single GPU. See my notes about setting <code>gpu_memory_fraction<\/code> and \"choosing the maximal number of processes\". Also note that: (1) The Pool map runs the loop for you, so you don't need an outer for-loop once you use it. (2) In your example, you should have something like <code>mdl=get_model(args)<\/code> before calling train()<\/p><\/li>\n<li><p>Weird tuple parenthesis: Pool only accepts a single argument, therefore we use a tuple to pass multiple arguments. See <a href=\"https:\/\/stackoverflow.com\/questions\/8521883\/multiprocessing-pool-map-and-function-with-two-arguments\/8523763#8523763\">multiprocessing.pool.map and function with two arguments<\/a> for more details. As suggested in one answer, you can make it more readable with<\/p>\n\n<pre><code>def train_mdl(params):\n    (x,y)=params\n    &lt; your code &gt;\n<\/code><\/pre><\/li>\n<li><p>As @Seven suggested, you can use CUDA_VISIBLE_DEVICES environment variable to choose which GPU to use for your process. You can do it from within your python script using the following on the beginning of the process function (<code>train_mdl<\/code>).<\/p>\n\n<pre><code>import os # the import can be on the top of the python script\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"{}\".format(gpu_id)\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p><strong>A better practice for executing your experiments<\/strong> would be to isolate your training\/evaluation code from the hyper parameters\/ model search code.\nE.g. have a script named <code>train.py<\/code>, which accepts a specific combination of hyper parameters and references to your data as arguments, and executes training for a single model.<\/p>\n\n<p>Then, to iterate through the all the possible combinations of parameters you can use a simple task (jobs) queue, and submit all the possible combinations of hyper-parametrs as separate jobs. The task queue will feed your jobs one at a time to your machine. Usually, you can also set the queue to execute number of processes concurrently (see details below).<\/p>\n\n<p>Specifically, I use <a href=\"http:\/\/vicerveza.homeunix.net\/~viric\/soft\/ts\/\" rel=\"noreferrer\">task spooler<\/a>, which is super easy to install and handful (doesn't requires admin privileges, details below).<\/p>\n\n<p>Basic usage is (see notes below about task spooler usage):<\/p>\n\n<pre><code>ts &lt;your-command&gt;\n<\/code><\/pre>\n\n<p>In practice, I have a separate python script that manages my experiments, set all the arguments per specific experiment and send the jobs to the <code>ts<\/code> queue.<\/p>\n\n<p><strong>Here are some relevant snippets of python code from my experiments manager:<\/strong><\/p>\n\n<p><code>run_bash<\/code> executes a bash command<\/p>\n\n<pre><code>def run_bash(cmd):\n    p = subprocess.Popen(cmd, shell=True, stdout=subprocess.PIPE, executable='\/bin\/bash')\n    out = p.stdout.read().strip()\n    return out  # This is the stdout from the shell command\n<\/code><\/pre>\n\n<p>The next snippet sets the number of concurrent processes to be run (see note below about choosing the maximal number of processes):<\/p>\n\n<pre><code>max_job_num_per_gpu = 2\nrun_bash('ts -S %d'%max_job_num_per_gpu)\n<\/code><\/pre>\n\n<p>The next snippet iterates through a list of all combinations of hyper params \/ model params. Each element of the list is a dictionary, where the keys are the command line arguments for the <code>train.py<\/code> script<\/p>\n\n<pre><code>for combination_dict in combinations_list:\n\n    job_cmd = 'python train.py ' + '  '.join(\n            ['--{}={}'.format(flag, value) for flag, value in combination_dict.iteritems()])\n\n    submit_cmd = \"ts bash -c '%s'\" % job_cmd\n    run_bash(submit_cmd)\n<\/code><\/pre>\n\n<p><strong>A note about about choosing the maximal number of processes:<\/strong><\/p>\n\n<p>If you are short on GPUs, you can use <code>gpu_memory_fraction<\/code> you found, to set the number of processes as <code>max_job_num_per_gpu=int(1\/gpu_memory_fraction)<\/code><\/p>\n\n<p><strong>Notes about task spooler (<code>ts<\/code>):<\/strong><\/p>\n\n<ol>\n<li><p>You could set the number of concurrent processes to run (\"slots\") with:<\/p>\n\n<p><code>ts -S &lt;number-of-slots&gt;<\/code><\/p><\/li>\n<li><p>Installing <code>ts<\/code> doesn't requires admin privileges. You can download and compile it from source with a simple <code>make<\/code>, add it to your path and you're done.<\/p><\/li>\n<li><p>You can set up multiple queues (I use it for multiple GPUs), with<\/p>\n\n<p><code>TS_SOCKET=&lt;path_to_queue_name&gt; ts &lt;your-command&gt;<\/code><\/p>\n\n<p>e.g.<\/p>\n\n<p><code>TS_SOCKET=\/tmp\/socket-ts.gpu_queue_1 ts &lt;your-command&gt;<\/code><\/p>\n\n<p><code>TS_SOCKET=\/tmp\/socket-ts.gpu_queue_2 ts &lt;your-command&gt;<\/code><\/p><\/li>\n<li><p>See <a href=\"http:\/\/vicerveza.homeunix.net\/~viric\/soft\/ts\/article_linux_com.html\" rel=\"noreferrer\">here<\/a> for further usage example<\/p><\/li>\n<\/ol>\n\n<p><strong>A note about automatically setting the path names and file names:<\/strong>\nOnce you separate your main code from the experiment manager, you will need an efficient way to generate file names and directory names, given the hyper-params. I usually keep my important hyper params in a dictionary and use the following function to generate a single chained string from the dictionary key-value pairs.\nHere are the functions I use for doing it:<\/p>\n\n<pre><code>def build_string_from_dict(d, sep='%'):\n    \"\"\"\n     Builds a string from a dictionary.\n     Mainly used for formatting hyper-params to file names.\n     Key-value pairs are sorted by the key name.\n\n    Args:\n        d: dictionary\n\n    Returns: string\n    :param d: input dictionary\n    :param sep: key-value separator\n\n    \"\"\"\n\n    return sep.join(['{}={}'.format(k, _value2str(d[k])) for k in sorted(d.keys())])\n\n\ndef _value2str(val):\n    if isinstance(val, float): \n        # %g means: \"Floating point format.\n        # Uses lowercase exponential format if exponent is less than -4 or not less than precision,\n        # decimal format otherwise.\"\n        val = '%g' % val\n    else:\n        val = '{}'.format(val)\n    val = re.sub('\\.', '_', val)\n    return val\n<\/code><\/pre>\n"},{"AnswerId":"42650838","CreationDate":"2017-03-07T14:27:52.287","ParentId":null,"OwnerUserId":"4834515","Title":null,"Body":"<p>As I understand, firstly tensorflow constructs a symbolic graph and infers the derivatives based on chain rule. Then allocates memory for all (necessary) tensors, including some inputs and outputs of layers for efficiency. When running a session, data will be loaded into the graph but in general, memory use will not change any more. <\/p>\n\n<p>The error you met, I guess, may be caused by constructing several models in one GPU. <\/p>\n\n<p>Isolating your training\/evaluation code from the hyper parameters is a good choice, as @user2476373 proposed. But I am using bash script directly, not task spooler (may be it's more convenient), e.g.<\/p>\n\n<pre><code>CUDA_VISIBLE_DEVICES=0 python train.py --lrn_rate 0.01 --weight_decay_rate 0.001 --momentum 0.9 --batch_size 8 --max_iter 60000 --snapshot 5000\nCUDA_VISIBLE_DEVICES=0 python eval.py \n<\/code><\/pre>\n\n<p>Or you can write a 'for' loop in the bash script, not necessarily in python script. Noting that I used <code>CUDA_VISIBLE_DEVICES=0<\/code> at beginning of the script (the index could be 7 if you have 8 GPUs in one machine). Because based on my experience, I've found that tensorflow uses all GPUs in one machine if I didn't specify operations use which GPU with the code like this<\/p>\n\n<pre><code>with tf.device('\/gpu:0'):\n<\/code><\/pre>\n\n<p>If you want to try multi-GPU implementation, there is some <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/inception_train.py#L233\" rel=\"nofollow noreferrer\">example<\/a>. <\/p>\n\n<p>Hope this could help you.<\/p>\n"},{"AnswerId":"56050913","CreationDate":"2019-05-09T01:15:28.893","ParentId":null,"OwnerUserId":"11116409","Title":null,"Body":"<p>An easy solution: Give each model a unique session and graph. <\/p>\n\n<p>It works for this platform: TensorFlow 1.12.0, Keras 2.1.6-tf, Python 3.6.7, Jupyter Notebook.<\/p>\n\n<p><strong>Key code:<\/strong><\/p>\n\n<pre><code>with session.as_default():\n    with session.graph.as_default():\n        # do something about an ANN model\n<\/code><\/pre>\n\n<p><strong>Full code:<\/strong><\/p>\n\n<pre><code>import tensorflow as tf\nfrom tensorflow import keras\nimport gc\n\ndef limit_memory():\n    \"\"\" Release unused memory resources. Force garbage collection \"\"\"\n    keras.backend.clear_session()\n    keras.backend.get_session().close()\n    tf.reset_default_graph()\n    gc.collect()\n    #cfg = tf.ConfigProto()\n    #cfg.gpu_options.allow_growth = True\n    #keras.backend.set_session(tf.Session(config=cfg))\n    keras.backend.set_session(tf.Session())\n    gc.collect()\n\n\ndef create_and_train_ANN_model(hyper_parameter):\n    print('create and train my ANN model')\n    info = { 'result about this ANN model' }\n    return info\n\nfor i in range(10):\n    limit_memory()        \n    session = tf.Session()\n    keras.backend.set_session(session)\n    with session.as_default():\n        with session.graph.as_default():   \n            hyper_parameter = { 'A set of hyper-parameters' }  \n            info = create_and_train_ANN_model(hyper_parameter)      \n    limit_memory()\n<\/code><\/pre>\n\n<p><strong>Inspired by this link:<\/strong> <a href=\"https:\/\/stackoverflow.com\/questions\/54652536\/keras-tensorflow-backend-error-tensor-input-10-specified-in-either-feed-de\">Keras (Tensorflow backend) Error - Tensor input_1:0, specified in either feed_devices or fetch_devices was not found in the Graph<\/a><\/p>\n"},{"AnswerId":"42734613","CreationDate":"2017-03-11T11:28:47.030","ParentId":null,"OwnerUserId":"7682922","Title":null,"Body":"<p>You probably don't want to do this.<\/p>\n\n<p>If you run thousands and thousands of models on your data, and pick the one that evaluates best, you are not doing machine learning; instead you are memorizing your data set, and there is no guarantee that the model you pick will perform at all outside that data set.<\/p>\n\n<p>In other words, that approach is similar to having a single model, which has thousands of degrees of liberty. Having a model with such high order of complexity is problematic, since it will be able to fit your data better than is actually warranted; such a model is annoyingly able to memorize any noise (outliers, measurement errors, and such) in your training data, which causes the model to perform poorly when the noise is even slightly different.<\/p>\n\n<p>(Apologies for posting this as an answer, the site wouldn't let me add a comment.)<\/p>\n"}]}
{"QuestionId":42427179,"AnswerCount":1,"Tags":"<python><neural-network><keras><deep-learning><keras-layer>","CreationDate":"2017-02-23T22:11:51.933","AcceptedAnswerId":null,"OwnerUserId":710734.0,"Title":"Keras intermediate layers output","Body":"<p>I'm trying to get the intermediate layers output when using functional API of Keras. I'm able to get the output when using the standard Sequential API, but not with the functional API. <\/p>\n\n<p>I'm working on this working toy example:<\/p>\n\n<pre><\/pre>\n\n<p>But if I use the functional API, it doesn't work. The outputs are not correct. For example it is outputting the initial input in the second layer:<\/p>\n\n<pre><\/pre>\n\n<p><strong>ANSWER:<\/strong> Well, I solve it myself. Apparently when using the functional API layer 0 is the input itself. And so everything is shifted one position forward. <\/p>\n","answers":[{"AnswerId":"52909537","CreationDate":"2018-10-20T19:48:53.320","ParentId":null,"OwnerUserId":"2099607","Title":null,"Body":"<p>The issue arise from the fact, as the OP suggested, that the layer with index 0 (i.e. <code>model.layers[0]<\/code>) corresponds to the input layer: \"when using the functional API layer 0 is the input itself. And so everything is shifted one position forward.\"<\/p>\n\n<p><em>Note: this answer is posted as community wiki as suggested in <a href=\"https:\/\/meta.stackoverflow.com\/a\/251598\/2099607\">accepted answer<\/a> of <a href=\"https:\/\/meta.stackoverflow.com\/q\/251597\/2099607\">\"Question with no answers, but issue solved in the comments (or extended in chat)\"<\/a>.<\/em><\/p>\n"}]}
{"QuestionId":42427540,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-23T22:38:43.823","AcceptedAnswerId":42427844.0,"OwnerUserId":3291312.0,"Title":"Trying to expand working code to read multiple JPGs instead of just one into a tensor","Body":"<p>I am trying to read in a whole directory into one tensor, with each file in the directory a 28x28 image that is flattened into one row, with each new row representing another image (images are titled 0001.jpg, etc). I have been able to successfully do this for one image but have had no luck creating any loop mechanism to load the files.<\/p>\n\n<p>I realize this is probably a simple solution but I have no clue how to go about it. If anyone has an example to point me to, or any help would be incredibly appreciated. Thank you.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42427844","CreationDate":"2017-02-23T23:03:24.763","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>To load multiple images into the same tensor, you can use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/batch\" rel=\"nofollow noreferrer\"><code>tf.train.batch()<\/code><\/a>. For example, to combine 32 images&mdash;which must have the same dimensions&mdash;into a single tensor, you can modify your program as follows:<\/p>\n\n<pre><code>filenames = tf.train.match_filenames_once(\"C:\/train_data\/*.jpg\")\nfilename_queue = tf.train.string_input_producer(filenames)\n\nimage_reader = tf.WholeFileReader()\n_, image_file = image_reader.read(filename_queue)\nimage_orig = tf.image.decode_jpeg(image_file)\nimages = tf.image.decode_jpeg(image_file)\n\n# `image_batch` contains 32 consecutive images, packed into a single tensor.\nimage_batch, = tf.train.batch((images,), 32)\n<\/code><\/pre>\n\n<p>If your images might have different sizes, you can use an <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\" rel=\"nofollow noreferrer\">image-processing function<\/a> like <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/image\/resize_image_with_crop_or_pad\" rel=\"nofollow noreferrer\"><code>tf.image.resize_image_with_crop_or_pad()<\/code><\/a> to convert each <code>image<\/code> to the same shape before passing it to <code>tf.train.batch()<\/code>.<\/p>\n"}]}
{"QuestionId":42429436,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-24T01:57:08.433","AcceptedAnswerId":42429580.0,"OwnerUserId":null,"Title":"Tensorflow vs Python","Body":"<p>I'm not an expert in this domain so far. I am a web developer, and I am using rather Python and MySQL as programming language. I've heard about TensorFlow. What are the advantages of Tensorflow compared to Python? What is the use of Tensorflow?<\/p>\n\n<p><strong>EDIT :<\/strong> \nI know TensorFlow is a machine learning framework. What is the advantage of using that framework instead of another one?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42429580","CreationDate":"2017-02-24T02:12:25.960","ParentId":null,"OwnerUserId":"3763831","Title":null,"Body":"<p>TensorFlow isn't a programming language, it is a machine learning framework and python is one of the languages it supports. You can't really compare tensorflow and python. Python is a language. <\/p>\n"}]}
{"QuestionId":42429627,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-24T02:19:11.127","AcceptedAnswerId":42429652.0,"OwnerUserId":6751201.0,"Title":"Running TensorFlow official CNN example error: ImportError: cannot import name 'cifar10'","Body":"<p>An error happened when I runing the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\" rel=\"nofollow noreferrer\">TensorFlow official CNN example<\/a>.<\/p>\n\n<pre><\/pre>\n\n<p>I download  zip files from  and run  in  .<\/p>\n","answers":[{"AnswerId":"42429652","CreationDate":"2017-02-24T02:21:55.880","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>You probably downloaded the unstable master version.<\/p>\n\n<p>I recommend you to reinstall TensorFlow as explained in the <a href=\"https:\/\/www.tensorflow.org\/install\/\" rel=\"nofollow noreferrer\">official documentation<\/a>, and try again.<\/p>\n"}]}
{"QuestionId":42430022,"AnswerCount":1,"Tags":"<c++><opencv><neural-network><deep-learning><caffe>","CreationDate":"2017-02-24T03:03:09.427","AcceptedAnswerId":null,"OwnerUserId":7614599.0,"Title":"Caffe convert_imageset: symbol lookup error","Body":"<p>I used caffe to train my CNN.When I transfered my data to the lmdb format using the tool , it report an error:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>How could I solve it??<\/p>\n","answers":[{"AnswerId":"42465976","CreationDate":"2017-02-26T07:40:14.520","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The symbol you are missing is the function:<\/p>\n\n<pre><code>caffe::ReadImageToDatum(std::basic_string&lt;char, std::char_traits&lt;char&gt;, std::allocator&lt;char&gt; &gt; const&amp;, int, int, int, int, int, bool, std::basic_string&lt;char, std::char_traits&lt;char&gt;, std::allocator&lt;char&gt; &gt; const&amp;, caffe::Datum*)\n<\/code><\/pre>\n\n<p>(You can use <a href=\"https:\/\/sourceware.org\/binutils\/docs-2.24\/binutils\/c_002b_002bfilt.html\" rel=\"nofollow noreferrer\"><code>c++filt<\/code><\/a> shell utility to recover this from the symbol you got in your error message)<\/p>\n\n<p>The missing function is declared in <code>io.hpp<\/code> file, but its implementation in <code>io.cpp<\/code> is conditioned on <code>USE_OPENCV<\/code> directive (see <a href=\"https:\/\/github.com\/BVLC\/caffe\/issues\/5329\" rel=\"nofollow noreferrer\">BVLC\/caffe issue #5329<\/a>).  <\/p>\n\n<p>Re-build your caffe with <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/Makefile.config.example#L11\" rel=\"nofollow noreferrer\"><code>USE_OPENCV := 1<\/code><\/a> in <code>'Makefile.config'<\/code>.<\/p>\n"}]}
{"QuestionId":42430132,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><keras-layer>","CreationDate":"2017-02-24T03:17:47.477","AcceptedAnswerId":42432042.0,"OwnerUserId":3167448.0,"Title":"How does one train a single layered Neural Net in Keras with cifar?","Body":"<p>I wanted to train a single layered neural network using the cifar data set and the keras framework. Since each image for the data set is 32 by 32 by 3 I wasn't quite sure how to process the image using a single layered network with no convolution. I think that flattening each image to have a data set of shape  is the right thing to do. Thus I did the following:<\/p>\n\n<pre><\/pre>\n\n<p>then I just made a single layered network that matched the input dimension:<\/p>\n\n<pre><\/pre>\n\n<p>the code seems to compile fine and everything. Thus, is this the correct way to train a single layered Neural Network on an image data set without convolution? I guess the main thing that is throwing me off is that the image is a 3D tensor but a singled layer net would just treat it as a 1D vector regardless of its shape. Right?<\/p>\n\n<p>Also, since Keras provided a  function it seemed unclear to me if that was a preferred method to use due to efficiency or some other reason. However, I wasn't able to make that one work.<\/p>\n\n<p>Also, this goes without saying but the softmax layer doesn't really count as another layer. Right? I <strong>want<\/strong> it to be single layer.<\/p>\n\n<hr>\n\n<p>whole code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42432042","CreationDate":"2017-02-24T06:17:26.397","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>You're doing everything right, this is the way if you only want one dense layer. It accepts only 1D tensors so you did well to reshape. <\/p>\n\n<p>The softmax isn't considered as a layer because it doesnt have any weights or parameters to train. <\/p>\n\n<p>Just out of curiosity, why do you use relu on the inputs? Aren't they already supposed to be between 0 and 1? <\/p>\n"}]}
{"QuestionId":42430222,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-24T03:28:53.273","AcceptedAnswerId":null,"OwnerUserId":2123328.0,"Title":"Tensorflow's provided linear regression example is giving bad results","Body":"<p>I'm following the <a href=\"https:\/\/www.tensorflow.org\/get_started\/get_started\" rel=\"nofollow noreferrer\">linear regression example<\/a> from the tensor flow get started tutorial (not the one using tf.contrib.learn) . While the provided example works, if I use any other data as an input to train on I always get the following printed out:<\/p>\n\n<pre><\/pre>\n\n<p>I have tried it with various different data sets. Even simple data sets like this one <\/p>\n\n<pre><\/pre>\n\n<p>I can implement linear regression on the data without any problem if I implement it only in numpy and get correct weight and bias values.<\/p>\n\n<p>I have tried adjusting the hyperparameters but still no luck. It always returns nan<\/p>\n\n<p>I have also literally copied the code from the site and just replaced the data values so I know there is no hidden typo. This has been driving me crazy. Please help.<\/p>\n","answers":[]}
{"QuestionId":42430331,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T03:39:23.717","AcceptedAnswerId":42430404.0,"OwnerUserId":5721911.0,"Title":"TensorFlow: construct a tensor with recursively defined elements?","Body":"<p>I would like to do something like:<\/p>\n\n<pre><\/pre>\n\n<p>to construct a tensor 'x'. Is such a construct possible? I looked at 'tf.while_loop' but does not seem to help.<\/p>\n","answers":[{"AnswerId":"42430404","CreationDate":"2017-02-24T03:46:43.503","ParentId":null,"OwnerUserId":"1269703","Title":null,"Body":"<pre><code>def f(x):\n    return tf.sin(x)\nx = []\nx.append(tf.constant(1.0))\nfor i in range(1,100):\n    x.append(f(i))\n<\/code><\/pre>\n\n<p>Is this what you want?\nThis is not recursive by the way. It's iterative.<\/p>\n"},{"AnswerId":"42633390","CreationDate":"2017-03-06T18:55:02.953","ParentId":null,"OwnerUserId":"5721911","Title":null,"Body":"<p>The solution in <a href=\"https:\/\/stackoverflow.com\/questions\/37697747\/typeerror-tensor-object-does-not-support-item-assignment-in-tensorflow\">TypeError: &#39;Tensor&#39; object does not support item assignment in TensorFlow<\/a> pretty much solves the problem as Atirag pointed out, but 'tf.pack()' seems to have been deprecated. Just for completeness, here is an up-to-date solution to the posted problem:<\/p>\n\n<pre><code>x0 = tf.Variable(tf.ones(1, dtype=tf.float32))\ndef f(x):\n    return tf.sin(x)\nx = [x0]\nfor i in range(1, 100):\n    x.append(f(x[i - 1]))\n\ntf.stack(x)\n<\/code><\/pre>\n"}]}
{"QuestionId":42430355,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-24T03:42:08.837","AcceptedAnswerId":null,"OwnerUserId":5657150.0,"Title":"Tensorflow Inception V3 unable to load compute graph","Body":"<p>I have retrained Tensorflow's Inception V3's last layer on a flower dataset. This was done using:<\/p>\n\n<pre><\/pre>\n\n<p>The training was successful, and then I ran:<\/p>\n\n<pre><\/pre>\n\n<p>This too ran fine and next I ran:<\/p>\n\n<pre><\/pre>\n\n<p>When I run this, I receive an error that says<\/p>\n\n<pre><\/pre>\n\n<p>Any help is greatly appreciated, thank you.<\/p>\n","answers":[{"AnswerId":"44732415","CreationDate":"2017-06-24T03:20:59.490","ParentId":null,"OwnerUserId":"7386732","Title":null,"Body":"<p>I have run this using following command instead of bazel and I found it easier.<\/p>\n\n<pre><code>python \/path_to_file\/label_image.py \/path_to_image\/image.jpeg\n<\/code><\/pre>\n\n<p>First make sure that graph is created after you run retrain.py and it is at the correct location. (default is inside \/tmp\/). If you want you can get the script <a href=\"https:\/\/github.com\/llSourcell\/tensorflow_image_classifier\/blob\/master\/src\/label_image.py\" rel=\"nofollow noreferrer\">here<\/a> and you can change the file locations as per your need.<\/p>\n\n<p>Note: If graph file is not created you may want to check <a href=\"https:\/\/stackoverflow.com\/questions\/42267743\/tensorflow-retrain-py-oserror-errno-2-no-such-file-or-directory\/44685809#44685809\">this<\/a> regarding running retrain.py<\/p>\n"}]}
{"QuestionId":42430525,"AnswerCount":1,"Tags":"<python><tensorflow><protocol-buffers>","CreationDate":"2017-02-24T03:59:59.647","AcceptedAnswerId":42816166.0,"OwnerUserId":7071375.0,"Title":"FixedLenSequenceFeature unexpected behaviour","Body":"<p>Tried to better understand the most popular sequential TFRecord tutorial:\n<a href=\"http:\/\/www.wildml.com\/2016\/08\/rnns-in-tensorflow-a-practical-guide-and-undocumented-features\/\" rel=\"nofollow noreferrer\">http:\/\/www.wildml.com\/2016\/08\/rnns-in-tensorflow-a-practical-guide-and-undocumented-features\/<\/a><\/p>\n\n<p>Where this came up in the serialization part:<\/p>\n\n<pre><\/pre>\n\n<p>why we couldn't use this?<\/p>\n\n<pre><\/pre>\n\n<p>Using the latter method will result in the following error (which I hardly understand)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42816166","CreationDate":"2017-03-15T16:58:58.337","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Assuming <code>ex<\/code> is a <code>tf.SequenceExample<\/code>, I get a different error from you:<\/p>\n\n<pre><code>&gt;&gt;&gt; ex = tf.SequenceExample()\n&gt;&gt;&gt; sequence = [1, 2, 3]\n&gt;&gt;&gt; ex.feature_lists.feature_list[\"tokens\"].int64_list.value.extend(sequence)\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nAttributeError: 'FeatureList' object has no attribute 'int64_list'\n<\/code><\/pre>\n\n<p>However, the following code at least appears to be syntactically correct, so I assume you mean the following:<\/p>\n\n<pre><code>&gt;&gt;&gt; ex.feature_lists.feature_list[\"tokens\"].feature.add().int64_list.value.extend(sequence)\n<\/code><\/pre>\n\n<p>Why does creating the example this way yield an error? The short answer is that the data is subtly different. When you do the following, you are creating an example whose <code>token<\/code> feature list has 22 scalar values:<\/p>\n\n<pre><code>for token, label in zip(sequence, labels):\n  fl_tokens.feature.add().int64_list.value.append(token)\n<\/code><\/pre>\n\n<p>Notice that you're calling <code>feature.add()<\/code> once for each element of <code>sequence<\/code>. By contrast, in the <code>extend()<\/code>-based version you're creating an example whose <code>token<\/code> feature list is a single value, which is a vector of 22 elements.<\/p>\n\n<p>While these might be mathematically equivalent, TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/parse_single_sequence_example\" rel=\"nofollow noreferrer\"><code>tf.parse_single_sequence_example()<\/code><\/a> function expects you to pass explicit shapes when you define the various features in the <code>sequence_features<\/code> argument, and will raise the error you are seeing if they do not match.<\/p>\n"}]}
{"QuestionId":42431904,"AnswerCount":0,"Tags":"<c++><caffe>","CreationDate":"2017-02-24T06:06:59.663","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Loading binaryproto file into caffe library","Body":"<p>I have a binaryproto file and trying to load into caffe library.\nThe whole API loading the binary file is as follow.<\/p>\n\n<pre><\/pre>\n\n<p>I always have error at  <\/p>\n\n<pre><\/pre>\n\n<p>I provide a correct file path to the function call.<\/p>\n\n<p>The error is as follow.<\/p>\n\n<pre><\/pre>\n\n<p>What could be wrong? There is no such file or directory error, but the file is there and the correct path is provided.<\/p>\n\n<p>EDIT:\nAfter reading the file at , I print the  and found as<\/p>\n\n<pre><\/pre>\n\n<p>That means image format at  is wrong (actual image has channel 3, width 640 and height 480).<\/p>\n\n<p>That could be because of two reasons\n(1)Saving the mean.binaryproto file has wrong format.\nMy program for saving mean.binaryproto file is<\/p>\n\n<pre><\/pre>\n\n<p>(2)Or \nReading the file into blob_proto has problem.<\/p>\n\n<p>How can I solve the problem?<\/p>\n","answers":[]}
{"QuestionId":42432496,"AnswerCount":0,"Tags":"<deep-learning><nvidia><torch><multi-gpu>","CreationDate":"2017-02-24T06:52:08.390","AcceptedAnswerId":null,"OwnerUserId":4741002.0,"Title":"Multi-GPU Batch Normalization in Torch","Body":"<p>Does torch synchronize the  and  between the clones present in the N GPUs during a multi-GPU training? It is being said that any Torch module that goes into  is optimized by the usage of NVIDIAs  library. <\/p>\n\n<p>But a few people claim that Torch doesn't synchronize the  and . I'd be really glad if someone shed some clarity on this topic.<\/p>\n","answers":[]}
{"QuestionId":42432526,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-24T06:54:58.057","AcceptedAnswerId":null,"OwnerUserId":5627463.0,"Title":"Best device placement strategies for RNNs?","Body":"<p>Consider a simple n-layer stacked RNN (LSTM, GRU etc.) for sequential prediction, what is the most efficient device placement strategy (if multiple GPUs are available)? I think if is preferred to dedicate one GPU to each layer of the RNN, but what about softmax and loss computation? Should they be placed on GPUs as well or CPUs? What is more, if we further add an attention mechanism, should the attention be placed on CPUs?<\/p>\n","answers":[{"AnswerId":"42514513","CreationDate":"2017-02-28T16:56:01.883","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>This is an active area of research; and in general it depends on the specifics of your network.  Probably you want to place the softmax and loss on the same GPU (if they'll fit); likely either on the GPU of the top layer, or on gpu:0.  Regarding attention, I don't think there's a very clear answer.  It depends on the size of your attention weights' sizes, etc.  In these cases, your best bet is just to try the two placements I suggested for the softmax\/loss.  This leaves you a grid search with 4 possible locations.  Try them all and report back :)<\/p>\n"}]}
{"QuestionId":42432618,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T07:02:31.173","AcceptedAnswerId":null,"OwnerUserId":6776821.0,"Title":"Tensorflow error: TypeError: Fetch argument None has invalid type <type 'NoneType'>","Body":"<p>I get the error \"TypeError: Fetch argument None has invalid type \" when I run the following codes:<\/p>\n\n<pre><\/pre>\n\n<p>Here is my network defintion:<\/p>\n\n<pre><\/pre>\n\n<p>I notice this <a href=\"https:\/\/stackoverflow.com\/questions\/39114832\/tensorflow-typeerror-fetch-argument-none-has-invalid-type-type-nonetype\">post<\/a> ask the same question, but the above codes seems do the same thing as the answer since I assign  to .<\/p>\n\n<p>Note: my tensorflow version is 0.12.1<\/p>\n\n<p>EDIT:\nHere is my error message:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42432842,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T07:17:21.933","AcceptedAnswerId":null,"OwnerUserId":3809822.0,"Title":"TensorFlow on Windows 10: \u201cnot a supported wheel on this platform\u201d error","Body":"<p>I was following the <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">guide steps<\/a> until get the \"not a supported wheel on this platform\" error. I checked the Anaconda is properly installed with the following version information:<\/p>\n\n<pre><\/pre>\n\n<p>Anyone know how to solve this?<\/p>\n","answers":[{"AnswerId":"42435996","CreationDate":"2017-02-24T10:14:52.120","ParentId":null,"OwnerUserId":"7568768","Title":null,"Body":"<p>I think it's only available for Python 3.5.x not 3.6.<\/p>\n\n<p>You can quickly create a 3.5 environment with <code>conda create -n tensorflow python=3.5<\/code><\/p>\n"}]}
{"QuestionId":42433317,"AnswerCount":1,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2017-02-24T07:51:26.557","AcceptedAnswerId":null,"OwnerUserId":3324702.0,"Title":"tensorflow session.run() hangs while attempting to restore rnn model based on tutorial code","Body":"<p>I've been walking through the RNN code in TenserFlow tutorial: <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/recurrent<\/a><\/p>\n\n<p>The original RNN code is here: <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/ptb_word_lm.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/tutorials\/rnn\/ptb\/ptb_word_lm.py<\/a><\/p>\n\n<p>I saved the trained RNN model as 'train-model' through<\/p>\n\n<pre><\/pre>\n\n<p>Now I'm trying to restore the saved model and run additional test with it by<\/p>\n\n<pre><\/pre>\n\n<p>It seem that the model is loaded correctly, but it hangs at the line<\/p>\n\n<pre><\/pre>\n\n<p>in function , when called for computing .   is unable to quit the program, and the GPU utilization is at 0%, so it is most probably blocked on something. <\/p>\n\n<p>Any help would be greatly appreciated!<\/p>\n","answers":[{"AnswerId":"45590217","CreationDate":"2017-08-09T12:05:46.447","ParentId":null,"OwnerUserId":"5949718","Title":null,"Body":"<p>Try installing Tensorflow from source. It is recommended because you can build the desired Tensorflow binary for the specific architecture (GPU, CUDA, cuDNN). <\/p>\n\n<p>This is even mentioned as one of the Best Practices for improving Tensorflow performance. Check the  <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide\" rel=\"nofollow noreferrer\">Tensorflow performance  Guide<\/a>. A small excerpt from the same :<\/p>\n\n<blockquote>\n  <p>Building from source with compiler optimizations for the target hardware and ensuring the latest CUDA platform and cuDNN libraries are installed results in the highest performing installs.<\/p>\n<\/blockquote>\n\n<p>The problem you mentioned typically occurs when the configured Compute Capability with which the Tensorflow binary was built is different than that of your GPU. But installing it from source gives you an option to configure the specific Compute Capability. Check the <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"nofollow noreferrer\">guide for installing from source<\/a>.<\/p>\n"}]}
{"QuestionId":42433443,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><keras><conv-neural-network>","CreationDate":"2017-02-24T07:59:25.773","AcceptedAnswerId":null,"OwnerUserId":5653822.0,"Title":"Questions about keras example pretrained_word_embeddings","Body":"<p>I have several questions about Keras example <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/pretrained_word_embeddings.py\" rel=\"nofollow noreferrer\">pretrained_word_embeddings<\/a> to increase a level of understanding how it works.<\/p>\n\n<ol>\n<li><p>Is it reasonable to use  layer in such model? <\/p><\/li>\n<li><p>Last  layer has to cover all output shape every time? At original model, last conv layer output is 35 and we set up maxpool the same 35 value.<\/p><\/li>\n<li><p>Am I right if to say that increase of value 128 (kernels number) will increase accuracy?<\/p><\/li>\n<li><p>Is it make sense to put additional conv layers to increase the accuracy? Even if it will decrease model training phase.<\/p><\/li>\n<\/ol>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"42434604","CreationDate":"2017-02-24T09:05:17.663","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So basically there is one simple answer to your questions - you need to test it:<\/p>\n\n<ol>\n<li>Adding <code>dropout<\/code> is usually a good thing. It introduces the reasonable amount of randomization and regularization. The downside is that you need to set the right value of its parameter - which sometimes might take a while. <\/li>\n<li>In my opinion - the value of a <code>MaxPooling1D<\/code> in the last pooling layer was set in this way in order to reduce the dimensionality of the next layer input. One may check if e.g. values like a half of the number in previous layer (in a presented case - e.g. <code>18<\/code>) - which only doulbes the size of the input to a next layer - could introduce any improvement.<\/li>\n<li>It's hard to say - if you e.g. have a small amount of data with a really rigid structure - to much parameters might seriously harm your training. The best way is to test different parameters values in either <strong>grid<\/strong> or <strong>random<\/strong> search paradigm. It's belived that <a href=\"http:\/\/www.jmlr.org\/papers\/volume13\/bergstra12a\/bergstra12a.pdf\" rel=\"nofollow noreferrer\">random search<\/a> does a better job :) <\/li>\n<\/ol>\n"}]}
{"QuestionId":42434500,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T09:00:12.620","AcceptedAnswerId":42434583.0,"OwnerUserId":5380429.0,"Title":"Getting mnist import error in tensorflow","Body":"<pre><\/pre>\n\n<p>I have just started to work with tensorflow and when I try to import the mnist dataset, I get the following error.<\/p>\n\n<pre><\/pre>\n\n<p>I have checked the python path and confirmed that the mnist dataset exists in one of the python paths(\/usr\/local\/lib\/python2.7\/dist-packages).<\/p>\n\n<pre><\/pre>\n\n<p>I even tried this method as suggested <a href=\"https:\/\/stackoverflow.com\/a\/33664749\">here<\/a> but again I get a import error.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42434583","CreationDate":"2017-02-24T09:04:08.177","ParentId":null,"OwnerUserId":"2283175","Title":null,"Body":"<p>Don't call your file <code>tensorflow.py<\/code>, call it something else. For instance <code>my_experiment1.py<\/code>. That should fix it!<\/p>\n"}]}
{"QuestionId":42434788,"AnswerCount":1,"Tags":"<tensorflow><reinforcement-learning>","CreationDate":"2017-02-24T09:15:15.703","AcceptedAnswerId":null,"OwnerUserId":7615953.0,"Title":"Efficient feeding of data for reinforcement learning algorithms","Body":"<p>I am currently implementing the deep double Q learning algorithm in TensorFlow. I have an experience replay buffer implemented based on NumPy arrays. However, some performance analysis shows that the feeding of the data from NumPy arrays to the graph using feed_dict is very inefficient. This is also pointed out in the documentation <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/performance\/performance_guide<\/a>.<\/p>\n\n<p>Does anyone have a suggestion as to have the feeding can be done more efficiently? With static datasets the feeding can be done with input pipelines, such as record readers. Experience replay buffer, however, change with time, making this type of feeding more challenging.<\/p>\n\n<p>Any answers are greatly appreciated, thank you!<\/p>\n","answers":[{"AnswerId":"44639180","CreationDate":"2017-06-19T20:15:00.587","ParentId":null,"OwnerUserId":"4023951","Title":null,"Body":"<p>The short answer is <strong>there is no way to speed it up<\/strong> if your data is a minimum set and is unique.  If your data has some redundancies or unneeded decoration that you can remove, then remove it prior to feeding it into your model, but assuming you've already done that and your data is unique in each batch, then there is nothing (in this regard) that you can do.<\/p>\n\n<p>That said, there are things you can do to improve your networks performance.  <\/p>\n\n<ul>\n<li>Your Q-learning algorithm may be able to be improved as per <a href=\"http:\/\/file:\/\/\/Users\/anton\/Downloads\/Wiering_98_fastonline.pdf\" rel=\"nofollow noreferrer\">this paper<\/a>, which basically says don't do a learning phase until your network as accumulated error past a threshold.  <\/li>\n<li>If you are reusing some of the training sets or replaying them, then you could use a load stage to load the training data into the GPU for quick playback.<\/li>\n<\/ul>\n"}]}
{"QuestionId":42435378,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-24T09:45:54.810","AcceptedAnswerId":null,"OwnerUserId":217019.0,"Title":"Dynamic shape and # of layers based on placeholders in TensorfFlow","Body":"<p>Is it possible to create a shape based on a placeholder?<\/p>\n\n<p>I have 2 use cases:<\/p>\n\n<pre><\/pre>\n\n<p>And:<\/p>\n\n<pre><\/pre>\n\n<p>The first one I guess I can \"fix\" with  (not sure about performance penalties in such a case).<\/p>\n\n<p>For the second one I have no idea? <\/p>\n\n<p><strong>The why:<\/strong> I want to build and export the Graph using Python and then read it for training\/predictions using the Java API. I don't want to prepare a file for every single hidden layer combination so I wanted to export only a single \"template\" graph, something like this:<\/p>\n\n<pre><\/pre>\n\n<p>And then the Java user would simply  an array with the hidden network config. But I am getting  when trying to generate the graph.<\/p>\n","answers":[{"AnswerId":"42435967","CreationDate":"2017-02-24T10:13:16.620","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Unfortunately I don't think what you want is possible. What would happen with a dynamic number of hidden layers? Randomly spawning extra (untrained) layers is not something that is usual with neural networks. <\/p>\n\n<p>Maybe clarifying your question would work. For example: lets say you either want to have two or three hidden layers, there are some possibilities. You could have two different sets of weights that either from layer two to layer X, or from layer two to layer three to layer X. If you want something like this, check tf.cond on this page: <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/control_flow_ops\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/control_flow_ops<\/a><\/p>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":42436266,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><keras>","CreationDate":"2017-02-24T10:27:23.247","AcceptedAnswerId":42444835.0,"OwnerUserId":3218279.0,"Title":"Netwok architecture understanding using keras code","Body":"<p>My data is of 68871 x 43, where the features are in the column no. 1-43 and each label is represented as a 1x21 vector<\/p>\n\n<p>In my keras code:<\/p>\n\n<ul>\n<li>print trainX.shape -----> (41311, 10, 43)  <\/li>\n<li>print trainY.shape -----> (41311, 21)<\/li>\n<li>print testX.shape  ------> (27538, 10, 43) <\/li>\n<li>print testY.shape  ------> (27538, 21)<\/li>\n<\/ul>\n\n<p>When I run the following keras code:<\/p>\n\n<pre><\/pre>\n\n<p>The code executes like this <a href=\"https:\/\/i.stack.imgur.com\/jSyjX.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jSyjX.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>My understanding for the keras toolkit is that, if the 3D tensor for trainX is of the shape (41311,10,43) then keras should create a lstm with 10 timestep. If this is so then why in the screenshot the samples run one by one, if the timestep is 10 then it should have run in the batches of 10 and get 10 prediction and then run the next 10 batches of samples.<\/p>\n\n<p>Anyone can answer me: why the screenshot shows the samples run one by one when the timestep is 10 ? \nThanks so much in advance.<\/p>\n","answers":[{"AnswerId":"42444835","CreationDate":"2017-02-24T17:31:49.230","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>By sample they understand a numpy array of shape (10,43). The first dimension is your number of samples. <\/p>\n\n<p>So what the network is doing is :\n1) split the input into batches of shape (batch, 10, 43)<\/p>\n\n<p>2) feed the lstm, one sample at a time. A sample being a sequence of 10 events\/tensors of length 43 each. So each sequence is 1 sample. <\/p>\n\n<p>Is that clear? <\/p>\n"}]}
{"QuestionId":42437115,"AnswerCount":8,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-02-24T11:08:11.443","AcceptedAnswerId":null,"OwnerUserId":5122790.0,"Title":"Tensorflow: Replacement for tf.nn.rnn_cell._linear(input, size, 0, scope)","Body":"<p>I am trying to get the SequenceGAN (<a href=\"https:\/\/github.com\/LantaoYu\/SeqGAN\" rel=\"nofollow noreferrer\">https:\/\/github.com\/LantaoYu\/SeqGAN<\/a>) from <a href=\"https:\/\/arxiv.org\/pdf\/1609.05473.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1609.05473.pdf<\/a> to run.<br>\nAfter fixing the obvious errors, like replacing  with , it still doesn't run, since the highway-network part requires the  function:<\/p>\n\n<pre><\/pre>\n\n<p>the  function doesn't appear to be there anymore in Tensorflow 1.0 or 0.12, and I have no clue what to replace it with. I can't find any new implementations of this, or any information on tensorflow's github or (unfortunately very sparse) documentation.<\/p>\n\n<p>Does anybody know the new pendant of the function?\nThanks a lot in advance!<\/p>\n","answers":[{"AnswerId":"52764340","CreationDate":"2018-10-11T15:57:26.827","ParentId":null,"OwnerUserId":"10490070","Title":null,"Body":"<p>tensorflow.python.ops.rnn_cell_impl._linear now is at tensorflow.contrib.rnn.python.ops.core_rnn_cell._linear. And I prefer to use tf.layers.Dense to replace.\nfor example, change<\/p>\n\n<pre><code>from tensorflow.contrib.rnn.python.ops import core_rnn_cell\ncore_rnn_cell._linear(states, length, bias=True)\n<\/code><\/pre>\n\n<p>to<\/p>\n\n<pre><code>tf.layers.Dense(units=length)(states)\n<\/code><\/pre>\n\n<p>I'm using tensorflow 1.6.<\/p>\n"},{"AnswerId":"43001802","CreationDate":"2017-03-24T14:15:29.923","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>The answer of ruoho ruotsi is almost correct:\nYet, the <em>definition<\/em> of <code>linear<\/code> is not located in <code>tf.contrib.rnn.basicRNNCell<\/code>, but in <code>tf.contrib.rnn.python.ops.rnn_cell<\/code>, or <code>tf.contrib.rnn.python.ops.core_rnn_cell_impl<\/code>, respectively.<\/p>\n\n<p>You can find their source code <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9967bf6702153631e27f52ac95ab822f73fe758d\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/9967bf6702153631e27f52ac95ab822f73fe758d\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py\" rel=\"nofollow noreferrer\">here<\/a>.<\/p>\n"},{"AnswerId":"53101886","CreationDate":"2018-11-01T13:04:23.927","ParentId":null,"OwnerUserId":"7429461","Title":null,"Body":"<p>Right now when you what to use <code>_linear<\/code> function, please use it like this: <\/p>\n\n<p><code>from tensorflow.contrib.rnn.python.ops import core_rnn_cell<\/code><\/p>\n\n<pre><code>core_rnn_cell._linear(output, size)\n<\/code><\/pre>\n\n<p>Here is the link of its corresponding source file. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell.py\" rel=\"nofollow noreferrer\">source file<\/a><\/p>\n\n<p>This is because after tf r1.5, the location of this function is changed.<\/p>\n\n<p>When you are using tf r1.4 or even older version, please use it like this:<\/p>\n\n<pre><code>from tensorflow.python.ops import rnn_cell_impl\n\nrnn_cell_impl._linear(output, size)\n<\/code><\/pre>\n\n<p>Here is the link of this older version of <code>_linear<\/code>. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.4\/tensorflow\/python\/ops\/rnn_cell_impl.py\" rel=\"nofollow noreferrer\">source file<\/a><\/p>\n\n<p>As for <code>tf.nn.rnn_cell._linear<\/code>, this function is not available a long time ago.<\/p>\n"},{"AnswerId":"44238429","CreationDate":"2017-05-29T08:57:47.280","ParentId":null,"OwnerUserId":"1894478","Title":null,"Body":"<p>As for now (1.2r0) you can replace it with \n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/fully_connected\" rel=\"nofollow noreferrer\"><code>tf.contrib.layers.fully_connected(inputs=[inputs, state], num_outputs=self._num_units, biases_initializer=tf.constant_initializer(0.0), activation_fn=None)<\/code><\/a>\nNote that <code>biases_initializer<\/code> defaults to all zeros for <code>_linear<\/code> and for no bias at all for <code>fully_connected<\/code>.<\/p>\n"},{"AnswerId":"55320605","CreationDate":"2019-03-24T03:59:07.557","ParentId":null,"OwnerUserId":"10399691","Title":null,"Body":"<p>Use <code>rnn_cell.LayerNormBasicLSTMCell._linear<\/code> instead of <code>_linear<\/code><\/p>\n"},{"AnswerId":"42517842","CreationDate":"2017-02-28T20:10:12.980","ParentId":null,"OwnerUserId":"3844056","Title":null,"Body":"<p>With version 1.0, stuff has moved all around. I've had similar hunts updating <code>tf.nn.rnn_cell.LSTMCell<\/code> to <code>tf.contrib.rnn.BasicLSTMCell<\/code>.<\/p>\n\n<p>For your case <code>tf.nn.rnn_cell._linear<\/code> now lives in <code>tf.contrib.rnn.python.ops.core_rnn_cell_impl<\/code> as well as the definition of the <code>BasicRNNCell<\/code>. Checking the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\/BasicRNNCell\" rel=\"nofollow noreferrer\">BasicRNNCell docs<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py#L113-L118\" rel=\"nofollow noreferrer\">source code<\/a>, we see at L113-L118 the use of _linear. <\/p>\n\n<pre><code>  def __call__(self, inputs, state, scope=None):\n    \"\"\"Most basic RNN: output = new_state = act(W * input + U * state + B).\"\"\"\n    with _checked_scope(self, scope or \"basic_rnn_cell\", reuse=self._reuse):\n      output = self._activation(\n          _linear([inputs, state], self._num_units, True))\n    return output, output\n<\/code><\/pre>\n\n<p>the _linear method is defined at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py#L854\" rel=\"nofollow noreferrer\">line 854<\/a> as a:<br>\n<code>Linear map: sum_i(args[i] * W[i]), where W[i] is a variable.<\/code><\/p>\n\n<p>Good luck!<\/p>\n"},{"AnswerId":"42661991","CreationDate":"2017-03-08T02:40:19.530","ParentId":null,"OwnerUserId":"7675704","Title":null,"Body":"<p>To solving this problem,we can define a linear() function.<\/p>\n\n<pre><code>def linear(input_, output_size, scope=None):\n    '''\n    Linear map: output[k] = sum_i(Matrix[k, i] * args[i] ) + Bias[k]\n    Args:\n        args: a tensor or a list of 2D, batch x n, Tensors.\n    output_size: int, second dimension of W[i].\n    scope: VariableScope for the created subgraph; defaults to \"Linear\".\n    Returns:\n    A 2D Tensor with shape [batch x output_size] equal to\n    sum_i(args[i] * W[i]), where W[i]s are newly created matrices.\n    Raises:\n    ValueError: if some of the arguments has unspecified or wrong shape.\n    '''\n\n    shape = input_.get_shape().as_list()\n    if len(shape) != 2:\n        raise ValueError(\"Linear is expecting 2D arguments: %s\" % str(shape))\n    if not shape[1]:\n        raise ValueError(\"Linear expects shape[1] of arguments: %s\" % str(shape))\n    input_size = shape[1]\n\n    # Now the computation.\n    with tf.variable_scope(scope or \"SimpleLinear\"):\n        matrix = tf.get_variable(\"Matrix\", [output_size, input_size], dtype=input_.dtype)\n        bias_term = tf.get_variable(\"Bias\", [output_size], dtype=input_.dtype)\n\n    return tf.matmul(input_, tf.transpose(matrix)) + bias_term\n\n\ndef highway(input_, size, num_layers=1, bias=-2.0, f=tf.nn.relu, scope='Highway'):\n    \"\"\"Highway Network (cf. http:\/\/arxiv.org\/abs\/1505.00387).\n    t = sigmoid(Wy + b)\n    z = t * g(Wy + b) + (1 - t) * y\n    where g is nonlinearity, t is transform gate, and (1 - t) is carry gate.\n    \"\"\"\n\n    with tf.variable_scope(scope):\n        for idx in range(num_layers):\n            g = f(linear(input_, size, scope='highway_lin_%d' % idx))\n\n            t = tf.sigmoid(linear(input_, size, scope='highway_gate_%d' % idx) + bias)\n\n            output = t * g + (1. - t) * input_\n            input_ = output\n\n    return output\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/github.com\/mkroutikov\/tf-lstm-char-cnn\/blob\/7e899e6992cbf9a96e6d791e5d364eaaeec339a2\/model.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/mkroutikov\/tf-lstm-char-cnn\/blob\/7e899e6992cbf9a96e6d791e5d364eaaeec339a2\/model.py<\/a><\/p>\n"},{"AnswerId":"46022279","CreationDate":"2017-09-03T10:04:48.673","ParentId":null,"OwnerUserId":"8409755","Title":null,"Body":"<p>I met this error while using SkFlow's TensorFlowDNNRegressor.\nThe first time I saw the answer of ruoho ruots, I am a bit confused. \nBut the next day I realized what he meant. <\/p>\n\n<p>Here is what I do:<\/p>\n\n<pre><code>from tensorflow.python.ops import rnn_cell_impl\n<\/code><\/pre>\n\n<p>replace <code>tf.nn.rnn_cell._linear<\/code> with <code>rnn_cell_impl._linear<\/code><\/p>\n"}]}
{"QuestionId":42437169,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-24T11:10:23.330","AcceptedAnswerId":42450565.0,"OwnerUserId":4590360.0,"Title":"Modify a 1D tensor on the basis of a condition","Body":"<p>I want to apply a condition to each element of a 1D tensor in tensorflow and hence modify the input tensor. For example, if the tensor is:<\/p>\n\n<p>,<\/p>\n\n<p>I want to check whether each element is greater than  or not. If it is then that element becomes  else . How could same be done in tensorflow?<\/p>\n\n<p>What I have tried until now is, writing a python function and then using it in tensorflow using , but it is not working. See below code -     <\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p>Is it possible to do this in  without writing any python function?<\/p>\n","answers":[{"AnswerId":"42450565","CreationDate":"2017-02-25T00:40:11.660","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>You could do the following:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import tensorflow as tf\n\ny_true = tf.Variable([0.0, 0.3, 0.0, 0.4, 0.0, 0.1, 0.0, 0.0])\ncomp_op = tf.greater(y_true, 0.1) # returns boolean tensor\ncast_op = tf.cast(comp_op, tf.int32) # casts boolean tensor into int32\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    print(sess.run(cast_op))\n<\/code><\/pre>\n\n<p>prints: <code>[0 1 0 1 0 0 0 0]<\/code><\/p>\n"}]}
{"QuestionId":42437467,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T11:26:23.793","AcceptedAnswerId":42438144.0,"OwnerUserId":2641038.0,"Title":"When does Tensorflow update weights and biases?","Body":"<p>When does tensorflow update weights and biases in the for loop?<\/p>\n\n<p>Below is the code from tf's github. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_softmax.py\" rel=\"noreferrer\">mnist_softmax.py<\/a><\/p>\n\n<pre><\/pre>\n\n<ol>\n<li>When does tensorflow update weights and biases?<\/li>\n<li>Does it update them when running ? If so, Does it mean, in this program, tf update weights and biases 1000 times?<\/li>\n<li>Or does it update them after finishing the whole for loop?<\/li>\n<li>If 2. is correct, my next question is, does tf update the model using different training data every time (since it uses next_batch(100). There are 1000*100 training data points in total. But all data points are considered only once individually. Am I correct or did I misunderstand something?<\/li>\n<li>If 3. is correct, is it weird that after just one update step the model had been trained?\nI think I must be misunderstanding something, It would be really great  if anyone can give me a hint or refer some material.<\/li>\n<\/ol>\n","answers":[{"AnswerId":"42438144","CreationDate":"2017-02-24T12:01:39.120","ParentId":null,"OwnerUserId":"3033441","Title":null,"Body":"<ol>\n<li>It updates weights every time you run the <code>train_step<\/code>.<\/li>\n<li>Yes, it is updating the weights 1000 times in this program.<\/li>\n<li>See above<\/li>\n<li>Yes, you are correct, it loads a mini-batch containing 100 points at once and uses it to compute gradients.<\/li>\n<li>It's not weird at all. You don't necessarily need to see the same data again and again, all that is required is that you have enough data for the network to converge. You can iterate multiple times over the same data if you want, but since this model doesn't have many parameters, it converges in a single epoch.<\/li>\n<\/ol>\n\n<p>Tensorflow works by creating a graph of the computations that are required for computing the output of a network. Each of the basic operations like matrix multiplication, addition, anything you can think of are nodes in this computation graph. In the tensorflow mnist example that you are following, the lines from 40-46 define the network architecture<\/p>\n\n<ul>\n<li>x: placeholder<\/li>\n<li>y_: placeholder<\/li>\n<li>W: Variable - This is learnt during training<\/li>\n<li>b: Variable - This is also learnt during training<\/li>\n<\/ul>\n\n<p>The network represents a simple linear regression model where the prediction is made using <code>y = W*x + b<\/code> (see line 43).<\/p>\n\n<p>Next, you configure the training procedure for your network. This code uses cross-entropy as the loss function to minimize (see line 57). The minimization is done using the gradient descent algorithm (see line 59).<\/p>\n\n<p>At this point, your network is fully constructed. Now you need to run these nodes so that actual computation if performed (no computation has been performed up till this point).<\/p>\n\n<p>In the loop where <code>sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})<\/code> is executed, tf computes the value of <code>train_step<\/code> which causes the GradientDescentOptimizer to try to minimize the <code>cross_entropy<\/code> and this is how training progresses.<\/p>\n"}]}
{"QuestionId":42437662,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-02-24T11:36:40.263","AcceptedAnswerId":42437970.0,"OwnerUserId":1556826.0,"Title":"Error in 2D Convolution in Keras","Body":"<p>I know this type of question has been asked many times here, but I could not figure out the answers from those questions. I have a 100x100 image in grayscale. I get the following error while trying to perform 2D Convolution in the very first layer. <\/p>\n\n<pre><\/pre>\n\n<p>I am getting an error like this: Error when checking model input: expected convolution2d_input_1 to have shape (None, 1, 100, 100) but got array with shape (32, 3, 100, 100). I am not sure where I am going wrong. <\/p>\n","answers":[{"AnswerId":"42437970","CreationDate":"2017-02-24T11:52:46.557","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Try:<\/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=16,\n    color_mode='grayscale',\n    class_mode='binary')\n\n    validation_generator = test_datagen.flow_from_directory(\n    test_data_dir,\n    target_size=(img_width, img_height),\n    batch_size=16,\n    color_mode='grayscale\n    class_mode='binary')\n<\/code><\/pre>\n"}]}
{"QuestionId":42438170,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-02-24T12:02:57.807","AcceptedAnswerId":42487319.0,"OwnerUserId":1978504.0,"Title":"Tensorflow: using a session\/graph in method","Body":"<p>\nMy situation is like this:<\/p>\n\n<p>I have a script that trains a tensorflow model. Inside this script I instantiate a class that feeds the training data. That classes's initialization in turn instantiates another class called \"image\" to do various operations for data augmentation and what not.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>My problem is that I'm trying to use tensorflow to do some operations within this image class by passing along either the session itself or the graph. But I've had little success.<\/p>\n\n<h2> Approach that works (but too slow) <\/h2>\n\n<p>What I have right now, but working painfully slow, is something like this (simplified):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This aproach works, but it creates a new Session for every image:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<h2> Approach that doesn't work (and should have been way faster) <\/h2>\n\n<p>What does not work, for example, and I can't figure out why, is to pass along the graph or session from my main script, like so:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This is the error I get:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Any help would be much appreciated!<\/p>\n","answers":[{"AnswerId":"42487319","CreationDate":"2017-02-27T13:43:59.163","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>How about instead of having an <code>Image<\/code> class you create an <code>ImageAugmenter<\/code> class that takes a session when initialized, and then processes your images using Tensorflow? You could do something like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\nimport numpy as np\n\nclass ImageAugmenter(object):\n    def __init__(self, sess):\n        self.sess = sess\n        self.im_placeholder = tf.placeholder(tf.float32, shape=[1,784,3])\n\n    def augment(self, image):\n        augment_op = tf.image.random_saturation(self.im_placeholder, 0.6, 0.8)\n        return self.sess.run(augment_op, {self.im_placeholder: image})\n\nclass DataFeed(object):\n    def __init__(self, data_dir, sess):\n        self.images = load_data(data_dir)\n        self.augmenter = ImageAugmenter(sess)\n\n    def process_data(self):\n        processed_images = []\n        for im in self.images:\n            processed_images.append(self.augmenter.augment(im))\n        return processed_images\n\ndef load_data(data_dir):\n    # True method would read images from disk\n    # This is just a mockup\n    images = []\n    images.append(np.random.random([1,784,3]))\n    images.append(np.random.random([1,784,3]))\n    return images\n\nif __name__ == \"__main__\":\n    TRAIN_DATA_DIR = '\/some\/dir\/'\n    sess = tf.Session()\n    data_feed = DataFeed(TRAIN_DATA_DIR, sess)\n    train_data = data_feed.process_data()\n    print(train_data)\n<\/code><\/pre>\n\n<p>With this you wouldn't be creating a new session for each image, and it should give you what you want.<\/p>\n\n<p>Note how <code>sess.run()<\/code> is called; The key I pass to its feed dict <em>is<\/em> the placeholder tensor defined above. According to your error trace, you are probably trying to call <code>sess.run()<\/code> from a part of your code where <code>im_placeholder<\/code> has not been defined, or it has been defined as other than a <code>tf.placeholder<\/code>.<\/p>\n\n<p>Additionally, you could further improve the code by changing the <code>ImageAugmenter.augment()<\/code> method to receive both the lower and upper parameters as input for the <code>tf.image.random_saturation()<\/code> method, or you could initialize the <code>ImageAugmenter<\/code> with a specific shape instead of having it hardcoded, for example.<\/p>\n"}]}
{"QuestionId":42439233,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-24T12:55:52.170","AcceptedAnswerId":42441617.0,"OwnerUserId":127480.0,"Title":"How to use freeze_graph.py tool in TensorFlow v1","Body":"<p>\nIs it possible to use the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"noreferrer\"><\/a> tool with models saved via  in TensorFlow v1? If so, how?<\/p>\n\n<p>I have code that looks roughly like this:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>This produces a directory containing:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Where  is a directory containing the files for intermediate training steps. The rest are files.<\/p>\n\n<p>My understanding is that this is a meta graph (the  file) and its variables (the  file) in saver v2 format. These files contain the data needed for the  tool but it is unclear how to tell the  tool to load the data from these files.<\/p>\n\n<p>All of these attempts produce the error message <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The  code includes the comment  next to where the  argument is used so I had thought the third of the above attempts would work but, alas, no.<\/p>\n","answers":[{"AnswerId":"42441617","CreationDate":"2017-02-24T14:51:23.373","ParentId":null,"OwnerUserId":"127480","Title":null,"Body":"<p>As @mrry pointed out in a comment, the answer to this particular question is to prefix the output prefix with <code>.\/<\/code>. When this was done I discovered it is also necessary to provide values for the <code>--input_graph<\/code> and <code>--output_name_names<\/code> arguments.<\/p>\n\n<p>The command now looks like<\/p>\n\n<pre><code>python freeze_graph.py \\\n    --input_graph output\/graph.pbtxt \\\n    --input_checkpoint .\/output-16640 \\\n    --output_graph \/tmp\/out \\\n    --output_node_names &lt;name&gt;\n<\/code><\/pre>\n\n<p>Unfortunately my graph contains variables for pre-loaded data which causes <code>freeze_graph.py<\/code> to fail with a message like <code>Attempting to use uninitialized value ...<\/code>; solving this subsequent problem is beyond the scope of this question.<\/p>\n"}]}
{"QuestionId":42439360,"AnswerCount":2,"Tags":"<python><numpy><casting><tensorflow><classification>","CreationDate":"2017-02-24T13:02:55.443","AcceptedAnswerId":42439638.0,"OwnerUserId":4834745.0,"Title":"How do I convert a numpy Array to a data type that tensorflow can classify?","Body":"<p>I am writing a Python program to detect the state of a chess board and I am using a sliding window to detect the position of each piece. My main program detects the chessboard within an image and passes its cropped picture to the my_sliding_window method. This is supposed to use Tensorflow to detect a piece in the sliding window. From <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/index.html?index=..%2F..%2Findex#0\" rel=\"nofollow noreferrer\">this<\/a> tutorial I saw that pictures are read like this:<\/p>\n\n<pre><\/pre>\n\n<p>But I don't want to read it from file as I already have the picture in a numpy array. How do make my numpy array in such a way that it can be classified by Tensorflow?<\/p>\n\n<p>Thank you.<\/p>\n\n<p>Code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"52049453","CreationDate":"2018-08-28T02:58:53.027","ParentId":null,"OwnerUserId":"8211166","Title":null,"Body":"<p>I solved the problem by using this one-liner:<\/p>\n\n<pre><code>image_data = cv2.imencode('.jpg', cv_image)[1].tostring()\n<\/code><\/pre>\n"},{"AnswerId":"42439638","CreationDate":"2017-02-24T13:18:05.340","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>You could use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/convert_to_tensor\" rel=\"nofollow noreferrer\"><code>tf.convert_to_tensor()<\/code><\/a> to convert your numpy array into a TensorFlow tensor:<\/p>\n\n<blockquote>\n  <p>This function converts Python objects of various types to Tensor objects. It accepts Tensor objects, numpy arrays, Python lists, and Python scalars.<\/p>\n<\/blockquote>\n\n<h2>Update<\/h2>\n\n<p>Ok, so what you're trying to do is to feed the numpy array <code>image_data<\/code>, with dimensions <code>[123, 82]<\/code> to the placeholder <code>DecodeJpeg\/contents:0<\/code>. However that placeholder was defined with <code>shape=()<\/code> meaning it only accepts 0D tensors as input (see <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/dims_types#shape\" rel=\"nofollow noreferrer\">tensor shapes<\/a>), hence throwing you an error.<\/p>\n\n<p>What the <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/index.html?index=..%2F..%2Findex#4\" rel=\"nofollow noreferrer\"><em>original code<\/em><\/a> does is to read an image as a dimensionless string with:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>image_data = tf.gfile.FastGFile(image_path, 'rb').read()\n<\/code><\/pre>\n\n<p>which is then fed to the <code>DecodeJpeg\/contents:0<\/code> placeholder in:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>predictions = sess.run(softmax_tensor, {'DecodeJpeg\/contents:0': image_data})\n<\/code><\/pre>\n\n<p>The easiest way to proceed and try to run your images through the pretrained graph would be to use the same <code>tf.gfile.FastGFile()<\/code> call for loading the images.<\/p>\n"}]}
{"QuestionId":42439735,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-24T13:23:07.213","AcceptedAnswerId":42440581.0,"OwnerUserId":7617153.0,"Title":"why the tf.Session() returns a wrong result?","Body":"<p>I'm a newbie to tensorflow and met the problem below. I'm very thankful for your answer.<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<p><strong>then the result is 0.0.<\/strong><\/p>\n","answers":[{"AnswerId":"42440581","CreationDate":"2017-02-24T14:01:27.240","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>How exactly did you define <code>sess<\/code>?<\/p>\n\n<p>I just ran your code within a <code>tf.Session()<\/code> context and it worked:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import tensorflow as tf\n\nwith tf.Session() as sess:\n    a = tf.placeholder(tf.float32)\n    b = tf.placeholder(tf.float32)\n    adder_node = a + b  \n\n    print(sess.run(adder_node, {a: 3, b: 4.5}))\n<\/code><\/pre>\n\n<p>Output: <code>7.5<\/code><\/p>\n"},{"AnswerId":"43306365","CreationDate":"2017-04-09T11:52:32.893","ParentId":null,"OwnerUserId":"4597696","Title":null,"Body":"<p>It can be a memory problem.<\/p>\n\n<p>If you're using notebooks, try stopping these notebooks.<\/p>\n\n<p>If you're using Docker-Nvidia, just try restarting the container:<\/p>\n\n<pre><code>nvidia-docker stop &lt;container-name&gt;\n<\/code><\/pre>\n\n<p>And then restart:<\/p>\n\n<pre><code>nvidia-docker start &lt;container-name&gt;\n<\/code><\/pre>\n"}]}
{"QuestionId":42439854,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-02-24T13:29:43.043","AcceptedAnswerId":null,"OwnerUserId":7617139.0,"Title":"Logistic regression: Bias and cost function in tensorflow","Body":"<p>The lectures Andrew Ng has said that <\/p>\n\n<p>J= -1\/m * sum (y*log(h(x))+(1-y)log(1-h(x)))<\/p>\n\n<p>but in <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners<\/a> they state:<\/p>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/13lb5.png\" alt=\"enter image description here\"><\/p>\n\n<p>Why did they use this formula?<\/p>\n","answers":[{"AnswerId":"42440404","CreationDate":"2017-02-24T13:54:04.220","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>That is the binary vs the categorical approach. Andrew Ng cost function is binary crossentropy (or logloss), while the one used in the tensorflow tutorial is the categorical crossentropy. They use one-hot vector encoding to use this accross multiple categories. So even for 2 categories you have a label like this one: [0, 1]. <\/p>\n\n<p>In the binary case 0 is the label for the false case. This does not exist in the categorical case where only the 1 matters, because of the softmax classifier, where all your elements need to sum up to 1 anyways.<\/p>\n"}]}
{"QuestionId":42440111,"AnswerCount":3,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-02-24T13:40:27.327","AcceptedAnswerId":null,"OwnerUserId":5434466.0,"Title":"How to find Tensorflow Serving version?","Body":"<p>To find Tensorflow version we can do that by:\npython -c 'import tensorflow as tf; print(tf.<strong>version<\/strong>)' <\/p>\n\n<p>Tensorflow Serving is a separate install, so how to find the version of Tensorflow Serving?<\/p>\n\n<p>Is it same as Tensorflow? Do not see any reference\/comments or documentation related to this.<\/p>\n","answers":[{"AnswerId":"56000139","CreationDate":"2019-05-06T06:54:26.213","ParentId":null,"OwnerUserId":"6778091","Title":null,"Body":"<p>After you have installed <code>tensorflow_model_server<\/code>, run this<\/p>\n\n<pre><code>tensorflow_model_server --version\n<\/code><\/pre>\n\n<p>you will get the version of tf-serving.\nIn my case, i get<\/p>\n\n<pre><code>TensorFlow ModelServer: 1.13.0-rc1+dev.sha.f16e777\nTensorFlow Library: 1.13.1\n<\/code><\/pre>\n"},{"AnswerId":"43431328","CreationDate":"2017-04-15T21:11:44.217","ParentId":null,"OwnerUserId":"4490749","Title":null,"Body":"<p>When building TF-Serving, you are going to do:<\/p>\n\n<p><code>git clone --recursive https:\/\/github.com\/tensorflow\/serving\ncd serving<\/code><\/p>\n\n<p>From there, you can submit:\n<code>git branch --list -a<\/code><\/p>\n\n<p>This will list all the possible TF-Serving versions. At the time of writing, I have:<\/p>\n\n<blockquote>\n  <p>remotes\/origin\/master<\/p>\n  \n  <p>remotes\/origin\/r0.5.1<\/p>\n<\/blockquote>\n\n<p>Then, you can checkout the branch you want before you build TF-Serving:\n<code>git checkout r0.5.1<\/code><\/p>\n"},{"AnswerId":"50768372","CreationDate":"2018-06-08T21:03:28.933","ParentId":null,"OwnerUserId":"7842888","Title":null,"Body":"<p>If one installs Tensor flow serving using <\/p>\n\n<ol>\n<li><code>apt-get<\/code>, we can find the installed version by using <code>apt list --installed<\/code><\/li>\n<li>For <code>pip<\/code> installation use <code>pip freeze<\/code><\/li>\n<li>For <code>conda<\/code> installation use <code>conda list<\/code><\/li>\n<\/ol>\n"}]}
{"QuestionId":42440274,"AnswerCount":1,"Tags":"<python><deep-learning><keras><convolution>","CreationDate":"2017-02-24T13:48:09.557","AcceptedAnswerId":42456188.0,"OwnerUserId":7103753.0,"Title":"Training on the merged layer in keras","Body":"<p>I am implementing following <a href=\"http:\/\/www.sciencedirect.com\/science\/article\/pii\/S1361841516300330\" rel=\"nofollow noreferrer\">this<\/a> paper by Mohammad Havaei. It uses following architecture:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Miu4O.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Miu4O.jpg\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I have modified some code from <a href=\"https:\/\/github.com\/naldeborgh7575\/brain_segmentation\/tree\/master\/code\" rel=\"nofollow noreferrer\">here<\/a> to do so.<\/p>\n\n<pre><\/pre>\n\n<p>I have used this alternate approach as Graph model  is deprecated in keras 1.0\nMy question is how do I train the model now?\nI have used this to train<\/p>\n\n<pre><\/pre>\n\n<p>In case I need to train separately two layers and then merge, how do I do that? <\/p>\n","answers":[{"AnswerId":"42456188","CreationDate":"2017-02-25T12:42:09.407","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<pre><code>from keras.layers import *\nfrom keras.models import Model\n\nprint 'Compiling two-path model...'\n\n# Input of the model\ninput_model = Input(shape=(4,33,33))\n# Local pathway\n#Add first convolution\nmodel_l = Convolution2D(64,7,7,\n                            border_mode='valid', \n                            activation='relu', \n                            W_regularizer=l1l2(l1=0.01, l2=0.01))(input_model)\nmodel_l = BatchNormalization(mode=0,axis=1)(model_l)\nmodel_l = MaxPooling2D(pool_size=(2,2),strides=(1,1))(model_l)\nmodel_l = Dropout(0.5)(model_l)\n#Add second convolution\nmodel_l = Convolution2D(64,3,3,\n                        border_mode='valid',\n                        W_regularizer=l1l2(l1=0.01, l2=0.01),\n                        input_shape=(4,33,33))(model_l)\nmodel_l = BatchNormalization(mode=0,axis=1)(model_l)\nmodel_l = MaxPooling2D(pool_size=(4,4),strides=(1,1))(model_l)\nmodel_l = Dropout(0.5)(model_l)\n\n#global pathway\nmodel_g = Convolution2D(160,12,12,\n                        border_mode='valid', \n                        activation='relu',\n                        W_regularizer=l1l2(l1=0.01, l2=0.01))(input_model)\nmodel_g = BatchNormalization(mode=0,axis=1)(model_g)\nmodel_g = MaxPooling2D(pool_size=(2,2), strides=(1,1))(model_g)\nmodel_g = Dropout(0.5)(model_g)\n\n# merge local and global pathways\n\nmerge = Merge(mode='concat', concat_axis=1)([model_l,model_g])\nmerge = Convolution2D(5,21,21,\n                      border_mode='valid',\n                      W_regularizer=l1l2(l1=0.01, l2=0.01))(merge)\nmerge = Flatten()(merge)\npredictions = Dense(5, activation='softmax')(merge)\n\nmodel_merged = Model(input=input_model,output=predictions)\nsgd = SGD(lr=0.001, decay=0.01, momentum=0.9)\nmodel_merged.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=['accuracy'])\n\nprint('Done')\nreturn model_merged\n<\/code><\/pre>\n\n<p>this is the equivalent of the network you posted but defined with the <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">Functional API<\/a><\/p>\n\n<p>As you can see, there is only 1 Input layer, used twice. You can then train it like you said :<\/p>\n\n<pre><code>model_merged.fit(X_train, Y_train, batch_size=self.batch_size, nb_epoch=self.n_epoch, validation_split=0.1, verbose=1)\n<\/code><\/pre>\n\n<p>does that help ?<\/p>\n"}]}
{"QuestionId":42440565,"AnswerCount":3,"Tags":"<python><tensorflow><lstm><recurrent-neural-network>","CreationDate":"2017-02-24T14:00:43.073","AcceptedAnswerId":null,"OwnerUserId":7071375.0,"Title":"How to feed back RNN output to input in tensorflow","Body":"<p>In case where suppose I have a trained RNN (e.g. language model), and I want to see what it would generate on its own, <strong>how should I feed its output back to its input?<\/strong><\/p>\n\n<p>I read the following related questions:<\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/36609920\/tensorflow-using-lstms-for-generating-text\">TensorFlow using LSTMs for generating text<\/a><\/p><\/li>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/34302042\/tensorflow-lstm-generative-model?rq=1\">TensorFlow LSTM Generative Model<\/a><\/p><\/li>\n<\/ul>\n\n<p>Theoretically it is clear to me, that in tensorflow we use truncated backpropagation, so we have to define the max step which we would like to \"trace\". Also we reserve a dimension for batches, therefore if I'd like to train a sine wave, I have to feed  inputs.<\/p>\n\n<p>The following code works:<\/p>\n\n<pre><\/pre>\n\n<p>The state size of the LSTM can be varied, also I experimented with feeding sine wave into the network and zeros, and in both cases it converged in ~500 iterations. So far I have understood that in this case the graph consists  number of LSTM cells sharing their parameters, and it is only up to me that I feed input to them as a <em>time series<\/em>. However when <strong>generating<\/strong> samples the network is explicitly depending on its previous output - meaning that I cannot feed the unrolled model at once. I tried to compute the state and output at every step:<\/p>\n\n<pre><\/pre>\n\n<p>However in this model it seems that there is no continuity between the LSTM cells. What is going on here?<\/p>\n\n<p>Do I have to initialize a zero array with i.e. 100 time steps, and assign each run's result into the array? Like feeding the network with this:<\/p>\n\n<p>run 0: <\/p>\n\n<p>run 1: <\/p>\n\n<p>run 1: <\/p>\n\n<p>etc...<\/p>\n\n<p><strong>What to do if I want to use this trained network to use its own output as its input in the following time step?<\/strong><\/p>\n","answers":[{"AnswerId":"54486072","CreationDate":"2019-02-01T19:32:28.340","ParentId":null,"OwnerUserId":"6037060","Title":null,"Body":"<p>I know I'm a bit late to the party but I think this gist could be useful:<\/p>\n\n<p><a href=\"https:\/\/gist.github.com\/CharlieCodex\/f494b27698157ec9a802bc231d8dcf31\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/CharlieCodex\/f494b27698157ec9a802bc231d8dcf31<\/a><\/p>\n\n<p>It lets you autofeed the input through a filter and back into the network as input. To make shapes match up <code>processing<\/code> can be set as a <code>tf.layers.Dense<\/code> layer.<\/p>\n\n<p>Please ask any questions!<\/p>\n\n<p>Edit:<\/p>\n\n<p>In your particular case, create a lambda which performs the processing of the <code>dynamic_rnn<\/code> outputs into your character vector space. Ex:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># if you have:\nW = tf.Variable( ... )\nB = tf.Variable( ... )\nYo, Ho = tf.nn.dynamic_rnn( cell , inputs , state )\nlogits = tf.matmul(W, Yo) + B\n ...\n# use self_feeding_rnn as\nprocess_yo = lambda Yo: tf.matmul(W, Yo) + B\nYo, Ho = self_feeding_rnn( cell, seed, initial_state, processing=process_yo)\n<\/code><\/pre>\n"},{"AnswerId":"42444842","CreationDate":"2017-02-24T17:32:12.743","ParentId":null,"OwnerUserId":"1601927","Title":null,"Body":"<p>You can use its own output (last state) as the next-step input (initial state).\nOne way to do this is to:<\/p>\n\n<ol>\n<li>use zero-initialized variables as the input state at every time step<\/li>\n<li>each time you completed a truncated sequence and got some output state, update the state variables with this output state you just got.<\/li>\n<\/ol>\n\n<p>The second can be done by either:<\/p>\n\n<ol>\n<li>fetching the states to python and feeding them back next time, as done in the <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/ptb\" rel=\"nofollow noreferrer\">ptb example in tensorflow\/models<\/a><\/li>\n<li>build an update op in the graph and add a dependency, as done in the <a href=\"https:\/\/github.com\/ppwwyyxx\/tensorpack\/blob\/master\/examples\/PennTreebank\/PTB-LSTM.py#L90\" rel=\"nofollow noreferrer\">ptb example in tensorpack<\/a>.<\/li>\n<\/ol>\n"},{"AnswerId":"47888477","CreationDate":"2017-12-19T13:51:34.363","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>If I understood you correctly, you want to find a way to feed the output of time step <code>t<\/code> as input to time step <code>t+1<\/code>, right? To do so, there is a relatively easy work around that you can use at <em>test time<\/em>:<\/p>\n\n<ol>\n<li>Make sure your input placeholders can accept a dynamic sequence length, i.e. the size of the time dimension is <code>None<\/code>.<\/li>\n<li>Make sure you are using <code>tf.nn.dynamic_rnn<\/code> (which you do in the posted example).<\/li>\n<li>Pass the initial state into <code>dynamic_rnn<\/code>.<\/li>\n<li>Then, at test time, you can loop through your sequence and feed each time step individually (i.e. max sequence length is 1). Additionally, you just have to carry over the internal state of the RNN. See pseudo code below (the variable names refer to your code snippet).<\/li>\n<\/ol>\n\n<p>I.e., change the definition of the model to something like this:<\/p>\n\n<pre><code>lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(state_size, forget_bias=1.)\nX = tf.placeholder_with_default(zero_x, [None, None, 1])  # [batch_size, seq_length, dimension of input]\nbatch_size = tf.shape(self.input_)[0]\ninitial_state = lstm_cell.zero_state(batch_size, dtype=tf.float32)\ndef_x = np.sin(np.linspace(0, 10, n_samples))[None, :, None]\nzero_x = np.zeros(n_samples)[None, :, None]\noutput, last_states = tf.nn.dynamic_rnn(inputs=X, cell=lstm_cell, dtype=tf.float64,\n    initial_state=initial_state)\npred = tf.contrib.layers.fully_connected(output, 1, activation_fn=tf.tanh)\n<\/code><\/pre>\n\n<p>Then you can perform inference like so:<\/p>\n\n<pre><code>fetches = {'final_state': last_state,\n           'prediction': pred}\n\ntoy_initial_input = np.array([[[1]]])  # put suitable data here\nseq_length = 20  # put whatever is reasonable here for you\n\n# get the output for the first time step\nfeed_dict = {X: toy_initial_input}\neval_out = sess.run(fetches, feed_dict)\noutputs = [eval_out['prediction']]\nnext_state = eval_out['final_state']\n\nfor i in range(1, seq_length):\n    feed_dict = {X: outputs[-1],\n                 initial_state: next_state}\n    eval_out = sess.run(fetches, feed_dict)\n    outputs.append(eval_out['prediction'])\n    next_state = eval_out['final_state']\n\n# outputs now contains the sequence you want\n<\/code><\/pre>\n\n<p>Note that this can also work for batches, however it can be a bit more complicated if you sequences of different lengths in the same batch.<\/p>\n\n<p>If you want to perform this kind of prediction not only at test time, but also at training time, it is also possible to do, but a bit more complicated to implement.<\/p>\n"}]}
{"QuestionId":42440681,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-24T14:06:10.770","AcceptedAnswerId":null,"OwnerUserId":2200571.0,"Title":"Tensorflow GraphDef cannot be larger than 2GB, even though it is not","Body":"<p>I have been getting this error when I try to save and restore a model.<\/p>\n\n<p>I have played around and changed the structure enough to reduce the size of the model to 1.2GB yet the error persists:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea why this might occur?<\/p>\n\n<p>Some code:<\/p>\n\n<pre><\/pre>\n\n<p>Trace of error:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42441393,"AnswerCount":0,"Tags":"<input><tensorflow><boolean>","CreationDate":"2017-02-24T14:41:23.390","AcceptedAnswerId":null,"OwnerUserId":7617408.0,"Title":"Tensorflow: learn evaluation of boolean features","Body":"<p>I want to train a model with tensorflow that has boolean features like this:<\/p>\n\n<pre><\/pre>\n\n<p>Looks like an easy task to me, but it turns out, it is no easy task for me. I have no idea how to do this, can not find anything similar (besides <a href=\"https:\/\/stackoverflow.com\/questions\/38443395\/read-boolean-features-for-tensorflow\" title=\"this\">this<\/a> ) on the web and I am not able to adjust one of the tensorflow examples to my needs.<\/p>\n\n<p>O.K. Here comes the code...<\/p>\n\n<pre><\/pre>\n\n<ol>\n<li>Running this code as shown yields an Accuracy: 0.000000<\/li>\n<li>When I use data in test_set that are contained in training_set (e.g.   ) I get an Accuracy: 0.500000 <\/li>\n<li>When I exchange the LinearClassifier with a DNNClassifier (see comments in code), I get Accuracy: 0.000000 for the disjunct test_set (see 1.) <\/li>\n<li>For contained test_set (see 2.) I get Accuracy: 1.000000<\/li>\n<\/ol>\n\n<p>I do not understand these results :-o\nI had expected an Accuracy of some 0.something on first step that increases with each step to approach an Accuracy of 1.0. And I expected this approach to be slower for disjunct trainings- and test-set.<\/p>\n\n<p>The parameter  in  does not seem to have any effect.<\/p>\n","answers":[]}
{"QuestionId":42441431,"AnswerCount":1,"Tags":"<tensorflow><keras><keras-layer>","CreationDate":"2017-02-24T14:42:49.530","AcceptedAnswerId":42446706.0,"OwnerUserId":2106858.0,"Title":"How to set the input of a Keras layer with a Tensorflow tensor?","Body":"<p>In my <a href=\"https:\/\/stackoverflow.com\/questions\/41672114\/add-tensorflow-pre-processing-to-existing-keras-model-for-use-in-tensorflow-ser\">previous question<\/a>, I used Keras'  to connect my Tensorflow pre-processing output tensor to my Keras model's input. However, <a href=\"https:\/\/github.com\/fchollet\/keras\/commit\/92e8a20761bedbde8fd56a02a165884e8132f045\" rel=\"noreferrer\">this method has been removed<\/a> after Keras version .<\/p>\n\n<p>How can I achieve this in newer Keras versions?<\/p>\n\n<p><strong>Example:<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42446706","CreationDate":"2017-02-24T19:23:35.090","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>After you are done with pre-processing, You can add the tensor as input layer by calling <code>tensor<\/code> param of <code>Input<\/code><\/p>\n\n<p>So in your case:<\/p>\n\n<pre><code>tf_embedding_input = ...    # pre-processing output tensor\n\n# Keras model\nmodel = Sequential()\nmodel.add(Input(tensor=tf_embedding_input)) \nmodel.add(Embedding(max_features, 128, input_length=maxlen))\n<\/code><\/pre>\n"}]}
{"QuestionId":42442869,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T15:50:41.870","AcceptedAnswerId":null,"OwnerUserId":3042790.0,"Title":"No module named layers","Body":"<p>I'm using tensorflow 1.0.0 and I want to access the tensorflow.layers module. The module seems to exist:<\/p>\n\n<pre><\/pre>\n\n<p>But when I try to import e.g. the dense function:<\/p>\n\n<pre><\/pre>\n\n<p>Why can python not access the module?<\/p>\n","answers":[{"AnswerId":"42448388","CreationDate":"2017-02-24T21:13:39.147","ParentId":null,"OwnerUserId":"121687","Title":null,"Body":"<p>Something is wrong with your installation or workspace:<\/p>\n\n<ul>\n<li>Make sure you don\u2019t have a directory called \u2018tensorflow\u201d in your Python Path. <\/li>\n<li>Install again the official tensorflow distro <code>pip install \u2014upgrade \u2014ignore-installed tensorflow)<\/code><\/li>\n<li>Make sure you are using the right tensorflow version <code>print(tensorflow.__version__)<\/code><\/li>\n<\/ul>\n"}]}
{"QuestionId":42442946,"AnswerCount":3,"Tags":"<python><windows><tensorflow>","CreationDate":"2017-02-24T15:54:18.673","AcceptedAnswerId":42482683.0,"OwnerUserId":774133.0,"Title":"TensorFlow installation: 'Native pip' on Windows","Body":"<p>I am desperately trying to install TensorFlow (and Keras) on a Windows 10 machine (64bit). I was able to have a working Keras installation on a Ubuntu Linux, but now need to have it working on Windows.<\/p>\n\n<p>I am following the official instructions here: <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a>\nspecifically those at \"Installing with native pip\".<\/p>\n\n<p>First, I downloaded python 3.5 from python.org. <\/p>\n\n<pre><\/pre>\n\n<p>Then I run . CUDA 8, cuDNN 5.1 and all the other requirements are satisfied.<\/p>\n\n<p>Now if I ask some details:<\/p>\n\n<pre><\/pre>\n\n<p>If I test it:<\/p>\n\n<pre><\/pre>\n\n<p>I have no error messages as a consequence of the . But then:<\/p>\n\n<pre><\/pre>\n\n<p>I have read the answers to other questions here on StackOverflow but could not find any pointer to a solution. In some cases, people had the same problem when using a file named , but I checked this and could not find any script with that filename.<\/p>\n\n<p>It was very easy under Linux and I thought it could not be this hard on Windows, but I was wrong.<\/p>\n\n<p>I followed the instructions... what did I miss? <\/p>\n\n<hr>\n\n<h1>Answers to comments.<\/h1>\n\n<p>?<\/p>\n\n<p>Note I am not running a script but using the prompt - nothing changes if I use a script.<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<h2>Update<\/h2>\n\n<p>I uninstalled tensorflow with the following command:<\/p>\n\n<pre><\/pre>\n\n<p>Now if I import , I receive an error:<\/p>\n\n<pre><\/pre>\n\n<p>So this behaviour should exclude the fact that another  file was causing the error I described. Am I right?<\/p>\n\n<hr>\n\n<p>Next, I installed the module  again:<\/p>\n\n<pre><\/pre>\n\n<p>So, now it seems to be working. The entire procedure was:<\/p>\n\n<ul>\n<li>install required software and libraries<\/li>\n<li>install  from python.org<\/li>\n<li>install  natively (using )<\/li>\n<li>uninstall <\/li>\n<li>install <\/li>\n<\/ul>\n\n<p>May be a python expert can understand what happened...<\/p>\n\n<hr>\n\n<p><em>PS. Yesterday I tried to install it using Anaconda, but had problems in subsequent steps when trying to install the Keras required packages. On Linux I did not have any problem with Keras either.<\/em><\/p>\n","answers":[{"AnswerId":"42462865","CreationDate":"2017-02-25T23:26:50.470","ParentId":null,"OwnerUserId":"2125449","Title":null,"Body":"<p>I had the same problem. To solve it, I made a brand new installation of python 3.5.2 &amp; tensorfow (pip3 install --upgrade tensorflow).\nI then write a helloMundo.py with the following content:<\/p>\n\n<pre><code>import tensorflow as tf\n# Simple hello world using TensorFlow\n# The value returned by the constructor represents the output\n# of the Constant op.\nhello = tf.constant('Hello, TensorFlow gr!')\n# Start tf session\nsess = tf.Session()\n# Run the op\nprint(sess.run(hello))} \n<\/code><\/pre>\n\n<p>I open the helloMundo.py from the python IDLE, then I open the menu Run -> Run Module (or F5) and it works.<\/p>\n\n<p>Try to put your code in a .py file.<\/p>\n\n<p>Regards,\nF. Kubha<\/p>\n"},{"AnswerId":"42443539","CreationDate":"2017-02-24T16:24:18.863","ParentId":null,"OwnerUserId":"4504245","Title":null,"Body":"<p>This might sound odd, but are any of your folders or script named tensorflow?<\/p>\n\n<pre><code>tensorflow\/script.py\n<\/code><\/pre>\n\n<p>or  <\/p>\n\n<pre><code>dir\/tensorflow.py\n<\/code><\/pre>\n\n<p>try changing those to not <code>tensorflow<\/code> and try again.<br>\nIf that did not resolve things let me know<\/p>\n\n<p>reasoning behind this and why I asked to see the output of <code>dir(tf)<\/code>:<\/p>\n\n<p>Here is the order in which a module will be searched when importing:<\/p>\n\n<blockquote>\n  <ol>\n  <li><p>The directory containing the input script (or the current directory when no file is specified). <\/p><\/li>\n  <li><p>PYTHONPATH (a list of directory names,\n  with the same syntax as the shell variable PATH). <\/p><\/li>\n  <li><p>The installation-dependent default.<\/p><\/li>\n  <\/ol>\n<\/blockquote>\n"},{"AnswerId":"42482683","CreationDate":"2017-02-27T09:58:10.730","ParentId":null,"OwnerUserId":"774133","Title":null,"Body":"<p>The solution was to uninstall and re-install <code>tensorflow-gpu<\/code>. Now it seems to be working.<\/p>\n\n<p>It does not find <code>cuDNN<\/code>, but at least the test script runs.<\/p>\n"}]}
{"QuestionId":42443009,"AnswerCount":1,"Tags":"<c++><tensorflow><eigen>","CreationDate":"2017-02-24T15:58:35.587","AcceptedAnswerId":42444525.0,"OwnerUserId":2412846.0,"Title":"C++ use of Eigen in tensorflow","Body":"<p>What is the relation between tensorflow and Eigen, particularly regarding the  datastructures?<\/p>\n\n<p>There are some older quotations (e.g. <a href=\"https:\/\/www.google.co.uk\/url?sa=t&amp;rct=j&amp;q=&amp;esrc=s&amp;source=web&amp;cd=3&amp;cad=rja&amp;uact=8&amp;ved=0ahUKEwj1mqPhiqnSAhUhBMAKHdMBCqMQFggrMAI&amp;url=http%3A%2F%2Fdownload.tensorflow.org%2Fpaper%2Fwhitepaper2015.pdf&amp;usg=AFQjCNEAFa6vGrlp_CKFlVxs0EHLsyhVtQ&amp;bvm=bv.147448319,bs.1,d.d2s\" rel=\"noreferrer\">here<\/a>) which state that tensorflow is using Eigen extensively (afaik a tensorflow guy has extended the Eigen code). More recent tensorflow documentation, however, seems to not explicitly refer to Eigen.<\/p>\n\n<p>Are the two  structures identical? Are they being updated concurrently? Is there any (possibly future) disadvantage in using the  over the ?<\/p>\n","answers":[{"AnswerId":"42444525","CreationDate":"2017-02-24T17:13:58.367","ParentId":null,"OwnerUserId":"1641621","Title":null,"Body":"<p><code>tensorflow::tensor<\/code> is just a thin wrapper around <code>Eigen::Tensor<\/code> with limited high-level features only. You can access to the underlying <code>Eigen::Tensor<\/code> of a <code>tensorflow::tensor<\/code> using the <code>tensor()<\/code> method. I guess that this information also solve your last two questions. <\/p>\n"}]}
{"QuestionId":42443104,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><keras><recurrent-neural-network>","CreationDate":"2017-02-24T16:02:55.033","AcceptedAnswerId":42443375.0,"OwnerUserId":2863057.0,"Title":"One to many LSTM in Keras","Body":"<p>Is it possible to implement a one-to-many LSTM in Keras?<br>\nIf yes, can you please provide me with a simple example?<\/p>\n","answers":[{"AnswerId":"42443375","CreationDate":"2017-02-24T16:16:24.873","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>It's possible with a <code>RepeatVector<\/code> layer. For example:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(10, input_shape=(1))\nmodel.add(RepeatVector(10))\nmodel.add(LSTM(1, return_sequences=True))\n<\/code><\/pre>\n\n<p>Then - the input shape is <code>(1)<\/code> and the output is <code>(10, 1)<\/code>.<\/p>\n"}]}
{"QuestionId":42443467,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-02-24T16:21:27.567","AcceptedAnswerId":42443594.0,"OwnerUserId":5538474.0,"Title":"How to print the values of tensors inside a while loop?","Body":"<p>I'm very new to tensorflow, and I couldn't figure this one out.<\/p>\n\n<p>I have this while loop:<\/p>\n\n<pre><\/pre>\n\n<p>And I'm evaluating it this way:<\/p>\n\n<pre><\/pre>\n\n<p>I want to check the value of  inside the while loop, but  doesn't seem to work, and  just tells me it's a tensor and gives me its shape. \nHow do I go about doing this?<\/p>\n\n<p>Thank you so much!<\/p>\n","answers":[{"AnswerId":"42443594","CreationDate":"2017-02-24T16:26:38.327","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Take a look at this webpage: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Print\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Print<\/a><\/p>\n\n<p>You can see that tf.Print is an identity operator with the side effect of printing data when evaluating. You should therefore use this line to print: <\/p>\n\n<p><code>reprs = tf.Print(reprs, [reprs])<\/code><\/p>\n\n<p>Hope this helps, and good luck!<\/p>\n"},{"AnswerId":"48420247","CreationDate":"2018-01-24T10:31:19.110","ParentId":null,"OwnerUserId":"3309610","Title":null,"Body":"<p>The approach suggested by rmeertens is the one I think is correct. I would just add (as a response to your comments) that if something is printing \"Tensor(\"while\/update_parent:0, ...... \" then that implies that that value in the graph is not being evaluated. <\/p>\n\n<p>You are likely seeing that as the output of your \"print(r.eval())\" statement, NOT the tf.Print() statement.<\/p>\n\n<p>Note that the output of tf.Print() appears in PyCharm (the IDE I am using) in red, while the output of a normal python print operation appears in black. So the tf.Print() output looks like a warning message. It could be that it is indeed printing out, but you are simply overlooking it.  <\/p>\n"}]}
{"QuestionId":42443665,"AnswerCount":0,"Tags":"<python><image><image-processing><theano>","CreationDate":"2017-02-24T16:29:57.180","AcceptedAnswerId":null,"OwnerUserId":7618017.0,"Title":"Python Script iGAN theano Error","Body":"<p>I'm trying to use software called iGAN created by junyanz from Berkeley. <a href=\"https:\/\/people.eecs.berkeley.edu\/~junyanz\/projects\/averageExplorer\/\" rel=\"nofollow noreferrer\">https:\/\/people.eecs.berkeley.edu\/~junyanz\/projects\/averageExplorer\/<\/a><\/p>\n\n<p><a href=\"https:\/\/github.com\/junyanz\/iGAN\/blob\/master\/README.md\" rel=\"nofollow noreferrer\">iGAN - Github<\/a><\/p>\n\n<p>I've installed the necessary dependencies but I seem to get stuck when I run the script. My goal is average a few thousand photos but first I need to get the script running. <\/p>\n\n<pre><\/pre>\n\n<p>I'm a photographer by trade and tinker with code as a hobby. Any help is greatly appreciated. <\/p>\n","answers":[]}
{"QuestionId":42443936,"AnswerCount":2,"Tags":"<keras><train-test-split>","CreationDate":"2017-02-24T16:43:23.863","AcceptedAnswerId":null,"OwnerUserId":3898714.0,"Title":"Keras split train test set when using ImageDataGenerator","Body":"<p>I have a single directory which contains sub-folders (according to labels) of images. I want to split this data into train and test set while using ImageDataGenerator in Keras. Although model.fit() in keras has argument validation_split for specifying the split, I could not find the same for model.fit_generator(). How to do it ?<\/p>\n\n<pre><\/pre>\n\n<p>I don't have separate directory for validation data, need to split it from the training data<\/p>\n","answers":[{"AnswerId":"52372042","CreationDate":"2018-09-17T16:21:41.200","ParentId":null,"OwnerUserId":"2835460","Title":null,"Body":"<p>Keras has now added Train \/ validation split from a single directory using ImageDataGenerator:<\/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    validation_split=0.2) # set validation split\n\ntrain_generator = train_datagen.flow_from_directory(\n    train_data_dir,\n    target_size=(img_height, img_width),\n    batch_size=batch_size,\n    class_mode='binary',\n    subset='training') # set as training data\n\nvalidation_generator = train_datagen.flow_from_directory(\n    train_data_dir, # same directory as training data\n    target_size=(img_height, img_width),\n    batch_size=batch_size,\n    class_mode='binary',\n    subset='validation') # set as validation data\n\nmodel.fit_generator(\n    train_generator,\n    steps_per_epoch = train_generator.samples \/\/ batch_size,\n    validation_data = validation_generator, \n    validation_steps = validation_generator.samples \/\/ batch_size,\n    epochs = nb_epochs)\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"noreferrer\">https:\/\/keras.io\/preprocessing\/image\/<\/a><\/p>\n"},{"AnswerId":"43220610","CreationDate":"2017-04-05T02:04:52.963","ParentId":null,"OwnerUserId":"690430","Title":null,"Body":"<p>I have a <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/6152\" rel=\"noreferrer\">PR<\/a> for it. One way is to hash the filenames and do a variant assignment.<\/p>\n\n<p>Example:<\/p>\n\n<pre><code># -*- coding: utf-8 -*-\n\"\"\"Train model using transfer learning.\"\"\"\nimport os\nimport re\nimport glob\nimport hashlib\nimport argparse\nimport warnings\n\nimport six\nimport numpy as np\nimport tensorflow as tf\nfrom tensorflow.python.platform import gfile\nfrom keras.models import Model\nfrom keras import backend as K\nfrom keras.optimizers import SGD\nfrom keras.layers import Dense, GlobalAveragePooling2D, Input\nfrom keras.applications.inception_v3 import InceptionV3\nfrom keras.preprocessing.image import (ImageDataGenerator, Iterator,\n                                       array_to_img, img_to_array, load_img)\nfrom keras.callbacks import ModelCheckpoint, TensorBoard, EarlyStopping\n\nRANDOM_SEED = 0\nMAX_NUM_IMAGES_PER_CLASS = 2 ** 27 - 1  # ~134M\nVALID_IMAGE_FORMATS = frozenset(['jpg', 'jpeg', 'JPG', 'JPEG'])\n# we chose to train the top 2 inception blocks\nBATCH_SIZE = 100\nTRAINABLE_LAYERS = 172\nINCEPTIONV3_BASE_LAYERS = len(InceptionV3(weights=None, include_top=False).layers)\n\nSTEPS_PER_EPOCH = 625\nVALIDATION_STEPS = 100\nMODEL_INPUT_WIDTH = 299\nMODEL_INPUT_HEIGHT = 299\nMODEL_INPUT_DEPTH = 3\nFC_LAYER_SIZE = 1024\n\n# Helper: Save the model.\ncheckpointer = ModelCheckpoint(\n    filepath='.\/output\/checkpoints\/inception.{epoch:03d}-{val_loss:.2f}.hdf5',\n    verbose=1,\n    save_best_only=True)\n\n# Helper: Stop when we stop learning.\nearly_stopper = EarlyStopping(patience=10)\n\n# Helper: TensorBoard\ntensorboard = TensorBoard(log_dir='.\/output\/')\n\n\ndef as_bytes(bytes_or_text, encoding='utf-8'):\n    \"\"\"Converts bytes or unicode to `bytes`, using utf-8 encoding for text.\n\n    # Arguments\n        bytes_or_text: A `bytes`, `str`, or `unicode` object.\n        encoding: A string indicating the charset for encoding unicode.\n\n    # Returns\n        A `bytes` object.\n\n    # Raises\n        TypeError: If `bytes_or_text` is not a binary or unicode string.\n    \"\"\"\n    if isinstance(bytes_or_text, six.text_type):\n        return bytes_or_text.encode(encoding)\n    elif isinstance(bytes_or_text, bytes):\n        return bytes_or_text\n    else:\n        raise TypeError('Expected binary or unicode string, got %r' %\n                        (bytes_or_text,))\n\n\nclass CustomImageDataGenerator(ImageDataGenerator):\n    def flow_from_image_lists(self, image_lists,\n                              category, image_dir,\n                              target_size=(256, 256), color_mode='rgb',\n                              class_mode='categorical',\n                              batch_size=32, shuffle=True, seed=None,\n                              save_to_dir=None,\n                              save_prefix='',\n                              save_format='jpeg'):\n        return ImageListIterator(\n            image_lists, self,\n            category, image_dir,\n            target_size=target_size, color_mode=color_mode,\n            class_mode=class_mode,\n            data_format=self.data_format,\n            batch_size=batch_size, shuffle=shuffle, seed=seed,\n            save_to_dir=save_to_dir,\n            save_prefix=save_prefix,\n            save_format=save_format)\n\n\nclass ImageListIterator(Iterator):\n    \"\"\"Iterator capable of reading images from a directory on disk.\n\n    # Arguments\n        image_lists: Dictionary of training images for each label.\n        image_data_generator: Instance of `ImageDataGenerator`\n            to use for random transformations and normalization.\n        target_size: tuple of integers, dimensions to resize input images to.\n        color_mode: One of `\"rgb\"`, `\"grayscale\"`. Color mode to read images.\n        classes: Optional list of strings, names of sudirectories\n            containing images from each class (e.g. `[\"dogs\", \"cats\"]`).\n            It will be computed automatically if not set.\n        class_mode: Mode for yielding the targets:\n            `\"binary\"`: binary targets (if there are only two classes),\n            `\"categorical\"`: categorical targets,\n            `\"sparse\"`: integer targets,\n            `None`: no targets get yielded (only input images are yielded).\n        batch_size: Integer, size of a batch.\n        shuffle: Boolean, whether to shuffle the data between epochs.\n        seed: Random seed for data shuffling.\n        data_format: String, one of `channels_first`, `channels_last`.\n        save_to_dir: Optional directory where to save the pictures\n            being yielded, in a viewable format. This is useful\n            for visualizing the random transformations being\n            applied, for debugging purposes.\n        save_prefix: String prefix to use for saving sample\n            images (if `save_to_dir` is set).\n        save_format: Format to use for saving sample images\n            (if `save_to_dir` is set).\n    \"\"\"\n\n    def __init__(self, image_lists, image_data_generator,\n                 category, image_dir,\n                 target_size=(256, 256), color_mode='rgb',\n                 class_mode='categorical',\n                 batch_size=32, shuffle=True, seed=None,\n                 data_format=None,\n                 save_to_dir=None, save_prefix='', save_format='jpeg'):\n        if data_format is None:\n            data_format = K.image_data_format()\n\n        classes = list(image_lists.keys())\n        self.category = category\n        self.num_class = len(classes)\n        self.image_lists = image_lists\n        self.image_dir = image_dir\n\n        how_many_files = 0\n        for label_name in classes:\n            for _ in self.image_lists[label_name][category]:\n                how_many_files += 1\n\n        self.samples = how_many_files\n        self.class2id = dict(zip(classes, range(len(classes))))\n        self.id2class = dict((v, k) for k, v in self.class2id.items())\n        self.classes = np.zeros((self.samples,), dtype='int32')\n\n        self.image_data_generator = image_data_generator\n        self.target_size = tuple(target_size)\n        if color_mode not in {'rgb', 'grayscale'}:\n            raise ValueError('Invalid color mode:', color_mode,\n                             '; expected \"rgb\" or \"grayscale\".')\n        self.color_mode = color_mode\n        self.data_format = data_format\n        if self.color_mode == 'rgb':\n            if self.data_format == 'channels_last':\n                self.image_shape = self.target_size + (3,)\n            else:\n                self.image_shape = (3,) + self.target_size\n        else:\n            if self.data_format == 'channels_last':\n                self.image_shape = self.target_size + (1,)\n            else:\n                self.image_shape = (1,) + self.target_size\n\n        if class_mode not in {'categorical', 'binary', 'sparse', None}:\n            raise ValueError('Invalid class_mode:', class_mode,\n                             '; expected one of \"categorical\", '\n                             '\"binary\", \"sparse\", or None.')\n        self.class_mode = class_mode\n        self.save_to_dir = save_to_dir\n        self.save_prefix = save_prefix\n        self.save_format = save_format\n\n        i = 0\n        self.filenames = []\n        for label_name in classes:\n            for j, _ in enumerate(self.image_lists[label_name][category]):\n                self.classes[i] = self.class2id[label_name]\n                img_path = get_image_path(self.image_lists,\n                                          label_name,\n                                          j,\n                                          self.image_dir,\n                                          self.category)\n                self.filenames.append(img_path)\n                i += 1\n\n        print(\"Found {} {} files\".format(len(self.filenames), category))\n        super(ImageListIterator, self).__init__(self.samples, batch_size, shuffle,\n                                                seed)\n\n    def next(self):\n        \"\"\"For python 2.x.\n\n        # Returns\n            The next batch.\n        \"\"\"\n        with self.lock:\n            index_array, current_index, current_batch_size = next(\n                self.index_generator)\n        # The transformation of images is not under thread lock\n        # so it can be done in parallel\n        batch_x = np.zeros((current_batch_size,) + self.image_shape,\n                           dtype=K.floatx())\n        grayscale = self.color_mode == 'grayscale'\n        # build batch of image data\n        for i, j in enumerate(index_array):\n            img = load_img(self.filenames[j],\n                           grayscale=grayscale,\n                           target_size=self.target_size)\n            x = img_to_array(img, data_format=self.data_format)\n            x = self.image_data_generator.random_transform(x)\n            x = self.image_data_generator.standardize(x)\n            batch_x[i] = x\n        # optionally save augmented images to disk for debugging purposes\n        if self.save_to_dir:\n            for i in range(current_batch_size):\n                img = array_to_img(batch_x[i], self.data_format, scale=True)\n                fname = '{prefix}_{index}_{hash}.{format}'.format(\n                    prefix=self.save_prefix,\n                    index=current_index + i,\n                    hash=np.random.randint(10000),\n                    format=self.save_format)\n                img.save(os.path.join(self.save_to_dir, fname))\n        # build batch of labels\n        if self.class_mode == 'sparse':\n            batch_y = self.classes[index_array]\n        elif self.class_mode == 'binary':\n            batch_y = self.classes[index_array].astype(K.floatx())\n        elif self.class_mode == 'categorical':\n            batch_y = np.zeros((len(batch_x), self.num_class),\n                               dtype=K.floatx())\n            for i, label in enumerate(self.classes[index_array]):\n                batch_y[i, label] = 1.\n        else:\n            return batch_x\n        return batch_x, batch_y\n\n\n# https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\ndef create_image_lists(image_dir, validation_pct=10):\n    \"\"\"Builds a list of training images from the file system.\n\n    Analyzes the sub folders in the image directory, splits them into stable\n    training, testing, and validation sets, and returns a data structure\n    describing the lists of images for each label and their paths.\n\n    # Arguments\n        image_dir: string path to a folder containing subfolders of images.\n        validation_pct: integer percentage of images reserved for validation.\n\n    # Returns\n        dictionary of label subfolder, with images split into training\n        and validation sets within each label.\n    \"\"\"\n    if not os.path.isdir(image_dir):\n        raise ValueError(\"Image directory {} not found.\".format(image_dir))\n    image_lists = {}\n    sub_dirs = [x[0] for x in os.walk(image_dir)]\n    sub_dirs_without_root = sub_dirs[1:]  # first element is root directory\n    for sub_dir in sub_dirs_without_root:\n        file_list = []\n        dir_name = os.path.basename(sub_dir)\n        if dir_name == image_dir:\n            continue\n        print(\"Looking for images in '{}'\".format(dir_name))\n        for extension in VALID_IMAGE_FORMATS:\n            file_glob = os.path.join(image_dir, dir_name, '*.' + extension)\n            file_list.extend(glob.glob(file_glob))\n        if not file_list:\n            warnings.warn('No files found')\n            continue\n        if len(file_list) &lt; 20:\n            warnings.warn('Folder has less than 20 images, which may cause '\n                          'issues.')\n        elif len(file_list) &gt; MAX_NUM_IMAGES_PER_CLASS:\n            warnings.warn('WARNING: Folder {} has more than {} images. Some '\n                          'images will never be selected.'\n                          .format(dir_name, MAX_NUM_IMAGES_PER_CLASS))\n        label_name = re.sub(r'[^a-z0-9]+', ' ', dir_name.lower())\n        training_images = []\n        validation_images = []\n        for file_name in file_list:\n            base_name = os.path.basename(file_name)\n            # Get the hash of the file name and perform variant assignment.\n            hash_name = hashlib.sha1(as_bytes(base_name)).hexdigest()\n            hash_pct = ((int(hash_name, 16) % (MAX_NUM_IMAGES_PER_CLASS + 1)) *\n                        (100.0 \/ MAX_NUM_IMAGES_PER_CLASS))\n            if hash_pct &lt; validation_pct:\n                validation_images.append(base_name)\n            else:\n                training_images.append(base_name)\n        image_lists[label_name] = {\n            'dir': dir_name,\n            'training': training_images,\n            'validation': validation_images,\n        }\n    return image_lists\n\n\n# https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/image_retraining\/retrain.py\ndef get_image_path(image_lists, label_name, index, image_dir, category):\n    \"\"\"\"Returns a path to an image for a label at the given index.\n\n    # Arguments\n      image_lists: Dictionary of training images for each label.\n      label_name: Label string we want to get an image for.\n      index: Int offset of the image we want. This will be moduloed by the\n      available number of images for the label, so it can be arbitrarily large.\n      image_dir: Root folder string of the subfolders containing the training\n      images.\n      category: Name string of set to pull images from - training, testing, or\n      validation.\n\n    # Returns\n      File system path string to an image that meets the requested parameters.\n    \"\"\"\n    if label_name not in image_lists:\n        raise ValueError('Label does not exist ', label_name)\n    label_lists = image_lists[label_name]\n    if category not in label_lists:\n        raise ValueError('Category does not exist ', category)\n    category_list = label_lists[category]\n    if not category_list:\n        raise ValueError('Label %s has no images in the category %s.',\n                         label_name, category)\n    mod_index = index % len(category_list)\n    base_name = category_list[mod_index]\n    sub_dir = label_lists['dir']\n    full_path = os.path.join(image_dir, sub_dir, base_name)\n    return full_path\n\n\ndef get_generators(image_lists, image_dir):\n    train_datagen = CustomImageDataGenerator(rescale=1. \/ 255,\n                                             horizontal_flip=True)\n\n    test_datagen = CustomImageDataGenerator(rescale=1. \/ 255)\n\n    train_generator = train_datagen.flow_from_image_lists(\n        image_lists=image_lists,\n        category='training',\n        image_dir=image_dir,\n        target_size=(MODEL_INPUT_HEIGHT, MODEL_INPUT_WIDTH),\n        batch_size=BATCH_SIZE,\n        class_mode='categorical',\n        seed=RANDOM_SEED)\n\n    validation_generator = test_datagen.flow_from_image_lists(\n        image_lists=image_lists,\n        category='validation',\n        image_dir=image_dir,\n        target_size=(MODEL_INPUT_HEIGHT, MODEL_INPUT_WIDTH),\n        batch_size=BATCH_SIZE,\n        class_mode='categorical',\n        seed=RANDOM_SEED)\n\n    return train_generator, validation_generator\n\n\ndef get_model(num_classes, weights='imagenet'):\n    # create the base pre-trained model\n    # , input_tensor=input_tensor\n    base_model = InceptionV3(weights=weights, include_top=False)\n\n    # add a global spatial average pooling layer\n    x = base_model.output\n    x = GlobalAveragePooling2D()(x)\n    # let's add a fully-connected layer\n    x = Dense(FC_LAYER_SIZE, activation='relu')(x)\n    # and a logistic layer -- let's say we have 2 classes\n    predictions = Dense(num_classes, activation='softmax')(x)\n\n    # this is the model we will train\n    model = Model(inputs=[base_model.input], outputs=[predictions])\n    return model\n\n\ndef get_top_layer_model(model):\n    \"\"\"Used to train just the top layers of the model.\"\"\"\n    # first: train only the top layers (which were randomly initialized)\n    # i.e. freeze all convolutional InceptionV3 layers\n    for layer in model.layers[:INCEPTIONV3_BASE_LAYERS]:\n        layer.trainable = False\n    for layer in model.layers[INCEPTIONV3_BASE_LAYERS:]:\n        layer.trainable = True\n\n    # compile the model (should be done after setting layers to non-trainable)\n    model.compile(optimizer='rmsprop', loss='categorical_crossentropy',\n                  metrics=['accuracy'])\n\n    return model\n\n\ndef get_mid_layer_model(model):\n    \"\"\"After we fine-tune the dense layers, train deeper.\"\"\"\n    # freeze the first TRAINABLE_LAYER_INDEX layers and unfreeze the rest\n    for layer in model.layers[:TRAINABLE_LAYERS]:\n        layer.trainable = False\n    for layer in model.layers[TRAINABLE_LAYERS:]:\n        layer.trainable = True\n\n    # we need to recompile the model for these modifications to take effect\n    # we use SGD with a low learning rate\n    model.compile(optimizer=SGD(lr=0.0001, momentum=0.9),\n                  loss='categorical_crossentropy',\n                  metrics=['accuracy'])\n\n    return model\n\n\ndef train_model(model, epochs, generators, callbacks=None):\n    train_generator, validation_generator = generators\n    model.fit_generator(\n        train_generator,\n        steps_per_epoch=STEPS_PER_EPOCH,\n        validation_data=validation_generator,\n        validation_steps=VALIDATION_STEPS,\n        epochs=epochs,\n        callbacks=callbacks)\n    return model\n\n\ndef main(image_dir, validation_pct):\n    sub_dirs = [x[0] for x in gfile.Walk(image_dir)]\n    num_classes = len(sub_dirs) - 1\n    print(\"Number of classes found: {}\".format(num_classes))\n\n    model = get_model(num_classes)\n\n    print(\"Using validation percent of %{}\".format(validation_pct))\n    image_lists = create_image_lists(image_dir, validation_pct)\n\n    generators = get_generators(image_lists, image_dir)\n\n    # Get and train the top layers.\n    model = get_top_layer_model(model)\n    model = train_model(model, epochs=10, generators=generators)\n\n    # Get and train the mid layers.\n    model = get_mid_layer_model(model)\n    _ = train_model(model, epochs=100, generators=generators,\n                    callbacks=[checkpointer, early_stopper, tensorboard])\n\n    # save model\n    model.save('.\/output\/model.hdf5', overwrite=True)\n\n\nif __name__ == '__main__':\n    parser = argparse.ArgumentParser()\n    parser.add_argument('--image-dir', required=True, help='data directory')\n    parser.add_argument('--validation-pct', default=10, help='validation percentage')\n    args = parser.parse_args()\n\n    os.makedirs('.\/output\/checkpoints\/', exist_ok=True)\n\n    main(**vars(args))\n<\/code><\/pre>\n"}]}
{"QuestionId":42444468,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><computer-vision><deep-learning>","CreationDate":"2017-02-24T17:11:25.083","AcceptedAnswerId":42452431.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Is there a way to locate the filenames of images encoded into TFRecord files?","Body":"<p>I am wondering if the filename information of the image encoded could be encoded into a TFRecord file while creating the tfrecord files, and if so, how could this information be decoded back? When decoded, is the filename a Tensor object?<\/p>\n","answers":[{"AnswerId":"42452431","CreationDate":"2017-02-25T05:44:43.630","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>Just like fabrizioM said, you have to store the sources in the tfrecords file if you want to use them. <a href=\"https:\/\/github.com\/MartinThoma\/algorithms\/tree\/master\/ML\/tf-image-loading\" rel=\"nofollow noreferrer\">Here is an example<\/a>:<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n\n\"\"\"Example for reading and writing tfrecords.\"\"\"\n\nimport tensorflow as tf\nfrom PIL import Image\nimport numpy as np\nimport scipy.misc\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\n\ndef write_images(filenames=['Aurelia-aurita-3.jpg'],\n                 labels=[0],\n                 tf_records_filename=\"example.tfrecords\"):\n    \"\"\"\n    Write images to tfrecords file.\n\n    Parameters\n    ----------\n    filenames : list of strings\n        List containing the paths to image files.\n    labels : list of integers\n    tf_records_filename : string\n        Where the file gets stored\n    \"\"\"\n    filename_queue = tf.train.string_input_producer(filenames)\n\n    reader = tf.WholeFileReader()\n    key, value = reader.read(filename_queue)\n\n    my_img = tf.image.decode_jpeg(value)\n\n    init_op = tf.initialize_all_variables()\n    with tf.Session() as sess:\n        sess.run(init_op)\n\n        # Start populating the filename queue.\n        coord = tf.train.Coordinator()\n        threads = tf.train.start_queue_runners(coord=coord)\n\n        writer = tf.python_io.TFRecordWriter(tf_records_filename)\n        for i in range(len(filenames)):\n            image = my_img.eval()  # image is an image tensor\n\n            image_raw = image.tostring()\n            rows = image.shape[0]\n            cols = image.shape[1]\n\n            if np.ndim(image) == 3:\n                depth = image.shape[2]\n            else:\n                depth = 1\n\n            example = tf.train.Example(features=tf.train.Features(feature={\n                'height': _int64_feature(rows),\n                'width': _int64_feature(cols),\n                'depth': _int64_feature(depth),\n                'label': _int64_feature(labels[i]),\n                'image_raw': _bytes_feature(image_raw),\n                'src': _bytes_feature(filenames[i])}))\n            writer.write(example.SerializeToString())\n        coord.request_stop()\n        coord.join(threads)\n\n\ndef read_and_decode(filename_queue):\n    \"\"\"Read and decode them from filename_queue.\"\"\"\n    reader = tf.TFRecordReader()\n    _, serialized_example = reader.read(filename_queue)\n    features = tf.parse_single_example(\n        serialized_example,\n        # Defaults are not specified since both keys are required.\n        features={\n            'image_raw': tf.FixedLenFeature([], tf.string),\n            'label': tf.FixedLenFeature([], tf.int64),\n            'height': tf.FixedLenFeature([], tf.int64),\n            'width': tf.FixedLenFeature([], tf.int64),\n            'depth': tf.FixedLenFeature([], tf.int64),\n            'src': tf.FixedLenFeature([], tf.string)\n        })\n    image = tf.decode_raw(features['image_raw'], tf.uint8)\n    label = tf.cast(features['label'], tf.int32)\n    height = tf.cast(features['height'], tf.int32)\n    width = tf.cast(features['width'], tf.int32)\n    depth = tf.cast(features['depth'], tf.int32)\n    # fn = tf.cast(features['filename'], tf.str)\n    return image, label, height, width, depth, features['src']\n\n\ndef get_all_records(record_filename):\n    \"\"\"Get all records from record_filename.\"\"\"\n    records = []\n    with tf.Session() as sess:\n        fn_queue = tf.train.string_input_producer([record_filename])\n        image, label, height, width, depth, src = read_and_decode(fn_queue)\n        image = tf.reshape(image, tf.stack([height, width, 3]))\n        init_op = tf.global_variables_initializer()\n        sess.run(init_op)\n        coord = tf.train.Coordinator()\n        threads = tf.train.start_queue_runners(coord=coord)\n        nr_of_images = 1\n        for i in range(nr_of_images):\n            example, label, src = sess.run([image, label, src])\n            img = Image.fromarray(example, 'RGB')\n            records.append({'image': img, 'label': label,\n                            'src': src})\n        coord.request_stop()\n        coord.join(threads)\n    return records\n\nwrite_images()\nrecords = get_all_records('example.tfrecords')\nprint(records[0]['src'])\nscipy.misc.imshow(records[0]['image'])\n<\/code><\/pre>\n"}]}
{"QuestionId":42444700,"AnswerCount":1,"Tags":"<ubuntu><torch><luajit>","CreationDate":"2017-02-24T17:23:40.407","AcceptedAnswerId":null,"OwnerUserId":7385862.0,"Title":"I am not able to install some luaJIT dependencies","Body":"<p>I am trying to install the following  luaJIT dependencies in my Ubuntu 16.0.4\nI try to do that by running the command below:<\/p>\n\n<pre><\/pre>\n\n<p>The system challenges my permissions and gives me back the partial list of errors shown below: <\/p>\n\n<pre><\/pre>\n\n<p>I would highly appreciate any help. <\/p>\n","answers":[{"AnswerId":"43878473","CreationDate":"2017-05-09T19:35:46.503","ParentId":null,"OwnerUserId":"7385862","Title":null,"Body":"<p>Colonel Thirty Two,<\/p>\n\n<p>I have found a solution for the problem. In fact a sudo should prefix the command, but the problem is when you do it as below:<\/p>\n\n<p>$ sudo for NAME in dpnn nn optim optnet csvigo cutorch cunn fblualib torchx tds; do luarocks install $NAME; done<\/p>\n\n<p>it gives you an error. <\/p>\n\n<p>The alternative is :\n$ for NAME in dpnn nn optim optnet csvigo cutorch cunn fblualib torchx tds; do sudo luarocks install $NAME; done\nand this should work for linux and os-x.<\/p>\n\n<p>or you can also install the dependencies one by one as below:<\/p>\n\n<p>sudo luarocks install dpnn\nsudo luarocks install nn\n..\n..\n..\nsudo luarocks install tds<\/p>\n"}]}
{"QuestionId":42444796,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T17:30:09.867","AcceptedAnswerId":null,"OwnerUserId":7618335.0,"Title":"How to turn off events.out.tfevents file in tf.contrib.learn Estimator","Body":"<p>When using <strong>estimator.Estimator<\/strong> in <strong>tensorflow.contrib.learn,<\/strong> after training and prediction there are these files in the <strong>modeldir<\/strong>: <\/p>\n\n<ul>\n<li>checkpoint<\/li>\n<li>events.out.tfevents.1487956647<\/li>\n<li>events.out.tfevents.1487957016<\/li>\n<li>graph.pbtxt<\/li>\n<li>model.ckpt-101.data-00000-of-00001<\/li>\n<li>model.ckpt-101.index<\/li>\n<li>model.ckpt-101.meta<\/li>\n<\/ul>\n\n<p>When the graph is complicated or the number of variables is big, the graph.pbtxt file and the events files can be very big. Is here a way to not write these files? Since model reloading only needs the checkpoint files removing them won't affect evaluation and prediction down the road.<\/p>\n","answers":[{"AnswerId":"53078259","CreationDate":"2018-10-31T07:25:10.597","ParentId":null,"OwnerUserId":"3876830","Title":null,"Body":"<p>On <a href=\"https:\/\/developers.google.com\/machine-learning\/crash-course\/\" rel=\"nofollow noreferrer\">Google's Machine Learning Crash Course<\/a> they use the following approach:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code># Create classifier:\nclassifier = tf.estimator.DNNRegressor(\n    feature_columns=feature_columns,\n    optimizer=optimizer\n)\n\n# Train it:\nclassifier.train(\n    input_fn=training_input_fn,\n    steps=steps\n)\n\n# Remove event files to save disk space.\n_ = map(os.remove, glob.glob(os.path.join(classifier.model_dir, 'events.out.tfevents*')))\n<\/code><\/pre>\n"},{"AnswerId":"51049687","CreationDate":"2018-06-26T18:52:09.990","ParentId":null,"OwnerUserId":"6098298","Title":null,"Body":"<p>I had the same issue and was not able to find any resolution for this while the events file kept on growing in size. My understanding is that this file stores the events generated by tensorflow. I went ahead and deleted this manually.\nInterestingly, it never got created again while the other files are getting updated when I run a train sequence.<\/p>\n"},{"AnswerId":"49691384","CreationDate":"2018-04-06T11:06:10.030","ParentId":null,"OwnerUserId":"5131441","Title":null,"Body":"<p>If you don't want the events.out.tfevents files to be written. Find in your code somethings like these and delete them.<\/p>\n\n<pre><code>tfFileWriter = tf.summary.FileWriter(os.getcwd())\ntfFileWriter.add_graph(sess.graph)\ntfFileWriter.close()\n<\/code><\/pre>\n"}]}
{"QuestionId":42445275,"AnswerCount":1,"Tags":"<machine-learning><keras><keras-layer>","CreationDate":"2017-02-24T17:56:45.427","AcceptedAnswerId":null,"OwnerUserId":2595774.0,"Title":"Merge outputs of different models with different input shapes","Body":"<p>I am new to Keras. I am trying to merge the output layers of three pretrained models in Keras. Each model has two separate inputs, but of different dimensions, and a Dense layer output.<\/p>\n\n<pre><\/pre>\n\n<p>Since model?.input returns a list of Tensors, this does not work. I tried different things and nothing seems to work. Is there an easy solution to this problem?<\/p>\n\n<p><strong>EDIT:<\/strong>\nAdapted working solution from indraforyou for multiple inputs in each model.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42446467","CreationDate":"2017-02-24T19:08:59.627","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p><s>The models object is not a callable function.<\/s> This should fix the problem:<\/p>\n\n<pre><code>x = merge([model1.output,\n           model2.output,\n           model3.output],\n          mode='concat', concat_axis=1)\n<\/code><\/pre>\n\n<p><strong>Updating the working code<\/strong><\/p>\n\n<pre><code>from keras.models import Model\nfrom keras.layers import Input, Dense, merge\n\n\ndef MyModel1():\n  inp = Input(batch_shape=(None,32,))\n  out = Dense(8)(inp)\n  return Model(input=inp, output=out)\n\ndef MyModel2():\n  inp = Input(batch_shape=(None,10,))\n  out = Dense(4)(inp)\n  return Model(input=inp, output=out)\n\ndef MyModel3():\n  inp = Input(batch_shape=(None,12,))\n  out = Dense(6)(inp)\n  return Model(input=inp, output=out)\n\nmodel1 = MyModel1()\nmodel2 = MyModel2()\nmodel3 = MyModel3()\n\nx = merge([model1.output,\n           model2.output,\n           model3.output],\n          mode='concat', concat_axis=1)\n\nx = Dense(2, activation='softmax')(x)\n\nmerged =  Model(input=[model1.input,\n                    model2.input,\n                    model3.input],\n             output=x)\n\nmerged.summary()\n<\/code><\/pre>\n"}]}
{"QuestionId":42445753,"AnswerCount":1,"Tags":"<python><tensorflow><jupyter-notebook><mnist>","CreationDate":"2017-02-24T18:24:17.157","AcceptedAnswerId":42445963.0,"OwnerUserId":4896087.0,"Title":"Where is the working directory?","Body":"<p>I ran this code on my machine using Jupyter Notebook, thinking it may not work since the relative directory doesn't exist...However, it turns out it worked. So the relative directory used  must be relative to current working dir which is . However, I was not able to find any related folders or files. I also checked , nothing there either...What am I missing? Where does the file go? Thanks!<\/p>\n\n<pre><\/pre>\n\n<p>EDIT:<\/p>\n\n<p>I did check current working dir using this code:<\/p>\n\n<pre><\/pre>\n\n<p>and this is the result:<\/p>\n\n<p>'C:\\Users\\george.liu\\OneDrive\\WorkingDir\\Temp'<\/p>\n","answers":[{"AnswerId":"42445963","CreationDate":"2017-02-24T18:36:48.450","ParentId":null,"OwnerUserId":"6451573","Title":null,"Body":"<p>On Linux <code>\/datasets\/<\/code> is an absolute path<\/p>\n\n<p>But on Windows machine, if you have a subdirectory <code>python<\/code> in your current drive, ex <code>D:\\python<\/code> and the current directory is <em>somewhere<\/em> in <code>D:<\/code> (not necessarily at root, could be: <code>D:\\python<\/code> already) you can do <code>os.chdir(\"\/python\")<\/code> and it works like if you did <code>os.chdir(r\"D:\\python\")<\/code><\/p>\n\n<p>So the slash is <em>not<\/em> ignored, it's just like <code>:\\<\/code> (root of current drive)<\/p>\n\n<p>In your case, python looks for <code>C:\\datasets\\ud730\\mnist<\/code> since the current directory is somewhere on drive <code>C:<\/code>.<\/p>\n\n<p>You can check that by printing <code>os.path.abspath('\/datasets\/ud730\/mnist')<\/code><\/p>\n"}]}
{"QuestionId":42446331,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-24T18:59:48.423","AcceptedAnswerId":45518731.0,"OwnerUserId":7618783.0,"Title":"How to launch TensorFlow Shell shown in XLA Demo","Body":"<p>In the recent video \"XLA: TensorFlow, Compiled!\" from the 2017 TensorFlow Dev Summit, the presenter shows a shell that resembles iPython and is able to view the JIT'ed assembly:\n<a href=\"https:\/\/youtu.be\/kAOanJczHA0?t=2m32s\" rel=\"nofollow noreferrer\">https:\/\/youtu.be\/kAOanJczHA0?t=2m32s<\/a><\/p>\n\n<p>I'm trying to recreate the demo so I can display the generated assembly for some other sample code.<\/p>\n\n<p>I assume the shell is just a wrapper that imports TensorFlow and drops to a iPython shell?\nI tried launching a script such as:<\/p>\n\n<pre><\/pre>\n\n<p>and dropping to a shell with:<\/p>\n\n<pre><\/pre>\n\n<p>But when I attempt to recreate the code snippet in the demo the assembly isn't displayed. Can anyone help reproduce it?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42525403","CreationDate":"2017-03-01T07:12:18.650","ParentId":null,"OwnerUserId":"7639691","Title":null,"Body":"<p>You're right that the shell is just a wrapper that imports TensorFlow and drops into an iPython shell.<\/p>\n\n<p>Note that at the moment, in order to get a version of TensorFlow that supports XLA, you need to compile TensorFlow from source.  When you run .\/configure you'll be asked whether to enable XLA, and you need to respond Y (yes).<\/p>\n\n<p>We're working hard to enable XLA by default, but that's what's necessary at the moment.  Hope this helps!<\/p>\n"},{"AnswerId":"45518731","CreationDate":"2017-08-05T05:52:15.203","ParentId":null,"OwnerUserId":"3594","Title":null,"Body":"<p>Support for \"vmodule\" logging <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/072b0c9c552837a97c967963628b3c30951388f9\" rel=\"nofollow noreferrer\">just landed in open source TensorFlow<\/a>, so it's possible to enable logging in specific translation units to see the assembly. Open source TensorFlow doesn't generally use flags, so instead there are some environment variables for controlling some of this debug-oriented behavior. I just built at head and ran on a GCE VM:<\/p>\n\n<p><code>cdleary@instance-1:~\/tensorflow$ TF_CPP_VMODULE=gpu_compiler=3 bazel-bin\/tensorflow\/compiler\/xla\/tests\/while_test_gpu\n<\/code><\/p>\n\n<p>And that spits out, as in the video, the PTX (virtual) assembly:<\/p>\n\n<p><code>2017-08-05 05:45:07.030245: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349] .version 4.2\n2017-08-05 05:45:07.030252: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349] .target sm_37\n2017-08-05 05:45:07.030257: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349] .address_size 64\n2017-08-05 05:45:07.030263: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349] \n2017-08-05 05:45:07.030269: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349]  \/\/ .globl   _greater_than\n2017-08-05 05:45:07.030275: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349] \n2017-08-05 05:45:07.030281: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349] .visible .entry _greater_than(\n2017-08-05 05:45:07.030293: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349]  .param .u64 _greater_than_param_0,\n2017-08-05 05:45:07.030300: I tensorflow\/compiler\/xla\/service\/gpu\/gpu_compiler.cc:349]  .param .u64 _greater_than_param_1\n[snip]\n<\/code><\/p>\n\n<p>Similarly, for CPU:<\/p>\n\n<p><code>cdleary@instance-1:~\/tensorflow$ TF_CPP_VMODULE=compiler_functor=3 bazel-bin\/tensorflow\/compiler\/xla\/tests\/while_test_cpu\n<\/code><\/p>\n\n<p>Produces:<\/p>\n\n<p><code>[snip]\n2017-08-05 05:49:43.819213: I tensorflow\/compiler\/xla\/service\/cpu\/compiler_functor.cc:121] NestedWhileWithScalarResult.v2:\n2017-08-05 05:49:43.819244: I tensorflow\/compiler\/xla\/service\/cpu\/compiler_functor.cc:121] 0x00000000   movq    (%rcx), %rax\n2017-08-05 05:49:43.819256: I tensorflow\/compiler\/xla\/service\/cpu\/compiler_functor.cc:121] 0x00000003   movl    $0, (%rax)\n2017-08-05 05:49:43.819263: I tensorflow\/compiler\/xla\/service\/cpu\/compiler_functor.cc:121] 0x00000009   movq    8(%rcx), %rcx\n2017-08-05 05:49:43.819284: I tensorflow\/compiler\/xla\/service\/cpu\/compiler_functor.cc:121] 0x0000000d   nopl    (%rax)\n2017-08-05 05:49:43.819299: I tensorflow\/compiler\/xla\/service\/cpu\/compiler_functor.cc:121] 0x00000010   movq    %rcx, 16(%rcx)\n2017-08-05 05:49:43.819314: I tensorflow\/compiler\/xla\/service\/cpu\/compiler_functor.cc:121] 0x00000014   movq    %rax, 24(%rcx)\n[snip]\n<\/code><\/p>\n\n<p>HTH!<\/p>\n"}]}
{"QuestionId":42446876,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><nvidia><tensorrt>","CreationDate":"2017-02-24T19:33:09.863","AcceptedAnswerId":null,"OwnerUserId":5861307.0,"Title":"How to run pretrained tensorflow model using Nvidia's Tensor RT on Jetson TX1?","Body":"<p>In the Nvidia's blog, they introduced their TensorRT as follows: <\/p>\n\n<blockquote>\n  <p>NVIDIA TensorRT\u2122 is a high performance neural network inference engine for production deployment of deep learning applications. TensorRT can be used to rapidly optimize, validate and deploy trained neural network for inference to hyperscale data centers, embedded, or automotive product platforms.<\/p>\n<\/blockquote>\n\n<p>So I am wondering, if I have a pre-trained Tensorflow model, can I use it in TensorRT in Jetson TX1 for inference?<\/p>\n","answers":[{"AnswerId":"48058059","CreationDate":"2018-01-02T08:55:46.677","ParentId":null,"OwnerUserId":"2648715","Title":null,"Body":"<p>From JetPack 3.1, NVIDIA has added TensorRT support for Tensorflow also. So, the trained TF model can be directly deployed in Jetson TX1\/TK1\/TX2<\/p>\n"},{"AnswerId":"45450738","CreationDate":"2017-08-02T03:01:03.380","ParentId":null,"OwnerUserId":"4715661","Title":null,"Body":"<p>Base on <a href=\"https:\/\/devtalk.nvidia.com\/default\/topic\/981654\/use-jetson-tx1-tensor-rt-to-run-my-tensorflow-model\/?offset=2\" rel=\"nofollow noreferrer\">this post<\/a> from Nvidia forum, it seems that you could use TensorRT for inference with caffemodel but not tensorflow model now. \nBeside tensorRT, building tensorflow on tx1 is another issue (refer here: <a href=\"https:\/\/github.com\/ugv-tracking\/cfnet\" rel=\"nofollow noreferrer\">https:\/\/github.com\/ugv-tracking\/cfnet<\/a>).<\/p>\n"}]}
{"QuestionId":42447036,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><theano><keras>","CreationDate":"2017-02-24T19:43:18.180","AcceptedAnswerId":null,"OwnerUserId":6522749.0,"Title":"Overfitting my model over my training data of a single sample","Body":"<p>I am trying to over-fit my model over my training data that consists of only a single sample. The training accuracy comes out to be 1.00. But, when I predict the output for my test data which consists of the same single training input sample, the results are not accurate. The model has been trained for 100 epochs and the loss ~ 1e-4. \nWhat could be the possible sources of error?<\/p>\n","answers":[{"AnswerId":"42456700","CreationDate":"2017-02-25T13:30:21.537","ParentId":null,"OwnerUserId":"3126298","Title":null,"Body":"<p>As mentioned in the comments of your post, it isn't possible to give specific advice without you first providing more details.<\/p>\n\n<p>Generally speaking, your approach to overfitting a tiny batch (in your case one image) is in essence providing three sanity checks, i.e. that:<\/p>\n\n<ol>\n<li>backprop is functioning<\/li>\n<li>the weight updates are doing their job<\/li>\n<li>the learning rate is in the correct order of magnitude<\/li>\n<\/ol>\n\n<p>As is pointed out by Andrej Karpathy in Lecture 5 of CS231n course at Stanford - <a href=\"https:\/\/youtu.be\/gYpoJMlgyXA?t=1h1m22s\" rel=\"nofollow noreferrer\">\"if you can't overfit on a tiny batch size, things are definitely broken\"<\/a>.<\/p>\n\n<p>This means, given your description, that your implementation is incorrect. I would start by checking each of those three points listed above. For example, alter your test somehow by picking several different images or a btach-size of 5 images instead of one. You could also revise your <code>predict function<\/code>, as that is where there is definitely some discrepancy, given you are getting zero error during training (and so validation?).<\/p>\n"}]}
{"QuestionId":42447177,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras><keras-layer>","CreationDate":"2017-02-24T19:52:02.603","AcceptedAnswerId":null,"OwnerUserId":3167448.0,"Title":"How does one store the information to be able to plot the error vs epochs in Keras?","Body":"<p>I wanted to plot the error vs the number of epochs in keras. To do that I found a really useful post <a href=\"http:\/\/machinelearningmastery.com\/display-deep-learning-model-training-history-in-keras\/\" rel=\"nofollow noreferrer\">here<\/a>. It works like a charm locally but if I want to run my scripts in some remote GPU\/cluster then its not as useful because I need to send myself the data to be able to actually make the plots. Is there a way to run something similar but store the errors in some file\/csv\/json anything and then be able to send that to myself with standard tools like rysnc\/scp\/dropbox?<\/p>\n\n<p>Right now the only idea I had was to pickle the file and then send it to myself and unpickle it but it doesn't work at all. I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>Is there a different way to do it using Keras and python? I just need to be able to store this type of data so that I can send it to my local computer somehow.<\/p>\n\n<hr>\n\n<p>Currently my totally reproducible error can be done as follow:<\/p>\n\n<p>store_data.py<\/p>\n\n<pre><\/pre>\n\n<p>and to plot it:<\/p>\n\n<p>plot.py<\/p>\n\n<pre><\/pre>\n\n<p>Also, if ur having issue with matplot lib due to framework builds just use the virtual env that already solves that problem for you with:<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"http:\/\/matplotlib.org\/faq\/osx_framework.html#virtualenv\" rel=\"nofollow noreferrer\">http:\/\/matplotlib.org\/faq\/osx_framework.html#virtualenv<\/a><\/p>\n\n<p><a href=\"http:\/\/matplotlib.org\/faq\/osx_framework.html\" rel=\"nofollow noreferrer\">http:\/\/matplotlib.org\/faq\/osx_framework.html<\/a><\/p>\n\n<p><a href=\"http:\/\/matplotlib.org\/faq\/virtualenv_faq.html\" rel=\"nofollow noreferrer\">http:\/\/matplotlib.org\/faq\/virtualenv_faq.html<\/a><\/p>\n\n<p>the  uses python3\u2019s implementation of virtualenv, and that implementation makes python a framework.<\/p>\n","answers":[{"AnswerId":"42447245","CreationDate":"2017-02-24T19:56:38.497","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Try pickling only loss dictionary:<\/p>\n\n<pre><code>pickle.dump( history.history, open( \"history.p\", \"wb\" ) )\n<\/code><\/pre>\n"}]}
{"QuestionId":42447725,"AnswerCount":0,"Tags":"<generator><keras>","CreationDate":"2017-02-24T20:30:02.950","AcceptedAnswerId":null,"OwnerUserId":2145297.0,"Title":"Keras fit_generator throwing ValueError","Body":"<p>So I'm trying to create a generator to iterate through a data set for use in training with Keras's fit_generator. Here's the definition of the generator, the model, and the call to fit_generator:<\/p>\n\n<pre><\/pre>\n\n<p>Keras will then successfully train 98\/100 training examples and throw this error<\/p>\n\n<pre><\/pre>\n\n<p>It seems like what's happening is that it popped of all of the training_idxs and it's still trying to get more until Keras exhaust the training examples in its internal queue. Is there a way to get it to stop trying to get more training examples from the generator?<\/p>\n","answers":[]}
{"QuestionId":42447975,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T20:45:23.760","AcceptedAnswerId":null,"OwnerUserId":2490125.0,"Title":"In TensorFlow, how can I selectively apply a function only to the maximum elements of each row of a tensor?","Body":"<p>I have a 2D tensor, and I want to apply a function only to the maximum elements of each row.  For example, if my tensor is:<\/p>\n\n<pre><\/pre>\n\n<p>and my function is to multiply by 2, then I want my result to be<\/p>\n\n<pre><\/pre>\n\n<p>My current approach is to use , like this:<\/p>\n\n<pre><\/pre>\n\n<p>The idea is to create a tensor which is the max of each row using tf.reduce_max, and then use tf.equal to create a boolean tensor which says whether each element of input is equal to its row max.  This then gets passed to tf.select as a mask to determine whether the function should be applied.  However, this doesn't seem to work - the max tensor is not of the same dimension as the input tensor, and I think there's some issue with broadcasting.  Also, this method may run into issues if there are two equal maximal elements.  I'd rather have a method that automatically breaks ties, perhaps using tf.arg_max.<\/p>\n","answers":[{"AnswerId":"42450412","CreationDate":"2017-02-25T00:18:08.033","ParentId":null,"OwnerUserId":"2490125","Title":null,"Body":"<p>I was able to modify the code in the question by using tf.expand_dims on the result of tf.reduce_max to get a tensor of the proper shape to be broadcast to match the input tensor.  The result looks like this:<\/p>\n\n<pre><code>tf.select(tf.equal(my_tensor,\n                   tf.expand_dims(tf.reduce_max(my_tensor,axis=1), axis=1)),\n          my_function(my_tensor),\n          my_tensor)\n<\/code><\/pre>\n\n<p>While this works, I suspect there is a cleaner method.<\/p>\n"}]}
{"QuestionId":42448796,"AnswerCount":0,"Tags":"<python><tensorflow><computer-vision>","CreationDate":"2017-02-24T21:43:06.797","AcceptedAnswerId":null,"OwnerUserId":1145744.0,"Title":"How to pass an image with feed_dict in Tensorflow session run","Body":"<p>I need to pass an image during session Run in Tensorflow<\/p>\n\n<p>I have tried <\/p>\n\n<pre><\/pre>\n\n<p>but doesn't work<\/p>\n\n<p>Returns me the error:<\/p>\n\n<pre><\/pre>\n\n<p> tensor has the \n(120 is the number of training classes in which images are grouped) <\/p>\n\n<p>the program load and labels input images with this method<\/p>\n\n<pre><\/pre>\n\n<p>than a batch load the record and feed the CNN<\/p>\n\n<p>EDIT<\/p>\n\n<p><strong>I don't use same symbol-key value please before flag as duplicate pay attention.<\/strong><\/p>\n","answers":[]}
{"QuestionId":42449339,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2017-02-24T22:27:23.197","AcceptedAnswerId":42659091.0,"OwnerUserId":4013859.0,"Title":"General way to test restored tensorflow model","Body":"<p>I am very, very brand new to Tensorflow, and need to write a script that tests a single example on a model restored from a checkpoint file. <\/p>\n\n<p>I was wondering if there was a general way to build a test function for a restored model without knowing all the minute details of the model. <\/p>\n\n<p>Further, in the last section of the code below, does this look like I'm headed in the right direction? If so, how does one build 'y' without knowing details of the model by heart?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42659091","CreationDate":"2017-03-07T21:56:34.183","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/extend\/estimators\" rel=\"nofollow noreferrer\">Estimator<\/a> class has a convenient set of utilities such that, if your model is wrapped around an estimator, loading and predicting from it is easy.<\/p>\n\n<p>Overall, without some kind of coordination, this will be hard.<\/p>\n"}]}
{"QuestionId":42449635,"AnswerCount":1,"Tags":"<amazon-ec2><keras><recurrent-neural-network>","CreationDate":"2017-02-24T22:53:14.610","AcceptedAnswerId":42463521.0,"OwnerUserId":6949914.0,"Title":"Why does my keras neural network model output different values on a different machine?","Body":"<p>I am using aws ec2 to train a model for a multi-label classification task. After training , I tested the model on the same machine which gives me good results (accuracy of 90+%). However, after I imported the saved model into my local machine (no GPU), it gives different results (accuracy is less than 5%). Any suggestions on why this is happening? Thanks.<\/p>\n\n<p>TL;DR: <strong><em>Keras\/tensorflow<\/em><\/strong> model produces different results when transfered <strong>from GPU<\/strong> machine <strong>to CPU<\/strong>.<\/p>\n","answers":[{"AnswerId":"42463521","CreationDate":"2017-02-26T00:58:55.307","ParentId":null,"OwnerUserId":"6949914","Title":null,"Body":"<p>Upon searching the net, I found the problem. It seems like that keras over tensorflow when running on a GPU tends to produce results that are not reproducible when transfered to a non-GPU machine. This most likely has something to do with the cuDNN installed. cuDNN's maxpooling and some convolution backward algorithm is non-deterministic - as said from a forum.<\/p>\n\n<p>Solutions I found say the use of <code>numpy.random.seed(seed_no)<\/code> right before calling any keras libraries. This works when you run the code <strong>on a CPU<\/strong>. Works with both <em>keras\/theano<\/em> and <em>keras\/tensorflow<\/em>.<\/p>\n\n<p>Solution for <strong>GPU users<\/strong> using keras over theano involves <a href=\"https:\/\/www.kaggle.com\/c\/ultrasound-nerve-segmentation\/discussion\/22637#130118\" rel=\"nofollow noreferrer\">modifying the .theanorc file into<\/a>:<\/p>\n\n<pre><code>[dnn.conv]\nalgo_bwd_filter = deterministic\nalgo_bwd_data = deterministic\n<\/code><\/pre>\n\n<p>Or <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2479#issuecomment-213987747\" rel=\"nofollow noreferrer\">using theano flags<\/a>: <code>THEANO_FLAGS=\"dnn.conv.algo_bwd_filter=deterministic,dnn.conv.algo_bwd_data=deterministic\" python rnn_model.py<\/code><\/p>\n\n<p>However, I haven't found any clear instructions yet of how to produce uniform results on a keras with tensorflow as back-end running on a GPU. <\/p>\n"}]}
{"QuestionId":42450156,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-24T23:46:45.907","AcceptedAnswerId":42450418.0,"OwnerUserId":395857.0,"Title":"Why is the value of a `tf.constant()` stored multiple times in memory in TensorFlow?","Body":"<p>I <a href=\"https:\/\/stackoverflow.com\/a\/35688187\/395857\">read<\/a> that (in TensorFlow):<\/p>\n\n<blockquote>\n  <p>the value of a <a href=\"https:\/\/web.archive.org\/web\/20170224234612\/https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"nofollow noreferrer\"><\/a> is stored multiple times in memory.<\/p>\n<\/blockquote>\n\n<p>Why is the value of a  stored multiple times in memory?<\/p>\n","answers":[{"AnswerId":"42450418","CreationDate":"2017-02-25T00:18:53.543","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Because data for a constant tensor is embedded into graph definition. This means this data is stored both in the client, which maintains the graph definition, and in the runtime, which allocates it's own memory for all tensors.<\/p>\n\n<p>IE, try<\/p>\n\n<pre><code>a = tf.constant([1,2])\ntf.get_default_graph().as_graph_def()\n<\/code><\/pre>\n\n<p>You'll see<\/p>\n\n<pre><code>    dtype: DT_INT32\n    tensor_shape {\n      dim {\n        size: 2\n      }\n    }\n    tensor_content: \"\\001\\000\\000\\000\\002\\000\\000\\000\"\n  }\n<\/code><\/pre>\n\n<p>The <code>tensor_content<\/code> field is the raw content, same as <code>np.array([1,2], dtype=np.int32).tobytes()<\/code>.<\/p>\n\n<p>Now, to see the runtime allocation, you can run with <code>export TF_CPP_MIN_LOG_LEVEL=1<\/code>.<\/p>\n\n<p>If you evaluate anything using <code>a<\/code> you'll see something like this<\/p>\n\n<pre><code>2017-02-24 16:13:58: I tensorflow\/core\/framework\/log_memory.cc:35] __LOG_MEMORY__ MemoryLogTensorOutput { step_id: 1 kernel_name: \"Const_1\/_1\" tensor { dtype: DT_INT32 shape { dim { size: 2 } } allocation_description { requested_bytes: 8 allocated_bytes: 256 allocator_name: \"cuda_host_bfc\" allocation_id: 1 ptr: 8605532160 } } }\n<\/code><\/pre>\n\n<p>This means the runtime asked to allocate 8 bytes, and TF actually allocated   256 bytes. (the choices on how much data to actually allocate are somewhat arbitrary at the moment - <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/d24e64227653ab3cab7405c79ca166063a42f0ff\/tensorflow\/core\/common_runtime\/bfc_allocator.cc#L302\" rel=\"noreferrer\">bfc_allocator.cc<\/a> )<\/p>\n\n<p>Having constants embedded in the graph makes it easier to do some graph-based optimizations like <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/b00fc538638f87ac45be9105057b9865f0f9418b\/tensorflow\/core\/common_runtime\/graph_optimizer.cc#L57\" rel=\"noreferrer\">constant folding<\/a> . But this also means that large constants are inefficient. Also, using large constants is a common cause of exceeding 2GB limit for size of graph.<\/p>\n"},{"AnswerId":"42450320","CreationDate":"2017-02-25T00:04:53.907","ParentId":null,"OwnerUserId":"2878004","Title":null,"Body":"<p>They are referring to the fact that when initializing the constant one copy of the constant is stored as a numpy array and another copy is stored in tensorflow. The two copies exist while it is initializing the constant. <\/p>\n"}]}
{"QuestionId":42450235,"AnswerCount":1,"Tags":"<python><theano><keras>","CreationDate":"2017-02-24T23:55:40.630","AcceptedAnswerId":null,"OwnerUserId":3849791.0,"Title":"Making my data fit Keras Sequential Models and Dense Layers and produce output","Body":"<p>I have structured data that looks like this.<\/p>\n\n<pre><\/pre>\n\n<p>I'm trying to predict the class FAILMODE. There's only 122 unique values in FAILMODE. Those are my classes.<\/p>\n\n<p>Based on all the other data in the rows, I want to have a one-hot matrix or even the class itself be a product of the computation on my test set. Here's my code so far-<\/p>\n\n<pre><\/pre>\n\n<p>Here is where the tutorial says-<\/p>\n\n<pre><\/pre>\n\n<p>I don't know what X or Y is so I just tried the following-<\/p>\n\n<pre><\/pre>\n\n<p>It resulted in this error-<\/p>\n\n<pre><\/pre>\n\n<p>Please be thorough with your answer. Thank you!<\/p>\n","answers":[{"AnswerId":"42450857","CreationDate":"2017-02-25T01:22:45.463","ParentId":null,"OwnerUserId":"1862500","Title":null,"Body":"<p>Regular neural network (including Keras Sequential model) only accepts float for data (X) and int or one-hot-encoding for label\/class (Y). So you need to convert your dataset to match the requirements. So what you may want to do:<\/p>\n\n<ol>\n<li>Map all categorical (string) values (eg. CAMPAIGN\/TRP, BNRBC1, XXIPNF) to float (it will be better if you can normalize the data)<\/li>\n<li>Put all the data columns (excluding the label column) in X<\/li>\n<li>Put the label column (should be 1 column only) inside Y and convert to one-hot-encoding using <a href=\"https:\/\/keras.io\/utils\/np_utils\/\" rel=\"nofollow noreferrer\">to_categorical<\/a> eg <code>Y = to_categorical(Y)<\/code><\/li>\n<li>Split the training and testing data using function such as <a href=\"http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.model_selection.train_test_split.html\" rel=\"nofollow noreferrer\">train_test_split<\/a> <code>X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.33)<\/code><\/li>\n<li>Train the model using <code>model.fit(X_train, Y_train, nb_epoch=100, batch_size=10)<\/code>. Adjust the <code>nb_epoch<\/code> and <code>batch_size<\/code> later on based on your expectation of speed and accuracy of the training<\/li>\n<li>Evaluate the accuracy using <code>scores = model.evaluate(self, X_test, Y_test, batch_size=10)<\/code><\/li>\n<\/ol>\n\n<p>You can also check this article to get idea how to convert categorical data to numeric <a href=\"http:\/\/fastml.com\/converting-categorical-data-into-numbers-with-pandas-and-scikit-learn\/\" rel=\"nofollow noreferrer\">http:\/\/fastml.com\/converting-categorical-data-into-numbers-with-pandas-and-scikit-learn\/<\/a><\/p>\n"}]}
{"QuestionId":42450246,"AnswerCount":1,"Tags":"<python-2.7><tensorflow>","CreationDate":"2017-02-24T23:56:37.937","AcceptedAnswerId":42518648.0,"OwnerUserId":2985049.0,"Title":"Tensorflow FailedPreconditionError while trying to save frozen model","Body":"<p>I'm training a CNN with word embeddings and for some reason I'm getting  exception whenever I try to save a frozen version of the model for later use.<\/p>\n\n<p>This is despite the fact that I call  just before training and I have no problem training and checkpointing the model. <\/p>\n\n<p>The problem occurs when I try to load a model from a checkpoint and save a frozen model. The function I'm using is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The error I get is:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42518648","CreationDate":"2017-02-28T21:01:03.883","ParentId":null,"OwnerUserId":"2985049","Title":null,"Body":"<p>Turns out I was constructing a <code>Saver<\/code> object before I made a <code>Session<\/code> so nothing from the session was being saved.<\/p>\n"}]}
{"QuestionId":42450378,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-25T00:13:51.650","AcceptedAnswerId":null,"OwnerUserId":7619803.0,"Title":"How XLA express matrix_solve","Body":"<p>I browsed through XLA's document and cannot find a good way to express more complicated matrix operations like matrix_solve, matrix_triangular_solve, cholesky and so on. How XLA handles this? I know there is a catch-call operation \"CallCustom\" but just wonder about better ways.<\/p>\n","answers":[{"AnswerId":"42525794","CreationDate":"2017-03-01T07:35:56.700","ParentId":null,"OwnerUserId":"7639691","Title":null,"Body":"<p>In general, the intention is for the actual computation to be specified in regular TensorFlow.  Then you turn on XLA either via Just-In-Time compilation (<a href=\"https:\/\/www.tensorflow.org\/performance\/xla\/jit\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/performance\/xla\/jit<\/a>), or Ahead-Of-Time compilation (...\/xla\/tfcompile).<\/p>\n\n<p>In terms of underlying support for matrix solvers, note that in addition to typical dense matrix operations, XLA does support some control flow primitives.  See <a href=\"https:\/\/www.tensorflow.org\/performance\/xla\/operation_semantics\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/performance\/xla\/operation_semantics<\/a> paying attention to the while loop construct (#while), and how to select output from different choices (#select).<\/p>\n\n<p>I haven't worked out whether this will yield a great result, but at a high-level it seems like the fundamental pieces are there.<\/p>\n\n<p>(Sorry for the abbreviated links; I can't seem to post more than 2)<\/p>\n"}]}
{"QuestionId":42450389,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-02-25T00:14:46.923","AcceptedAnswerId":42451067.0,"OwnerUserId":665023.0,"Title":"How a filter in convolutional neural network can generate multiple channels?","Body":"<p>I am learning convolutional neural network with Tensorflow. <\/p>\n\n<p>I have some doubts regarding <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\"><\/a>. One of its parameters is :<\/p>\n\n<blockquote>\n  <p>a filter \/ kernel tensor of shape [filter_height, filter_width,  in_channels, out_channels]<\/p>\n<\/blockquote>\n\n<p>I do not understand what is the meaning of .<\/p>\n\n<p>Suppose input image is . So the size is 3xx and the channel is 1.<br>\nThen we have a filter , which means after the filtering, we will have an image of size 2x2 (\"valid\" padding) with 5 channels.<\/p>\n\n<p>Where are the 5 channels from? From my understanding, the filtering can only have 1 channel generated. Is Tensorflow using 5 different filter functions here?<\/p>\n","answers":[{"AnswerId":"42451067","CreationDate":"2017-02-25T01:54:16.400","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>The <code>filter<\/code> argument to the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\"><code>tf.nn.conv2d<\/code><\/a> function, as you quoted, is a 4D tensor of dimensions <code>[filter_height, filter_width, in_channels, out_channels]<\/code>. This tensor represents a stack of <code>out_channels<\/code> filters of dimension <code>filter_height x filter_width<\/code>, to be applied over an image with <code>in_channels<\/code> channels.<\/p>\n\n<p>The parameters, <code>filter_height<\/code>, <code>filter_width<\/code> and <code>out_channels<\/code> are <em>defined by you<\/em>, whereas <code>input_channels<\/code> is dependent on your input to <code>tf.nn.conv2d<\/code>.<\/p>\n\n<p>In other words, a filter tensor with dimensions <code>[2, 2, 1, 5]<\/code>, represents <code>5<\/code> different <code>2 x 2<\/code> filters to be applied over a <code>1<\/code>-channel input, but you could perfectly change it to <code>[2, 2, 1, 7]<\/code>, or whatever else gives you better results.<\/p>\n\n<p>To further illustrate, in the following gif you have a <code>[3, 3, 1, 1]<\/code> tensor filter convolving over a <code>[1, 5, 5, 1]<\/code> image. This means you have only <code>1<\/code> filter being convolved over the image.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/I7DBr.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/I7DBr.gif\" alt=\"Convolution GIF\"><\/a><\/p>\n\n<p><a href=\"http:\/\/deeplearning.stanford.edu\/wiki\/index.php\/Feature_extraction_using_convolution\" rel=\"nofollow noreferrer\">GIF source<\/a><\/p>\n"}]}
{"QuestionId":42451004,"AnswerCount":0,"Tags":"<python><neural-network><keras>","CreationDate":"2017-02-25T01:46:37.257","AcceptedAnswerId":null,"OwnerUserId":6201224.0,"Title":"Accuracy rate of Keras neural network not changing","Body":"<p>I am a complete newbie to Neural Networks and keras and I am trying to build a neural network classifier using Keras to predict gender on the basis of height. When I run the following code, the accuracy does not improve upon each iteration and the predicted class is always 0. <\/p>\n\n<p>Since this is a homework assignment, the professor has opted to keep the data private, but my data merely consists of a pandas DataFrame with two columns: gender and height where gender is coded as a string of 'female' and 'male' and height is a number in inches.<\/p>\n\n<pre><\/pre>\n\n<p>Possible problems\/approaches I have tried:<\/p>\n\n<ol>\n<li>My data is pretty balanced. The proportion of females and males are almost perfectly split 50%, so this should not be an issue.<\/li>\n<li>Changing the learning rate (I have tried 0.1, 0.01, 0.001, and 0.0001).<\/li>\n<li>Changing the optimizer.<\/li>\n<\/ol>\n","answers":[]}
{"QuestionId":42451062,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-25T01:53:59.747","AcceptedAnswerId":null,"OwnerUserId":7619988.0,"Title":"TensorFlow: SKCompat Depreciation Warning","Body":"<p>NOTE: My first question on here. Pardon for lack of details or info. More than happy to clarify if needed. <\/p>\n\n<p>I'm running TensorFlow 1.0.0 on Mac and I keep getting this warning when using the learn.Estimator class<\/p>\n\n<blockquote>\n  <p>WARNING:tensorflow:From :25: calling\n  fit (from tensorflow.contrib.learn.python.learn.estimators.estimator)\n  with y is deprecated and will be removed after 2016-12-01.\n  Instructions for updating: Estimator is decoupled from Scikit Learn\n  interface by moving into separate class SKCompat. Arguments x, y and\n  batch_size are only available in the SKCompat class, Estimator will\n  only accept input_fn. Example conversion:   est = Estimator(...) ->\n  est = SKCompat(Estimator(...))<\/p>\n<\/blockquote>\n\n<p>I've tried looking this class up and there is zero information in regards to it. Full code is posted here<\/p>\n\n<p><a href=\"https:\/\/github.com\/austinmwhaley\/DeepFarm\/blob\/master\/prototype_1.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/austinmwhaley\/DeepFarm\/blob\/master\/prototype_1.ipynb<\/a><\/p>\n\n<p>Please let  me know if there is any other info that anyone needs<\/p>\n","answers":[{"AnswerId":"43659392","CreationDate":"2017-04-27T13:31:32.360","ParentId":null,"OwnerUserId":"3053918","Title":null,"Body":"<p>Or you simply use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/estimator\/Estimator\" rel=\"nofollow noreferrer\">updated Estimator API of TensorFlow r1.1<\/a><\/p>\n\n<p>The API for the model definition is quite similar with some small changes in parameters, return type or function name only. Here is an example that I have used:<\/p>\n\n<pre><code>def model_fn():\n    def _build_model(features, labels, mode, params):\n        # 1. Configure the model via TensorFlow operations\n        # Connect the first hidden layer to input layer (features) with relu activation\n        y = tf.contrib.layers.fully_connected(features, num_outputs=64, activation_fn=tf.nn.relu,\n                                              weights_initializer=tf.contrib.layers.xavier_initializer())\n        y = tf.contrib.layers.fully_connected(y, num_outputs=64, activation_fn=tf.nn.relu,\n                                              weights_initializer=tf.contrib.layers.xavier_initializer())\n        y = tf.contrib.layers.fully_connected(y, num_outputs=1, activation_fn=tf.nn.sigmoid,\n                                              weights_initializer=tf.contrib.layers.xavier_initializer())\n\n        predictions = y\n\n        # 2. Define the loss function for training\/evaluation\n        if mode == tf.estimator.ModeKeys.TRAIN or mode == tf.estimator.ModeKeys.EVAL:\n            loss = tf.reduce_mean((predictions - labels) ** 2)\n        else:\n            loss = None\n\n        if mode != tf.estimator.ModeKeys.PREDICT:\n            eval_metric_ops = {\n                \"rmse\": tf.metrics.root_mean_squared_error(tf.cast(labels, tf.float32), predictions),\n                \"accuracy\": tf.metrics.accuracy(tf.cast(labels, tf.float32), predictions),\n                \"precision\": tf.metrics.precision(tf.cast(labels, tf.float32), predictions)\n            }\n        else:\n            eval_metric_ops = None\n\n        # 3. Define the training operation\/optimizer\n        if mode == tf.estimator.ModeKeys.TRAIN:\n            train_op = tf.contrib.layers.optimize_loss(\n                loss=loss,\n                global_step=tf.contrib.framework.get_global_step(),\n                learning_rate=0.001,\n                optimizer=\"Adam\")\n        else:\n            train_op = None\n\n        if mode == tf.estimator.ModeKeys.PREDICT:\n            predictions_dict = {\"pred\": predictions}\n        else:\n            predictions_dict = None\n\n        # 5. Return predictions\/loss\/train_op\/eval_metric_ops in ModelFnOps object\n        return tf.estimator.EstimatorSpec(mode=mode,\n                                          predictions=predictions_dict,\n                                          loss=loss,\n                                          train_op=train_op,\n                                          eval_metric_ops=eval_metric_ops)\n    return _build_model\n<\/code><\/pre>\n\n<p>And you can use this model then like this:<\/p>\n\n<pre><code>e = tf.estimator.Estimator(model_fn=model_fn(), params=None)\ne.train(input_fn=input_fn(), steps=1000)\n<\/code><\/pre>\n\n<p>An example of an input-function for TensorFlow r1.1 can be found in my answer <a href=\"https:\/\/stackoverflow.com\/questions\/42792518\/setting-batch-size-when-using-input-fn-for-tf-contrib-learn-estimator\/43658096#43658096\">here<\/a>.<\/p>\n"},{"AnswerId":"42913748","CreationDate":"2017-03-20T20:52:14.203","ParentId":null,"OwnerUserId":"7739713","Title":null,"Body":"<p>You can import SKCompat from tensorflow.contrib.learn.python:<\/p>\n\n<pre><code>from tensorflow.contrib.learn.python import SKCompat\n<\/code><\/pre>\n\n<p>And then wrap your estimator with SKCompat() e.g. like this:<\/p>\n\n<pre><code>classifier = SKCompat(tf.contrib.learn.LinearClassifier(args))\n<\/code><\/pre>\n"}]}
{"QuestionId":42451088,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-25T01:56:18.950","AcceptedAnswerId":null,"OwnerUserId":3627660.0,"Title":"Tensorflow : TypeError: expected string or bytes-like object","Body":"<p>I am beginner in tensorflow , I am trying to add summaries to a code of neural network from this link <a href=\"https:\/\/pythonprogramming.net\/rnn-tensorflow-python-machine-learning-tutorial\/\" rel=\"nofollow noreferrer\">https:\/\/pythonprogramming.net\/rnn-tensorflow-python-machine-learning-tutorial\/<\/a>\nI got an error but I could not know what is wrong?\nhere is the code<\/p>\n\n<pre><\/pre>\n\n<p>here is the error <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42518036","CreationDate":"2017-02-28T20:22:30.323","ParentId":null,"OwnerUserId":"7128661","Title":null,"Body":"<p><a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/summary\/scalar\" rel=\"nofollow noreferrer\"><code>tf.summary.scalar<\/code><\/a> expects a name as the first argument, not an array. This should work instead:<\/p>\n\n<pre><code>tf.summary.scalar(value.op.name, value)\n<\/code><\/pre>\n"}]}
{"QuestionId":42451275,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-25T02:24:44.090","AcceptedAnswerId":null,"OwnerUserId":6043772.0,"Title":"TypeError: concat() got an unexpected keyword argument 'axis'","Body":"<pre><\/pre>\n\n<p>I tried on tensorflow version r0.11 and r0.12 to pre-train fine tuning inception_resnet_v2 and resnet_150 . Both raised this error. Can you help me? Thanks<\/p>\n","answers":[]}
{"QuestionId":42451347,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-25T02:36:56.717","AcceptedAnswerId":null,"OwnerUserId":7620007.0,"Title":"Tensorboard shows nothing in chrome","Body":"<p>I'm new to use the tensorflow. when I start to use the tensorboard,I run the demo code , it runs with no error.but when I input  and open the chrome it shows nothing. So I follow the guide to find the error.\nI input<\/p>\n\n<pre><\/pre>\n\n<p>it shows<\/p>\n\n<pre><\/pre>\n\n<p>I input <\/p>\n\n<pre><\/pre>\n\n<p>it shows<\/p>\n\n<pre><\/pre>\n\n<p>but when I input<\/p>\n\n<pre><\/pre>\n\n<p>it shows<\/p>\n\n<pre><\/pre>\n\n<p>So I can't watch anything in the chrome.<\/p>\n\n<p>Could anyone help me?<\/p>\n\n<p>Thank you so much!<\/p>\n","answers":[]}
{"QuestionId":42451356,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-02-25T02:37:48.573","AcceptedAnswerId":null,"OwnerUserId":7045779.0,"Title":"How to select tensor element by other index tensor in tensorflow","Body":"<p>Let me give some example that I want.<\/p>\n\n<pre><\/pre>\n\n<p>I want to make the new tensor to select element of num_tensor by index_tensor.<\/p>\n\n<pre><\/pre>\n\n<p>Thank you for any help.<\/p>\n","answers":[]}
{"QuestionId":42453058,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><neural-network><deep-learning><keras>","CreationDate":"2017-02-25T07:03:43.623","AcceptedAnswerId":null,"OwnerUserId":3993741.0,"Title":"Is it possible to load neural network weights onto a new model with the same architecture, but of different size, in Keras? If so, how?","Body":"<p>For example if you have weights from a convolutional neural net layer of filter size 64x64, and you want to load these pre-trained weights onto a new convolutional neural net layer of different size (say 32x32 or 128x128), how can this be achieved?  Can you transform (oversample, undersample, repeat, average, interpolate, truncate, etc.) the weights to fit onto any size layer?  <\/p>\n\n<p>Are there examples of this in Keras?  <\/p>\n\n<p>Are there best practices?<\/p>\n","answers":[]}
{"QuestionId":42453131,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-25T07:13:25.340","AcceptedAnswerId":null,"OwnerUserId":7620584.0,"Title":"How do I print predictions in Tensorflow","Body":"<p>Ok, I worked through a sample of predicting whether or not a particular product quality was good or bad on Tensorflow.  My final section of code looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>It puts out a nice set of 10 epochs that top out at 0.527 accuracy, which I assume to be 52.7% accuracy.  <\/p>\n\n<pre><\/pre>\n\n<p>The problem is that now, I want to feed back in just 1 row of data from (presumably) a numpy array into Tensorflow to get a prediction.  How do I do this?  I assume it follows a pattern that looks like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42457443","CreationDate":"2017-02-25T14:41:50.987","ParentId":null,"OwnerUserId":"5176181","Title":null,"Body":"<p>It depends on your graph. If you check your <code>accuracy<\/code> node, it may looks like<\/p>\n\n<pre><code>tf.reduce_mean(tf.equal(my_prediction, correct_label))\n<\/code><\/pre>\n\n<p>where <code>correct_label<\/code> would be something related to <code>y_train<\/code>,\nand <code>my_prediction<\/code> would be some tf node, probably look like <code>tf.round(...)<\/code> depend on your implementation.<\/p>\n\n<p>What you have to do here is to find that <code>my_prediction<\/code> out, then you can get a prediction using<\/p>\n\n<pre><code>output = sess.run([my_prediction], feed_dict={X: [input]})\n<\/code><\/pre>\n"}]}
{"QuestionId":42453138,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-25T07:14:50.983","AcceptedAnswerId":null,"OwnerUserId":7620627.0,"Title":"Tensorflow automatically deletes old model files","Body":"<p>I use tf.train.Supervisor to start a session to train my model, and save model parameter every 1000 steps. But it seems that tensorflow would automatically delete old model files. Only 5 recent models were saved. Once model.ckpt-6000 is produced, the model.ckpt-1000 is deleted. But I can not find any documents on this operation.<\/p>\n","answers":[{"AnswerId":"42458231","CreationDate":"2017-02-25T15:58:21.537","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>Here are the docs:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver<\/a><\/p>\n\n<p>It has a variable called max_to_keep which defaults to 5.  So just call <\/p>\n\n<pre><code>tf.train.Saver(max_to_keep=100000) \n<\/code><\/pre>\n\n<p>or some other large number that encompasses all that might be created.  No way to save all models as an option that I know of.<\/p>\n\n<p>Update: as Ferran Par\u00e9s points out in a comment just set max_to_keep=None to save all of the outputs.<\/p>\n\n<pre><code>tf.train.Saver(max_to_keep=None) \n<\/code><\/pre>\n"}]}
{"QuestionId":42453546,"AnswerCount":3,"Tags":"<python><machine-learning><deep-learning><keras>","CreationDate":"2017-02-25T08:02:42.697","AcceptedAnswerId":null,"OwnerUserId":541106.0,"Title":"Keras model gets constant loss and accuracy","Body":"<p>I am trying to train a keras CNN against the Street View House Numbers Dataset. You can find the project <a href=\"https:\/\/github.com\/asindico\/digitRecognition\/tree\/9faa600f956e264ca61afe48d64c833f8b58711c\" rel=\"nofollow noreferrer\">here<\/a>.\nThe problem is that during training neither loss nor accuracy change over time. I have tried with 1 Channel (Gray Scale) images, with RGB (3 channels) images, with wider (50,50) and smaller (28,28) images, with more or less filters in the convolutional layers, with wider and smaller patches in the pooling layers, with and without dropout, with bigger and smaller batches, with smaller and bigger learning step for the optimizers, with different optimizers, ...<\/p>\n\n<p>Still the training gets stuck to constant loss and accuracy<\/p>\n\n<p>Here is how I prepared the data<\/p>\n\n<pre><\/pre>\n\n<p>Here is the model<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42456811","CreationDate":"2017-02-25T13:41:16.157","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>why are the labels for 104 : <code>[10 10  1 10  4]<\/code> ? I believe it should be <code>[10 10 1 0 4]<\/code>, no?<\/p>\n\n<p>In my opinion : either you have a problem with the input data (the preparation might be wrong), or you have an architecture not suitable for this problem.<\/p>\n\n<p>It is training, you can see on the notebook there is a change of loss between epoch 1 and 2. So it's not a training issue.<\/p>\n"},{"AnswerId":"42456556","CreationDate":"2017-02-25T13:16:31.093","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>In cases like this it is most of the time a wrong training set. I would recommend you to take a look at the actual images and labels you feed into the network. Additionally look at the actual colorbar of the images. This means seeing how their values are distributed. This often leads to the solution. Anyways, if you are able to map them, then so will the computer given a good learning rate.<\/p>\n"},{"AnswerId":"42465134","CreationDate":"2017-02-26T05:37:49.063","ParentId":null,"OwnerUserId":"1862500","Title":null,"Body":"<p>As my comment above, I'd suggest to avoid training a model to predict 5 digits combination. It would be far more efficient to train the model to predict a single number. I tried to build quick example based on Keras example <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/cifar10_cnn.py\" rel=\"nofollow noreferrer\">cifar10_cnn.py<\/a> on <a href=\"http:\/\/ufldl.stanford.edu\/housenumbers\/\" rel=\"nofollow noreferrer\">MNIST SHVN format 2 (cropped digits)<\/a>:<\/p>\n\n<pre><code>import numpy as np\nimport scipy.io as sio\nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout, Activation, Flatten\nfrom keras.layers import Convolution2D, MaxPooling2D\nfrom keras.utils.np_utils import to_categorical\n\n# parameters\nnb_epoch = 10\nbatch_size = 32\n\n# load data\nnb_classes = 10\ntrain_data = sio.loadmat('train_32x32.mat')\ntest_data = sio.loadmat('test_32x32.mat')\nX_train = train_data['X'].T \/ 255\nX_test = test_data['X'].T \/ 255\ny_train = to_categorical(train_data['y'] % nb_classes)\ny_test = to_categorical(test_data['y'] % nb_classes)\n\n# model\nmodel = Sequential()\nmodel.add(Convolution2D(32, 3, 3, border_mode='same', input_shape=X_train.shape[1:]))\nmodel.add(Activation('relu'))\nmodel.add(Convolution2D(32, 3, 3))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\nmodel.add(Convolution2D(64, 3, 3, border_mode='same'))\nmodel.add(Activation('relu'))\nmodel.add(Convolution2D(64, 3, 3))\nmodel.add(Activation('relu'))\nmodel.add(MaxPooling2D(pool_size=(2, 2)))\nmodel.add(Dropout(0.25))\nmodel.add(Flatten())\nmodel.add(Dense(512))\nmodel.add(Activation('relu'))\nmodel.add(Dropout(0.5))\nmodel.add(Dense(nb_classes))\nmodel.add(Activation('softmax'))\nmodel.compile(loss='categorical_crossentropy', optimizer='rmsprop', metrics=['accuracy'])\n\n# train\nmodel.fit(X_train, y_train, batch_size=batch_size, nb_epoch=nb_epoch, validation_data=(X_test, y_test), shuffle=True)\n<\/code><\/pre>\n\n<p>Once you trained the model, train another model to <a href=\"http:\/\/www.pyimagesearch.com\/2017\/02\/13\/recognizing-digits-with-opencv-and-python\/\" rel=\"nofollow noreferrer\">recognize\/extract each number from an image<\/a> using library such as <a href=\"https:\/\/pypi.python.org\/pypi\/opencv-python\" rel=\"nofollow noreferrer\">OpenCV<\/a><\/p>\n"}]}
{"QuestionId":42453622,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-25T08:12:38.287","AcceptedAnswerId":null,"OwnerUserId":4993842.0,"Title":"Tensorflow padding with 1s instead of zeroes","Body":"<p>Tensorflow functions seem to use only zeroes for padding (for example, the  function). <\/p>\n\n<p>Is there an easy way (that requires minimal changes in existing code) to run the same functions with paddings of 1s or any other arbitrary numbers instead of zeroes?<\/p>\n","answers":[{"AnswerId":"42458126","CreationDate":"2017-02-25T15:49:43.010","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>The easiest way would be to change the padding in conv2d from \"SAME\" to \"VALID\" and perform the padding yourself.<\/p>\n\n<pre><code>img = tf.ones([img_height+padding,img_width+padding, channels])\nimg[padding\/\/2:img_height, padding\/\/2:img_width] = old_img\nresult = conv2d(img, kernel, padding=\"VALID\")\n<\/code><\/pre>\n"}]}
{"QuestionId":42453999,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-25T09:00:11.137","AcceptedAnswerId":null,"OwnerUserId":3091242.0,"Title":"Plotting Train\/Val Accuracies on same graph when using queues","Body":"<p>My input pipeline consists of a queue created using . This forces me to make separate input queues for train and validation data. Hence for getting predictions, need to pass both the queues through the model (with  the second time).<\/p>\n\n<p>Since now I have separate tensors for the train and validation prediction, I have to have separate  for them and they will come in separate graphs in tensorboard.<\/p>\n\n<p>How can I make two different scalar summaries plotted in the same graph in 1.0?<\/p>\n","answers":[]}
{"QuestionId":42454973,"AnswerCount":1,"Tags":"<python><image><numpy><matrix><tensorflow>","CreationDate":"2017-02-25T10:41:20.543","AcceptedAnswerId":null,"OwnerUserId":3262787.0,"Title":"Tensorflow: Slice Input image and grayscale","Body":"<p>I am new to Tensorflow and am trying to slice an image to then grayscale it. My code so far looks like this. I am first taking the input, then extract image data into a tensor, then slice the middle part of the image. The weights etc. can be disregarded for this step, just wanted to give the reader context:<\/p>\n\n<pre><\/pre>\n\n<p>Whenever I run this, however, I get an error stating:<\/p>\n\n<pre><\/pre>\n\n<p>I have read the documentation on slicing and resizing, but I am still confused. As said, I'm new to tensorflow, would be grateful for any help :)<\/p>\n\n<p>I tried commenting out to debug, but I'm not even sure if that caused the problems. I am always getting the above error.<\/p>\n","answers":[{"AnswerId":"42457993","CreationDate":"2017-02-25T15:37:12.227","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>Update:<\/p>\n\n<p>Looks like your using an older version of tensorflow and didn't state it though.  To fix your error change it to be like so.<\/p>\n\n<pre><code>    inputs = tf.image.resize_images(inputs, 84, 84)\n<\/code><\/pre>\n\n<p>where in this case the first 84 is the new height and the second one is the width.<\/p>\n\n<p>Old Answer Valid for tensorflow 1.0:\nI think your issue is that <\/p>\n\n<pre><code>inputs = tf.image.resize_images(inputs, [84, 84, 3])\n<\/code><\/pre>\n\n<p>should be<\/p>\n\n<pre><code>inputs = tf.image.resize_images(inputs, [84, 84])\n<\/code><\/pre>\n\n<p>as size should only describe height and width, and should not include number of channels.<\/p>\n"}]}
{"QuestionId":42455414,"AnswerCount":1,"Tags":"<tensorflow><out-of-memory><ubuntu-14.04>","CreationDate":"2017-02-25T11:23:50.013","AcceptedAnswerId":null,"OwnerUserId":6728986.0,"Title":"CUDA_ERROR_OUT_OF_MEMORY ubuntu 14.04 cuda8","Body":"<p>I am using tensorflow with cuda8 on ubuntu 14.04\nMy CPU: GeForce GT 740M\nI am a newbie to GPUs\nSometimes, after I have run the same script several times on the gpu, I will get a memory error, which will be gone the next time I reboot.\nThanks for sharing your expertise with me. I dont really know how to solve this problem.<\/p>\n\n<p>Here is the error message:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42457932","CreationDate":"2017-02-25T15:30:59.980","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>There are many reasons you could be getting this issue.  <\/p>\n\n<ul>\n<li>Check if you're using the GPU to also run X server because it crashed from the start.  Check with nvidia-smi to see how much space you actually have to work with.<\/li>\n<li>Make sure you have the appropriate CUDA drivers and toolkit version for the tensorflow you are running (367.35 or newer and toolkit 8.0)<\/li>\n<li>Is your card supported? (I think it should work but nvidia likes to be sneaky about supporting old hardware where they lock you out as a way to buy newer nvidia GPUs). After double checking your card is supported. Needs CUDA compute >= 3.0<\/li>\n<li>You can debug your code with the tensorflow debugger.<\/li>\n<li>Last but not least as comments have suggested it seems like your GPU resources aren't being freed after your software has ended.  Make sure you kill the process as the GPU will free the resources after the program calls exit().  <\/li>\n<\/ul>\n"}]}
{"QuestionId":42456461,"AnswerCount":7,"Tags":"<windows><tensorflow><anaconda>","CreationDate":"2017-02-25T13:07:44.953","AcceptedAnswerId":43349812.0,"OwnerUserId":3359302.0,"Title":"Tensorflow: Not a supported wheel on this platform","Body":"<p>I've installed both anaconda 2 and 3 in a Windows 64bit machine. I'm trying to install Tensorflow with anaconda as per the instructions <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"noreferrer\">here<\/a>. However, I'm getting the following error while doing a pip install.<\/p>\n\n<pre><\/pre>\n\n<p>My anaconda info:<\/p>\n\n<pre><\/pre>\n\n<p>I would really appreciate any help since I'm banging my head on this since last 8 hours.<\/p>\n\n<h1>Edit<\/h1>\n\n<p>I've removed Anaconda 2 and tried to install tensorflow again. Getting the same error. Please find the anaconda info.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43707687","CreationDate":"2017-04-30T14:49:45.753","ParentId":null,"OwnerUserId":"7862197","Title":null,"Body":"<p>Believe me I have survived it.<\/p>\n\n<p>First, please ensure that your system is Win7 64Bit and your Python version is 3.5.2(64bit), I didn't try Python 3.5.<\/p>\n\n<p>Second, don't believe Anaconda and PyCharm. Of course I know they can both be used to install packages. However, the packages installed by them can't be synchronized to cmd sometimes.<br>\nwhat I want to say is: Ensure the content of cmd is the same as the website of TensorFlow.<br>\nI wanted to upload more pictures to tell you how to do, but stackoverflow didn't allow it.<br>\nDetails are shown as follows:\n<img src=\"https:\/\/i.stack.imgur.com\/mt9Eb.png\" alt=\"enter image description here\">\n<img src=\"https:\/\/i.stack.imgur.com\/SCrrj.png\" alt=\"enter image description here\"><\/p>\n"},{"AnswerId":"43349812","CreationDate":"2017-04-11T15:09:59.043","ParentId":null,"OwnerUserId":"30889","Title":null,"Body":"<p>Explicitly use python 3.5 when you create your Anaconda environment.  IE:<\/p>\n\n<p>create -n tensorflow python=3.5<\/p>\n\n<p>then run the following command to install tensorflow:<\/p>\n\n<p>pip install --ignore-installed --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.1-cp35-cp35m-win_amd64.whl\" rel=\"noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.1-cp35-cp35m-win_amd64.whl<\/a><\/p>\n\n<p>Then you can successful do the Hello Tensorflow example from <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a><\/p>\n"},{"AnswerId":"43714686","CreationDate":"2017-05-01T05:27:30.500","ParentId":null,"OwnerUserId":"418234","Title":null,"Body":"<p>I had the same problem, I uninstalled Anaconda and installed python 3.5.2 from python.org and I was able to run Hello Tensorflow.<\/p>\n"},{"AnswerId":"42467264","CreationDate":"2017-02-26T10:20:50.907","ParentId":null,"OwnerUserId":"1342236","Title":null,"Body":"<p>I've encountered the same problem recently when trying to install tensorflow v1.0 on azure vm. None of anaconda versions worked for me. <br>\nSo I decided to try upgrading tensorflow to v1.0 on my local pc to see if the error will be reproducible. This resulted in the same error (even though I've installed r0.12 successfully just couple weeks ago).<\/p>\n\n<p>Reading tensorflow installation notes carefully I've noticed the following sentence:<\/p>\n\n<blockquote>\n  <p>The Anaconda installation is community supported, not officially\n  supported.<\/p>\n<\/blockquote>\n\n<p>So I've decided to try <a href=\"https:\/\/www.python.org\/downloads\/release\/python-352\/\" rel=\"nofollow noreferrer\">python 3.5.2 from python.org<\/a> and this worked immediately.<\/p>\n\n<p>I believe the problem here is that the latest tensorflow version (1.0) does not work with anaconda yet. \nNevertheless if you still want to use anaconda you can try <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.12\/get_started\/os_setup\" rel=\"nofollow noreferrer\">tensorflow r0.12<\/a>, but keep in mind that doing so is not recommended and there were breaking changes from r0.12 to 1.0<\/p>\n"},{"AnswerId":"43734283","CreationDate":"2017-05-02T09:19:15.907","ParentId":null,"OwnerUserId":"4814427","Title":null,"Body":"<p>Use python 3.5<\/p>\n\n<p><code>conda create -n tensorflow python=3.5<\/code><\/p>\n"},{"AnswerId":"44636005","CreationDate":"2017-06-19T16:52:35.207","ParentId":null,"OwnerUserId":"7697439","Title":null,"Body":"<p>Try this.<\/p>\n\n<p>pip install --ignore-installed --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.1-cp36-cp36m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.1-cp36-cp36m-win_amd64.whl<\/a><\/p>\n\n<p>this is the command that runs when you install with native pip. This change worked out for me<\/p>\n"},{"AnswerId":"42457638","CreationDate":"2017-02-25T15:01:15.373","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>On Microsoft Windows, TensorFlow needs Python 3.5 64-bit. You seem to use Python 2.7.<\/p>\n"}]}
{"QuestionId":42456505,"AnswerCount":0,"Tags":"<parallel-processing><tensorflow><distributed>","CreationDate":"2017-02-25T13:11:48.560","AcceptedAnswerId":null,"OwnerUserId":4646269.0,"Title":"Model parallel for arch that doesn't fit in 1 GPU","Body":"<p>I have a tensorflow model that is too big to fit into a single TITAN X GPU. I do however have access to 4 such GPUs. Does tensorflow support a model parallel training paradigm where different GPUs are training different layers of the same net?<\/p>\n\n<p>If so, where can I find resources that help me do do this?<\/p>\n","answers":[]}
{"QuestionId":42456608,"AnswerCount":0,"Tags":"<python><machine-learning><tensorflow><nlp><deep-learning>","CreationDate":"2017-02-25T13:22:23.850","AcceptedAnswerId":null,"OwnerUserId":2150541.0,"Title":"TensorFlow: How to customizingly gather \/ split a tensor?","Body":"<p>Please allow me to introduce the background briefly. I am working on a Bi-LSTM-CRF architecture to do a sequence labeling task. Instead of feeding features of individual steps encoded by LSTM into CRF layer, I want to collect feature vectors belonging to the same segment, and averaging them, then send the segment features into CRF. This method was proposed in the paper <em>Neural Models for Sequence Chunking<\/em>.<\/p>\n\n<p>Now we have a 3D tensor  <em>[batch_size, num_steps, num_features]<\/em> and a 2D tensor  <em>[batch_size, num_steps]<\/em> indicating the IOB labels of the sequence. How could I gather feature vectors belonging to the same segment and transform it to a 4D output tensor <em>[batch_size, max_segment_num, max_segment_len, num_features]<\/em>? The segment boundary is determined by the IOB labels, where <em>B<\/em> indicates the beginning and <em>I<\/em> the inner token.<\/p>\n\n<pre><\/pre>\n\n<p>I have investigated a few builtin functions like  or , but to my knowledge none of them supports a custom way. I also looked into the masking functions but didn't figure out how they can help.<\/p>\n\n<p>I tried manually manipulating the tensor and assign to a newly created 4D tensor, but was stuck there. The tensor X in the definition of the model is a placeholder, and is fed with data in training. Hence, the following statement fails, complaining that we must feed data to the placeholder now.<\/p>\n\n<pre><\/pre>\n\n<p>Could anyone give me a hint? How could I solve the problem or is there an easier way using builtin functions? Thanks in advance.<\/p>\n","answers":[]}
{"QuestionId":42458071,"AnswerCount":1,"Tags":"<compilation><tensorflow><opencl><fedora-25>","CreationDate":"2017-02-25T15:44:15.993","AcceptedAnswerId":null,"OwnerUserId":4429063.0,"Title":"TensorFlow: Compile Error for OpenCl Install from Source","Body":"<p>Hello I have followed the Instructions to install from source and after running .\/configure with clang\/++ and ComputeCpp path provided and Yes to OpenCL.<\/p>\n\n<p>With Bazel command :<\/p>\n\n<pre><\/pre>\n\n<p>I get the following compile error:<\/p>\n\n<pre><\/pre>\n\n<p>I looked up the Symbol and it belongs to Rust which is not installed on my system nor can I see how to install it or the library required by this Install of TensorFlow. It only happens when I install for OpenCL. So it seems it is used for OpenCL along with ComputeCpp. I am installing on Fedora 25 (Yey)<\/p>\n\n<p>Any ideas would be greatly appreciated. Also how do you stop it from downloading everytime you run .\/Configure can it not just read that it already has the files downloaded and if so where?<\/p>\n\n<p>Thanks\nHbyte<\/p>\n","answers":[{"AnswerId":"42647000","CreationDate":"2017-03-07T11:21:38.777","ParentId":null,"OwnerUserId":"4220358","Title":null,"Body":"<p>First off all, Fedora 25 is not officially supported by ComputeCpp yet. Although, we will do all we can to help.<\/p>\n\n<p>You shouldn't need to change any of the bazel build files.\nCould you confirm where is ComputeCpp installed?<\/p>\n\n<blockquote>\n  <p>I inserted ompute++ and have now found that header files located in my \/usr\/include path are not being used. <\/p>\n<\/blockquote>\n\n<p>Do you mean that your system headers are not picked after configuring TF to use OpenCL?<\/p>\n\n<p>From TensorFlow set up point of view you should point \n<code>.\/configure<\/code> to what's the location.. there is no need to copy headers \/ binaries anywhere.<\/p>\n\n<p>for instance my ComputeCpp is in my home directory<\/p>\n\n<p><code>Do you wish to build TensorFlow with OpenCL support? [y\/N] y\nOpenCL support will be enabled for TensorFlow\nDo you wish to build TensorFlow with CUDA support? [y\/N] \nNo CUDA support will be enabled for TensorFlow\nwhich: no clang++-3.6 in (\/usr\/lib64\/ccache:\/usr\/local\/bin:\/usr\/bin:\/usr\/local\/sbin:\/usr\/sbin:\/home\/luke\/.local\/bin:\/home\/luke\/bin)\nPlease specify which C++ compiler should be used as the host C++ compiler. [Default is ]: \/usr\/bin\/clang++\nPlease specify which C compiler should be used as the host C compiler. [Default is \/usr\/local\/bin\/clang-3.6]: \/usr\/bin\/clang\nPlease specify the location where ComputeCpp for SYCL 1.2 is installed. [Default is \/usr\/local\/computecpp]: \/home\/luke\/ComputeCpp-CE-0.1.2-Linux\n<\/code>\nAs of the whl package, it boils down to Google, but as far as I am aware TensorFlow is not officially supporting Fedora 25 either.<\/p>\n\n<p>If you have a fix for the build system please create Pull Request :)<\/p>\n\n<p>What happens if you choose GCC for your host compiler?<\/p>\n\n<p>Would it be possible to open an issue on GitHub to track the progress of this?<\/p>\n"}]}
{"QuestionId":42458415,"AnswerCount":1,"Tags":"<tensorflow><python-3.5><mnist>","CreationDate":"2017-02-25T16:14:36.083","AcceptedAnswerId":42461595.0,"OwnerUserId":3221293.0,"Title":"Is there any way to only import the MNIST images with 0's and 1's?","Body":"<p>I am just starting out with tensorflow and I want to test something only on the 0's and 1's from the MNIST images. Is there a way to import only these images?<\/p>\n","answers":[{"AnswerId":"42461595","CreationDate":"2017-02-25T21:07:39.517","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Assuming you are using <\/p>\n\n<p><code>from tensorflow.examples.tutorials.mnist import input_data<\/code><\/p>\n\n<p>No, there is no function or argument in that file... What you can do is load all data, and select only the ones and zeros. <\/p>\n"}]}
{"QuestionId":42458734,"AnswerCount":1,"Tags":"<python><pandas><numpy><keras>","CreationDate":"2017-02-25T16:42:42.567","AcceptedAnswerId":42458882.0,"OwnerUserId":3849791.0,"Title":"How do I convert all strings (like \"Fault\") and into a unique float?","Body":"<p>I have a DataFrame that has , , and  (strings with characters) items in it. I want a unique float for every unique object like so-<\/p>\n\n<pre><\/pre>\n\n<p>to<\/p>\n\n<pre><\/pre>\n\n<p>Also, will it work on all of the columns or would I have to do column by column?<\/p>\n\n<p>Last question, will it also convert all the 's to 's as well?<\/p>\n","answers":[{"AnswerId":"42458882","CreationDate":"2017-02-25T16:56:39.363","ParentId":null,"OwnerUserId":"4686625","Title":null,"Body":"<p>As mentioned by Jon, you could make use of <a href=\"http:\/\/pandas.pydata.org\/pandas-docs\/stable\/generated\/pandas.Series.factorize.html?highlight=factorize#pandas.Series.factorize\" rel=\"nofollow noreferrer\"><code>Series.factorize<\/code><\/a>. <\/p>\n\n<pre><code>(s.factorize()[0]+1).astype('float')\n<\/code><\/pre>\n\n<p>To perform this column-wise over an entire DataFrame, just use <code>apply<\/code>. <\/p>\n\n<p><strong>Demo<\/strong><\/p>\n\n<pre><code>&gt;&gt;&gt; s = pd.Series(['Exhaust', 'Fault', 'Probation', 5, int,\n                   'Exhaust', int, 'Fault', 'Motor'])\n\n&gt;&gt;&gt; s\n0          Exhaust\n1            Fault\n2        Probation\n3                5\n4    &lt;class 'int'&gt;\n5          Exhaust\n6    &lt;class 'int'&gt;\n7            Fault\n8            Motor\ndtype: object\n\n&gt;&gt;&gt; (s.factorize()[0]+1).astype('float')\narray([ 1.,  2.,  3.,  4.,  5.,  1.,  5.,  2.,  6.])\n<\/code><\/pre>\n\n<hr>\n\n<p>A NumPy solution may be to use the <code>return_inverse<\/code> keyword arg of <code>np.unique<\/code>, <\/p>\n\n<pre><code>(np.unique(s, return_inverse=True)[1]+1).astype('float')\n<\/code><\/pre>\n\n<p>however from some rough benchmarking the Pandas solution may be faster. <\/p>\n"}]}
{"QuestionId":42459726,"AnswerCount":2,"Tags":"<python><keras>","CreationDate":"2017-02-25T18:07:52.707","AcceptedAnswerId":null,"OwnerUserId":3849791.0,"Title":"Why does my keras model terminate and freeze my notebook if I do more than one epoch?","Body":"<p>I did 1  with batch sizes of 10 and it successfully completed. The accuracy rate was absolutely horrible coming in at a whopping 27%. I want to make it run on more than one epoch to see if the accuracy will, ideally, be above 80% or so, but it keeps freezing my Jupyter Notebook if I try to make it do more than one epoch. How can I fix this?<\/p>\n\n<p>My backend is Theano just for clarification.<\/p>\n\n<p>There is definitely a correlation between performance and batch_size. I tried doing batch_size=1 and it took 12s of horrifying, daunting, unforgivable time out of my day to do 1 epoch.<\/p>\n","answers":[{"AnswerId":"42460023","CreationDate":"2017-02-25T18:33:46.053","ParentId":null,"OwnerUserId":"3849791","Title":null,"Body":"<p>It takes time to run through the epochs and sometimes it looks like it freezes, but it still runs and if you wait long enough it will finish. Increasing the batch size makes it run through the epochs faster.<\/p>\n"},{"AnswerId":"53147988","CreationDate":"2018-11-05T03:36:59.087","ParentId":null,"OwnerUserId":"8768111","Title":null,"Body":"<p>You probably have to look into more factors.\nLook at the system resources, e.g CPU, Memory, Disk IO.  (If you use linux, run <code>sar<\/code> command)<\/p>\n\n<p>For me, I had other problem with frozen notebook, and it turns out to be the issue of low memory.<\/p>\n"}]}
{"QuestionId":42459931,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-25T18:25:08.313","AcceptedAnswerId":42494381.0,"OwnerUserId":6842947.0,"Title":"Accessing epoch value across multiple threads using input_producer\/limit_epochs\/epochs:0 local variable","Body":"<p>I tried to extract the current epoch number while reading data using multiple cpu threads. However during a trial code I observed an output which did not make any sense. Consider the code below :<\/p>\n\n<pre><\/pre>\n\n<p>The  function is defined as below :<\/p>\n\n<pre><\/pre>\n\n<p>The output I obtain is the following :<\/p>\n\n<pre><\/pre>\n\n<p>The last number in each line corresponds to the value of  local variable.<\/p>\n\n<p>For a first trial, I kept only 10 images in the queue meaning I should get a total of 40 lines of output, which I get.<\/p>\n\n<ul>\n<li>However, I should get equal number of 1,2, 3 and 4 as the last character in each line, since each filename should be extracted in each of the 4 epochs.<\/li>\n<\/ul>\n\n<p>Why am I getting the same number 4 in all the lines ?<\/p>\n\n<p><strong>Further Information<\/strong><\/p>\n\n<ul>\n<li>I tried using range(1) (for a single thread), and still the same observation.<\/li>\n<li>Don't bother with the digit '0'. It is simply the label of the corresponding file. I saved the image file names in such a way.<\/li>\n<\/ul>\n","answers":[{"AnswerId":"42494381","CreationDate":"2017-02-27T19:41:16.860","ParentId":null,"OwnerUserId":"6842947","Title":null,"Body":"<p>I did a lot of experiments and finally concluded the following :<\/p>\n\n<p>I used to believe that -<\/p>\n\n<blockquote>\n  <p><code>tf.train.string_input_producer()<\/code> enqueues a queue epoch-wise.<br>\n  Meaning that, first one complete epoch is enqueued (in multiple<br>\n  stages if capacity is less than the number of filenames) and then<br>\n  further epochs are enqueued.<\/p>\n<\/blockquote>\n\n<p>It is not really the case.<\/p>\n\n<blockquote>\n  <p>When <code>tf.start_queue_runners()<\/code> is executed, all the epochs are\n  enqueued together (in multiple stages if capacity is less than number\n  of filenames). The local variable <code>epochs:0<\/code> is used by <code>tf.train.string_input_producer<\/code> to maintain the epoch that is being enqueued. Once <code>epochs:0<\/code> reaches <code>num_epochs<\/code>, it remains constant and no matter how many threads are dequeuing from the queue, it does not change.<\/p>\n<\/blockquote>\n\n<p>When you capture the value of <code>epochs:0<\/code> it gives you the instantaneous value of the counter <code>epochs<\/code> and it tells you that at that time which epoch of the dataset is being enqueued. It does not tell you that which epoch of the dataset are you dequeuing. <\/p>\n\n<p>So it is a bad idea to get the value of the current epoch from the <code>epochs:0<\/code> <code>local_variable<\/code>.<\/p>\n"}]}
{"QuestionId":42460530,"AnswerCount":1,"Tags":"<python><pandas><numpy><keras>","CreationDate":"2017-02-25T19:19:26.707","AcceptedAnswerId":42461298.0,"OwnerUserId":3849791.0,"Title":"How do I make keras predict something other than a one-hot matrix?","Body":"<p>I have a dictionary of 122 unique values. I'm feeding the program over 45,000 records with 33 data points to refer to when making a prediction about what the output should be. What I've noticed is that it's only predicting . I need it to predict 1's 2's 3's ... up until 122. All are floats as well, so I don't know if that matters.<\/p>\n\n<p>Here is my code-<\/p>\n\n<pre><\/pre>\n\n<p>This is how I intend on associating the output with the actual string value (like 'Exhaust')<\/p>\n\n<pre><\/pre>\n\n<p>I use this to turn all data into unique floats in the DataFrame.<\/p>\n\n<pre><\/pre>\n\n<p>Thanks for your help!<\/p>\n","answers":[{"AnswerId":"42461298","CreationDate":"2017-02-25T20:34:31.073","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>Your model is wrong in this case, change it to:<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(Dense(12, input_dim=7, init='uniform', activation='relu'))\nmodel.add(Dense(7, init='uniform', activation='relu'))\nmodel.add(Dense(122, init='uniform', activation='softmax'))\n\nmodel.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\n<\/code><\/pre>\n\n<p>And make sure the input data is one-hot encoded with the function keras.utils.np_utils.to_categorical. Then the model should train.<\/p>\n\n<p>To then recover the integer class indices, when you predict with the model, you take the argmax of the output probability distribution, so you get the array index with the maximum probability.<\/p>\n\n<pre><code>x = some test point.\ny = model.predict(x)[0]\n\nclass = np.argmax(y)\n<\/code><\/pre>\n"}]}
{"QuestionId":42461406,"AnswerCount":1,"Tags":"<python-3.x><machine-learning><keras><conv-neural-network>","CreationDate":"2017-02-25T20:45:41.063","AcceptedAnswerId":42461764.0,"OwnerUserId":1144431.0,"Title":"Loading a pre trained Keras model and predicting","Body":"<p>I've cobbled together a simple neural network using some Keras examples, on the basic Kaggle Cat vs Dog data (<a href=\"https:\/\/www.kaggle.com\/c\/dogs-vs-cats-redux-kernels-edition\/data\" rel=\"nofollow noreferrer\">https:\/\/www.kaggle.com\/c\/dogs-vs-cats-redux-kernels-edition\/data<\/a>). I was able to train and save a model using<\/p>\n\n<pre><\/pre>\n\n<p>But when I try loading the model to predict, I get <\/p>\n\n<pre><\/pre>\n\n<p>Here's where the images are saved in my PyCharm solution. There are 1-12500 .jpg files in the Test directory, 11500 labeled .jpgs in each of the Training set cat and dog directories, and 1000 labeled .jpgs in each of the validate directories.\n<a href=\"https:\/\/i.stack.imgur.com\/NIXlG.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/NIXlG.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>And here's my code<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42461764","CreationDate":"2017-02-25T21:22:54.873","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>I believe that what you feed to the <code>predict_generator<\/code> isn't in the right format. <\/p>\n\n<p>When you predict values, in opposition with training and evaluation, you don't want to feed the labels. \nTherefore I would try to change your <code>test_generator<\/code> to this :<\/p>\n\n<pre><code>test_generator = train_datagen.flow_from_directory(\n     'test',\n     target_size=(64, 64),\n     batch_size=32,\n     #This will not output the targets. \n     class_mode=None)\n<\/code><\/pre>\n\n<p>You can find <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\">documentation about the ImageDataGenerator here<\/a>. <\/p>\n"}]}
{"QuestionId":42461673,"AnswerCount":1,"Tags":"<python><neural-network><keras><embedding>","CreationDate":"2017-02-25T21:15:16.420","AcceptedAnswerId":42462568.0,"OwnerUserId":999346.0,"Title":"Apply a shared Embedding layer on a set of documents in keras","Body":"<p>I am trying to create a model in which I want to predict the order of a certain set of documents given a certain query. My idea was basically to use a shared embedding layer for both the query and the documents, then merge the two \"branches\" using a cosine similarity between each document and the query (using a custom lambda). The loss function would then compute the difference between the expected position and the predicted similarity.<\/p>\n\n<p>My question is: Is there a way to create Embeddings for a set of textual features (provided that they have the same length)?<\/p>\n\n<p>I can properly transform my query in a \"doc2vec-like embedding\" by applying Embedding + Convolution1D + GlobalMaxPooling1D, but I had no luck using the same strategy on the sets of documents (and Reshaping + 2D convolutions don't really make sense to me given that I am working with textual data).<\/p>\n\n<p>Note that a constraint I have is that I need to use the same Embedding layer for both my query and the set of documents (I am using the Keras' functional apis to do so).<\/p>\n\n<p>[EDIT, adding sample code]<\/p>\n\n<pre><\/pre>\n\n<p><strong>[SOLUTION]<\/strong> thanks to Nassim Ben, use  like this to apply recurrently a Layer to all the dimensions of a layer like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42462568","CreationDate":"2017-02-25T22:52:54.103","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Alright. If I understand correctly the situation, you have 50 text snippets of length 50 that you want to embed. <\/p>\n\n<p>After doing the word embeddings, you find yourself with a Tensor T of shape (50,50,emb_size). \nWhay I would do is to use a LSTM layer in a TimeDistributed wrapper. Adding those lines after <code>emb(T)<\/code> :<\/p>\n\n<pre><code>right = TimeDistributed(LSTM(5))(right)\n<\/code><\/pre>\n\n<p>This will apply the same LSTM to each of the 50 documents and output a final state of length 5 at the end of each document processing. The shape of right after this step is (50,5). You have embedded each document in a length 5 vector. \nThe advantage of TimeDistributed is that the LSTM applied to each document will share the same weights so your documents will be 'treated' the same way. You can find documentation about LSTM <a href=\"https:\/\/keras.io\/layers\/recurrent\/#lstm\" rel=\"nofollow noreferrer\">here<\/a> and about TimeDistributed <a href=\"https:\/\/keras.io\/layers\/wrappers\/\" rel=\"nofollow noreferrer\">here<\/a>. <\/p>\n\n<p>I hope this helps a bit. <\/p>\n"}]}
{"QuestionId":42461873,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-02-25T21:35:15.587","AcceptedAnswerId":42462448.0,"OwnerUserId":665023.0,"Title":"feature number in tensorflow tf.nn.conv2d","Body":"<p>In the Tensorflow example \"Deep MNIST for Experts\" <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/pros\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/get_started\/mnist\/pros<\/a><\/p>\n\n<p>I am not clear how to determine the feature number specified in weight of activation function.<\/p>\n\n<p>For example:<\/p>\n\n<blockquote>\n  <p>We can now implement our first layer. It will consist of convolution,\n  followed by max pooling. The convolution will compute 32 features for\n  each 5x5 patch.<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>Why 32 is picked here?<\/p>\n\n<blockquote>\n  <p>In order to build a deep network, we stack several layers of this\n  type. The second layer will have 64 features for each 5x5 patch.<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>Again, why 64 is picked\uff1f<\/p>\n\n<blockquote>\n  <p>Now that the image size has been reduced to 7x7, we add a\n  fully-connected layer with 1024 neurons to allow processing on the\n  entire image.<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>Why 1024 here?<\/p>\n\n<p>Thanks<\/p>\n","answers":[{"AnswerId":"42462448","CreationDate":"2017-02-25T22:41:30.457","ParentId":null,"OwnerUserId":"6842947","Title":null,"Body":"<p>32 and 64 are number of filters in the respective layers.\n1024 is the number of output neurons in the fully connected layer.\nYour question basically is about the reason behind the choice of these hyperparameters. <\/p>\n\n<p>There is no mathematical or programming reason behind these specific choices. These have been picked up after experiments as they delivered a good accuracy over MNIST dataset.\nYou can change these numbers and that is one way by which you can modify a model. \nUnfortunately you cannot yet explore the reason for the choice behind these parameters within TensorFlow or any other literature source.<\/p>\n"},{"AnswerId":"42468742","CreationDate":"2017-02-26T12:51:11.910","ParentId":null,"OwnerUserId":"5209752","Title":null,"Body":"<p>Each of these filters will actually do something, like check for edges, check for colour change, or right-shift, left-shit the image, sharpen, blur etc.\nEach of these filters are actually working on finding out the meaning of the image by sharpening, enhancing, smoothening, intensifying etc.<\/p>\n\n<p>For e.g. check this link which explains the meaning of these filters \n<a href=\"http:\/\/setosa.io\/ev\/image-kernels\/\" rel=\"nofollow noreferrer\">http:\/\/setosa.io\/ev\/image-kernels\/<\/a><\/p>\n\n<p>So all these filters are actually neurons where the output will be max-pooled and eventually fed into a FC layer after some activation.<\/p>\n\n<p>If you are looking for just understanding the filters, that is another approach. However if you are looking to learn how conv. architectures work but since these are tried and tested filters over the dataset, you hsould just go with it for now.<\/p>\n\n<p>The filters also learn through Backprop.<\/p>\n"}]}
{"QuestionId":42462431,"AnswerCount":2,"Tags":"<python-3.5><keras>","CreationDate":"2017-02-25T22:39:29.400","AcceptedAnswerId":47958486.0,"OwnerUserId":5352399.0,"Title":"OSError: broken data stream when reading image file","Body":"<p>I am trying to read an image file using Image package of Keras. <\/p>\n\n<p>Here is my code.<\/p>\n\n<pre><\/pre>\n\n<p>When I run the code, I get the following error.<\/p>\n\n<pre><\/pre>\n\n<p>Please note, if I convert  to , then the given code works perfectly. But I have several thousands of pictures and I can't convert all of them to png format. I tried several things after searching for solution in web but couldn't get rid of the problem.<\/p>\n\n<p>Any help would be appreciated!<\/p>\n","answers":[{"AnswerId":"42463493","CreationDate":"2017-02-26T00:54:14.560","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>According to <a href=\"https:\/\/github.com\/python-pillow\/Pillow\/issues\/1057#issuecomment-219211437\" rel=\"nofollow noreferrer\">here<\/a> <code>Pillow<\/code> upgrade by <code>pip install Pillow --upgrade<\/code> should solve this issue. <\/p>\n\n<p>If you are still facing the problem you can use <code>mogrify<\/code> to batch convert all your images. <code>mogrify -format png *.jpg<\/code><\/p>\n"},{"AnswerId":"47958486","CreationDate":"2017-12-24T06:18:38.173","ParentId":null,"OwnerUserId":"5254777","Title":null,"Body":"<p>Use this at the beginning of your code:<\/p>\n\n<pre><code>from PIL import Image, ImageFile\nImageFile.LOAD_TRUNCATED_IMAGES = True\n<\/code><\/pre>\n\n<p>I found it <a href=\"https:\/\/github.com\/python-pillow\/Pillow\/issues\/1510\" rel=\"noreferrer\">here<\/a>. And this is working for me.<\/p>\n"}]}
{"QuestionId":42462459,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-25T22:42:10.563","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"Why do TensorFlow programs that use the GPU have a much higher virtual size than TensorFlow programs running on CPU only?","Body":"<p>On my Ubuntu 14.04.5 LTS x64 server, I have 3 TensorFlow programs that run on the GPU, and 2 TensorFlow programs that run on the CPU:<\/p>\n\n<pre><\/pre>\n\n<p> is the same as  except that the former runs on the CPU while the latter runs on the GPU.<\/p>\n\n<p>Why do TensorFlow processes that use the GPU have a much higher  <a href=\"https:\/\/serverfault.com\/q\/48582\/126950\">virtual size<\/a> ( column) than TensorFlow processes running on CPU only (776GB vs. ~12GB)?<\/p>\n","answers":[{"AnswerId":"42462771","CreationDate":"2017-02-25T23:15:33.090","ParentId":null,"OwnerUserId":"1845248","Title":null,"Body":"<p>As your link indicates a bit, the virtual size is based on the virtual memory addresses allocated to the program and does not (accurately) represent the amount of data that is stored or used by the program. Virtual memory addresses may point to locations on the GPU that are 4 bytes each while the addresses could each represent 1MB on the CPU, thus inflating the represented virtual size by a factor of ~33.000 for those addresses.<\/p>\n"}]}
{"QuestionId":42462470,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-25T22:44:04.593","AcceptedAnswerId":null,"OwnerUserId":7466373.0,"Title":"Install tensorflow from sources","Body":"<p>For resolving \"AttributeError: 'module' object has no attribute 'GRUCell'\" i was told to install tensorflow from sources. I tried a lot from the tensorflow website but it did not happen. I need to execute sequence to sequence models. <\/p>\n\n<p>Can anyone please help me.<\/p>\n\n<p>Thanks a lot in advance<\/p>\n","answers":[]}
{"QuestionId":42462864,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-25T23:26:49.337","AcceptedAnswerId":null,"OwnerUserId":2318074.0,"Title":"Distributed batch normalization","Body":"<p>I have a deep neural net duplicated across multiple GPUs for training, as in the tensorflow CIFAR10 tutorial. How can I add in batch normalization when each GPU can see only a portion of each batch?<\/p>\n","answers":[]}
{"QuestionId":42462895,"AnswerCount":1,"Tags":"<android><tensorflow>","CreationDate":"2017-02-25T23:30:59.553","AcceptedAnswerId":null,"OwnerUserId":5244666.0,"Title":"How can I use tf.Session.run() on Android?","Body":"<p>I want to make Tensorflow predictions on the Android. I found example Tensorflow Demo app for Android. But how can I use tf.Session.run,  tf.nn.softmax and other functions?<\/p>\n","answers":[{"AnswerId":"42519002","CreationDate":"2017-02-28T21:24:30.173","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>If you look at the Inference Library on Android, you'll see how you can load and run graphs:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/TensorFlowImageClassifier.java#L151\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/android\/src\/org\/tensorflow\/demo\/TensorFlowImageClassifier.java#L151<\/a><\/p>\n\n<p>Here is the definition for that library:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/android\/java\/org\/tensorflow\/contrib\/android\/TensorFlowInferenceInterface.java\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/android\/java\/org\/tensorflow\/contrib\/android\/TensorFlowInferenceInterface.java<\/a><\/p>\n\n<p>Does that help?<\/p>\n"}]}
{"QuestionId":42463594,"AnswerCount":4,"Tags":"<python><machine-learning><terminal><tensorflow><cpu>","CreationDate":"2017-02-26T01:11:34.817","AcceptedAnswerId":null,"OwnerUserId":7338578.0,"Title":"CPU instructions not compiled with TensorFlow","Body":"<p>MacBook Air: OSX El Capitan<\/p>\n\n<p>When I run TensorFlow code in terminal (), I get a longer than normal waiting time to get back output followed by these error messages: <\/p>\n\n<p><\/p>\n\n<p>I have no clue how to fix this. I would like to get TensorFlow to just work on this pip3 install. So I followed the path to: <\/p>\n\n<p>Do I need to edit the code here? Or is there an alternate way to get TensorFlow to compile with these instructions?<\/p>\n\n<p>I installed TensorFlow using .<\/p>\n","answers":[{"AnswerId":"42463698","CreationDate":"2017-02-26T01:28:44.940","ParentId":null,"OwnerUserId":"6842947","Title":null,"Body":"<p><strong>NOTE<\/strong> : These are not error messages but mere warning messages.<\/p>\n\n<p>The best way to maximise TF performance (apart from writing good code !!), is to compile it from the <a href=\"https:\/\/www.tensorflow.org\/performance\/performance_guide#build_and_install_from_source\" rel=\"noreferrer\">sources<\/a><\/p>\n\n<p>When you do that, TF would ask you for a variety of options which will also involve options for these instructions.<\/p>\n\n<p>In my own experience, compilation from the source is better in performance on an average.<\/p>\n\n<p>If you are doing some intensive processing that could be done on a GPU then that might also explain your waiting time.\nFor GPU support you would need to do <code>pip3 install tensorflow-gpu<\/code><\/p>\n"},{"AnswerId":"45816696","CreationDate":"2017-08-22T11:42:57.893","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>These are warning which mean it may be faster to build tensorflow on your pc from source.<\/p>\n\n<p>However if you want to disable them, you may use the code below<\/p>\n\n<pre><code>import os\nos.environ['TF_CPP_MIN_LOG_LEVEL']='2'\nimport tensorflow as tf\n<\/code><\/pre>\n\n<p>this should silence the warnings. 'TF_CPP_MIN_LOG_LEVEL' represents the Tensorflow environment variable responsible for logging. Also if you are on Ubuntu you may use this code below<\/p>\n\n<pre><code>export TF_CPP_MIN_LOG_LEVEL=2 \n<\/code><\/pre>\n\n<p>I hope this helps.<\/p>\n"},{"AnswerId":"42975991","CreationDate":"2017-03-23T12:24:52.090","ParentId":null,"OwnerUserId":"5825372","Title":null,"Body":"<blockquote>\n  <p>Those are simply warnings. They are just informing you if you build TensorFlow from source it can be faster on your machine. Those instructions are not enabled by default on the builds available I think to be compatible with more CPUs as possible.\n  If you have any other doubts regarding this please feel free to ask, otherwise this can be closed.<\/p>\n<\/blockquote>\n\n<p><code>Try export TF_CPP_MIN_LOG_LEVEL=2<\/code><\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7778\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7778<\/a><\/p>\n"},{"AnswerId":"42980553","CreationDate":"2017-03-23T15:33:04.513","ParentId":null,"OwnerUserId":"4352606","Title":null,"Body":"<p>You can also compile using bazel with opt arguments:<\/p>\n\n<pre><code>bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-msse4.1 --copt=-msse4.2 \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>I think you can find something in this discussion: <a href=\"https:\/\/stackoverflow.com\/questions\/41293077\/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions\">How to compile Tensorflow with SSE4.2 and AVX instructions?<\/a><\/p>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":42463896,"AnswerCount":1,"Tags":"<python><tensorflow><classification><prediction>","CreationDate":"2017-02-26T02:04:16.230","AcceptedAnswerId":42597280.0,"OwnerUserId":6265962.0,"Title":"Prediction using InceptionV3 in Tensorflow","Body":"<p>I have trained InceptionV3 model in tensor flow on my own dataset. I have the checkpoint file and the graph (.meta) from training. I am using these files to classify the label of a new image. Till now I have the following:<\/p>\n\n<p><strong>inception_v3 code from TFslim<\/strong><\/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\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p><strong>Prediction code<\/strong><\/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\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>This gives 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-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I am fairly new to Tensorflow so any help here is much appreciated. I am doing something wrong, but just cannot figure it out. Thanks in advance :)<\/p>\n\n<p><strong>EDIT 1<\/strong><\/p>\n\n<p>I restarted my session by reloading the graph and weights into the session. I froze the graph and saved it as frozen_graph.pb. Following is my code:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I load the frozen graph into a new session. Following are my node names:<\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I use the input node and final node (predictions) to classify a new image. Following is my code for 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\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>I get 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-html lang-html prettyprint-override\"><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>Not sure of what is going wrong now. I see my nodes are correct. Any help on this is much appreciated. <\/p>\n","answers":[{"AnswerId":"42597280","CreationDate":"2017-03-04T14:44:06.843","ParentId":null,"OwnerUserId":"6265962","Title":null,"Body":"<p>Finally I figured out how to resolve this issue. We have to freeze the graph in order to produce a .pb file that lets us re-use the nodes for prediction. Following is my code to do 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>def freeze_graph(model_folder):\r\n    # We retrieve our checkpoint fullpath\r\n    checkpoint = tf.train.get_checkpoint_state(model_folder)\r\n    input_checkpoint = checkpoint.model_checkpoint_path\r\n\r\n    # We precise the file fullname of our freezed graph\r\n    absolute_model_folder = \"\/\".join(input_checkpoint.split('\/')[:-1])\r\n    output_graph = absolute_model_folder + \"\/frozen_model.pb\"\r\n\r\n    # Before exporting our graph, we need to precise what is our output node\r\n    # This is how TF decides what part of the Graph he has to keep and what part it can dump\r\n    # NOTE: this variable is plural, because you can have multiple output nodes\r\n    output_node_names = \"tower_0\/logits\/predictions\"\r\n\r\n    # We clear devices to allow TensorFlow to control on which device it will load operations\r\n    clear_devices = True\r\n\r\n    # We import the meta graph and retrieve a Saver\r\n    saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=clear_devices)\r\n\r\n    # We retrieve the protobuf graph definition\r\n    graph = tf.get_default_graph()\r\n    input_graph_def = graph.as_graph_def()\r\n\r\n    # We start a session and restore the graph weights\r\n    with tf.Session() as sess:\r\n        saver.restore(sess, input_checkpoint)\r\n        for op in  sess.graph.get_operations():\r\n                print(op.name)\r\n        # We use a built-in TF helper to export variables to constants\r\n        output_graph_def = graph_util.convert_variables_to_constants(\r\n            sess, # The session is used to retrieve the weights\r\n            input_graph_def # The graph_def is used to retrieve the nodes \r\n           ,output_node_names.split(\",\")) # The output node names are used to select the usefull nodes\r\n\r\n\r\n        # Finally we serialize and dump the output graph to the filesystem\r\n        with tf.gfile.GFile(output_graph, \"wb\") as f:\r\n            f.write(output_graph_def.SerializeToString())\r\n        print(\"%d ops in the final graph.\" % len(output_graph_def.node))\r\n\r\n\r\nif __name__ == '__main__':\r\n    parser = argparse.ArgumentParser()\r\n    parser.add_argument(\"--model_folder\", type=str, help=\"Model folder to export\")\r\n    args = parser.parse_args()\r\n\r\n    freeze_graph(args.model_folder)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>After this, we can load the graph. Following is my code for 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>def load_graph(frozen_graph_filename):\r\n    # We load the protobuf file from the disk and parse it to retrieve the \r\n    # unserialized graph_def\r\n    with tf.gfile.GFile(frozen_graph_filename, \"rb\") as f:\r\n        graph_def = tf.GraphDef()\r\n        graph_def.ParseFromString(f.read())\r\n\r\n    # Then, we can use again a convenient built-in function to import a graph_def into the \r\n    # current default Graph\r\n    with tf.Graph().as_default() as graph:\r\n        tf.import_graph_def(\r\n            graph_def,\r\n            input_map=None,\r\n            return_elements=None,\r\n            name=\"prefix\",\r\n            op_dict=None,\r\n            producer_op_list=None\r\n        )\r\n    return graph<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>This will create your protocol buffer (frozen_graph.pb), which we can lo<\/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>if __name__ == '__main__':\r\n    # Let's allow the user to pass the filename as an argument\r\n    parser = argparse.ArgumentParser()\r\n    parser.add_argument(\"--frozen_model_filename\", default=\"\/serving\/frozen_model.pb\", type=str, help=\"Frozen model file to import\")\r\n    parser.add_argument(\"--image_name\",type=str,help=\"Image to test\")\r\n    args = parser.parse_args()\r\n\r\n    # Create test batch\r\n    image_data = create_test_batch(args.image_name)\r\n    # We use our \"load_graph\" function\r\n    graph = load_graph(args.frozen_model_filename)\r\n\r\n    # We can verify that we can access the list of operations in the graph\r\n    #for op in graph.get_operations():\r\n        #print(op.name)\r\n        # prefix\/Placeholder\/inputs_placeholder\r\n        # ...\r\n        # prefix\/Accuracy\/predictions\r\n\r\n    # We access the input and output nodes \r\n    x = graph.get_tensor_by_name('prefix\/batch_processing\/Reshape:0') # Input tensor\r\n    y = graph.get_tensor_by_name('prefix\/tower_0\/logits\/predictions:0') # Output tensor\r\n\r\n    # We launch a Session\r\n    with tf.Session(graph=graph) as sess:\r\n        # Note: we didn't initialize\/restore anything, everything is stored in the graph_def\r\n        y_out = sess.run(y, feed_dict={\r\n            x:image_data}) # &lt; 45    \r\n        print(y_out)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>My input node (x) expects a batch of size 64 and size 299x299x3 so I hacked my way by duplicating my test image 64 times and creating an input batch. I do this in the following way:<\/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 create_test_batch(input_image):\r\n        data = []\r\n        img = cv2.imread(input_image) # Read the test image\r\n        img_yuv = cv2.cvtColor(img, cv2.COLOR_BGR2YUV) # Convert RGB image to YUV\r\n        # equalize the histogram of the Y channel\r\n        img_yuv[:,:,0] = cv2.equalizeHist(img_yuv[:,:,0])\r\n        # convert the YUV image back to RGB format\r\n        img_output = cv2.cvtColor(img_yuv, cv2.COLOR_YUV2BGR)\r\n        img_resize = cv2.resize(img_output,(299,299)) # Resize the image acceptable by InceptionV3 model\r\n        for i in range(0,64):\r\n                data.append(img_resize) # Create a batch of 64 images \r\n                #data.append(np.resize((ndimage.imread('\/serving\/'+input_image)),(299,299,3)))\r\n        print np.shape(data)\r\n        return data<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n\n<p>If there is a better way of solving the input batch problem, I will appreciate an answer.<\/p>\n"}]}
{"QuestionId":42464018,"AnswerCount":2,"Tags":"<python><tensorflow><argparse>","CreationDate":"2017-02-26T02:25:06.253","AcceptedAnswerId":42464700.0,"OwnerUserId":5214063.0,"Title":"FLAGS = None meaning?","Body":"<p>I am new to python &amp; tensorFlow, and I am following <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/mnist\/mnist_softmax.py\" rel=\"noreferrer\">this MNIST tutorial<\/a> on tensorFlow documentation. <\/p>\n\n<p>In the first bit, I don't know what FLAGS = None does here. I searched in Google, and came back empty. Seems like this is too obvious to others? <\/p>\n\n<pre><\/pre>\n\n<p>So what is FLAGS and how it is used? \ne.g., FLAGS.data_dir<\/p>\n\n<p>Any help would be appreciated!<\/p>\n","answers":[{"AnswerId":"42464099","CreationDate":"2017-02-26T02:36:42.390","ParentId":null,"OwnerUserId":"6622817","Title":null,"Body":"<p>This was the full code you're looking at: I'll explain:<\/p>\n\n<pre><code>from __future__ import absolute_import\nfrom __future__ import division\nfrom __future__ import print_function\n\nimport argparse\nimport sys\n\nfrom tensorflow.examples.tutorials.mnist import input_data\n\nimport tensorflow as tf\n\nFLAGS = None   #Adds a default value to FLAGS\n\n\ndef main(_):  #Everything inside the function is not checked until it's called\n  mnist = input_data.read_data_sets(FLAGS.data_dir, one_hot=True) #FLAGS is not None anymore because it got changed below\n\n  x = tf.placeholder(tf.float32, [None, 784])\n  W = tf.Variable(tf.zeros([784, 10]))\n  b = tf.Variable(tf.zeros([10]))\n  y = tf.matmul(x, W) + b\n\n  y_ = tf.placeholder(tf.float32, [None, 10])\n\n  cross_entropy = tf.reduce_mean(\n      tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y))\n  train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)\n\n  sess = tf.InteractiveSession()\n  tf.global_variables_initializer().run()\n  # Train\n  for _ in range(1000):\n    batch_xs, batch_ys = mnist.train.next_batch(100)\n    sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})\n\n  # Test trained model\n  correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))\n  accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\n  print(sess.run(accuracy, feed_dict={x: mnist.test.images,\n                                      y_: mnist.test.labels}))\n\nif __name__ == '__main__': \n  parser = argparse.ArgumentParser()\n  parser.add_argument('--data_dir', type=str, default='\/tmp\/tensorflow\/mnist\/input_data',\n                      help='Directory for storing input data')\n\n  FLAGS, unparsed = parser.parse_known_args() #Here it changed the value of FLAGS to the first thing returned from parser.parse_known_args()\n\n  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed) #runs the app (calling main)\n<\/code><\/pre>\n\n<p>what was happening is that FLAGS got changed here:\n<code>FLAGS, unparsed = parser.parse_known_args()<\/code><\/p>\n"},{"AnswerId":"42464700","CreationDate":"2017-02-26T04:19:11.560","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>Initializing <code>FLAGS=None<\/code> is just a way of initializing the global constant.  If left as is it will raise an error in <code>main<\/code>, since <code>None<\/code> does not have any attributes.<\/p>\n\n<p>But if set via an <code>argparse<\/code> <code>parser<\/code> as shown in the fuller examples, it is a simple object with a variety of attributes.  <code>main<\/code> assumes one those attributes is called <code>data_dir<\/code>.<\/p>\n\n<p>If after the <\/p>\n\n<pre><code>FLAGS, unparsed = parser.parse_known_args()\nprint(FLAGS)\n<\/code><\/pre>\n\n<p>you should see <code>Namespace(data_dir='a directory', ....)<\/code>, where the value for <code>data_dir<\/code> was parsed from the command line.<\/p>\n"}]}
{"QuestionId":42464109,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><runtime-error><classification>","CreationDate":"2017-02-26T02:37:59.687","AcceptedAnswerId":42468652.0,"OwnerUserId":7201349.0,"Title":"Tensorflow ERROR: 'FileWriter' method","Body":"<pre><\/pre>\n\n<p>Hi, I'm using a python3.6 sample code from a YouTube tutorial with Tensorflow 1.0.<\/p>\n\n<p>I get an error at line 36 with the FileWriter method as the following:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Does anybody know how to fix this error and what the '\/home\/sergo\/work\/logs' pathname argument means?<\/strong><\/p>\n\n<p>Any help is very much appreciated.<\/p>\n\n<p>Thank you!<\/p>\n\n<p>(I'm using Mac 0SX Yosemite 10.10.5)<\/p>\n","answers":[{"AnswerId":"42468652","CreationDate":"2017-02-26T12:43:45.930","ParentId":null,"OwnerUserId":"5209752","Title":null,"Body":"<p>In the line, \nsummary_writer = tf.summary.FileWriter('\/home\/sergo\/work\/logs',graph_def = sess.graph_def)   #I GET AN ERROR IN THIS FileWriter method<\/p>\n\n<p>it explicitly mentions a path \/home\/sergo\/work\/logs<\/p>\n\n<p>This path may not exist in your local machine\nso please provide a path which exists and that should resolve your issue<\/p>\n\n<p>Thank you<\/p>\n"}]}
{"QuestionId":42464288,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><chatbot>","CreationDate":"2017-02-26T03:06:51.733","AcceptedAnswerId":42466486.0,"OwnerUserId":5636645.0,"Title":"What is the expected time of training for the following Seq2Seq model?","Body":"<p>Hello I am trying to test the following Seq2Seq model in order to get a Chatbot, I am following this tutorial:<\/p>\n\n<p><a href=\"http:\/\/suriyadeepan.github.io\/2016-06-28-easy-seq2seq\/\" rel=\"nofollow noreferrer\">http:\/\/suriyadeepan.github.io\/2016-06-28-easy-seq2seq\/<\/a><\/p>\n\n<p>And this is the main code:<\/p>\n\n<p><a href=\"https:\/\/github.com\/suriyadeepan\/easy_seq2seq\" rel=\"nofollow noreferrer\">https:\/\/github.com\/suriyadeepan\/easy_seq2seq<\/a><\/p>\n\n<p>The problem tha I am having is in the trainnig time, after to download the appropiate corpus I ran the following code to train:<\/p>\n\n<pre><\/pre>\n\n<p>Following the repository instructions, and the model became to train, this is the output, the main problem is that my computer has been computing the result for approximately 2 days and 9 hours, using all the processors, the specifications of my computer are the followings :<\/p>\n\n<pre><\/pre>\n\n<p>Taking this facts I would like to appreciate feedback for someone that has trained this model and if I have a kind of bug or if it is normal since is a very complex model, besides to know if my computer is going to be able to compute this data, below is the output that I am getting, thanks a lot for the support,<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42466486","CreationDate":"2017-02-26T08:45:51.187","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Training deep models on a CPU takes forever. If you plan on actually using deep learning technology you will have to get a gpu or use a pretrained one and even then I would recommend a gpu, because predicting is just that much faster. <\/p>\n"}]}
{"QuestionId":42464567,"AnswerCount":2,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2017-02-26T03:56:41.767","AcceptedAnswerId":null,"OwnerUserId":7623659.0,"Title":"How to install Tensorflow on Windows 10 with anaconda?","Body":"<p>I know that there exists a link for installing Tensorflow for python 3.5 on<br>\n<a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">Windows Installation link<\/a>. There also a similar question on <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">StackOverflow link<\/a> also, but it case when I use this command:<\/p>\n\n<pre><\/pre>\n\n<p>But it said Wheel package needs to be updated. So I updated it using pip and ran the application once again. The output was<\/p>\n\n<pre><\/pre>\n\n<p>So how do I solve this problem ? Please help.\nI've tried to install the cpu version of TensorFlow.<\/p>\n","answers":[{"AnswerId":"43866901","CreationDate":"2017-05-09T10:05:53.273","ParentId":null,"OwnerUserId":"7872066","Title":null,"Body":"<p>The executable, <a href=\"https:\/\/www.continuum.io\/downloads#windows\" rel=\"nofollow noreferrer\">Anaconda<\/a> for python 3.5 is not available on the official website. \nAn alternative to downloading that version is to download the latest version of Anaconda(3.6 as of 9 May, 2017), open your cmd shell in windows and execute the following commands. Note that the activate command is not fully supported in Windows Powershell. Click <a href=\"https:\/\/github.com\/conda\/conda\/issues\/1519\" rel=\"nofollow noreferrer\">here<\/a> to see why. <\/p>\n\n<pre><code>conda create --name tensorflow python=3.5\nactivate tensorflow\nconda install -c conda-forge tensorflow=1.0.0\n<\/code><\/pre>\n\n<p>The answer has been borrowed from Anaconda Public Google Group.<\/p>\n"},{"AnswerId":"42464593","CreationDate":"2017-02-26T03:59:56.927","ParentId":null,"OwnerUserId":"7623490","Title":null,"Body":"<pre><code>conda upgrade conda\nconda upgrade --all\nconda create -n tensorflow python=3.5.\nactivate tensorflow\nconda install -c conda-forge tensorflow\n<\/code><\/pre>\n\n<p>This will create an \"environment\" that will contain all of your packages you need (the example above is just tensorflow) and you will be able to import that library <strong>while you are in that environment<\/strong> anaconda is really just used to manage packages and segregate projects that require different packages<br>\n<br>\nWhen your finished with your environment, to close out use:<\/p>\n\n<pre><code>deactivate\n<\/code><\/pre>\n\n<p>these commands are slightly different on OSX\/Linux so be sure to look them up if you are on a different operating system\n<br>\n<br>\nIf you are using TF for some machine learning then you will probably want these packages in your environment as well:<\/p>\n\n<pre><code>conda install pandas matplotlib jupyter notebook scipy scikit-learn\n<\/code><\/pre>\n\n<p>Place that line between (activate tensorflow) and (conda install)<\/p>\n"}]}
{"QuestionId":42464581,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-02-26T03:58:49.010","AcceptedAnswerId":42465090.0,"OwnerUserId":7623868.0,"Title":"Missing `set_input` in keras","Body":"<p>everyone!<\/p>\n\n<p>I notice that in <a href=\"https:\/\/blog.keras.io\/keras-as-a-simplified-interface-to-tensorflow-tutorial.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/keras-as-a-simplified-interface-to-tensorflow-tutorial.html<\/a> it says that we can set tensorflow op as input of keras model like: . But I find that keras does not have set_input function: <\/p>\n\n<pre><\/pre>\n\n<p>But I get:<\/p>\n\n<pre><\/pre>\n\n<p>What may be the problem?<\/p>\n","answers":[{"AnswerId":"42465090","CreationDate":"2017-02-26T05:28:58.780","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>I guess <code>set_input()<\/code> method is removed in latest versions of Keras. If you see <a href=\"https:\/\/keras.io\/layers\/containers\/\" rel=\"nofollow noreferrer\">this documentation<\/a> of Keras, there is a function called <code>set_input()<\/code> function of <code>keras.layers.containers.Sequential<\/code> class. But its source code is no longer available on Github. <\/p>\n\n<p>If you look at the source code of <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/layers\/core.py#L699\" rel=\"nofollow noreferrer\">Dense<\/a> layer class in Keras, you will see that there is no such method called <code>set_input()<\/code> as well. If you also see the source of <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/engine\/topology.py#L218\" rel=\"nofollow noreferrer\">abstract class Layer<\/a> which is the base class for <code>Dense<\/code> layer, you will see there is no such function called <code>set_input()<\/code>. <\/p>\n\n<p>So, we can conclude, <code>set_input()<\/code> method is probably no longer available in Keras.<\/p>\n"}]}
{"QuestionId":42464742,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-26T04:25:58.273","AcceptedAnswerId":null,"OwnerUserId":3650053.0,"Title":"how to weight gradient in different examples","Body":"<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Assume I have a dataset , each row is an example with dim = 3, then I have a 3 * 2 weight matrix , my output is x_data * w. Then I want to get the gradient of the weight matrix. So I use  . But it seems to sum the gradients for each example directly. What I want is to sum the gradients of each example by a distribution. So I use a   as the distribution. It works well. However my problem is I don't know what to do if coef is also a tensorflow variable. When it also depends on w, I can't use  simply.<\/p>\n","answers":[{"AnswerId":"42529927","CreationDate":"2017-03-01T10:59:53.780","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>When you create your <code>coef<\/code> variable you can pass it the <code>trainable=False<\/code> parameter. With this your optimizer will not modify it during training. Citing the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"nofollow noreferrer\">official documentation<\/a>:<\/p>\n\n<blockquote>\n  <p><code>trainable<\/code>: If <code>True<\/code>, the default, also adds the variable to the graph\n  collection <code>GraphKeys.TRAINABLE_VARIABLES<\/code>. This collection is used as\n  the default list of variables to use by the <code>Optimizer<\/code> classes.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42466123,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-02-26T07:58:29.090","AcceptedAnswerId":42476325.0,"OwnerUserId":7624367.0,"Title":"Receiving random cost output on tensorflow regression- python","Body":"<p>I am relatively new to tensorflow and I have attempted to adapt some code from a tutorial to process my own data. <\/p>\n\n<p>The data can be found here: <a href=\"https:\/\/github.com\/z12332\/tensorflow-test-1\/blob\/master\/export.csv\" rel=\"nofollow noreferrer\">https:\/\/github.com\/z12332\/tensorflow-test-1\/blob\/master\/export.csv<\/a><\/p>\n\n<p>Keep in mind, that the dataset being fed consists of only columns 9 though 27 (with nan's converted to 0) and column 30 as the labels<\/p>\n\n<p>here is the link to the tutorial code: <a href=\"https:\/\/github.com\/llSourcell\/How_to_use_Tensorflow_for_classification-LIVE\/blob\/master\/demo.ipynb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/llSourcell\/How_to_use_Tensorflow_for_classification-LIVE\/blob\/master\/demo.ipynb<\/a><\/p>\n\n<p>I am able to get the program to run without an error message, but for some reason, it outputs 200 training steps of relatively random cost values. As and example, here are the first few steps:<\/p>\n\n<pre><\/pre>\n\n<p>Here is my code in full, for anyone knows why this is happening or how to debug:<\/p>\n\n<pre><\/pre>\n\n<p>Could the random values be from the missing values (changed to 0 in the dataset)? Also, how would I now apply the test_X,Y dataframes to predict? <\/p>\n\n<p>Either way Im here to learn so thanks for your help!<\/p>\n","answers":[{"AnswerId":"42476325","CreationDate":"2017-02-27T01:13:14.687","ParentId":null,"OwnerUserId":"4356204","Title":null,"Body":"<p>Here is a working version of the code, but before I'll offer some notes.<\/p>\n\n<p>1) One do some reading at <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/beginners\" rel=\"nofollow noreferrer\">tensorflow mnist tutorial<\/a>. \nIm particular see why your placeholder sizes are not correct and why \nwe are going to use a one-hot-encoded version of the labels for this task.<\/p>\n\n<p>2) Consider using cross-entropy cost. It is a more well suited cost\n for this multiclass task.<\/p>\n\n<p>3) try not to be too underwhelmed by the performance of this basic model\n (it does not perform well). Consider exploring the data looking for import features and also look around for what the state of the art performance on this dataset\nmight be.<\/p>\n\n<pre><code>import tensorflow as tf\nimport pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\n\ndf = pd.read_csv('\/Users\/benny\/desktop\/export.csv')\ndata_ = df.iloc[1:,9:27]\ndata_['CRISPR'] = df.iloc[:,30]\ndata_ = data_.drop(['Diseases'],axis=1)\n# we will need the number of classes to be predict\n# the nunique methods gets us the number of unique labels\nnclasses = data[\"CRISPR\"].nunique()\n# lets collect of labels here. We'll one-hot-encode them\n# using pandas.get_dummies()\ninputY = pd.get_dummies(data_.iloc[:, -1])\n\ndim = 16\nlearning_rate = 0.0000001\ndisplay_step = 50\n\nX = tf.placeholder(tf.float32, [None, dim])\n# Y should define the shape of your labels. \n# As discussed we're going to need one hot encoded labels for\n# this prediction task. this line does not define the shape of your input.\n# we'll define later\n# Y = tf.placeholder(tf.float32)\n\n\ntrain_X = data_.iloc[:200, :-2].as_matrix()\ntrain_X = train_X.fillna(value=0)\ntrain_Y = inputY[:200].as_matrix()\n\ntest_X = data_.iloc[200:320, :-2].as_matrix()\ntest_Y = inputY[200:320].as_matrix()\n\nn_samples = train_Y.size\n\n# Its important we get the shape of the weight and bias matrices\n# correct, the version in code is:\n# W = tf.Variable(tf.zeros([dim]), name=\"weight\")\n# that wont work since we want to be able to multiply [X, W]\n# to produce a evidence vector for each each example. \n# the shape of X is [200 x dim] - there should be a weight for each \n# feature and there 10 classes so  W is [dim, nclasses],\nW = tf.Variable(tf.zeros([dim, nclasses]))\n\n# for the bias, there should be one for each class.\n# b = tf.Variable(tf.zeros([1]), name=\"bias\")\nb = tf.Variable(tf.zeros([nclasses]))\n\n# the correct operation here is tf.matmal, I suspect you introduced\n# this to make your early matrix definiton work in the graph\n# activation = tf.add(tf.mul(X, W), b)\nactivation = tf.add(tf.matmul(X, W), b) \n\n# you forgot the actual model! Assuming you want to do\n# softmax classification let do:\ny = tf.nn.softmax(activation)\n# Now let's define or input labels ( we could have called them Y )\n# as you had them. Notice what we are saying here is expect a matrix\n# of floats with any number of examples and nclasses number of columns\n# which is exactly the size of train_Y.  \ny_ = tf.placeholder(tf.float32, [None, nclasses])\n\n\n# we define the cost reflect the fact that our model output is called\n# y not activations (anymore)\n# cost = tf.reduce_sum(tf.pow(activation-Y, 2))\/(2*n_samples)\ncost = tf.reduce_sum(tf.pow(y_ - y, 2))\/(2*n_samples)\noptimizer =  tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)\n\nhm_epochs = 1000\n\ninit = tf.initialize_all_variables()\nsess = tf.Session()\nsess.run(init)\n\n# I imagine what you want to do here is stochastic gradient descent. \n# I am not sure this is the way to do it. End to check the code I \n# will train over the entire training data for 1000 repetitions, \n# similar to the tutorial code.\n# .....\nfor i in range(hm_epochs):\n    sess.run(optimizer, feed_dict={X: train_X,\n                               y_: train_Y})\n\n    if (i) % display_step == 0:\n        cc = sess.run(cost, feed_dict={X: train_X,y_: train_Y})\n        print \"Training step:\", '%04d' % (i), \"cost=\", \"{:.9f}\".format(cc)\n\n\n# To check the accuracy ( this is one way of measuring the performance \n# of an algorithm on a classification task) we will do: (This is \n# adapted from the tensflow mnist example [code][1]) \n\ncorrect_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))\nprint(sess.run(accuracy, feed_dict={X: test_X,\n                              y_: test_Y}))\n<\/code><\/pre>\n"}]}
{"QuestionId":42466644,"AnswerCount":3,"Tags":"<machine-learning><neural-network><keras><lstm><recurrent-neural-network>","CreationDate":"2017-02-26T09:07:52.517","AcceptedAnswerId":42579854.0,"OwnerUserId":2499909.0,"Title":"Dimension Error while using LSTM after Convolutional neural network in Keras","Body":"<p>In keras 1.2.2, I have made a dataset which has these dimensions:<\/p>\n\n<pre><\/pre>\n\n<p>Here, 2000 is the number of instances (batches of data), 100 is the number of samples in each batch, 32 is the image rows and cols, and 3 is the number of channels (RGB).<\/p>\n\n<p>I have written this code which applies an LSTM after a CNN, however, I get this error:<\/p>\n\n<blockquote>\n  <p>ValueError: Input 0 is incompatible with layer lstm_layer: expected ndim=3, found ndim=2<\/p>\n<\/blockquote>\n\n<p>This is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42466722","CreationDate":"2017-02-26T09:17:48.210","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Are you sure that <code>fit()<\/code> is the right method to use? I would use <code>train_on_batch<\/code> since your training data is already split in batches.  <\/p>\n\n<p>To better control your input_shape, I would advise you to explicitely define it in your first layer. Instead of <code>X_train.shape[1:]<\/code>, use <code>(32,32,3)<\/code> to avoid any surprises. <\/p>\n\n<p>Does this help you?<\/p>\n"},{"AnswerId":"55761076","CreationDate":"2019-04-19T11:22:10.207","ParentId":null,"OwnerUserId":"10615173","Title":null,"Body":"<p>Output dimension of LSTM should be the same with your timesteps.\nTry this:<\/p>\n\n<pre><code>model.add(LSTM(output_dim=timesteps, return_sequences=True))\n<\/code><\/pre>\n\n<p>It works for me.<\/p>\n"},{"AnswerId":"42579854","CreationDate":"2017-03-03T13:13:27.867","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Try exchanging each <code>Convolution2D(...)<\/code> to:<\/p>\n\n<pre><code>TimeDistributed(Convolution2D(...))\n<\/code><\/pre>\n\n<p>You need to let your model know that your data is sequential and you want to apply some layer to each element in your sequence. This is what <code>TimeDistributed<\/code> wrapper is for.<\/p>\n"}]}
{"QuestionId":42466917,"AnswerCount":0,"Tags":"<tensorflow><magenta>","CreationDate":"2017-02-26T09:42:49.803","AcceptedAnswerId":null,"OwnerUserId":7624621.0,"Title":"drums_rnn_generate\uff1acommand not found","Body":"<p>when I tried to use magenta models\u2014\u2014\"<a href=\"https:\/\/github.com\/tensorflow\/magenta\/tree\/master\/magenta\/models\/drums_rnn\" rel=\"nofollow noreferrer\">Drums RNN<\/a>\" to generate my drum tracks, I got this info:<\/p>\n\n<pre><\/pre>\n\n<p>and here is my .sh file<\/p>\n\n<pre><\/pre>\n\n<p>I have installed magenta environment successfully,and activated magenta. it's strange that I can use command  which is from \"<a href=\"https:\/\/github.com\/tensorflow\/magenta\/tree\/master\/magenta\/models\/melody_rnn\" rel=\"nofollow noreferrer\">melody_rnn<\/a>\" without any problem.\nDoes someone get the same problem like mine?I want to know where the problem is and how to sovle it.Thanks anyway.<\/p>\n","answers":[]}
{"QuestionId":42467734,"AnswerCount":1,"Tags":"<python><image-processing><keras>","CreationDate":"2017-02-26T11:11:50.560","AcceptedAnswerId":42492239.0,"OwnerUserId":1394017.0,"Title":"Keras - How to use ImageDataGenerator without deforming aspect ratio","Body":"<p>I have a folder with images of different sizes. I tried to use ImageDataGenerator with flow_from_directory for batch loading \/ data augmentation.<\/p>\n\n<p>Is there a way to keep the aspect ratios of my images ? It seems like the images are being stretched to target_size : I would like to \"pad\" my images without deforming them (filling the gaps with a constant value)<\/p>\n\n<p>Here's my code :<\/p>\n\n<pre><\/pre>\n\n<p>Images are being stretched to (256,256).<\/p>\n","answers":[{"AnswerId":"42492239","CreationDate":"2017-02-27T17:36:58.230","ParentId":null,"OwnerUserId":"1394017","Title":null,"Body":"<p>I found an answer to my question.<\/p>\n\n<p>There's currently no way to keep aspect ratio when using ImageDataGenerator \/ flow_from_directory, but a <a href=\"https:\/\/github.com\/fchollet\/keras\/pull\/4987\" rel=\"noreferrer\">pull request<\/a> is opened on Github to add this feature.<\/p>\n"}]}
{"QuestionId":42468070,"AnswerCount":1,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-02-26T11:47:28.443","AcceptedAnswerId":42519049.0,"OwnerUserId":5801843.0,"Title":"Use subfolders in training set to retrain Inception's Final Layer","Body":"<p>In the example doc to <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining#training_on_your_own_categories\" rel=\"nofollow noreferrer\">retrain Inception's Final Layer<\/a> the training image data is stored in folders. Each folder defines the category for the containing images through its name.<\/p>\n\n<p>In the following example  belongs to the category  and  to the category :<\/p>\n\n<pre><\/pre>\n\n<p>Can I achieve higher granularity through subfolders? For example:<\/p>\n\n<pre><\/pre>\n\n<p>The doc is not clear about that.<\/p>\n","answers":[{"AnswerId":"42519049","CreationDate":"2017-02-28T21:27:01.113","ParentId":null,"OwnerUserId":"5708323","Title":null,"Body":"<p>No, unfortunately you can't create sub-classes like that. Instead, you can try naming your classes like this:<\/p>\n\n<pre><code>roses_alba -&gt; rose-alba-1.jpg\nroses_damascena -&gt; rose-damascena-1.jpg\nroses_damascena -&gt; rose-damascena-2.jpg\n<\/code><\/pre>\n\n<p>I'd need to know more about what your requirements are though to understand if that meets your needs.<\/p>\n"}]}
{"QuestionId":42468088,"AnswerCount":0,"Tags":"<visual-studio><tensorflow><anaconda>","CreationDate":"2017-02-26T11:49:13.903","AcceptedAnswerId":null,"OwnerUserId":4037200.0,"Title":"cannot import tensorflow (installed with anaconda) in Visual Studio IDE","Body":"<p>Follow <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">these steps<\/a>, I successfully installed Tensorflow 1.0.0 of GPU version on my win10 laptop, using anaconda virtual environment (I know little about this, just follow the tips)<\/p>\n\n<p>I run the HelloWorld code in interactive mode like this:<\/p>\n\n<pre><\/pre>\n\n<p>However, I cannot run it when I write in anaconda's editor or Visual Studio 2015 (community edition). It says <\/p>\n\n<blockquote>\n  <p>no module named tensorflow<\/p>\n<\/blockquote>\n\n<p>My config in VS is:\n<a href=\"https:\/\/i.stack.imgur.com\/tBFrt.png\" rel=\"nofollow noreferrer\">pic<\/a><\/p>\n\n<p>So, can anyone give some tips on how to code tensorflow program happily in VS? Thx!<\/p>\n","answers":[]}
{"QuestionId":42468292,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-26T12:09:58.070","AcceptedAnswerId":42474169.0,"OwnerUserId":2885553.0,"Title":"GradientDescentOptimizer got wrong result","Body":"<p><strong>I want use gradient descent to solve equation set, but I got wrong result everytime, so I check my code and written a numpy edition, in this edition I provide explicit loss gradient and I can get currect result.<\/strong><\/p>\n\n<p>So I don't understand why GradientDescentOptimizer can not work.<\/p>\n\n<p><strong>here is my code without tf:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>And here is my code with tf:<\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I test these 2 codes with very simple equation set:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Use gradient descent<\/strong>\n<strong>So at 1st compute, the delta x = (0.7, 1.8)<\/strong><\/p>\n\n<p>But unfortunately my code with tf give the <\/p>\n\n<pre><\/pre>\n\n<p>And my code without tf give the<\/p>\n\n<pre><\/pre>\n\n<p><strong>Absolutely code without tf is right, but why tf comput gradient may less 0.05 then currect result?<\/strong>\n<strong>I think this is the reason my code without tf can solve the equation set, but tf edition can not solve equation set currently.<\/strong><\/p>\n\n<p><strong>Can someone tell me why tf give a incurrent gradiant? Thanks<\/strong><\/p>\n\n<p>My platform is Win10 + tensorflow-gpu v1.0<\/p>\n","answers":[{"AnswerId":"42474169","CreationDate":"2017-02-26T20:56:12.130","ParentId":null,"OwnerUserId":"6791223","Title":null,"Body":"<p>You forgot to reshape <code>_b<\/code> in your tensorflow implementation. So you are subtracting a row from a column at this line:         <code>loss = tf.reduce_mean(tf.square(tf.matmul(_a, _x) - _b))<\/code>.<\/p>\n\n<p><strong>EDIT<\/strong>: do not use reduce operations (such a mean or sum) without specifying a reduction axis. By default,  reduction operations in numpy and tensorflow reduce along all dimensions, so you keep getting a single number regardless dimensions of the input array. That could lead to many obscure bugs like that one.<\/p>\n"}]}
{"QuestionId":42468502,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-26T12:29:49.363","AcceptedAnswerId":null,"OwnerUserId":7624998.0,"Title":"tf.TFRecordReader returns multiples copies of the same data with only 1 epoch","Body":"<p>I am trying to evaluate a model as fast as possible. I get my examples from an only TFRecords file and it seems to be extremely slow, so I searched for any explanation here and I found a sample code from Yaroslav Bulatov (<a href=\"https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/ericyue-slowreader\/benchmark.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yaroslavvb\/stuff\/blob\/master\/ericyue-slowreader\/benchmark.py<\/a>).<\/p>\n\n<p>I have replaced the tf.train.shuffle_batch call by tf.train.batch because I need to read only 1 epoch and I don't mind if the samples are shuffled or not. When enqueue_many=False, the result is right, however, when I tried enqueue_many=True with 2 enqueued items, I got the same sample duplicated. <\/p>\n\n<p>The key code is here:<\/p>\n\n<pre><\/pre>\n\n<p>And the complete proof of concept is here:<\/p>\n\n<pre><\/pre>\n\n<p>I would like to prevent two calls to a reader from returning the same TFRecord when used in enqueue_many context. The expected behaviour would be sequential vectors [[0,1,2,3...15],[16,17...]...], however I get [[0,1,2,3...15],[0,1,2,3...15],[16,17...],[16,17...]...<\/p>\n\n<p>My output is this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42517621","CreationDate":"2017-02-28T19:55:42.690","ParentId":null,"OwnerUserId":"7624998","Title":null,"Body":"<p>I put this question as an issue in tensorflow's github and luckily @yaroslavvb answered quickly and gave me the solution. Just in case you are stuck in the same place that I was, you must know that the problem is related to the optimization options. It is an known bug of TF 1.0 and it has been solved in the master branch.<\/p>\n\n<p>You can find more information here: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7916\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7916<\/a><\/p>\n"}]}
{"QuestionId":42468588,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-26T12:37:29.900","AcceptedAnswerId":null,"OwnerUserId":5737155.0,"Title":"What is the influence of an unused ops in tensorflow?","Body":"<p>I build a graph in tensorflow like this:<\/p>\n\n<pre><\/pre>\n\n<p>Then send this word embedding to the neural network model. The result seems OK.\nBut I want to add another embedding, for example, pos-tags, I add it like this:<\/p>\n\n<pre><\/pre>\n\n<p>The two embeddings were connected to a new embedding. But It is strange to me that alougth I only set a pos ebedding like this and do not use it the model, in other words, the embedding is only word embedding, add the above codes then result changed a lot. In my understanding, if i don't send  self.pos_embedding to any actual variables, the ops of self.pos_embedding will not excute. But the results changed a lot. <\/p>\n\n<p>The below is my model build function. The loss function value will be changed even I only add this line: self.test_variable=tf.constant(0). Why this line of code changed the loss function value?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42470319,"AnswerCount":1,"Tags":"<python><tensorflow><output><lstm>","CreationDate":"2017-02-26T15:21:48.673","AcceptedAnswerId":null,"OwnerUserId":6917400.0,"Title":"Output of Tensorflow LSTM-Cell","Body":"<p>(using python)<\/p>\n\n<p>I've got a question on Tensorflow LSTM-Implementation. There are currently several implementations in TF, but I use: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<ul>\n<li>where n_units is the amount of 'parallel' LSTM-Cells. <\/li>\n<\/ul>\n\n<p>Then to get my output I call: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<ul>\n<li>where (as )  is of shape <\/li>\n<li>where  is my batch_size<\/li>\n<li>where  is the amount of timesteps my RNN will go through<\/li>\n<li>where  is the length of one of my input vectors (vector fed into the network on one specific timestep on one specific batch)<\/li>\n<\/ul>\n\n<p>I expect rnn_outputs to be of shape  as I have not specified another output size. \nDocumentation of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn\" rel=\"noreferrer\"><\/a> tells me that output is of shape . \nThe documentation of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\/BasicLSTMCell\" rel=\"noreferrer\"><\/a> does have a property , which is defaulted to n_units (the amount of LSTM-cells I use). <\/p>\n\n<p>So does each LSTM-Cell only output a scalar for every given timestep? I would expect it to output a vector of the length of the input vector. This seems not to be the case from how I understand it right now, so I am confused. Can you tell me whether that's the case or how I could change it to output a vector of size of the input vector per single lstm-cell maybe?<\/p>\n","answers":[{"AnswerId":"46418974","CreationDate":"2017-09-26T06:09:37.627","ParentId":null,"OwnerUserId":"4341842","Title":null,"Body":"<p>I think the primary confusion is on the terminology of the LSTM cell's argument: <code>num_units<\/code>. Unfortunately it doesn't mean, as the name suggests, \"the no. of LSTM cells\" that should be equal to your time-steps. They actually correspond to the number of dimensions in the hidden state (cell state + hidden state vector). \nThe call to <code>dynamic_rnn()<\/code> returns a tensor of shape: <code>[batch_size, time_steps, output_size]<\/code> where, <br><br><\/p>\n\n<p>(Please note this) output_size = num_units; if (num_proj = None) in the lstm cell <br>\nwhere as, output_size = num_proj; if it is defined. <br><br><\/p>\n\n<p>Now, typically, you will extract the last time_step's result and project it to the size of output dimensions using a <code>mat-mul + biases<\/code> operation manually, or use the num_proj argument in the LSTM cell. <br> \nI have been through the same confusion and had to look really deep to get it cleared. Hope this answer clears some of it.<\/p>\n"}]}
{"QuestionId":42470604,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2017-02-26T15:46:38.833","AcceptedAnswerId":42474160.0,"OwnerUserId":4338302.0,"Title":"Keras Callback EarlyStopping comparing training and validation loss","Body":"<p>I'm fitting a neural network in <em>Python<\/em> <em>Keras<\/em>. <\/p>\n\n<p>To avoid overfitting I would like to monitor the training\/validation loss and create a proper callback which stops computations when training loss is too much less than validation loss. <\/p>\n\n<p>An example of a callback is: <\/p>\n\n<pre><\/pre>\n\n<p><strong>Is there any way to stop training when training loss too little compared to validation loss?<\/strong><\/p>\n\n<p>Thank you in advance<\/p>\n","answers":[{"AnswerId":"42474160","CreationDate":"2017-02-26T20:55:18.580","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>You can create a custom callback class for your purpose. <\/p>\n\n<p>I have created one that should correspond to your need :<\/p>\n\n<pre><code>class CustomEarlyStopping(Callback):\n    def __init__(self, ratio=0.0,\n                 patience=0, verbose=0):\n        super(EarlyStopping, self).__init__()\n\n        self.ratio = ratio\n        self.patience = patience\n        self.verbose = verbose\n        self.wait = 0\n        self.stopped_epoch = 0\n        self.monitor_op = np.greater\n\n    def on_train_begin(self, logs=None):\n        self.wait = 0  # Allow instances to be re-used\n\n    def on_epoch_end(self, epoch, logs=None):\n        current_val = logs.get('val_loss')\n        current_train = logs.get('loss')\n        if current_val is None:\n            warnings.warn('Early stopping requires %s available!' %\n                          (self.monitor), RuntimeWarning)\n\n        # If ratio current_loss \/ current_val_loss &gt; self.ratio\n        if self.monitor_op(np.divide(current_train,current_val),self.ratio):\n            self.wait = 0\n        else:\n            if self.wait &gt;= self.patience:\n                self.stopped_epoch = epoch\n                self.model.stop_training = True\n            self.wait += 1\n\n    def on_train_end(self, logs=None):\n        if self.stopped_epoch &gt; 0 and self.verbose &gt; 0:\n            print('Epoch %05d: early stopping' % (self.stopped_epoch))\n<\/code><\/pre>\n\n<p>I took the liberty to interpret that you wanted to stop if the ratio between the <code>train_loss<\/code> and the <code>validation_loss<\/code> goes under a certain <strong>ratio<\/strong> threshold. This ratio argument should be between <code>0.0<\/code> and <code>1.0<\/code>. However, <code>1.0<\/code> is dangerous as the validation loss and the training loss might fluctuate a lot in an erratic way at the beginning of the training.<\/p>\n\n<p>You can add a patience argument which will wait to see if the breaking of your threshold is staying for a certain number of epochs. <\/p>\n\n<p>The way to use this is for exampe :<\/p>\n\n<pre><code>callbacks = [CustomEarlyStopping(ratio=0.5, patience=2, verbose=1), \n            ... Other callbacks ...]\n...\nmodel.fit(..., callbacks=callbacks)\n<\/code><\/pre>\n\n<p>In this case it will stop if the training loss stays lower than <code>0.5*val_loss<\/code> for more than 2 epochs.<\/p>\n\n<p>Does that help you? <\/p>\n"}]}
{"QuestionId":42471183,"AnswerCount":1,"Tags":"<python><c++><tensorflow><gpu>","CreationDate":"2017-02-26T16:39:19.227","AcceptedAnswerId":null,"OwnerUserId":7625888.0,"Title":"specificy gpu devices failed when using tensorflow c++ api","Body":"<p>I trained my tf model in python:<\/p>\n\n<pre><\/pre>\n\n<p>and use the command line to abstract the model:<\/p>\n\n<pre><\/pre>\n\n<p>and during test phase, I set the device as follow:<\/p>\n\n<pre><\/pre>\n\n<p>but someting is wrong:<\/p>\n\n<pre><\/pre>\n\n<p>so,how can I use gpu i correctly??<\/p>\n\n<p>anyone could help??<\/p>\n","answers":[{"AnswerId":"42514386","CreationDate":"2017-02-28T16:49:20.303","ParentId":null,"OwnerUserId":"2084543","Title":null,"Body":"<p>Is it possible you're using a version of TensorFlow without GPU support enabled?  If you're building a binary you may need to add additional BUILD rules from \/\/tensorflow that enable GPU support.  Also ensure you enabled GPU support when running configure.<\/p>\n\n<p><strong>EDIT<\/strong>: Can you file a bug on TF's github issues with:<\/p>\n\n<p>1) your BUILD rule<\/p>\n\n<p>2) much more of your code so we can see how you're building your model and creating your session<\/p>\n\n<p>3) how you ran configure<\/p>\n\n<p>While this API is not yet marked \"public\"; we want to see if there's indeed a bug you are running into so we can fix it.<\/p>\n"}]}
{"QuestionId":42471523,"AnswerCount":1,"Tags":"<python><numpy><tensorflow><mnist>","CreationDate":"2017-02-26T17:04:19.373","AcceptedAnswerId":null,"OwnerUserId":7625987.0,"Title":"How can I generate a proper MNIST image?","Body":"<p>Hey guys so I've been working on a tensorflow project and I want to take a took at the test images from the MNIST database. Below is the gist of my code for converting the original data(ubyte?) into 2d numpy:<\/p>\n\n<pre><\/pre>\n\n<p>However when the img gets shown <a href=\"https:\/\/i.stack.imgur.com\/RrA8c.png\" rel=\"nofollow noreferrer\">here<\/a>, it looks nothing like a normal number so I figure I must have messed up somewhere, but can't pinpoint it other than when the numpy array gets reshaped to [28, 28] from [784]. Any clues?<\/p>\n\n<p>EDIT: So apparently if I use matplotlib instead of PIL it works fine:<\/p>\n","answers":[{"AnswerId":"42471959","CreationDate":"2017-02-26T17:44:57.763","ParentId":null,"OwnerUserId":"4234910","Title":null,"Body":"<p>Multiply the data by 255 and convert to np.uint8 <a href=\"http:\/\/pillow.readthedocs.io\/en\/3.4.x\/handbook\/concepts.html\" rel=\"nofollow noreferrer\">(uint8 for mode 'L')<\/a> have it work.<\/p>\n\n<pre><code>def gen_image(arr):\n    two_d = (np.reshape(arr, (28, 28)) * 255).astype(np.uint8)\n    img = Image.fromarray(two_d, 'L')\n    return img\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/stackoverflow.com\/a\/10967471\/4234910\">This answer helps.<\/a><\/p>\n"}]}
{"QuestionId":42471921,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-02-26T17:42:00.253","AcceptedAnswerId":42472053.0,"OwnerUserId":null,"Title":"executing function in TensorFlow","Body":"<p>I have some questions regarding this <a href=\"https:\/\/github.com\/nlintz\/TensorFlow-Tutorials\/blob\/master\/03_net.py\" rel=\"nofollow noreferrer\">Code<\/a> - neural networks in TensorFlow. <\/p>\n\n<pre><\/pre>\n\n<ul>\n<li><p>After a single run of the loop at line 37, how do i call <strong>model()<\/strong> with <strong>X[0]<\/strong> and the newly learnt <strong>w_h<\/strong> and <strong>w_o<\/strong> , so that I can see the function returns<\/p><\/li>\n<li><p>Similarly, how to print the values of <strong>h<\/strong> inside the <strong>model()<\/strong> function ? <\/p><\/li>\n<\/ul>\n\n<p>Thanks in advance. I am new to tensorFlow :)<\/p>\n","answers":[{"AnswerId":"51321767","CreationDate":"2018-07-13T09:19:09.660","ParentId":null,"OwnerUserId":"2182770","Title":null,"Body":"<p>Regarding the second question:<\/p>\n\n<ul>\n<li>Similarly, how to print the values of h inside the model() function ?<\/li>\n<\/ul>\n\n<p>With the function <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Print\" rel=\"nofollow noreferrer\">tf.Print<\/a><\/p>\n\n<pre><code>def model(X, w_h, w_o):\n    print_h = tf.nn.sigmoid(tf.matmul(X, w_h)) # this is a basic mlp, think 2 stacked logistic regressions\n    h = tf.Print(print_h,[print_h]) # add a node to the execution graph that prints h when ever h is needed and executed in the graph\n    return tf.matmul(h, w_o) # note that we dont take the softmax at the end because our cost fn does that for us\n<\/code><\/pre>\n\n<p>Good explanation can be found here: <a href=\"https:\/\/towardsdatascience.com\/using-tf-print-in-tensorflow-aa26e1cff11e\" rel=\"nofollow noreferrer\">https:\/\/towardsdatascience.com\/using-tf-print-in-tensorflow-aa26e1cff11e<\/a><\/p>\n"},{"AnswerId":"42472053","CreationDate":"2017-02-26T17:53:54.517","ParentId":null,"OwnerUserId":"7394828","Title":null,"Body":"<p>The <code>feed_dict<\/code> translates the placeholders into real value(s). So provide a single entry for <code>feed_dicts<\/code> and evaluate <code>py_x<\/code>.<\/p>\n\n<p>The following should work:<\/p>\n\n<p><strong>For the result (px_y):<\/strong><\/p>\n\n<pre><code>print(sess.run(py_x, feed_dict={X: [yoursample]}))\n<\/code><\/pre>\n\n<p><strong>For <code>h<\/code><\/strong> it's (almost) the same. But as in the linked code <code>h<\/code> is a private member of <code>model()<\/code> you'll need a reference to <code>h<\/code> in order to evaluate it. The easiest way most probably is to replace lines:<\/p>\n\n<pre><code>(14) return tf.matmul(h, w_o)\nwith\n(14) return (tf.matmul(h, w_o), h)\n\n(26) py_x = model(X, w_h, w_o)\nwith\n(26) py_x, h = model(X, w_h, w_o)\n<\/code><\/pre>\n\n<p>and use:<\/p>\n\n<pre><code>print(sess.run(h, feed_dict={X: [yoursample]}))\n<\/code><\/pre>\n\n<p>or alternatively (evaluate multiple variables):<\/p>\n\n<pre><code>py_val, h_val = sess.run([py_x, h], feed_dict={X: [yoursample]})\nprint(py_val, h)\n<\/code><\/pre>\n\n<p><strong>Explanation:<\/strong>\nBy the way we told Tensorflow how our Network is constructed we did not need an explicit reference to the (inner\/hidden) variable <code>h<\/code>. But in order to evaluate it we do need the reference to define <em>what exactly<\/em> to evaluate.<\/p>\n\n<p>There are other ways to get the variables out of the guts of Tensorflow, but as we explicitly create this variable a few lines above I'd avoid dropping something into a black box and ask the very same black box later on to give it back. <\/p>\n"}]}
{"QuestionId":42472274,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-26T18:13:56.457","AcceptedAnswerId":null,"OwnerUserId":5037585.0,"Title":"IndexError: list index out of range on opening pickled file","Body":"<p>I am getting  error while un-pickling a data file which I pickled in another program. I am using Python 3.5.2 and Tensorflow v1.0. Below is the code.<\/p>\n\n<p>Pickling code:<\/p>\n\n<pre><\/pre>\n\n<p>Un-pickling code:<\/p>\n\n<pre><\/pre>\n\n<p>Output is:<\/p>\n\n<pre><\/pre>\n\n<p>Even if I try to print out train_x[0], using below code, I still get the exact same error.<\/p>\n\n<pre><\/pre>\n\n<p>I don't understand this, what is the error and how to correct it?<\/p>\n","answers":[]}
{"QuestionId":42472286,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><neural-network><video-processing>","CreationDate":"2017-02-26T18:14:33.860","AcceptedAnswerId":42494822.0,"OwnerUserId":2953509.0,"Title":"TensorFlow video processing, changes detection","Body":"<p>I'm newbie with machine learning, and I have only basic knowledge in neural networks.<\/p>\n\n<p>I have pretty clear task:<\/p>\n\n<p><strong>1.<\/strong> Video stream shows static picture (white area with yellow squares)\n(in different videos squares located in different places)<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/q412J.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/q412J.png\" alt=\"video stream content example\"><\/a><\/p>\n\n<p><strong>2.<\/strong> In some moment content of the video changes, and starts to show white area without some of the yellow squares.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/7tARN.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/7tARN.png\" alt=\"video content changed\"><\/a><\/p>\n\n<p><strong>3.<\/strong> I need to create mechanism which can determines and somehow indicates that changes.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/mlvBa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mlvBa.png\" alt=\"changes detected\"><\/a><\/p>\n\n<p>I'm going to use for that task TensorFlow framework. Could anybody push me in right direction? Or I'll be very happy to see list of steps to overcome the problem.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42494822","CreationDate":"2017-02-27T20:07:59.860","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>If you know how the static picture looks beforehand, may be some background-subtraction would work? Basically you just subtract the static picture from every frame and check the content of the result. If the resulting picture is empty (zeros or close to it up to some threshold) there is no change to detect. If the resulting picture contains a region that is non-zero (may be above or below a certain manually tuned threshold), you detected a change in that region.<\/p>\n"}]}
{"QuestionId":42472416,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-26T18:27:51.320","AcceptedAnswerId":42474730.0,"OwnerUserId":3849791.0,"Title":"How do I prevent the TypeError: unsupported operand type(s) for -: 'NoneType' and 'int' when using keras EarlyStopping?","Body":"<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>and I'm getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>It doesn't produce that error when I don't use .<\/p>\n\n<p>Anyone have a fix?<\/p>\n","answers":[{"AnswerId":"42474730","CreationDate":"2017-02-26T21:56:44.120","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>If you think about it : you ask to monitor a validation loss without using a validation during the training. <\/p>\n\n<p>Use<\/p>\n\n<pre><code>model.fit(X_train, Y_train, nb_epoch=20, batch_size=20, validation_split=0.2, callbacks=[EarlyStopping(monitor='val_loss', patience=4)])\n<\/code><\/pre>\n\n<p>For example if you want to have a validation. It will use 20% of your data as validation set. You won't be training on those samples, just validating your model at the end of each epoch. <\/p>\n\n<p>And as mentionned in your other question about this code : change the last activation to a softmax to use with <code>categorical_crossentropy<\/code>. Or switch the objective to <code>binary_crossentropy<\/code> depending on your needs. <\/p>\n"}]}
{"QuestionId":42472447,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-26T18:30:31.610","AcceptedAnswerId":42472932.0,"OwnerUserId":3849791.0,"Title":"How do I prevent loss: nan while I'm fitting my keras model?","Body":"<p>Here is my code:<\/p>\n\n<pre><\/pre>\n\n<p>It was running the Epochs and getting better in accuracy but then the loss started being nan and the accuracy went way down. I used  and got an error from that as well.<\/p>\n\n<p>Anyone got a fix?<\/p>\n","answers":[{"AnswerId":"42472932","CreationDate":"2017-02-26T19:11:21.913","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>If you are using <code>categorical_crossentropy<\/code> as loss function then the last layer of the model should be <code>softmax<\/code>. <\/p>\n\n<p>Here you are using <code>sigmoid<\/code> which has the chance of making all dimensions of output close to 0 which will result in loss to overflow and hence <code>nan<\/code>.<\/p>\n"}]}
{"QuestionId":42473034,"AnswerCount":2,"Tags":"<python><tensorflow><tflearn>","CreationDate":"2017-02-26T19:20:06.787","AcceptedAnswerId":null,"OwnerUserId":6688554.0,"Title":"When I used tflearn to predict a test case, it says \"Cannot feed value of shape (32, 32, 3) for Tensor u'input\/X:0', which has shape '(?, 32, 32, 3)\"","Body":"<pre><\/pre>\n\n<p>When i try to predict, It is showing error:\n\"Cannot feed value of shape (32, 32, 3) for Tensor u'input\/X:0', which has shape '(?, 32, 32, 3)\".<\/p>\n\n<p>Please Someone help.<\/p>\n","answers":[{"AnswerId":"42473106","CreationDate":"2017-02-26T19:26:24.880","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>You need to use <code>tf.expand_dims<\/code> on the input for which you want to make the prediction:<\/p>\n\n<pre><code># 't2' is a tensor of shape [32, 32, 3]\nshape(expand_dims(t2, axis=0)) ==&gt; [1, 32, 32, 3]\n<\/code><\/pre>\n"},{"AnswerId":"42480551","CreationDate":"2017-02-27T08:04:09.080","ParentId":null,"OwnerUserId":"6060417","Title":null,"Body":"<p>Here's the longform verison of what Frank said,<\/p>\n\n<p>You are feeding a 3 dimensional tensor when the input accepts a four dimensional tensor. <\/p>\n\n<p>The '?' in the shape corresponds to batch size which you define in the <code>model.fit()<\/code> function. During training this is equal to 100.<\/p>\n\n<p>When you are predicting the network still expects a 4 dimensional tensor similar to training but you are feeding just one single image which is 3 dimensional(shape = [32,32,3]). Since you are trying to predict the class of just one image, which corresponds to a shape [1,32,32,3]. So if you change your code to:<\/p>\n\n<pre><code>print(model.predict(test_x[1].reshape([1,32,32,3])))\n<\/code><\/pre>\n\n<p>This will print the scores for each of the ten classes.<\/p>\n"}]}
{"QuestionId":42473052,"AnswerCount":1,"Tags":"<python><windows><tensorflow><gpu><keras>","CreationDate":"2017-02-26T19:21:24.630","AcceptedAnswerId":42474827.0,"OwnerUserId":448357.0,"Title":"TensorFlow 1.0 does not see GPU on Windows (but Theano does)","Body":"<p>I have a running installation of Keras &amp; Theano on Windows (by following this <a href=\"https:\/\/github.com\/philferriere\/dlwin\" rel=\"noreferrer\">tutorial<\/a>). Now I've tried to switch the backend to <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"noreferrer\">Tensorflow<\/a> which worked quite fine.<\/p>\n\n<p>The only issue I have, is that <a href=\"https:\/\/stackoverflow.com\/questions\/38559755\/how-to-get-current-available-gpus-in-tensorflow\">Tensorflow does not detect my GPU<\/a>, which Theano in contrast does:<\/p>\n\n<pre><\/pre>\n\n<p>yields no results but when running with Theano backend, it works quite nicely:<\/p>\n\n<pre><\/pre>\n\n<p>Apparently there is some configuration missing, but I don't know what. For Theano to run correctly, I needed a file called  with the following content:<\/p>\n\n<pre><\/pre>\n\n<p>Maybe something similar is missing or maybe I need to <a href=\"https:\/\/github.com\/philferriere\/dlwin#switching-between-cpu-and-gpu-mode\" rel=\"noreferrer\">add environment variables like for Theano?<\/a>. Possibly <a href=\"https:\/\/stackoverflow.com\/questions\/33659047\/where-is-the-configure-of-tensorflow-and-how-to-enable-the-gpu-support\">related question<\/a> on Linux (?).<\/p>\n\n<p>The full installation log (which included a strange exception) can be found in this <a href=\"https:\/\/gist.github.com\/apacha\/a595c244f90a27aced56f67f7598d90d\" rel=\"noreferrer\">Gist<\/a>.<\/p>\n\n<p>Any ideas, how to make the GPU visible to Tensorflow?<\/p>\n","answers":[{"AnswerId":"42474827","CreationDate":"2017-02-26T22:08:53.017","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>Installing both tensorflow and tensorflow-gpu on the same machine might cause issues at the moment.<\/p>\n\n<p>Install either tensorflow (for cpu only) or tensorflow-gpu (for gpu only) for version 1.0<\/p>\n"}]}
{"QuestionId":42473745,"AnswerCount":1,"Tags":"<python><windows><tensorflow><gpu>","CreationDate":"2017-02-26T20:18:53.880","AcceptedAnswerId":42495051.0,"OwnerUserId":7626593.0,"Title":"Remove all traces of tensorflow-gpu installation on Win10","Body":"<p>I'm having an issue reverting to tensorflow-cpu from gpu on Windows 10 64 bit, Python 3.5.3. <\/p>\n\n<p>I'm using <a href=\"https:\/\/github.com\/riga\/tfdeploy\" rel=\"nofollow noreferrer\">riga\/tfdeploy<\/a> to distribute trained models, which is not compatible with tf-gpu. I installed tf-gpu for an experiment, then reverted to cpu, all using pip install\/uninstall. Now I'm getting error messages about unknown ops like <strong>RealDiv<\/strong> and <strong>VariableV2<\/strong>. When starting a tensorflow session I now get a bunch of messages that did not appear before installing and uninstalling the GPU version, like:<\/p>\n\n<p><\/p>\n\n<p>My question is: How can I remove any trace of tf-gpu from my system and get back to tf-cpu? I'm using tensorflow-1.0.0-cp35-cp35m-win_amd64.whl (V1.0.0)<\/p>\n\n<p>I read this: <\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/36902457\/how-to-uninstall-tensorflow-completely\">How to uninstall TensorFlow completely?<\/a><\/p>\n\n<p>And several related issues on GitHub, but haven't been able to go back to tf-cpu.<\/p>\n\n<p>Steps to reproduce:<\/p>\n\n<ol>\n<li>On Win 10 64 bit, Python 3.5.3, pip install tensorflow<\/li>\n<li>run a tf.Session() - no messages about ops like BestSplits, RealDiv etc., tfdeploy runs fine.<\/li>\n<li>pip install tensorflow-gpu<\/li>\n<li>pip uninstall tensorflow-gpu<\/li>\n<li>Some trace of these ops remains registered, causing warnings like  while runninng tf, and also causing frameworks like tfdeploy to crash <\/li>\n<\/ol>\n\n<p>I have tried completely reinstalling Python 3, deleting all site-packages etc. I've installed and uninstalled CUDA. The ops seem to be registered somewhere, leading to different behavior of tf-cpu after installing and uninstalling tf-gpu.<\/p>\n\n<p>Any pointers on getting rid of\/unregistering these ops is appreciated!<\/p>\n","answers":[{"AnswerId":"42495051","CreationDate":"2017-02-27T20:22:19.977","ParentId":null,"OwnerUserId":"4847576","Title":null,"Body":"<p>There was a bug on the PYPI packages generating OpKernel errors. The fix for that as of now is to uninstall TensorFlow, download a <a href=\"http:\/\/ci.tensorflow.org\/view\/Nightly\/job\/nightly-win\/97\/\" rel=\"nofollow noreferrer\">nightly build<\/a> and install it, while the PYPI packages are not replaced.<\/p>\n\n<p>Please see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7500#issuecomment-280194967\" rel=\"nofollow noreferrer\">this issue on Github<\/a> for more details. <\/p>\n"}]}
{"QuestionId":42474666,"AnswerCount":2,"Tags":"<python><windows><tensorflow><pip>","CreationDate":"2017-02-26T21:50:33.730","AcceptedAnswerId":42474711.0,"OwnerUserId":3890459.0,"Title":"Tensorflow installation error on Windows 7","Body":"<p>I am just about to install tensorflow to Anaconda environment, but I have some problems. Once I put the following command<\/p>\n\n<pre><\/pre>\n\n<p>I get the error message: tensorflow_gpu_1.0.0-cp35-cp35m-win_x86_64.whl is not a supported wheel on this platform. Then I upgraded pip with this command:<\/p>\n\n<pre><\/pre>\n\n<p>But nothing happened. I have python 3.5.2 and Anaconda 4.2.0 (64 bit).<\/p>\n\n<p>Thanks in advance<\/p>\n","answers":[{"AnswerId":"42474711","CreationDate":"2017-02-26T21:55:04.177","ParentId":null,"OwnerUserId":"2350575","Title":null,"Body":"<p>Install with <code>amd64<\/code> instead of <code>x86_64<\/code>:<\/p>\n\n<p><code>pip install --ignore-installed --upgrade  https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl<\/code><\/p>\n"},{"AnswerId":"42968313","CreationDate":"2017-03-23T06:00:09.717","ParentId":null,"OwnerUserId":"7733102","Title":null,"Body":"<p>I had a similar error on pip install.  Tried the above and lots of files downloaded.\nBut at the end got this error:\nIOError: [Errno 2] No such file or directory: 'c:\\users\\oskars~1.rie\\appdata\\local\\temp\\pip-crhbnr-build\\setup.py<\/p>\n"}]}
{"QuestionId":42475212,"AnswerCount":1,"Tags":"<c++><tensorflow><eigen>","CreationDate":"2017-02-26T22:50:31.367","AcceptedAnswerId":null,"OwnerUserId":2131200.0,"Title":"C++ Eigen: dynamic tensor","Body":"<p>I would like to implement a C++ class that has a vector of tensors as member. The dimensions of the tensors are not predefined but will take values according to some input data. Moreover, the ranks of the tensors can be different. Something like this:<\/p>\n\n<pre><\/pre>\n\n<p>In <a href=\"https:\/\/bitbucket.org\/eigen\/eigen\/src\/677c9f1577810e869f4f09881cabc3e503a810c1\/unsupported\/Eigen\/CXX11\/src\/Tensor\/README.md?at=default&amp;fileviewer=file-view-default#markdown-header-class-tensorref\" rel=\"nofollow noreferrer\">Eigen<\/a>, however, there is no such  type for dynamic tensors.<\/p>\n\n<p>For constructing each tensor, I will read a vector of data  that represents a tensor of dimension  ( times). Something like this:<\/p>\n\n<pre><\/pre>\n\n<p>Is it possible to do that?<\/p>\n\n<p>Thank you very much in advance for your help!<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>As Yaroslav Bulatov pointed out, Eigen does not support dynamic rank and thus the supported ranks have to be written out explicitly. In my code:<\/p>\n\n<pre><\/pre>\n\n<p>Yaroslav also suggested that using macros can help to void code duplication. I'm not familiar with C++ macros, thus any help would be very appreciated.<\/p>\n\n<p>Thanks for your help!<\/p>\n","answers":[{"AnswerId":"46372388","CreationDate":"2017-09-22T20:00:48.957","ParentId":null,"OwnerUserId":"2502544","Title":null,"Body":"<p>You can check out the xtensor C++ template library, which supports both dynamic and static dimensionality.<\/p>\n\n<p><a href=\"http:\/\/xtensor.readthedocs.io\/en\/latest\/\" rel=\"nofollow noreferrer\">http:\/\/xtensor.readthedocs.io\/en\/latest\/<\/a><\/p>\n\n<p>xtensor has an API that is very similar to that of numpy, including vectorization, broadcasting, universal function. There is a numpy to xtensor cheat sheet here: <a href=\"http:\/\/xtensor.readthedocs.io\/en\/latest\/numpy.html\" rel=\"nofollow noreferrer\">http:\/\/xtensor.readthedocs.io\/en\/latest\/numpy.html<\/a><\/p>\n\n<p>Finally, you can try it live in a C++ Jupyter notebook by clicking on the binder badge at the top of <a href=\"https:\/\/github.com\/QuantStack\/xtensor\/\" rel=\"nofollow noreferrer\">https:\/\/github.com\/QuantStack\/xtensor\/<\/a><\/p>\n\n<p>xtensor also comes with bindings for the main languages of scientific computing (R, Julia, Python).<\/p>\n"}]}
{"QuestionId":42475381,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-26T23:08:45.357","AcceptedAnswerId":42476116.0,"OwnerUserId":1647057.0,"Title":"Add dropout layers between pretrained dense layers in keras","Body":"<p>In , there is a VGG16 model pre-trained on imagenet.<\/p>\n\n<pre><\/pre>\n\n<p>This model has the following structure.<\/p>\n\n<pre><\/pre>\n\n<p>I would like to fine-tune this model with dropout layers between the dense layers (fc1, fc2 and predictions), while keeping all the pre-trained weights of the model intact. I know it's possible to access each layer individually with , but I haven't found anywhere how to add new layers between the existing layers.<\/p>\n\n<p>What's the best practice of doing this?<\/p>\n","answers":[{"AnswerId":"42476116","CreationDate":"2017-02-27T00:44:44.343","ParentId":null,"OwnerUserId":"1647057","Title":null,"Body":"<p>I found an answer myself by using <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"noreferrer\">Keras functional API<\/a><\/p>\n\n<pre><code>from keras.applications import VGG16\nfrom keras.layers import Dropout\nfrom keras.models import Model\n\nmodel = VGG16(weights='imagenet')\n\n# Store the fully connected layers\nfc1 = model.layers[-3]\nfc2 = model.layers[-2]\npredictions = model.layers[-1]\n\n# Create the dropout layers\ndropout1 = Dropout(0.85)\ndropout2 = Dropout(0.85)\n\n# Reconnect the layers\nx = dropout1(fc1.output)\nx = fc2(x)\nx = dropout2(x)\npredictors = predictions(x)\n\n# Create a new model\nmodel2 = Model(input=model.input, output=predictors)\n<\/code><\/pre>\n\n<p><code>model2<\/code> have the dropout layers as I wanted<\/p>\n\n<pre><code>____________________________________________________________________________________________________\nLayer (type)                     Output Shape          Param #     Connected to                     \n====================================================================================================\ninput_1 (InputLayer)             (None, 3, 224, 224)   0                                            \n____________________________________________________________________________________________________\nblock1_conv1 (Convolution2D)     (None, 64, 224, 224)  1792        input_1[0][0]                    \n____________________________________________________________________________________________________\nblock1_conv2 (Convolution2D)     (None, 64, 224, 224)  36928       block1_conv1[0][0]               \n____________________________________________________________________________________________________\nblock1_pool (MaxPooling2D)       (None, 64, 112, 112)  0           block1_conv2[0][0]               \n____________________________________________________________________________________________________\nblock2_conv1 (Convolution2D)     (None, 128, 112, 112) 73856       block1_pool[0][0]                \n____________________________________________________________________________________________________\nblock2_conv2 (Convolution2D)     (None, 128, 112, 112) 147584      block2_conv1[0][0]               \n____________________________________________________________________________________________________\nblock2_pool (MaxPooling2D)       (None, 128, 56, 56)   0           block2_conv2[0][0]               \n____________________________________________________________________________________________________\nblock3_conv1 (Convolution2D)     (None, 256, 56, 56)   295168      block2_pool[0][0]                \n____________________________________________________________________________________________________\nblock3_conv2 (Convolution2D)     (None, 256, 56, 56)   590080      block3_conv1[0][0]               \n____________________________________________________________________________________________________\nblock3_conv3 (Convolution2D)     (None, 256, 56, 56)   590080      block3_conv2[0][0]               \n____________________________________________________________________________________________________\nblock3_pool (MaxPooling2D)       (None, 256, 28, 28)   0           block3_conv3[0][0]               \n____________________________________________________________________________________________________\nblock4_conv1 (Convolution2D)     (None, 512, 28, 28)   1180160     block3_pool[0][0]                \n____________________________________________________________________________________________________\nblock4_conv2 (Convolution2D)     (None, 512, 28, 28)   2359808     block4_conv1[0][0]               \n____________________________________________________________________________________________________\nblock4_conv3 (Convolution2D)     (None, 512, 28, 28)   2359808     block4_conv2[0][0]               \n____________________________________________________________________________________________________\nblock4_pool (MaxPooling2D)       (None, 512, 14, 14)   0           block4_conv3[0][0]               \n____________________________________________________________________________________________________\nblock5_conv1 (Convolution2D)     (None, 512, 14, 14)   2359808     block4_pool[0][0]                \n____________________________________________________________________________________________________\nblock5_conv2 (Convolution2D)     (None, 512, 14, 14)   2359808     block5_conv1[0][0]               \n____________________________________________________________________________________________________\nblock5_conv3 (Convolution2D)     (None, 512, 14, 14)   2359808     block5_conv2[0][0]               \n____________________________________________________________________________________________________\nblock5_pool (MaxPooling2D)       (None, 512, 7, 7)     0           block5_conv3[0][0]               \n____________________________________________________________________________________________________\nflatten (Flatten)                (None, 25088)         0           block5_pool[0][0]                \n____________________________________________________________________________________________________\nfc1 (Dense)                      (None, 4096)          102764544   flatten[0][0]                    \n____________________________________________________________________________________________________\ndropout_1 (Dropout)              (None, 4096)          0           fc1[0][0]                        \n____________________________________________________________________________________________________\nfc2 (Dense)                      (None, 4096)          16781312    dropout_1[0][0]                  \n____________________________________________________________________________________________________\ndropout_2 (Dropout)              (None, 4096)          0           fc2[1][0]                        \n____________________________________________________________________________________________________\npredictions (Dense)              (None, 1000)          4097000     dropout_2[0][0]                  \n====================================================================================================\nTotal params: 138,357,544\nTrainable params: 138,357,544\nNon-trainable params: 0\n____________________________________________________________________________________________________\n<\/code><\/pre>\n"}]}
{"QuestionId":42475391,"AnswerCount":1,"Tags":"<python><tensorflow><scheduling>","CreationDate":"2017-02-26T23:09:49.937","AcceptedAnswerId":42480679.0,"OwnerUserId":7623701.0,"Title":"Neural Network model in Tensorflow to solve task scheduling problems","Body":"<p>I am new to Tensorflow and trying to build a Neural Network model in Tensorflow to solve Task Scheduling Problem.<\/p>\n\n<p>I built the model with 2 hidden layers, 36 nodes in the input layer, and 22 nodes in the output layer. All the values in the nodes(in both input and output layer) are normalized floating point numbers(values between 0.0 and 1.0). I followed the example online to build the model as I need to import data from csv file: <a href=\"http:\/\/tneal.org\/post\/tensorflow-iris\/TensorFlowIris\/\" rel=\"nofollow noreferrer\">http:\/\/tneal.org\/post\/tensorflow-iris\/TensorFlowIris\/<\/a><\/p>\n\n<p>I was initially using 9 samples of data to train the network and got overfitting results, so I increased the number of samples to 1000, but the result became weird, and it is not even overfitting anymore(when the same data set are used for both training and testing, the prediction and actual values for output are not the same).<\/p>\n\n<p>When I adjusted the value for the learning rate, the predicted results were changed, and I had even got some negative or very large values. I had also tried to change the optimizer, number of nodes in the hidden layer, the cost function, but still didn't get any improvements.<\/p>\n\n<p>Here is the script I had wrote in python:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the result:\n<a href=\"https:\/\/i.stack.imgur.com\/lPQyw.png\" rel=\"nofollow noreferrer\">Prediction values above and actual values below<\/a><\/p>\n\n<p>Can someone tell me what maybe the cause of the problems in this model?\nThank you.<\/p>\n","answers":[{"AnswerId":"42480679","CreationDate":"2017-02-27T08:12:46.533","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>This is a very broad question, your problem could be related to a lot of things (data, optimization not tuned and\/or did not converge, model design insufficient, bugs in the implementation, etc. ...)<\/p>\n\n<p>A possibility is to take a step back and implement the simplest model you can think of that would still solve the problem. E.g. if you have a regression problem (which I am not sure you do - the <code>'StartingTimeX'<\/code> might be a regression problem, but the <code>'ProcessorForNodeX'<\/code> looks more like a classification problem), you could start off with a simple linear regression model. If this model gives you results that are good enough for your application, there's nothing more to do. Generally in Machine Learning, the simplest model to solve the task at hand is the one you should aim for (<a href=\"https:\/\/en.wikipedia.org\/wiki\/Occam&#39;s_razor\" rel=\"nofollow noreferrer\">Occam's razor<\/a>).<\/p>\n\n<p>If the simple model is not sufficient enough (e.g. because it does not generalize well enough), you can think about how to improve the results, e.g. by using more sophisticated models.<\/p>\n"}]}
{"QuestionId":42475710,"AnswerCount":1,"Tags":"<python><random><tensorflow>","CreationDate":"2017-02-26T23:48:31.327","AcceptedAnswerId":null,"OwnerUserId":1938262.0,"Title":"Tile Random Tensor Without Loosing Randomness in TensorFlow","Body":"<p>In the following example I have a random row tensor  (which for the sake of the example is a simple  but in my real code is a complicated graph instead, whose shape I cannot change).<\/p>\n\n<p>It seems like that in TensorFlow the  operation is run only after the value of  is computed, indeed all the rows of the tiled tensor  are the same. Is there a way to tile a tensor without loosing the randomness of each tile?<\/p>\n\n<pre><\/pre>\n\n<p>Result (all row the same):<\/p>\n\n<pre><\/pre>\n\n<p>Desired result (all rows computed separately):<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42475779","CreationDate":"2017-02-26T23:57:16.757","ParentId":null,"OwnerUserId":"6842947","Title":null,"Body":"<p>You get all rows as same because <code>tf.tile<\/code> does replication of the first argument. In your case, the node named <code>r<\/code> outputs a random tensor. This tensor is fed to <code>tf.tile<\/code> which replicates it (i.e copies the value) three times.<\/p>\n\n<p>If you do not want that you can directly pass a specified shape to <code>tf.random_normal()<\/code> to generate a tensor which is entirely random and matches your shape requirement.<\/p>\n"}]}
{"QuestionId":42475810,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-27T00:02:18.543","AcceptedAnswerId":42475859.0,"OwnerUserId":5102599.0,"Title":"TensorFlow - Unable to get the value of a variable","Body":"<p>I'm new to Tensorflow. I have two tensors\/tensorflow objects, X &amp; Y which basically are two matrices. I perform some computation on them and want to print their value.<\/p>\n\n<pre><\/pre>\n\n<p>I want to get the value of . I'm unable to print it.<\/p>\n\n<p>Tried doing this :<\/p>\n\n<pre><\/pre>\n\n<p>But to no avail ! Throws the following error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42475859","CreationDate":"2017-02-27T00:06:31.850","ParentId":null,"OwnerUserId":"6842947","Title":null,"Body":"<p>The error message speaks for itself. X and Y  are <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"nofollow noreferrer\">placeholders<\/a>. A placeholder is just an object which is expected to contain a value at runtime. In your present case, you need to do something like following :<\/p>\n\n<p><code>v = sess.run(output, feed_dict = {X : matrix_for_X, Y: matrix_for_Y})<\/code><\/p>\n"}]}
{"QuestionId":42476251,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-27T01:03:40.447","AcceptedAnswerId":42476368.0,"OwnerUserId":192936.0,"Title":"TensorFlow FixedLengthRecordReader Placeholder data type error","Body":"<p>I am trying to create a CNN with TensorFlow using the FixedLengthRecordReader.<\/p>\n\n<pre><\/pre>\n\n<p>In the last line I get the following:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas?<\/p>\n","answers":[{"AnswerId":"42476368","CreationDate":"2017-02-27T01:19:34.893","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>You need to pass an image (batch of images) of dimensions <code>[?, IMAGE_SIZE, IMAGE_SIZE, IMAGE_DEPTH]<\/code> and a label (batch of labels) of dimension <code>[?, 1]<\/code> to your <code>sess.run()<\/code> call as a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run\" rel=\"nofollow noreferrer\"><code>feed_dict<\/code><\/a>.<\/p>\n\n<p>Since you have a batch of images named <code>batch_images<\/code> and a batch of labels named <code>batch_labels<\/code> your last lines should be:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>images, labels = sess.run([batch_images, batch_labels])\n_, loss_value = sess.run([train_op, cost], feed_dict={X: images, Y: labels})\n<\/code><\/pre>\n"},{"AnswerId":"42476375","CreationDate":"2017-02-27T01:21:11.337","ParentId":null,"OwnerUserId":"6842947","Title":null,"Body":"<p>You have defined <code>tf.placeholder()<\/code> <code>X<\/code> and <code>Y<\/code> but you are not providing any value to those placeholders when using <code>sess.run()<\/code>. Do the following :<\/p>\n\n<pre><code>for step in range(1000):\n  img, lbl = sess.run([batch_images, batch_labels])\n  _, loss_value = sess.run([train_op, cost],feed_dict={X: batch_images,\\\n                           Y: batch_labels})\n<\/code><\/pre>\n"}]}
{"QuestionId":42476493,"AnswerCount":1,"Tags":"<python><python-2.7><ubuntu><tensorflow>","CreationDate":"2017-02-27T01:36:48.700","AcceptedAnswerId":null,"OwnerUserId":4282481.0,"Title":"pip install tensorflow-gpu installing in python 3.5","Body":"<p>I am trying to install tensorflow for python 2.7 on Ubuntu 16. I am using pip install tensorflow-gpu and I get the following message in the terminal:<\/p>\n\n<pre><\/pre>\n\n<p>Requirement already satisfied: tensorflow-gpu in \/usr\/local\/lib\/python3.5\/dist-packages\nRequirement already satisfied: wheel>=0.26 in \/usr\/lib\/python3\/dist-packages (from tensorflow-gpu)\nRequirement already satisfied: six>=1.10.0 in \/usr\/lib\/python3\/dist-packages (from tensorflow-gpu)\nRequirement already satisfied: numpy>=1.11.0 in \/usr\/local\/lib\/python3.5\/dist-packages (from tensorflow-gpu)\nRequirement already satisfied: protobuf>=3.1.0 in \/usr\/local\/lib\/python3.5\/dist-packages (from tensorflow-gpu)\nRequirement already satisfied: setuptools in \/usr\/local\/lib\/python3.5\/dist-packages (from protobuf>=3.1.0->tensorflow-gpu)<\/p>\n\n<p>When I ftry to  It says  I am guessing since it is looking in the python 2.7 packages.<\/p>\n\n<p>Shouldn't it be looking in python 2.7 dist-packges?<\/p>\n","answers":[{"AnswerId":"47950774","CreationDate":"2017-12-23T07:48:34.887","ParentId":null,"OwnerUserId":"5361479","Title":null,"Body":"<p>I would suggest you to use anaconda and run the following command<\/p>\n\n<p><code>conda install -c anaconda tensorflow-gpu<\/code><\/p>\n\n<ul>\n<li>Anaconda will make your life easier... hope this helps\nThis will also install the <code>cuda toolkit<\/code> and <code>cudnn<\/code> for you and you are good to go<\/li>\n<\/ul>\n"}]}
{"QuestionId":42477023,"AnswerCount":0,"Tags":"<deep-learning><caffe>","CreationDate":"2017-02-27T02:46:20.053","AcceptedAnswerId":null,"OwnerUserId":2467772.0,"Title":"Difference between Datum class and Blob class in caffe library","Body":"<p>Caffe has two data class Datum class and Blob.<\/p>\n\n<pre><\/pre>\n\n<p>The reason why I query is <\/p>\n\n<p>I am structuring two nets using caffe libraries.\nOne net will go through the input image and extract features.<\/p>\n\n<p>Then the second net will take the feature as input and do forward for the detection.<\/p>\n\n<p>What I am confusing is whether I should keep the feature extracted using the first net in Blob format or Datum to input into the second net. Because feature_extraction example in the caffe library changed Blob data to Datum to save the data. Please look at the following code.<\/p>\n\n<pre><\/pre>\n\n<p>Inside the code, num_mini_batches is 10, batch_size is 50 and dim_features is 4096. dim_features is clear it is output, so it is depth of the output volume.<\/p>\n\n<pre><\/pre>\n\n<p>Could somebody explain me for these two queries?<\/p>\n","answers":[]}
{"QuestionId":42477182,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-02-27T03:06:21.213","AcceptedAnswerId":null,"OwnerUserId":7035612.0,"Title":"caffe | How to compute an element-wise weighted sum of multiple input blobs?","Body":"<p>I have multiple convolution layers which have the same output shape.<br>\nHow can I compute an element-wise weighted sum of these layers' outputs in caffe?<\/p>\n","answers":[{"AnswerId":"42479197","CreationDate":"2017-02-27T06:32:25.537","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>For element-wise operations use <a href=\"http:\/\/caffe.help\/manual\/layers\/eltwise.html\" rel=\"nofollow noreferrer\"><code>\"Eltwise\"<\/code><\/a> layer. To achieve a \"weighted sum\" you can use the <code>coeff<\/code> parameter:<\/p>\n\n<pre><code>layer {\n  name: \"wsum\"\n  type: \"Eltwise\"\n  bottom: \"x1\"\n  bottom: \"x2\"\n  bottom: \"x3\"\n  top: \"y\"\n  eltwise_param {\n    operation: SUM\n    coeff: 0.5 # weight for x1 \n    coeff: 0.3 # weight for x2\n    coeff: 0.2 # weight for x3\n  }\n}\n<\/code><\/pre>\n\n<p>Note that <code>'coeff'<\/code> parameter is only applicapble for <code>SUM<\/code> operation.<\/p>\n"}]}
{"QuestionId":42477652,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-02-27T04:05:55.587","AcceptedAnswerId":null,"OwnerUserId":3849791.0,"Title":"How do I make model.fit(...) print one epoch at a time with keras?","Body":"<p>My epochs come out one by one like so-<\/p>\n\n<pre><\/pre>\n\n<p>It holds the previous epoch on the screen. Is there a way that I can just see them pop up and delete the previous epoch?<\/p>\n","answers":[{"AnswerId":"42478426","CreationDate":"2017-02-27T05:33:24.880","ParentId":null,"OwnerUserId":"6769011","Title":null,"Body":"<p>BaseLogger and ProgbarLogger callbacks kick in by default and therefore you see such output.\nThese callback are automatically applied to every Keras model. The History object gets returned by the fit method of models. You need to catch hold of that object and then configure the behavior (pop-up and delete) you want.<\/p>\n\n<pre><code>hist = model.fit(X_train, Y_train, batch_size=batch_size, nb_epoch=nb_epoch,\n           verbose=2, validation_data=(X_test, Y_test))\n\n# Then explore the hist object\nhist.history #gives you a dictionary\nhist.epoch #gives you a list\n<\/code><\/pre>\n"}]}
{"QuestionId":42478081,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><anaconda><upgrade>","CreationDate":"2017-02-27T04:57:03.720","AcceptedAnswerId":null,"OwnerUserId":1130295.0,"Title":"upgrading to gpu in tensorflow with anaconda gives error message","Body":"<p>I used the technique described in the answer to How to install Tensorflow on Windows 10 with anaconda?  That worked for me.  I then tried to run an upgrade to GPU with the following msg<\/p>\n\n<p>pip install --ignore-installed --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl<\/a><\/p>\n\n<p>tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl is not a supported wheel on this platform.<\/p>\n","answers":[{"AnswerId":"42479147","CreationDate":"2017-02-27T06:29:00.987","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There seems to be an error in the Windows installation instructions, which refers to a non-existent URL. You can enter one of the following commands instead:<\/p>\n\n<pre><code>pip install --ignore-installed --upgrade tensorflow-gpu\n\npip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n"}]}
{"QuestionId":42478339,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-27T05:22:21.210","AcceptedAnswerId":42480364.0,"OwnerUserId":4333347.0,"Title":"How to use a trained model on different inputs","Body":"<p>I implemented a relatively straightforward logistic regression function. I save all the necessary variables such as weights, bias, x, y, etc. and then I run the training algorithm...<\/p>\n\n<pre><\/pre>\n\n<p>The model is saved and the  and  of the trained model is displayed...<\/p>\n\n<pre><\/pre>\n\n<p>This is all fine and dandy. However, now I want to be able to predict any given image using the trained model. For example, I want to feed it picture of say  and see what it predicts it to be.<\/p>\n\n<p>I have another module that restores the model. First we load the variables...<\/p>\n\n<pre><\/pre>\n\n<p>This is good. Now I want to compare one image to the model and get a prediction. In this example, I take the first image from the test dataset  and I attempt to compare it to the model.<\/p>\n\n<pre><\/pre>\n\n<p>I know this will not work. I get the error...<\/p>\n\n<blockquote>\n  <p>ValueError: Cannot feed value of shape (784,) for Tensor 'Placeholder:0', which has shape '(?, 784)'<\/p>\n<\/blockquote>\n\n<p>I am at a loss for ideas. This question is rather long, if a straightforward answer is not possible, some guidance as to the steps I may take to do this is appreciated.<\/p>\n","answers":[{"AnswerId":"42480364","CreationDate":"2017-02-27T07:51:34.663","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>Your input placeholder must be of size <code>(?, 784)<\/code>, the question mark meaning variable size which is probably the batch size. You are feeding an input of size <code>(784,)<\/code> which does not work as the error message states.<\/p>\n\n<p>In your case, during prediction time, the batch size is just 1, so the following should work:<\/p>\n\n<pre><code>import numpy as np\n...\nx_in = np.expand_dims(mnist.test.images[0], axis=0)\nclassification = sess.run(tf.argmax(pred, 1), feed_dict={x:x_in})\n<\/code><\/pre>\n\n<p>Assuming that the input image is available as a numpy array. If it is already a tensor, the corresponding function is <code>tf.expand_dims(..)<\/code>.<\/p>\n"}]}
{"QuestionId":42478954,"AnswerCount":0,"Tags":"<python-2.7><tensorflow><ubuntu-16.04>","CreationDate":"2017-02-27T06:13:24.030","AcceptedAnswerId":null,"OwnerUserId":7628023.0,"Title":"Extra evaluation time when using GPU compared to CPU","Body":"<p>Since I am still very new to Tensorflow, I don't know what caused this problem. I am currently using Tensorflow to do text classification. I needed the timing to evaluate 1 data since I need precision. When I time the evaluation time in CPU, the time required is constant while there seems to be additional time when using GPU. <\/p>\n\n<p>Here is the statistics:<\/p>\n\n<ul>\n<li><p>Using GPU:<\/p>\n\n<pre><\/pre><\/li>\n<li><p>Using CPU :<\/p>\n\n<pre><\/pre><\/li>\n<\/ul>\n","answers":[]}
{"QuestionId":42479024,"AnswerCount":2,"Tags":"<tensorflow><autodiff>","CreationDate":"2017-02-27T06:19:07.477","AcceptedAnswerId":null,"OwnerUserId":7129836.0,"Title":"how tensorflow handles complex gradient?","Body":"<p>Let <strong>z<\/strong> is a complex variable, <strong>C(z)<\/strong> is its conjugation.\nIn complex analysis theory, the derivative of <strong>C(z)<\/strong> w.r.t <strong>z<\/strong> don't exist. But in tesnsorflow, we can calculate <strong>dC(z)\/dz<\/strong> and the result is just <strong>1<\/strong>.\nHere is an example:<\/p>\n\n<pre><\/pre>\n\n<p>The input <strong>X<\/strong> is<\/p>\n\n<pre><\/pre>\n\n<p>and the result <strong>Z<\/strong> is<\/p>\n\n<pre><\/pre>\n\n<p>I don't understand why the gradient is set to be <strong>1<\/strong>?\nAnd I want to know <strong>how tensorflow handles the complex gradients in general<\/strong>.<\/p>\n","answers":[{"AnswerId":"57085573","CreationDate":"2019-07-18T00:27:34.297","ParentId":null,"OwnerUserId":"11800006","Title":null,"Body":"<p>Bit late, but I came across this issue recently too.<\/p>\n\n<p>The key point is that TensorFlow <em>defines<\/em> the \"gradient\" of a complex-valued function f(z) of a complex variable as \"the gradient of the real map F: (x,y) -> Re(f(x+iy)), expressed as a complex number\" (the gradient of that real map is a vector in R^2, so we can express it as a complex number in the obvious way).<\/p>\n\n<p>Presumably the reason for that definition is that in TF one is usually concerned with gradients for the purpose of running gradient descent on a loss function, and in particular for identifying the direction of maximum increase\/decrease of that loss function. Using the above definition of gradient means that a complex-valued function of complex variables can be used as a loss function in a standard gradient descent algorithm, and the result will be that the real part of the function gets minimised (which seems to me a somewhat reasonable interpretation of \"optimise this complex-valued function\").<\/p>\n\n<p>Now, to your question, an equivalent way to write that definition of gradient is<\/p>\n\n<p>gradient(f) := dF\/dx + idF\/dy = conj(df\/dz + dconj(f)\/dz)<\/p>\n\n<p>(you can easily verify that using the definition of d\/dz). That's how TensorFlow handles complex gradients. As for the case of f(z):=conj(z), we have df\/dz=0 (as you mention) and dconj(f)\/dz=1, giving gradient(f)=1.<\/p>\n\n<p>I wrote up a longer explanation here, if you're interested: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3348#issuecomment-512101921\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3348#issuecomment-512101921<\/a><\/p>\n"},{"AnswerId":"58099604","CreationDate":"2019-09-25T13:31:01.530","ParentId":null,"OwnerUserId":"5931672","Title":null,"Body":"<h2><strong>How?<\/strong><\/h2>\n\n<p>The equation used by Tensorflow for the gradient is:<\/p>\n\n<p><img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=%5Cnabla_z%20f%20%3D%20%5Cleft(%20%5Cfrac%7B%5Cpartial%20f%7D%7B%5Cpartial%20z%7D%20%2B%20%5Cfrac%7B%5Cpartial%20f*%7D%7B%5Cpartial%20z%7D%20%5Cright)*%3D2%5Cfrac%7B%5Cpartial%20Real(f)%7D%7B%5Cpartial%20z*%7D\" alt=\"tf-grad-def\"><\/p>\n\n<p>Where the '*' means conjugate.<\/p>\n\n<p>When using the definition of the partial derivatives wrt z and z* it uses Wirtinger Calculus. Wirtinger calculus enables to calculate the derivative wrt a complex variable for non-holomorphic functions. The Wirtinger definition is:<\/p>\n\n<p><img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=%5Cfrac%7B%5Cpartial%20f%7D%7B%5Cpartial%20z%7D%20%3D%20%5Cfrac%7B1%7D%7B2%7D%5Cleft(%20%5Cfrac%7B%5Cpartial%20f%7D%7B%5Cpartial%20x%7D%20-%20j%20%5Cfrac%7B%5Cpartial%20f%7D%7B%5Cpartial%20y%7D%20%5Cright)\" alt=\"wirtinger\"><\/p>\n\n<h2><strong>Why this definition?<\/strong><\/h2>\n\n<p>When using for example Complex-Valued Neural Networks (CVNN) the gradients will be used over non-holomorphic, real-valued scalar function of one or several complex variables, tensorflow definition of a gradient can then be written as:<\/p>\n\n<p><img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=2%5Cfrac%7B%5Cpartial%20f%7D%7B%5Cpartial%20z*%7D%20%3D%20%5Cleft(%20%5Cfrac%7B%5Cpartial%20f%7D%7B%5Cpartial%20x%7D%20%2B%20j%20%5Cfrac%7B%5Cpartial%20f%7D%7B%5Cpartial%20y%7D%20%5Cright)%20\" alt=\"\"> <\/p>\n\n<p>This definition corresponds with the literature of CVNN like for example chapter 4 section 4.3 of <a href=\"https:\/\/books.google.fr\/books?hl=en&amp;lr=&amp;id=B2_bucoS5VAC&amp;oi=fnd&amp;pg=PT9&amp;dq=Complex-Valued+Neural+Networks:+Advances+and+Applications&amp;ots=eDbcbA-Zq5&amp;sig=AAcoG9_26T9IsnpGNOKoelEdWAM&amp;redir_esc=y#v=onepage&amp;q=Complex-Valued%20Neural%20Networks%3A%20Advances%20and%20Applications&amp;f=false\" rel=\"nofollow noreferrer\">this book<\/a> or <a href=\"https:\/\/link.springer.com\/chapter\/10.1007\/978-3-642-24955-6_66\" rel=\"nofollow noreferrer\">Amin et al.<\/a> (between countless examples).<\/p>\n"}]}
{"QuestionId":42479902,"AnswerCount":6,"Tags":"<python><memory><pytorch><torch><tensor>","CreationDate":"2017-02-27T07:21:10.363","AcceptedAnswerId":42482819.0,"OwnerUserId":5352399.0,"Title":"How does the \"view\" method work in PyTorch?","Body":"<p>I am confused about the method  in the following code snippet.<\/p>\n\n<pre><\/pre>\n\n<p>My confusion is regarding the following line.<\/p>\n\n<pre><\/pre>\n\n<p>What does  function do? I have seen its usage in many places, but I can't understand how it interprets its parameters. <\/p>\n\n<p>What happens if I give negative values as parameters to the  function? For example, what happens if I call, ?<\/p>\n\n<p>Can anyone explain the main principle of  function with some examples?<\/p>\n","answers":[{"AnswerId":"48650355","CreationDate":"2018-02-06T19:15:51.373","ParentId":null,"OwnerUserId":"2701950","Title":null,"Body":"<p>Let's do some examples, from simpler to more difficult.<\/p>\n\n<ol>\n<li><p>The <code>view<\/code> method returns a tensor with the same data as the <code>self<\/code> tensor (which means that the returned tensor has the same number of elements), but with a different shape.  For example:<\/p>\n\n<pre><code>a = torch.arange(1, 17)  # a's shape is (16,)\n\na.view(4, 4) # output below\n  1   2   3   4\n  5   6   7   8\n  9  10  11  12\n 13  14  15  16\n[torch.FloatTensor of size 4x4]\n\na.view(2, 2, 4) # output below\n(0 ,.,.) = \n1   2   3   4\n5   6   7   8\n\n(1 ,.,.) = \n 9  10  11  12\n13  14  15  16\n[torch.FloatTensor of size 2x2x4]\n<\/code><\/pre><\/li>\n<li><p>Assuming that <code>-1<\/code> is not one of the parameters, when you multiply them together, the result must be equal to the number of elements in the tensor.  If you do: <code>a.view(3, 3)<\/code>, it will raise a <code>RuntimeError<\/code> because shape (3 x 3) is invalid for input with 16 elements. In other words: 3 x 3 does not equal 16 but 9.<\/p><\/li>\n<li><p>You can use <code>-1<\/code> as one of the parameters that you pass to the function, but only once. All that happens is that the method will do the math for you on how to fill that dimension.  For example <code>a.view(2, -1, 4)<\/code> is equivalent to <code>a.view(2, 2, 4)<\/code>. [16 \/ (2 x 4) = 2]<\/p><\/li>\n<li><p>Notice that the returned tensor <strong>shares the same data<\/strong>. If you make a change in the \"view\" you are changing the original tensor's data:<\/p>\n\n<pre><code>b = a.view(4, 4)\nb[0, 2] = 2\na[2] == 3.0\nFalse\n<\/code><\/pre><\/li>\n<li><p>Now, for a more complex use case.  The documentation says that each new view dimension must either be a subspace of an original dimension, or only span <strong><em>d, d + 1, ..., d + k<\/em><\/strong> that satisfy the following contiguity-like condition that for all <strong><em>i = 0, ..., k - 1, stride[i] = stride[i + 1] x size[i + 1]<\/em><\/strong>. Otherwise, <code>contiguous()<\/code> needs to be called before the tensor can be viewed.  For example:<\/p>\n\n<pre><code>a = torch.rand(5, 4, 3, 2) # size (5, 4, 3, 2)\na_t = a.permute(0, 2, 3, 1) # size (5, 3, 2, 4)\n\n# The commented line below will raise a RuntimeError, because one dimension\n# spans across two contiguous subspaces\n# a_t.view(-1, 4)\n\n# instead do:\na_t.contiguous().view(-1, 4)\n\n# To see why the first one does not work and the second does,\n# compare a.stride() and a_t.stride()\na.stride() # (24, 6, 2, 1)\na_t.stride() # (24, 2, 1, 6)\n<\/code><\/pre>\n\n<p>Notice that for <code>a_t<\/code>, <strong><em>stride[0] != stride[1] x size[1]<\/em><\/strong> since <strong><em>24 != 2 x 3<\/em><\/strong><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"58897267","CreationDate":"2019-11-17T03:34:34.690","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<h2><a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.view\" rel=\"nofollow noreferrer\"><strong><code>torch.Tensor.view()<\/code><\/strong><\/a><\/h2>\n\n<p>Simply put, <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.view\" rel=\"nofollow noreferrer\"><strong><code>torch.Tensor.view()<\/code><\/strong><\/a> which is inspired by <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.ndarray.reshape.html\" rel=\"nofollow noreferrer\"><strong><code>numpy.ndarray.reshape()<\/code><\/strong><\/a> or <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html#numpy.reshape\" rel=\"nofollow noreferrer\"><strong><code>numpy.reshape()<\/code><\/strong><\/a>, creates a <em>new view<\/em> of the tensor, as long as the new shape is compatible with the shape of the original tensor. <\/p>\n\n<p>Let's understand this in detail using a concrete example.<\/p>\n\n<pre><code>In [43]: t = torch.arange(18) \n\nIn [44]: t \nOut[44]: \ntensor([ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16, 17])\n<\/code><\/pre>\n\n<p>With this tensor <code>t<\/code> of shape <code>(18,)<\/code>, new <em>views<\/em> can <strong>only<\/strong> be created for the following shapes:<\/p>\n\n<p><strong><code>(1, 18)<\/code><\/strong> or equivalently  <strong><code>(1, -1)<\/code><\/strong> or <strong><code>(-1, 18)<\/code><\/strong><br>\n   <strong><code>(2, 9)<\/code><\/strong>  or equivalently  <strong><code>(2, -1)<\/code><\/strong> or <strong><code>(-1, 9)<\/code><\/strong><br>\n   <strong><code>(3, 6)<\/code><\/strong>  or equivalently  <strong><code>(3, -1)<\/code><\/strong> or <strong><code>(-1, 6)<\/code><\/strong><br>\n   <strong><code>(6, 3)<\/code><\/strong>  or equivalently  <strong><code>(6, -1)<\/code><\/strong> or <strong><code>(-1, 3)<\/code><\/strong><br>\n   <strong><code>(9, 2)<\/code><\/strong>  or equivalently  <strong><code>(9, -1)<\/code><\/strong> or <strong><code>(-1, 2)<\/code><\/strong><br>\n   <strong><code>(18, 1)<\/code><\/strong> or equivalently  <strong><code>(18, -1)<\/code><\/strong> or <strong><code>(-1, 1)<\/code><\/strong>  <\/p>\n\n<p>As we can already observe from the above shape tuples, the multiplication of the elements of the shape tuple (e.g. <code>2*9<\/code>, <code>3*6<\/code> etc.) <em>must always<\/em> be equal to the total number of elements in the original tensor (<code>18<\/code> in our example). <\/p>\n\n<p>Another thing to observe is that we used a <code>-1<\/code> in one of the places in each of the shape tuples. By using a <code>-1<\/code>, we are being lazy in doing the computation ourselves and rather delegate the task to PyTorch to do calculation of that value for the shape when it creates the new <em>view<\/em>. One important thing to note is that we can <strong>only<\/strong> use a single <code>-1<\/code> in the shape tuple. The remaining values should be explicitly supplied by us. Else PyTorch will complain by throwing a <code>RuntimeError<\/code>:<\/p>\n\n<blockquote>\n  <p>RuntimeError: only one dimension can be inferred<\/p>\n<\/blockquote>\n\n<p>So, with all of the above mentioned shapes, PyTorch will always return a <em>new view<\/em> of the original tensor <code>t<\/code>. This basically means that it just changes the stride information of the tensor for each of the new views that are requested.<\/p>\n\n<p>Below are some examples illustrating how the strides of the tensors are changed with each new <em>view<\/em>.<\/p>\n\n<pre><code># stride of our original tensor `t`\nIn [53]: t.stride() \nOut[53]: (1,)\n<\/code><\/pre>\n\n<p>Now, we will see the strides for the new <em>views<\/em>:<\/p>\n\n<pre><code># shape (1, 18)\nIn [54]: t1 = t.view(1, -1)\n# stride tensor `t1` with shape (1, 18)\nIn [55]: t1.stride() \nOut[55]: (18, 1)\n\n# shape (2, 9)\nIn [56]: t2 = t.view(2, -1)\n# stride of tensor `t2` with shape (2, 9)\nIn [57]: t2.stride()       \nOut[57]: (9, 1)\n\n# shape (3, 6)\nIn [59]: t3 = t.view(3, -1) \n# stride of tensor `t3` with shape (3, 6)\nIn [60]: t3.stride() \nOut[60]: (6, 1)\n\n# shape (6, 3)\nIn [62]: t4 = t.view(6,-1)\n# stride of tensor `t4` with shape (6, 3)\nIn [63]: t4.stride() \nOut[63]: (3, 1)\n\n# shape (9, 2)\nIn [65]: t5 = t.view(9, -1) \n# stride of tensor `t5` with shape (9, 2)\nIn [66]: t5.stride()\nOut[66]: (2, 1)\n\n# shape (18, 1)\nIn [68]: t6 = t.view(18, -1)\n# stride of tensor `t6` with shape (18, 1)\nIn [69]: t6.stride()\nOut[69]: (1, 1)\n<\/code><\/pre>\n\n<p>So that's the magic of the <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.view\" rel=\"nofollow noreferrer\"><code>view()<\/code><\/a> function. It just changes the strides of the (original) tensor for each of the new <em>views<\/em>, as long as the shape of the new <em>view<\/em> is compatible with the original shape.<\/p>\n\n<p>Another interesting thing one might observe from the strides tuples is that the value of the element in the 0<sup>th<\/sup> position is equal to the value of the element in the 1<sup>st<\/sup> position of the shape tuple.<\/p>\n\n<pre><code>In [74]: t3.shape \nOut[74]: torch.Size([3, 6])\n                        |\nIn [75]: t3.stride()    |\nOut[75]: (6, 1)         |\n          |_____________|\n<\/code><\/pre>\n\n<p>This is because:<\/p>\n\n<pre><code>In [76]: t3 \nOut[76]: \ntensor([[ 0,  1,  2,  3,  4,  5],\n        [ 6,  7,  8,  9, 10, 11],\n        [12, 13, 14, 15, 16, 17]])\n<\/code><\/pre>\n\n<p>the stride <code>(6, 1)<\/code> says that to go from one element to the next element along the 0<sup>th<\/sup> dimension, we have to <em>jump<\/em> or take 6 steps. (i.e. to go from <code>0<\/code> to <code>6<\/code>, one has to take 6 steps.) But to go from one element to the next element in the 1<sup>st<\/sup> dimension, we just need only one step (for e.g. to go from <code>2<\/code> to <code>3<\/code>).<\/p>\n\n<p>Thus, the strides information is at the heart of how the elements are accessed from memory for performing the computation.<\/p>\n\n<hr>\n\n<h2><a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.reshape\" rel=\"nofollow noreferrer\"><strong>torch.reshape()<\/strong><\/a><\/h2>\n\n<p>This function would return a <em>view<\/em> and is exactly the same as using <a href=\"https:\/\/pytorch.org\/docs\/stable\/tensors.html#torch.Tensor.view\" rel=\"nofollow noreferrer\"><code>torch.Tensor.view()<\/code><\/a> as long as the new shape is compatible with the shape of the original tensor. Otherwise, it will return a copy. <\/p>\n\n<p>However, the notes of <a href=\"https:\/\/pytorch.org\/docs\/stable\/torch.html#torch.reshape\" rel=\"nofollow noreferrer\"><code>torch.reshape()<\/code><\/a> warns that: <\/p>\n\n<blockquote>\n  <p>contiguous inputs and inputs with compatible strides can be reshaped without copying, but one should not depend on the copying vs. viewing behavior.<\/p>\n<\/blockquote>\n"},{"AnswerId":"42482819","CreationDate":"2017-02-27T10:04:34.830","ParentId":null,"OwnerUserId":"2704763","Title":null,"Body":"<p>The view function is meant to reshape the tensor. <\/p>\n\n<p>Say you have a tensor<\/p>\n\n<pre><code>import torch\na = torch.range(1, 16)\n<\/code><\/pre>\n\n<p><code>a<\/code> is a tensor that has 16 elements from 1 to 16(included). If you want to reshape this tensor to make it a <code>4 x 4<\/code> tensor then you can use <\/p>\n\n<pre><code>a = a.view(4, 4)\n<\/code><\/pre>\n\n<p>Now <code>a<\/code> will be a <code>4 x 4<\/code> tensor. <em>Note that after the reshape the total number of elements need to remain the same. Reshaping the tensor <code>a<\/code> to a <code>3 x 5<\/code> tensor would not be appropriate.<\/em><\/p>\n\n<h3>What is the meaning of parameter -1?<\/h3>\n\n<p>If there is any situation that you don't know how many rows you want but are sure of the number of columns, then you can specify this with a -1. (<em>Note that you can extend this to tensors with more dimensions. Only one of the axis value can be -1<\/em>). This is a way of telling the library: \"give me a tensor that has these many columns and you compute the appropriate number of rows that is necessary to make this happen\".<\/p>\n\n<p>This can be seen in the neural network code that you have given above. After the line <code>x = self.pool(F.relu(self.conv2(x)))<\/code> in the forward function, you will have a 16 depth feature map. You have to flatten this to give it to the fully connected layer. So you tell pytorch to reshape the tensor you obtained to have specific number of columns and tell it to decide the number of rows by itself.<\/p>\n\n<p>Drawing a similarity between numpy and pytorch, <code>view<\/code> is similar to numpy's <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.reshape.html\" rel=\"noreferrer\">reshape<\/a> function. <\/p>\n"},{"AnswerId":"54687473","CreationDate":"2019-02-14T09:49:56.297","ParentId":null,"OwnerUserId":"11061263","Title":null,"Body":"<p>I figured it out that <code>x.view(-1, 16 * 5 * 5)<\/code> is equivalent to <code>x.flatten(1)<\/code>, where the parameter 1 indicates the flatten process starts from the 1st dimension(not flattening the 'sample' dimension)\nAs you can see, the latter usage is semantically more clear and easier to use, so I prefer <code>flatten()<\/code>.<\/p>\n"},{"AnswerId":"55880999","CreationDate":"2019-04-27T13:22:18.467","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<blockquote>\n  <p>What is the meaning of parameter -1?<\/p>\n<\/blockquote>\n\n<p>You can read <code>-1<\/code> as dynamic number of parameters or \"anything\". Because of that there can be only one parameter <code>-1<\/code> in <code>view()<\/code>.<\/p>\n\n<p>If you ask <code>x.view(-1,1)<\/code> this will output tensor shape <code>[anything, 1]<\/code> depending on the number of elements in <code>x<\/code>. For example:<\/p>\n\n<pre><code>import torch\nx = torch.tensor([1, 2, 3, 4])\nprint(x,x.shape)\nprint(\"...\")\nprint(x.view(-1,1), x.view(-1,1).shape)\nprint(x.view(1,-1), x.view(1,-1).shape)\n<\/code><\/pre>\n\n<p>Will output:<\/p>\n\n<pre><code>tensor([1, 2, 3, 4]) torch.Size([4])\n...\ntensor([[1],\n        [2],\n        [3],\n        [4]]) torch.Size([4, 1])\ntensor([[1, 2, 3, 4]]) torch.Size([1, 4])\n<\/code><\/pre>\n"},{"AnswerId":"55155611","CreationDate":"2019-03-14T05:37:30.493","ParentId":null,"OwnerUserId":"2579263","Title":null,"Body":"<p><code>weights.reshape(a, b)<\/code> will return a new tensor with the same data as weights with size (a, b) as in it copies the data to another part of memory.<\/p>\n\n<p><code>weights.resize_(a, b)<\/code> returns the same tensor with a different shape. However, if the new shape results in fewer elements than the original tensor, some elements will be removed from the tensor (but not from memory). If the new shape results in more elements than the original tensor, new elements will be uninitialized in memory.<\/p>\n\n<p><code>weights.view(a, b)<\/code> will return a new tensor with the same data as weights with size (a, b)<\/p>\n"}]}
{"QuestionId":42479954,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><keras><gmm>","CreationDate":"2017-02-27T07:24:16.953","AcceptedAnswerId":null,"OwnerUserId":5217799.0,"Title":"Implement Gaussian Mixture Model using keras","Body":"<p>I am trying to implement Gaussian Mixture Model using keras with tensorflow backend. Is there any guide or example on how to implement it?<\/p>\n","answers":[{"AnswerId":"42481143","CreationDate":"2017-02-27T08:40:54.947","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Are you sure that it is what you want? you want to integrate a GMM into a neural network? <\/p>\n\n<p>Tensorflow and Keras are libraries to create, train and use neural networks models. The Gaussian Mixture Model is <strong>not<\/strong> a neural network.<\/p>\n"}]}
{"QuestionId":42480069,"AnswerCount":1,"Tags":"<python><matrix><tensorflow><reduce>","CreationDate":"2017-02-27T07:32:55.230","AcceptedAnswerId":42480834.0,"OwnerUserId":6476080.0,"Title":"Reducing two tensors in Tensorflow","Body":"<p>I have two tensors.<\/p>\n\n<ol>\n<li>A tensor <img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=%5Clambda\" alt=\"\\sum_{n=1}^N\"> of shape (1,N)<\/li>\n<li>A tensor <img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=z\" alt=\"\\sum_{n=1}^N\"> of shape (N,T)<\/li>\n<\/ol>\n\n<p>What I want to calculate is the following scalar:<\/p>\n\n<p><img src=\"https:\/\/chart.googleapis.com\/chart?cht=tx&amp;chl=%5Csum_%7Bn=1%7D%5E%7BN%7D%20%5Cfrac%7Bz_%7Bn,t%7D%7D%7B%5Clambda_%7Bn%7D%7D%20%2B%20%5Cfrac%7B1-z_%7Bn,t%7D%7D%7B1-%5Clambda_%7Bn%7D%7D\" alt=\"\\sum_{n=1}^N\"><\/p>\n\n<p> seemed helpful, but I couldn't get my head around combining the two tensors and reduce functions to get what I want. Can someone help me how to write the above equation in ?<\/p>\n","answers":[{"AnswerId":"42480834","CreationDate":"2017-02-27T08:22:58.863","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>Does this work?<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nN = 10\nT = 20\nl = tf.constant(np.random.randn(1, N), dtype=tf.float32)\nz = tf.constant(np.random.randn(N, T), dtype=tf.float32)\n\nwith tf.Session() as sess:\n    # swap axis for broadcasting to work\n    l = tf.transpose(l, [1, 0])\n    z_div_l = tf.divide(z, l)\n    z_div_l_2 = tf.divide(1.0 - z, 1.0 - l)\n    result = tf.reduce_sum(tf.add(z_div_l, z_div_l_2), axis=0)\n    eval_result = sess.run(result)\n    print('{}\\n{}'.format(eval_result.shape, eval_result))\n<\/code><\/pre>\n\n<p>This calculates the above expression for every <code>t<\/code> from 0 to T-1, so it is not a scalar but a vector of size <code>(T,)<\/code>. Your question mentions you want to compute just one scalar, but the sum is only over <code>N<\/code> and not over <code>T<\/code>, so I assumed you just want this expression to be evaluated for every <code>t<\/code>.<\/p>\n"}]}
{"QuestionId":42480111,"AnswerCount":7,"Tags":"<python><pytorch>","CreationDate":"2017-02-27T07:35:42.180","AcceptedAnswerId":42616812.0,"OwnerUserId":5352399.0,"Title":"Model summary in pytorch","Body":"<p>Is there any way, I can print the summary of a model in PyTorch like  method does in Keras as follows?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42616812","CreationDate":"2017-03-06T02:55:05.703","ParentId":null,"OwnerUserId":"4385382","Title":null,"Body":"<p>While you will not get as detailed information about the model as in Keras' model.summary, simply printing the model will give you some idea about the different layers involved and their specifications.<\/p>\n\n<p>For instance:<\/p>\n\n<pre><code>from torchvision import models\nmodel = models.vgg16()\nprint(model)\n<\/code><\/pre>\n\n<p>The output in this case would be something as follows:<\/p>\n\n<pre><code>VGG (\n  (features): Sequential (\n    (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (1): ReLU (inplace)\n    (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (3): ReLU (inplace)\n    (4): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))\n    (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (6): ReLU (inplace)\n    (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (8): ReLU (inplace)\n    (9): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))\n    (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (11): ReLU (inplace)\n    (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (13): ReLU (inplace)\n    (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (15): ReLU (inplace)\n    (16): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))\n    (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (18): ReLU (inplace)\n    (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (20): ReLU (inplace)\n    (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (22): ReLU (inplace)\n    (23): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))\n    (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (25): ReLU (inplace)\n    (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (27): ReLU (inplace)\n    (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (29): ReLU (inplace)\n    (30): MaxPool2d (size=(2, 2), stride=(2, 2), dilation=(1, 1))\n  )\n  (classifier): Sequential (\n    (0): Dropout (p = 0.5)\n    (1): Linear (25088 -&gt; 4096)\n    (2): ReLU (inplace)\n    (3): Dropout (p = 0.5)\n    (4): Linear (4096 -&gt; 4096)\n    (5): ReLU (inplace)\n    (6): Linear (4096 -&gt; 1000)\n  )\n)\n<\/code><\/pre>\n\n<p>Now you could, as mentioned by <a href=\"https:\/\/stackoverflow.com\/users\/2704763\/kashyap\">Kashyap<\/a>, use the <code>state_dict<\/code> method to get the weights of the different layers. But using this listing of the layers would perhaps provide more direction is creating a helper function to get that Keras like model summary! Hope this helps!<\/p>\n"},{"AnswerId":"56762410","CreationDate":"2019-06-25T21:43:16.177","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>In order to use torchsummary type:<\/p>\n\n<pre><code>from torchsummary import summary\n<\/code><\/pre>\n\n<p>Install it first if you don't have it.<\/p>\n\n<pre><code>pip install torchsummary \n<\/code><\/pre>\n\n<p>And then you can try it, but note from some reason it is not working unless I set model to cuda <code>alexnet.cuda<\/code>:<\/p>\n\n<pre><code>from torchsummary import summary\nhelp(summary)\nimport torchvision.models as models\nalexnet = models.alexnet(pretrained=False)\nalexnet.cuda()\nsummary(alexnet, (3, 224, 224))\nprint(alexnet)\n<\/code><\/pre>\n\n<p>The <code>summary<\/code> must take the input size and batch size is set to -1 meaning any batch size we provide.<\/p>\n\n<p>If we set <code>summary(alexnet, (3, 224, 224), 32)<\/code> this means use the <code>bs=32<\/code>.<\/p>\n\n<pre><code>summary(model, input_size, batch_size=-1, device='cuda')\n<\/code><\/pre>\n\n<hr>\n\n<p>Out:<\/p>\n\n<pre><code>Help on function summary in module torchsummary.torchsummary:\n\nsummary(model, input_size, batch_size=-1, device='cuda')\n\n----------------------------------------------------------------\n        Layer (type)               Output Shape         Param #\n================================================================\n            Conv2d-1           [32, 64, 55, 55]          23,296\n              ReLU-2           [32, 64, 55, 55]               0\n         MaxPool2d-3           [32, 64, 27, 27]               0\n            Conv2d-4          [32, 192, 27, 27]         307,392\n              ReLU-5          [32, 192, 27, 27]               0\n         MaxPool2d-6          [32, 192, 13, 13]               0\n            Conv2d-7          [32, 384, 13, 13]         663,936\n              ReLU-8          [32, 384, 13, 13]               0\n            Conv2d-9          [32, 256, 13, 13]         884,992\n             ReLU-10          [32, 256, 13, 13]               0\n           Conv2d-11          [32, 256, 13, 13]         590,080\n             ReLU-12          [32, 256, 13, 13]               0\n        MaxPool2d-13            [32, 256, 6, 6]               0\nAdaptiveAvgPool2d-14            [32, 256, 6, 6]               0\n          Dropout-15                 [32, 9216]               0\n           Linear-16                 [32, 4096]      37,752,832\n             ReLU-17                 [32, 4096]               0\n          Dropout-18                 [32, 4096]               0\n           Linear-19                 [32, 4096]      16,781,312\n             ReLU-20                 [32, 4096]               0\n           Linear-21                 [32, 1000]       4,097,000\n================================================================\nTotal params: 61,100,840\nTrainable params: 61,100,840\nNon-trainable params: 0\n----------------------------------------------------------------\nInput size (MB): 18.38\nForward\/backward pass size (MB): 268.12\nParams size (MB): 233.08\nEstimated Total Size (MB): 519.58\n----------------------------------------------------------------\nAlexNet(\n  (features): Sequential(\n    (0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))\n    (1): ReLU(inplace)\n    (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)\n    (3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))\n    (4): ReLU(inplace)\n    (5): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)\n    (6): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (7): ReLU(inplace)\n    (8): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (9): ReLU(inplace)\n    (10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\n    (11): ReLU(inplace)\n    (12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)\n  )\n  (avgpool): AdaptiveAvgPool2d(output_size=(6, 6))\n  (classifier): Sequential(\n    (0): Dropout(p=0.5)\n    (1): Linear(in_features=9216, out_features=4096, bias=True)\n    (2): ReLU(inplace)\n    (3): Dropout(p=0.5)\n    (4): Linear(in_features=4096, out_features=4096, bias=True)\n    (5): ReLU(inplace)\n    (6): Linear(in_features=4096, out_features=1000, bias=True)\n  )\n)\n<\/code><\/pre>\n"},{"AnswerId":"54581226","CreationDate":"2019-02-07T19:50:57.440","ParentId":null,"OwnerUserId":"2579263","Title":null,"Body":"<p>Simply print the model after defining an object for the model class<\/p>\n\n<pre><code>class RNN(nn.Module):\n    def __init__(self, input_dim, embedding_dim, hidden_dim, output_dim):\n        super().__init__()\n\n        self.embedding = nn.Embedding(input_dim, embedding_dim)\n        self.rnn = nn.RNN(embedding_dim, hidden_dim)\n        self.fc = nn.Linear(hidden_dim, output_dim)\n    def forward():\n        ...\n\nmodel = RNN(input_dim, embedding_dim, hidden_dim, output_dim)\nprint(model)\n<\/code><\/pre>\n"},{"AnswerId":"42612738","CreationDate":"2017-03-05T19:09:50.403","ParentId":null,"OwnerUserId":"1690724","Title":null,"Body":"<p>AFAK there is no model.summary() like equivalent in pytorch<\/p>\n\n<p>Meanwhile you can refer <a href=\"https:\/\/github.com\/szagoruyko\/functional-zoo\/blob\/master\/visualize.py\" rel=\"nofollow noreferrer\">script<\/a> by szagoruyko, which gives a nice visualizaton like in <a href=\"https:\/\/github.com\/szagoruyko\/functional-zoo\/blob\/master\/resnet-18-export.ipynb\" rel=\"nofollow noreferrer\">resnet18-example<\/a><\/p>\n\n<p>Cheers<\/p>\n"},{"AnswerId":"49989438","CreationDate":"2018-04-23T20:26:57.510","ParentId":null,"OwnerUserId":"4335937","Title":null,"Body":"<p>Yes, you can get exact Keras representation, using <a href=\"https:\/\/github.com\/sksq96\/pytorch-summary\" rel=\"noreferrer\">pytorch-summary<\/a> package.<\/p>\n\n<h3>Example for VGG16<\/h3>\n\n<pre><code>from torchvision import models\nfrom torchsummary import summary\n\nvgg = models.vgg16()\nsummary(vgg, (3, 224, 224))\n\n----------------------------------------------------------------\n        Layer (type)               Output Shape         Param #\n================================================================\n            Conv2d-1         [-1, 64, 224, 224]           1,792\n              ReLU-2         [-1, 64, 224, 224]               0\n            Conv2d-3         [-1, 64, 224, 224]          36,928\n              ReLU-4         [-1, 64, 224, 224]               0\n         MaxPool2d-5         [-1, 64, 112, 112]               0\n            Conv2d-6        [-1, 128, 112, 112]          73,856\n              ReLU-7        [-1, 128, 112, 112]               0\n            Conv2d-8        [-1, 128, 112, 112]         147,584\n              ReLU-9        [-1, 128, 112, 112]               0\n        MaxPool2d-10          [-1, 128, 56, 56]               0\n           Conv2d-11          [-1, 256, 56, 56]         295,168\n             ReLU-12          [-1, 256, 56, 56]               0\n           Conv2d-13          [-1, 256, 56, 56]         590,080\n             ReLU-14          [-1, 256, 56, 56]               0\n           Conv2d-15          [-1, 256, 56, 56]         590,080\n             ReLU-16          [-1, 256, 56, 56]               0\n        MaxPool2d-17          [-1, 256, 28, 28]               0\n           Conv2d-18          [-1, 512, 28, 28]       1,180,160\n             ReLU-19          [-1, 512, 28, 28]               0\n           Conv2d-20          [-1, 512, 28, 28]       2,359,808\n             ReLU-21          [-1, 512, 28, 28]               0\n           Conv2d-22          [-1, 512, 28, 28]       2,359,808\n             ReLU-23          [-1, 512, 28, 28]               0\n        MaxPool2d-24          [-1, 512, 14, 14]               0\n           Conv2d-25          [-1, 512, 14, 14]       2,359,808\n             ReLU-26          [-1, 512, 14, 14]               0\n           Conv2d-27          [-1, 512, 14, 14]       2,359,808\n             ReLU-28          [-1, 512, 14, 14]               0\n           Conv2d-29          [-1, 512, 14, 14]       2,359,808\n             ReLU-30          [-1, 512, 14, 14]               0\n        MaxPool2d-31            [-1, 512, 7, 7]               0\n           Linear-32                 [-1, 4096]     102,764,544\n             ReLU-33                 [-1, 4096]               0\n          Dropout-34                 [-1, 4096]               0\n           Linear-35                 [-1, 4096]      16,781,312\n             ReLU-36                 [-1, 4096]               0\n          Dropout-37                 [-1, 4096]               0\n           Linear-38                 [-1, 1000]       4,097,000\n================================================================\nTotal params: 138,357,544\nTrainable params: 138,357,544\nNon-trainable params: 0\n----------------------------------------------------------------\nInput size (MB): 0.57\nForward\/backward pass size (MB): 218.59\nParams size (MB): 527.79\nEstimated Total Size (MB): 746.96\n----------------------------------------------------------------\n<\/code><\/pre>\n"},{"AnswerId":"48659275","CreationDate":"2018-02-07T08:39:25.967","ParentId":null,"OwnerUserId":"2326672","Title":null,"Body":"<p>Simplest to remember (not as pretty as Keras):<\/p>\n\n<pre><code>print(model)\n<\/code><\/pre>\n\n<p>This also work:<\/p>\n\n<pre><code>repr(model)\n<\/code><\/pre>\n\n<p>If you just want the number of parameters:<\/p>\n\n<pre><code>sum([param.nelement() for param in model.parameters()])\n<\/code><\/pre>\n\n<p>From: <a href=\"https:\/\/discuss.pytorch.org\/t\/is-there-similar-pytorch-function-as-model-summary-as-keras\/2678\/3\" rel=\"nofollow noreferrer\">Is there similar pytorch function as model.summary() as keras? (forum.PyTorch.org)<\/a><\/p>\n"},{"AnswerId":"45528544","CreationDate":"2017-08-06T04:33:58.830","ParentId":null,"OwnerUserId":"221742","Title":null,"Body":"<p>This will show a model's weights and parameters (but not output shape).<\/p>\n\n<pre><code>from torch.nn.modules.module import _addindent\nimport torch\nimport numpy as np\ndef torch_summarize(model, show_weights=True, show_parameters=True):\n    \"\"\"Summarizes torch model by showing trainable parameters and weights.\"\"\"\n    tmpstr = model.__class__.__name__ + ' (\\n'\n    for key, module in model._modules.items():\n        # if it contains layers let call it recursively to get params and weights\n        if type(module) in [\n            torch.nn.modules.container.Container,\n            torch.nn.modules.container.Sequential\n        ]:\n            modstr = torch_summarize(module)\n        else:\n            modstr = module.__repr__()\n        modstr = _addindent(modstr, 2)\n\n        params = sum([np.prod(p.size()) for p in module.parameters()])\n        weights = tuple([tuple(p.size()) for p in module.parameters()])\n\n        tmpstr += '  (' + key + '): ' + modstr \n        if show_weights:\n            tmpstr += ', weights={}'.format(weights)\n        if show_parameters:\n            tmpstr +=  ', parameters={}'.format(params)\n        tmpstr += '\\n'   \n\n    tmpstr = tmpstr + ')'\n    return tmpstr\n\n# Test\nimport torchvision.models as models\nmodel = models.alexnet()\nprint(torch_summarize(model))\n\n# # Output\n# AlexNet (\n#   (features): Sequential (\n#     (0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2)), weights=((64, 3, 11, 11), (64,)), parameters=23296\n#     (1): ReLU (inplace), weights=(), parameters=0\n#     (2): MaxPool2d (size=(3, 3), stride=(2, 2), dilation=(1, 1)), weights=(), parameters=0\n#     (3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2)), weights=((192, 64, 5, 5), (192,)), parameters=307392\n#     (4): ReLU (inplace), weights=(), parameters=0\n#     (5): MaxPool2d (size=(3, 3), stride=(2, 2), dilation=(1, 1)), weights=(), parameters=0\n#     (6): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)), weights=((384, 192, 3, 3), (384,)), parameters=663936\n#     (7): ReLU (inplace), weights=(), parameters=0\n#     (8): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)), weights=((256, 384, 3, 3), (256,)), parameters=884992\n#     (9): ReLU (inplace), weights=(), parameters=0\n#     (10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)), weights=((256, 256, 3, 3), (256,)), parameters=590080\n#     (11): ReLU (inplace), weights=(), parameters=0\n#     (12): MaxPool2d (size=(3, 3), stride=(2, 2), dilation=(1, 1)), weights=(), parameters=0\n#   ), weights=((64, 3, 11, 11), (64,), (192, 64, 5, 5), (192,), (384, 192, 3, 3), (384,), (256, 384, 3, 3), (256,), (256, 256, 3, 3), (256,)), parameters=2469696\n#   (classifier): Sequential (\n#     (0): Dropout (p = 0.5), weights=(), parameters=0\n#     (1): Linear (9216 -&gt; 4096), weights=((4096, 9216), (4096,)), parameters=37752832\n#     (2): ReLU (inplace), weights=(), parameters=0\n#     (3): Dropout (p = 0.5), weights=(), parameters=0\n#     (4): Linear (4096 -&gt; 4096), weights=((4096, 4096), (4096,)), parameters=16781312\n#     (5): ReLU (inplace), weights=(), parameters=0\n#     (6): Linear (4096 -&gt; 1000), weights=((1000, 4096), (1000,)), parameters=4097000\n#   ), weights=((4096, 9216), (4096,), (4096, 4096), (4096,), (1000, 4096), (1000,)), parameters=58631144\n# )\n<\/code><\/pre>\n\n<p>Edit: isaykatsman has a pytorch PR to add a <code>model.summary()<\/code> that is exactly like keras <a href=\"https:\/\/github.com\/pytorch\/pytorch\/pull\/3043\/files\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/pytorch\/pull\/3043\/files<\/a><\/p>\n"}]}
{"QuestionId":42480494,"AnswerCount":1,"Tags":"<python><recursion><theano>","CreationDate":"2017-02-27T08:01:00.887","AcceptedAnswerId":null,"OwnerUserId":7628502.0,"Title":"Python\/Theano: Is it possible to construct truly recursive theano functions?","Body":"<p>For example, I can define a recursive Python lambda function for computing the Fibonacci sequence as follows:<\/p>\n\n<pre><\/pre>\n\n<p>However, if I try to turn this into a Theano function, Theano won't accept fn because fn invokes the Boolean operation \">\". So this code crashes:<\/p>\n\n<pre><\/pre>\n\n<p>But if I replace the Boolean operator with theano.tensor.gt(z,1), the code goes into an infinite recursion, so theano.tensor.gt(z,1) isn't serving the role of \">\":<\/p>\n\n<pre><\/pre>\n\n<p>Running this results in \"maximum recursion depth exceeded\". What's wrong? I get the same \"maximum recursion depth exceeded\" error if I replace the definition of fn with<\/p>\n\n<pre><\/pre>\n\n<p>PS I am NOT looking to do this using theano.scan... because I want to learn to do this calculation recursively without resorting to an explicit loop.<\/p>\n\n<p>--Ken<\/p>\n","answers":[{"AnswerId":"51859927","CreationDate":"2018-08-15T13:44:09.190","ParentId":null,"OwnerUserId":"9947010","Title":null,"Body":"<p>In Theano you can make a recursion using <code>outputs_info<\/code> parameter of <code>theano.scan(fn=myfunc(), outputs_info=...)<\/code> and passing previous outputs of <code>myfunc()<\/code> as arguments in the next iteration of <code>myfunc()<\/code>.<\/p>\n\n<p>In case of Fibonacci sequence the code may look like this:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><code>import numpy as np\nimport theano\nimport theano.tensor as T\n\n# placeholder for the number of elements in the Fibonacci sequence\nt_N = T.iscalar('N')\n\n# first two elements for Fibonacci sequence\ninitial = np.array([1,1], dtype=np.int32)\nt_initial = theano.shared(initial)\n\ndef fibonacci_iter(prev1_value, prev2_value):\n    return prev1_value + prev2_value\n\n# Iterate N-2 times over fibonacci() function\n# ('taps': [-2,-1] means take two previous values in the sequence of outputs):\noutputs, updates = theano.scan(\n    fn=fibonacci_iter,\n    outputs_info = [{'initial': t_initial, 'taps': [-2,-1]}],  # a list of dicts or a dict\n    n_steps=t_N-2)\n\n# compiled function:\nfibonacci = theano.function(\n    inputs=[t_N],\n    outputs=outputs)\n\nn = 10\nfibonacci_seq = fibonacci(n)\nprint(np.concatenate([initial, fibonacci_seq]))\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[ 1  1  2  3  5  8 13 21 34 55]\n<\/code><\/pre>\n\n<p>Reference:<br>\n<a href=\"http:\/\/deeplearning.net\/software\/theano\/library\/scan.html#theano.scan\" rel=\"nofollow noreferrer\">http:\/\/deeplearning.net\/software\/theano\/library\/scan.html#theano.scan<\/a><\/p>\n"}]}
{"QuestionId":42481858,"AnswerCount":0,"Tags":"<tensorflow><neural-network>","CreationDate":"2017-02-27T09:18:32.693","AcceptedAnswerId":null,"OwnerUserId":6430408.0,"Title":"using checkpoint to predict, but each time get different result","Body":"<blockquote>\n  <p>I am a newer to tensorflow. I have trained a CNN model , and get\n  checkpoint files.<\/p>\n  \n  <p>Then,I freeze the checkpoint files to \"pb\" proto. Referenced the website:<a href=\"https:\/\/blog.metaflow.fr\/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc#.nxb8t2ylh\" rel=\"nofollow noreferrer\">https:\/\/blog.metaflow.fr\/tensorflow-how-to-freeze-a-model-and-serve-it-with-a-python-api-d4f3596b3adc#.nxb8t2ylh<\/a><\/p>\n  \n  <p>Next,I load graph from pb file,and predict with it .But, the result is different at each predict(for example:the same picture may got 1,1.04,1.2,1.3 these scores)  , why? Did I miss save some variables?<\/p>\n<\/blockquote>\n","answers":[]}
{"QuestionId":42481938,"AnswerCount":4,"Tags":"<python-2.7><tensorflow><ubuntu-14.04><anaconda>","CreationDate":"2017-02-27T09:22:49.280","AcceptedAnswerId":42483264.0,"OwnerUserId":7515188.0,"Title":"AttributeError: 'module' object has no attribute 'merge_all_summaries'","Body":"<p>Ubuntu 14.04. <\/p>\n\n<p>Python 2.7.13 :: Anaconda custom (64-bit)<\/p>\n\n<p>I installed Tensorflow follow the tutorial: <a href=\"https:\/\/www.tensorflow.org\/install\/\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/<\/a><\/p>\n\n<p>when I enter <\/p>\n\n<blockquote>\n  <p>~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/examples\/tutorials\/mnist<\/p>\n<\/blockquote>\n\n<p>and attempt to run the already existed python file:<\/p>\n\n<blockquote>\n  <p>fully_connected_feed.py<\/p>\n<\/blockquote>\n\n<p>I met the below <strong>AttributeError<\/strong>:<\/p>\n\n<pre><\/pre>\n\n<p>But the same code run successfully in other computer. So I think that must be the configuration problem in my computer.<\/p>\n\n<p>I have followed the same steps to install tensorflow many times and use it to run Deep Learning for a period of time. But it was the first time I met such problem.<\/p>\n\n<p>There are many suggestions in Google say such kind of AttributeError maybe the problem with <em>python version<\/em>. But it's not.<\/p>\n","answers":[{"AnswerId":"57559890","CreationDate":"2019-08-19T15:36:31.833","ParentId":null,"OwnerUserId":"2173097","Title":null,"Body":"<p>This worked for me<\/p>\n\n<pre><code>import tensorflow as tf\ntf.merge_all_summaries = tf.compat.v1.summary.merge_all\ntf.train.SummaryWriter = tf.compat.v1.summary.FileWriter\n<\/code><\/pre>\n"},{"AnswerId":"42585827","CreationDate":"2017-03-03T18:14:48.437","ParentId":null,"OwnerUserId":"4533188","Title":null,"Body":"<p>In <a href=\"https:\/\/stackoverflow.com\/a\/40066895\/4533188\">https:\/\/stackoverflow.com\/a\/40066895\/4533188<\/a> you can read up on an answer to a similar problem: The answer is to migrate as appropriate. Check out <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/migration<\/a>. There you see that<\/p>\n\n<pre><code>- tf.merge_summary\n    - should be renamed to tf.summary.merge\n- tf.train.SummaryWriter\n    - should be renamed to tf.summary.FileWriter\n<\/code><\/pre>\n\n<p>(Actually SummaryWriter has also been changed.) So you should be able to solve your problem if you write before your code<\/p>\n\n<pre><code>import tensorflow as tf\ntf.merge_all_summaries = tf.summary.merge_all\ntf.train.SummaryWriter = tf.summary.FileWriter\n<\/code><\/pre>\n\n<p>(I had the same issue in <a href=\"https:\/\/stackoverflow.com\/q\/42585620\/4533188\">Keras + TensorFlow: \u201cmodule &#39;tensorflow&#39; has no attribute &#39;merge_all_summaries&#39;&#39;\u201d<\/a>.)<\/p>\n"},{"AnswerId":"42483264","CreationDate":"2017-02-27T10:24:41.290","ParentId":null,"OwnerUserId":"7515188","Title":null,"Body":"<p>I have solven this problem with the help of teammates.<\/p>\n\n<p>It's the <strong>tensorflow versions<\/strong> problem.<\/p>\n\n<p>Tensorflow version in the former computer is <strong>0.12.1<\/strong>. Tensorflow version in the computer met AttributeError is <strong>1.0.0<\/strong>. This new tensorflow version has change some <strong>Python api<\/strong>, So the AttributeError met.<\/p>\n\n<hr>\n\n<p>In this computer which met AttributeError:<\/p>\n\n<pre><code>:~$ python\nPython 2.7.13 |Anaconda custom (64-bit)| (default, Dec 20 2016, 23:09:15) \n[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\nAnaconda is brought to you by Continuum Analytics.\nPlease check out: http:\/\/continuum.io\/thanks and https:\/\/anaconda.org\n&gt;&gt;&gt; import tensorflow as tf\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcublas.so.8.0 locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcudnn.so.5 locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcufft.so.8.0 locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcuda.so.1 locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcurand.so.8.0 locally\n&gt;&gt;&gt; __tf.version__\nTraceback (most recent call last):\n  File \"&lt;stdin&gt;\", line 1, in &lt;module&gt;\nNameError: name '__tf' is not defined\n&gt;&gt;&gt; tf.__version__\n'1.0.0'\n&gt;&gt;&gt; \n<\/code><\/pre>\n\n<p>In the former computer:<\/p>\n\n<pre><code>:~$ python\nPython 2.7.13 |Anaconda 2.4.1 (64-bit)| (default, Dec 20 2016, 23:09:15) \n[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux2\nType \"help\", \"copyright\", \"credits\" or \"license\" for more information.\nAnaconda is brought to you by Continuum Analytics.\nPlease check out: http:\/\/continuum.io\/thanks and https:\/\/anaconda.org\n&gt;&gt;&gt; import tensorflow as tf\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcublas.so locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcudnn.so locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcufft.so locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcuda.so.1 locally\nI tensorflow\/stream_executor\/dso_loader.cc:128] successfully opened CUDA library libcurand.so locally\n&gt;&gt;&gt; tf.__version__\n'0.12.1'\n&gt;&gt;&gt; \n<\/code><\/pre>\n"},{"AnswerId":"50798695","CreationDate":"2018-06-11T13:19:05.060","ParentId":null,"OwnerUserId":"9162147","Title":null,"Body":"<p>You need to update depreicated  Tensorflow API in the python files you are running i.e. fully_connected_feed.py<\/p>\n\n<p>Look for tf.merge_all_summaries() and replace it with tf.summary.merge_all()<\/p>\n\n<p>Present version of Tensorflow uses <\/p>\n\n<pre><code>tf.summary.merge_all()\n<\/code><\/pre>\n\n<p>instead of the old version which used<\/p>\n\n<pre><code>tf.merge_all_summaries()\n<\/code><\/pre>\n"}]}
{"QuestionId":42482345,"AnswerCount":1,"Tags":"<python><numpy><caffe><pycaffe>","CreationDate":"2017-02-27T09:43:23.960","AcceptedAnswerId":42482465.0,"OwnerUserId":6799037.0,"Title":"How to save img loaded by caffe.io.load_image","Body":"<pre><\/pre>\n\n<p>here is the code.I load the image use caffe.io.load_img and want to save it after processing,but there is an error:<\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>how to save it,please\uff1f<\/p>\n","answers":[{"AnswerId":"42482465","CreationDate":"2017-02-27T09:49:12.417","ParentId":null,"OwnerUserId":"5540305","Title":null,"Body":"<p>You can use <code>PIL<\/code> instead to save the image. I don't think <code>caffe<\/code> has any method exposed for saving the image. <\/p>\n\n<p><strong>Edit<\/strong> - Yeah, <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/python\/caffe\/io.py\" rel=\"nofollow noreferrer\">there's no<\/a> function to save image.<\/p>\n\n<pre><code>from PIL import Image \nimg = Image.fromarray(result.astype('uint8')) # convert image to uint8\nimg.save(path+'.png')\n<\/code><\/pre>\n"}]}
{"QuestionId":42482493,"AnswerCount":0,"Tags":"<installation><tensorflow><runtime-error><ubuntu-14.04>","CreationDate":"2017-02-27T09:50:33.640","AcceptedAnswerId":null,"OwnerUserId":7603807.0,"Title":"can't install tensorflow on Ubuntu 14.04.5 LTS with CUDA_ERROR_UNKNOWN","Body":"<p>I have installed the Tensorflow many times but still met errors. The system always tells me that:<\/p>\n\n<blockquote>\n  <p>failed call to cuInit: CUDA_ERROR_UNKNOWN\" and \"kernel driver does not appear to be running on this host:\/proc\/driver\/nvidia\/version does not exist<\/p>\n<\/blockquote>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/MR9xl.jpg\" alt=\"Here is the a picture of the terminal\"> <\/p>\n","answers":[]}
{"QuestionId":42482500,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-27T09:50:59.220","AcceptedAnswerId":null,"OwnerUserId":5122632.0,"Title":"How to implement new pooling technique in Tensorflow?","Body":"<p>I thought of implementing a new pooling layer, how can I add it to tensorflow as a new layer.\nCouldn't understand the Tensorflow codebase.\nAny leads on how to begin shall be helpful.<\/p>\n","answers":[{"AnswerId":"42483893","CreationDate":"2017-02-27T10:52:51.640","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>If you want to add a new tensorflow operation for pooling you can take a look at: <a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/extend\/adding_an_op<\/a>. <\/p>\n\n<p>You might also want to look at their github and add a pull request if multiple people could use this: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow<\/a>.<\/p>\n"}]}
{"QuestionId":42482537,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-02-27T09:52:42.800","AcceptedAnswerId":null,"OwnerUserId":7261750.0,"Title":"How to get the probability of predict result in classification in tensorflow estimator.py?","Body":"<p>I am using the tensorflow to predict ctr of ads with the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/wide_and_deep\" rel=\"nofollow noreferrer\">google wide&amp;deep model<\/a>.\nAbsolutely, this is a classification problem which to predict the ad will be click(1) or not(0).<\/p>\n\n<p>But the result of 1 or 0 should be computed with a probability such as if p(x=1) >0.2, the result will be 1, and otherwise 0.<\/p>\n\n<p>My question is how to get the probability.<\/p>\n\n<p>I have print much information in estimator.py[both use wide_n_deep model or wide model only], function predict and some functions it called in estimator.py or graph_actions.py. <\/p>\n\n<p>But it doesn't work.(The predict result is also 0 or 1, not the probability of 0 or 1.)<\/p>\n\n<p>In API DOCS, a method named predict_proba will return the probability. But when I use it as , it returns an array with infinite length. How to use this method?<\/p>\n\n<p>My code was modified based on: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/wide_n_deep_tutorial.py\" rel=\"nofollow noreferrer\">wide_n_deep_tutorial.py<\/a><\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"44466477","CreationDate":"2017-06-09T20:49:14.807","ParentId":null,"OwnerUserId":"5361356","Title":null,"Body":"<p>Replace the \n<code>pred_proba = m.predict_proba(x=None, input_fn=lambda: input_fn(df_test))<\/code><\/p>\n\n<p>By\n<code>pred_proba = m.predict_proba(input_fn=lambda: input_fn(df_test))<\/code><\/p>\n"}]}
{"QuestionId":42482724,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><keras><tensorboard>","CreationDate":"2017-02-27T10:00:19.900","AcceptedAnswerId":42488795.0,"OwnerUserId":6254201.0,"Title":"keras using tensorflow as backend :Cannot interpret feed_dict key as Tensor: Can not convert a int into a Tensor","Body":"<p>I am trying to use tensorboard in keras. Following is my code :<\/p>\n\n<pre><\/pre>\n\n<p>BUT error happens as follows:<\/p>\n\n<blockquote>\n  <p>File\n  \"\/home\/jasontian\/enter\/lib\/python3.5\/site-packages\/spyder\/utils\/site\/sitecustomize.py\",\n  line 866, in runfile execfile(filename, namespace)<\/p>\n  \n  <p>File\n  \"\/home\/jasontian\/enter\/lib\/python3.5\/site-packages\/spyder\/utils\/site\/sitecustomize.py\",\n  line 102, in execfile exec(compile(f.read(), filename, 'exec'),\n  namespace)<\/p>\n  \n  <p>File \"\/media\/jasontian\/keras_tf.py\", line 111, in  history =\n  c_model.fit(X_train, y_train,\n  batch_size=128,nb_epoch=3,shuffle=True,verbose=1,validation_split=0.25,callbacks=[tensor_board])<\/p>\n  \n  <p>File\n  \"\/home\/jasontian\/enter\/lib\/python3.5\/site-packages\/keras\/engine\/training.py\",\n  line 1196, in fit initial_epoch=initial_epoch)   File\n  \"\/home\/jasontian\/enter\/lib\/python3.5\/site-packages\/keras\/engine\/training.py\",\n  line 911, in _fit_loop callbacks.on_epoch_end(epoch, epoch_logs)<\/p>\n  \n  <p>File\n  \"\/home\/jasontian\/enter\/lib\/python3.5\/site-packages\/keras\/callbacks.py\",\n  line 76, in on_epoch_end callback.on_epoch_end(epoch, logs)<\/p>\n  \n  <p>File\n  \"\/home\/jasontian\/enter\/lib\/python3.5\/site-packages\/keras\/callbacks.py\",\n  line 653, in on_epoch_end result = self.sess.run([self.merged],\n  feed_dict=feed_dict)<\/p>\n  \n  <p>File\n  \"\/home\/jasontian\/enter\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\",\n  line 766, in run run_metadata_ptr)<\/p>\n  \n  <p>File\n  \"\/home\/jasontian\/enter\/lib\/python3.5\/site-packages\/tensorflow\/python\/client\/session.py\",\n  line 921, in _run\n  + e.args[0]) TypeError: Cannot interpret feed_dict key as Tensor: Can not convert a int into a Tensor.<\/p>\n<\/blockquote>\n\n<p>At first I thought it might be  (it is float64), but I found it works well in an example. \nupdate:the shape of X_train is (39209,32,32,1).\nSo how can I solve this? <\/p>\n","answers":[{"AnswerId":"42488795","CreationDate":"2017-02-27T14:52:51.797","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Doesn't it work without a tf session? If you don't really need the session you can try this :<\/p>\n\n<pre><code>from keras.layers import merge, Dropout, Convolution2D, MaxPooling2D, Input, Dense, Flatten, Merge\nfrom keras.models import Model\nfrom keras.callbacks import EarlyStopping, ReduceLROnPlateau,TensorBoard\nimport pickle\nfrom sklearn.utils import shuffle\nimport numpy as np\nimport random\nfrom keras.optimizers import Adam\n\nnp.random.seed(1000) \n\ndef load_pickled_data(file, columns):\n    with open(file, mode='rb') as f:\n        dataset = pickle.load(f)\n    return tuple(map(lambda c: dataset[c], columns))\n\ntrain_preprocessed_dataset_file = \"train.p\"\ntest_preprocessed_dataset_file = \"test.p\"\n\nX_train, y_train_64 = load_pickled_data(train_preprocessed_dataset_file, columns = ['features', 'labels'])\nX_test, y_test_64 = load_pickled_data(test_preprocessed_dataset_file, columns = ['features', 'labels'])\n\ny_train = y_train_64.astype(np.float32)\ny_test = y_test_64.astype(np.float32)\n\n\n###CNN model###\ninput_img = Input(shape=(32, 32, 1))\n\nconv_1 = Convolution2D(32, 5, 5, border_mode='same',     activation='relu')(input_img)\npool_1 = MaxPooling2D((2, 2))(conv_1)\npool_1 = Dropout(0.1)(pool_1)\n\nconv_2 = Convolution2D(64, 5, 5, border_mode='same', activation='relu')(pool_1)\npool_2 = MaxPooling2D((2, 2))(conv_2)\npool_2 = Dropout(0.2)(pool_2)\n\nconv_3 = Convolution2D(128, 5, 5, border_mode='same', activation='relu')(pool_2)\npool_3 = MaxPooling2D((2, 2))(conv_3)\npool_3 = Dropout(0.3)(pool_3)\npool_3 = Flatten()(pool_3)\n\npool_1 = MaxPooling2D((4, 4))(pool_1)\npool_1 = Flatten()(pool_1)\n\npool_2 = MaxPooling2D((2, 2))(pool_2)\npool_2 =Flatten()(pool_2)\n\nall_features =  merge([pool_1, pool_2, pool_3], mode='concat')\n\nlogits = Dense(500,activation='relu')(all_features)\nlogits = Dropout(0.5)(logits)\nres = Dense(43,activation='softmax')(logits)\n\nc_model = Model(input_img, res)\nc_model.summary()\n\nadam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08)\nc_model.compile(loss='categorical_crossentropy', optimizer= adam, metrics=['accuracy'])\ntensor_board = TensorBoard(log_dir='.\/logs', histogram_freq=1)\n\nhistory = c_model.fit(X_train, y_train, batch_size=128,nb_epoch=3,shuffle=True,verbose=1,validation_split=0.25,callbacks=[tensor_board])\n\nloss_and_metrics = c_model.evaluate(X_test, y_test, batch_size=128)\n<\/code><\/pre>\n\n<p>If your keras default backend is Tensorflow, you don't have to specify it. <\/p>\n"}]}
{"QuestionId":42483644,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-27T10:41:57.583","AcceptedAnswerId":null,"OwnerUserId":6107874.0,"Title":"Error imporing tensorflow in python 3.4 \"Failed to load the native TensorFlow runtime\"","Body":"<p>Traceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/<strong>init<\/strong>.py\", line 61, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 28, in \n    _pywrap_tensorflow = swig_import_helper()\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow', fp, pathname, description)\n  File \"\/usr\/lib\/python3.4\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\nImportError: \/opt\/Xilinx\/Vivado\/2015.1\/lib\/lnx64.o\/libstdc++.so.6: version `GLIBCXX_3.4.19' not found (required by \/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/_pywrap_tensorflow.so)<\/p>\n\n<p>During handling of the above exception, another exception occurred:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"image_resize.py\", line 10, in \n    import tensorflow as tf\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/<strong>init<\/strong>.py\", line 24, in \n    from tensorflow.python import *\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/<strong>init<\/strong>.py\", line 72, in \n    raise ImportError(msg)\nImportError: Traceback (most recent call last):\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/<strong>init<\/strong>.py\", line 61, in \n    from tensorflow.python import pywrap_tensorflow\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 28, in \n    _pywrap_tensorflow = swig_import_helper()\n  File \"\/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/pywrap_tensorflow.py\", line 24, in swig_import_helper\n    _mod = imp.load_module('_pywrap_tensorflow', fp, pathname, description)\n  File \"\/usr\/lib\/python3.4\/imp.py\", line 243, in load_module\n    return load_dynamic(name, filename, file)\nImportError: \/opt\/Xilinx\/Vivado\/2015.1\/lib\/lnx64.o\/libstdc++.so.6: version `GLIBCXX_3.4.19' not found (required by \/usr\/local\/lib\/python3.4\/dist-packages\/tensorflow\/python\/_pywrap_tensorflow.so)<\/p>\n\n<p>Failed to load the native TensorFlow runtime.<\/p>\n\n<p>See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#import_error\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/get_started\/os_setup.md#import_error<\/a><\/p>\n\n<p>for some common reasons and solutions.  Include the entire stack trace\nabove this error message when asking for help.<\/p>\n\n<p>i am using ubuntu 14.4\nand python version 3.4<\/p>\n","answers":[{"AnswerId":"42518468","CreationDate":"2017-02-28T20:48:39.993","ParentId":null,"OwnerUserId":"7128661","Title":null,"Body":"<p>It looks like you installed Xilinx which includes an older version of libstdc++. If you remove it from your <code>LD_LIBRARY_PATH<\/code>, that should solve the issue.<\/p>\n"},{"AnswerId":"49155001","CreationDate":"2018-03-07T15:01:59.170","ParentId":null,"OwnerUserId":"8657905","Title":null,"Body":"<p>I'm using Ubuntu 14.4 with python 3.5. I run the following and the python3 shell imported tensorflow perfectly fine.<\/p>\n\n<pre><code>sudo -H pip3 install tensorflow\n<\/code><\/pre>\n"}]}
{"QuestionId":42485004,"AnswerCount":1,"Tags":"<tensorflow><protocol-buffers>","CreationDate":"2017-02-27T11:46:29.400","AcceptedAnswerId":null,"OwnerUserId":7629474.0,"Title":"Using inception-v3 checkpoint file in tensorflow","Body":"<p>In one of my project, I used a public pre-trained inception-v3 model available here : <a href=\"http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz<\/a>.<\/p>\n\n<p>I only want to use last feature vector (output of pool_3\/_reshape:0). By looking at script example classify_image.py, I can successfully pass an image throught the Deep DNN, extract the bottleneck tensor () and use it for further purpose.<\/p>\n\n<p>I recently saw that there were a more recent trained inception model. Checkpoint of training is available here : <a href=\"http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-v3-2016-03-01.tar.gz\" rel=\"nofollow noreferrer\">http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-v3-2016-03-01.tar.gz<\/a>.<\/p>\n\n<p>I would like to use this new pretrained instead of the old one. However file format is different. The \"old model\" uses a graph def in ProtocolBuffer form (classify_image_graph_def.pb) that is easily reusable. The \"new one\" only provides a checkpoint format, and I'm struggling to insert it into my code.<\/p>\n\n<p>Is there an easy way to convert a checkpoint file to a ProtocolBuffer file that could be then used to create a graph?<\/p>\n","answers":[{"AnswerId":"43734575","CreationDate":"2017-05-02T09:35:47.330","ParentId":null,"OwnerUserId":"3958070","Title":null,"Body":"<p>It seems you have to use <code>freeze_graph.py<\/code>:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/freeze_graph.py<\/a><\/p>\n\n<p>The script converts checkpoint variables into Const ops in a standalone GraphDef file.\nThis script is designed to take a GraphDef proto, a SaverDef proto, and a set of variable values stored in a checkpoint file, and output a GraphDef with all of the variable ops converted into const ops containing the values of the\nvariables.\nIt's useful to do this when we need to load a single file in C++, especially in environments like mobile or embedded where we may not have access to the RestoreTensor ops and file loading calls that they rely on.<\/p>\n\n<p>An example of command-line usage is:<\/p>\n\n<pre><code>bazel build tensorflow\/python\/tools:freeze_graph &amp;&amp; \\\nbazel-bin\/tensorflow\/python\/tools\/freeze_graph \\\n--input_graph=some_graph_def.pb \\\n--input_checkpoint=model.ckpt-8361242 \\\n--output_graph=\/tmp\/frozen_graph.pb --output_node_names=softmax\n<\/code><\/pre>\n"}]}
{"QuestionId":42485717,"AnswerCount":2,"Tags":"<python><numpy><theano>","CreationDate":"2017-02-27T12:20:52.243","AcceptedAnswerId":42485794.0,"OwnerUserId":4730816.0,"Title":"Calculate saturated values after clipping in Theano","Body":"<p>I am using theano.clip to limit values of my numpy array. For e.g.<\/p>\n\n<pre><\/pre>\n\n<p><strong><em>Output:<\/em><\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I want to calculate how many values were clipped after the clipping operation. Is it possible?<\/p>\n","answers":[{"AnswerId":"42485785","CreationDate":"2017-02-27T12:24:59.247","ParentId":null,"OwnerUserId":"2902280","Title":null,"Body":"<p>If your array's name is <code>a<\/code>, you can do<\/p>\n\n<pre><code>np.logical_or(a &gt;= 1, a &lt;= -1).sum()\n<\/code><\/pre>\n\n<p>You won't count elements twice, since <code>- max_val &lt; max_val<\/code>. However, this requires two passes on <code>a<\/code>.<\/p>\n"},{"AnswerId":"42485794","CreationDate":"2017-02-27T12:25:34.053","ParentId":null,"OwnerUserId":"3293881","Title":null,"Body":"<p>Here's one approach with <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.count_nonzero.html\" rel=\"nofollow noreferrer\"><code>np.count_nonzero<\/code><\/a> on a mask of values beyond the limits computed with comparison against the min and max limits -<\/p>\n\n<pre><code>np.count_nonzero((array &lt; -max_val) | (array &gt; max_val))\n<\/code><\/pre>\n\n<p><code>np.count_nonzero<\/code> is meant for performance, as it operates on a mask\/boolean array to get the total count <a href=\"https:\/\/stackoverflow.com\/a\/38687313\/3293881\"><code>pretty efficiently<\/code><\/a>.<\/p>\n\n<p>Alternatively, a shorter version using absolute values as the min and max limits as in this case they are just negative and positive values of the same limiting number -<\/p>\n\n<pre><code>np.count_nonzero(np.abs(array) &gt; max_val)\n<\/code><\/pre>\n\n<p>Sample run -<\/p>\n\n<pre><code>In [267]: array\nOut[267]: \narray([[ 1., -1., -3.,  1.,  1.],\n       [ 3., -4., -5.,  0., -1.],\n       [ 8., -3., -7., -3., -3.],\n       [ 8.,  2., -2., -3., -3.],\n       [ 7.,  0.,  0.,  1.,  0.]])\n\nIn [268]: max_val = np.array([2.0]).astype('float32')\n\nIn [269]: np.count_nonzero((array &lt; -max_val) | (array &gt; max_val))\nOut[269]: 13\n\nIn [270]: np.count_nonzero(np.abs(array) &gt; max_val)\nOut[270]: 13\n<\/code><\/pre>\n"}]}
{"QuestionId":42486004,"AnswerCount":1,"Tags":"<csv><machine-learning><tensorflow><neural-network><keras>","CreationDate":"2017-02-27T12:37:23.340","AcceptedAnswerId":null,"OwnerUserId":6011073.0,"Title":"How to save the encoded output in Keras","Body":"<p>I want to save the encoded part just before decoder stage in autoencoder model of Keras with Tensorflow backend. <\/p>\n\n<p>For instance;<\/p>\n\n<pre><\/pre>\n\n<p>I want to save the  as csv file after the autoencoder training. Suppose the output of  will be .<\/p>\n\n<p>Thank you<\/p>\n","answers":[{"AnswerId":"42486644","CreationDate":"2017-02-27T13:10:13.603","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Try:<\/p>\n\n<pre><code>autoencoder = Model(input_img, decoded)\nencoder = Model(input_img, encoded1)\n\nautoencoder.compile(loss=my_loss, optimizer=my_optimizer)\nautoencoder.fit(my_data, my_data, .. rest of fit params)\n\nnumpy.savetxt(\"encoded1.csv\", encoder.predict(x), delimiter=\",\")\n<\/code><\/pre>\n\n<p>Moreover - I don't know what kind of data do you have but I advise you to use <code>linear<\/code> activation is a last layer and <code>mse<\/code> loss function. <\/p>\n"}]}
{"QuestionId":42487109,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-02-27T13:32:54.830","AcceptedAnswerId":42487230.0,"OwnerUserId":1014196.0,"Title":"Mean subtraction of input","Body":"<p>Is it when I define a binaryproto mean file in a prototxt file, it will subtract the mean in the background (Caffe internally does that) from the input automatically? Or should I write an implementation (that implements the mean subtraction from input) to do that?<\/p>\n","answers":[{"AnswerId":"42487230","CreationDate":"2017-02-27T13:39:31.237","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>When defining <code>mean_file<\/code> in <code>\"Data\"<\/code> layer parameters caffe will subtract the mean values from the inputs. It is already implemented in caffe for training and validation.<\/p>\n\n<p>Note that when you deploy your network you will need to manually subtract the mean.<\/p>\n"}]}
{"QuestionId":42487255,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe><shuffle>","CreationDate":"2017-02-27T13:40:56.190","AcceptedAnswerId":42487588.0,"OwnerUserId":1014196.0,"Title":"Caffe's way of doing data shuffling","Body":"<p>Is shuffling done by setting the flag <em>--shuffle<\/em> as below as found in <em>create_imagenet.sh<\/em> ? :<\/p>\n\n<pre><\/pre>\n\n<p>I mean I don't need to shuffle it manually afterwards, if the flag does it already. What about the label, is it shuffled automatically in the generated <em>lmdb<\/em> file?<\/p>\n","answers":[{"AnswerId":"42487588","CreationDate":"2017-02-27T13:57:57.420","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Using <a href=\"https:\/\/stackoverflow.com\/a\/31431716\/1714410\"><code>convert_imageset<\/code><\/a> tool creates a <strong>copy<\/strong> of your training\/validation data in a binary database file (either in <a href=\"\/questions\/tagged\/lmdb\" class=\"post-tag\" title=\"show questions tagged &#39;lmdb&#39;\" rel=\"tag\">lmdb<\/a> or <a href=\"\/questions\/tagged\/leveldb\" class=\"post-tag\" title=\"show questions tagged &#39;leveldb&#39;\" rel=\"tag\">leveldb<\/a> format). The data encoded in the dataset includes pairs of example and its corresponding label.<br>\nTherefore, when <code>shuffle<\/code>-ing the dataset the labels are shuffled with the data to maintain the correspondence between data and its ground-truth label.<br>\nThere is no need to shuffle the data again during training. <\/p>\n"}]}
{"QuestionId":42487352,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-27T13:45:29.237","AcceptedAnswerId":null,"OwnerUserId":6620662.0,"Title":"Tensor data type as string?","Body":"<p>I have the following Tensorflow code:<\/p>\n\n<pre><\/pre>\n\n<p>The last statement fails with the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I think the problem is related to the Tensor datatype vs. a string datatype. How can I correctly indicate to the os.path.basename function that the image_filename is a string?<\/p>\n","answers":[{"AnswerId":"42487463","CreationDate":"2017-02-27T13:51:56.160","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>The problem is that match_filenames_once returns <\/p>\n\n<blockquote>\n  <p>A variable that is initialized to the list of files matching pattern.<\/p>\n<\/blockquote>\n\n<p>(See here: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/match_filenames_once\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/match_filenames_once<\/a>). <\/p>\n\n<p>os.path.basename and string.split are functions that work on strings, not on tensors. <\/p>\n\n<p>What I recommend you to do is load the images outside of your tensorflow pipeline, this makes labeling easier for you I think. <\/p>\n"}]}
{"QuestionId":42487411,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><recurrent-neural-network><tflearn>","CreationDate":"2017-02-27T13:48:57.010","AcceptedAnswerId":null,"OwnerUserId":5092995.0,"Title":"How to do word classification with Tensorflow tflearn","Body":"<p>I searched on the internet for a decent documentation or examples with some description, but found nothing.\nWhat I'm trying to do, is to classify a word or little phrases (max. 3 words), and give them a priority \"tag\". (For example: Stephen Hawking => 1, Mr. Red => 0)\nI tried to do this with Multilayer perceptron, but there I cannot use words as test data. So I'm trying to do it with RNN, using tflearn library for TensorFlow.\nThe problem there is that I cannot find any documentation about how I should pre-process the data, or how I should give the data to the LSTM layer.<\/p>\n\n<p>I wrote a little bit of code:<\/p>\n\n<pre><\/pre>\n\n<p>The data has the shape => \"name\";0\/1<\/p>\n\n<p>When I execute this code I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>Can someone please help me understand those layer\/algorithm a little bit better?<\/p>\n\n<p>Many thanks in advance<\/p>\n","answers":[{"AnswerId":"42699352","CreationDate":"2017-03-09T15:34:01.777","ParentId":null,"OwnerUserId":"7676016","Title":null,"Body":"<p>From what i've understood, LSTM takes 3D input -> [batch_size, sequence_length, vector_representation]\nIn your case, you can set the sequence length to 3 (the maximum length of a sequence). I usually make a mapping from the word into the 300d GloVe representation. So the input data will look like this:<\/p>\n\n<p><code>net = tflearn.input_data(shape=[None, 3, 300])<\/code><\/p>\n\n<p>I'm not sure with the error, but i think you also need to specify the regression like this :<\/p>\n\n<pre><code>net = tflearn.layers.core.fully_connected(net, 2, activation='softmax')\nnet = tflearn.layers.estimator.regression(net, optimizer='adam', learning_rate=LEARNING_RATE,\n                     loss='categorical_crossentropy', name='target')\n<\/code><\/pre>\n"}]}
{"QuestionId":42487614,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><normalization><caffe>","CreationDate":"2017-02-27T13:59:10.647","AcceptedAnswerId":42489644.0,"OwnerUserId":1014196.0,"Title":"Caffe's input data normalization","Body":"<p>As far as I understood if I want to apply normalization of data [0, 255] to [0, 1), I can pass a parameter of scale: 0.00390625 in the <em>transform_param<\/em> of <em>prototxt<\/em> file. However, what if I would want to perform normalization of data to [-0.5, 0.5] (in order to have 0 mean distribution) at <em>prototxt<\/em> level?<\/p>\n\n<p>Consider my <em>train_val.prototxt<\/em> file looks like below:<\/p>\n\n<pre><\/pre>\n\n<p>Mathematically speaking, I think it looks like below:<\/p>\n\n<pre><\/pre>\n\n<p>but I'm clueless how to map into the scale value of <em>transform_param<\/em> since the scale value has no notion of negative value (signed value). Perhaps, Caffe has different mechanism to achieve this (instead of doing it via scale operation). <\/p>\n\n<p>From my understanding obtained from Udacity's lecture on deep learning, it mentioned that it is a good practise to always normalize the input so as to have a <em>0<\/em> mean. Can I relax this sort of guide? Any significant effect if I would ignore it?<\/p>\n","answers":[{"AnswerId":"42489644","CreationDate":"2017-02-27T15:33:50.427","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Looking at <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/proto\/caffe.proto#L410-L431\" rel=\"nofollow noreferrer\"><code>caffe.proto<\/code><\/a> comments on <code>TransformationParameter<\/code><\/p>\n\n<blockquote>\n<pre><code>\/\/ For data pre-processing, we can do simple scaling and subtracting the\n\/\/ data mean, if provided. Note that the mean subtraction is always carried\n\/\/ out before scaling.\n<\/code><\/pre>\n<\/blockquote>\n\n<p>mean subtraction is done <strong>before<\/strong> scaling, therefore, after subtracting <code>mean_file<\/code>,  your data is approx in range <code>[-128. 128]<\/code>. scale it by 1\/256 and you are done, your data should be roughly in range <code>[-.5, .5]<\/code>.<\/p>\n"}]}
{"QuestionId":42487908,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-27T14:14:07.133","AcceptedAnswerId":42493778.0,"OwnerUserId":1392241.0,"Title":"tensorflow tf.Operation run method usage","Body":"<p>I am exploring tensorflow. \nI have the following problem that I will illustrate in a small code fragment.\nI am not searching for the best way, I am just exploring all options<\/p>\n\n<pre><\/pre>\n\n<p>So my question here is, what is the purpose of tf.Operation.run and how I should be using it. What is wrong with my code? Shouldn't c_operation.run return the value of the resulting operation.<\/p>\n\n<p>Does the associated tensor (c here) holds the values? I did not find a way to extract them (aside from using sess.run(c)<\/p>\n","answers":[{"AnswerId":"47197616","CreationDate":"2017-11-09T08:58:40.410","ParentId":null,"OwnerUserId":"6092557","Title":null,"Body":"<p>As in ash's answer, <code>tf.Operation.run<\/code> or <code>sess.run(op)<\/code> does not return the value.<br>\nYou either use a <code>tf.Tensor<\/code>(like <code>c<\/code>), or get the value in <code>tf.Operation.outputs<\/code>.<br>\nNotice that it's a list of <code>tf.Tensor<\/code>.<br>\nIn other words, <code>c<\/code> is actually equal to <code>c_operation.outputs[0]<\/code><\/p>\n"},{"AnswerId":"42493778","CreationDate":"2017-02-27T19:04:40.523","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>In your sample code, <code>c<\/code> is a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor\" rel=\"nofollow noreferrer\"><code>tf.Tensor<\/code><\/a> while <code>c_operation<\/code> is a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation\" rel=\"nofollow noreferrer\"><code>tf.Operation<\/code><\/a>. A <code>tf.Operation<\/code> represents a computation that produces 0 or more <code>tf.Tensor<\/code>s. <\/p>\n\n<p>Calling <code>run<\/code> on a <code>tf.Operation<\/code> executes all operations in the graph required to produce inputs for this operation, but doesn't return anything (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation#run\" rel=\"nofollow noreferrer\">documentation<\/a>). Calling <code>eval<\/code> on a <code>tf.Tensor<\/code> executes the operation that produces it and returns its value (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor#eval\" rel=\"nofollow noreferrer\">documentation<\/a>). In general, if you're interested in the value, you'd want to call <code>eval<\/code> on the <code>tf.Tensor<\/code>.<\/p>\n\n<p>For example:<\/p>\n\n<pre><code># The following two lines are equivalent\nv = sess.run(c)\nv = c.eval(session=sess)\n\n# The following two lines are equivalent, and neither returns a value\nsess.run(c_operation)\nc_operation.run(session=sess)\n<\/code><\/pre>\n"}]}
{"QuestionId":42488041,"AnswerCount":2,"Tags":"<scope><tensorflow><neural-network><recurrent-neural-network>","CreationDate":"2017-02-27T14:20:30.143","AcceptedAnswerId":42580771.0,"OwnerUserId":4776477.0,"Title":"How to combine FCNN and RNN in Tensorflow?","Body":"<p>I want to make a Neural Network, which would have recurrency (for example, LSTM) at some layers and normal connections (FC) at others.\nI cannot find a way to do it in Tensorflow.\nIt works, if I have only FC layers, but I don't see how to add just one recurrent layer properly.<\/p>\n\n<p>I create a network in a following way : <\/p>\n\n<pre><\/pre>\n\n<p>It should process the batches in a sequential order. I have a function for processing just one time-step, which will be called later, by a function, which process the whole sequence :<\/p>\n\n<pre><\/pre>\n\n<p>The following function should take sequence of batches as input and produce sequence of batches as an output:<\/p>\n\n<pre><\/pre>\n\n<p>The issue is with a variable scopes and their property \"reuse\".<\/p>\n\n<p>If I run this code as it is I am getting the following error:\n' Variable Train\/process_sequence\/basic_lstm_cell\/weights does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope? '<\/p>\n\n<p>If I comment out the line, which tell it to reuse variables ( tf.get_variable_scope().reuse_variables() ) I am getting the following error:\n'Variable Train\/process_sequence\/basic_lstm_cell\/weights already exists, disallowed. Did you mean to set reuse=True in VarScope?'<\/p>\n\n<p>It seems, that we need \"reuse=None\" for the weights of the LSTM cell to be initialized and we need \"reuse=True\" in order to call the LSTM cell. <\/p>\n\n<p>Please, help me to figure out the way to do it properly.<\/p>\n","answers":[{"AnswerId":"42518973","CreationDate":"2017-02-28T21:22:26.623","ParentId":null,"OwnerUserId":"5709722","Title":null,"Body":"<p>I think the problem is that you're creating variables with tf.Variable. Please, use tf.get_variable instead -- does this solve your issue?<\/p>\n"},{"AnswerId":"42580771","CreationDate":"2017-03-03T13:55:20.207","ParentId":null,"OwnerUserId":"4776477","Title":null,"Body":"<p>It seems that I have solved this issue using the hack from the official Tensorflow RNN example (<a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/recurrent<\/a>) with the following code<\/p>\n\n<pre><code>with tf.variable_scope(\"RNN\"):\n  for time_step in range(num_steps):\n    if time_step &gt; 0: tf.get_variable_scope().reuse_variables()\n    (cell_output, state) = cell(inputs[:, time_step, :], state)\n    outputs.append(cell_output)\n<\/code><\/pre>\n\n<p>The hack is that when we run LSTM first time, tf.get_variable_scope().reuse is set to False, so that the new LSTM cell is created. When we run it next time, we set tf.get_variable_scope().reuse to True, so that we are using the LSTM,  which was already created.<\/p>\n"}]}
{"QuestionId":42488070,"AnswerCount":1,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2017-02-27T14:21:42.360","AcceptedAnswerId":42621017.0,"OwnerUserId":7097823.0,"Title":"How to design the label for tensorflow's ctc loss layer","Body":"<p>I just started using ctc loss layer in tensorflow(r1.0) and got a little bit confused with the \"labels\" input<\/p>\n\n<p>In tensorflow's API document, it says<\/p>\n\n<blockquote>\n  <p>labels: An int32 SparseTensor. labels.indices[i, :] == [b, t] means labels.values[i] stores the id for (batch b, time t). labels.values[i] must take on values in [0, num_labels)<\/p>\n<\/blockquote>\n\n<ol>\n<li>Is [b,t] and values[i] mean there is a label \"values[i]\" at \"t\" of sequence \"b\" in the batch?<\/li>\n<li>It says value must be in [0,num_labels), but for a sparse tensor, almost everywhere is 0 excepted for some specified places, so I don't really know how should the sparse tensor for ctc be like<\/li>\n<li>And for example, if I have a short video of hand gesture, and it has a label \"1\",should I label the output of all timesteps as \"1\", or only label the last timestep as \"1\" and take other as \"blank\"?<\/li>\n<\/ol>\n\n<p>thanks!<\/p>\n","answers":[{"AnswerId":"42621017","CreationDate":"2017-03-06T08:40:27.543","ParentId":null,"OwnerUserId":"7652692","Title":null,"Body":"<p>To address your questions:<br>\n1. The notation in the documentation here seems a bit misleading, as the output label index <code>t<\/code> need not be the same as the input time slice, it's simply the index to the output sequence. A different letter could be used because the input and output sequences are not explicitly aligned. Otherwise, your assertion seems correct. I give an example below.<\/p>\n\n<ol start=\"2\">\n<li><p>Zero is a valid class in your sequence output label. The so-called blank label in TensorFlow's CTC implementation is the last (largest) class, which should probably not be in your ground truth labels anyhow. So if you were writing a binary sequence classifier, you'd have three classes, 0 (say \"off\"), 1 (\"on\") and 2 (\"blank\" output of CTC).<\/p><\/li>\n<li><p>CTC Loss is for labeling sequence input with <strong>sequence<\/strong> output. If you only have \na single class label output for the sequence input, you're probably better off using a softmax cross entropy loss on the output of the last time step of the RNN cell.<\/p><\/li>\n<\/ol>\n\n<p>If you do end up using CTC loss, you can see how I've constructed the training sequence through a reader here: <a href=\"https:\/\/stackoverflow.com\/questions\/42578532\/how-to-generate-read-sparse-sequence-labels-for-ctc-loss-within-tensorflow\">How to generate\/read sparse sequence labels for CTC loss within Tensorflow?<\/a>.<\/p>\n\n<p>As an example, after I batch two examples that have label sequences <code>[44, 45, 26, 45, 46, 44, 30, 44]<\/code> and <code>[5,  8, 17,  4, 18, 19, 14, 17, 12]<\/code>, respectively, I get the following result from evaluating the (batched) SparseTensor:<\/p>\n\n<pre><code>SparseTensorValue(indices=array([[0, 0],\n       [0, 1],\n       [0, 2],\n       [0, 3],\n       [0, 4],\n       [0, 5],\n       [0, 6],\n       [0, 7],\n       [1, 0],\n       [1, 1],\n       [1, 2],\n       [1, 3],\n       [1, 4],\n       [1, 5],\n       [1, 6],\n       [1, 7],\n       [1, 8]]), values=array([44, 45, 26, 45, 46, 44, 30, 44,  5,  8, 17,  4, 18, 19, 14, 17, 12], dtype=int32), dense_shape=array([2, 9]))\n<\/code><\/pre>\n\n<p>Notice how the rows of the indices in the sparse tensor value correspond to the batch number and the columns correspond to the sequence index for that particular label. The values themselves are the sequence label classes. The rank is 2 and the size of the last dimension (nine in this case) is the length of the longest sequence. <\/p>\n"}]}
{"QuestionId":42488475,"AnswerCount":1,"Tags":"<tensorflow><time-series><lstm><dropout>","CreationDate":"2017-02-27T14:40:07.447","AcceptedAnswerId":42491902.0,"OwnerUserId":6534643.0,"Title":"Tensorflow LSTM Dropout Implementation","Body":"<ul>\n<li>How specifically does tensorflow apply <em>dropout<\/em> when calling tf.nn.rnn_cell.DropoutWrapper() ?<\/li>\n<\/ul>\n\n<p>Everything I read about applying dropout to rnn's references this paper by <a href=\"https:\/\/arxiv.org\/pdf\/1409.2329.pdf\" rel=\"noreferrer\">Zaremba et. al<\/a> which says don't apply dropout between recurrent connections. Neurons should be dropped out randomly before or after LSTM layers, but not inter-LSTM layers. Ok.<\/p>\n\n<ul>\n<li>The question I have is how are the neurons turned off with respect to time? <\/li>\n<\/ul>\n\n<p>In the paper that everyone cites, it seems that a random 'dropout mask' is applied at each timestep, rather than generating one random 'dropout mask' and reusing it, applying it to all the timesteps in a given layer being dropped out. Then generating a new 'dropout mask' on the next batch.<\/p>\n\n<p>Further, and probably what matters more at the moment, how does tensorflow do it? I've checked the tensorflow api and tried searching around for a detailed explanation but have yet to find one. <\/p>\n\n<ul>\n<li>Is there a way to dig into the actual tensorflow source code?<\/li>\n<\/ul>\n","answers":[{"AnswerId":"42491902","CreationDate":"2017-02-27T17:19:20.670","ParentId":null,"OwnerUserId":"6933420","Title":null,"Body":"<p>You can check the implementation <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py#L475\" rel=\"nofollow noreferrer\">here.<\/a><\/p>\n\n<p>It uses the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/python\/ops\/nn_ops.py#L1885\" rel=\"nofollow noreferrer\">dropout op<\/a> on the input into the RNNCell, then on the output, with the keep probabilities you specify. <\/p>\n\n<p>It seems like each sequence you feed in gets a new mask for input, then for output. No changes inside of the sequence. <\/p>\n"}]}
{"QuestionId":42488615,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><cublas>","CreationDate":"2017-02-27T14:45:29.410","AcceptedAnswerId":null,"OwnerUserId":4650448.0,"Title":"Failed to create CUBLAS handle. Tensorflow interaction with OpenCV","Body":"<p>I am trying to use a PlayStation Eye Camera for a deep reinforcement learning project. The network, TensorFlow installation (0.11) and CUDA (8.0) are functional because I have been able to train the network on a simulation.<\/p>\n\n<p>Now when I am trying to read in images from the real camera the network code crashes with the error below. Is there a mistake in my OpenCV installation (3.2.0) or is there another problem? I would be eternally grateful because I have not been finding any information about this problem.<\/p>\n\n<pre><\/pre>\n\n<p>Relevant code from the camera class:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47467185","CreationDate":"2017-11-24T06:00:46.997","ParentId":null,"OwnerUserId":"1841143","Title":null,"Body":"<p>Maybe the following command helps:<\/p>\n\n<pre><code>sudo rm -rf .nv\/\n<\/code><\/pre>\n\n<p>Good luck.<\/p>\n"}]}
{"QuestionId":42488731,"AnswerCount":2,"Tags":"<tensorflow><jupyter-notebook><keras>","CreationDate":"2017-02-27T14:50:21.847","AcceptedAnswerId":42511033.0,"OwnerUserId":180387.0,"Title":"Getting 'undefined symbol: cudnnCreate' in Jupyter notebook, but not on commandline","Body":"<p>I am running sample deep network script using keras with tensorflow with a GPU. It's on ubuntu. Everything runs fine from the commandline; e.g., I can copy any example network into a python file and run it and have verified the GPU is being used.<\/p>\n\n<p>Problem: I've started up a Jupyter notebook to play around with some scripts and now I get a notification that the notebook's kernel has died and the following error shows up in the log:<\/p>\n\n<p><\/p>\n\n<p>I have a feeling that the notebook is having trouble finding the cuda libs. Are there some path variables that the notebook isn't getting? I have tried to set some using the os library (LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64, for example), so no avail.<\/p>\n","answers":[{"AnswerId":"42498399","CreationDate":"2017-02-28T00:41:15.843","ParentId":null,"OwnerUserId":"7632696","Title":null,"Body":"<p>My guess is that you didn't install cuDNN properly. Note that it's a separate installation step from CUDA:\n<a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow noreferrer\">https:\/\/developer.nvidia.com\/cudnn<\/a><\/p>\n\n<p>Your command-line job probably doesn't rely on cuDNN (perhaps it's not using convolutions) so you don't get that failure.<\/p>\n"},{"AnswerId":"42511033","CreationDate":"2017-02-28T14:13:19.593","ParentId":null,"OwnerUserId":"180387","Title":null,"Body":"<p>As I suspected, and as was suggested by Jonathan Hseu in a comment, my Jupyter notebook was using different environment variables than I was using from the commandline. Namely, my .bashrc file has the following line, which points to the proper cuDNN library:<\/p>\n\n<p><code>export LD_LIBRARY_PATH=\/usr\/local\/cuda\/lib64<\/code><\/p>\n\n<p>I (misguidedly) was running Jupyter notebook as a different user, so this environment variable was never getting set. Running it as myself, using my .bashrc, solved my problem.<\/p>\n"}]}
{"QuestionId":42488938,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2017-02-27T15:00:29.143","AcceptedAnswerId":null,"OwnerUserId":5232723.0,"Title":"Operation u'init_27' has been marked as not fetchable. tensorflow","Body":"<p>I'm triying to train my network using a convolutional neural network based on keras tensorflow this is my code i have an error on the function compile but i don't know why<\/p>\n\n<pre><\/pre>\n\n<p>error: <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42489019,"AnswerCount":1,"Tags":"<tensorflow><theano><keras><tensorboard><autoencoder>","CreationDate":"2017-02-27T15:04:23.800","AcceptedAnswerId":42507812.0,"OwnerUserId":7561641.0,"Title":"how to use TensorBoard callback AND TensorBoard server?","Body":"<p><a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">keras blog autoencoder code <\/a>\nI am trying to run the code for Convolutional Autoencode from<\/p>\n\n<p><a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-autoencoders-in-keras.html<\/a><\/p>\n\n<pre><\/pre>\n\n<p>after running it I run this code for training :<\/p>\n\n<pre><\/pre>\n\n<p>now I want to plot the result I using callback ! I type this <\/p>\n\n<pre><\/pre>\n\n<p>in my terminal and it successfully switch back to theano but when I run <\/p>\n\n<pre><\/pre>\n\n<p>it still imply that not switch back to tensorflow. Does anyone know how to fix it?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42507812","CreationDate":"2017-02-28T11:39:42.397","ParentId":null,"OwnerUserId":"7568768","Title":null,"Body":"<p>To switch to the Tensorflow backend you have to edit the <code>keras.json<\/code> file located in <code>~\/.keras<\/code>.<\/p>\n\n<p>You should see a line <code>\"backend\": \"theano\"<\/code>, change \"theano\" to \"tensorflow\" and if Tensorflow is properly installed it should work and the line \"Using TensorFlow backend.\" should appear when you import Keras.<\/p>\n"}]}
{"QuestionId":42489489,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-02-27T15:26:57.027","AcceptedAnswerId":null,"OwnerUserId":1506145.0,"Title":"Tensorflow - pick values from indicies, what is the operation called?","Body":"<h1>An example<\/h1>\n\n<p>Suppose I have a tensor  with shape <\/p>\n\n<pre><\/pre>\n\n<p>And a tensor  with shape  which describes what values to be selected in the innermost dimension<\/p>\n\n<pre><\/pre>\n\n<p>Then the result will be a  matrix with these values<\/p>\n\n<pre><\/pre>\n\n<p>What is this operation called in tensorflow and how to do it?<\/p>\n\n<h1>General<\/h1>\n\n<p>Note that the above shape  is only an example, it can be any dimension. Some conditions for this operation:<\/p>\n\n<ul>\n<li><\/li>\n<li><\/li>\n<li><\/li>\n<\/ul>\n","answers":[{"AnswerId":"42491301","CreationDate":"2017-02-27T16:52:02.140","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>I think you can emulate this with <code>tf.gather_nd<\/code>. You will just have to convert \"your\" indices to a representation that is suitable for <code>tf.gather_nd<\/code>. The following example here is tied to your specific example, i.e. input tensors of shape <code>(2, 2, 2)<\/code> but I think this gives you an idea how you could write the conversion for input tensors with arbitrary shape, although I am not sure how easy it would be to implement this (haven't thought about it too long). Also, I'm not claiming that this is the easiest possible solution.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nvalues = np.array([[[0, 1], [2, 3]], [[4, 5], [6, 7]]])\nvalues_tf = tf.constant(values)\nindices = np.array([[1, 0], [0, 0]])\n\nconverted_idx = []\nfor k in range(values.shape[0]):\n    outer = []\n    for l in range(values.shape[1]):\n        inds = [k, l, indices[k][l]]\n        outer.append(inds)\n        print(inds)\n    converted_idx.append(outer)\n\nwith tf.Session() as sess:\n    result = tf.gather_nd(values_tf, converted_idx)\n    print(sess.run(result))\n<\/code><\/pre>\n\n<p>This prints<\/p>\n\n<pre><code>[[1 2]\n [4 6]]\n<\/code><\/pre>\n\n<p><strong>Edit<\/strong>: To handle arbitrary shapes here is a recursive solution that should work (only tested on your example):<\/p>\n\n<pre><code>def convert_idx(last_dim_vals, ori_indices, access_to_ori, depth):\n    if depth == len(last_dim_vals.shape) - 1:\n        inds = access_to_ori + [ori_indices[tuple(access_to_ori)]]\n        return inds\n\n    outer = []\n    for k in range(ori_indices.shape[depth]):\n        inds = convert_idx(last_dim_vals, ori_indices, access_to_ori + [k], depth + 1)\n        outer.append(inds)\n    return outer\n<\/code><\/pre>\n\n<p>You can use this together with the original code I posted like so:<\/p>\n\n<pre><code>...\nconverted_idx = convert_idx(values, indices, [], 0)\nwith tf.Session() as sess:\n    result = tf.gather_nd(values_tf, converted_idx)\n    print(sess.run(result))\n<\/code><\/pre>\n"}]}
{"QuestionId":42489804,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-27T15:41:34.377","AcceptedAnswerId":null,"OwnerUserId":886724.0,"Title":"SyncReplicasOptimizer does not synchronize workers?","Body":"<p>\nI am trying to use SyncReplicasOptimizer to train a model by a synchronized SGD. I followed the example in mnist_replicas.py. I am using 1 ps and 2 workers.<\/p>\n\n<p>I added a variable  and an operation to increase it by 1 (This is just to understand what is going on):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>then after creating the session by the chief I do the following in my training loop :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>My understanding of SyncReplicasOptimizer as explained <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard3\/tf.train.SyncReplicasOptimizer.md\" rel=\"nofollow noreferrer\">here<\/a> is that at every execution of my training loop, it will wait for my 2 workers to end their backpropagation, average the gradients and apply them.<\/p>\n\n<p>So I am expecting that both workers print :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Which is the case for my chief worker, but the other one prints :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>and that patterns changes between executions.<\/p>\n\n<p>Am I misunderstanding something?<\/p>\n\n<p>Moreover, It happens that the chief worker ends, whereas the other worker does not get further after printing :<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>and keeps looping. It seems to be a race condition where the chief worker ends, and the variable  is somehow corrupted, so the other worker never sees it incremented to  and then never stops.\nAny idea to avoid that?<\/p>\n\n<p>Thanks for help<\/p>\n","answers":[{"AnswerId":"42635441","CreationDate":"2017-03-06T21:00:40.267","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>There are a few things you are observing, and I think we can unpack them to understand things better.<\/p>\n\n<p>Sync replicas works by adding barriers to the training process. One of the barriers prevents the worker from finishing its step until the chief worker has updated the weights from the previous iteration. If the chief worker exits, then, it might be that the other workers are never unblocked, hence you seeing the non-chief worker stuck at epoch 4 (it is actually stuck trying to run the train_op, I think).<\/p>\n\n<p>Also, what session are you using? If you're using a MonitoredTrainingSession with the correct hook, it was not designed to allow you to call session.run more than once per iteration in the training loop, so this might be creating the nondeterminism you're seeing in what gets printed in the normal workers. To see the right behavior I'd do something like<\/p>\n\n<pre><code>if is_chief:\n  inc_epoch_op = tf.assign_add(epoch, 1)\nelse:\n  inc_epoch_op = epoch\n\n_, __, ep = sess.run([train_op, global_step, inc_epoch_op])\nprint \"epoch:\", ep\n<\/code><\/pre>\n\n<p>Is this still weird?<\/p>\n"}]}
{"QuestionId":42490299,"AnswerCount":1,"Tags":"<tensorflow><neural-network><classification><conv-neural-network><tf-slim>","CreationDate":"2017-02-27T16:02:58.093","AcceptedAnswerId":42490953.0,"OwnerUserId":6671287.0,"Title":"Classifying images using pre-trained (Tensorflow) CNN","Body":"<p>I have trained alexnet_v2 on my own dataset, and now would like to use it within another application. This should be very simple and I've tried to implement it in a number of ways but either I get errors I can't work around or (in the case of the code below) it hangs indefinitely.<\/p>\n\n<p>Ideally, I'd like it in C++ (but the C++ API seems unreliable, or at least has outdated documentation in many places, so python is acceptable), and I'd like to be classifying large groups of images (for example: providing the program with 80 images of animals and returning whether any of them show a cat).<\/p>\n\n<p>Am I going about this the right way with the code below? If so, how can I fix it.<\/p>\n\n<p>If not, are there any working examples of a better way?<\/p>\n\n<p>Many thanks.<\/p>\n\n<pre><\/pre>\n\n<p>EDIT: After adding line in bold, the program no longer hangs. However I'm getting a placeholder error:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: You must feed a value for placeholder tensor\n  'Placeholder' with dtype string [[Node: Placeholder =\n  Placeholderdtype=DT_STRING, shape=[],\n  _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n<\/blockquote>\n\n<p>I've double checked the spelling, and as far as I can see I'm feeding it correctly. What's wrong?<\/p>\n","answers":[{"AnswerId":"42490953","CreationDate":"2017-02-27T16:35:59.413","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>tf.train.batch()<\/code> function uses background threads to prefetch examples, but you need to add an explicit command (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/start_queue_runners\" rel=\"nofollow noreferrer\"><code>tf.train.start_queue_runners(sess)<\/code><\/a>) to start these threads. Rewriting the last part of your code as follows should stop it hanging:<\/p>\n\n<pre><code>with tf.Session() as sess:\n  sess.run(initializer)\n  init_fn(sess)\n\n  # Starts background threads for input preprocessing.\n  tf.train.start_queue_runners(sess)\n\n  image_np, pred_np = sess.run(\n      [image_tensor, pred],\n      feed_dict={image_filename_placeholder: image_filename})\n<\/code><\/pre>\n"}]}
{"QuestionId":42490755,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-02-27T16:26:07.917","AcceptedAnswerId":42491100.0,"OwnerUserId":5744566.0,"Title":"TensorFlow placeholder dimension - what's the difference?","Body":"<p>New to TensorFlow but am stuck with this placeholder declaration question. What exactly is the difference between defining a placeholder x as:<\/p>\n\n<pre><\/pre>\n\n<p>as opposed to this? <\/p>\n\n<pre><\/pre>\n\n<p>I'm thinking in terms of matrices. So say, suppose variable x is fed 10 values, and seq_size is 3 - the first gives 10x3 and the second gives a 10x3x1. Why would tensorflow consider them differently? <\/p>\n","answers":[{"AnswerId":"42491100","CreationDate":"2017-02-27T16:42:18.907","ParentId":null,"OwnerUserId":"2642845","Title":null,"Body":"<p>Tensorflow will consider it differently for shape validation purpose.\nThe matrix multiplication for instance with a matrix of size 3x4 will not be possible with the second version as the dimensions 1 and 3 do not match. Tensorflow will be able to detect that at graph construction. <\/p>\n\n<p>Also, on the code readability side, it may be good for general understanding to have extra dimension 1 if this dimension might be changed in the future. For instance if your data points are univariate time series, using <\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, [None, seq_size, 1])\n<\/code><\/pre>\n\n<p>will make it easier to extend your result to multivariate time series with dimension <code>d&gt;1<\/code> with <\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, [None, seq_size, d])\n<\/code><\/pre>\n\n<p>as all the code already have this extra dimension accounted for.<\/p>\n"},{"AnswerId":"42516401","CreationDate":"2017-02-28T18:41:18.277","ParentId":null,"OwnerUserId":"7368783","Title":null,"Body":"<pre><code>x = tf.placeholder(tf.float32, [None, seq_size, 1])\n<\/code><\/pre>\n\n<p>Here 'x' is a placeholder to hold tensor of size [anything, seq_size, 1]. This holds good with matrix operations where some multi-dimensional operations can be carried out easily by up-casing them to higher dimensional matrix. <\/p>\n\n<p>P.S: No. of elements in array of shape [None, seq_size] &amp; [None, seq_size, 1] can be same. They can be reshaped to each other easily.<\/p>\n"}]}
{"QuestionId":42492589,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-27T17:56:21.943","AcceptedAnswerId":null,"OwnerUserId":886724.0,"Title":"Distributed tensorflow on distributed data : synchronize workers","Body":"<p>I am using distributed tensorflow for a deep model calibration.<\/p>\n\n<p>My data is spread on different machines (every machine contains part of the data). I created different workers, each one operating on the data of a its corresponding machine.<\/p>\n\n<p>My calibration loop looks like this :<\/p>\n\n<pre><\/pre>\n\n<p>Now, every N steps, I want to compute the loss on my validation data set. This validation data set is also located on different machines (every machine contains a part of the whole validation data set).<\/p>\n\n<p>So I want all my workers to compute the validation loss on the validation data of their corresponding machine, and then the chief worker to make an average of these losses.<\/p>\n\n<p>So the chief worker has to run its own validation loss computation on its machines, wait for the other workers to do the same on their corresponding machines, and then sum and average the results.<\/p>\n\n<p>Is there a simple manner to do this?       <\/p>\n","answers":[]}
{"QuestionId":42492824,"AnswerCount":3,"Tags":"<python><machine-learning><scikit-learn><neural-network><keras>","CreationDate":"2017-02-27T18:09:45.617","AcceptedAnswerId":42495410.0,"OwnerUserId":3849791.0,"Title":"Can I send callbacks to a KerasClassifier?","Body":"<p>I want the classifier to run faster and stop early if the patience reaches the number I set. In the following code it does 10 iterations of fitting the model.<\/p>\n\n<pre><\/pre>\n\n<p>Here is the resulting error-<\/p>\n\n<pre><\/pre>\n\n<p>I changed the cross_val_score in the following-<\/p>\n\n<pre><\/pre>\n\n<p>and now I get this error-<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"http:\/\/machinelearningmastery.com\/dropout-regularization-deep-learning-models-keras\/\" rel=\"noreferrer\">This code came from here<\/a>. The code is by far the most accurate I've used so far. The problem is that there is no defined  anywhere in the code. It also takes forever to fit. The  operation occurs at the  and there's no parameters to throw a callback in there.<\/p>\n\n<p>How do I go about doing this?\nAlso, how do I run the model trained on a test set?<\/p>\n\n<p>I need to be able to save the trained model for later use...<\/p>\n","answers":[{"AnswerId":"42495410","CreationDate":"2017-02-27T20:43:29.583","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Reading <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/wrappers\/scikit_learn.py\" rel=\"noreferrer\">from here<\/a>, which is the source code of KerasClassifier, you can pass it the arguments of fit and they should be used.\nI don't have your dataset so I cannot test it, but you can tell me if this works and if not I will try and adapt the solution. Change this line :<\/p>\n\n<pre><code>estimators.append(('mlp', KerasClassifier(build_fn=create_baseline, nb_epoch=300, batch_size=16, verbose=0, callbacks=[...your_callbacks...])))\n<\/code><\/pre>\n\n<p>A small explaination of what's happening : KerasClassifier is taking all the possibles arguments for <code>fit<\/code>, <code>predict<\/code>, <code>score<\/code> and uses them accordingly when each method is called. They made a function that filters the arguments that should go to each of the above functions that can be called in the pipeline. \nI guess there are several <code>fit<\/code> and <code>predict<\/code> calls inside the <code>StratifiedKFold<\/code> step to train on different splits everytime. <\/p>\n\n<p>The reason why it takes forever to fit and it fits 10 times is because one fit is doing 300 epochs, as you asked. So the KFold is repeating this step over the different folds :<\/p>\n\n<ul>\n<li>calls <code>fit<\/code> with all the parameters given to <code>KerasClassifier<\/code> (300 epochs and batch size = 16). It's training on 9\/10 of your data and using 1\/10 as validation. <\/li>\n<\/ul>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>Ok, so I took the time to download the dataset and try your code... First of all you need to correct a \"few\" things in your network : <\/p>\n\n<ul>\n<li><p>your input have a 60 features. You clearly show it in your data prep :<\/p>\n\n<pre><code>X = dataset[:,:60].astype(float)\n<\/code><\/pre>\n\n<p>so why would you have this :<\/p>\n\n<pre><code>model.add(Dropout(0.2, input_shape=(33,)))\n<\/code><\/pre>\n\n<p>please change to : <\/p>\n\n<pre><code>model.add(Dropout(0.2, input_shape=(60,)))\n<\/code><\/pre><\/li>\n<li><p>About your targets\/labels. You changed the objective from the original code (<code>binary_crossentropy<\/code>) to <code>categorical_crossentropy<\/code>. But you didn't change your Y array. So either do this in your data preparation :<\/p>\n\n<pre><code>from keras.utils.np_utils import to_categorical\nencoded_Y = to_categorical(encoder.transform(Y))\n<\/code><\/pre>\n\n<p>or change your objective back to <code>binary_crossentropy<\/code>.<\/p><\/li>\n<li><p>Now the network's output size : 122 on the last dense layer? your dataset obviously has 2 categories so why are you trying to output 122 classes? it won't match the target. Please change back your last layer to :<\/p>\n\n<pre><code>model.add(Dense(2, init='normal', activation='softmax'))\n<\/code><\/pre>\n\n<p>if you choose to use <code>categorical_crossentropy<\/code>, or <\/p>\n\n<pre><code>model.add(Dense(1, init='normal', activation='sigmoid'))\n<\/code><\/pre>\n\n<p>if you go back to <code>binary_crossentropy<\/code>.<\/p><\/li>\n<\/ul>\n\n<p>So now that your network compiles, I could start to troubleshout.<\/p>\n\n<p><strong>here is your solution<\/strong><\/p>\n\n<p>So now I could get the real error message. It turns out that when you feed <code>fit_params=whatever<\/code> in the <code>cross_val_score()<\/code> function, you are feeding those parameters to a pipeline. In order to know to which part of the pipeline you want to send those parameters you have to specify it like this :<\/p>\n\n<pre><code>fit_params={'mlp__callbacks':calls}\n<\/code><\/pre>\n\n<p>Your error was saying that the process couldn't unpack <code>'callbacks'.split('__', 1)<\/code> into 2 values. It was actually looking for the name of the pipeline's step to apply this to.<\/p>\n\n<p>It should be working now :)<\/p>\n\n<pre><code>results = cross_val_score(pipeline, X, encoded_Y, cv=kfold, fit_params={'mlp__callbacks':calls})\n<\/code><\/pre>\n\n<p>BUT, you should be aware of what's happening here... the cross validation actually calls the <code>create_baseline()<\/code> function to recreate the model from scratch 10 times an trains it 10 times on different parts of the dataset. So it's not doing epochs as you were saying, it's doing 300 epochs 10 times. \nWhat is also happening as a consequence of using this tool : since the models are always differents, it means the <code>fit()<\/code> method is applied 10 times on different models, therefore, the callbacks are also applied 10 different times and the files saved by <code>ModelCheckpoint()<\/code> get overriden and you find yourself only with the best model of the last run.<\/p>\n\n<p>This is intrinsec to the tools you use, I don't see any way around this. This comes as consequence to using different general tools that weren't especially thought to be used together with all the possible configurations.<\/p>\n"},{"AnswerId":"46913176","CreationDate":"2017-10-24T14:28:01.233","ParentId":null,"OwnerUserId":"3103059","Title":null,"Body":"<p>This is what I have done<\/p>\n\n<pre><code>results = cross_val_score(estimator, X, Y, cv=kfold,\n                      fit_params = {'callbacks': [checkpointer,plateau]})\n<\/code><\/pre>\n\n<p>and has worked so far<\/p>\n"},{"AnswerId":"42495388","CreationDate":"2017-02-27T20:42:22.567","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Try:<\/p>\n\n<pre><code>estimators.append(('mlp', \n                   KerasClassifier(build_fn=create_model2,\n                                   nb_epoch=300,\n                                   batch_size=16,\n                                   verbose=0,\n                                   callbacks=[list_of_callbacks])))\n<\/code><\/pre>\n\n<p>where <code>list_of_callbacks<\/code> is a list of callbacks you want to apply. You could find details <a href=\"https:\/\/keras.io\/scikit-learn-api\/\" rel=\"nofollow noreferrer\">here<\/a>. It's mentioned there that parameters fed to <code>KerasClassifier<\/code> could be legal fitting parameters.<\/p>\n\n<p>It's also worth to mention that if you are using multiple runs with GPUs there might be a problem due to several reported memory leaks especially when you are using theano. I also noticed that running multiple fits consequently may show results which seem to be not independent when using <code>sklearn<\/code> API. <\/p>\n\n<p><strong>Edit:<\/strong><\/p>\n\n<p>Try also:<\/p>\n\n<pre><code>results = cross_val_score(pipeline, X, encoded_Y, cv=kfold, fit_params = {'mlp__callbacks': calls})\n<\/code><\/pre>\n\n<p>Instead of putting callbacks list in a wrapper instantiation.<\/p>\n"}]}
{"QuestionId":42492834,"AnswerCount":1,"Tags":"<tensorflow><pip><bazel>","CreationDate":"2017-02-27T18:10:09.097","AcceptedAnswerId":42506875.0,"OwnerUserId":560624.0,"Title":"Redownloaded custom built tensorflow wheel wont install","Body":"<p>I've built Tensorflow with custom SIMD extensions and created a wheel for it. If I simply do  on the box that I built it on, that works. However if I upload the  file to cloud storage, and do  I get this error:<\/p>\n\n<pre><\/pre>\n\n<p>Do I need to configure my build differently somehow?<\/p>\n","answers":[{"AnswerId":"42506875","CreationDate":"2017-02-28T10:57:32.077","ParentId":null,"OwnerUserId":"481649","Title":null,"Body":"<p>(capturing the solution as an answer)<\/p>\n\n<p>The URL used for the download is not correct. The base url needed to be <code>storage.googleapis.com<\/code><\/p>\n"}]}
{"QuestionId":42492840,"AnswerCount":0,"Tags":"<python><input><encoding><tensorflow><recurrent-neural-network>","CreationDate":"2017-02-27T18:10:42.397","AcceptedAnswerId":null,"OwnerUserId":1269703.0,"Title":"Understanding tensorflow RNN encoder input for translation task","Body":"<p>I'm following this tutorial (<a href=\"https:\/\/theneuralperspective.com\/2016\/11\/20\/recurrent-neural-networks-rnn-part-3-encoder-decoder\/\" rel=\"nofollow noreferrer\">https:\/\/theneuralperspective.com\/2016\/11\/20\/recurrent-neural-networks-rnn-part-3-encoder-decoder\/<\/a>) to implement a RNN using tensorflow.<\/p>\n\n<p>I have prepared the input data for both encoder and decoder as described but I'm having trouble understanding how the dynamic_rnn function on tensorflow is expecting the input for this particular task so I'm going to describe the data I have for now in hopes someone has an idea.<\/p>\n\n<p>I have 20 sentences in english and in french. As is described on the above link, each sentence is padded according to the longest sentence in each language and tokens for GO and EOS are added where expected.<\/p>\n\n<p>Since not all the code is available on the link I tried doing some of the missing code for myself.  This is what my encoder code looks like<\/p>\n\n<pre><\/pre>\n\n<p>My embeddings are of shape [20, 60, 60] but I think this is incorrect. If I look into the documentation for dynamic_rnn (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn<\/a>) then the inputs are expected to be of size [batch, max_time, ...]. If I look into this example here (<a href=\"https:\/\/medium.com\/@erikhallstrm\/using-the-dynamicrnn-api-in-tensorflow-7237aba7f7ea#.4hrf7z1pd\" rel=\"nofollow noreferrer\">https:\/\/medium.com\/@erikhallstrm\/using-the-dynamicrnn-api-in-tensorflow-7237aba7f7ea#.4hrf7z1pd<\/a>) then the inputs are supposed to be [batch, truncated_backprop_length, state_size].<\/p>\n\n<p>The max length of a sentence for the encoder is 60.  This is where I am confused about what exactly should the 2 last numbers on the inputs should be. Should max_time be also the max length for my sentence sequence or should that be another quantity. Also should the last number be also the max length of the sentences' sequences or the length of the current input without the padding?<\/p>\n","answers":[]}
{"QuestionId":42492900,"AnswerCount":1,"Tags":"<python><python-2.7><list><numpy><keras>","CreationDate":"2017-02-27T18:13:02.817","AcceptedAnswerId":null,"OwnerUserId":3584858.0,"Title":"Convert list of list into numpy array without commas between each list element","Body":"<p>I have a list of list that looks like this:.<\/p>\n\n<p>I would like to convert it into a numpy array that looks like this:.<\/p>\n\n<p>Is there any way I can achieve this? I don't seem to be able to mimic the X_train dataset structure given by  even though .<\/p>\n","answers":[{"AnswerId":"42494781","CreationDate":"2017-02-27T20:05:35.980","ParentId":null,"OwnerUserId":"901925","Title":null,"Body":"<p>You need to explain the <code>keras<\/code> target better.  The presence or not of commas is an indication of the nature of the data structure, but it is really just a display convention.<\/p>\n\n<p>Python list uses the comma delimiter all the time:<\/p>\n\n<pre><code>In [751]: alist = [[1,2],[3,4]]\nIn [752]: alist\nOut[752]: [[1, 2], [3, 4]]\n<\/code><\/pre>\n\n<p>A numpy array can be displayed with and without the comma.<\/p>\n\n<pre><code>In [753]: arr=np.array(arr)\nIn [754]: arr\nOut[754]: \narray([[1, 2],\n       [3, 4]])\nIn [755]: str(arr)\nOut[755]: '[[1 2]\\n [3 4]]'\nIn [756]: print(arr)\n[[1 2]\n [3 4]]\n<\/code><\/pre>\n\n<p>To get your mix of commas and space I'd have to make an object dtype array:<\/p>\n\n<pre><code>In [757]: a=np.empty((2,), dtype=object)\nIn [758]: a\nOut[758]: array([None, None], dtype=object)\nIn [759]: print(a)\n[None None]\nIn [760]: a[0]=[1,2]\nIn [761]: a[1]=[3,4]\nIn [762]: a\nOut[762]: array([[1, 2], [3, 4]], dtype=object)\nIn [763]: print(a)\n[[1, 2] [3, 4]]\n<\/code><\/pre>\n\n<p>I can imagine the <code>keras<\/code> package constructing such an array, but I wouldn't recommend it to a beginner.<\/p>\n\n<hr>\n\n<p>From the keras imdb documentation<\/p>\n\n<blockquote>\n  <p>X_train, X_test: list of sequences, which are lists of indexes (integers). If the nb_words argument was specific, the maximum possible index value is nb_words-1. If the maxlen argument was specified, the largest possible sequence length is maxlen.<\/p>\n<\/blockquote>\n\n<p>It says list of lists, but it could be object dtype array of lists.  Sounds like those inner lists vary in length, which explains why it isn't a 2d array.<\/p>\n\n<pre><code> np.array([[1,2], [3,4,5]]) \n<\/code><\/pre>\n\n<p>Produces such an array.<\/p>\n"}]}
{"QuestionId":42493734,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-27T19:01:58.960","AcceptedAnswerId":43814166.0,"OwnerUserId":5265397.0,"Title":"How to initialize the LSTM cell in TensorFlow?","Body":"<p>I am trying to run the following code in TF v1.0 and it throws an error. I have created a lstm, and then defined the state variable to pass it to lstm_cell in order to calculate the outputs, but state cannot be initialized:<\/p>\n\n<pre><\/pre>\n\n<p>Here it the traceback:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43814166","CreationDate":"2017-05-05T21:45:06.063","ParentId":null,"OwnerUserId":"2335276","Title":null,"Body":"<p>This is because in the latest versions of Tensorflow, the return of <code>state_size<\/code> property of a default <code>BasicLSTMCell<\/code> is a <code>LSTMStateTuple<\/code> (a Python Tuple). <\/p>\n\n<p>If you check the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/contrib\/rnn\/python\/ops\/core_rnn_cell_impl.py\" rel=\"nofollow noreferrer\">source code<\/a>, you could see that the same number of units is returned in both elements of the tuple (it was concatenated along the same axis in previous versions) and should be considered when initalizing the cell states. <\/p>\n\n<p>Therefore, this should do the trick:<\/p>\n\n<pre><code>state = tf.zeros([20, lstm_cell.state_size[0]*2)\n<\/code><\/pre>\n"}]}
{"QuestionId":42494235,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-02-27T19:32:53.683","AcceptedAnswerId":42501708.0,"OwnerUserId":2832033.0,"Title":"Two top blobs in train.prototxt when using caffe.NetSpec() in Python","Body":"<p>I am using <em>caffe.NetSpec<\/em> in  to define and export in architecture using the following code :<\/p>\n\n<pre><\/pre>\n\n<p>But, when generating the , two Top blobs appears in the  layer as follow :<\/p>\n\n<pre><\/pre>\n\n<p>What's going wrong here? \nThanks<\/p>\n","answers":[{"AnswerId":"42501708","CreationDate":"2017-02-28T06:27:56.823","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>How about using a <code>NetSpec()<\/code> object?<\/p>\n\n<pre><code>import caffe\n\nns = caffe.NetSpec() # use this object to store the layers\nns.data, ns.label = L.Data(name='data',  ntop=2, \n                           data_param={'source':'', 'batch_size': 32})\nns.conv1_1 = L.Convolution(ns.data, name='conv1_1',\n                    convolution_param=\n                    {'kernel_size':3,'num_output':64,'pad':1},\n                    param=[{'lr_mult':1, 'decay_mult':1},\n                           {'lr_mult':2,'decay_mult':0}])\nprint str(ns.to_proto()) # print the net stored in ns object\n<\/code><\/pre>\n"}]}
{"QuestionId":42494479,"AnswerCount":0,"Tags":"<tensorflow><nvprof><nvvp>","CreationDate":"2017-02-27T19:46:51.990","AcceptedAnswerId":null,"OwnerUserId":1514682.0,"Title":"nvprof shows error with TensorFlow","Body":"<p>I am trying to run nvprof with .\nI am using following command <\/p>\n\n<pre><\/pre>\n\n<p>It starts the application but after sometime it shows following errors and application exits.<\/p>\n\n<pre><\/pre>\n\n<p>Any idea whats going wrong. File runs just fine without using nvprof.\n<br\/>\nNote: I have 0.12.0 version installed and I am on IBM PPC64le machine.<\/p>\n","answers":[]}
{"QuestionId":42494517,"AnswerCount":1,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-02-27T19:49:46.510","AcceptedAnswerId":42499178.0,"OwnerUserId":6115356.0,"Title":"Image captioning giving weak results","Body":"<p>I am trying to build an image captioning model.<\/p>\n\n<pre><\/pre>\n\n<p>I have tried running this model on the all 5 captions of the first 100 images of FLickr8k test dataset for 50 epochs. All captions are prepended with  and concatenated with . \nTo generate the caption I am giving the input image and  as the initial word. With each iteration I predict the probability distribution over the vocabulary and obtain the next word. In the next iteration I give  PredictedWord as the input and generate the probability distribution again.<\/p>\n\n<p>What happens is that I get the same probability distribution in every timestep. <\/p>\n\n<p>My question is that:<\/p>\n\n<ol>\n<li>Is my model too small to generate captions?<\/li>\n<li>Is the training data too small?<\/li>\n<li>Is the number of epochs too small?<\/li>\n<li>Is my entire approach wrong?<\/li>\n<\/ol>\n","answers":[{"AnswerId":"42499178","CreationDate":"2017-02-28T02:37:07.053","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>Before answering your question, I want to ask, what did you mean by iteration in the following statement?<\/p>\n\n<pre><code>What happens is that I get the same probability distribution in every iteration.\n<\/code><\/pre>\n\n<p>Given an image and the initial word, you should get the next word which should be given as input to generate the next word and this process should go on until you get a special token (ex., EOC) that represent the end of the caption.<\/p>\n\n<blockquote>\n  <ol>\n  <li>Is my model too small to generate captions?<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>I would say No but may be this model is small to generate good captions.<\/p>\n\n<blockquote>\n  <ol start=\"2\">\n  <li>Is the training data too small?<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>Yes, only 100 images is not enough to train a image caption generating neural network.<\/p>\n\n<blockquote>\n  <ol start=\"3\">\n  <li>Is the number of epochs too small?<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>No, 50 epochs is not too small. You can probably try by tuning other parameters, for example, learning rate!<\/p>\n\n<blockquote>\n  <ol start=\"4\">\n  <li>Is my entire approach wrong?<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>No, your approach is not wrong. You can augment your approach to generate good captions for images. You should find good examples in web, just go through them, I believe you will get ideas from them.<\/p>\n"}]}
{"QuestionId":42494557,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-27T19:52:08.283","AcceptedAnswerId":null,"OwnerUserId":2628128.0,"Title":"TensorFlow New Op: AttributeError: 'module' object has no attribute 'custom_op'","Body":"<p>I am creating new ops (<a href=\"https:\/\/www.tensorflow.org\/extend\/adding_an_op\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/extend\/adding_an_op<\/a>) for TensorFlow (r1.0) running both on x86 and ARMv7.<\/p>\n\n<p>Minor code modifications are necessary to run TensorFlow on ARMv7, but this guide helps a lot:\n<a href=\"https:\/\/github.com\/samjabrahams\/tensorflow-on-raspberry-pi\/blob\/master\/GUIDE.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/samjabrahams\/tensorflow-on-raspberry-pi\/blob\/master\/GUIDE.md<\/a>.<\/p>\n\n<p>But I noticed that the <strong>custom operations do not work on my ARMv7 installation<\/strong> of TensorFlow.<\/p>\n\n<p>For example, when I test my custom operation in a Python script on ARMv7:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error (that does not show up on x86):<\/p>\n\n<pre><\/pre>\n\n<p>I further investigated the issue, and apparently there is not my custom operation in the .so library file.<\/p>\n\n<pre><\/pre>\n\n<p>The same commands on x86 have the following output:<\/p>\n\n<pre><\/pre>\n\n<p>Does anybody have similar issue? Can we consider this a bug?<\/p>\n","answers":[{"AnswerId":"53540645","CreationDate":"2018-11-29T13:55:34.197","ParentId":null,"OwnerUserId":"5952028","Title":null,"Body":"<p>I hit a similar issue with a similar error message when I tried to load my new op, however, my problem was I tried to register a customized op that had the same op name as tensorflow, and that led to a name collision. Changing the name fixed it without recompiling TF. <\/p>\n\n<p>The error message I encountered:\n<code>AttributeError: module '6e237d88703da016805889179d3f5baa' has no attribute 'custom_op'<\/code><\/p>\n"},{"AnswerId":"42536614","CreationDate":"2017-03-01T16:09:46.283","ParentId":null,"OwnerUserId":"2628128","Title":null,"Body":"<p>Apparently, recompiling and re-installing the TF made it works.<\/p>\n"}]}
{"QuestionId":42494678,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><neural-network><deep-learning><recurrent-neural-network>","CreationDate":"2017-02-27T19:59:28.247","AcceptedAnswerId":null,"OwnerUserId":6317571.0,"Title":"Language Modelling with RNN and LSTM Cell in Tensorflow","Body":"<p>My RNN for language modelling is predicting only \"the\" \"and\" and \"unknown\" what's wrong with my code? <\/p>\n\n<p>Here I define the hyper parameters: <\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Here I generate the data (my input is given by word embeddings long 100 calculated with Word2Vec):<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Define the placeholders:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Inputs and desired outputs:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Forward pass:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Training:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[]}
{"QuestionId":42494695,"AnswerCount":0,"Tags":"<python><tensorflow><sequence>","CreationDate":"2017-02-27T20:00:26.613","AcceptedAnswerId":null,"OwnerUserId":7631801.0,"Title":"tf.train.import_meta_graph('model.meta') cannot handle seq2seq models with attention?","Body":"<p>Environment:\nUbuntu 16.04;\nTensorFlow v1.0.0 (CPU)<\/p>\n\n<p>When attempting to import a saved graph using \"tf.train.import_meta_graph('model.meta'),\" I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>This error isn't thrown when I retrain my model <strong>without attention<\/strong> and import the graph with the same line of code.\nIs loading a model trained with attention not currently supported? Here is what my attention implementation looks like:<\/p>\n\n<pre><\/pre>\n\n<p>Thanks!<\/p>\n","answers":[]}
{"QuestionId":42495549,"AnswerCount":0,"Tags":"<caffe><multi-gpu>","CreationDate":"2017-02-27T20:52:22.907","AcceptedAnswerId":null,"OwnerUserId":1245262.0,"Title":"Shouldn't Caffe run more quickly on 2 GPU's than on 1?","Body":"<p>We've just gotten a multi-gpu machine at work, and I'm trying to verify that 2 GPU's on Caffe are better than 1. To do this, I'm using the quick train example of the CIFAR-10 dataset. So, far, I'm finding that 2GPU's slows things down  and I don't understand why. <\/p>\n\n<p>The caffe version I'm running is:<\/p>\n\n<pre><\/pre>\n\n<p>The topology of our GPU's is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The processes they are each handling are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I trained on the CIFAR-10 dataset using this basic script:<\/p>\n\n<pre><\/pre>\n\n<p>with the slight variations of:<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>I would've expected the fastest result to be obtained by , followed by , then by . Instead, I saw the exact opposite.<\/p>\n\n<p>What I saw was this,<\/p>\n\n<p>For :<\/p>\n\n<pre><\/pre>\n\n<p>For :<\/p>\n\n<pre><\/pre>\n\n<p>For :<\/p>\n\n<pre><\/pre>\n\n<p>So, my resulting times to train are:<\/p>\n\n<pre><\/pre>\n\n<p>Clearly, there's something I don't understand about running multiple GPU's with Caffe. I had expected using 2 GPU's would give me a speed-up of about 1.8X. What am I not understanding here?<\/p>\n","answers":[]}
{"QuestionId":42495930,"AnswerCount":6,"Tags":"<tensorflow><out-of-memory><gpu><vram>","CreationDate":"2017-02-27T21:17:05.677","AcceptedAnswerId":45478028.0,"OwnerUserId":6917400.0,"Title":"Tensorflow OOM on GPU","Body":"<p>i'm training some Music Data on a LSTM-RNN in Tensorflow and encountered some Problem with GPU-Memory-Allocation which i don't understand: I encounter an OOM when there actually seems to be just about enough VRAM still available. \nSome background: \nI'm working on Ubuntu Gnome 16.04, using a GTX1060 6GB, Intel Xeon E3-1231V3 and 8GB RAM. \nSo now first the part of the error-message which i can understand, in the and i will add the whole error message in the end again for anyone who might ask for it to help: <\/p>\n\n<blockquote>\n  <p>I tensorflow\/core\/common_runtime\/bfc_allocator.cc:696] 8 Chunks of\n  size 256 totalling 2.0KiB I\n  tensorflow\/core\/common_runtime\/bfc_allocator.cc:696] 1 Chunks of size\n  1280 totalling 1.2KiB I\n  tensorflow\/core\/common_runtime\/bfc_allocator.cc:696] 5 Chunks of size\n  44288 totalling 216.2KiB I\n  tensorflow\/core\/common_runtime\/bfc_allocator.cc:696] 5 Chunks of size\n  56064 totalling 273.8KiB I\n  tensorflow\/core\/common_runtime\/bfc_allocator.cc:696] 4 Chunks of size\n  154350080 totalling 588.80MiB I\n  tensorflow\/core\/common_runtime\/bfc_allocator.cc:696] 3 Chunks of size\n  813400064 totalling 2.27GiB I\n  tensorflow\/core\/common_runtime\/bfc_allocator.cc:696] 1 Chunks of size\n  1612612352 totalling 1.50GiB I\n  tensorflow\/core\/common_runtime\/bfc_allocator.cc:700] Sum Total of\n  in-use chunks: 4.35GiB I\n  tensorflow\/core\/common_runtime\/bfc_allocator.cc:702] Stats:  <\/p>\n  \n  <p>Limit:        5484118016 <\/p>\n  \n  <p>InUse:        4670717952 <\/p>\n  \n  <p>MaxInUse:     5484118016 <\/p>\n  \n  <p>NumAllocs:    29 <\/p>\n  \n  <p>MaxAllocSize: 1612612352<\/p>\n  \n  <p>W tensorflow\/core\/common_runtime\/bfc_allocator.cc:274]\n  *********************___________*__***************************************************xxxxxxxxxxxxxx W tensorflow\/core\/common_runtime\/bfc_allocator.cc:275] Ran out of\n  memory trying to allocate 775.72MiB.  See logs for memory state. W\n  tensorflow\/core\/framework\/op_kernel.cc:993] Resource exhausted: OOM\n  when allocating tensor with shape[14525,14000]<\/p>\n<\/blockquote>\n\n<p>So i can read that there is a maximum of 5484118016 bytes to be allocated, \n4670717952 bytes are allready in use, and another 777.72MB = 775720000 bytes are to be allocated. 5484118016 bytes - 4670717952 bytes - 775720000 bytes = 37680064 bytes according to my calculator. \nSo there should still be 37MB of free VRAM after allocating the space for the new Tensor he wants to push in there. This seems also to be quite legit to me, as Tensorflow would probably (i guess?) not try to allocate more VRAM than there is still available and just put the rest of the data on hold in RAM or something. <\/p>\n\n<p>Now i guess there is just some big error in my thinking, but i would be quite gratefull if someone could explain to me, what this error is. The obvious solving-strategy to my problem is to just make my batches a bit smaller, having them each at around 1.5GB probably just is too big. Still i would love to know what the actual problem is there. <\/p>\n\n<p>edit: I found something telling me to try: <\/p>\n\n<pre><\/pre>\n\n<p>which still does not work, but as the tensorflow documentation lacks any explanation of what <\/p>\n\n<pre><\/pre>\n\n<p>would be, i would love to ask you guys.<\/p>\n\n<p>Adding the rest of the error message for anyone interested: <\/p>\n\n<p>Sorry for the long copy\/paste, but maybe someone would need\/want to see it, <\/p>\n\n<p>Thank you so much in advance, \nLeon<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"47796256","CreationDate":"2017-12-13T15:06:09.310","ParentId":null,"OwnerUserId":"7338066","Title":null,"Body":"<p>I resolve this issue by reducing <code>batch_size=52<\/code>\nOnly to reduce memory use is to reduce batch_size. <\/p>\n\n<blockquote>\n  <p>Batch_size depends on your gpu graphics card, size of VRAM, Cache memory etc.<\/p>\n<\/blockquote>\n\n<p>Please prefer this <a href=\"https:\/\/stackoverflow.com\/questions\/39076388\/tensorflow-deep-mnist-resource-exhausted-oom-when-allocating-tensor-with-shape\">Another Stack Overflow Link<\/a><\/p>\n"},{"AnswerId":"48160378","CreationDate":"2018-01-09T01:39:45.180","ParentId":null,"OwnerUserId":"2277037","Title":null,"Body":"<p>When encountering OOM on GPU I believe changing <code>batch size<\/code> is the right option to try at first. <\/p>\n\n<blockquote>\n  <p>For different GPU you may need different batch size based on the GPU\n  memory you have. <\/p>\n<\/blockquote>\n\n<p>Recently I faced the similar type of problem, tweaked a lot to do the different type of experiment.<\/p>\n\n<p>Here is the link to the <a href=\"https:\/\/stackoverflow.com\/questions\/48007984\/training-broke-with-resourceexausted-error\">question<\/a> (also some tricks are included).<\/p>\n\n<p>However, while reducing the size of the batch you may find that your training gets slower. So if you have multiple GPU you may use them. To check about your GPU you can write on the terminal,<\/p>\n\n<pre><code>nvidia-smi\n<\/code><\/pre>\n\n<p>It will show you necessary information about your gpu rack.<\/p>\n"},{"AnswerId":"45478028","CreationDate":"2017-08-03T07:50:00.627","ParentId":null,"OwnerUserId":"8035148","Title":null,"Body":"<p>Try to take a look at this<\/p>\n\n<blockquote>\n  <p>Be careful not to run the evaluation and training binary on the same\n  GPU or else you might run out of memory. Consider running the\n  evaluation on a separate GPU if available or suspending the training\n  binary while running the evaluation on the same GPU.<\/p>\n<\/blockquote>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/tutorials\/deep_cnn\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/deep_cnn<\/a><\/p>\n"},{"AnswerId":"57052118","CreationDate":"2019-07-16T07:29:23.877","ParentId":null,"OwnerUserId":"10227573","Title":null,"Body":"<p>Had same OOM problem running model permutations one after another. It seems after completing one model, then defining and running a new model, that GPU memory is NOT completely cleared of previous model(s), and something is building up in memory and causing eventual OOM error. <\/p>\n\n<p>Answer from g-eoj to another problem:<\/p>\n\n<pre><code>keras.backend.clear_session()\n<\/code><\/pre>\n\n<p>should clear the previous model. From <a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/backend\/<\/a>\nDestroys the current TF graph and creates a new one. Useful to avoid clutter from old models \/ layers.\nAfter running and saving one model, clear the session, then run next model.<\/p>\n"},{"AnswerId":"51505733","CreationDate":"2018-07-24T18:56:00.190","ParentId":null,"OwnerUserId":"9265084","Title":null,"Body":"<p>I came across the same problem. I shut down all the anaconda prompt windows and cleared all the python tasks. Reopened an Anaconda prompt window and executed the train.py file. It worked for me the next time. The Anaconda and Python terminals were taking up the memory which doesn't leave space for the training process. <\/p>\n\n<p>Also, try to reduce the batch size of the training process if the above approach doesn't work. <\/p>\n\n<p>Hope this helps <\/p>\n"},{"AnswerId":"55230018","CreationDate":"2019-03-18T21:04:11.717","ParentId":null,"OwnerUserId":"5020654","Title":null,"Body":"<p>I have recently had a very similar error and it was due to accidentally having a training process running in the background while trying to train in a different process. Stopping one fixed the error immediately.<\/p>\n"}]}
{"QuestionId":42496084,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-27T21:27:58.563","AcceptedAnswerId":42496257.0,"OwnerUserId":734263.0,"Title":"Code using tensorflow.mul() fails after upgrading to tensorflow v1.0","Body":"<p>This code runs on Tensorflow v0.12.1, but fails on my new installation on TF v1.0. Is it that this function is deprecated? What's the function I should use? (Tensorflow up and running so I believe it's not a misconfiguration)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42496257","CreationDate":"2017-02-27T21:39:22.157","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <code>tf.mul()<\/code> function has been renamed to <code>tf.multiply()<\/code> in TensorFlow 1.0.<\/p>\n"},{"AnswerId":"50565352","CreationDate":"2018-05-28T11:27:37.580","ParentId":null,"OwnerUserId":"8076093","Title":null,"Body":"<p>For your kind information , <code>tf.mul<\/code>, <code>tf.sub<\/code> and <code>tf.neg<\/code> were used in the prior version of tensor-flow.  They have been deprecated.<br>\nYou can instead use <code>tf.multiply<\/code>, <code>tf.subtract<\/code> and <code>tf.negative<\/code>. \nThanks.<\/p>\n"}]}
{"QuestionId":42496747,"AnswerCount":1,"Tags":"<matrix><indexing><lua><max><torch>","CreationDate":"2017-02-27T22:13:10.517","AcceptedAnswerId":42963505.0,"OwnerUserId":4752277.0,"Title":"Torch - Lua \/ Get max index from matrix","Body":"<p>I am trying to code a neural net for domineering<br>\nThe input is a matrix of 8 x 8 x 3. I organised the matrix as following :<br>\nThe first depth is for the state of the game, the second depth is flipped board and the last depth is the player plane<br>\nThe output is 8 x 8 is the best game to play aka the move to learn (generated by Monte Carlo Tree Search)<\/p>\n\n<p>Then the network is a 8 x 8 tensor with the probability of being the best game to play, \nI need to get the index (x,y) of the max probability of the tensor for me to <\/p>\n\n<p>I tried with the function torch.max(tensor, 2) and torch.max(tensor?1) but I didn't get what I need. <\/p>\n\n<p>Can someone have any clue to help me ? <\/p>\n\n<p>Thank you a lot !<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42963505","CreationDate":"2017-03-22T22:03:36.440","ParentId":null,"OwnerUserId":"1235026","Title":null,"Body":"<p>To get the maximum element (best_row, best_col) of out, <\/p>\n\n<pre><code>-- First get the maximum element and indices per row\nmaxP_per_row, bestColumn_per_row = torch.max(out,2)\n-- then get the best element and the best row\nbest_p, best_row = torch.max(maxP_per_row, 1)\n-- then find the best column for the best row\nbest_col = bestColumn_per_row[best_row]\n<\/code><\/pre>\n\n<p>You can do the same for target. Hope this helps.<\/p>\n"}]}
{"QuestionId":42497216,"AnswerCount":2,"Tags":"<tensorflow><recurrent-neural-network>","CreationDate":"2017-02-27T22:48:22.293","AcceptedAnswerId":null,"OwnerUserId":5265397.0,"Title":"What is the equivalent of tf.nn.rnn in new versions of TensorFlow?","Body":"<p>I used to create the RNN network, in version 0.8 of TensorFlow, using: <\/p>\n\n<pre><\/pre>\n\n<p> is not available anymore, and it sounds it has been moved to . What is the exact code to create RNN network out of a ?<\/p>\n\n<p>Or, in the case that I have an stacked LSTM, <\/p>\n\n<pre><\/pre>\n\n<p>What is the replacement for  in new versions of TensorFlow?<\/p>\n","answers":[{"AnswerId":"42497900","CreationDate":"2017-02-27T23:50:08.083","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>You should use <a href=\"https:\/\/web.archive.org\/web\/20170227234909\/https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn\" rel=\"nofollow noreferrer\"><code>tf.nn.dynamic_rnn<\/code><\/a>.<\/p>\n\n<p>FYI: <a href=\"https:\/\/stackoverflow.com\/q\/42356027\/395857\">What is the upside of using <code>tf.nn.rnn<\/code> instead of <code>tf.nn.dynamic_rnn<\/code> in TensorFlow?<\/a><\/p>\n"},{"AnswerId":"42539696","CreationDate":"2017-03-01T18:51:43.657","ParentId":null,"OwnerUserId":"3844056","Title":null,"Body":"<p><code>tf.nn.rnn<\/code> is equivalent to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/static_rnn\" rel=\"nofollow noreferrer\"><code>tf.nn.static_rnn<\/code><\/a>.<\/p>\n\n<p>Note: before <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.4\/RELEASE.md#release-120\" rel=\"nofollow noreferrer\">version 1.2 of TensorFlow<\/a>, the <em>namespace<\/em> <code>tf.nn.static_rnn<\/code> did not exist, but only <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.1\/api_docs\/python\/\" rel=\"nofollow noreferrer\"><code>tf.contrib.rnn.static_rnn<\/code><\/a> (which is now an <em>alias<\/em> for <code>tf.nn.static_rnn<\/code>).<\/p>\n"}]}
{"QuestionId":42497340,"AnswerCount":6,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-02-27T22:58:54.727","AcceptedAnswerId":42497456.0,"OwnerUserId":4188354.0,"Title":"How to convert one-hot encodings into integers?","Body":"<p>I have a numpy array data set with shape (100,10). Each row is a one-hot encoding. I want to transfer it into a nd-array with shape (100,) such that I transferred each vector row into a integer that denote the index of the nonzero index. Is there a quick way of doing this using numpy or tensorflow?<\/p>\n","answers":[{"AnswerId":"51473937","CreationDate":"2018-07-23T07:50:16.503","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<p>While I strongly suggest to use numpy for speed, <a href=\"http:\/\/mpu.readthedocs.io\/en\/latest\/ml.html#mpu.ml.one_hot2indices\" rel=\"nofollow noreferrer\"><code>mpu.ml.one_hot2indices(one_hots)<\/code><\/a> shows how to do it without numpy. Simply <code>pip install mpu --user --upgrade<\/code>.<\/p>\n\n<p>Then you can do<\/p>\n\n<pre><code>&gt;&gt;&gt; one_hot2indices([[1, 0], [1, 0], [0, 1]])\n[0, 0, 1]\n<\/code><\/pre>\n"},{"AnswerId":"54052897","CreationDate":"2019-01-05T14:18:13.183","ParentId":null,"OwnerUserId":"10615173","Title":null,"Body":"<p>You can use this simple code:<\/p>\n\n<pre><code>a=[[0,0,0,0,0,1,0,0,0,0]]\nj=0\nfor i in a[0]:\n    if i==1:\n        print(j)\n    else:\n        j+=1\n<\/code><\/pre>\n\n<p><strong>5<\/strong><\/p>\n"},{"AnswerId":"53390097","CreationDate":"2018-11-20T09:40:34.360","ParentId":null,"OwnerUserId":"7928119","Title":null,"Body":"<pre><code>def int_to_onehot(n, n_classes):\n    v = [0] * n_classes\n    v[n] = 1\n    return v\n\ndef onehot_to_int(v):\n    return v.index(1)\n\n\n&gt;&gt;&gt; v = int_to_onehot(2, 5)\n&gt;&gt;&gt; v\n[0, 0, 1, 0, 0]\n\n\n&gt;&gt;&gt; i = onehot_to_int(v)\n&gt;&gt;&gt; i\n2\n<\/code><\/pre>\n"},{"AnswerId":"42497761","CreationDate":"2017-02-27T23:35:37.750","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>You can use  <a href=\"https:\/\/web.archive.org\/web\/20170227233359\/https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.argmax.html\" rel=\"noreferrer\">numpy.argmax<\/a> or  <a href=\"https:\/\/web.archive.org\/web\/20170227233407\/https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/argmax\" rel=\"noreferrer\">tf.argmax<\/a>. Example:<\/p>\n\n<pre><code>import numpy as np  \na  = np.array([[0,1,0,0],[1,0,0,0],[0,0,0,1]])\nprint('np.argmax(a, axis=1): {0}'.format(np.argmax(a, axis=1)))\n<\/code><\/pre>\n\n<p>output:<\/p>\n\n<pre><code>np.argmax(a, axis=1): [1 0 3]\n<\/code><\/pre>\n\n<p>You may also want to look at  <a href=\"https:\/\/web.archive.org\/web\/20170227234444\/http:\/\/scikit-learn.org\/stable\/modules\/generated\/sklearn.preprocessing.LabelBinarizer.html#sklearn.preprocessing.LabelBinarizer.inverse_transform\" rel=\"noreferrer\"><code>sklearn.preprocessing.LabelBinarizer.inverse_transform<\/code><\/a>.<\/p>\n"},{"AnswerId":"57642488","CreationDate":"2019-08-25T00:48:17.267","ParentId":null,"OwnerUserId":"10173662","Title":null,"Body":"<p>What I do in these cases is something like this. The idea is to interpret the one-hot vector as an index of a 1,2,3,4,5... array.<\/p>\n\n<pre><code># Define stuff\nimport numpy as np\none_hots = np.zeros([100,10])\nfor k in range(100):\n    one_hots[k,:] = np.random.permutation([1,0,0,0,0,0,0,0,0,0])\n\n# Finally, the trick\nramp = np.tile(np.arange(0,10),[100,1])\nintegers = ramp[one_hots==1].ravel()\n<\/code><\/pre>\n\n<p>I prefer this trick because I feel <code>np.argmax<\/code> and other suggested solutions may be slower than indexing (although indexing may consume more memory)<\/p>\n"},{"AnswerId":"42497456","CreationDate":"2017-02-27T23:09:27.160","ParentId":null,"OwnerUserId":"2989337","Title":null,"Body":"<p>As pointed out by Frank Demoncourt, since a one_hot only has a single 1 and the rest are zeros, you can use argmax for this particular example.  In general, if you want to find a value in a numpy array, you'll probabaly want to consult <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.where.html\" rel=\"noreferrer\">numpy.where<\/a>.  Also, this stack exchange question:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/432112\/is-there-a-numpy-function-to-return-the-first-index-of-something-in-an-array\">Is there a NumPy function to return the first index of something in an array?<\/a><\/p>\n\n<p>Since a one-hot vector is a vector with all 0s and a single 1, you can do something like this:<\/p>\n\n<pre><code>&gt;&gt;&gt; import numpy as np\n&gt;&gt;&gt; a = np.array([[0,1,0,0],[1,0,0,0],[0,0,0,1]])\n&gt;&gt;&gt; [np.where(r==1)[0][0] for r in a]\n[1, 0, 3]\n<\/code><\/pre>\n\n<p>This just builds a list of the index which is 1 for each row.  The [0][0] indexing is just to ditch the structure (a tuple with an array) returned by <code>np.where<\/code> which is more than you asked for.<\/p>\n\n<p>For any particular row, you just want to index into a.  For example in the zeroth row the 1 is found in index 1.<\/p>\n\n<pre><code>&gt;&gt;&gt; np.where(a[0]==1)[0][0]\n1\n<\/code><\/pre>\n"}]}
{"QuestionId":42497396,"AnswerCount":1,"Tags":"<tensorflow><cmusphinx>","CreationDate":"2017-02-27T23:03:36.327","AcceptedAnswerId":null,"OwnerUserId":7627282.0,"Title":"g2p-seq2seq-cmudict model not found","Body":"<p>I am attempting to run the g2p-seq2seq tool in the interest of adding words to a dictionary for cmusphinx.<\/p>\n\n<p>I am following the instructions on <a href=\"https:\/\/github.com\/cmusphinx\/g2p-seq2seq\" rel=\"nofollow noreferrer\">the g2p github page<\/a>.<\/p>\n\n<p>I have installed TensorFlow using the virtualenv method as specified on the tensorflow install page.<\/p>\n\n<p>I am in the virtualenv environment using  as per specified on the tensorflow page.<\/p>\n\n<p>But when I attempt to run the command:<br>\n<\/p>\n\n<p>I receive <\/p>\n\n<p>But I have the \"g2p-seq2seq-cmudict\" folder downloaded from the link on the github page and it is in the \"Documents\" folder where I am running the command from.<\/p>\n\n<p>And lastly here is a picture for further understanding <a href=\"https:\/\/i.imgur.com\/ZjSatBC.png\" rel=\"nofollow noreferrer\">screenie<\/a><\/p>\n","answers":[{"AnswerId":"42618329","CreationDate":"2017-03-06T05:42:10.147","ParentId":null,"OwnerUserId":"5939045","Title":null,"Body":"<p>first you sould to determine the file of your dictionary will you build,<\/p>\n\n<p>example in <\/p>\n\n<p>cd \/home\/cmudict-master\/ pick your dictionary file -> cmudict.dict<\/p>\n\n<p>then you sould have some folder to include the model will be created after g2p training your dictionary<\/p>\n\n<p>example in \n\/home\/cmudict-master\/english-model<\/p>\n\n<p>for that example the command will be right here<\/p>\n\n<p>g2p-seq2seq --train cmudict.dict --model english-model<\/p>\n\n<p>hope it will done<\/p>\n"}]}
{"QuestionId":42497521,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-27T23:14:39.213","AcceptedAnswerId":42498143.0,"OwnerUserId":7542570.0,"Title":"Use of softmax with tf.nn.weighted_cross_entropy_with_logits","Body":"<p>Is it necessary to use tf.nn.softmax() to get the softmax of logits before using ?  <\/p>\n\n<p>I am doing binary classification on an imbalanced set, and have my  value set to [1.0, 15.0] to compensate for the latter class being underrepresented in the data.<\/p>\n\n<p>The other similar op  explicitly says not to use softmax beforehand, but the weighted version does not specify.  I have tried it both with and without, and when I use the softmax before the model does not learn (e.g. AUC converges to ).<\/p>\n\n<p>The last layer in my model is using elu activation on a  tensor.  My labels are coded as  for the first class, and  for the second.<\/p>\n","answers":[{"AnswerId":"42498143","CreationDate":"2017-02-28T00:15:22.600","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<blockquote>\n  <p>Is it necessary to use tf.nn.softmax() to get the softmax of logits before using tf.nn.weighted_cross_entropy_with_logits()?<\/p>\n<\/blockquote>\n\n<p>You shouldn't because <a href=\"https:\/\/web.archive.org\/web\/20170228001302\/https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/weighted_cross_entropy_with_logits\" rel=\"nofollow noreferrer\"><code>weighted_cross_entropy_with_logits<\/code><\/a> applies the sigmoid to the targets.<\/p>\n\n<p>You can see that in the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/bd7f1f4955cee1df4f6dbd0baba783fbcf845bc3\/tensorflow\/python\/ops\/nn_impl.py#L222-L242\" rel=\"nofollow noreferrer\">code<\/a>:<\/p>\n\n<pre><code>def weighted_cross_entropy_with_logits(targets, logits, pos_weight, name=None):\n  \"\"\"Computes a weighted cross entropy.\n\n     For brevity, let `x = logits`, `z = targets`, `q = pos_weight`.\n\n    The loss is:\n    qz * -log(sigmoid(x)) + (1 - z) * -log(1 - sigmoid(x))\n  = qz * -log(1 \/ (1 + exp(-x))) + (1 - z) * -log(exp(-x) \/ (1 + exp(-x)))\n  = qz * log(1 + exp(-x)) + (1 - z) * (-log(exp(-x)) + log(1 + exp(-x)))\n  = qz * log(1 + exp(-x)) + (1 - z) * (x + log(1 + exp(-x))\n  = (1 - z) * x + (qz +  1 - z) * log(1 + exp(-x))\n  = (1 - z) * x + (1 + (q - 1) * z) * log(1 + exp(-x))\n\n  Setting `l = (1 + (q - 1) * z)`, to ensure stability and avoid overflow,\n  the implementation uses\n  (1 - z) * x + l * (log(1 + exp(-abs(x))) + max(-x, 0))\n\n    \"\"\"\n    # The logistic loss formula from above is\n    #   (1 - z) * x + (1 + (q - 1) * z) * log(1 + exp(-x))\n    # For x &lt; 0, a more numerically stable formula is\n    #   (1 - z) * x + (1 + (q - 1) * z) * log(1 + exp(x)) - l * x\n    # To avoid branching, we use the combined version\n    #   (1 - z) * x + l * (log(1 + exp(-abs(x))) + max(-x, 0))\n    log_weight = 1 + (pos_weight - 1) * targets\n    return math_ops.add(\n        (1 - targets) * logits,\n        log_weight * (math_ops.log1p(math_ops.exp(-math_ops.abs(logits))) +\n                      nn_ops.relu(-logits)),\n        name=name)\n<\/code><\/pre>\n\n<p>FYI <a href=\"https:\/\/stats.stackexchange.com\/a\/254071\/12359\">Softmax vs Sigmoid function in Logistic classifier?<\/a><\/p>\n"}]}
{"QuestionId":42497836,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-02-27T23:43:26.650","AcceptedAnswerId":null,"OwnerUserId":665023.0,"Title":"Tensorflow variable initialization in a loop","Body":"<p>I am trying a Tensorflow example:<\/p>\n\n<pre><\/pre>\n\n<p>But the output is unexpected. I expected it to output:<\/p>\n\n<pre><\/pre>\n\n<p>BUt the actual output is:<\/p>\n\n<pre><\/pre>\n\n<p>Why is this? session.run(model) will initialize variable each time, is this statement correct?<\/p>\n","answers":[{"AnswerId":"43073706","CreationDate":"2017-03-28T15:28:06.887","ParentId":null,"OwnerUserId":"1560300","Title":null,"Body":"<p>It seems as you have to initialize it inside the loop as such<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>import tensorflow as tf\n\nfor i in range(5):\n    X = tf.Variable(0, name='x')\n    model = tf.global_variables_initializer()\n    with tf.Session() as sess:\n        sess.run(model)\n        X += 1\n        print(X.eval())\n<\/code><\/pre>\n"},{"AnswerId":"42498041","CreationDate":"2017-02-28T00:04:31.223","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<blockquote>\n  <p>session.run(model) will initialize variable each time<\/p>\n<\/blockquote>\n\n<p>That is correct. The issue is that each time <code>x = x + 1<\/code> a new addition is created in the graph, which explains the results you obtain.<\/p>\n\n<p>Graph after the first iteration:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ZsiCR.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ZsiCR.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After the second iteration:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/T5s1e.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/T5s1e.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After the third iteration:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/xECkM.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/xECkM.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After the fourth iteration:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DnDVI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DnDVI.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>After the fifth iteration:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DJZrw.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DJZrw.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>The code I use, mostly taken from Yaroslav Bulatov's answer in <a href=\"https:\/\/stackoverflow.com\/q\/42257015\/395857\">How can I list all Tensorflow variables a node depends on?<\/a>:<\/p>\n\n<pre><code>import tensorflow as tf\n\nimport matplotlib.pyplot as plt\nimport networkx as nx\n\ndef children(op):\n  return set(op for out in op.outputs for op in out.consumers())\n\ndef get_graph():\n  \"\"\"Creates dictionary {node: {child1, child2, ..},..} for current\n  TensorFlow graph. Result is compatible with networkx\/toposort\"\"\"\n\n  ops = tf.get_default_graph().get_operations()\n  return {op: children(op) for op in ops}\n\ndef plot_graph(G):\n    '''Plot a DAG using NetworkX'''        \n    def mapping(node):\n        return node.name\n    G = nx.DiGraph(G)\n    nx.relabel_nodes(G, mapping, copy=False)\n    nx.draw(G, cmap = plt.get_cmap('jet'), with_labels = True)\n    plt.show()\n\n\nx = tf.Variable(0, name='x')\nmodel = tf.global_variables_initializer()\nwith tf.Session() as session:\n    for i in range(5):\n        session.run(model)\n        x = x + 1\n        print(session.run(x))\n\n        plot_graph(get_graph())\n<\/code><\/pre>\n"}]}
{"QuestionId":42497910,"AnswerCount":1,"Tags":"<terminal><tensorflow><bazel>","CreationDate":"2017-02-27T23:51:04.627","AcceptedAnswerId":null,"OwnerUserId":2872779.0,"Title":"Tensorflow: Quantize graph not working","Body":"<p>I've been following this tutorial to quantize the graph for iOS: <a href=\"https:\/\/petewarden.com\/2016\/05\/03\/how-to-quantize-neural-networks-with-tensorflow\/\" rel=\"nofollow noreferrer\">https:\/\/petewarden.com\/2016\/05\/03\/how-to-quantize-neural-networks-with-tensorflow\/<\/a><\/p>\n\n<p>I run this in in terminal:<\/p>\n\n<pre><\/pre>\n\n<p>However, all it outputs is the following:<\/p>\n\n<pre><\/pre>\n\n<p>Why isn't it completing the command? Does my computer require a GPU?<\/p>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>Running the same command in a docker image outputs the following:<\/p>\n\n<pre><\/pre>\n\n<p><strong>Update<\/strong><\/p>\n\n<p>For anyone who encounters this, just run the command to quantize your graph in terminal without using docker, it may take a while (Mine took about an hour) but it should work and it doesn't require GPU. <\/p>\n","answers":[{"AnswerId":"42498396","CreationDate":"2017-02-28T00:40:47.523","ParentId":null,"OwnerUserId":"2872779","Title":null,"Body":"<p>Never mind, it just took an hour to output and wasn't updating.<\/p>\n"}]}
{"QuestionId":42497925,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><keras>","CreationDate":"2017-02-27T23:52:33.470","AcceptedAnswerId":42499697.0,"OwnerUserId":5896419.0,"Title":"Keras ValueError: Dimensions must be equal","Body":"<p>The goal of the model is to categorically classify video inputs by the word articulated with them. Each input has the dimensionality 45 frames, 1 gray color channel, 100 pixel rows, and 150 pixel columns (45, 1, 100, 150), while each corresponding output is a one hot encoded representation of one of 3 possible words (e.g. \"yes\" => [0, 0, 1]).<\/p>\n\n<p>During the compilation of the model, the following error occurs:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the script used to train the model:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42499697","CreationDate":"2017-02-28T03:32:43.653","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>According to <a href=\"https:\/\/keras.io\/layers\/convolutional\/#convolution2d\" rel=\"nofollow noreferrer\">Convolution2D<\/a> in Keras, the following should be the shape of input and filter.<\/p>\n\n<pre><code>shape of input = [batch, in_height, in_width, in_channels]\nshape of filter = [filter_height, filter_width, in_channels, out_channels]\n<\/code><\/pre>\n\n<p>So, the meaning of the error you are getting -<\/p>\n\n<pre><code>ValueError: Dimensions must be equal, but are 1 and 3 for 'Conv2D_94' (op: 'Conv2D') with \ninput shapes: [?,100,150,1], [3,3,3,32].\n<\/code><\/pre>\n\n<p><code>[?,100,150,1]<\/code> means <code>in_channels<\/code> value is 1 whereas <code>[3,3,3,32]<\/code> means <code>in_channels<\/code> value is 3. Thats why you are getting the error - <code>Dimensions must be equal, but are 1 and 3<\/code>.<\/p>\n\n<p>So you can change the shape of the filter to <code>[3, 3, 1, 32]<\/code>.<\/p>\n"}]}
{"QuestionId":42498040,"AnswerCount":1,"Tags":"<python><dictionary><tensorflow>","CreationDate":"2017-02-28T00:04:26.977","AcceptedAnswerId":42498463.0,"OwnerUserId":7632501.0,"Title":"Tensorflow runtime dictionary lookup, hendling output from tf.parse_single_sequence_example","Body":"<p>I am using tensorflow parse_single_sequence_example for decoding the record_string from TFRecordReader. It returns two dicts, one for context_features and other for sequence_features.  <\/p>\n\n<pre><\/pre>\n\n<p>context_parsed and sequence_parsed are both dictionaries. How do I get the tensor object associated with the keys. If I do the following fetch operation<\/p>\n\n<pre><\/pre>\n\n<p>It fails and understandably so.<\/p>\n\n<pre><\/pre>\n\n<p>How do I fetch the context_parsed['output'] tensor? How do I feed that tensor to some placeholder in my graph?<\/p>\n\n<pre><\/pre>\n\n<p>I add the above line and try to fetch it but it does not work and the terminal just gets stuck in ipython.<\/p>\n\n<pre><\/pre>\n\n<p>I am also adding the output of tf.contrib.learn.run_n<\/p>\n\n<pre><\/pre>\n\n<p>----> 1 context = tf.contrib.learn.run_n(out, n=1, feed_dict=None)<\/p>\n\n<pre><\/pre>\n\n<p><strong>How do I fetch the context_parsed['output'] tensor? How do I feed that tensor to some placeholder in my graph?<\/strong><\/p>\n","answers":[{"AnswerId":"42498463","CreationDate":"2017-02-28T00:48:05.820","ParentId":null,"OwnerUserId":"7632501","Title":null,"Body":"<p>I figured out the mistake I was doing. Actually, I didn't start the new thread for TFReader() and because of it, the terminal gets hanged up.<\/p>\n\n<pre><code>with tf.Session() as sess:\n    sess.run(tf.initialize_all_variables())\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n    print(sess.run(context_parsed['length']))\n    coord.join(threads)\n<\/code><\/pre>\n\n<p>The output is printed as 6<\/p>\n"}]}
{"QuestionId":42498274,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-02-28T00:29:01.723","AcceptedAnswerId":null,"OwnerUserId":7632204.0,"Title":"How to get real prediction from TensorFlow","Body":"<p>I'm really new to TensorFlow and MI in general. I've been reading a lot, been searching days, but haven't really found anything useful, so..<\/p>\n\n<p>My main problem is that every single tutorial\/example uses images\/words\/etc., and the outcome is just a vector with numbers between 0 and 1 (yeah, that's the probability). Like that beginner tutorial, where they want to identify numbers in images. So the result set is just a \"map\" to every single digit's (0-9) probability (kind of). Here comes my problem: I have no idea what the result could be, it could be 1, 2, 999, anything.<\/p>\n\n<p>So basically:<\/p>\n\n<pre><\/pre>\n\n<p>This really is just a simplified example. I would always have the same number of inputs and outputs, but how can I get the predicted values back with TensorFlow, not just something like ? I'm not really sure how clear my question is, if it's not, please ask.<\/p>\n","answers":[]}
{"QuestionId":42498768,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><gmm>","CreationDate":"2017-02-28T01:49:07.760","AcceptedAnswerId":null,"OwnerUserId":5217799.0,"Title":"Implement GMM using tensorflow","Body":"<p>I am trying to implement Gaussian Mixture Model using Tensorflow.  Is there any guide or example on how to implement it?<\/p>\n\n<p>I have tried to implement GMM from the tensorflow example <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/contrib\/factorization\/python\/ops\/gmm.py\" rel=\"nofollow noreferrer\">here<\/a>. But I think its buggy. <\/p>\n","answers":[{"AnswerId":"43929462","CreationDate":"2017-05-12T04:46:13.737","ParentId":null,"OwnerUserId":"4350596","Title":null,"Body":"<p>You can take a look at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/factorization\/python\/ops\/gmm_test.py\" rel=\"nofollow noreferrer\">gmm_test<\/a> of tensorflow. Simply use the gmm model is ok I think<\/p>\n\n<p>In the official test cases you can find many details, which I think is better than document<\/p>\n\n<p>But when I use it, there's a bug happened, I'm still investing it, hope you can help me<\/p>\n"}]}
{"QuestionId":42498876,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-28T02:04:12.040","AcceptedAnswerId":42504176.0,"OwnerUserId":5398272.0,"Title":"TensorFlow: \"No gradients provided for any variable\" and partial_run","Body":"<p><strong>Problem<\/strong><\/p>\n\n<p>Using Tensorflow's partial_run() method doesn't work as I expected. I use it towards the bottom of the supplied code, and I believe it's giving me the attached error.<\/p>\n\n<p>The general flow of data is that I need to get a prediction from the model, use that prediction in some non-tensorflow code (to program a software synthesiser) to then get audio features (MFCCS, RMS, FFT) after playing a midi note, which can be finally passed to the cost function to check how close the predicted patch was to recreating a desired sound supplied as the current example. <\/p>\n\n<p><strong>Code - omitted preprocessing<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>ERROR<\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42504176","CreationDate":"2017-02-28T08:56:29.770","ParentId":null,"OwnerUserId":"2130551","Title":null,"Body":"<p>The immediate error you are experiencing:<\/p>\n\n<blockquote>\n  <p>No gradients provided for any variable, check your graph for ops that do not support gradients, between variables<\/p>\n<\/blockquote>\n\n<p>is because there is no gradient path from your <code>cost<\/code> to your weights. This is because there are placeholders and calculations happening outside of your graph in between the weights and cost. Thus, there does not exist a path of gradients from cost to the weights.<\/p>\n\n<p>In other words, think about the setup.<\/p>\n\n<pre><code>Weights -&gt; prediction -&gt; get_features -&gt; calculate cost.\n<\/code><\/pre>\n\n<p>Now, think about the back-propagation, we can calculate the gradient of the cost, but we have no gradients from the cost to get_features or from get_features to prediction, because get_features isn't part of the graph:<\/p>\n\n<pre><code>Weights &lt;- prediction &lt;-\/- get_features &lt;-\/- calculate cost.\n<\/code><\/pre>\n\n<p>So, the weights will never be able to learn.\nYou need to somehow get a path from your cost back to the prediction if you want this setup to work, perhaps simulating the gradient of get_features in a backwards path of your graph. There may be an cleaner way, but I can't think of one right this moment.<\/p>\n\n<p>Hope that helps!<\/p>\n"}]}
{"QuestionId":42499340,"AnswerCount":2,"Tags":"<tensorflow><convolution>","CreationDate":"2017-02-28T02:56:07.133","AcceptedAnswerId":null,"OwnerUserId":6430408.0,"Title":"List of Tensors when single Tensor expected","Body":"<p>I use concat to get tensors as the input of CNN. But got the error:      List of Tensors when single Tensor expected<\/p>\n\n<pre><\/pre>\n\n<p>I have searched the answers for these question, add \n    before return, but still get the same error ,why?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"46648489","CreationDate":"2017-10-09T14:10:20.373","ParentId":null,"OwnerUserId":"7323453","Title":null,"Body":"<p>Maybe check again the type of <code>image_height, image_width<\/code> because sometimes it is necessary to cast these into an integer dtype, e.g. <code>tf.cast(image_height, tf.int32)<\/code><\/p>\n"},{"AnswerId":"42499634","CreationDate":"2017-02-28T03:25:57.113","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>The error <code>List of Tensors when single Tensor expected<\/code>  comes from the fact you wrote <code>tf.concat(0,[image_val,image])<\/code> instead of <code>tf.concat([image_val,image],0)<\/code>.<\/p>\n"}]}
{"QuestionId":42499592,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><lstm><word-sense-disambiguation>","CreationDate":"2017-02-28T03:21:49.150","AcceptedAnswerId":42512916.0,"OwnerUserId":7632928.0,"Title":"ResourceExhaustedError :OOM when allocating tensor with shape []","Body":"<pre><\/pre>\n\n<p>This is my code.It keeps going this mistake 'ResourceExhaustedError:OOM when allocating tensor with shape'\nI used guppy.Then got this.<a href=\"https:\/\/i.stack.imgur.com\/Mi9NH.png\" rel=\"nofollow noreferrer\">result of guppy<\/a><\/p>\n\n<p>Why the variable of tensorflow take so much space.<\/p>\n","answers":[{"AnswerId":"42512916","CreationDate":"2017-02-28T15:38:43.733","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem was caused by this line in the training loop: <\/p>\n\n<pre><code>while s + batch_size &lt; ran:\n    # ...\n    batch_xs1 = tf.nn.embedding_lookup(embedding_matrix, batch_id)\n<\/code><\/pre>\n\n<p>Calling the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/embedding_lookup\" rel=\"nofollow noreferrer\"><code>tf.nn.embedding_lookup()<\/code><\/a> function adds nodes to the TensorFlow graph, and&mdash;because these are never garbage collected&mdash;doing so in a loop causes a memory leak.<\/p>\n\n<p>The actual cause of the memory leak is probably the <code>embedding_matrix<\/code> NumPy array in the argument to <code>tf.nn.embedding_lookup()<\/code>. TensorFlow tries to be helpful and convert all NumPy arrays in the arguments to a function into <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"nofollow noreferrer\"><code>tf.constant()<\/code><\/a> nodes in the TensorFlow graph. However, in a loop, this will end up with multiple separate copies of the <code>embedding_matrix<\/code> copied into TensorFlow and then onto scarce GPU memory.<\/p>\n\n<p>The simplest solution is to move the <code>tf.nn.embedding_lookup()<\/code> call outside the training loop. For example:<\/p>\n\n<pre><code>def while_loop(s,e,step):\n  batch_id_placeholder = tf.placeholder(tf.int32)\n  batch_xs1 = tf.nn.embedding_lookup(embedding_matrix, batch_id_placeholder)\n\n  while s+batch_size&lt;ran:\n    batch_id=file_id[s:e]\n    batch_col=label_matrix[s:e]                                             \n\n    batch_label = csc_matrix((data, (batch_row, batch_col)), shape=(batch_size, n_classes))\n    batch_label = batch_label.toarray()\n\n    batch_xs=sess.run(batch_xs1, feed_dict={batch_id_placeholder: batch_id})\n<\/code><\/pre>\n"},{"AnswerId":"53826364","CreationDate":"2018-12-18T04:23:20.837","ParentId":null,"OwnerUserId":"10804221","Title":null,"Body":"<p>I recently had this problem with TF + Keras and previously with Darknet with yolo v3.\nMy dataset contained very large images for the memory of my two GTX 1050s.\nI had to resize the images to be smaller.\nOn average, a 1024x1024 image needs 6GB per GPU.<\/p>\n"}]}
{"QuestionId":42500034,"AnswerCount":1,"Tags":"<python><multithreading><tensorflow>","CreationDate":"2017-02-28T04:06:36.603","AcceptedAnswerId":null,"OwnerUserId":4317770.0,"Title":"Tensorflow multiple samples without batching","Body":"<p>I have a graph that unfortunately has some nodes that do not support batching.(custom ops that have yet to be fleshed out) I've been able to have multiple threads calling sess.run and throwing the data in through the feed_dict. I have now converted my data to tfrecords to properly utilize queues, but still can't find a way to tell it to run multiple instances of the graph in parallel without just having multiple threads calling sess.run(). I assume the tensorflow developers have created a more \"pythonic\" way to do so somewhere, but I have yet to find it. How do I do this within tensorflow?<\/p>\n\n<p>edit: even with the data batched, the previous question stands, as my computation spends half its time on the cpu, and half on the gpu, thus regardless of batching, one will be waiting on the other for half the time. I'd like to have the graph train multiple samples asynchronously to fill that space.<\/p>\n\n<p>edit 2: I guess I have to put pseudocode here for people who don't want to read the text above.<\/p>\n\n<pre><\/pre>\n\n<p>I'd like to do this in a \"pipeline\" manner, where as soon as the cpu op is done, it starts on another sample.<\/p>\n","answers":[{"AnswerId":"42500593","CreationDate":"2017-02-28T05:04:24.187","ParentId":null,"OwnerUserId":"3098092","Title":null,"Body":"<pre><code>import tensorflow as tf\n\ninput_example = get_input_example()\ncpu_output = some_cpu_only_op(input_example)\n\ncpu_output_batch = tf.train.batch(input_example, batch_size, num_threads)\ngpu_output_batch = a_bunch_of_gpu_ops(cpu_output_batch)\n\nwith tf.Session() as sess:\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n\n    for i in range(num_train_steps):\n        output_values = sess.run(gpu_output_batch)\n        do_stuff_with(output_values)\n\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n\n<p><code>input_example<\/code> will have to be fed by a queue.<\/p>\n"}]}
{"QuestionId":42500053,"AnswerCount":1,"Tags":"<python><caffe>","CreationDate":"2017-02-28T04:08:41.003","AcceptedAnswerId":null,"OwnerUserId":5545794.0,"Title":"How to convert VGG16 caffemodel in model zoo from old definition \"layers\" to \"layer\"?","Body":"<p>I want to add a new python layer to the VGG16 model, But is seems that Python layer <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/O2xSt0Th-qI\" rel=\"nofollow noreferrer\">does not support old layer definition \"layers\"<\/a>, and I want use the pre-trained VGG16 caffemodel trained using prototxt with \"layers\",but it can not be used in prototxt with \"layer\". So is there any to convert the caffemodal, making it applicable with \"layer\"?<\/p>\n","answers":[{"AnswerId":"42500534","CreationDate":"2017-02-28T04:59:22.587","ParentId":null,"OwnerUserId":"5545794","Title":null,"Body":"<p>just using the <strong>tools\/upgrate_net_proto_text<\/strong> to convert \"layers\" to \"layer\", and the caffemodel can be used by both of them.<\/p>\n"}]}
{"QuestionId":42500383,"AnswerCount":1,"Tags":"<python><c++><neural-network><deep-learning><caffe>","CreationDate":"2017-02-28T04:43:58.390","AcceptedAnswerId":42501608.0,"OwnerUserId":2467772.0,"Title":"How to create Data layer in caffe?","Body":"<p>I have a caffe code in Python version.\nThat python code has a  like<\/p>\n\n<pre><\/pre>\n\n<p>I like to crate in c++ caffe as<\/p>\n\n<pre><\/pre>\n\n<p>My query is how to create a data layer with size 300 x 250 with 0 value initialized.<\/p>\n\n<p>I looked at data layers in caffe and available data layers are<\/p>\n\n<pre><\/pre>\n\n<p>Which one is suitable for the Data layer what I like to have and how to initialize 0 value for each cell?<\/p>\n","answers":[{"AnswerId":"42501608","CreationDate":"2017-02-28T06:20:58.643","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>It seems like <a href=\"http:\/\/caffe.help\/manual\/layers\/dummydata.html\" rel=\"nofollow noreferrer\"><code>\"DummyData\"<\/code><\/a> layer will do the job for you:<\/p>\n\n<pre><code>layer {\n  type: \"DummyData\"\n  name: \"hidden_seed\"\n  top: \"hidden_seed\"\n  dummy_data_param {\n    shape { dim: 300 dim: 250 }  # your desired data shape\n    data_filler { type: \"constant\" value: 0 } # fill with all zeros\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":42500739,"AnswerCount":2,"Tags":"<python><tensorflow><distributed><lstm><grpc>","CreationDate":"2017-02-28T05:17:51.083","AcceptedAnswerId":null,"OwnerUserId":7583479.0,"Title":"Between-graph replication version of PTB rnn model is slower than single-gpu version ( even in tf 1.0.0 )","Body":"<p>I changed the PTB model (you can found it in tensorflow\/models\/tutorials\/rnn\/ptb) to a between-graph version, but this distributed version ( with 1 ps server, 2 worker ) has no speedup effect even if the ps and workers are in single machine. the timeline profiling shows significant delay between GPU jobs and CPU jobs for distributed version. below are codes and timeline graphs:<\/p>\n\n<pre><\/pre>\n\n<p>The timeline of original single-gpu verion ( with 1 ps server, 2 worker ) is as follow ( one iteration ):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QPXFa.jpg\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>The timeline of worker 0 of distributed verion is as follow ( one iteration ), worker 1 is has similar timeline:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/WKnwF.jpg\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>The machine has two Tesla m40 gpus, the single-gpu verion performance is about 11000 wps ( with GPU util is about 60% ), while the between graph version is only 6000 wps for each worker ( GPU util is about more than 30% for each gpu ), so the speedup is only 1.09 for two workers (two gpus). \nAt the same time, i also made a multi-gpu verions of PTB model (without distibuted between-graph or in-graph frameworks ) which has the 1.6+ speedup in the same machine. so what caused the poor perfromance for distributed version ? <\/p>\n\n<p>The run command:\nPS:  CUDA_VISIBLE_DEVICES=\"\" python ptb_word_lm_dist.py  --data_path=\/data\/simple-examples\/data\/ --model small --job_name=ps --task_index=0<\/p>\n\n<p>Worker 0\uff1a CUDA_VISIBLE_DEVICES=0 python ptb_word_lm_dist.py --data_path=\/data\/simple-examples\/data\/ --model small --job_name=worker --task_index=0<\/p>\n\n<p>Worker 1: CUDA_VISIBLE_DEVICES=1 python ptb_word_lm_dist.py  --data_path=\/data\/simple-examples\/data\/ --model small --job_name=worker --task_index=1<\/p>\n\n<p>( Tested with tensorflow 0.12. A modified version for tf 1.0.0 has worse performance )<\/p>\n","answers":[{"AnswerId":"43271777","CreationDate":"2017-04-07T07:09:03.887","ParentId":null,"OwnerUserId":"2994475","Title":null,"Body":"<p>Maybe the bottleneck is disk IO, because of the single-gpu verion performance is about 11000 wps ( with GPU util is about 60% )<\/p>\n"},{"AnswerId":"42519678","CreationDate":"2017-02-28T22:11:27.207","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>What if you do this using single process tensorflow? IE, single process with <code>CUDA_VISIBLE_DEVICES=0,1<\/code>? It would make sense to try that first, to rule out issues that are unrelated to distributed TensorFlow.<\/p>\n\n<p>I've seen a situation with distributed TensorFlow running model from <a href=\"https:\/\/github.com\/rafaljozefowicz\/lm\" rel=\"nofollow noreferrer\">https:\/\/github.com\/rafaljozefowicz\/lm<\/a> where 8 GPU on one worker was 8 times faster than 8 TensorFlow workers. <\/p>\n\n<p>Some digging found that this issue was a large contributor <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6116\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6116<\/a><\/p>\n\n<p>What happens now is that sending large Tensors through gRPC is extremely inefficient. The fix is in master but not yet in TF 1.0, so you have to get a latest nightly to try it out.<\/p>\n"}]}
{"QuestionId":42501198,"AnswerCount":0,"Tags":"<python><caffe><pycaffe>","CreationDate":"2017-02-28T05:53:43.630","AcceptedAnswerId":null,"OwnerUserId":7633523.0,"Title":"how to output all object proposals in py-faster rcnn","Body":"<p>I tried to run py-faster-rcnn  and got the following results.<\/p>\n\n<pre><\/pre>\n\n<p>But it only displayed one object proposal when I save the image result. How could I save the results of all 89 object proposals?<\/p>\n","answers":[]}
{"QuestionId":42501472,"AnswerCount":1,"Tags":"<tensorflow><google-cloud-ml><google-cloud-ml-engine>","CreationDate":"2017-02-28T06:12:34.283","AcceptedAnswerId":42502264.0,"OwnerUserId":4392784.0,"Title":"Disable eval using tf.learn.Experiment","Body":"<p>What's the best way to disable evaluation if I'm using tf.learn.Experiment?<\/p>\n\n<p>I'm running this <a href=\"https:\/\/github.com\/googlegenomics\/cloudml-examples\" rel=\"nofollow noreferrer\">model<\/a> which constructs an experiment.<\/p>\n\n<pre><\/pre>\n\n<p>To debug an issue with evaluation, I'd like to prevent evluation from running. Is there an easy way to do this?<\/p>\n","answers":[{"AnswerId":"42502264","CreationDate":"2017-02-28T07:05:34.043","ParentId":null,"OwnerUserId":"1399222","Title":null,"Body":"<p>The answer is going to depend on the method you invoke on <code>Experiment<\/code>. Presumably you are going to call train_and_evaluate, e.g., if <code>TF_CONFIG<\/code>'s task type is set to \"master\" (cf <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/contrib\/learn\/python\/learn\/learn_runner.py#L135\" rel=\"nofollow noreferrer\">this code<\/a>).<\/p>\n\n<p>In that case, you'll want to set <code>min_eval_frequency<\/code> to <code>0<\/code> or <code>None<\/code> (cf <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.0\/tensorflow\/contrib\/learn\/python\/learn\/experiment.py#L453\" rel=\"nofollow noreferrer\">this code<\/a>)<\/p>\n"}]}
{"QuestionId":42501715,"AnswerCount":5,"Tags":"<python><python-3.x><numpy><tensorflow>","CreationDate":"2017-02-28T06:28:12.393","AcceptedAnswerId":null,"OwnerUserId":5217799.0,"Title":"Alternative of numpy.linalg.pinv in tensorflow","Body":"<p>I am looking for a alternative of numpy.linalg.pinv in tensorflow.\nSo far I have found that tensorflow have only  which throws an error if matrix in not invertible. <\/p>\n","answers":[{"AnswerId":"47155138","CreationDate":"2017-11-07T10:08:52.553","ParentId":null,"OwnerUserId":"1380707","Title":null,"Body":"<p>TensorFlow provides an SVD op, so you can compute the pseudo-inverse from that quite easily:<\/p>\n\n<pre><code>def pinv(A, b, reltol=1e-6):\n  # Compute the SVD of the input matrix A\n  s, u, v = tf.svd(A)\n\n  # Invert s, clear entries lower than reltol*s[0].\n  atol = tf.reduce_max(s) * reltol\n  s = tf.boolean_mask(s, s &gt; atol)\n  s_inv = tf.diag(tf.concat([1. \/ s, tf.zeros([tf.size(b) - tf.size(s)])], 0))\n\n  # Compute v * s_inv * u_t * b from the left to avoid forming large intermediate matrices.\n  return tf.matmul(v, tf.matmul(s_inv, tf.matmul(u, tf.reshape(b, [-1, 1]), transpose_a=True)))\n<\/code><\/pre>\n"},{"AnswerId":"51087168","CreationDate":"2018-06-28T15:58:58.497","ParentId":null,"OwnerUserId":"10006663","Title":null,"Body":"<p>An alternative to support batching for faster computing is to replace <code>tf.diag<\/code> with <code>tf.matrix_diag<\/code>. Something like this:<\/p>\n\n<pre><code>def pinv(a, rcond=1e-15):\n    s, u, v = tf.svd(a)\n    # Ignore singular values close to zero to prevent numerical overflow\n    limit = rcond * tf.reduce_max(s)\n    non_zero = tf.greater(s, limit)\n\n    reciprocal = tf.where(non_zero, tf.reciprocal(s), tf.zeros(s.shape))\n    lhs = tf.matmul(v, tf.matrix_diag(reciprocal))\n    return tf.matmul(lhs, u, transpose_b=True)\n<\/code><\/pre>\n"},{"AnswerId":"42505372","CreationDate":"2017-02-28T09:51:30.530","ParentId":null,"OwnerUserId":"2721160","Title":null,"Body":"<p>I am not aware of numpy.linalg.pinv alternatives in tensorflow, yet regularization is an alternative in case a matrix is not invertible. An example: <\/p>\n\n<pre><code>try: \n    result = tf.matrix_inverse(input, adjoint=None, name=None)\nexcept: \n    input += np.identity((input.shape))* c \n    result = tf.matrix_inverse(input, adjoint=None, name=None)\n<\/code><\/pre>\n\n<p>Where c is a constant and should be very small such as c = 0.000001<\/p>\n\n<p>Yet matrix inversion is rather expensive computationally and should only be done if neccesary. For more information see: <a href=\"https:\/\/www.johndcook.com\/blog\/2010\/01\/19\/dont-invert-that-matrix\/\" rel=\"nofollow noreferrer\">do not invert that matrix<\/a><\/p>\n"},{"AnswerId":"44617373","CreationDate":"2017-06-18T16:54:39.253","ParentId":null,"OwnerUserId":"3921457","Title":null,"Body":"<p>You can combine the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/py_func\" rel=\"nofollow noreferrer\">tf.py_func<\/a> of <code>tensorflow<\/code> with <code>numpy<\/code>'s <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.12.0\/reference\/generated\/numpy.linalg.pinv.html\" rel=\"nofollow noreferrer\">pseudo-inverse<\/a> as follows:<\/p>\n\n<pre><code>return tf.py_func(np.linalg.pinv, [input], tf.float32)\n<\/code><\/pre>\n"},{"AnswerId":"50100560","CreationDate":"2018-04-30T12:28:20.783","ParentId":null,"OwnerUserId":"2628369","Title":null,"Body":"<p>An implementation that follows numpy's <a href=\"https:\/\/docs.scipy.org\/doc\/numpy-1.14.0\/reference\/generated\/numpy.linalg.pinv.html\" rel=\"nofollow noreferrer\"><code>pinv<\/code><\/a> implementation:<\/p>\n\n<pre><code>def pinv(a, rcond=1e-15):\n    s, u, v = tf.svd(a)\n    # Ignore singular values close to zero to prevent numerical overflow\n    limit = rcond * tf.reduce_max(s)\n    non_zero = tf.greater(s, limit)\n\n    reciprocal = tf.where(non_zero, tf.reciprocal(s), tf.zeros(s.shape))\n    lhs = tf.matmul(v, tf.matrix_diag(reciprocal))\n    return tf.matmul(lhs, u, transpose_b=True)\n<\/code><\/pre>\n\n<p>This supports both single and batch computations of pseudo-inverses:<\/p>\n\n<pre><code># Pseudo-inverse of one (4, 3) matrix, has shape (3, 4)\npinv(tf.random_normal((4, 3)))\n\n# Pseudo-inverses of two (4, 3) matrices, has shape (2, 3, 4)\npinv(tf.random_normal((2, 4, 3)))\n<\/code><\/pre>\n\n<p>Thank you @Andy Tsai for the hint!<\/p>\n"}]}
{"QuestionId":42502759,"AnswerCount":0,"Tags":"<amazon-web-services><keras><convolution>","CreationDate":"2017-02-28T07:35:24.637","AcceptedAnswerId":null,"OwnerUserId":7346966.0,"Title":"What is the optimal way to train neural networks on Amazon Web Services for training neural networks?","Body":"<p>Maybe this is a stupid question, but I can't solve my problem.<\/p>\n\n<p>I have a dataset contains a lot of images divided on classes (like <a href=\"https:\/\/challenge.kitware.com\/#challenge\/n\/ISIC_2017%3A_Skin_Lesion_Analysis_Towards_Melanoma_Detection\" rel=\"nofollow noreferrer\">this<\/a> or <a href=\"https:\/\/www.kaggle.com\/c\/diabetic-retinopathy-detection\/data\" rel=\"nofollow noreferrer\">this<\/a>). The size of the data is more than 5 Gb.\nI need to train the deep convolution neural network (e.g. <a href=\"http:\/\/www.cv-foundation.org\/openaccess\/content_cvpr_2015\/papers\/Szegedy_Going_Deeper_With_2015_CVPR_paper.pdf\" rel=\"nofollow noreferrer\">Inception<\/a>).<\/p>\n\n<p>Training of such a network requires a huge amount of computing resources that are not available in my computer. So, I decided to train it on Amazon GPU instance. I've made an instance like in <a href=\"http:\/\/machinelearningmastery.com\/develop-evaluate-large-deep-learning-models-keras-amazon-web-services\/\" rel=\"nofollow noreferrer\">this nice tutorial<\/a>. I connected it through WinSCP (explained <a href=\"http:\/\/docs.rightscale.com\/faq\/How_to_Transfer_Files_from_EC2_to_Desktop_Using_WinSCP.html\" rel=\"nofollow noreferrer\">here<\/a>). But I have a problem associated with uploading the data on my instance EBS volume. When I try to do this whith WinSCP it takes a lot of time (more than 10 hours). The speed of my internet connection is 10 Mb\/sec, but the speed of uploading files on ec2 is 50-60 Kb\/sec. I have no idea, why.<\/p>\n\n<p>Could you suggest the best way to solve my problem? For example, is it possible to create a disc without running instance, upload data to it and then attach it to the instance? If so, advise links to tutorials. <\/p>\n\n<p>The instance is Ubuntu, my operation system is Windows 10.<\/p>\n\n<p>Many thanks!<\/p>\n","answers":[]}
{"QuestionId":42503112,"AnswerCount":1,"Tags":"<tensorflow><gpu>","CreationDate":"2017-02-28T07:55:56.643","AcceptedAnswerId":42800320.0,"OwnerUserId":4917824.0,"Title":"GPU utilization mostly 0% during training","Body":"<p>(GTX 1080, Tensorflow 1.0.0)<\/p>\n\n<p>During the training  output (below) suggests that the GPU <strong>utilization<\/strong> is 0% most of the time (despite <strong>usage<\/strong> of GPU). Regarding the time I already train, that seems to be the case. Once in a while it peaks up to 100% or similar, for a second though.<\/p>\n\n<pre><\/pre>\n\n<p>The situation occurs to me as I described in <a href=\"https:\/\/github.com\/thtrieu\/darkflow\/issues\/67\" rel=\"noreferrer\">this<\/a> issue.\nThe problem can be replicated either with the code from that github repository or by following <a href=\"https:\/\/www.tensorflow.org\/tutorials\/image_retraining\" rel=\"noreferrer\">this<\/a> simple retraining example from tensorflow's website and passing restricted <a href=\"https:\/\/www.tensorflow.org\/tutorials\/using_gpu\" rel=\"noreferrer\">per_process_gpu_memory_fraction<\/a> (less than 1.0) like that in the session:<\/p>\n\n<pre><\/pre>\n\n<p>Question 1: How to really use (utilize) GPU during the training while using &lt;1.0 of the GPU?<\/p>\n\n<p>Question 2: How to really use full GPU (not setting it to &lt;1.0) with my graphic card?<\/p>\n\n<p>Help&amp;hints appreciated!<\/p>\n","answers":[{"AnswerId":"42800320","CreationDate":"2017-03-15T02:56:20.013","ParentId":null,"OwnerUserId":"4410813","Title":null,"Body":"<p>When you create a graph that is bigger than the memory of the GPU TensorFlow falls back to the CPU, to it uses RAM and CPU instead of the GPU. So just remove the option for <code>per_process_gpu_memory_fraction<\/code> and decrement the batch size. Most probably the examples uses a big batch size because it was trained in more than one GPU or in a CPU with >32Gb, which it is not your case. It can also be the optimizer algorithm you chose. SGD uses less memory than other algorithms, try to set it first. With 8Gb in the GPU you can try a batch size of 16 and SGD, it should work. Then you can increase the batch size or use other algoritms like RMSprop.<\/p>\n\n<p>If it is still not working you probably are doing something else. Like for example you are saving a checkpoint in every iteration. Saving a checkpoints is done in the CPU and probably takes much more time than a simple iteration in the GPU. That could be the reason you are seeing spikes in the GPU usage.<\/p>\n"}]}
{"QuestionId":42504652,"AnswerCount":2,"Tags":"<python-3.x><machine-learning><tensorflow>","CreationDate":"2017-02-28T09:19:32.473","AcceptedAnswerId":42504743.0,"OwnerUserId":4058157.0,"Title":"How to update Tensorflow on mac?","Body":"<blockquote>\n  <blockquote>\n    <p>My current Tensorflow version is 0.10.0. I need to update it to latest version.Latest one is 1.0.0.Is there any method that I can use in\n    terminal?<\/p>\n  <\/blockquote>\n  \n  <p>This is the python code which I used to find version<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42504743","CreationDate":"2017-02-28T09:23:21.137","ParentId":null,"OwnerUserId":"826970","Title":null,"Body":"<p>If you are using anaconda;<\/p>\n\n<pre><code>&gt;conda update tensorflow\n<\/code><\/pre>\n\n<p>If you are using pip;<\/p>\n\n<pre><code>&gt;pip install tensorflow --upgrade\n<\/code><\/pre>\n\n<p>If you want to force install a specific version on conda;<\/p>\n\n<pre><code>&gt;conda install tensorflow=1.0.0\n<\/code><\/pre>\n\n<p>If you want to force install a specific version on pip;<\/p>\n\n<pre><code>&gt;pip install 'tensorflow==1.0.0' --force-reinstall\n<\/code><\/pre>\n"},{"AnswerId":"47765300","CreationDate":"2017-12-12T04:20:13.770","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>In anaconda cloud, first pick a channel which has the latest version of tensorflow binary. The latest versions are usually available at the channel <code>conda-forge<\/code><\/p>\n\n<pre><code>conda update -f -c conda-forge tensorflow\n<\/code><\/pre>\n\n<p>This will upgrade your installation to the very latest version available. As of this writing version <code>1.4.0-py36_0<\/code><\/p>\n"}]}
{"QuestionId":42504669,"AnswerCount":3,"Tags":"<python><tensorflow><neural-network><keras><python-multiprocessing>","CreationDate":"2017-02-28T09:20:05.117","AcceptedAnswerId":42506478.0,"OwnerUserId":3431836.0,"Title":"Keras + Tensorflow and Multiprocessing in Python","Body":"<p>I'm using Keras with Tensorflow as backend.<\/p>\n\n<p>I am trying to save a model in my main process and then load\/run (i.e. call ) within another process.<\/p>\n\n<p>I'm currently just trying the naive approach from the docs to save\/load the model: <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"noreferrer\">https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model<\/a>.<br>\nSo basically:<\/p>\n\n<ol>\n<li> in main process<\/li>\n<li> in child process<\/li>\n<li> in child process<\/li>\n<\/ol>\n\n<p>However, it simply hangs on the  call.<\/p>\n\n<p>Searching around I've discovered this potentially related answer suggesting that Keras can only be utilized in one process: <a href=\"https:\/\/stackoverflow.com\/questions\/38176827\/using-multiprocessing-with-theano\">using multiprocessing with theano<\/a> but am unsure if this is true (can't seem to find much on this).<\/p>\n\n<p>Is there a way to accomplish my goal?  A high level description or short example is greatly appreciated.<\/p>\n\n<p>Note: I've attempted approaches along the lines of passing a graph to the process but failed since it seems tensorflow graphs aren't pickable (related SO post for that here: <a href=\"https:\/\/stackoverflow.com\/questions\/34900246\/tensorflow-passing-a-session-to-a-python-multiprocess\">Tensorflow: Passing a session to a python multiprocess<\/a>).  If there is indeed a way to pass the tensorflow graph\/model to the child process then I am open to that as well.<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"44300391","CreationDate":"2017-06-01T06:48:39.790","ParentId":null,"OwnerUserId":"1314209","Title":null,"Body":"<p>I created one simple example to show how to run Keras model in multiple processes with multiple gpus. Hope this sample could help you.\n<a href=\"https:\/\/github.com\/yuanyuanli85\/Keras-Multiple-Process-Prediction\" rel=\"noreferrer\">https:\/\/github.com\/yuanyuanli85\/Keras-Multiple-Process-Prediction<\/a><\/p>\n"},{"AnswerId":"42506478","CreationDate":"2017-02-28T10:40:25.473","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>From my experience - the problem lies in loading <code>Keras<\/code> to one process and then spawning a new process when the <code>keras<\/code> has been loaded to your main environment. But for some applications (like e.g. training a mixture of <code>Keras<\/code>models) it's simply better to have all of this things in one process. So what I advise is the following (a little bit cumbersome - but working for me) approach:<\/p>\n\n<ol>\n<li><p><strong>DO NOT LOAD KERAS TO YOUR MAIN ENVIRONMENT<\/strong>. If you want to load Keras \/ Theano \/ TensorFlow do it only in the function environment. E.g. <strong>don't<\/strong> do this:<\/p>\n\n<pre><code>import keras\n\ndef training_function(...):\n    ...\n<\/code><\/pre>\n\n<p>but do the following:<\/p>\n\n<pre><code>def training_function(...):\n    import keras\n    ...\n<\/code><\/pre><\/li>\n<li><p><strong>Run work connected with each model in a separate process:<\/strong> I'm usually creating workers which are making the job (like e.g. training, tuning, scoring) and I'm running them in separate processes. What is nice about it that whole memory used by this process is completely <strong>freed<\/strong> when your process is done. This helps you with loads of memory problems which you usually come across when you are using multiprocessing or even running multiple models in one process. So this looks e.g. like this:<\/p>\n\n<pre><code>def _training_worker(train_params):\n    import keras\n    model = obtain_model(train_params)\n    model.fit(train_params)\n    send_message_to_main_process(...)\n\ndef train_new_model(train_params):\n    training_process = multiprocessing.Process(target=_training_worker, args = train_params)\n    training_process.start()\n    get_message_from_training_process(...)\n    training_process.join()\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Different approach is simply preparing different scripts for different model actions. But this may cause memory errors especially when your models are memory consuming. <strong>NOTE<\/strong> that due to this reason it's better to make your execution strictly sequential.<\/p>\n"},{"AnswerId":"53839801","CreationDate":"2018-12-18T19:27:23.983","ParentId":null,"OwnerUserId":"8418305","Title":null,"Body":"<p>I created a decorator that fixed my code.<\/p>\n\n<pre><code>from multiprocessing import Pipe, Process\n\ndef child_process(func):\n    \"\"\"Makes the function run as a separate process.\"\"\"\n    def wrapper(*args, **kwargs):\n        def worker(conn, func, args, kwargs):\n            conn.send(func(*args, **kwargs))\n            conn.close()\n        parent_conn, child_conn = Pipe()\n        p = Process(target=worker, args=(child_conn, func, args, kwargs))\n        p.start()\n        ret = parent_conn.recv()\n        p.join()\n        return ret\nreturn wrapper\n\n@child_process\ndef keras_stuff():\n    \"\"\" Keras stuff here\"\"\"\n<\/code><\/pre>\n"}]}
{"QuestionId":42504864,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-28T09:28:10.170","AcceptedAnswerId":42512100.0,"OwnerUserId":2455061.0,"Title":"How to finish this very easy distributed training example?","Body":"<p>I'm using tensorflow version 0.12.1, and following <a href=\"https:\/\/www.tensorflow.org\/deploy\/distributed\" rel=\"nofollow noreferrer\">this doc<\/a>. <\/p>\n\n<p>What I want to do is to add 1 to  in every worker.<\/p>\n\n<p>My goal is to print results that are >, but I'm only getting .<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n","answers":[{"AnswerId":"42512100","CreationDate":"2017-02-28T15:00:00.363","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem is actually independent of distributed execution, and stems from these two lines:<\/p>\n\n<pre><code>  count = tf.Variable(0)\n  count = tf.add(count,tf.constant(1))\n<\/code><\/pre>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/add\" rel=\"nofollow noreferrer\"><code>tf.add()<\/code><\/a> op is a pure functional op, which creates a new tensor with its output each time it runs, rather than modifying its input. If you want the value to increase, and that increase to be visible across workers, you must use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable#assign_add\" rel=\"nofollow noreferrer\"><code>tf.Variable.assign_add()<\/code><\/a> method instead, as follows:<\/p>\n\n<pre><code>  count = tf.Variable(0)\n  increment_count = count.assign_add(1)\n<\/code><\/pre>\n\n<p>Then call <code>sess.run(increment_count)<\/code> inside your training loop to increment the value of the <code>count<\/code> variable.<\/p>\n"}]}
{"QuestionId":42505260,"AnswerCount":3,"Tags":"<machine-learning><tensorflow><neural-network><classification><deep-learning>","CreationDate":"2017-02-28T09:46:43.807","AcceptedAnswerId":null,"OwnerUserId":231624.0,"Title":"Image classification, narrow domain with custom labels","Body":"<p>Let's suppose I would like to <strong>classify motorbikes by model<\/strong>. <\/p>\n\n<ul>\n<li>there are <strong>couple of hundreds models of motorbikes<\/strong> I'm interested in.<\/li>\n<li>I do have <strong>tens, sometimes hundreds<\/strong> of pictures of each motorbike model.<\/li>\n<\/ul>\n\n<p>Can you please point me to the <strong>practical example<\/strong> that demonstrates how to <strong>train model on your data<\/strong> and then use it to classify images? It needs to be a <strong>deep learning<\/strong> model, not simple logistic regression.<\/p>\n\n<p>I'm not sure about it, but it seems like I can't use pre-trained neural net because it has been trained on wide range of objects like cat, human, cars etc. They may be not too good at distinguishing the motorbike nuances I'm interested in.<\/p>\n\n<p>I found couple of such examples (tensorflow has one), but sadly, all of them were using pre-trained model. None of it had example how to train it on your own dataset.<\/p>\n","answers":[{"AnswerId":"42509037","CreationDate":"2017-02-28T12:40:48.807","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>In cases like yours you either use transfer learning or fine tuning. If you have more then thousand images of motorbikes I would use fine tuning and if you have less transfer learning. <\/p>\n\n<p>Fine tuning is using a pre trained model and using a different classifier part. Then the new classifier part maybe the last 1-2 layers of the trained model are trained to your dataset. <\/p>\n\n<p>Transfer learning means using a pre trained model and letting it output features for an input image. Now you use a new classifier based on those features. Maybe a SVM or a logistic regression.<\/p>\n\n<p>An example for this can be seen here: <a href=\"https:\/\/github.com\/cpra\/dlvc2016\/blob\/master\/lectures\/lecture10.pdf\" rel=\"nofollow noreferrer\">https:\/\/github.com\/cpra\/dlvc2016\/blob\/master\/lectures\/lecture10.pdf<\/a>. slide 33.<\/p>\n"},{"AnswerId":"56284308","CreationDate":"2019-05-24T00:05:58.387","ParentId":null,"OwnerUserId":"11530294","Title":null,"Body":"<p>What you want is pretty EZ. Follow the darknet YOLO implementation<\/p>\n\n<p>Instruction: <code>https:\/\/pjreddie.com\/darknet\/yolo\/<\/code><\/p>\n\n<p>Code <code>https:\/\/github.com\/pjreddie\/darknet<\/code><\/p>\n\n<p>Training YOLO on COCO\nYou can train YOLO from scratch if you want to play with different training regimes, hyper-parameters, or datasets. Here's how to get it working on the COCO dataset.<\/p>\n\n<p>Get The COCO Data\nTo train YOLO you will need all of the COCO data and labels. The script scripts\/get_coco_dataset.sh will do this for you. Figure out where you want to put the COCO data and download it, for example:<\/p>\n\n<pre><code>cp scripts\/get_coco_dataset.sh data\ncd data\nbash get_coco_dataset.sh\n<\/code><\/pre>\n\n<p>Add your data inside and make sure it is same as testing samples. \nNow you should have all the data and the labels generated for Darknet.<\/p>\n\n<p>Then call training script with the pre-trained weight. <\/p>\n\n<p>Keep in mind that only training on your motorcycle may not result in good estimation. There would be biased result coming out, I red it somewhere b4. <\/p>\n\n<p>The rest is all inside the link. Good luck <\/p>\n"},{"AnswerId":"56284524","CreationDate":"2019-05-24T00:43:21.967","ParentId":null,"OwnerUserId":"51292","Title":null,"Body":"<p>This paper <a href=\"http:\/\/Quick,%20Draw!%20Doodle%20Recognition\" rel=\"nofollow noreferrer\">Quick, Draw! Doodle Recognition<\/a> from a <a href=\"https:\/\/www.kaggle.com\/c\/quickdraw-doodle-recognition\" rel=\"nofollow noreferrer\">kaggle challenge<\/a> may be similar enough to what you are doing. The code is on <a href=\"https:\/\/github.com\/kristineguo\/quickdraw\" rel=\"nofollow noreferrer\">github<\/a>. You may need some <a href=\"https:\/\/medium.com\/ymedialabs-innovation\/data-augmentation-techniques-in-cnn-using-tensorflow-371ae43d5be9\" rel=\"nofollow noreferrer\">data augmentation<\/a> if you only have a few hundred images for each category.<\/p>\n"}]}
{"QuestionId":42505293,"AnswerCount":0,"Tags":"<python-3.x><tensorflow><skflow><gmm>","CreationDate":"2017-02-28T09:48:20.127","AcceptedAnswerId":null,"OwnerUserId":5217799.0,"Title":"Input is not invertible. [[Node: MatrixInverse_2=MatrixInverse[T=DT_FLOAT, adjoint=false, _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](add_138)]]","Body":"<p>I am trying to implement Gaussian Mixture Model using Tensorflow's tf.learn(skflow) but I am getting the following error.<\/p>\n\n<pre><\/pre>\n\n<p>Error StackTrace :- <\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42506340,"AnswerCount":0,"Tags":"<keras>","CreationDate":"2017-02-28T10:32:49.317","AcceptedAnswerId":null,"OwnerUserId":826983.0,"Title":"Is there an alternative to fit_generator() without having to set samples_per_epoch?","Body":"<p>I do not understand the design of the  method. In my example the number of samples may vary and I am not able to determine the number of samples beforehand unless going through the whole data set beforehand.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>In my scenario I have recorded laps. Since those laps are not finished equally fast, the number of <em>samples<\/em> in each lap may vary.<\/p>\n\n<p>Imho it would be nice if I could simply throw a  exception to indicate the end of an epoch. Even better would be a  and a  exception to control epochs and the whole training based on these exception but I guess that's not possible.<\/p>\n\n<p>However, I'd like to know if I can use something like  without having to know the number of samples of each epoch.<\/p>\n","answers":[]}
{"QuestionId":42506384,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-28T10:35:46.583","AcceptedAnswerId":null,"OwnerUserId":6659511.0,"Title":"Using 'read_batch_record_features' with an Estimator","Body":"<p>(I'm using tensorflow 1.0 and Python 2.7)<\/p>\n\n<p>I'm having trouble getting an Estimator to work with queues. Indeed, if I use the deprecated SKCompat interface with custom data files and a given batch size, the model trains properly. I'm trying to use the new interface with an input_fn that batches features out of TFRecord files (equivalent to my custom data files). The scripts runs properly but the loss value doesn't change after 200 or 300 steps. It seems that the model is looping on a small input batch (this would explain why the loss converges so fast).<\/p>\n\n<p>I have a 'run.py' script that looks like the following:<\/p>\n\n<pre><\/pre>\n\n<p>My inputToModel function looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Any help or suggestions is welcome !                      <\/p>\n","answers":[{"AnswerId":"42508760","CreationDate":"2017-02-28T12:26:55.750","ParentId":null,"OwnerUserId":"5436892","Title":null,"Body":"<p>Try to use: <code>tf.global_variables_initializer().run()<\/code><\/p>\n\n<p>I wanna do a similar thing but I do not know how to use Estimator API with multi-threading. There is an Experiment class for serving too - might be useful<\/p>\n\n<p><strong>delete<\/strong> line <code>session = tf.Session(graph=tf.get_default_graph())<\/code>\nand <code>session.close()<\/code> and try:<\/p>\n\n<pre><code>with tf.Session() as sess:\n  tf.global_variables_initializer().run()\n<\/code><\/pre>\n"}]}
{"QuestionId":42506487,"AnswerCount":0,"Tags":"<python><machine-learning><deep-learning><theano><cross-entropy>","CreationDate":"2017-02-28T10:40:55.243","AcceptedAnswerId":null,"OwnerUserId":3109123.0,"Title":"I'm having nan in Cross entropy which is not reasonable","Body":"<p>I'm tracing back the problem in my network to detect the reason of having nan from the cross entropy. <\/p>\n\n<p>I took the values from the network where exactly I start to have Nan and test them alone on another function. <\/p>\n\n<p>Here I'm testing the 50 values together to find which one gave the NaN.  <\/p>\n\n<pre><\/pre>\n\n<p>and I have this result. <\/p>\n\n<pre><\/pre>\n\n<p>So in order to see why the nan exist here I test the function on hand and on wolfarma and on theano itself. I took the one before the last and test it alone by: <\/p>\n\n<pre><\/pre>\n\n<p>and the result is Not Nan. <\/p>\n\n<pre><\/pre>\n\n<p>So this number in the first function give NaN while it shouldn't ? <\/p>\n","answers":[]}
{"QuestionId":42507030,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-28T11:04:49.223","AcceptedAnswerId":42537180.0,"OwnerUserId":5718092.0,"Title":"Implementing Attention in Tensorflow","Body":"<p>I would like to check if my attention implementation is correct in TensorFlow.<\/p>\n\n<p>Basically, I am using the attention mentioned in <a href=\"https:\/\/arxiv.org\/pdf\/1509.06664v1.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1509.06664v1.pdf<\/a>. (Just the baseline attention, not the word-for-word attention). So far I implemented it without using the last hidden state h_N.<\/p>\n\n<pre><\/pre>\n\n<p>This graph compiles, runs and trains properly. (loss is decreasing etc) but the performance is lower than I would expect. I would appreciate if I could get a check if I'm doing it right. <\/p>\n\n<p>Generally,<\/p>\n\n<p>1) For multiplications that are [?, seq_len, dim] matrix multiplied by [dim, dim]. Is it correct to use tf.reshape from [?, seq_len, dim] to [-1,dim] and apply matmul of shapes [-1, dim] with [dim, dim] and then reshape back to [?, seq_len, dim] after matmul?<\/p>\n\n<p>2) I notice I obtain (?, seq_len) for a, the attention vector. So I am required to do (?, seq_len) x (?, dim, seq_len). <\/p>\n\n<p>Is it correct to cast and expand_dims from (?, seq_len) to (?, seq_len, 1) and then do a matmul (I think this was what batch_matmul does in previous versions).<\/p>\n\n<p>Thanks in advance! <\/p>\n","answers":[{"AnswerId":"42537180","CreationDate":"2017-03-01T16:36:54.140","ParentId":null,"OwnerUserId":"1812460","Title":null,"Body":"<p>Not sure <code>tf.einsum<\/code> in TF1.0 is implemented efficiently, but it would make the computation quite elegant.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nbatch_size = 3\nseq_len = 5\ndim = 2\n# [batch_size x seq_len x dim]  -- hidden states\nY = tf.constant(np.random.randn(batch_size, seq_len, dim), tf.float32)\n# [batch_size x dim]            -- h_N\nh = tf.constant(np.random.randn(batch_size, dim), tf.float32)\n\ninitializer = tf.random_uniform_initializer()\nW = tf.get_variable(\"weights_Y\", [dim, dim], initializer=initializer)\nw = tf.get_variable(\"weights_w\", [dim], initializer=initializer)\n\n# [batch_size x seq_len x dim]  -- tanh(W^{Y}Y)\nM = tf.tanh(tf.einsum(\"aij,jk-&gt;aik\", Y, W))\n# [batch_size x seq_len]        -- softmax(Y w^T)\na = tf.nn.softmax(tf.einsum(\"aij,j-&gt;ai\", M, w))\n# [batch_size x dim]            -- Ya^T\nr = tf.einsum(\"aij,ai-&gt;aj\", Y, a)\n\nwith tf.Session() as sess:\n    sess.run(tf.global_variables_initializer())\n    a_val, r_val = sess.run([a, r])\n    print(\"a:\", a_val, \"\\nr:\", r_val)\n<\/code><\/pre>\n"}]}
{"QuestionId":42507766,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><convolution>","CreationDate":"2017-02-28T11:37:46.500","AcceptedAnswerId":42510408.0,"OwnerUserId":7368783.0,"Title":"Why am I getting only one channeled-output through the tf.nn.conv2d?","Body":"<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I am just trying to use convolution and apply some blur effect initially. Yeah I know that my kernel values aren't right. But my question is, I am giving an input image that has 3 channels. How could I get an output image of 3 channels. Well. I tried. But all I get is some one channeled values alone. <\/p>\n","answers":[{"AnswerId":"42510408","CreationDate":"2017-02-28T13:44:02.663","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>You are passing a kernel of shape <code>[3, 3, 3, 1]<\/code> to <code>tf.nn.conv2d()<\/code>. If you want to get a 3-channel image output from your convolution, the fourth dimension of your kernel (called <code>out_channels<\/code> in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\">official documentation<\/a>) should be <code>3<\/code> instead of <code>1<\/code>; <code>[3, 3, 3, 3]<\/code> for example.<\/p>\n\n<p>You could also take a look at the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/conv2d\" rel=\"nofollow noreferrer\"><code>conv2d<\/code> documentation<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/34619177\/what-does-tf-nn-conv2d-do-in-tensorflow\">this question<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/42450389\/how-a-filter-in-convolutional-neural-network-can-generate-multiple-channels\">this question<\/a> to better understand Tensorflow's <code>conv2d<\/code> method.<\/p>\n"}]}
{"QuestionId":42508283,"AnswerCount":1,"Tags":"<python><keras><keras-layer>","CreationDate":"2017-02-28T12:02:10.887","AcceptedAnswerId":42510771.0,"OwnerUserId":2182857.0,"Title":"Loading Keras layers from config","Body":"<p>I am trying to use the \"layer_from_config\" Keras utilty to load layers from previously saved configurations, as described in:\n<a href=\"https:\/\/keras.io\/layers\/about-keras-layers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/about-keras-layers\/<\/a><\/p>\n\n<p>For starters, I'm trying to use it on a basic model<\/p>\n\n<pre><\/pre>\n\n<p>As expected,  returns a dict type object, and printing it reads<\/p>\n\n<pre><\/pre>\n\n<p>However, when I run the code above I'm getting the following error message<\/p>\n\n<pre><\/pre>\n\n<p>So, what am I doing wrong?<\/p>\n","answers":[{"AnswerId":"42510771","CreationDate":"2017-02-28T14:01:01.960","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Ok this is a weird case, maybe coming from an update. Here is the way it's working :<\/p>\n\n<p>if you <code>print(model.layers[1].get_config())<\/code> :<\/p>\n\n<pre><code>{'trainable': True, 'name': 'activation_1', 'activation': 'relu'}\n<\/code><\/pre>\n\n<p>if you <code>print(model.get_config()[1])<\/code> : <\/p>\n\n<pre><code>{'config': {'trainable': True, 'name': 'activation_1', 'activation': 'relu'}, 'class_name': 'Activation'}\n<\/code><\/pre>\n\n<p>So the <code>model.get_config()<\/code> is the one containing a list of dictionnaries that <code>layer_from_config()<\/code> will accept. <\/p>\n\n<p>Instead of getting the list of layers and then its config which is in a \"bad\" format, you have to get the model config which is a list of layers config with the right format.<\/p>\n\n<p>Their doc isn't up to date it seems. Either they should adapt it, or they should adapt the code of <code>layer.get_config()<\/code>. <\/p>\n\n<p>Anyway, you can use it now :)<\/p>\n"}]}
{"QuestionId":42509462,"AnswerCount":1,"Tags":"<r><tensorflow><rstudio>","CreationDate":"2017-02-28T13:02:01.273","AcceptedAnswerId":null,"OwnerUserId":2640926.0,"Title":"tensorflow in R got \"invalid version specification\"","Body":"<p>I'm a newbie R-lang and want to execute tensorflow in mac(OS Sierra 10.12.2) and RStudio(v1.0.136)<\/p>\n\n<p>According to the <a href=\"https:\/\/rstudio.github.io\/tensorflow\/using_tensorflow_api.html#tensor_shapes\" rel=\"nofollow noreferrer\">manual<\/a>, I wrote below code.<\/p>\n\n<p>code:<\/p>\n\n<pre><\/pre>\n\n<p>but I got this error:<\/p>\n\n<pre><\/pre>\n\n<p>To confirm env,  got below code.<\/p>\n\n<pre><\/pre>\n\n<p>but python in console<\/p>\n\n<pre><\/pre>\n\n<p>if I execute:<\/p>\n\n<pre><\/pre>\n\n<p>got<\/p>\n\n<pre><\/pre>\n\n<p>Does anyone have solutions to remove error in R?<\/p>\n","answers":[{"AnswerId":"59062161","CreationDate":"2019-11-27T02:30:39.143","ParentId":null,"OwnerUserId":"6874836","Title":null,"Body":"<p>I have circumvented the <em>invalid version specification<\/em> error by following these steps:<\/p>\n\n<ol>\n<li><p>started tensorboard from conda tensorflow environment (make sure you have tensorboard in this environment):<\/p>\n\n<p><code>conda activate tensorflow<\/code><\/p><\/li>\n<\/ol>\n\n<p>then entered tensorboard command<\/p>\n\n<pre><code>  tensorboard --logdir 'logs'\n<\/code><\/pre>\n\n<p>As result it showed me that tensorboard would be available at <em>localhost:6006<\/em>:<\/p>\n\n<p><code>Tensorboard 2.0.0 at http:\/\/localhost:6006<\/code><\/p>\n\n<ol start=\"2\">\n<li><p>entered the address in web browser and the tensorboard opened empty<\/p><\/li>\n<li><p>then, run keras session on R Studio having set training history with<\/p>\n\n<p><code>callbacks = callback_tensorboard('logs\/run_a')<\/code><\/p><\/li>\n<\/ol>\n\n<p>and, while the training epochs were still running, I refreshed tensorboard in web browser which forced the loading of new data from the <em>'logs'<\/em> directory. This way I got the training history.<\/p>\n"}]}
{"QuestionId":42509811,"AnswerCount":2,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-02-28T13:17:37.057","AcceptedAnswerId":null,"OwnerUserId":6744309.0,"Title":"gentlely way to read tfrecords data into batches","Body":"<p><strong>all the follow question are based on tensorflow 1.0 API<\/strong><\/p>\n\n<p>I am now be able to write images under directory which named by class name, and this is I generate tfrecords codes:<\/p>\n\n<pre><\/pre>\n\n<p>this is the main method, so here I stored height, widht, channels(this value con't readout) etc.<\/p>\n\n<p>And I am able read tfrecords out, this is my code:<\/p>\n\n<pre><\/pre>\n\n<p>All goes fair, however, the problem occurs <strong>when I try load tfrecords data into batches and feed it into network<\/strong><\/p>\n\n<p>Here is all the code I load batches:<\/p>\n\n<pre><\/pre>\n\n<p>I run this program, got this error:<\/p>\n\n<pre><\/pre>\n\n<p><strong>apparently, program didn't got tfrecords file at all.<\/strong><\/p>\n\n<p>I have tried this:\n1. I thought it maybe  not right, I change it into both relative path and absolute path, either works;\n2. I places tfrecords file just beside script, and write tfrecords file name directly didn't work.<\/p>\n\n<p>So, basicly, I got this problem:<\/p>\n\n<p><strong>1. What's the officially and reasonable way to write a as short as possiable program load tfrecords file into batches and feed into network<\/strong><\/p>\n\n<p><strong>2. BTW, what's the simplest and elegantest way to write tensorflow layer? slim is a good choice, original way is ugly and complicated!<\/strong><\/p>\n","answers":[{"AnswerId":"42526772","CreationDate":"2017-03-01T08:31:37.037","ParentId":null,"OwnerUserId":"6744309","Title":null,"Body":"<p>For anyone maybe occured same questions, I made some mistakes in above code.\nSimply do not using <code>decode_raw<\/code>, using <code>tf.image.decode_jpeg<\/code> instead, and my code function<\/p>\n\n<pre><code>def inputs(train, batch_size, num_epochs):\n    if not num_epochs:\n        num_epochs = None\n    filenames = ['.\/data\/tiny_5_tfrecords\/train-00000-of-00002',\n                 '.\/data\/tiny_5_tfrecords\/train-00001-of-00002']\n    print(filenames)\n    with tf.name_scope('input'):\n        filename_queue = tf.train.string_input_producer(filenames, num_epochs=num_epochs)\n    print(filename_queue)\n    image, label = read_and_decode(filename_queue)\n    images, sparse_labels = tf.train.shuffle_batch(\n        [image, label],\n        batch_size=batch_size,\n        num_threads=2,\n        capacity=1000 + 3 * batch_size,\n        min_after_dequeue=1000)\n\n    return images, sparse_labels\n<\/code><\/pre>\n\n<p>I missed a tab for the last 2 lines.<\/p>\n"},{"AnswerId":"45838192","CreationDate":"2017-08-23T11:14:51.647","ParentId":null,"OwnerUserId":"4620643","Title":null,"Body":"<p>I don't know if your own answer implies what I'm writing here since I don't fully understand your own answer. However, the issue causing the ValueError is that <code>shuffle_batch<\/code> needs the image sizes to all be the same and known. Thus what would fix the issue is in the function <code>read_and_decode<\/code>, simply make <code>image_shape<\/code> have the same values for all images, for example:<br>\n<code>image_shape = tf.stack([FLAGS.target_image_height, FLAGS.target_image_width, 3])<\/code><br>\nand then using:<br>\n<code>image = tf.reshape(image, image_shape)<\/code><br>\nto make all images have the same size.<\/p>\n"}]}
{"QuestionId":42509878,"AnswerCount":3,"Tags":"<tensorflow><sampling>","CreationDate":"2017-02-28T13:20:00.453","AcceptedAnswerId":null,"OwnerUserId":7276519.0,"Title":"what is the difference between sampled_softmax_loss and nce_loss in tensorflow?","Body":"<p>i notice there are two functions about negative Sampling in tensorflow to compute the loss (<strong>sampled_softmax_loss<\/strong> and <strong>nce_loss<\/strong>). the paramaters of these two function are similar, but i really want to know what is the difference between the two?<\/p>\n","answers":[{"AnswerId":"42511026","CreationDate":"2017-02-28T14:12:57.103","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Sampled softmax tries to normalise over all samples in your output. Having a non-normal distribution (logarithmic over your labels) this is not an optimal loss function. Note that although they have the same parameters, they way you use the function is different. Take a look at the documentation here: <a href=\"https:\/\/github.com\/calebchoo\/Tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard4\/tf.nn.nce_loss.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/calebchoo\/Tensorflow\/blob\/master\/tensorflow\/g3doc\/api_docs\/python\/functions_and_classes\/shard4\/tf.nn.nce_loss.md<\/a> and read this line: <\/p>\n\n<blockquote>\n  <p>By default this uses a log-uniform (Zipfian) distribution for sampling, so your labels must be sorted in order of decreasing frequency to achieve good results. For more details, see log_uniform_candidate_sampler.<\/p>\n<\/blockquote>\n\n<p>Take a look at this paper where they explain why they use it for word embeddings: <a href=\"http:\/\/papers.nips.cc\/paper\/5165-learning-word-embeddings-efficiently-with-noise-contrastive-estimation.pdf\" rel=\"nofollow noreferrer\">http:\/\/papers.nips.cc\/paper\/5165-learning-word-embeddings-efficiently-with-noise-contrastive-estimation.pdf<\/a><\/p>\n\n<p>Hope this helps!<\/p>\n"},{"AnswerId":"50572847","CreationDate":"2018-05-28T20:03:58.620","ParentId":null,"OwnerUserId":"6934176","Title":null,"Body":"<p>Check out this documentation from TensorFlow <a href=\"https:\/\/www.tensorflow.org\/extras\/candidate_sampling.pdf\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/extras\/candidate_sampling.pdf<\/a><\/p>\n\n<p>They seem pretty similar, but sampled softmax is only applicable for a single label while NCE extends to the case where your labels are a multiset. NCE can then model the expected counts rather than presence\/absence of a label.  I'm not clear on an exact example of when to use the sampled_softmax.<\/p>\n"},{"AnswerId":"43320139","CreationDate":"2017-04-10T09:57:18.290","ParentId":null,"OwnerUserId":"5915270","Title":null,"Body":"<p>Sample softmax is all about selecting a sample of the given number and try to get the softmax loss. Here the main objective is to make the result of the sampled softmax equal to our true softmax. So algorithm basically concentrate lot on selecting the those samples from the given distribution.\nOn other hand NCE loss is more of selecting noise samples and try to mimic the true softmax. It will take only one true class and a K noise classes.  <\/p>\n"}]}
{"QuestionId":42510813,"AnswerCount":1,"Tags":"<initialization><keras><convolution>","CreationDate":"2017-02-28T14:02:36.453","AcceptedAnswerId":null,"OwnerUserId":7325313.0,"Title":"how to initialize layers by numpy array in keras","Body":"<p>I want to convert a pre-trained caffe model to keras, then i need to initialize the layers , layer by layer. \nI saved the weights and biases in a mat file and I loaded them to python workspace. \nI know \"weights\" parameter get the numpy array but not how?\nThanks <\/p>\n","answers":[{"AnswerId":"42511288","CreationDate":"2017-02-28T14:25:00.340","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>You can get more information about how to set the weight of a model in the <a href=\"https:\/\/keras.io\/layers\/about-keras-layers\/\" rel=\"nofollow noreferrer\">Keras Layers Documentation<\/a>. Basically you use :<\/p>\n\n<blockquote>\n  <p><code>layer.set_weights(weights)<\/code>: sets the weights of the layer from a list of Numpy arrays (with the same shapes as the output of <code>get_weights<\/code>).<\/p>\n<\/blockquote>\n\n<p>Or you can directly initialize them when you create the layer. Every layer has a parameter <code>weights<\/code> that you can set with a numpy array. Read <a href=\"https:\/\/keras.io\/layers\/core\/\" rel=\"nofollow noreferrer\">each layer's documentation<\/a> to feed the right weights format. For example, <code>Dense()<\/code> layers accept this format for the parameter <code>weights<\/code> :<\/p>\n\n<blockquote>\n  <p>List of Numpy arrays to set as initial weights. The list should have 2 elements, of shape (input_dim, output_dim) and (output_dim,) for weights and biases respectively. <a href=\"https:\/\/keras.io\/layers\/core\/#dense\" rel=\"nofollow noreferrer\">source<\/a><\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42512679,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-02-28T15:26:35.713","AcceptedAnswerId":null,"OwnerUserId":7585135.0,"Title":"How can I have visualized embeddings on tensorboard? Not MNIST data","Body":"<p>I'm trying to create visualized graph on Tensorboard embeddings, I'm using csv data, not MNIST data, the data in csv is like follows:<\/p>\n\n<pre><\/pre>\n\n<p>The data in first column like 0.366782506 is sample input_data x, and \"0,1\" is the one-hot label y. while 0\n\n<p>I tried to take reference on how to creat visualized graph by embedding projector on Tensorboard, but I found examples only by using MNIST data, so I'm looking for help if anyone can give a guidance on how to create a visualized embedding graph on Tensorboard.\nI can have SCALAR, GRAPH and HISTOGRAM visulized on Tensorboard with my code as following:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you ver much!<\/p>\n","answers":[]}
{"QuestionId":42512981,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><keras>","CreationDate":"2017-02-28T15:41:21.177","AcceptedAnswerId":null,"OwnerUserId":999346.0,"Title":"How to use tf.nn.top_k in a keras loss function?","Body":"<p>I have some issues making this custom loss function (it checks if  data is ordered coherently with the real ordering indices provided by ) work:<\/p>\n\n<pre><\/pre>\n\n<p>I can properly run it with sample data:<\/p>\n\n<pre><\/pre>\n\n<p>But somehow it doesn't work if I use it in , as it raises:<\/p>\n\n<pre><\/pre>\n\n<p>Note that there is no \"None\" values in my training test set if I change  to  the model compiles fine and starts training properly (but it's clearly not the loss function I want).<\/p>\n\n<p>I have the subtle feeling that there might be something wrong in using  in a loss function as I don't see how it could be differentiable, but I don't have better ideas for evaluating differences in predicted ordering. Hints\/ideas\/papers\/references? :)<\/p>\n","answers":[{"AnswerId":"42513438","CreationDate":"2017-02-28T16:01:21.503","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>This might be voted down as I won't really fix your code but here goes nothing :<\/p>\n\n<p>I don't believe, indeed, that you can use top_k as an objective function. Just like you can't use the accuracy as an objective function. <\/p>\n\n<p>The reason is mathematical. Even if keras, tensorflow, theano and co. are awesome tools for AI and allows everybody to fiddle with neural nets, the latters still remain very complex mathematical tools. Those maths are well hidden under the hood but you should be aware of them when trying to go further than prefabricated tools. <\/p>\n\n<p>What happens when you train a network is that you compute how wrong the network is on an example and you backpropagate that error to learn from it. The algorithms behind that backpropagation are optimizers, more precisely they are <strong>gradient<\/strong> based optimizers. Computing a gradient requires to differentiate the function that we are optimizing, the loss\/objective function. It means that the objective <strong>must be differentiable<\/strong>. The accuracy isn't a differentiable function, it takes as an input a real number between 0 and 1 and outputs a step-like function : 0 if x&lt;0.5 and 1 if x>0.5. That function isnt differentiable because we can't get its gradient in 0.5. The top_k function is some kind of accuracy function. So indeed in my opinion you cannot use it in an objective, because under the hood, the smart tensorflow has to compute the gradients of your function.<\/p>\n\n<p>I hope this helps :)<\/p>\n"}]}
{"QuestionId":42513183,"AnswerCount":1,"Tags":"<machine-learning><caffe><training-data>","CreationDate":"2017-02-28T15:50:09.587","AcceptedAnswerId":null,"OwnerUserId":1014196.0,"Title":"Caffe's test accuracy during validation phase being constant when training a network","Body":"<p>I wonder why my test accuracy keeps on getting a constant value of 0.5. I use CaffeNet network with only change in the fully connected layer's parameter where I configured <em>num_output: 2<\/em>. <\/p>\n\n<p>My training set contains 1000 positive and 1000 negative examples whereas my validation set has 1000 positive and 1000 negative examples as well. The dataset contains images of person (whole body RGB colored). I've defined a mean file and scale value in the data layer. My network is trained to learn a person or not (binary classifier).<\/p>\n\n<p>A snippet of my solver information looks like below:<\/p>\n\n<pre><\/pre>\n\n<p>The training output is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I also tried to change the value of <em>test_iter<\/em> to 40 (instead of previously set to 80) after following <a href=\"https:\/\/stackoverflow.com\/questions\/33780779\/caffe-solver-prototxt-values-setting-strategy\/33786620#33786620\">this link<\/a> and <a href=\"https:\/\/stackoverflow.com\/questions\/35456689\/caffe-train-network-accuracy-1-constant-accuracy-issue\">this one<\/a> if the parameter is related, but it still didn't resolve. Also, I tried to reshuffle the data by regenerating the dataset using a modified <em>create_imagenet.sh<\/em> script but the issue still remains.<\/p>\n\n<p>Every time I changed value in the solver, I always changed the fully connected layer's name as well. Is this a correct way? <\/p>\n\n<p>The number of epoch here is ~10. Is it possible culprit? Does this kind of problem fall under over-fitting issue?<\/p>\n\n<p>Any hints or suggestions are welcome.<\/p>\n\n<p>EDITED:<\/p>\n\n<p>I turned on the debug info in the solver and found the loss is infinitesimal. Can I deduce that it's not learning much or at all then? The log with the debug info is as below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42524980","CreationDate":"2017-03-01T06:45:49.347","ParentId":null,"OwnerUserId":"773226","Title":null,"Body":"<pre><code>data_layer.cpp:73] Restarting data prefetching from start.\n<\/code><\/pre>\n\n<p>The above message occurs when the .txt file that is given as input to data layer reached the end of file.<\/p>\n\n<p>This message can occur frequently when:<\/p>\n\n<ol>\n<li>You gave the wrong .txt file to data layer<\/li>\n<li>The format of the .txt file is not as expected by Caffe<\/li>\n<li>Very few number of data is present in the file.<\/li>\n<\/ol>\n"}]}
{"QuestionId":42513613,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><lstm><recurrent-neural-network>","CreationDate":"2017-02-28T16:09:03.440","AcceptedAnswerId":42518245.0,"OwnerUserId":4492495.0,"Title":"TensorFlow dynamic_rnn regressor: ValueError dimension mismatch","Body":"<p>I would like to build a toy LSTM model for regression. <a href=\"http:\/\/mourafiq.com\/2016\/05\/15\/predicting-sequences-using-rnn-in-tensorflow.html\" rel=\"noreferrer\">This<\/a> nice tutorial is already too complicated for a beginner. <\/p>\n\n<p>Given a sequence of length , predict the next value. Consider  and the sequences: <\/p>\n\n<pre><\/pre>\n\n<p>the target values should be:<\/p>\n\n<pre><\/pre>\n\n<p>I define the following model:<\/p>\n\n<pre><\/pre>\n\n<p>We instantiating the LSTM model with  I get the following:<\/p>\n\n<pre><\/pre>\n\n<p>1) Do you know what the problem is?<\/p>\n\n<p>2) Will multiplying the LSTM output by the weights yield the regression? <\/p>\n","answers":[{"AnswerId":"42518245","CreationDate":"2017-02-28T20:35:28.820","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>As discussed in the comments, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn\" rel=\"noreferrer\"><code>tf.nn.dynamic_rnn(cell, inputs, ...)<\/code><\/a> function expects a list of three-dimensional tensors<sup>*<\/sup> as its <code>inputs<\/code> argument, where the dimensions are interpreted by default as <code>batch_size<\/code> x <code>num_timesteps<\/code> x <code>num_features<\/code>. (If you pass <code>time_major=True<\/code>, they are interpreted as <code>num_timesteps<\/code> x <code>batch_size<\/code> x <code>num_features<\/code>.) Therefore the preprocessing you've done in the original placeholder is unnecessary, and you can pass the oriding <code>X<\/code> value directly to <code>tf.nn.dynamic_rnn()<\/code>.<\/p>\n\n<hr>\n\n<p><sup>*<\/sup> Technically it can accept complicated nested structures in addition to lists, but the leaf elements must be three-dimensional tensors.<sup>**<\/sup><\/p>\n\n<p><sup>**<\/sup> Investigating this turned up a bug in the implementation of <code>tf.nn.dynamic_rnn()<\/code>. In principle, it should be sufficient for the inputs to have at least two dimensions, but the <code>time_major=False<\/code> path assumes that they have exactly three dimensions when it transposes the input into the time-major form, and it was the error message that this bug inadvertently causes that showed up in your program. We're working on getting that fixed.<\/p>\n"}]}
{"QuestionId":42514191,"AnswerCount":1,"Tags":"<docker><tensorflow><bazel><syntaxnet><docker-toolbox>","CreationDate":"2017-02-28T16:39:07.547","AcceptedAnswerId":42555059.0,"OwnerUserId":7636711.0,"Title":"Exhausted Virtual Memory Installing SyntaxNet Using Docker Toolbox","Body":"<p>I exhausted my virtual memory when trying to install <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/syntaxnet\" rel=\"nofollow noreferrer\">SyntaxNet<\/a> from this <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/syntaxnet\/Dockerfile\" rel=\"nofollow noreferrer\">Dockerfile<\/a> using the Docker Toolbox. I received this message when compiling the Dockerfile:<\/p>\n\n<pre><\/pre>\n\n<p>I have a feeling this could be resolved by changing Bazel's default jobs limit with (for example) , however I'm not sure where I would put that in the Dockerfile.<\/p>\n","answers":[{"AnswerId":"42555059","CreationDate":"2017-03-02T12:05:17.800","ParentId":null,"OwnerUserId":"1097141","Title":null,"Body":"<p>There are two possibilities: You could either modify the Dockerfile so that it creates a <code>~\/.bazelrc<\/code> that contains the following text:<\/p>\n\n<pre><code>build --jobs=1\n<\/code><\/pre>\n\n<p>Note that this works, even though the Dockerfile runs <code>bazel test<\/code> (as opposed to <code>bazel build<\/code>), because <code>build<\/code> flags in the <code>.bazelrc<\/code> also apply to Bazel's <code>test<\/code> command.<\/p>\n\n<p>The other possibility would be to modify the <code>RUN<\/code> command in the Dockerfile to include the <code>--jobs=1<\/code> parameter, e.g. <code>RUN [...] &amp;&amp; bazel test --jobs=1 --genrule_strategy=standalone [...]<\/code>.<\/p>\n\n<p>Bazel should then spawn not more than a single child process during the build. You can verify this by running \"ps axuf\" on your host and looking at the process tree of your container. If you modified the <code>RUN<\/code> cmd, you should also see the <code>--jobs=1<\/code> parameter on Bazel's command-line.<\/p>\n"}]}
{"QuestionId":42514206,"AnswerCount":3,"Tags":"<tensorflow><python-multithreading>","CreationDate":"2017-02-28T16:39:46.750","AcceptedAnswerId":42648211.0,"OwnerUserId":6842947.0,"Title":"TensorFlow : Enqueuing and dequeuing a queue from multiple threads","Body":"<p>The problem I am trying to solve is as follows :\nI have a list  of filenames. I have defined a <\/p>\n\n<ul>\n<li> with its  and . <\/li>\n<li>This  is expected to be filled by  for a specified  number of times. <\/li>\n<li>A number of threads are expected to work in parallel. Each thread dequeues an element from the  and does some operations on it and enqueues it to another queue. I have got that part right.<\/li>\n<li>However once  of  have been processed and the  is empty, provided that the current epoch , the queue must again be filled up and the threads must work again.<\/li>\n<\/ul>\n\n<p>The good news is : I have got it working in a certain case (See <strong>PS<\/strong> at the end !!)<\/p>\n\n<p>The bad news is : I think that there is a better way of doing this.<\/p>\n\n<p>The method I am using to do this now is as follows (I have simplified the functions and have removed e image processing based preprocessing and subsequent enqueuing  but the heart of the processing remains the same !!) : <\/p>\n\n<pre><\/pre>\n\n<p>The work function is as follows : <\/p>\n\n<pre><\/pre>\n\n<p>So, although this works, I have a feeling that there is a better and cleaner way to achieve this. So, in a nutshell my questions are :<\/p>\n\n<ol>\n<li>Is there a simpler and cleaner way of achieving this task in TensorFlow ?<\/li>\n<li>Is there any problem with this code's logic ? I am not very experienced with multithreading scenarios, so any obvious faults which have skipped my attention would be very helpful to me.<\/li>\n<\/ol>\n\n<p>P.S : It seems that this code is not Perfect after all. When I ran with 1.2 million images and 200 threads, it ran. However when I run it for 10 images and 20 threads, it gives the following error :<\/p>\n\n<pre><\/pre>\n\n<p>I thought I got that covered by . What the hell is going on here ?<\/p>\n","answers":[{"AnswerId":"42635530","CreationDate":"2017-03-06T21:05:47.620","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>I recommend having a single thread calling enqueue_many epochs times enqueue the correct number of images. It can then close the queue. This would let you simplify your work function and other threads.<\/p>\n"},{"AnswerId":"42662788","CreationDate":"2017-03-08T04:03:19.867","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>I think the GIL will prevent any actual parallelism from being done in those threads.<\/p>\n\n<p>To get performance with tensorflow you need to keep your data in tensorflow.<\/p>\n\n<p>Tensor Flow's <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">reading data guide<\/a> explains how to address a very similar sort of problem. <\/p>\n\n<p>More specifically, you seem to have rewritten a significant chunk of <code>string_input_producer<\/code>.<\/p>\n"},{"AnswerId":"42648211","CreationDate":"2017-03-07T12:23:04.273","ParentId":null,"OwnerUserId":"6842947","Title":null,"Body":"<p>I finally found out the answer. The problem was that multiple threads were clashing together on various points in the <code>work()<\/code> function.\nThe following <code>work()<\/code> function works perfectly.<\/p>\n\n<pre><code>def work(coord, val, sess, epoch, maxepochs, incrementepoch, supplyimg, q, lock, close_op):\n    print('I am thread number %s'%(threading.current_thread().name))\n    print('I can see a queue with size %d'%(sess.run(q)))\n    while not coord.should_stop():\n        lock.acquire()\n        if sess.run(q) &gt; 0:\n            filename, currepoch = sess.run([val, epoch])\n            filename = filename.decode(encoding='UTF-8')\n            tid = threading.current_thread().name\n            print(filename + ' ' + str(currepoch) + ' thread ' + str(tid))\n        elif sess.run(epoch) &lt; maxepochs:\n            print('Thread %s has acquired the lock'%(threading.current_thread().name))\n            print(\"The previous epoch = %d\"%(sess.run(epoch)))\n            sess.run([incrementepoch, supplyimg])\n            sz = sess.run(q)\n            print(\"The new epoch = %d\"%(sess.run(epoch)))\n            print(\"The new queue size = %d\"%(sz))\n    else:\n            coord.request_stop()\n        lock.release()\n\n    return None\n<\/code><\/pre>\n"}]}
{"QuestionId":42514960,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><keras><autoencoder>","CreationDate":"2017-02-28T17:17:17.487","AcceptedAnswerId":42519480.0,"OwnerUserId":7606144.0,"Title":"Keras autoencoder","Body":"<p>I've worked a long time ago with neural networks in Java and now I'm trying to learn to use TFLearn and Keras in Python. <\/p>\n\n<p>I'm trying to build an autoencoder, but as I'm experiencing problems the code I show you hasn't got the bottleneck characteristic (this should make the problem even easier).<\/p>\n\n<p>On the following code I create the network, the dataset (two random variables), and after train it plots the correlation between each predicted variable with its input.<\/p>\n\n<p>What the network should learn, is to output the same input that receives.<\/p>\n\n<pre><\/pre>\n\n<p>Even though some times the result is acceptable, many others isn't, I know neural networks have weight random initialization and therefore it may converge to different solutions, but I think this is too much and there may be some mistake in my code.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/N36IW.png\" rel=\"nofollow noreferrer\">Sometimes correlation is acceptable<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/GEZrl.png\" rel=\"nofollow noreferrer\">Others is quite lost<\/a><\/p>\n\n<p>**<\/p>\n\n<h2>UPDATE:<\/h2>\n\n<p>**<\/p>\n\n<p>Thanks Marcin Mo\u017cejko!<\/p>\n\n<p>Indeed that was the problem, my original question was because I was trying to build an autoencoder, so to be coherent with the title here comes an example of autoencoder (just making a more complex dataset and changing the activation functions):<\/p>\n\n<pre><\/pre>\n\n<p>For this example I used TanH activation function, but I tried with others and worked aswell. \nThe dataset has now 6 variables but the autoencoder has a bottleneck of 2 neurons; as long as variables 2 to 5 are formed combining variables 0 and 1, the autoencoder only needs to pass the information of those two and learn the functions to generate the other variables on the decoding phase.\nThe example above shows how all functions are learnt but one, the division... I don't know yet why.<\/p>\n","answers":[{"AnswerId":"42519480","CreationDate":"2017-02-28T21:55:44.190","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>I think that your case is relatively easy to explain why your network might fail to learn an identity function. Let's go through your example:<\/p>\n\n<ol>\n<li>Your input comes from <code>2d<\/code> space - and it doesn't lie on a <code>1d<\/code> or <code>0d<\/code> submanifold - due to uniform distribiution. From this it's easy to see that in order to get an identity function from your autoencoder <strong>every<\/strong> layer should be able to represent a function which range is at least <strong>two dimensional<\/strong>, beacuse the output of your last layer should also lie on a <code>2d<\/code> manifold. <\/li>\n<li><p>Let's go through your network and check if it satisfy the condtion need:<\/p>\n\n<pre><code>inputLayer = Input(shape=(2,))\nautoencoder = Dense(4, activation='relu')(inputLayer)\nautoencoder = Dense(4, activation='relu')(autoencoder)\nautoencoder = Dense(2, activation='relu')(autoencoder) # Possible problems here\n<\/code><\/pre>\n\n<p>You may see that the bottleneck might cause a problem - for this layer it might be hard to satisfy the condition from the first point. For this layer - in order to get the 2-dimensional output range you need to have weights which will make all the examples not falling into saturation region of <code>relu<\/code> (in this case all this samples will be squashed to <code>0<\/code> in one of the units - what makes impossible for range to be \"fully\" <code>2d<\/code>). So basically - the probability that this will not happen is relatively small. Also the probability that backpropagation will not move this unit to this region also cannot be neglected.<\/p><\/li>\n<\/ol>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>In a comment the question was asked why optimizer fail to prevent or undo the saturation. It's an example of a one of the important <code>relu<\/code> downsides - once an example falls into a <code>relu<\/code> saturation region - this example doesn't directly take part in learning of a given unit. It could affect it by influencing previous units - but due to <code>0<\/code> derivative - this influence is not direct. So basically <em>unsaturating an example<\/em> comes from a side effect - not the direct action of an optimizer.<\/p>\n"},{"AnswerId":"42518810","CreationDate":"2017-02-28T21:11:54.803","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>This is a really cool usecase to see and study the difficulties of training a Neural Net. <\/p>\n\n<p>Indeed I see multiple possibilities : <\/p>\n\n<p>1) If there are very different results between 2 different runs, it can come from the initialization. But it can also come from your data set which isn't the same at every run. <\/p>\n\n<p>2) Something that will make it difficult for the network to learn the correlation is your activations, more precisely the sigmoid. I would change that non linearity to another 'relu'. No reason to use a Sigmoid here, I know it looks linear around 0 but it isnt really linear. To produce a 0.7, the raw output will have to be quite high. The \"easy\" relation that you have in mind for this network isnt so easy since you constrain it. <\/p>\n\n<p>3) if it is sometimes not good, maybe it needs more epochs to converge? <\/p>\n\n<p>4) maybe you'd need a bigger dataset? In theory you're good since there is 900 samples for &lt; 200 parameters in your network but who knows...<\/p>\n\n<p>I can't try all of this since I'm on my phone right now but feel free to try and troubleshout with the hints I gave you :)  I hope this helps. <\/p>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>After some trials, as Marcin Mo\u017cejko was saying, the issue comes from the activations. You can read his answer to have more info on what is going wrong. The way to fix it is to change your activations. If you are a fan of <code>'relu'<\/code>, you can use a special version of this activation. For this, use a <a href=\"https:\/\/keras.io\/layers\/advanced-activations\/#leakyrelu\" rel=\"nofollow noreferrer\">LeakyRelu layer<\/a> and do not set an activation to the previous layer, like this :<\/p>\n\n<pre><code>    autoencoder = Dense(inputs*2)(inputLayer)\n    autoencoder = LeakyReLU(alpha=0.3)(autoencoder)\n<\/code><\/pre>\n\n<p>This will solve the case where you get stuck in a nonoptimal solution.\nOtherwise, as I said above, you can try not to use any non-linearities. Your loss will go down way faster and doesn't get stuck. <\/p>\n\n<p>Keep in mind that the non linearities were introduced to get the networks to find more complex patterns in the data. In your case you have the simplest linear pattern.<\/p>\n"}]}
{"QuestionId":42514986,"AnswerCount":1,"Tags":"<python><nlp><keras><word2vec><word-embedding>","CreationDate":"2017-02-28T17:18:24.420","AcceptedAnswerId":42518294.0,"OwnerUserId":5236675.0,"Title":"Product merge layers with Keras functionnal API for Word2Vec model","Body":"<p>I am trying to implement a Word2Vec CBOW with negative sampling with Keras, following the code found <a href=\"https:\/\/github.com\/abaheti95\/Deep-Learning\/blob\/master\/word2vec\/keras\/cbow_model.py\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<pre><\/pre>\n\n<p>However, I get an  error during the merge part because Embedding layer is a 3D tensor while cbow is only 2 dimensions. I assume I need to reshape the embedding (which is [?, 1, 100]) to [1, 100] but I can't find how to reshape with the functional API.\nI am using the Tensorflow backend.<\/p>\n\n<p>Also, if someone can point to an other implementation of CBOW with Keras (Gensim free), I would love to have a look to it!<\/p>\n\n<p>Thank you!<\/p>\n\n<p>EDIT: Here is the error<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42518294","CreationDate":"2017-02-28T20:38:59.100","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<pre><code>ValueError: Shape must be rank 2 but is rank 3 for 'MatMul' (op: 'MatMul') with input shapes: [?,1,100], [?,100].\n<\/code><\/pre>\n\n<p>Indeed you need to reshape the <code>word_embedding<\/code> tensor. Two ways to do it :<\/p>\n\n<ul>\n<li><p>Either you use the <code>Reshape()<\/code> layer, imported from <code>keras.layers.core<\/code>, this is done like :<\/p>\n\n<pre><code>word_embedding = Reshape((100,))(word_embedding)\n<\/code><\/pre>\n\n<p>the argument of <code>Reshape<\/code> is a tuple with the target shape.<\/p><\/li>\n<li><p>Or you can use <code>Flatten()<\/code> layer, also imported from <code>keras.layers.core<\/code>, used like this :<\/p>\n\n<pre><code>word_embedding = Flatten()(word_embedding)\n<\/code><\/pre>\n\n<p>taking nothing as an argument, it will just remove \"empty\" dimensions.<\/p><\/li>\n<\/ul>\n\n<p>Does this help you? <\/p>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>Indeed the second <code>merge()<\/code> is a bit more tricky. The <code>dot<\/code> merge in Keras only accepts tensors of the same rank, so same <code>len(shape)<\/code>.\nSo what you will do is use a <code>Reshape()<\/code> layer to add back that 1 empty dimension, then use the feature <code>dot_axes<\/code> instead of <code>concat_axis<\/code> which is not relevant for a <code>dot<\/code> merge.\nThis is what I propose you for the solution :<\/p>\n\n<pre><code>word_embedding = shared_embedding_layer(word_index)\n# Shape output = (None,1,emb_size)\ncontext_embeddings = shared_embedding_layer(context)\n# Shape output = (None, 2*window_size, emb_size)\nnegative_words_embedding = shared_embedding_layer(negative_samples)\n# Shape output = (None, negative, emb_size)\n\n# Now the context words are averaged to get the CBOW vector\ncbow = Lambda(lambda x: K.mean(x, axis=1),\n                     output_shape=(EMBEDDING_DIM,))(context_embeddings)\n# Shape output = (None, emb_size)\ncbow = Reshape((1,emb_size))(cbow)\n# Shape output = (None, 1, emb_size)\n\n# Context is multiplied (dot product) with current word and negative\n# sampled words\nword_context_product = merge([word_embedding, cbow], mode='dot')\n# Shape output = (None, 1, 1)\nword_context_product = Flatten()(word_context_product)\n# Shape output = (None,1)\nnegative_context_product = merge([negative_words_embedding, cbow], mode='dot',dot_axes=[2,2])\n# Shape output = (None, negative, 1)\nnegative_context_product = Flatten()(negative_context_product)\n# Shape output = (None, negative)\n<\/code><\/pre>\n\n<p>Is it working? :)<\/p>\n\n<p>The problem comes from the rigidity of TF regarding the matrix multiplication. Merge with \"dot\" mode calls the backend <code>batch_dot()<\/code> function and, as opposed to Theano, TensorFlow requires the matrix to have the same rank : <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/matmul\" rel=\"nofollow noreferrer\">read here<\/a>. <\/p>\n"}]}
{"QuestionId":42515641,"AnswerCount":0,"Tags":"<memory><tensorflow><tegra>","CreationDate":"2017-02-28T17:55:35.807","AcceptedAnswerId":null,"OwnerUserId":582967.0,"Title":"DMA between CPU and GPU in TensorFlow","Body":"<p>I am running TensorFlow on NVidia Jetson TX1 and faced memory shortage when I train large network such as GoogleNet.<\/p>\n\n<p>CPU and GPU in TX1 do not have separate memory and they share one memory. However, It seems that TensorFlow is trying to allocate separate memory space and copy from CPU side to GPU side. Thus it requests 2x memory than it really needs. <\/p>\n\n<p>In my opinion, this situation can be handled by something like DMA access between CPU and GPU. As far as I know, TensorFlow utilizes DMA between GPUs (not sure which one handles this. TensorFlow? or GPU driver?). Can I use DMA between CPU and GPU also in TensorFlow? or any other suggestions?<\/p>\n\n<p>EDIT: I just found that there is <em><a href=\"http:\/\/arrayfire.com\/zero-copy-on-tegra-k1\/\" rel=\"noreferrer\">Zero Copy<\/a><\/em> feature in CUDA which is what I exactly wanted. However, can I use this feature in TensorFlow?<\/p>\n","answers":[]}
{"QuestionId":42516237,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><caffe><conv-neural-network><gradient-descent>","CreationDate":"2017-02-28T18:30:33.747","AcceptedAnswerId":null,"OwnerUserId":2995747.0,"Title":"How does changing batch size results in different prediction time?","Body":"<p>I trained a data set(~8000 images) using <em>Caffe<\/em> and a batch size of 5 with <em>Alex net<\/em> network. This results in a prediction time of (800-900)ms. Then i changed the batch size to 56(maximum my machine can support) and the prediction time reduced to (200-300)ms on <em>cpu<\/em>. <\/p>\n\n<p>I can understand changing batch size using <em>stochastic gradient descent<\/em> can decrease training time and I know for <em>Alex net<\/em> I should be using batch size of 256, but I am using 56 because of my low configuration machine.  <\/p>\n\n<p>But how the batch size is affecting the prediction time on a single test data?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42516447","CreationDate":"2017-02-28T18:44:16.890","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>If those are also your prediction batch sizes, then the speed increase is merely parallelization of the scoring process.  However, if the speed is measured properly with identical batch sizes during prediction ...<\/p>\n\n<p>This depends a lot on your hardware and any short-cuts involved.  Since you haven't displayed the models you trained, it's hard to tell.  One hypothesis I have is that your second model managed to eliminate more of the trained parameters from affecting the final decision (i.e. weight = 0.0), and that your software optimizations or hardware short-cuts subsequently sped up the computations.<\/p>\n\n<p>Another possibility is that the larger model is actually richer, such that compiling the model results in using on-chip matrix operations rather than individual sparse-matrix operations (which could be slower, if you got unlucky).  I doubt that this is the case.<\/p>\n"}]}
{"QuestionId":42516484,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-02-28T18:46:55.203","AcceptedAnswerId":42545427.0,"OwnerUserId":3775565.0,"Title":"Why does my TensorFlow Neural Network for XOR only have an accuracy of around 0.5?","Body":"<p>I Wrote a Neural Network in TensorFlow for the XOR input. I have used 1 hidden layer with 2 units and softmax classification. The input is of the form &lt;1, x_1, x_2, zero, one> , where <\/p>\n\n<ul>\n<li>1 is the bias<\/li>\n<li>x_1 and x_2 are either between 0 and 1 for all the combination {00, 01, 10, 11}. Selected to be normally distributed around 0 or 1<\/li>\n<li>zero: is 1 if the output is zero<\/li>\n<li>one: is 1 if the output is one<\/li>\n<\/ul>\n\n<p>The accuracy is always around 0.5. What has gone wrong? Is the architecture of the neural network wrong, or is there something with the code?<\/p>\n\n<pre><\/pre>\n\n<p><strong>Check the comments section for the updated code<\/strong><\/p>\n","answers":[{"AnswerId":"42545427","CreationDate":"2017-03-02T01:50:43.433","ParentId":null,"OwnerUserId":"3775565","Title":null,"Body":"<p>The problem was with the randomly assigned weights. <a href=\"http:\/\/pastebin.com\/4NZdx3bD\" rel=\"nofollow noreferrer\">Here<\/a> is the modified version, obtained after a series of trail-and-error.<\/p>\n"}]}
{"QuestionId":42516680,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-02-28T18:58:15.047","AcceptedAnswerId":null,"OwnerUserId":7637346.0,"Title":"I can't solve it about Tensorflow, please let me know what is wrong","Body":"<p>I can't solve it about Tensorflow, please let me know what is wrong...<\/p>\n\n<p>while I studying basic tensorflow, I got a problem. but it's so weired...<\/p>\n\n<p>other guys can run it but only me, I can't run it because of error<\/p>\n\n<p>there are some code I learned and error message.  Please let me know what is wrong and how can i pix it.<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>TypeError: Expected float32, got {1, 2, 3} of type 'set' instead.<\/p>\n","answers":[{"AnswerId":"42519392","CreationDate":"2017-02-28T21:49:36.213","ParentId":null,"OwnerUserId":"2505209","Title":null,"Body":"<p>x,y should be of type array\/list not set. Use following, <\/p>\n\n<pre><code>x = [1, 2, 3]\ny = [1, 2, 3]\n<\/code><\/pre>\n\n<p>Hope this helps.<\/p>\n"}]}
{"QuestionId":42517083,"AnswerCount":1,"Tags":"<python><tensorflow><training-data>","CreationDate":"2017-02-28T19:22:05.393","AcceptedAnswerId":42519236.0,"OwnerUserId":6347839.0,"Title":"Basic Tensorflow Example - Prediction of a Line","Body":"<p>I'm trying to create this super simple example with Tensorflow and I clearly don't fully understand the API for Tensorflow. <\/p>\n\n<p>I have the following code. It's not mine originally - I found it from some demo, but I can't remember where I found it, or else I would give the author credit. Apologies. <\/p>\n\n<h3>Saving the Trained Line Model<\/h3>\n\n<pre><\/pre>\n\n<p>Ok that's all fine. I just want to load in the model and then query my model to get a predicted value. Here is my attempted code:<\/p>\n\n<h3>Loading and Querying the Trained Line Model<\/h3>\n\n<pre><\/pre>\n\n<p>This is a really basic question, but how can I just pass in a value and use my line almost like a function. Do I need to change the way the line model is being constructed? My guess is that the computation graph is not set up where the output is an actual variable that we can get. Is this correct? If so, how should I modify this program?<\/p>\n","answers":[{"AnswerId":"42519236","CreationDate":"2017-02-28T21:39:10.310","ParentId":null,"OwnerUserId":"2505209","Title":null,"Body":"<p>You have to create tensorflow graph again and load saved weights into it. I added couple of lines to your code and it gives desired outputs. Please check it. <\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\nsess = tf.Session() \nnew_saver = tf.train.import_meta_graph('linemodel.meta')\nnew_saver.restore(sess, tf.train.latest_checkpoint('.\/')) # latest checkpoint\nall_vars = tf.global_variables()\n\n# load saved weights into new variables\nW = all_vars[0]\nb = all_vars[1]\n\n# build TF graph\nx = tf.placeholder(tf.float32)\ny = tf.add(tf.multiply(W,x),b)\n\n# Session\ninit = tf.global_variables_initializer()\nprint(sess.run(all_vars))\nsess.run(init)    \nfor i in range(2):\n    x_ip = np.random.rand(10).astype(np.float32) # batch_size : 10\n    vals = sess.run(y,feed_dict={x:x_ip})\n    print vals\n<\/code><\/pre>\n\n<p>Output:<\/p>\n\n<pre><code>[array([ 0.1000001], dtype=float32), array([ 0.29999995], dtype=float32)]\n\n[-0.21707924 -0.18646611 -0.00732027 -0.14248954 -0.54388255 -0.33952206  -0.34291503 -0.54771954 -0.60995424 -0.91694558]\n[-0.45050886 -0.01207681 -0.38950539 -0.25888413 -0.0103816  -0.10003483 -0.04783082 -0.83299863 -0.53189355 -0.56571382]\n<\/code><\/pre>\n\n<p>I hope this helps.<\/p>\n"}]}
{"QuestionId":42517396,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2017-02-28T19:41:14.950","AcceptedAnswerId":null,"OwnerUserId":665023.0,"Title":"Tensorflow debugger cannot work on windows 10","Body":"<p>My code is fairly simple:<\/p>\n\n<pre><\/pre>\n\n<p>But when I ran it:<\/p>\n\n<pre><\/pre>\n\n<p>I searched it online and found \"_curses\" is not supported on Windows 10. Does that mean I cannot use Tensorflow Debugger on Windows 10?<\/p>\n","answers":[{"AnswerId":"43021241","CreationDate":"2017-03-25T19:54:48.520","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>TensorFlow Debugger (tfdbg) is not fully functional in version 1.0.x. TensorFlow 1.1.0rc0 has just came out. Please try it out:<\/p>\n\n<pre><code>pip.exe install --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.1.0rc0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>If there are still issues, please report them at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues<\/a><\/p>\n\n<p>Also, it's perhaps worth pointing out the following:<\/p>\n\n<ol>\n<li>The command-line interface of tfdbg is different between Windows and Linux\/Mac. Windows does not have official <code>curses<\/code> support, so it uses the <code>readline<\/code> interface.<\/li>\n<li><p>Because of 1., readline is a required dependency of tfdbg on Windows:<\/p>\n\n<p>pip.exe install pyreadline<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"43356124","CreationDate":"2017-04-11T21:01:16.027","ParentId":null,"OwnerUserId":"1780735","Title":null,"Body":"<p>According to: <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/debugger\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/debugger<\/a>\nYou have to write in cmd the following:<\/p>\n\n<pre><code>pip install pyreadline\n<\/code><\/pre>\n"},{"AnswerId":"43681806","CreationDate":"2017-04-28T13:47:58.667","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>I just verified TensorFlow 1.1.0 on a Windows Server 2016 Machine. tfdbg works. Here are the thing that I did, step by step<\/p>\n\n<ol>\n<li>Install Anaconda 3 with python version 3.5<\/li>\n<li>Open a command window as administrator.<\/li>\n<li>\"C:\\Program Files\\Anaconda3\\Scripts\\pip.exe\" install pyreadline<\/li>\n<li>\"C:\\Program Files\\Anaconda3\\Scripts\\pip.exe\" install  --ignore-installed --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.1.0-cp35-cp35m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.1.0-cp35-cp35m-win_amd64.whl<\/a><\/li>\n<li>Open Python shell with \"C:\\Program Files\\Anaconda3\\python.exe\"<\/li>\n<li>In the Python shell, enter the following simple tensorflow code to launch a TFDBG terminal.<\/li>\n<\/ol>\n\n<blockquote>\n<pre><code>import tensorflow as tf\nfrom tensorflow.python import debug as tf_debug\na = tf.Variable(42.0, name=\"a\")\nsess = tf.Session()\nsess = tf_debug.LocalCLIDebugWrapperSession(sess)\nsess.run(a.initializer)\n<\/code><\/pre>\n<\/blockquote>\n"},{"AnswerId":"42518301","CreationDate":"2017-02-28T20:39:18.573","ParentId":null,"OwnerUserId":"6232024","Title":null,"Body":"<p>Not an answer but cut and pasted the code listed above and got a different error from ui_factory.py as follows:<\/p>\n\n<pre><code>    ValueError: Exhausted all fallback ui_types.\n<\/code><\/pre>\n\n<p>Running Python 3.5.3, numpy 1.12.0, and TensorFlow 1.0.0 on Windows 10.<\/p>\n"}]}
{"QuestionId":42517452,"AnswerCount":2,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-02-28T19:44:50.710","AcceptedAnswerId":42518670.0,"OwnerUserId":1868775.0,"Title":"TensorFlow - Split and Squeeze","Body":"<p>I'm new to TensorFlow and I'm formatting some data to feed into a Recurrent Neural Network. My data is given by a 3d tensor fed into the placeholder . I want to split  along the 3rd dimension, and for this I have (note that  corresponds to the length of  along the 3rd dimension):<\/p>\n\n<pre><\/pre>\n\n<p>Although, as I have tried with :<\/p>\n\n<pre><\/pre>\n\n<p>If the  is a 3d  then  will return a list of  arrays with dimension 3, such that the 3rd dimension is singleton. With  I know that I can easily solve this using  together with a list comprehension to remove the singleton dimension:<\/p>\n\n<pre><\/pre>\n\n<p>But how can I do the same on TF?<\/p>\n","answers":[{"AnswerId":"42518670","CreationDate":"2017-02-28T21:02:43.520","ParentId":null,"OwnerUserId":"3734670","Title":null,"Body":"<p>Try to use the squeeze function of Tensorflow (tf.squeeze) and the scan function of Tensorflow (tf.scan) instead of a list comprehension.    <\/p>\n\n<pre><code>tf.scan(lambda a, x_i: tf.squeeze(x_i, [2]), x, initializer=tf.constant(0, shape=[n_dim0, n_dim1]))\n<\/code><\/pre>\n"},{"AnswerId":"46159837","CreationDate":"2017-09-11T16:00:58.167","ParentId":null,"OwnerUserId":"420385","Title":null,"Body":"<p>You may be looking for the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/unstack\" rel=\"nofollow noreferrer\"><code>tf.unstack<\/code><\/a> op:<\/p>\n\n<pre><code>x = tf.unstack(x, axis=2)\n<\/code><\/pre>\n"}]}
{"QuestionId":42517926,"AnswerCount":1,"Tags":"<tensorflow><pruning>","CreationDate":"2017-02-28T20:14:53.517","AcceptedAnswerId":null,"OwnerUserId":6232024.0,"Title":"How to freeze\/lock weights of one TensorFlow variable (e.g., one CNN kernel of one layer)","Body":"<p>I have a TensorFlow CNN model that is performing well and we would like to implement this model in hardware; i.e., an FPGA. It's a relatively small network but it would be ideal if it were smaller. With that goal, I've examined the kernels and find that there are some where the weights are quite strong and there are others that aren't doing much at all (the kernel values are all close to zero). This occurs specifically in layer 2, corresponding to the tf.Variable() named, \"W_conv2\". W_conv2 has shape [3, 3, 32, 32]. I would like to freeze\/lock the values of W_conv2[:, :, 29, 13] and set them to zero so that the rest of the network can be trained to compensate. Setting the values of this kernel to zero effectively removes\/prunes the kernel from the hardware implementation thus achieving the goal stated above.<\/p>\n\n<p>I have found similar questions with suggestions that generally revolve around one of two approaches; <\/p>\n\n<p>Suggestion #1: <\/p>\n\n<pre><\/pre>\n\n<p>Implementing this suggestion freezes the entire variable. I want to freeze just a slice, specifically W_conv2[:, :, 29, 13].<\/p>\n\n<p>Suggestion #2:<\/p>\n\n<pre><\/pre>\n\n<p>Again, implementing this suggestion does not allow the use of slices. For instance, if I try the inverse of my stated goal (optimize only a single kernel of a single variable) as follows:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Slicing tf.Variables() isn't possible in the way that I've tried it here. The only thing that I've tried which comes close to doing what I want is using .assign() but this is extremely inefficient, cumbersome, and caveman-like as I've implemented it as follows (after the model is trained):<\/p>\n\n<pre><\/pre>\n\n<p>The model was started in Keras then moved to TensorFlow since Keras didn't seem to have a mechanism to achieve the desired results. I'm starting to think that TensorFlow doesn't allow for pruning but find this hard to believe; it just needs the correct implementation.<\/p>\n","answers":[{"AnswerId":"51926528","CreationDate":"2018-08-20T08:10:53.103","ParentId":null,"OwnerUserId":"3782865","Title":null,"Body":"<p>A possible approach is to initialize these specific weights with zeros, and modify the minimization process such that gradients won't be applied to them. It can be done by replacing the call to <code>minimize()<\/code> with something like:<\/p>\n\n<pre><code>W_conv2_weights = np.ones((3, 3, 32, 32))\nW_conv2_weights[:, :, 29, 13] = 0\nW_conv2_weights_const = tf.constant(W_conv2_weights)\n\noptimizer = tf.train.RMSPropOptimizer(0.001)\n\nW_conv2_orig_grads = tf.gradients(loss, W_conv2)\nW_conv2_grads = tf.multiply(W_conv2_weights_const, W_conv2_orig_grads)\nW_conv2_train_op = optimizer.apply_gradients(zip(W_conv2_grads, W_conv2))\n\nrest_grads = tf.gradients(loss, rest_of_vars)\nrest_train_op = optimizer.apply_gradients(zip(rest_grads, rest_of_vars))\n\ntf.group([rest_train_op, W_conv2_train_op])\n<\/code><\/pre>\n\n<p>I.e, <\/p>\n\n<ol>\n<li>Preparing a constant Tensor for canceling the appropriate gradients<\/li>\n<li>Compute gradients only for <code>W_conv2<\/code>, then multiply element-wise with the constant <code>W_conv2_weights<\/code> to zero the appropriate gradients and only then apply gradients.<\/li>\n<li>Compute and apply gradients \"normally\" to the rest of the variables.<\/li>\n<li>Group the 2 train ops to a single training op.<\/li>\n<\/ol>\n"}]}
{"QuestionId":42518411,"AnswerCount":0,"Tags":"<python><tensorflow><conv-neural-network>","CreationDate":"2017-02-28T20:45:20.560","AcceptedAnswerId":null,"OwnerUserId":3681744.0,"Title":"TensorFlow, Multi-Phase Training","Body":"<p>I want to have a multi-layer ConvNet such that at every phase I only train one layer of it. My code sketch looks like the following:<\/p>\n\n<pre><\/pre>\n\n<p>The problem is that this code is not working and I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>How can I implement such function which implements a multi-layer ConvNet with the ability to get the output of each layer in different training phases, in TensorFlow?<\/p>\n","answers":[]}
{"QuestionId":42518744,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-02-28T21:07:58.150","AcceptedAnswerId":42533220.0,"OwnerUserId":5102599.0,"Title":"Making a list and appending to it in TensorFlow","Body":"<p>I am new to TensorFlow. I'm not able to understand how to create a dynamic \"pythonic\" list in TensorFlow. Basically, I perform some computation on a tensor object () and append it to a \"list\" , which I want to be a tensor with shape <\/p>\n\n<p>I want to do something like this:<\/p>\n\n<pre><\/pre>\n\n<p>I want  to be a Tensor with shape , basically a column vector which is a tensor object. If I run the code above, I instead get a python list of TensorObjects.<\/p>\n","answers":[{"AnswerId":"42519051","CreationDate":"2017-02-28T21:27:10.820","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>I think the real answer here is that you have to change your way of thinking a bit. In Tensorflow you first define a computational graph, and then evaluate it with values as input. <\/p>\n\n<p>For your example: if X is indeed a Tensor with shape (100,1), tf.log(X) would work best. \nIf you want to create a dynamic list like your code suggests, it would be better to create this list BEFORE passing it to the graph. <\/p>\n\n<p>Let me know if this helps!<\/p>\n"},{"AnswerId":"42533220","CreationDate":"2017-03-01T13:35:59.173","ParentId":null,"OwnerUserId":"5587428","Title":null,"Body":"<p>If you want to convert <code>X<\/code> to a (100,) tensor you could add <code>X = tf.stack(X)<\/code> after your for loop:<\/p>\n\n<pre><code>X = []\nfor i in range(100):\n  q = tf.log(train_data[i]) \n  print(q)    #Tensor(\"Log:0\", shape=(), dtype=float32) \n  X.append(q)\nX = tf.stack(X)\n<\/code><\/pre>\n\n<p>This is a useful construct where you may want to <code>tf.unstack<\/code> some tensor, loop over the resulting list, and then  use <code>tf.stack<\/code> to get back to a single tensor. <\/p>\n"}]}
{"QuestionId":42518883,"AnswerCount":1,"Tags":"<python><csv><tensorflow>","CreationDate":"2017-02-28T21:16:27.763","AcceptedAnswerId":42519141.0,"OwnerUserId":3451457.0,"Title":"Generalized Tensorflow CSV Load","Body":"<p>I am trying to load a large CSV file into a TensorFlow pipeline where each row represents a a flattened 256*256 image. I am trying to generalize the TensorFlow CSV files read example but am having trouble generalizing the decode_csv function to accomadate all of my columns:<\/p>\n\n<pre><\/pre>\n\n<p>However I am getting the error<\/p>\n\n<pre><\/pre>\n\n<p>How do I generalize the col1,...colN section?<\/p>\n","answers":[{"AnswerId":"42519141","CreationDate":"2017-02-28T21:33:19.263","ParentId":null,"OwnerUserId":"3451457","Title":null,"Body":"<p>Silly mistake on my part. tf.pack can take a list argument so there is no need to to separate output of tf.decode_csv into different cols. It can simply be done with <\/p>\n\n<pre><code>col = tf.decode_csv(csv_row, record_defaults=record_defaults)\nfeatures = tf.pack(col)\n<\/code><\/pre>\n"}]}
{"QuestionId":42519010,"AnswerCount":1,"Tags":"<python><tensorflow><tensorflow-serving>","CreationDate":"2017-02-28T21:25:09.847","AcceptedAnswerId":42520221.0,"OwnerUserId":5434466.0,"Title":"How to do batching in Tensorflow Serving?","Body":"<p>Deployed Tensorflow Serving and ran test for Inception-V3. Works fine.<\/p>\n\n<p>Now, would like to do batching for serving for Inception-V3. \nE.g. would like to send 10 images for prediction instead of one.<\/p>\n\n<p>How to do that? Which files to update (inception_saved_model.py or inception_client.py)? What those update look like? and how are the images passed to the serving -is it passed as a folder containing images or how?<\/p>\n\n<p>Appreciate some insight into this issue. Any code snippet related to this will be extremely helpful.<\/p>\n\n<p>=================================<\/p>\n\n<p>Updated inception_client.py<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42520221","CreationDate":"2017-02-28T22:51:20.170","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You should be able to compute predictions for a batch of images with a small change to the request construction code in <a href=\"https:\/\/github.com\/tensorflow\/serving\/blob\/f100a35aff0e9050a93371877d88f478db1569fe\/tensorflow_serving\/example\/inception_client.py\" rel=\"noreferrer\"><code>inception_client.py<\/code><\/a>. The following lines in that file create a request with a \"batch\" containing a single image (note <code>shape=[1]<\/code>, which means \"a vector of length 1\"):<\/p>\n\n<pre><code>with open(FLAGS.image, 'rb') as f:\n  # See prediction_service.proto for gRPC request\/response details.\n  data = f.read()\n  request = predict_pb2.PredictRequest()\n  request.model_spec.name = 'inception'\n  request.model_spec.signature_name = 'predict_images'\n  request.inputs['images'].CopyFrom(\n      tf.contrib.util.make_tensor_proto(data, shape=[1]))\n  result = stub.Predict(request, 10.0)  # 10 secs timeout\n  print(result)\n<\/code><\/pre>\n\n<p>You can pass more images in the same vector to run predictions on a batch of data. For example, if <code>FLAGS.image<\/code> were a comma-separated list of filenames:<\/p>\n\n<pre><code>request = predict_pb2.PredictRequest()\nrequest.model_spec.name = 'inception'\nrequest.model_spec.signature_name = 'predict_images'\n\n# Build a batch of images.\nimage_data = []\nfor image in FLAGS.image.split(','):\n  with open(image, 'rb') as f:\n    image_data.append(f.read())\n\nrequest.inputs['images'].CopyFrom(\n    tf.contrib.util.make_tensor_proto(image_data, shape=[len(image_data)]))\n\nresult = stub.Predict(request, 10.0)  # 10 secs timeout\nprint(result)\n\n if __name__ == '__main__':\n   tf.app.run()\n<\/code><\/pre>\n"}]}
{"QuestionId":42520104,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><theano><keras>","CreationDate":"2017-02-28T22:41:35.507","AcceptedAnswerId":42520312.0,"OwnerUserId":5896419.0,"Title":"Keras\/Theano VGG16 AttributeError: 'Model' object has no attribute 'ndim'","Body":"<p>I am trying to build a Keras model with a Theano backend that connects a time-distributed VGG16 network to an LSTM layer, then finally to a series of dense layers. However, I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Here is the section of the code I use to build the model:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42520312","CreationDate":"2017-02-28T22:57:35.310","ParentId":null,"OwnerUserId":"5896419","Title":null,"Body":"<p>The solution was to use cnn_base.output as an input to the GlobalAveragePooling2D layer:<\/p>\n\n<pre><code>cnn_out = GlobalAveragePooling2D()(cnn_base.output)\n<\/code><\/pre>\n"}]}
{"QuestionId":42520418,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-02-28T23:06:25.330","AcceptedAnswerId":42521116.0,"OwnerUserId":1868775.0,"Title":"How to multiply list of tensors by single tensor on TensorFlow?","Body":"<p>I am implementing an RNN and contrarily to the examples I have found which minimize only the cost for the output in the last step<\/p>\n\n<pre><\/pre>\n\n<p>However, the object  is a list of  with  elements, so the  throws the error <\/p>\n\n<blockquote>\n  <p>ValueError: Shape must be rank 2 but is rank 3 for 'MatMul' (op:\n  'MatMul') with input shapes: [100,128,16], [16,1].<\/p>\n<\/blockquote>\n\n<p>. How can I do this multiplication?<\/p>\n","answers":[{"AnswerId":"42521116","CreationDate":"2017-03-01T00:21:48.223","ParentId":null,"OwnerUserId":"1868775","Title":null,"Body":"<p>The solution is to <code>tf.stack<\/code> the list of tensors into a 3d tensor and then use <code>tf.map_fn<\/code> to apply the multiplication operation on each 2d tensor along dimension 0:<\/p>\n\n<pre><code>    # Transform the list into a 3D tensor with dimensions (n_timesteps, batch_size, N_HIDDEN)\n    outputs = tf.stack(outputs)\n\n    def pred_fn(current_output):\n        return tf.matmul(current_output, weights['out']) + biases['out']\n    # Use tf.map_fn to apply pred_fn to each tensor in outputs, along dimension 0 (timestep dimension)\n    pred = tf.map_fn(pred_fn, outputs)\n<\/code><\/pre>\n"}]}
{"QuestionId":42520805,"AnswerCount":1,"Tags":"<android><tensorflow><raspberry-pi3><android-things>","CreationDate":"2017-02-28T23:47:08.340","AcceptedAnswerId":null,"OwnerUserId":1207906.0,"Title":"How to run TensorFlow Inference on Android Things for example classifying images?","Body":"<p>Is TensorFlow supported on Android Things? Can TensorFlow Android samples be ported to run on Android Things? If so, what\u2019s the simplest way to do it?<\/p>\n","answers":[{"AnswerId":"42520959","CreationDate":"2017-03-01T00:03:48.243","ParentId":null,"OwnerUserId":"1207906","Title":null,"Body":"<p>The short answer: Yes, you can indeed run TensorFlow on embedded devices running Android Things such as Raspberry Pi 3.  We have ported the TensorFlow image classification app to Android Things. It is available at: <a href=\"https:\/\/github.com\/androidthings\/sample-tensorflow-imageclassifier\" rel=\"nofollow noreferrer\">https:\/\/github.com\/androidthings\/sample-tensorflow-imageclassifier<\/a>.<\/p>\n\n<p>The long answer: In order to run TensorFlow you first need to build model graph and run training against large input dataset. Once that\u2019s done, you run inference based on the trained model to make intelligent decisions.  The first part in training is resources hungry and intensive, typically done beforehand while the inference part is relatively lightweight. It is the inference part that has been ported to run on devices running Android Things.<\/p>\n\n<p>The following summarizes what you need to do when integrating TensorFlow into your Android Things project:<\/p>\n\n<ul>\n<li><p>First, add the TensorFlow Android Inference library file into your project: Tensorflow-Android-Inference-alpha-debug.aar<\/p><\/li>\n<li><p>Next, add a dependency in your app\u2019s build.gradle file like this: \ncompile(name: 'Tensorflow-Android-Inference-alpha-debug', ext: 'aar')<\/p><\/li>\n<\/ul>\n\n<p>To utilize the TensorFlow Android Inference library, you instantiate an org.tensorflow.contrib.android.TensorFlowInferenceInterface instance in order to expose the following methods:<\/p>\n\n<ul>\n<li>initializeTensorFlow: initialize TensorFlow object using a model graph as input<\/li>\n<li>fillNodeFloat: copy input data into TensorFlow input array<\/li>\n<li>runInference: run inference and save results in TensorFlow output array<\/li>\n<li>readNodeFloat: read from TensorFlow output array and save into your own array<\/li>\n<\/ul>\n\n<p>Check out the implementation in TensorFlowImageClassifier.java of the image classifier sample to see how you can pass inputs to TensorFlow, run inference, and then extract output labels from TensorFlow. For example in our image classification sample, our app can detect what breeds of dogs when shown an image of a dog.<\/p>\n\n<p>The model graph in the sample is built using Google Inception V3 TensorFlow model with a training set of 1.2 million images from ImageNet.  If you like to build your own model graph, be sure to update the model file, the label file, and input\/output names accordingly in your classifier implementation.<\/p>\n\n<p>In short the approach outlined above using TensorFlow Android Inference library as a gradle dependency offers a quick and easy way to add TensorFlow core functionalities into any Android projects, even complex projects such as <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/android\" rel=\"nofollow noreferrer\">the TensorFlow Android samples<\/a> that have native code and require NDK integration.  This approach also extends host platforms to those that currently do not have full Bazel support like Windows.<\/p>\n"}]}
{"QuestionId":42521005,"AnswerCount":2,"Tags":"<keras><batch-normalization>","CreationDate":"2017-03-01T00:09:45.373","AcceptedAnswerId":42524528.0,"OwnerUserId":5352399.0,"Title":"How the number of parameters associated with BatchNormalization layer is 2048?","Body":"<p>I have the following code.<\/p>\n\n<pre><\/pre>\n\n<p>And this is the model summary:<\/p>\n\n<pre><\/pre>\n\n<p>The size of the input for the BatchNormalization (BN) layer is 512. According to <a href=\"https:\/\/keras.io\/layers\/normalization\/\" rel=\"noreferrer\">Keras documentation<\/a>, shape of the output for BN layer is same as input which is 512.<\/p>\n\n<p>Then how the number of parameters associated with BN layer is 2048?<\/p>\n","answers":[{"AnswerId":"42524528","CreationDate":"2017-03-01T06:15:52.517","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>The batch normalization in Keras implements <a href=\"http:\/\/jmlr.org\/proceedings\/papers\/v37\/ioffe15.pdf\" rel=\"noreferrer\">this paper<\/a>. <\/p>\n\n<p>As you can read there, in order to make the batch normalization work during training, they need to keep track of the distributions of each normalized dimensions. To do so, since you are in <code>mode=0<\/code>by default, they compute 4 parameters per feature on the previous layer. Those parameters are making sure that you properly propagate and backpropagate the information. <\/p>\n\n<p>So <code>4*512 = 2048<\/code>, this should answer your question. <\/p>\n"},{"AnswerId":"45420579","CreationDate":"2017-07-31T16:04:54.613","ParentId":null,"OwnerUserId":"3517598","Title":null,"Body":"<p>These 2048 parameters are in fact <code>[gamma weights, beta weights, moving_mean(non-trainable), moving_variance(non-trainable)]<\/code>, each having 512 elements (the size of the input layer).<\/p>\n"}]}
{"QuestionId":42521166,"AnswerCount":1,"Tags":"<c++><tensorflow><protocol-buffers>","CreationDate":"2017-03-01T00:26:58.013","AcceptedAnswerId":42678809.0,"OwnerUserId":3002273.0,"Title":"Tensorflow Operation Documentation","Body":"<p>In the Tensorflow <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/ops\/array_ops.cc#L2519\" rel=\"nofollow noreferrer\">source<\/a>, I see,<\/p>\n\n<pre><\/pre>\n\n<p>In Python, I can do the following,<\/p>\n\n<pre><\/pre>\n\n<p>Note that I am getting in Python the Protobuf definition (I removed some for brevity) of the TF operation. I want to get the documentation section of the definition that I see in the C++ code. How do I get it?<\/p>\n","answers":[{"AnswerId":"42678809","CreationDate":"2017-03-08T18:11:11.017","ParentId":null,"OwnerUserId":"3002273","Title":null,"Body":"<p>That was painful. You need to patch TF and Protobuf<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8207\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8207<\/a>\n<a href=\"https:\/\/github.com\/google\/protobuf\/issues\/2798\" rel=\"nofollow noreferrer\">https:\/\/github.com\/google\/protobuf\/issues\/2798<\/a><\/p>\n\n<p>And then you need to additionally comment out the line,\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a3e636c0f561e2ac6d9f8a0044fbe09acb003803\/tensorflow\/python\/framework\/python_op_gen.cc#L708\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/a3e636c0f561e2ac6d9f8a0044fbe09acb003803\/tensorflow\/python\/framework\/python_op_gen.cc#L708<\/a><\/p>\n\n<p>Rebuilding and running gave,<\/p>\n\n<pre><code>bazel build --config=opt \/\/tensorflow\/tools\/pip_package:build_pip_package &amp;&amp; \\\nbazel-bin\/tensorflow\/tools\/pip_package\/build_pip_package \/tmp\/tensorflow_pkg &amp;&amp; \\\nsudo pip install --upgrade \/tmp\/tensorflow_pkg\/tensorflow-1.0.0*.whl\n<\/code><\/pre>\n\n<p>And executing my test,<\/p>\n\n<pre><code>$ python -c \"from tensorflow.python.ops import gen_array_ops; print gen_array_ops._InitOpDefLibrary()._ops['BroadcastGradientArgs'].op_def\"\nname: \"BroadcastGradientArgs\"\ninput_arg {\n  name: \"s0\"\n  type_attr: \"T\"\n}\ninput_arg {\n  name: \"s1\"\n  type_attr: \"T\"\n}\noutput_arg {\n  name: \"r0\"\n  type_attr: \"T\"\n}\noutput_arg {\n  name: \"r1\"\n  type_attr: \"T\"\n}\nattr {\n  name: \"T\"\n  type: \"type\"\n  default_value {\n    type: DT_INT32\n  }\n  allowed_values {\n    list {\n      type: DT_INT32\n      type: DT_INT64\n    }\n  }\n}\nsummary: \"Return the reduction indices for computing gradients of s0 op s1 with broadcast.\"\ndescription: \"This is typically used by gradient computations for a broadcasting operation.\"\n<\/code><\/pre>\n\n<p>That took way too much Regex ...<\/p>\n"}]}
{"QuestionId":42521400,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><cross-entropy>","CreationDate":"2017-03-01T00:58:22.963","AcceptedAnswerId":42521820.0,"OwnerUserId":7638401.0,"Title":"Calculating Cross Entropy in TensorFlow","Body":"<p>I am having a hard time with calculating cross entropy in tensorflow. In particular, I am using the function:<\/p>\n\n<pre><\/pre>\n\n<p>Using what is seemingly simple code, I can only get it to return a zero<\/p>\n\n<pre><\/pre>\n\n<p>returns<\/p>\n\n<pre><\/pre>\n\n<p>My understanding of cross entropy is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Where p(x) is the true probability of event x and q(x) is the predicted probability of event x. <\/p>\n\n<p>There if input any two numbers for p(x) and q(x) are used such that <\/p>\n\n<pre><\/pre>\n\n<p>there should be a nonzero cross entropy. I am expecting that I am using tensorflow incorrectly. Thanks in advance for any help.<\/p>\n","answers":[{"AnswerId":"42521880","CreationDate":"2017-03-01T01:58:26.930","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>In addition to Don's answer (+1), <a href=\"https:\/\/stackoverflow.com\/a\/36086477\/395857\">this answer written by mrry<\/a> may interest you, as it gives the formula to calculate the cross entropy in TensorFlow:<\/p>\n\n<blockquote>\n  <p>An alternative way to write:<\/p>\n\n<pre><code>xent = tf.nn.softmax_cross_entropy_with_logits(logits, labels)\n<\/code><\/pre>\n  \n  <p>...would be:<\/p>\n\n<pre><code>softmax = tf.nn.softmax(logits)\nxent = -tf.reduce_sum(labels * tf.log(softmax), 1)\n<\/code><\/pre>\n  \n  <p>However, this alternative would be (i) less numerically stable (since\n  the softmax may compute much larger values) and (ii) less efficient\n  (since some redundant computation would happen in the backprop). For\n  real uses, we recommend that you use\n  <code>tf.nn.softmax_cross_entropy_with_logits()<\/code>.<\/p>\n<\/blockquote>\n"},{"AnswerId":"42521820","CreationDate":"2017-03-01T01:49:38.460","ParentId":null,"OwnerUserId":"49329","Title":null,"Body":"<p>Like they say, you can't spell \"softmax_cross_entropy_with_logits\" without \"softmax\". Softmax of <code>[0.45]<\/code> is <code>[1]<\/code>, and <code>log(1)<\/code> is <code>0<\/code>.<\/p>\n\n<blockquote>\n  <p>Measures the probability error in discrete classification tasks in which the\n  classes are mutually exclusive (each entry is in exactly one class).  For\n  example, each CIFAR-10 image is labeled with one and only one label: an image\n  can be a dog or a truck, but not both.<\/p>\n  \n  <p><strong>NOTE:<\/strong>  While the classes are mutually exclusive, their probabilities\n  need not be.  All that is required is that each row of <code>labels<\/code> is\n  a valid probability distribution.  If they are not, the computation of the\n  gradient will be incorrect.<\/p>\n  \n  <p>If using exclusive <code>labels<\/code> (wherein one and only\n  one class is true at a time), see <code>sparse_softmax_cross_entropy_with_logits<\/code>.<\/p>\n  \n  <p><strong>WARNING:<\/strong> This op expects unscaled logits, since it performs a <code>softmax<\/code>\n  on <code>logits<\/code> internally for efficiency.  Do not call this op with the\n  output of <code>softmax<\/code>, as it will produce incorrect results.<\/p>\n  \n  <p><code>logits<\/code> and <code>labels<\/code> must have the same shape <code>[batch_size, num_classes]<\/code>\n  and the same dtype (either <code>float16<\/code>, <code>float32<\/code>, or <code>float64<\/code>).<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42522006,"AnswerCount":2,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-03-01T02:15:10.180","AcceptedAnswerId":42528371.0,"OwnerUserId":3718030.0,"Title":"How to input cifar10 into inceptionv3 in keras","Body":"<p>I am trying to classify CIFAR10 images using pre-trained imagenet weights for the Inception v3. I am using the following code. <\/p>\n\n<pre><\/pre>\n\n<p>But it gives me an error like 'Negative dimension' at an intermediate conv layer.<\/p>\n\n<p>This does not happen when I use VGG16 network.<\/p>\n\n<p>I am using keras with tensorflow backend and tf dim ordernig.<\/p>\n","answers":[{"AnswerId":"42528371","CreationDate":"2017-03-01T09:47:54.107","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>In <a href=\"https:\/\/keras.io\/applications\/#inceptionv3\" rel=\"nofollow noreferrer\">documentation<\/a> of this layer one may found that minimal shape of the input is <code>(150, 150, 3)<\/code> (with <code>tf<\/code> dim ordering). Your input is much smaller. This minimal size comes from multiple <code>pooling<\/code> and <code>valid<\/code> border mode which makes the output of each layers smaller - and if it is smaller than a certain size - it's impossible to perform neither pooling nor convolution.<\/p>\n"},{"AnswerId":"42528359","CreationDate":"2017-03-01T09:47:23.297","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Inception network is trained on 224x224 sized images and their down sampling path goes down to something below 10x10. Therefore for 32,32,3 images the downsampling leads to negative dimension sizes. Now you can do multiple things. First you could resize every image in the cifar10 dataset to 224x224 and pass this tensor into the inception model. You could remove some downsampling filters of the network. Then it would still work. Third you could do Zero padding to increase the image size without changing the resolution.<\/p>\n"}]}
{"QuestionId":42522663,"AnswerCount":2,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2017-03-01T03:31:27.783","AcceptedAnswerId":42523167.0,"OwnerUserId":7614637.0,"Title":"install Tensorflow 1.0 on windows by Anaconda error","Body":"<p>By following the instructions on <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">Tensorflow website<\/a>, I tried to install cpu-only version Tensorflow by Anaconda. However, the 4th step <\/p>\n\n<pre><\/pre>\n\n<p>outputs an error: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42522887","CreationDate":"2017-03-01T03:54:27.997","ParentId":null,"OwnerUserId":"4741406","Title":null,"Body":"<p>I think your problems could be one of this:<\/p>\n\n<ol>\n<li>your python version is not Python3.5<\/li>\n<li>Your system is not x64 windows.<\/li>\n<\/ol>\n\n<p>So you can change install method below.<\/p>\n\n<p>You can install <code>tensorflow<\/code> with <code>pip<\/code> direct  like this:<\/p>\n\n<pre><code>pip install tensorflow\n<\/code><\/pre>\n\n<p>but this works only pip version > 9,\nif you're using pip version under 9, you have to upgrade pip first.<\/p>\n\n<p>you can upgrade your pip with this:<\/p>\n\n<pre><code>pip install -U pip\n<\/code><\/pre>\n\n<blockquote>\n  <p>if you've installed python for all user, you might run this code with administrator account.<\/p>\n<\/blockquote>\n"},{"AnswerId":"42523167","CreationDate":"2017-03-01T04:24:39.357","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The command on the website is currently wrong. Replace <code>x86_64<\/code> at the end of the URL with <code>amd64<\/code> to get the correct command: <\/p>\n\n<pre><code>pip install --ignore-installed --upgrade https:\/\/storage.googleapis.com\/tensorflow\/windows\/cpu\/tensorflow-1.0.0-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>Alternatively you can use the following simpler command to install the latest CPU version:<\/p>\n\n<pre><code>pip install --ignore-installed --upgrade tensorflow\n<\/code><\/pre>\n"}]}
{"QuestionId":42523189,"AnswerCount":1,"Tags":"<numpy><tensorflow>","CreationDate":"2017-03-01T04:27:33.773","AcceptedAnswerId":42523798.0,"OwnerUserId":364772.0,"Title":"What is the best way to check if two Tensor shapes are the same including None?","Body":"<p>I'd like to check if two tensors have the same shape.<\/p>\n\n<p>Suppose I have some tensors like this:<\/p>\n\n<pre><\/pre>\n\n<p>I added . However, this fails, perhaps due to the None. Indeed  = , and also  is . They look same to me.<\/p>\n\n<p>If there is no None, it works OK.<\/p>\n\n<pre><\/pre>\n\n<p>How can I ignore None in the shape checking?<\/p>\n\n<p>In summary:<\/p>\n\n<pre><\/pre>\n\n<p>I want to get the assertion in line 7 to be True.<\/p>\n","answers":[{"AnswerId":"42523798","CreationDate":"2017-03-01T05:20:53.113","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can use <code>a.shape.as_list() == b.shape.as_list()<\/code> to compare two <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/TensorShape\" rel=\"nofollow noreferrer\"><code>tf.TensorShape<\/code><\/a> objects for \"equality\". However, you should be careful when doing this, because if the two shapes contain <code>None<\/code> in the same location then the tensors with those shapes are not guaranteed to have the same size in that dimension.<\/p>\n\n<p>(It would be useful to be able to represent \"symbolic\" dimensions like <code>batch_size<\/code> in a <code>tf.TensorShape<\/code>, which would make equality testing more useful. We're looking at extensions to the API to allow this in future versions of TensorFlow.)<\/p>\n"}]}
{"QuestionId":42524886,"AnswerCount":2,"Tags":"<tensorflow><autoencoder>","CreationDate":"2017-03-01T06:40:38.437","AcceptedAnswerId":null,"OwnerUserId":7039446.0,"Title":"the weight of encoder do not change when training autoencoder using tensorflow","Body":"<p>The coder implementing autoencoder is shown as following:<\/p>\n\n<pre><\/pre>\n\n<p>When I print the encoder, decoder weight and the loss.\nThe decoder and loss weight changes when training but the encoder weight remain the same as shown as following and I don't know why. Somebody help. <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"45361985","CreationDate":"2017-07-27T22:24:37.467","ParentId":null,"OwnerUserId":"7886651","Title":null,"Body":"<p>The weights always behave in that manner. That is, they always have gaussian distribution. Note that your input could follow any distribution in high dimensions. In addition, it seems that if you different types of distribution you end up having a Gaussian distribution (This is from probability theory). As a result, the distribution of the weights will somehow generalize and will continue following a Gaussian distribution. Note also that, with <code>Batch Normalization<\/code> the aim is to force the output of the <code>activation functions<\/code> to follow a <strong>Gaussian distribution<\/strong>. This is the general intuition. <\/p>\n\n<p>In addition, <code>l2_regulization<\/code> is in some way forcing the weights to follow a <code>Gaussian Distribution<\/code>. <\/p>\n\n<p>Finally, printing the <code>weights<\/code> as you are doing is wrong. Rather you should use <code>Tensorboard<\/code>.  <\/p>\n\n<p>Hope this answer helps.<\/p>\n"},{"AnswerId":"42635645","CreationDate":"2017-03-06T21:11:52.027","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>In general I recommend inspecting the graph on Tensorboard to make sure that it looks like you expect it to (for example, that there are gradient updates for the encoder weights.<\/p>\n\n<p>In your case it could be that <code>encoder_w[0, 0]<\/code> doesn't change much because the its gradients happen to be small, and so is the learning rate.<\/p>\n"}]}
{"QuestionId":42525226,"AnswerCount":1,"Tags":"<android><machine-learning><tensorflow>","CreationDate":"2017-03-01T07:01:26.920","AcceptedAnswerId":null,"OwnerUserId":7639646.0,"Title":"Can TensorFlow Models be Updated on the fly for Android Apps","Body":"<p>We have a requirement where models need to be updated on the flow . We could use some API to get the latest models . Is updating models possible when we use TensorFlow for android . If not are there any suggested alternatives ?<\/p>\n","answers":[{"AnswerId":"42525788","CreationDate":"2017-03-01T07:35:42.823","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow models are represented by <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/d90f60a7a445c59fbc26ba47e53873ecb2022950\/tensorflow\/core\/framework\/graph.proto\" rel=\"nofollow noreferrer\"><code>GraphDef<\/code><\/a> protocol buffers (assuming you have <a href=\"https:\/\/www.tensorflow.org\/extend\/tool_developers\/#freezing\" rel=\"nofollow noreferrer\">frozen<\/a> the weights into the graph; otherwise, they typically require a separate checkpoint file). Therefore you could fetch an updated model via an RPC interface\u2014such as <a href=\"http:\/\/www.grpc.io\/\" rel=\"nofollow noreferrer\">gRPC<\/a>, which handles protocol buffers as standard\u2014and use it to create a new TensorFlow session in your Android app.<\/p>\n"}]}
{"QuestionId":42525797,"AnswerCount":1,"Tags":"<tensorflow><lstm>","CreationDate":"2017-03-01T07:36:14.737","AcceptedAnswerId":42525985.0,"OwnerUserId":6079200.0,"Title":"How to change the activation function in LSTM cell in Tensorflow","Body":"<p>I am trying to change the activation function in the LSTM cell from the new 1.0 release of Tensorflow but am having difficulty.<\/p>\n\n<p>There is tf.contrib.rnn.LSTMcell which the API states should allow for changing functions but it does not seem to be implemented yet for this cell.<\/p>\n\n<p>Furthermore, tf.contrib.rnn.BasicLSTMCell, which also should allow for different activation functions doesn't seem to exist anymore.<\/p>\n\n<p>Do I just need to wait or is there another solution? <\/p>\n","answers":[{"AnswerId":"42525985","CreationDate":"2017-03-01T07:48:31.290","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>When you instantiate both <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\/LSTMCell\" rel=\"nofollow noreferrer\"><code>tf.contrib.rnn.LSTMcell<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\/BasicLSTMCell\" rel=\"nofollow noreferrer\"><code>tf.contrib.rnn.BasicLSTMCell<\/code><\/a> you can pass the activation function as the <code>activation<\/code> parameter. If you look at the linked documentation, you'll see, for example, that the constructor's signature for <code>BasicLSTMCell<\/code> is <\/p>\n\n<p><code>__init__(num_units, forget_bias=1.0, input_size=None, state_is_tuple=True, activation=tf.tanh)<\/code><\/p>\n"}]}
{"QuestionId":42525844,"AnswerCount":1,"Tags":"<lua><neural-network><deep-learning><torch><nvidia-digits>","CreationDate":"2017-03-01T07:39:49.210","AcceptedAnswerId":42595600.0,"OwnerUserId":5285417.0,"Title":"Classification with torch model exported from digits - lua 5.1","Body":"<p>I'm very new to deep learning and i'm trying to obtain a classification with lua.<\/p>\n\n<p>I've installed digits with torch and lua 5.1 and i've train the following model :<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vDegn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vDegn.png\" alt=\"image of the model\"><\/a><\/p>\n\n<p>After that, i've made a classification with the digits server to test the exemple and here is the result : \n<a href=\"https:\/\/i.stack.imgur.com\/bHcT8.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bHcT8.png\" alt=\"result of classification\"><\/a><\/p>\n\n<p>I've exported the model and now i'm trying to do a classification with the following lua code : <\/p>\n\n<pre><\/pre>\n\n<p>But here is the output :<\/p>\n\n<pre><\/pre>\n\n<p>And this is actually not the classification provided by digits...<\/p>\n","answers":[{"AnswerId":"42595600","CreationDate":"2017-03-04T11:43:27.380","ParentId":null,"OwnerUserId":"5285417","Title":null,"Body":"<p>OK, after searching in the digits code source, it looked like i've missed two things :<\/p>\n\n<ul>\n<li><p>you have to get  the mean image in the job folder and make the following pre-process : <\/p>\n\n<p>print '==> Preprocessing'\nfor i=1,im_mean:size(1) do\n    im[i]:csub(im_mean[i])\nend<\/p><\/li>\n<li><p>and the fact that i had to load my images in this way and multiply every pixel to 255.<\/p><\/li>\n<\/ul>\n\n<p>local im = image.load(image_url):type('torch.FloatTensor'):contiguous();\n    im:mul(255)<\/p>\n\n<p>Here is the total anwser : <\/p>\n\n<pre><code>require 'image'\nrequire 'nn'\nrequire 'torch'\nrequire 'paths'\n\nlocal function main()\n\nprint '==&gt; Downloading image and network'\nlocal image_url = '\/home\/delpech\/mnist\/test\/7\/03079.png'\nlocal network_url = '\/home\/delpech\/models\/snapshot_30_Model.t7'\nlocal mean_url = '\/home\/delpech\/models\/mean.jpg'\n\nprint '==&gt; Loading network'\nlocal net = torch.load(network_url)\nnet:evaluate();\n\nprint '==&gt; Loading synsets'\nprint 'Loads mapping from net outputs to human readable labels'\nlocal synset_words = {}\nfor line in io.lines'\/home\/delpech\/models\/labels.txt' do table.insert(synset_words, line) end\n\nprint '==&gt; Loading image and imagenet mean'\nlocal im = image.load(image_url):type('torch.FloatTensor'):contiguous();--:contiguous()\nim:mul(255)\nlocal I = image.scale(im,28,28,'bilinear'):float()\n\n\nlocal im_mean =  image.load(mean_url):type('torch.FloatTensor'):contiguous();\nim_mean:mul(255)\nlocal Imean = image.scale(im,28,28,'bilinear'):float()\n\nprint '==&gt; Preprocessing'\nfor i=1,im_mean:size(1) do\n    im[i]:csub(im_mean[i])\nend\n\nlocal _,classes = net:forward(im):sort(true);\nfor i=1,10 do\n  print('predicted class '..tostring(i)..': ', synset_words[classes[i]])\nend\n\nend \n\n\nmain()\n<\/code><\/pre>\n"}]}
{"QuestionId":42527048,"AnswerCount":2,"Tags":"<python-2.7><tensorflow>","CreationDate":"2017-03-01T08:46:13.713","AcceptedAnswerId":null,"OwnerUserId":7602053.0,"Title":"ValueError: Tensor must be from the same graph as Tensor","Body":"<p>I am trying to bulid a graph in tensorflow, but encountering the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: Tensor(transformation_0\/output\/output: 0\", shape=(),\n  dtype=float32 ) must be from the same graph as\n  Tensor(\"variables\/total_output: 0\", shape=(), dtype=float32_ref)<\/p>\n<\/blockquote>\n\n<p>Here is the code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43276669","CreationDate":"2017-04-07T11:08:59.317","ParentId":null,"OwnerUserId":"7832279","Title":null,"Body":"<p>have you tried:<\/p>\n\n<p>1) changing<\/p>\n\n<pre><code>graph =tf.Graph() \nwith graph.as_default()\n<\/code><\/pre>\n\n<p>for:<\/p>\n\n<pre><code>with tf.Session() as sess:\n<\/code><\/pre>\n\n<p>2) And removing:<\/p>\n\n<pre><code>sess = tf.Session(graph=graph)\n<\/code><\/pre>\n\n<p>I was having the same error and those changes solve it.<\/p>\n"},{"AnswerId":"46629307","CreationDate":"2017-10-08T08:59:28.610","ParentId":null,"OwnerUserId":"8739349","Title":null,"Body":"<p>try this,delete shape=[None] <\/p>\n\n<pre><code>a = tf.placeholder(tf.float32, name=\"input_placeholder_a\")\n<\/code><\/pre>\n"}]}
{"QuestionId":42527492,"AnswerCount":3,"Tags":"<tensorflow><gpu><cudnn>","CreationDate":"2017-03-01T09:07:33.843","AcceptedAnswerId":42532586.0,"OwnerUserId":3747801.0,"Title":"TensorFlow: How to verify that it is running on GPU","Body":"<p>I am looking for a simple way of verifying that my  graphs are actually running on the GPU. <\/p>\n\n<p>PS. It would also be nice to verify that the  library is used.<\/p>\n","answers":[{"AnswerId":"42532586","CreationDate":"2017-03-01T13:05:57.333","ParentId":null,"OwnerUserId":"7615953","Title":null,"Body":"<p>There are several ways to view op placement.<\/p>\n\n<ol>\n<li><p>Add RunOptions and RunMetadata to the session call and view the placement of ops and computations in Tensorboard. See code here: <a href=\"https:\/\/www.tensorflow.org\/get_started\/graph_viz\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/get_started\/graph_viz<\/a><\/p><\/li>\n<li><p>Specify the log_device_placement option in a session ConfigProto. This logs to console which device the operations are placed on. <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/ConfigProto\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/ConfigProto<\/a><\/p><\/li>\n<li><p>View GPU usage in the terminal using nvidia-smi.<\/p><\/li>\n<\/ol>\n"},{"AnswerId":"42547017","CreationDate":"2017-03-02T04:41:48.603","ParentId":null,"OwnerUserId":"2505209","Title":null,"Body":"<p>When you import TF in Python <\/p>\n\n<pre><code>import tensorflow as tf\n<\/code><\/pre>\n\n<p>You will get these logs which indicate usage of CUDA libraries<\/p>\n\n<pre><code>I tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcublas.so.8.0 locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcudnn.so.5 locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcufft.so.8.0 locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcuda.so.1 locally\nI tensorflow\/stream_executor\/dso_loader.cc:135] successfully opened CUDA library libcurand.so.8.0 locally\n<\/code><\/pre>\n\n<p>Also, when you build a graph and run a session with log_device_placement in Config Proto, you will get these logs (shows it found a GPU device):<\/p>\n\n<pre><code>I tensorflow\/stream_executor\/cuda\/cuda_gpu_executor.cc:910] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\nI tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:885] Found device 0 with properties: \nname: GeForce GTX 1060 6GB\nmajor: 6 minor: 1 memoryClockRate (GHz) 1.759\npciBusID 0000:01:00.0\nTotal memory: 5.93GiB\nFree memory: 4.94GiB\nI tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:906] DMA: 0 \nI tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:916] 0:   Y \nI tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:975] Creating TensorFlow device (\/gpu:0) -&gt; (device: 0, name: GeForce GTX 1060 6GB, pci bus id: 0000:01:00.0)\n<\/code><\/pre>\n"},{"AnswerId":"47593782","CreationDate":"2017-12-01T12:49:38.013","ParentId":null,"OwnerUserId":"133374","Title":null,"Body":"<p>There is a related <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2502\" rel=\"nofollow noreferrer\">TensorFlow upstream issue<\/a>. Basically it says that the Python API does not expose such information yet.<\/p>\n\n<p>The C++ API however does. E.g. there is <code>tensorflow::KernelsRegisteredForOp()<\/code>. I wrote a small Python wrapper around that and then implemented <code>supported_devices_for_op<\/code> <a href=\"https:\/\/github.com\/rwth-i6\/returnn\/blob\/master\/TFUtil.py\" rel=\"nofollow noreferrer\">here<\/a> (in <a href=\"https:\/\/github.com\/rwth-i6\/returnn\/commit\/ed47af101e0f750232c81944be154ff980eb70d9\" rel=\"nofollow noreferrer\">this commit<\/a>).<\/p>\n"}]}
{"QuestionId":42528563,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-03-01T09:57:04.467","AcceptedAnswerId":null,"OwnerUserId":7640063.0,"Title":"\"No module named tensorflow\" error after installation","Body":"<p>After installing the TensorFlow CPU version with virtualenv (in Ubuntu 14.0), when I try to test the installation with Python (not from the source directory), I get the following error:<\/p>\n\n<blockquote>\n  <p>No module named tensorflow<\/p>\n<\/blockquote>\n\n<p>Any explanation?<\/p>\n\n<p>I am following the exact sequence of steps from the TensorFlow installation page.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"54137643","CreationDate":"2019-01-10T22:11:46.310","ParentId":null,"OwnerUserId":"10606373","Title":null,"Body":"<p>Tensorflow does work with python 2.7 <a href=\"https:\/\/www.tensorflow.org\/install\/pip\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/pip<\/a>. Just wanted to add it here so no one get it wrong. Cant comment since I don't have the reputation yet :( <\/p>\n\n<p>Check your env and pip to make sure you have installed it for the correct python version in the env. <\/p>\n"},{"AnswerId":"44032587","CreationDate":"2017-05-17T18:42:40.387","ParentId":null,"OwnerUserId":"7099433","Title":null,"Body":"<p>Tensoflow works only with Python 3.5.x version<\/p>\n"}]}
{"QuestionId":42528639,"AnswerCount":1,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-03-01T10:01:34.143","AcceptedAnswerId":null,"OwnerUserId":7008017.0,"Title":"Keras CNN training accuracy is good but test accuracy is very low","Body":"<p>Please give me any comment for these CNN results.\nI have used 2000 training images and 400 test images.\nTraining accuracy is perfect but test accuracy is very low. \nI think it because there is much variation between training and test images.\nAnyone have a good idea for this case?<\/p>\n\n<p>[<img src=\"https:\/\/i.stack.imgur.com\/D66kV.png\" alt=\"Accuracy plot\">]<\/p>\n","answers":[{"AnswerId":"42528821","CreationDate":"2017-03-01T10:10:49.723","ParentId":null,"OwnerUserId":"3260576","Title":null,"Body":"<p>This is clear case of over-fitting. How many learneable parameters you have? For example VGGnet has 138M parameters and in this case its not very hard to see some neuron in the network must have sort of memorized a training image as it is and thus your network is not generalizing well.<\/p>\n\n<p><s>To fix that, first of all you can try a simpler model if task is simple like discriminating between shapes<\/s>. Also you can increase training data via transformations like swapping color channels (if it doesnt impact the output class), flipping or rotating image to make your net generalize better. Include L1\/L2 regularization in your loss function and try dropouts.<\/p>\n"}]}
{"QuestionId":42529560,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-01T10:43:11.567","AcceptedAnswerId":42534471.0,"OwnerUserId":2191652.0,"Title":"Tensorflow: What happens if loss becomes nan?","Body":"<p>During my training the loss becomes  from time to time. However this doesn't seem to affect my neural network training. Does anyone know if tensorflow has some internal handling of  values?  <\/p>\n\n<p>The loss values before and after the  are stable so I guess the learning rate is fine. I rather think that some of the images seem not to contain any sensible information, but I'm not 100% sure yet.<\/p>\n\n<p>Did anyone have similar experience?\nI'm using tensorflow 0.12.1<\/p>\n","answers":[{"AnswerId":"42534471","CreationDate":"2017-03-01T14:35:25.107","ParentId":null,"OwnerUserId":"524436","Title":null,"Body":"<p>Usually NaNs are \"infecting\", so once you see one pop up suddenly everything that is updated\/computed using the NaN value becomes NaN as well.<\/p>\n\n<p>What happens in your case is, I guess, that the NaN value only happens in the very last step of computing the loss and is not affecting the gradients (which don't need the value of the loss). If you use something like cross entropy, do you get exact 0s\/1s in your predictions? For numerical stability, it can be good to use TF's loss functions that have the softmax\/sigmoid integrated and not do the softmax\/sigmoid yourself.<\/p>\n"}]}
{"QuestionId":42529598,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-03-01T10:44:47.143","AcceptedAnswerId":47512401.0,"OwnerUserId":5436892.0,"Title":"Estimator with Coordinator as an input function for reading input data in distributed fashion in tensorflow","Body":"<p>The CNN cifar10 tutorial (tensor flow tutorials) gives an example of low-level API use for reading data as an independent job to train model (with multiple GPU).\nIs it possible to use high-level API Estimator with low-level threading support and multi\/single GPUs training?<\/p>\n\n<p>I am looking for a way to combine both: <\/p>\n\n<ul>\n<li><p>The custom Estimator from high-level API, details <a href=\"https:\/\/www.tensorflow.org\/extend\/estimators\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/extend\/estimators<\/a><\/p><\/li>\n<li><p>input_fn as a queue, which gives the same functionality which is described in <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/reading_data<\/a> for Coordinator class<\/p>\n\n<pre><\/pre><\/li>\n<\/ul>\n\n<p>It is not straightforward to me!<\/p>\n","answers":[{"AnswerId":"47512401","CreationDate":"2017-11-27T13:44:32.393","ParentId":null,"OwnerUserId":"4952954","Title":null,"Body":"<p>I push a code to <a href=\"https:\/\/github.com\/BobLiu20\/Classification_Nets\/blob\/master\/tensorflow\/training\/train_estimator.py\" rel=\"nofollow noreferrer\">here<\/a>.<br>\nIt is support input_fn as a queue when using estimator. High-level API Estimator with low-level threading support and multi\/single GPUs training. And easy to customer code all of you need.<\/p>\n"}]}
{"QuestionId":42530216,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><keras>","CreationDate":"2017-03-01T11:13:19.803","AcceptedAnswerId":42531757.0,"OwnerUserId":3813428.0,"Title":"How to access weight variables in Keras layers in tensor form for clip_by_weight?","Body":"<p>I'm implementing WGAN and need to clip weight variables.<\/p>\n\n<p>I'm currently using <em>Tensorflow<\/em> with <em>Keras<\/em> as high-level API. Thus building layers with Keras to avoid manually creation and initialization of variables.<\/p>\n\n<p>The problem is WGAN need to clip weight varibales, This can be done using  once I got those weight variable tensors, but I don't know to how to get them safely.<\/p>\n\n<p>One possible solution maybe using  to get all trainable variables. But I don't know how to get only <strong>weight<\/strong> variable without <strong>bias<\/strong> variables.<\/p>\n\n<p>Another solution is , but it get  arrays, although I can clip them with  APIs and set them using , but this may need CPU-GPU corporation, and may not be a good choice since clip operation needs to be performed on each train step.<\/p>\n\n<p>The only way I know is access them directly using <strong>exact<\/strong> variable names which I can get from TF lower level APIs or TensorBoard, but this is may not be safe since naming rule of Keras is not guaranteed to be stable.<\/p>\n\n<p>Is there any clean way to perform  only on those s with Tensorflow and Keras?<\/p>\n","answers":[{"AnswerId":"42531757","CreationDate":"2017-03-01T12:28:41.143","ParentId":null,"OwnerUserId":"2217210","Title":null,"Body":"<p>You can use constraints(<a href=\"https:\/\/keras.io\/constraints\/\" rel=\"nofollow noreferrer\">here<\/a>) class to implement new constraints on parameters. <\/p>\n\n<p>Here is how you can easily implement clip on weights and use it in your model.<\/p>\n\n<pre><code>from keras.constraints import Constraint\nfrom keras import backend as K\n\nclass WeightClip(Constraint):\n    '''Clips the weights incident to each hidden unit to be inside a range\n    '''\n    def __init__(self, c=2):\n        self.c = c\n\n    def __call__(self, p):\n        return K.clip(p, -self.c, self.c)\n\n    def get_config(self):\n        return {'name': self.__class__.__name__,\n                'c': self.c}\n\nimport numpy as np\nfrom keras.models import Sequential\nfrom keras.layers import Dense\n\nmodel = Sequential()\nmodel.add(Dense(30, input_dim=100, W_constraint = WeightClip(2)))\nmodel.add(Dense(1))\n\nmodel.compile(loss='mse', optimizer='rmsprop')\n\nX = np.random.random((1000,100))\nY = np.random.random((1000,1))\n\nmodel.fit(X,Y)\n<\/code><\/pre>\n\n<p>I have tested the running of the above code, but not the validity of the constraints. You can do so by getting the model weights after training using <code>model.get_weights()<\/code> or <code>model.layers[idx].get_weights()<\/code> and checking whether its abiding the constraints.<\/p>\n\n<p>Note: The constrain is not added to all the model weights .. but just to the weights of the specific layer its used and also <code>W_constraint<\/code> adds constrain to <code>W<\/code> param and b_constraint to <code>b<\/code> (bias) param<\/p>\n"}]}
{"QuestionId":42530487,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-01T11:26:58.170","AcceptedAnswerId":null,"OwnerUserId":2492058.0,"Title":"Distributing a separate dependency on Tensorflow","Body":"<p>I have some custom Tensorflow code I have in the  subdirectory of the project (all other parts of the code are standard Tensorflow from the official distribution).<\/p>\n\n<p>I would like to be able to distribute this code as an external dependency on Tensorflow, such that I can distribute the library via  and depend on the binary packages available for Tensorflow in  as well.<\/p>\n\n<p>My main goal is that I don't want to have users of my code have to compile the full Tensorflow tree (with my custom code only in ) just to get my custom code \/ module.<\/p>\n\n<p>Is this possible to do, and if so how?<\/p>\n","answers":[]}
{"QuestionId":42530721,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-01T11:39:04.743","AcceptedAnswerId":null,"OwnerUserId":6434823.0,"Title":"ImportError: No module named inception","Body":"<p>Im using imagenet train script (<a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/imagenet_train.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/imagenet_train.py<\/a>) to train my own image data set to classify in tensorflow (in Oracle VM Virtual box) and im getting an error as below <\/p>\n\n<p>shri@shri-VirtualBox:~\/Desktop\/Test\/inception_test\/models\/inception\/inception$ python imagenet_train.py\nTraceback (most recent call last):\n  File \"imagenet_train.py\", line 25, in \n    from inception import inception_train\nImportError: No module named inception<\/p>\n\n<p>Could someone please help understand the problem and how do i fix it?<\/p>\n\n<p>Regards,\nShri<\/p>\n","answers":[{"AnswerId":"42536052","CreationDate":"2017-03-01T15:43:09.507","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This sounds like a <a href=\"https:\/\/docs.python.org\/2\/tutorial\/modules.html#the-module-search-path\" rel=\"nofollow noreferrer\">Python module search path<\/a> issue. The <code>import<\/code> statements in the particular script <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/bc70271a51fe2e051b5d06edc6b9fd94880761d5\/inception\/inception\/imagenet_train.py\" rel=\"nofollow noreferrer\"><code>imagenet_train.py<\/code><\/a> and other scripts in that directory assume that they can find the other scripts in a submodule called <code>inception<\/code>, but when you run the script from the same directory, Python can't find that submodule.<\/p>\n\n<p>The easiest way to fix this is to change to the parent directory (<code>cd ~\/Desktop\/Test\/inception_test\/models\/inception<\/code>) and run the script as follows:<\/p>\n\n<pre><code>$ python inception\/inception_train.py\n<\/code><\/pre>\n"}]}
{"QuestionId":42530784,"AnswerCount":1,"Tags":"<python><tensorflow><virtualenv>","CreationDate":"2017-03-01T11:41:33.643","AcceptedAnswerId":null,"OwnerUserId":2502429.0,"Title":"File changes not updating in Tensorflow","Body":"<p>I recently installed TensorFlow on my mac with the recommended settings <a href=\"https:\/\/www.tensorflow.org\/install\/install_mac\" rel=\"nofollow noreferrer\">here<\/a>. If I save changes to a .py file, I have to deactivate\/activate in order to see them when I run the code.<\/p>\n\n<p>Why this is happening, and is there any way I can fix this?<\/p>\n","answers":[{"AnswerId":"42542288","CreationDate":"2017-03-01T21:22:00.113","ParentId":null,"OwnerUserId":"2502429","Title":null,"Body":"<p>Figured it out!<\/p>\n\n<ol>\n<li>Go to the directory where your code is saved.<\/li>\n<li><em>Then<\/em> run <code>source ~\/tensorflow\/bin\/activate<\/code><\/li>\n<\/ol>\n"}]}
{"QuestionId":42532061,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-03-01T12:42:43.173","AcceptedAnswerId":null,"OwnerUserId":7641250.0,"Title":"tensorflow: run model evaluation over multiple checkpoints","Body":"<p>In my current project I train a model and save checkpoints every 100 iteration steps. The checkpoint files are all saved to the same directory (model.ckpt-100, model.ckpt-200 , model.ckpt-300 etc). And after that I would like to evalute the model based on validation data for all the saved checkpoints, not just the latest one.<\/p>\n\n<p>Currently my piece of code for restoring the checkpoint file looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>However, this restores only the latest saved checkpoint file. So how do I write a loop over all saved checkpoint files? I tried to get a list of the checkpoint files with saver.last_checkpoints, however, the returned list is empty.<\/p>\n\n<p>Any help would be highly appreciated, thanks in advance!<\/p>\n","answers":[{"AnswerId":"42535291","CreationDate":"2017-03-01T15:09:00.990","ParentId":null,"OwnerUserId":"3867406","Title":null,"Body":"<p>You can iterate through the files in the directory :<\/p>\n\n<pre><code>import os\n\ndir_path = '.\/' #change that to wherever your files are\nckpt_files = [f for f in os.listdir(dir_path) if os.path.isfile(\n    os.path.join(dir_path, f)) and 'ckpt' in f]\n\nfor ckpt_file in ckpt_files:\n    saver.restore(sess, dir_path + ckpt_file)\n    global_step = ckpt.model_checkpoint_path.split('\/')[-1].split('-')[-1]\n    print('Succesfully loaded model from %s at step=%s.' %\n          (ckpt.model_checkpoint_path, global_step))\n\n    # Do your thing\n<\/code><\/pre>\n\n<p>add more conditions in the above list comprehension to be more selective like : <code>and 'meta' not in f<\/code> and so on depending on what's in that dir and the saver version you have<\/p>\n"},{"AnswerId":"53595586","CreationDate":"2018-12-03T14:12:56.280","ParentId":null,"OwnerUserId":"6383205","Title":null,"Body":"<h2>Fastest solution:<\/h2>\n\n<p><code>tensor2tensor<\/code> has a module <code>utils<\/code> with a script <a href=\"https:\/\/github.com\/tensorflow\/tensor2tensor\/blob\/master\/tensor2tensor\/utils\/avg_checkpoints.py\" rel=\"nofollow noreferrer\"><code>avg_checkpoints.py<\/code><\/a> that saves the averaged weights in a new checkpoint. Let us say you have a list of checkpoints that you want to average. You have 2 options for usage:<\/p>\n\n<ol>\n<li><p>From command line<\/p>\n\n<pre><code>TRAIN_DIR=path_to_your_model_folder\nFNC_PATH=path_to_tensor2tensor+'\/utils\/avg.checkpoints.py'\nCKPTS=model.ckpt-10000,model.ckpt-20000,model.ckpt-100000\n\npython3 $FNC_PATH --prefix=$TRAIN_DIR --checkpoints=$CKPTS \\ \n    --output_path=\"${TRAIN_DIR}averaged.ckpt\"\n<\/code><\/pre><\/li>\n<li><p>From your own code (use <code>os.system<\/code>):<\/p>\n\n<pre><code>import os\nos.system(\n    \"python3 \"+FNC_DIR+\" --prefix=\"+TRAIN_DIR+\" --checkpoints=\"+CKPTS+\n    \" --output_path=\"+TRAIN_DIR+\"averaged.ckpt\"\n)\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>As an alternative to specifying a list of checkpoints and using the <code>--checkpoints<\/code> argument, you can just use <code>--num_checkpoints=10<\/code> to average the last 10 checkpoints.<\/p>\n\n<h2>If you do not want a dependency on <code>tensor2tensor<\/code>:<\/h2>\n\n<p>Here is a code snippet that does not rely on <code>tensor2tensor<\/code>, but can still average a <strong>variable number of checkpoints<\/strong> (as opposed to ted's answer). Assume <code>steps<\/code> is a list of checkpoints that should be merged (e.g. <code>[10000, 20000, 30000, 40000]<\/code>).<\/p>\n\n<p>Then:<\/p>\n\n<pre><code># Restore all sessions and save the weight matrices\nvalues = []\nfor step in steps:\n    tf.reset_default_graph()\n    path = model_path+'\/model.ckpt-'+str(step)\n    with tf.Session() as sess:\n        saver = tf.train.import_meta_graph(path+'.meta')\n        saver.restore(sess, path)\n        values.append(sess.run(tf.all_variables()))\n\n# Average weights\nvariables = tf.all_variables()\nall_assign = []\nfor ind, var in enumerate(variables):\n    weights = np.concatenate(\n        [np.expand_dims(w[ind],axis=0)  for w in values],\n        axis=0\n    )\n    all_assign.append(tf.assign(var, np.mean(weights, axis=0))\n<\/code><\/pre>\n\n<p>Then you can proceed, however you prefer, e.g. saving the averaged checkpoint:<\/p>\n\n<pre><code># Now save the new values into a separate checkpoint\nwith tf.Session() as sess_test:\n    sess_test.run(all_assign)\n    saver = tf.train.Saver() \n    saver.save(sess_test, model_path+'\/average_'+str(num_checkpoints))\n<\/code><\/pre>\n"},{"AnswerId":"53596293","CreationDate":"2018-12-03T14:56:18.760","ParentId":null,"OwnerUserId":"10688944","Title":null,"Body":"<p>Best solution, follow this <a href=\"https:\/\/github.com\/vonclites\/checkmate\" rel=\"nofollow noreferrer\">link<\/a><\/p>\n\n<p>I've been using this for a long while, it as neat.\nYou can save model according to whatever the metric you like.\nIt deletes the older checkpoints and loads the best checkpoints when restoring model.<\/p>\n\n<p>If you are using accuracy as a metric set maximize= True to save model according to maximum accuracy.\nIf you are evaluating model on the basis of validation loss, you can set maximize flag to False to so that the model with least validation loss is saved.<\/p>\n"}]}
{"QuestionId":42532386,"AnswerCount":3,"Tags":"<python><neural-network><deep-learning><keras>","CreationDate":"2017-03-01T12:56:56.797","AcceptedAnswerId":42539325.0,"OwnerUserId":7553189.0,"Title":"How to work with multiple inputs for LSTM in Keras?","Body":"<p>I'm trying to predict the water usage of a population.<\/p>\n\n<p>I have 1 main input:<\/p>\n\n<ul>\n<li>Water volume<\/li>\n<\/ul>\n\n<p>and 2 secondary inputs:<\/p>\n\n<ul>\n<li>Temperature<\/li>\n<li>Rainfall<\/li>\n<\/ul>\n\n<p>In theory they have a relation with the water supply. <\/p>\n\n<p>It must be said that each rainfall and temperature data correspond with the water volume. So this is a time series problem.<\/p>\n\n<p>The problem is that I don't know how to use 3 inputs from just one .csv file, with 3 columns, each one for each input, as the code below is made. When I have just one input (e.g.water volume) the network works more or less good with this code, but not when I have more than one. (So if you run this code with the csv file below, it will show a dimension error).<\/p>\n\n<p>Reading some answers from:<\/p>\n\n<ul>\n<li><a href=\"http:\/\/machinelearningmastery.com\/time-series-prediction-lstm-recurrent-neural-networks-python-keras\/\" rel=\"noreferrer\">Time Series Prediction with LSTM Recurrent Neural Networks in Python with Keras<\/a><\/li>\n<li><a href=\"http:\/\/machinelearningmastery.com\/time-series-forecast-study-python-annual-water-usage-baltimore\/\" rel=\"noreferrer\">Time Series Forecast Case Study with Python: Annual Water Usage in Baltimore<\/a><\/li>\n<\/ul>\n\n<p>it seems to be that many people have the same problem.<\/p>\n\n<p>The code:<\/p>\n\n<p><strong>EDIT:<\/strong> Code has been updated<\/p>\n\n<pre><\/pre>\n\n<p>This is the csv file I have created, if it helps.<\/p>\n\n<p><a href=\"https:\/\/pastebin.com\/qZv5HygU\" rel=\"noreferrer\">datos.csv<\/a><\/p>\n\n<p>After changing the code, I fixed all the errors, but I'm not really sure about the results. This is a zoom in the prediction plot:\n<a href=\"https:\/\/i.stack.imgur.com\/gEJHZ.png\" rel=\"noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/gEJHZ.png\" alt=\"Zoom on plot\"><\/a><\/p>\n\n<p>which shows that there is a \"displacement\" in the values predicted and in the real ones. When there is a max in the real time series, there is a min in the forecast for the same time, but it seems like it corresponds to the previous time step.<\/p>\n","answers":[{"AnswerId":"54026565","CreationDate":"2019-01-03T16:54:45.073","ParentId":null,"OwnerUserId":"6004997","Title":null,"Body":"<p>You can change what you are optimizing, for maybe better results. For example, try predicting binary 0,1 if there will be a 'spike up' for the next day. Then feed the probability of a 'spike up' as a feature to predict the usage itself.<\/p>\n"},{"AnswerId":"42539325","CreationDate":"2017-03-01T18:31:01.437","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Change <\/p>\n\n<pre><code>a = dataset[i:(i + look_back), 0]\n<\/code><\/pre>\n\n<p>To <\/p>\n\n<pre><code>a = dataset[i:(i + look_back), :]\n<\/code><\/pre>\n\n<p>If you want the 3 features in your training data. <\/p>\n\n<p>Then use <\/p>\n\n<pre><code>model.add(LSTM(4, input_shape=(look_back,3)))\n<\/code><\/pre>\n\n<p>To specify that you have <code>look_back<\/code> time steps in your sequence, each with 3 features. <\/p>\n\n<p>It should run<\/p>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>Indeed, <code>sklearn.preprocessing.MinMaxScaler()<\/code>'s function : <code>inverse_transform()<\/code> takes an input which has the same shape as the object you fitted. So you need to do something like this :<\/p>\n\n<pre><code># Get something which has as many features as dataset\ntrainPredict_extended = np.zeros((len(trainPredict),3))\n# Put the predictions there\ntrainPredict_extended[:,2] = trainPredict\n# Inverse transform it and select the 3rd column.\ntrainPredict = scaler.inverse_transform(trainPredict_extended)[:,2]\n<\/code><\/pre>\n\n<p>I guess you will have other issues like this below in your code but nothing that you can't fix :) the ML part is fixed and you know where the error comes from. Just check the shapes of your objects and try to make them match.<\/p>\n"},{"AnswerId":"51411847","CreationDate":"2018-07-18T22:28:41.180","ParentId":null,"OwnerUserId":"10049949","Title":null,"Body":"<p>The displacement could be due to the lag in predicting maximums\/minimums given the randomness in the data.<\/p>\n"}]}
{"QuestionId":42533280,"AnswerCount":0,"Tags":"<python><deep-learning><caffe><pycaffe><matcaffe>","CreationDate":"2017-03-01T13:38:39.240","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"Getting error for adding InfogainLoss layer?","Body":"<p>I followed the instructions <a href=\"https:\/\/stackoverflow.com\/questions\/27632440\/infogainloss-layer\/27645934#27645934\">here<\/a> to create weight matrix for classes and add the Infogainloss function, which is able to tackle my imbalance data. The few last layers <\/p>\n\n<pre><\/pre>\n\n<p>And edited\/added the files that have been included in this <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/2805\/files#diff-e0bd8d028edf0f88febfebe3bf703904\" rel=\"nofollow noreferrer\">link<\/a>. However, it is giving an error while creating the layer in the network:<\/p>\n\n<pre><\/pre>\n\n<ol>\n<li>How can I solve this error? from which point this error stem from?<\/li>\n<li><p>And what is this parameter exactly doing? <\/p>\n\n<p><\/p><\/li>\n<\/ol>\n\n<p>I would really appreciate if someone could post any suggestion.\nThanks.<\/p>\n","answers":[]}
{"QuestionId":42533809,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deconvolution>","CreationDate":"2017-03-01T14:03:53.070","AcceptedAnswerId":null,"OwnerUserId":7641674.0,"Title":"How could I read image from a directory as input and output while traing a CNN model in Tensorflow?","Body":"<p>I want to use CNN to solve the deblurring task, and I have training data that is a directory of png images and a corresponding text file containing the files name. <\/p>\n\n<p>As the data is too large to add to the memory with one step, and is there any API or some method to make it possible that I could read the blury image as input and its ground-truth as expected result to train?<\/p>\n\n<p>I have spent quite a few time to solve this, but I got confused after read the APIs in the online API introductions.<\/p>\n","answers":[{"AnswerId":"42594179","CreationDate":"2017-03-04T09:06:47.720","ParentId":null,"OwnerUserId":"7641674","Title":null,"Body":"<p>The method is not that confused. The tensorflow provides TFrecords file to make good use of the memory.<\/p>\n\n<pre><code>def create_cord():\n\n    writer = tf.python_io.TFRecordWriter(\"train.tfrecords\")\n    for index in xrange(66742):\n        blur_file_name = get_file_name(index, True)\n        orig_file_name = get_file_name(index, False)\n        blur_image_path = cwd + blur_file_name\n        orig_image_path = cwd + orig_file_name\n\n        blur_image = Image.open(blur_image_path)\n        orig_image = Image.open(orig_image_path)\n\n        blur_image = blur_image.resize((IMAGE_HEIGH, IMAGE_WIDTH))\n        orig_image = orig_image.resize((IMAGE_HEIGH, IMAGE_WIDTH))\n\n        blur_image_raw = blur_image.tobytes()\n        orig_image_raw = orig_image.tobytes()\n        example = tf.train.Example(features=tf.train.Features(feature={\n        \"blur_image_raw\": tf.train.Feature(bytes_list=tf.train.BytesList(value=[blur_image_raw])),\n        'orig_image_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=[orig_image_raw]))\n    }))\n    writer.write(example.SerializeToString())\n    writer.close()\n<\/code><\/pre>\n\n<p>to read the dataset:<\/p>\n\n<pre><code>def read_and_decode(filename):\n    filename_queue = tf.train.string_input_producer([filename])\n\n    reader = tf.TFRecordReader()\n    _, serialized_example = reader.read(filename_queue)\n    features = tf.parse_single_example(serialized_example,\n                                   features={\n                                       'blur_image_raw':    tf.FixedLenFeature([], tf.string),\n                                       'orig_image_raw': tf.FixedLenFeature([], tf.string),\n                                   })\n\n    blur_img = tf.decode_raw(features['blur_image_raw'], tf.uint8)\n    blur_img = tf.reshape(blur_img, [IMAGE_WIDTH, IMAGE_HEIGH, 3])\n    blur_img = tf.cast(blur_img, tf.float32) * (1. \/ 255) - 0.5\n\n    orig_img = tf.decode_raw(features['blur_image_raw'], tf.uint8)\n    orig_img = tf.reshape(orig_img, [IMAGE_WIDTH, IMAGE_HEIGH, 3])\n    orig_img = tf.cast(orig_img, tf.float32) * (1. \/ 255) - 0.5\n\n    return blur_img, orig_img\n\n\nif __name__ == '__main__':\n\n    #  create_cord()\n\n    blur, orig = read_and_decode(\"train.tfrecords\")\n    blur_batch, orig_batch = tf.train.shuffle_batch([blur, orig],\n                                                batch_size=3, capacity=1000,\n                                                min_after_dequeue=100)\n    init = tf.global_variables_initializer()\n    with tf.Session() as sess:\n        sess.run(init)\n     # \u542f\u52a8\u961f\u5217\n        threads = tf.train.start_queue_runners(sess=sess)\n        for i in range(3):\n            v, l = sess.run([blur_batch, orig_batch])\n            print(v.shape, l.shape)\n<\/code><\/pre>\n"}]}
{"QuestionId":42534059,"AnswerCount":1,"Tags":"<c++><tensorflow><eigen><eigen3>","CreationDate":"2017-03-01T14:15:40.470","AcceptedAnswerId":null,"OwnerUserId":2131200.0,"Title":"Vector of Eigen tensors with different ranks","Body":"<p>My question: Is it possible to create a vector of Eigen tensors with different ranks?<\/p>\n\n<p>My goal is to be able to create a vector that contains objects whose types take values among  (defined below). The objects can have different types.<\/p>\n\n<p>Thanks in advance for your help!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42536197","CreationDate":"2017-03-01T15:49:40.967","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Since you tagged this question <a href=\"https:\/\/stackoverflow.com\/questions\/tagged\/tensorflow\">[tensorflow]<\/a>, I assume that you are using <code>Eigen::Tensor<\/code> as part of a TensorFlow program. The <a href=\"https:\/\/stackoverflow.com\/a\/13462096\/3574081\">answer<\/a> to which <a href=\"https:\/\/stackoverflow.com\/questions\/42534059\/vector-of-eigen-tensors-with-different-ranks#comment72204390_42534059\">Aziuth pointed in their comment<\/a> suggests using a wrapper class to hold the tensors, and fortunately TensorFlow comes with its own wrapper class:  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/tensor\" rel=\"nofollow noreferrer\"><code>tensorflow::Tensor<\/code><\/a>.<\/p>\n\n<p>You can create a <code>std::vector&lt;tensorflow::Tensor&gt;<\/code>, use TensorFlow's allocator to allocate memory for the different tensors, and access the wrapped <code>Eigen::Tensor<\/code> objects using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/tensor#tensor_18\" rel=\"nofollow noreferrer\"><code>tensorflow::Tensor::tensor()<\/code><\/a> method.<\/p>\n"}]}
{"QuestionId":42535795,"AnswerCount":1,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2017-03-01T15:32:23.430","AcceptedAnswerId":null,"OwnerUserId":7574973.0,"Title":"Implementing stack denoising autoencoder with tensorflow","Body":"<p>I was trying to implement a stack denoising autoencoder in tensorflow. Here is the code I got. It worked with one layer, but when I tried to stack it(by changing the list of parameter n_neuron). It doesn't work anymore. I was trying to debug it for a long time but still couldn't get the answer.    <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42537002","CreationDate":"2017-03-01T16:29:15.510","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Can you try this?<\/p>\n\n<pre><code>n_neuron = [n_visible,500,400] #n_visible is input layer size, the numbers after are hidden size neuorn unit nunmbers\n<\/code><\/pre>\n\n<p>This works perfectly for me on my computer. If it does not work for you, please let us know what error you get. <\/p>\n"}]}
{"QuestionId":42535815,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-01T15:33:27.257","AcceptedAnswerId":null,"OwnerUserId":3052875.0,"Title":"How to add more layers to Convolutional Neural Network text classification TensorFlow example?","Body":"<p>According to the documentation, the model presented in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/text_classification_character_cnn.py\" rel=\"nofollow noreferrer\">this example<\/a> is similar to the following paper:\n\"<a href=\"https:\/\/arxiv.org\/abs\/1509.01626\" rel=\"nofollow noreferrer\">Character-level Convolutional Networks for Text Classification<\/a>\"<\/p>\n\n<p>I found that the original model (presented in the paper) contains 9 layers deep with 6 convolutional layers and 3 fully-connected layers, but the implemented example contains only two convolutional layers:<\/p>\n\n<pre><\/pre>\n\n<p>If anybody can help me to extend this model for more layers?<\/p>\n","answers":[{"AnswerId":"42544781","CreationDate":"2017-03-02T00:32:21.327","ParentId":null,"OwnerUserId":"2505209","Title":null,"Body":"<ol>\n<li><p>Similar to BVLC Caffenet : <\/p>\n\n<p>def bvlc_caffenet(imgs,weights,biases):\n   # mean subtraction\n   mean = tf.constant([123.68, 116.779, 103.939], dtype=tf.float32, shape=[1, 1, 1, 3], name='img_mean')\n   images = imgs-mean<\/p>\n\n<p>#conv1\n   conv1 = tf.nn.conv2d(images,weights['c1'], [1, 3, 3, 1], padding='VALID')\n   out1  = tf.nn.relu(tf.nn.bias_add(conv1, biases['b1']))\n   pool1 = tf.nn.max_pool(out1,ksize=[1,3,3,1], strides=[1,2,2,1],padding='VALID')<\/p>\n\n<p>#conv2\n   conv2 = tf.nn.conv2d(pool1,weights['c2'], [1, 1, 1, 1], padding='VALID')\n   out2  = tf.nn.relu(tf.nn.bias_add(conv2, biases['b2']))\n   pool2 = tf.nn.max_pool(out2,ksize=[1,3,3,1], strides=[1,2,2,1],padding='VALID')<\/p>\n\n<p>#conv3\n   conv3 = tf.nn.conv2d(pool2,weights['c3'], [1, 1, 1, 1], padding='VALID')\n   out3  = tf.nn.relu(tf.nn.bias_add(conv3, biases['b3']))<\/p>\n\n<p>#conv4\n   conv4 = tf.nn.conv2d(out3,weights['c4'], [1, 1, 1, 1], padding='VALID')\n   out4  = tf.nn.relu(tf.nn.bias_add(conv4, biases['b4']))<\/p>\n\n<p>#conv5\n   conv5 = tf.nn.conv2d(out4,weights['c5'], [1, 1, 1, 1], padding='VALID')\n   out5  = tf.nn.relu(tf.nn.bias_add(conv5, biases['b5']))\n   pool5 = tf.nn.max_pool(out5,ksize=[1,3,3,1], strides=[1,2,2,1],padding='VALID')<\/p>\n\n<p>#flattening\n   shape = int(np.prod(pool5.get_shape()[1:]))\n   pool5_flat = tf.reshape(pool5, [-1, shape]) <\/p>\n\n<p>#fc6\n   fc6  = tf.matmul(pool5_flat,weights['f6'])\n   out6 = tf.nn.relu(tf.nn.bias_add(fc6,biases['b6']))\n   out6 = tf.nn.dropout(out6,0.5)<\/p>\n\n<p>#fc7\n   fc7  = tf.matmul(out6,weights['f7'])\n   out7 = tf.nn.relu(tf.nn.bias_add(fc7,biases['b7']))\n   out7 = tf.nn.dropout(out7,0.5)<\/p>\n\n<p>#fc8\n   fc8  = tf.matmul(out7,weights['f8'])\n   out8 = tf.nn.relu(tf.nn.bias_add(fc8,biases['b8']))\n   out8 = tf.nn.dropout(out8,0.5)<\/p>\n\n<p>probs = tf.nn.softmax(out8)\n   return probs<\/p>\n\n<h1>Initialized Weights and Biases for the Network<\/h1>\n\n<p>weights = {\n'c1': tf.Variable(tf.truncated_normal([7,7,3,96],stddev=0.1)),\n'c2': tf.Variable(tf.truncated_normal([5,5,96,256],stddev=0.1)),\n'c3': tf.Variable(tf.truncated_normal([3,3,256,384],stddev=0.1)),\n'c4': tf.Variable(tf.truncated_normal([3,3,384,384],stddev=0.1)),\n'c5': tf.Variable(tf.truncated_normal([3,3,384,256],stddev=0.1)),\n'f6': tf.Variable(tf.truncated_normal([4096,2048],stddev=0.1)),\n'f7': tf.Variable(tf.truncated_normal([2048,2048],stddev=0.1)),\n'f8': tf.Variable(tf.truncated_normal([2048,1000],stddev=0.1))\n}\nbiases = {\n'b1' : tf.Variable(tf.constant(0.1, shape=[96])),\n'b2' : tf.Variable(tf.constant(0.1, shape=[256])),\n'b3' : tf.Variable(tf.constant(0.1, shape=[384])),\n'b4' : tf.Variable(tf.constant(0.1, shape=[384])),\n'b5' : tf.Variable(tf.constant(0.1, shape=[256])),\n'b6' : tf.Variable(tf.constant(0.1, shape=[2048])),\n'b7' : tf.Variable(tf.constant(0.1, shape=[2048])),\n'b8' : tf.Variable(tf.constant(0.1, shape=[1000]))\n}<\/p><\/li>\n<\/ol>\n\n<p>Or <\/p>\n\n<ol start=\"2\">\n<li>follow this (another format) : <a href=\"https:\/\/www.cs.toronto.edu\/~frossard\/vgg16\/vgg16.py\" rel=\"nofollow noreferrer\">https:\/\/www.cs.toronto.edu\/~frossard\/vgg16\/vgg16.py<\/a><\/li>\n<\/ol>\n\n<p>Are these helpful ?<\/p>\n"}]}
{"QuestionId":42536784,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-01T16:17:44.933","AcceptedAnswerId":null,"OwnerUserId":6461097.0,"Title":"Load Tensorflow Model and reuse it by the another functions","Body":"<p>I have to load model in initialization and then reuse it by some function. In my architecture function run predict on several datasets by one model and I thought that load model from hard drive for each dataset is not a good solution.<\/p>\n\n<p>I need something like this, that sharing session (or model) between functions:<\/p>\n\n<pre><\/pre>\n\n<p>Is there any proper way to do this?<\/p>\n","answers":[{"AnswerId":"42544888","CreationDate":"2017-03-02T00:45:26.953","ParentId":null,"OwnerUserId":"2505209","Title":null,"Body":"<p>You can define model architecture and load its weights as a class. Use it for various datasets in different functions. <\/p>\n\n<pre><code>class vgg16:\ndef __init__(self, imgs, weights=None, sess=None):\n    self.imgs = imgs\n    self.convlayers()\n    self.fc_layers()\n    self.probs = tf.nn.softmax(self.fc3l)\n    if weights is not None and sess is not None:\n        self.load_weights(weights, sess)\n\ndef convlayers(self):\n    self.parameters = []\n      ::\n      ::\n\nif __name__ == '__main__':\n    sess = tf.Session()\n    imgs = tf.placeholder(tf.float32, [None, 224, 224, 3])\n    vgg = vgg16(imgs, 'vgg16_weights.npz', sess)\n\n    # run inference here on multiple datasets\n<\/code><\/pre>\n\n<p>Check here : <a href=\"https:\/\/www.cs.toronto.edu\/~frossard\/vgg16\/vgg16.py\" rel=\"nofollow noreferrer\">https:\/\/www.cs.toronto.edu\/~frossard\/vgg16\/vgg16.py<\/a><\/p>\n\n<p>Does that helps you ? <\/p>\n"}]}
{"QuestionId":42537244,"AnswerCount":1,"Tags":"<tensorflow><neural-network><non-linear-regression>","CreationDate":"2017-03-01T16:39:47.307","AcceptedAnswerId":null,"OwnerUserId":7634931.0,"Title":"Got stuck adding 9 random numbers with tensorflow \/ Python","Body":"<p>Sorry Newbee in Tensorflow and Python\nI implemented this code to learn the sum of 9 random numbers. I'm getting an error which I can't understand.Unfortunately I could not find a similar problem in tutorials our here ...<\/p>\n\n<pre><\/pre>\n\n<p>And I'm getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>Any help is appreciated.<\/p>\n","answers":[{"AnswerId":"42537878","CreationDate":"2017-03-01T17:10:24.080","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The code defines <code>y<\/code> twice:<\/p>\n\n<pre><code>y = tf.placeholder(tf.float32, shape=[n_samples])\n# ...\ny = tf.matmul(x, W) + b\n<\/code><\/pre>\n\n<p>Since <code>y<\/code> is just a normal Python variable, the second assignment overwrites the placeholder with the output of your bias addition. When you come to feed a value for <code>y<\/code>, TensorFlow interprets this as attempting to feed a replacement value for the result of <code>tf.matmul(x, W) + b<\/code>, and not the original <code>tf.placeholder()<\/code>.<\/p>\n\n<p>To fix this problem, use a different Python variable name for the placeholder <code>y<\/code> and result of <code>tf.matmul(x, W) + b<\/code>.<\/p>\n"}]}
{"QuestionId":42537552,"AnswerCount":0,"Tags":"<windows><tensorflow>","CreationDate":"2017-03-01T16:54:16.797","AcceptedAnswerId":null,"OwnerUserId":258483.0,"Title":"Numerous op_kernel.cc:943 error messages while running tensorflow on Windows","Body":"<p>Have installed Tensorflow on Windows. Simple check example causes the following<\/p>\n\n<pre><\/pre>\n\n<p>I.e. first run of session causes numerous error messges in  and the result is output not exactly as expected:<\/p>\n\n<blockquote>\n  <p>b'Hello, TensorFlow!'<\/p>\n<\/blockquote>\n\n<p>i.e. decorated with  and quotes.<\/p>\n\n<p>What can be the reason?<\/p>\n","answers":[]}
{"QuestionId":42537650,"AnswerCount":0,"Tags":"<tensorflow><clang>","CreationDate":"2017-03-01T16:58:39.557","AcceptedAnswerId":null,"OwnerUserId":1930462.0,"Title":"Compiling tensorflow: undefined reference to `clSetUserEventStatus@OPENCL_1.1'","Body":"<p><i>PS: Question is at the end, the following is just background information that might help.<\/i><\/p>\n\n<p>I'm trying to compile tensorflow, but getting this error:<\/p>\n\n<pre><\/pre>\n\n<p>By using \"strace\" tool, i've isolated the exact statement that's failing to be:<\/p>\n\n<pre><\/pre>\n\n<p>strace also confirms that this command IS reading , which contains the aforementioned  symbol.<\/p>\n\n<p>I verified that  contains  by checking the output of the  command:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><b>So here is my question<\/b>: clang (and\/or its linker) is reading the file that contains the symbol that it's complaining about... why is it \"missing something\" and complaining that it's undefined?<\/p>\n\n<p>How can I get this compile to work?<\/p>\n","answers":[]}
{"QuestionId":42537674,"AnswerCount":1,"Tags":"<machine-learning><lua><torch>","CreationDate":"2017-03-01T16:59:55.870","AcceptedAnswerId":42555217.0,"OwnerUserId":1984680.0,"Title":"Transform AAABBB table of tensors into nested ABABAB table of tensors in Torch","Body":"<p>I am trying to solve this problem for quite some time and still cannot come up with a solution, maybe somebody can help me out here. I have the following <strong>AAABBB<\/strong> input table to the neural network:<\/p>\n\n<pre><\/pre>\n\n<p>The above table gets preprocessed and then needs to be transformed into <strong>nested<\/strong> <strong>ABABAB<\/strong> input table:<\/p>\n\n<pre><\/pre>\n\n<p>How can I transform <strong>AAABBB<\/strong> table to <strong>ABABAB<\/strong> <strong>nested<\/strong> table <strong>within<\/strong> the network using <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/table.md\" rel=\"nofollow noreferrer\">Torch table layers<\/a>?  <\/p>\n","answers":[{"AnswerId":"42555217","CreationDate":"2017-03-02T12:12:26.630","ParentId":null,"OwnerUserId":"1984680","Title":null,"Body":"<p>It appears that <strong>dpnn<\/strong> package contains a <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/doc\/containers.md#nn.Containers\" rel=\"nofollow noreferrer\">Container<\/a> that does exactly that.\n<a href=\"https:\/\/github.com\/Element-Research\/dpnn#nn.ZipTable\" rel=\"nofollow noreferrer\">ZipTable<\/a> zips a table of tables into a table of tables.<\/p>\n\n<p>Here is a working example of how to convert <strong>AAABBB<\/strong> table into <strong>ABABAB nested<\/strong> table.<\/p>\n\n<pre><code>require 'dpnn'\n\naaa = torch.DoubleTensor(3,32,200)\nbbb = torch.DoubleTensor(3,32,54)\n\nmodel = nn.Sequential()\npar = nn.ParallelTable()\npar:add(nn.SplitTable(1))\npar:add(nn.SplitTable(1))\nmodel:add(par)\nmodel:add(nn.ZipTable())\nmodel:forward({aaa,bbb})\n<\/code><\/pre>\n"}]}
{"QuestionId":42538535,"AnswerCount":1,"Tags":"<machine-learning><neural-network><caffe><conv-neural-network><function-fitting>","CreationDate":"2017-03-01T17:45:21.257","AcceptedAnswerId":42541707.0,"OwnerUserId":2439346.0,"Title":"Is it necessary to define the backward function for the frozen layers?","Body":"<p>I have a function fitting network with 4 hidden layers.<\/p>\n\n<p>I need to find suitable weights for the <strong>first<\/strong> and the <strong>third<\/strong> layers,\nwhile the <strong>second<\/strong> and the <strong>fourth<\/strong> layers are some kinds of <em>normalization<\/em> layers and do not need to be learned, so I just froze them by setting their learning rate to zero.<\/p>\n\n<p>My question is:<\/p>\n\n<p>Should I define <strong>backward function<\/strong> for those two frozen layers?<\/p>\n\n<p>I <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/src\/caffe\/layers\/pooling_layer.cpp#L230\" rel=\"nofollow noreferrer\">saw<\/a> in caffe that the pooling layer that does not have any learnable parameter has the backward function.<\/p>\n\n<p>Thanks in advance,<\/p>\n","answers":[{"AnswerId":"42541707","CreationDate":"2017-03-01T20:45:19.100","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>Yes, you need a backward pass, otherwise your learning would stop at this layer (nothing below it will learn). Even for non-learnable layers you need to compute <strong>valid gradients<\/strong>.<\/p>\n"}]}
{"QuestionId":42539094,"AnswerCount":2,"Tags":"<theano><theano-cuda>","CreationDate":"2017-03-01T18:17:07.017","AcceptedAnswerId":null,"OwnerUserId":5232808.0,"Title":"Theano GPU Installation Issue","Body":"<p>I have installed Anaconda, theano , GPU Toolkit ver 8. I am getting this error.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42942949","CreationDate":"2017-03-22T04:46:49.977","ParentId":null,"OwnerUserId":"4492375","Title":null,"Body":"<p>See <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/lasagne-users\/Fg3o6seitOM\" rel=\"nofollow noreferrer\">discussion<\/a> here.<\/p>\n\n<p>I reinstalled <a href=\"http:\/\/deeplearning.net\/software\/libgpuarray\/installation.html\" rel=\"nofollow noreferrer\">libgpuarray<\/a> and it worked for me.<\/p>\n"},{"AnswerId":"42579089","CreationDate":"2017-03-03T12:37:46.733","ParentId":null,"OwnerUserId":"7653479","Title":null,"Body":"<p>Your cuda driver is faulty. Install another one, preferably 373.06. Using your current driver will result in wrong computations. DO NOT force the driver load<\/p>\n"}]}
{"QuestionId":42539370,"AnswerCount":2,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2017-03-01T18:33:59.320","AcceptedAnswerId":42543815.0,"OwnerUserId":137783.0,"Title":"Cuda issue in TensorFlow 1.0 tutorial looks like TF can't find CUPTI\/lib64?","Body":"<p>This question has nothing to do with the warnings SSE AVX etc.. I've included the output for completeness. The issue is the fail on some cuda libs, I think, at the end, the machine has a NVIDA 1070 card and has the Cuda libs that are used earlier in the process but something is missing at the end?\nI pip installed release 1.0 of TensorFlow\nI also downloaded the repo separately to get the most up to date tutorials.\nThis tutorial specifically to get instances of all of Tensorboard capabilities..\nAttempting to run the Minst_with_summaries.py from the tensorFlow tutorials in the repo (I copied the file out of the repo into a working directory) and I'm using Anaconda and Python 3.6 I get the following:<\/p>\n\n<pre><\/pre>\n\n<p>Looks to me that the installation of TensorFlow may be missing some stuff See that last several lines above?\nHow to fix? \nAlso reference this issue on GitHub: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7975\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7975<\/a><\/p>\n\n<p>The answer was posted on GitHub and it seems there is an install bug that can be fixed by:<\/p>\n\n<pre><\/pre>\n\n<p>Would be helpful if @mrry would reopen so others can see the correct resolution.<\/p>\n","answers":[{"AnswerId":"42543815","CreationDate":"2017-03-01T23:04:08.230","ParentId":null,"OwnerUserId":"137783","Title":null,"Body":"<p>Also reference this issue on GitHub: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7975\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7975<\/a><\/p>\n\n<p>You may try the apt-get install that the git-hub issue suggests but that did not do it for me: This did:<\/p>\n\n<p>The answer was posted on GitHub and it seems there is an install bug that can be fixed by:<\/p>\n\n<pre><code>adding \/usr\/local\/cuda\/extras\/CUPTI\/lib64 to your LD_LIBRARY_PATH\n<\/code><\/pre>\n\n<p>you can do that by editing your .bash profile<\/p>\n"},{"AnswerId":"43302377","CreationDate":"2017-04-09T02:23:35.433","ParentId":null,"OwnerUserId":"6521116","Title":null,"Body":"<p>I encounter this on my windows. I solved by add <code>C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v8.0\\extras\\CUPTI\\libx64<\/code> to environment variables.<\/p>\n"}]}
{"QuestionId":42540522,"AnswerCount":1,"Tags":"<tensorflow><lstm>","CreationDate":"2017-03-01T19:39:03.207","AcceptedAnswerId":null,"OwnerUserId":1546029.0,"Title":"How to visualize BPTT gradients in tensorflow","Body":"<p>I'd like to view the gradients of my Tensorflow LSTM over time, for example, by plotting the norm of the gradient from t=N through t=0. The question is, how can I get the gradient per timestep out of Tensorflow?<\/p>\n","answers":[{"AnswerId":"42543862","CreationDate":"2017-03-01T23:06:59.770","ParentId":null,"OwnerUserId":"5951780","Title":null,"Body":"<p>In the graph define:<\/p>\n\n<pre><code>params = tf.trainable_variables()\nopt = get_optimizer(optimizer)(self.learning_rate)\ngradients = tf.gradients(self.losses, params)\nself.gradient_norm = tf.global_norm(gradients)\nself.updates = opt.apply_gradients(zip(gradients, params), global_step=self.global_step)\n<\/code><\/pre>\n\n<p>Then eval self.gradient_norm or feed it to a SummariesWriter<\/p>\n"}]}
{"QuestionId":42541088,"AnswerCount":2,"Tags":"<python-2.7><tensorflow>","CreationDate":"2017-03-01T20:10:07.883","AcceptedAnswerId":null,"OwnerUserId":3988991.0,"Title":"Difficulty Understanding TensorFlow Computations","Body":"<p>I'm new to TensorFlow and have difficulty understanding how the computations works. I could not find the answer to my question on the web.<\/p>\n\n<p>For the following piece of code, the last time I print \"d\" in the for loop of the \"train_neural_net()\" function, I'm expecting the values to be identical to when I print \"test_distance.eval\". But they are way different. Can anyone tell me why this is happening? Isn't TensorFlow supposed to cache the Variable results learned in the for loop and use them when I run \"test_distance.eval\"?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42542266","CreationDate":"2017-03-01T21:20:37.897","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Each time you call the functions <code>neural_network_model1()<\/code> or <code>neural_network_model2()<\/code>, you create a new set of variables, so there are four sets of variables in total.<\/p>\n\n<ul>\n<li>The call to <code>sess.run(tf.global_variables_initializer())<\/code> initializes all four sets of variables.<\/li>\n<li><p>When you train in the <code>for<\/code> loop, you only update the first two sets of variables, created with these lines:<\/p>\n\n<pre><code>prediction1 = neural_network_model1(x1)\nprediction2 = neural_network_model2(x2)\n<\/code><\/pre><\/li>\n<li><p>When you evaluate with <code>test_distance.eval()<\/code>, the tensor <code>test_distance<\/code> depends only on the variables that were created in the last two sets of variables, which were created with these lines:<\/p>\n\n<pre><code>test_result1 = neural_network_model1(x3)\ntest_result2 = neural_network_model2(x4)\n<\/code><\/pre>\n\n<p>These variables were never updated in the training loop, so the evaluation results will be based on the random initial values.<\/p><\/li>\n<\/ul>\n\n<p>TensorFlow does include some code for sharing weights between multiple calls to the same function, using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_scope\" rel=\"nofollow noreferrer\"><code>with tf.variable_scope(...):<\/code><\/a> blocks. For more information on how to use these, see the <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variable_scope\" rel=\"nofollow noreferrer\">tutorial on variables and sharing<\/a> on the TensorFlow website.<\/p>\n"},{"AnswerId":"42542428","CreationDate":"2017-03-01T21:31:18.527","ParentId":null,"OwnerUserId":"6203497","Title":null,"Body":"<p>You don't need to define two function for generating models, you can use <code>tf.name_scope<\/code>, and pass a model name to the function to use it as a prefix for variable declaration. On the other hand, you defined two variables for distance, first is <code>distance<\/code> and second is <code>test_distance<\/code> . But your model will learn from train data to minimize <code>cost<\/code> which is only related to first distance variable. Therefore, <code>test_distance<\/code> is never used and the model which is related to it, will never learn anything! Again there is no need for two distance functions. You only need one. When you want to calculate <strong>train distance<\/strong>, you should <strong>feed it with train data<\/strong> and when you want to calculate <strong>test distance<\/strong> you should <strong>feed it with test data<\/strong>.\nAnyway, if you want second distance to work, you should declare another <code>optimizer<\/code> for it and also you have to learn it as you have done for first one. Also you should consider the fact that models are learning base on their initial values and training data. Even if you feed both models with exactly same training batches, you can't expect to have exactly similar characteristics models since initial values for weights are different and this could cause falling into different local minimum of error surface. At the end notice that whenever you call <code>neural_network_model1<\/code> or <code>neural_network_model2<\/code> you will generate new weights and biases, because <code>tf.Variable<\/code> is generating new variables for you.<\/p>\n"}]}
{"QuestionId":42541323,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-01T20:22:14.153","AcceptedAnswerId":null,"OwnerUserId":7643546.0,"Title":"Error Running Tensorflow with XLA","Body":"<p>I'm trying to run a simple tensorflow example using JIT compilation, as shown in <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/experimental\/xla\/jit\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/experimental\/xla\/jit<\/a> . Instead of the mnist_softmax_xla sample, I'm using the following code:<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting an error:<\/p>\n\n<pre><\/pre>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"42727974","CreationDate":"2017-03-10T21:26:35.363","ParentId":null,"OwnerUserId":"7643546","Title":null,"Body":"<p>Turns out I was running with Python2. It now works when running with Python3<\/p>\n"}]}
{"QuestionId":42542893,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-03-01T21:57:43.490","AcceptedAnswerId":null,"OwnerUserId":7643851.0,"Title":"Using tf.py_func() as a reader on queues gives out of range errors","Body":"<p>I am trying to create an input pipeline using tf.py_func() as a reader on queues. My data is distributed as following:<\/p>\n\n<pre><\/pre>\n\n<p>I have a file that contains paths to these data files in the following format:<\/p>\n\n<pre><\/pre>\n\n<p>I am creating three queues out of them, and then using readers and decoders to parse the elements in queues. Since tensorflow has defined readers like tf.read_file() and decoders such as tf.image.decode_jpeg() and tf.image.decode_png() for <strong>jpg<\/strong> and <strong>png<\/strong> formats but none for <strong>npz<\/strong> format, so I am using a python wrapper to parse the queue for activations files (third column in the file above):<\/p>\n\n<pre><\/pre>\n\n<p>Now, if I try to construct a wrapper using <strong>tf.py_func()<\/strong> around the python function above and use it to as a reader, I ran into out of range errors:<\/p>\n\n<pre><\/pre>\n\n<p>Following is the invocation of different readers and decoders on input queues:<\/p>\n\n<pre><\/pre>\n\n<p>And ImageReader is defined as:<\/p>\n\n<pre><\/pre>\n\n<p>When I try to call ImageReader.dequeue(1), I ran into out of range error. I think I am not correctly using the python wrapper in this setting. Any suggestions?<\/p>\n\n<p>Below is a sample code to reproduce the problem:<\/p>\n\n<pre><\/pre>\n\n<p>where 'list.txt' simply consists of names of 'x' number of files and '.\/data' contains those files. For simplicity I used 'txt' files.<\/p>\n","answers":[]}
{"QuestionId":42543863,"AnswerCount":2,"Tags":"<windows><tensorflow><gpu><anaconda>","CreationDate":"2017-03-01T23:07:11.563","AcceptedAnswerId":null,"OwnerUserId":7644067.0,"Title":"installing tensorflow with gpu in windows fails with exception","Body":"<p>I have search all over and I now Im on the verge of just giving up. \nThe cpu version of tensorflow installs just fine but I want to use the gpu. I have before used theano with gpu with CUDA8 and cuDNN 5.1 so they should be ok.\nI have also tried to install the nightly build of tensorflow with the same result.<\/p>\n\n<p>Any suggestions on what to check next. I'm this close to throw it all away, install virtualbox and ubuntu and run everything from there (don't want to though because of prestigious reasons :) ).  <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42544134","CreationDate":"2017-03-01T23:29:55.733","ParentId":null,"OwnerUserId":"4847576","Title":null,"Body":"<p>It seems a problem with Anaconda itself and not with TensorFlow, could you please try:<\/p>\n\n<ol>\n<li>Uninstall TensorFlow<\/li>\n<li>Install TensorFlow with <code>conda install<\/code> <\/li>\n<\/ol>\n\n<p>Also see <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7556\" rel=\"nofollow noreferrer\">this issue on Github<\/a>.\nI'd suggest as well, but with a caveat because I didn't test this one, to download <a href=\"http:\/\/ci.tensorflow.org\/view\/Nightly\/job\/nightly-win\/99\/\" rel=\"nofollow noreferrer\">a nightly build<\/a> and install it instead of the PYPI package.<\/p>\n\n<p>edit: now that I saw @mrry already had pointed you the direction to take so that might work.<\/p>\n"},{"AnswerId":"42556617","CreationDate":"2017-03-02T13:20:18.823","ParentId":null,"OwnerUserId":"7644067","Title":null,"Body":"<p>You where both on the right track. I did a clean install of Anaconda3 and created two new environments. One for theano and one for tensorflow. <\/p>\n\n<p>With the clean install I could use<br>\n<code>pip install tensorflow-gpu<\/code><br>\nNow both theano and tensorflow works with GPU.<\/p>\n\n<p>Thanks <\/p>\n"}]}
{"QuestionId":42543894,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-03-01T23:10:19.233","AcceptedAnswerId":42544476.0,"OwnerUserId":5947872.0,"Title":"Can you save a Tensorflow variable as a string, or to a stream?","Body":"<p>I'm farmilliar with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Saver\" rel=\"nofollow noreferrer\">tf.train.Saver<\/a>, which allows me to save variables to disk.<\/p>\n\n<p>But is it possible to get them directly as a string? Or write them to an open stream? Something like:<\/p>\n\n<pre><\/pre>\n\n<p>Or do I have to write to disk, then read it back?<\/p>\n","answers":[{"AnswerId":"42544476","CreationDate":"2017-03-02T00:02:58.807","ParentId":null,"OwnerUserId":"6203497","Title":null,"Body":"<p>You can convert any tensor to Numpy array. Then use Numpy array as you wish! for example your model has <code>W_layer1<\/code> as weights for first layer. You can convert it to Numpy like : <code>W_numpy = W_layer1.eval()<\/code> which is mentioned also here: <a href=\"https:\/\/stackoverflow.com\/questions\/34097281\/how-can-i-convert-a-tensor-into-a-numpy-array-in-tensorflow\">Tensor to Numpy array<\/a><\/p>\n"}]}
{"QuestionId":42545379,"AnswerCount":0,"Tags":"<tensorflow><keras><loss>","CreationDate":"2017-03-02T01:43:18.403","AcceptedAnswerId":null,"OwnerUserId":1058511.0,"Title":"why do you need to compile a keras \/ tensorflow model that was loaded from a saved copy?","Body":"<p>I understand how you save model and weights files and load them back up. I don't understand why the compilation with optimizers<\/p>\n\n<pre><\/pre>\n\n<p>Since you are no longer during gradient descent, why do you need the loss function and optimizer?<\/p>\n","answers":[]}
{"QuestionId":42545790,"AnswerCount":0,"Tags":"<docker><tensorflow>","CreationDate":"2017-03-02T02:33:22.887","AcceptedAnswerId":null,"OwnerUserId":5270944.0,"Title":"How to access a Tensorflow docker instance from the \"outside\" without Jupyter - for distributed Tensorflow","Body":"<p>I have nvidia-docker and the tensorflow-gpu image functioning fine on Ubuntu 14.04, Python 2.7. <\/p>\n\n<p>At this point, I can only figure out how to access Tensorflow in two ways - by running the following command, I'm able to access Tensorflow from the host machine in Jupyter notebook with the port 8888.<\/p>\n\n<pre><\/pre>\n\n<p>By running the following command I'm able to run an interactive shell in which I can run Python scripts or the REPL.<\/p>\n\n<pre><\/pre>\n\n<p>My problem is that I wish to run Tensorflow distributed across a few computers, and as far as I'm aware Jupyter won't work well for this because code needs to be run in the notebook itself.\nThe interactive shell won't really work either because it doesn't map to any port or allow external connections.<\/p>\n\n<p>Has anyone had luck with running distributed Tensorflow from the docker image? How did you do it?<\/p>\n","answers":[]}
{"QuestionId":42546365,"AnswerCount":1,"Tags":"<tensorflow><neural-network><deep-learning>","CreationDate":"2017-03-02T03:35:10.847","AcceptedAnswerId":null,"OwnerUserId":7503833.0,"Title":"How to restore variables of a particular scope from a saved checkpoint in tensorflow?","Body":"<pre><\/pre>\n\n<p>But saver.restore only has options to restore the entire graph. I would like to restore only those variables that are in a specific scope. <\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42977504","CreationDate":"2017-03-23T13:29:45.100","ParentId":null,"OwnerUserId":"2781994","Title":null,"Body":"<p>Assume you have Google's model of InceptionNet in scope <code>InceptionV1<\/code> and you want to load it except for the last layer in scope <code>InceptionRetrained<\/code> you want to retrain.<\/p>\n\n<p>Assuming you already started retraining the last layer and you created <em>last_layer.ckpt<\/em> file by <code>saver2.save(session, 'last_layer.ckpt')<\/code>, here is how to restore the net from both checkpoints.<\/p>\n\n<pre><code>saver1 = tf.train.Saver(var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='InceptionV1'))\nsaver1.restore(session, 'inception_model_from_google.ckpt')\n\nsaver2 = tf.train.Saver(var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='InceptionRetrained'))\nsaver2.restore(session, 'last_layer.ckpt')\n<\/code><\/pre>\n\n<hr>\n\n<p>If you are retraining only the last layer, don't forget to disable propagation of the gradient up the network (saves time) by calling the optimizer with <code>var_list<\/code> argument.<\/p>\n\n<pre><code>tf.train.Optimizer(0.0001).minimize(\n            loss, var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Inceptionretrained'))\n<\/code><\/pre>\n"}]}
{"QuestionId":42546469,"AnswerCount":2,"Tags":"<python><python-2.7><numpy><matplotlib><tensorflow>","CreationDate":"2017-03-02T03:44:39.900","AcceptedAnswerId":42547869.0,"OwnerUserId":3813674.0,"Title":"How do I generate sharp images Cifar-10","Body":"<p>I'm working with tensorflow and trying to visualize the input\/output of an autoencoder on Cifar-10. <\/p>\n\n<p>I'm following this answer here: <a href=\"https:\/\/stackoverflow.com\/questions\/35995999\/why-cifar-10-images-are-not-displayed-properly-using-matplotlib\/40144107#40144107\">Why CIFAR-10 images are not displayed properly using matplotlib?<\/a><\/p>\n\n<p>Here's the result from running their code with slight modification (change the figsize to 5,5):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/W6hKI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/W6hKI.png\" alt=\"Visualized Images\"><\/a><\/p>\n\n<p>However, this is still nowhere as sharp and clear as the images in the original page: <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow noreferrer\">https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html<\/a><\/p>\n\n<p>How can I do better?<\/p>\n","answers":[{"AnswerId":"42547869","CreationDate":"2017-03-02T05:58:10.667","ParentId":null,"OwnerUserId":"3804434","Title":null,"Body":"<p>There might be two issues here:<\/p>\n\n<p><strong>Issue 1:<\/strong><\/p>\n\n<p>It looks like your color channels (red, green, blue) are mixed. That would explain why the colors are so weird. If that is the case you will need to swap the color channels in your array as I have shown below.<\/p>\n\n<pre><code>import numpy as np\nimport matplotlib.pyplot as plt\nfrom matplotlib.cbook import get_sample_data\n\nrgb_image = plt.imread(get_sample_data(\"grace_hopper.png\", asfileobj=False))\n\n# correct color channels (R, G, B)\nplt.figure()\nplt.imshow(rgb_image)\nplt.axis('off')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0F0jg.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0F0jg.png\" alt=\"natural color\"><\/a><\/p>\n\n<pre><code># swapped color channels (R, B, G)\nrgb_image = rgb_image[:, :, [0, 2, 1]]\nplt.figure()\nplt.imshow(rgb_image)\nplt.axis('off')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/SEpiS.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/SEpiS.png\" alt=\"swapped color channels\"><\/a><\/p>\n\n<p><strong>Issue 2:<\/strong><\/p>\n\n<p>Matplotlib's <code>plt.imshow<\/code> has a keyword argument <code>interpolation<\/code> which defaults to <code>None<\/code> if not specified. Matplotlib then refers to the your local style sheet to determine the default interpolation behavior. Depending on your style sheet, this may result in interpolation being applied and therefore distortion of your image. See the <a href=\"http:\/\/matplotlib.org\/api\/pyplot_api.html#matplotlib.pyplot.imshow\" rel=\"nofollow noreferrer\">documentation for imshow for more details<\/a>. <\/p>\n\n<p>If you want to guarantee that Matplotlib does not interpolate your image, you should specify <code>interpolation=\"none\"<\/code> in <code>plt.imshow<\/code>. This is confusing because the default NoneType value of <code>None<\/code> yields different behavior than the string value of <code>\"none\"<\/code>. <\/p>\n\n<pre><code>red = np.zeros((100, 100, 3), dtype=np.uint8)\nred[:, :, 0] = 255\nred[40:60, 40:60, :] = 255\n\n# with interpolation\nplt.figure()\nplt.imshow(red, interpolation='bicubic') \nplt.axis('off')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/cG9RB.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cG9RB.png\" alt=\"with interpolation\"><\/a>    <\/p>\n\n<pre><code># without interpolation\nplt.figure()\nplt.imshow(red, interpolation='none') \nplt.axis('off')\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/DTUDH.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DTUDH.png\" alt=\"without interpolation\"><\/a><\/p>\n"},{"AnswerId":"50910557","CreationDate":"2018-06-18T13:24:20.570","ParentId":null,"OwnerUserId":"5576121","Title":null,"Body":"<p>Maybe you should do something like this. The images are very small, of the size of 32px in height and width, hence the they will be sharper only when in the size of thumbnails. I have interpolated it using bicubic transformation here. But you can change it to 'none' so that instead of blurry, you will get a pixelated image.<\/p>\n\n<pre><code>def unpickle(file):\n    with open(file, 'rb') as fo:\n        dict1 = pickle.load(fo, encoding='bytes')\n    return dict1\n\npd_tr = pd.DataFrame()\ntr_y = pd.DataFrame()\n\nfor i in range(1,6):\n    data = unpickle('data\/data_batch_' + str(i))\n    pd_tr = pd_tr.append(pd.DataFrame(data[b'data']))\n    tr_y = tr_y.append(pd.DataFrame(data[b'labels']))\n    pd_tr['labels'] = tr_y\n\ntr_x = np.asarray(pd_tr.iloc[:, :3072])\ntr_y = np.asarray(pd_tr['labels'])\nts_x = np.asarray(unpickle('data\/test_batch')[b'data'])\nts_y = np.asarray(unpickle('data\/test_batch')[b'labels'])    \nlabels = unpickle('data\/batches.meta')[b'label_names']\n\ndef plot_CIFAR(ind):\n    arr = tr_x[ind]\n    R = arr[0:1024].reshape(32,32)\/255.0\n    G = arr[1024:2048].reshape(32,32)\/255.0\n    B = arr[2048:].reshape(32,32)\/255.0\n\n    img = np.dstack((R,G,B))\n    title = re.sub('[!@#$b]', '', str(labels[tr_y[ind]]))\n    fig = plt.figure(figsize=(3,3))\n    ax = fig.add_subplot(111)\n    ax.imshow(img,interpolation='bicubic')\n    ax.set_title('Category = '+ title,fontsize =15)\n\nplot_CIFAR(4)\n<\/code><\/pre>\n"}]}
{"QuestionId":42546538,"AnswerCount":1,"Tags":"<tensorflow><nlp><stanford-nlp><sentiment-analysis><syntaxnet>","CreationDate":"2017-03-02T03:52:24.517","AcceptedAnswerId":null,"OwnerUserId":825707.0,"Title":"How dependency parse tree is used for sentiment analysis?","Body":"<p>With the announcement from Google on release of Parsey McParseface <a href=\"https:\/\/research.googleblog.com\/2016\/05\/announcing-syntaxnet-worlds-most.html\" rel=\"nofollow noreferrer\">syntaxnet<\/a>\nwhich is claimed to be the most accurate dependency parser. I want to understand how this parser can be used for more accurate sentiment analysis ? If someone can share some blogs or research papers or tutorials which can help me understand the overall flow.<\/p>\n","answers":[{"AnswerId":"42548087","CreationDate":"2017-03-02T06:15:12.870","ParentId":null,"OwnerUserId":"4356204","Title":null,"Body":"<p>Good question, Im not expert, in fact I got intrigued when you asked the question. <\/p>\n\n<p><strong>td;lr;<\/strong> more accurate dependency parsers would in allow one propagate sentiment through a graph and thus better determine sentiment polarity, at least in theory.<\/p>\n\n<p>It seems from my reading that sentiment analysis using dependency tree graphs propagate the independent (prior -- sentiment you might get from a lexicon) sentiment of words to compose overall sentiment polarity of the text. <\/p>\n\n<p>This approach uses the composition of language (its grammatical structure) to determine sentiment. This is somewhat* opposed to a statistical (naives bayes, logistics regression, neural networks) approach to understanding sentiment.<\/p>\n\n<p>Here's the paper I scanned:<\/p>\n\n<p><a href=\"http:\/\/www.aaai.org\/ocs\/index.php\/FLAIRS\/FLAIRS14\/paper\/viewFile\/7869\/7837\" rel=\"nofollow noreferrer\">http:\/\/www.aaai.org\/ocs\/index.php\/FLAIRS\/FLAIRS14\/paper\/viewFile\/7869\/7837<\/a><\/p>\n\n<p>For a deeper exploration of whats possible, this might help:<\/p>\n\n<p><a href=\"https:\/\/arxiv.org\/pdf\/1401.6330.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1401.6330.pdf<\/a><\/p>\n\n<p>A more through introduction to dependency parsing if you're interested might be <a href=\"https:\/\/web.stanford.edu\/~jurafsky\/slp3\/14.pdf\" rel=\"nofollow noreferrer\">https:\/\/web.stanford.edu\/~jurafsky\/slp3\/14.pdf<\/a> <\/p>\n\n<p>*somewhat in the sense that (in particular) convolution networks do learn a certain composition of language so do rnns.<\/p>\n"}]}
{"QuestionId":42547456,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-02T05:23:31.507","AcceptedAnswerId":42585542.0,"OwnerUserId":null,"Title":"Are there any downsides of creating TensorFlow placeholders for variable sized vs. fixed sized inputs?","Body":"<p>I am wondering is there are any substantial downsides (e.g., regarding computational efficiency, memory...) in creating TensorFlow placeholders for variable sized inputs (vs. fixed sized inputs)?<\/p>\n\n<p>Say, I am doing mini-batch learning and initialize the graph with a placeholder where I assume a fixed batch_size upfront:<\/p>\n\n<pre><\/pre>\n\n<p>Alternatively, I can initialize the placeholder variable so that it accepts variable sized inputs:<\/p>\n\n<pre><\/pre>\n\n<p>I am not that familiar with the low level tensorflow implementations under the hood, but wouldn't the latter have to check dimensions, allocate memory, and create new arrays at each iteration to account for a case where my minibatch size changes during training? So, depending on the implementation, wouldn't that be computationally wasteful if I am working with a fixed batch dimension?<\/p>\n","answers":[{"AnswerId":"42585542","CreationDate":"2017-03-03T17:58:36.023","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>There is a definite tension between providing fully defined shapes (which can yield better performance) and allowing dimensions to vary (which makes the dataflow graph easier to reuse). As you suspect, there are some downsides to using variable-shaped <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a> ops to represent input in a TensorFlow model:<\/p>\n\n<ul>\n<li><p>TensorFlow is often able to simplify the dataflow graph when shapes are fully known. For example, a call to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/shape\" rel=\"noreferrer\"><code>tf.shape(x)<\/code><\/a> returns a <code>tf.Tensor<\/code> containing the true dynamic shape of a tensor <code>x<\/code>. If that shape is fully defined at graph construction time, TensorFlow will replace the shape computation with a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"noreferrer\"><code>tf.constant()<\/code><\/a>, and this will be used in constant folding optimizations to reduce the amount of work done at runtime.<\/p><\/li>\n<li><p>As an extreme case, the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/experimental\/xla\/\" rel=\"noreferrer\">XLA compiler<\/a> requires that all input tensor shapes be fully defined before code is generated, so that it can generate much more efficient kernel code where array bounds (etc.) are compile-time constants. XLA recompiles the kernel code for each combination of input shapes, so using fixed-size tensors will avoid the recompilation overhead.<\/p><\/li>\n<li><p>TensorFlow's memory allocator currently <em>does<\/em> allocate new arrays for every intermediate and output tensor at each call to <code>tf.Session.run()<\/code>. However, the underlying memory allocators (the BFC allocator for GPU memory, and <code>tcmalloc<\/code> or <code>jemalloc<\/code> for CPU memory) tend to perform better if they have a static distribution of allocation requests (since the requests can be satisfied from buffers that were recently freed).<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":42547718,"AnswerCount":3,"Tags":"<python><tensorflow><neural-network><deep-learning>","CreationDate":"2017-03-02T05:47:03.240","AcceptedAnswerId":42548510.0,"OwnerUserId":1496362.0,"Title":"Reader doesn't seem to be installed in TensorFlow","Body":"<p>I have installed TensorFlow through pip for python 2.7. <\/p>\n\n<p>I am trying to run some example RNN models, one of which requires the import of reader. <\/p>\n\n<pre><\/pre>\n\n<p>After reading a bit, I tried a popular suggestions: <\/p>\n\n<pre><\/pre>\n\n<p>I ran: <\/p>\n\n<pre><\/pre>\n\n<p>Isn't reader supposed to be part of tensorflow? How can I install this module? <\/p>\n\n<p>I need it because in the code I am trying to run, it includes commands like this: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43401204","CreationDate":"2017-04-13T20:31:43.543","ParentId":null,"OwnerUserId":"4152567","Title":null,"Body":"<ol>\n<li>Download the reader from github:\n<a href=\"https:\/\/raw.githubusercontent.com\/tensorflow\/models\/master\/tutorials\/rnn\/ptb\/reader.py\" rel=\"nofollow noreferrer\">https:\/\/raw.githubusercontent.com\/tensorflow\/models\/master\/tutorials\/rnn\/ptb\/reader.py<\/a><\/li>\n<\/ol>\n\n<p>and add the file path and module to your python project:<\/p>\n\n<pre><code>import os\n\nimport sys\n\nscriptpath = \"\/My\/Path\/To\/Module\/File\/\" \n\nsys.path.append(os.path.abspath(scriptpath))\n\nimport reader\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li><p>Otherwise, A file can be read using:<\/p>\n\n<p>from tensorflow import read_file<\/p><\/li>\n<\/ol>\n\n<p>instead of:<\/p>\n\n<pre><code>from tensorflow.models.rnn.ptb import reader\n<\/code><\/pre>\n"},{"AnswerId":"42548510","CreationDate":"2017-03-02T06:42:28.677","ParentId":null,"OwnerUserId":"1496362","Title":null,"Body":"<p>Tried installing tensorflow for python3, but same issue. <\/p>\n\n<p>So I just put this reader.py file in my main folder: <a href=\"http:\/\/programtalk.com\/vs2\/?source=python\/7331\/tensorflow_with_latest_papers\/reader.py\" rel=\"nofollow noreferrer\">http:\/\/programtalk.com\/vs2\/?source=python\/7331\/tensorflow_with_latest_papers\/reader.py<\/a> and that seems to work. <\/p>\n"},{"AnswerId":"42548892","CreationDate":"2017-03-02T07:07:18.463","ParentId":null,"OwnerUserId":"708620","Title":null,"Body":"<p>I don't think <code>reader.py<\/code> is part of <a href=\"http:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">http:\/\/github.com\/tensorflow\/tensorflow<\/a>, but it's included in <a href=\"http:\/\/github.com\/tensorflow\/models\" rel=\"nofollow noreferrer\">http:\/\/github.com\/tensorflow\/models<\/a>. It's used in  <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/ptb\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/tutorials\/rnn\/ptb<\/a> and there's a commit message 3 months ago for <code>reader.py<\/code> that says \"Moving example models from github.com\/tensorflow\/tensorflow to github.com\/tensorflow\/models\".<\/p>\n\n<p>In the example code <code>ptb_word_lm.py<\/code> of the tutorials\/rnn\/ptb, there's a <code>import reader<\/code> and I'm able to run <code>ptb_word_lm.py<\/code> successfully. <\/p>\n"}]}
{"QuestionId":42547865,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-03-02T05:57:54.797","AcceptedAnswerId":null,"OwnerUserId":3051460.0,"Title":"Is the first and second convolution equivalent","Body":"<p>I have the first .prototxt of a convolution as follows:<\/p>\n\n<pre><\/pre>\n\n<p>And the second .prototxt of a convolution as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The different between first and second convolution are the addition definition  and three times . Is the first and second convolution equivalent? Given that the input data is 3D data.<\/p>\n","answers":[]}
{"QuestionId":42548084,"AnswerCount":0,"Tags":"<python><neural-network><stanford-nlp><keras>","CreationDate":"2017-03-02T06:14:40.333","AcceptedAnswerId":null,"OwnerUserId":6354442.0,"Title":"Training inputs in keras","Body":"<p>I have word-vectors(Glove vectors) pre trained over general twitter dataset. I have a set of tweets and I want to retrain the word-vectors parallelly with the task in hand. This requires updating the inputs in the MLP. How do I retrain input in sequential MLP in keras?<\/p>\n\n<p>Edit :\nGlove pre-trained(on twitter) vector are already provided on their website . I have some users and their tweets and I have a classification task . So I am trying to convert words of the tweets into wordvectors and taking their average to get word vector representation of tweet. But as adopted in standard literature , I want to retrain the wordvectors on the new network(Word vectors are inputs now). How can I implement this in keras<\/p>\n","answers":[]}
{"QuestionId":42548392,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><skflow>","CreationDate":"2017-03-02T06:35:39.280","AcceptedAnswerId":null,"OwnerUserId":5217799.0,"Title":"How to debug skflow code (tensorflow) gmm_ops.py?","Body":"<p>Hi I am new to tensorflow. I want to debug Tensorflow (skflow) gmm_ops.py (Gaussian Mixture Model). I am getting ERROR:tensorflow:Model diverged with loss = NaN.\nHow should I do it ? Is there any example?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42635706","CreationDate":"2017-03-06T21:15:51.550","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Usually a NanLoss means something overflowed or underflowed during training. Things such as normalizing the examples or processing a subset of the data tend to help debugging what could have caused this.<\/p>\n"}]}
{"QuestionId":42549546,"AnswerCount":1,"Tags":"<text><tensorflow><readfile>","CreationDate":"2017-03-02T07:48:17.680","AcceptedAnswerId":42552380.0,"OwnerUserId":7457101.0,"Title":"reading textfile returning empty variable in tensorflow","Body":"<p>I have a text file which has 110 rows and 1024 columns of float values. I am trying to load the textfile and it doesnt read any thing. <\/p>\n\n<pre><\/pre>\n\n<p>printing the shape of the variable returns . <\/p>\n","answers":[{"AnswerId":"42552380","CreationDate":"2017-03-02T10:07:17.947","ParentId":null,"OwnerUserId":"2283175","Title":null,"Body":"<p>Firstly - <code>tf.shape(a) == []<\/code> doesn't mean that variable is empty. All scalars and strings have shape <code>[]<\/code>. <\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/dims_types\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/dims_types<\/a><\/p>\n\n<p>May be you can check \"rank\" instead - it would be 0 for scalars and strings.\nOther than that it looks like <code>string_input_producer<\/code> is a queue and it needs additional wiring to make ti work.<\/p>\n\n<p>Please try this<\/p>\n\n<pre><code>filename = '300_faults.txt'\nfilename_queue = tf.train.string_input_producer([filename])\nreader = tf.TextLineReader()\n_,a = reader.read(filename_queue)\n#x = np.loadtxt('300_faults.txt')  # working\n#a = tf.constant(x,tf.float32)     # working\n\nmodel = tf.initialize_all_variables()\nwith tf.Session() as session:\n    session.run(model)\n\n    coord = tf.train.Coordinator()\n    threads = tf.train.start_queue_runners(coord=coord)\n\n    print(session.run(tf.shape(a)))\n\n    print(session.run((a)))\n    coord.request_stop()\n    coord.join(threads)\n<\/code><\/pre>\n"}]}
{"QuestionId":42550998,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-03-02T09:06:34.247","AcceptedAnswerId":null,"OwnerUserId":3755376.0,"Title":"How to ignore log print a unused layer in caffe","Body":"<p>I have an prototxt as follows:<\/p>\n\n<pre><\/pre>\n\n<p>where the second layer contains the scale factor of 2 of the original data, I named it as . During training, I just used the , and  without using . Hence, when I run training, the value of  always printing in my terminal. How can I ignore this? This log looks like<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42551699","CreationDate":"2017-03-02T09:38:31.127","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><strong>General solution:<\/strong><br>\nIf you wan to silence a <code>\"top\"<\/code> layer you can use <a href=\"http:\/\/caffe.help\/manual\/layers\/silence.html\" rel=\"nofollow noreferrer\"><code>\"Silence\"<\/code><\/a> layer:<\/p>\n\n<pre><code>layer {\n  type: \"Silence\"\n  name: \"silence_this_layer_for_me\"\n  bottom: \"label_scale2\"\n}\n<\/code><\/pre>\n\n<p>And that's it! You won;t hear from <code>label_scale2<\/code> again<\/p>\n\n<p><strong>A solution for HDF5Data<\/strong><br>\nAlternatively, you do not have to expose all <code>datasets<\/code> of an <code>hdf5<\/code> input. You can simply <a href=\"https:\/\/stackoverflow.com\/a\/37419537\/1714410\">comment out<\/a> the redundant <code>\"top\"<\/code>: <\/p>\n\n<pre><code>layer {\n  name: \"data_scale2\"\n  type: \"HDF5Data\"\n  top: \"data_scale2\"\n  # top: \"label_scale2\" \n  include {\n    phase: TRAIN\n  }\n  hdf5_data_param {\n    source: \".\/train_scale2.txt\"\n    batch_size: 2\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":42551108,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-02T09:11:54.247","AcceptedAnswerId":null,"OwnerUserId":7646250.0,"Title":"I need to optimize tensorflow calling","Body":"<p>I am coding a reinforcement learning agent with the Python API of tensorflow. I need to call tensorflow millions of times during execution to evaluate tensors (with no training) that are fed from outside tensorflow, and much less often I need to send a large training set. The evaluation calls take more than 1 millisecond each (GPU GTX 970, i7-4790K CPU), and together they take more than half of my code running time. I would like to know if I could somehow reduce the time it takes tensorflow to evaluate data without training, or reduce the overhead of calling tensorflow.<\/p>\n\n<p>Something that may slow the whole evaluation process is that I use two inputs, I apply a convolutional layer to one and then concat to the other using tf.concat before applying some dense layers. Would this make the evaluation process particularly slow? <\/p>\n\n<p>Some example code:<\/p>\n\n<pre><\/pre>\n\n<p>Then, to evaluate my function, I call millions of times this:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42551282,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-02T09:20:34.703","AcceptedAnswerId":42555684.0,"OwnerUserId":7646272.0,"Title":"don't understand mnist example in tensorflow","Body":"<p>I don't understand the mnist example in 'deep mnist for experts' in Tensorflow.<\/p>\n\n<p>In order to build a deep network, we stack several layers of this type. The second layer will have 64 features for each 5x5 patch.<\/p>\n\n<pre><\/pre>\n\n<p>I don't know why outputchannel is 64.<\/p>\n\n<p>I think we need to 32*2 * 5*5 filter for 64 outputchannel, so<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p>first of all, i'm sorry for not good english\nmaybe you hard to understand my ask\nso i', write sudo code logic <\/p>\n\n<pre><\/pre>\n\n<p>i think it make 62 output feature using 32*2 filter and save memory\nwhat part is wrong?<\/p>\n\n<p>thank you<\/p>\n","answers":[{"AnswerId":"42559648","CreationDate":"2017-03-02T15:37:15.177","ParentId":null,"OwnerUserId":"6824263","Title":null,"Body":"<p>It's bigger because you are using your filter(5x5) and using it on every filter for every inputdata. So it would look like:\n32*2*1 (one because you are using one filter)<\/p>\n"},{"AnswerId":"42555684","CreationDate":"2017-03-02T12:34:30.880","ParentId":null,"OwnerUserId":"7219743","Title":null,"Body":"<p>The fact is that the output dimension is independent from that of the input.<\/p>\n\n<p>In fact each step of the convolution is a tensor product between W_conv2 (W) and any <code>(n * m)<\/code> portion of the input matrix (I): if W has dimension <code>(n * m * k * h)<\/code> and I <code>(n * m * k)<\/code>, the result of the product I*W is a vector with dimension <code>h<\/code>. <\/p>\n\n<p><code>n, m, and k<\/code> must be equal in I and W, but you don't have any limitation on <code>h<\/code>.<\/p>\n"}]}
{"QuestionId":42551391,"AnswerCount":2,"Tags":"<java><c><compilation><tensorflow><native>","CreationDate":"2017-03-02T09:25:25.847","AcceptedAnswerId":null,"OwnerUserId":1439333.0,"Title":"Tensorflow Java API in windows","Body":"<p>I was trying to configure Tensorflow API for java in windows.<\/p>\n\n<p>As per the read me\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/README.md\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/java\/README.md<\/a><\/p>\n\n<p>It says we have to build the native library will need to be built from source for windows. But its not have detailed instructions for it.\nAnyone has any luck getting it compiled?\nIs there any steps I can follow?<\/p>\n","answers":[{"AnswerId":"42552998","CreationDate":"2017-03-02T10:36:09.483","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>From the official <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"nofollow noreferrer\">website<\/a>:<\/p>\n\n<blockquote>\n  <p>We don't officially support building TensorFlow on Windows; however,\n  you may try to build TensorFlow on Windows if you don't mind using the\n  highly experimental Bazel on Windows or TensorFlow CMake build.<\/p>\n<\/blockquote>\n\n<p>Also, a related github issue: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/17<\/a>, where one user <em>rongjiecomputer<\/em> mentioned (in Sep 2016):<\/p>\n\n<blockquote>\n  <p>Tensorflow requires Bazel to build from source, If I am not mistaken,\n  Bazel is a build system like GNU Make, but not a compiler.<\/p>\n  \n  <p>Based on what I saw as I briefly scanned through the code, source code\n  of Tensorflow itself uses mostly standard C++ library for things like\n  threading so it should be no problem to compile on Windows, essential\n  third-party libraries it uses all have Windows support.<\/p>\n  \n  <p>Therefore, I think the main problem lies on lack of build method for\n  Windows rather than source code itself. There are works on using CMake\n  to build instead of Bazel but not complete yet. If someone can\n  translate Bazel build rules to CMake's, I think we will be able to\n  build it on Windows.<\/p>\n<\/blockquote>\n\n<p>So I assume, if Bazel Windows works as it should, you could follow the same steps as for building it on other platforms with Bazel. I haven't tried it out my-self though.<\/p>\n"},{"AnswerId":"42657749","CreationDate":"2017-03-07T20:30:26.067","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/ci_build\/windows\/libtensorflow_cpu.sh\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/tools\/ci_build\/windows\/libtensorflow_cpu.sh<\/a> for the script that will be used to build the release binaries for Windows in the next release of TensorFlow.<\/p>\n\n<p>That shows how to use <code>bazel<\/code> (and some hacks) to generate <code>tensorflow_jni.dll<\/code>.<\/p>\n\n<p><strong>UPDATE<\/strong>: The <a href=\"https:\/\/maven-badges.herokuapp.com\/maven-central\/org.tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">TensorFlow for Java maven package<\/a> now supports Windows and starting with release 1.1, binaries for <code>tensorflow_jni.dll<\/code> are also available.<\/p>\n"}]}
{"QuestionId":42551440,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-02T09:27:21.327","AcceptedAnswerId":null,"OwnerUserId":3832688.0,"Title":"Error when trying to build tensorflow","Body":"<p>I am currently trying to build tensorflow from source, as I will need to use the C++ API. As a first guideline I tried to follow the installing from source guideline at <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/get_started\/os_setup<\/a>.\nHowever, when I am running:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error<\/p>\n\n<pre><\/pre>\n\n<p>Did anyone else experienced this error and can give a hint on how to resolve it? <\/p>\n","answers":[{"AnswerId":"42554655","CreationDate":"2017-03-02T11:47:42.307","ParentId":null,"OwnerUserId":"3832688","Title":null,"Body":"<p>Okay I resolved it on my own.\nThe problem was, that I cloned the repository with Eclipse. This seems to cause some issue. Solution: Delete folder, clone it with git in terminal and everything works. <\/p>\n"}]}
{"QuestionId":42551628,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><neural-network><keras>","CreationDate":"2017-03-02T09:35:44.217","AcceptedAnswerId":42578612.0,"OwnerUserId":2473992.0,"Title":"On loading the saved Keras sequential model, my test data gives low accuracy in the beginning","Body":"<p>I am creating a simple sequential Keras model which will take 10k inputs in a batch of 100. Each input has 3 columns and the corresponding output is sum of that row.<\/p>\n\n<p>Sequential model has 2 layers- LSTM(Stateful=true) , Dense.<\/p>\n\n<p>Now, after compiling and fitting the model, I am saving it in 'model.h5' file.<\/p>\n\n<p>Then, I read the saved model, and call model.predict with a test data (size=10k , batch_size = 100).<\/p>\n\n<p><strong>Problem:<\/strong> the prediction doesn't work properly for first 400-500 inputs and for the rest its working perfectly fine with very low val_loss.<\/p>\n\n<p><strong>Case1:<\/strong> I make the LSTM layer Stateless(i.e. Stateful=False)\nIn this case Keras is providing very accurate outputs for all the test data.<\/p>\n\n<p><strong>Case2:<\/strong> Instead of saving and then reading again, if I directly apply model.predict on the model created, all the outputs are coming accurately.<\/p>\n\n<p>But, I need Stateful=True, also, I want to save my model and then resume work on that model later.<\/p>\n\n<p>1.Is there any way to solve this?<\/p>\n\n<p>2.Also, when I am providing test data, how is the model's accuracy increasing? ( because the first 400-500 tests provide inaccurate results and the rest are pretty accurate)<\/p>\n","answers":[{"AnswerId":"42578612","CreationDate":"2017-03-03T12:13:08.433","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Your problem seems to come from losing the hidden states of your cells. During model building they might be reset and this might cause the problem.<\/p>\n\n<p>So (it's a little bit cumbersome), but you could save and load also a <code>states<\/code> of your network:<\/p>\n\n<ol>\n<li><p>How to save? (assuming that <code>i-th<\/code> layer is a <code>recurrent<\/code>one):<\/p>\n\n<pre><code>hidden_state = model.layers[i].states[0].eval()\ncell_state = model.layers[i].states[0].eval()\n\nnumpy.save(\"some name\", hidden_state)\nnumpy.save(\"some other name\", cell_state)\n<\/code><\/pre><\/li>\n<li><p>Now when you can reload the hidden state, <a href=\"https:\/\/stackoverflow.com\/questions\/42415909\/initializing-lstm-hidden-state-tensorflow-keras\/42420389#42420389\">here<\/a> you can read on how to set the hidden state in a layer.<\/p><\/li>\n<\/ol>\n\n<p>Of course - it's the best to pack all of this methods in some kind of object and e.g. class constructor methods.<\/p>\n"}]}
{"QuestionId":42552540,"AnswerCount":3,"Tags":"<python><tensorflow><neural-network><recurrent-neural-network>","CreationDate":"2017-03-02T10:14:06.190","AcceptedAnswerId":42574342.0,"OwnerUserId":1496362.0,"Title":"TensorFLow: tf.contrib.rnn module object is not callable","Body":"<p>I am trying to run a tutorial (<a href=\"http:\/\/r2rt.com\/recurrent-neural-networks-in-tensorflow-ii.html\" rel=\"nofollow noreferrer\">http:\/\/r2rt.com\/recurrent-neural-networks-in-tensorflow-ii.html<\/a>) and have some issue with modules that have moved.<\/p>\n\n<p>I researched a bit and found that some classes have moved and fixed:<\/p>\n\n<pre><\/pre>\n\n<p>to<\/p>\n\n<pre><\/pre>\n\n<p>This works now.<\/p>\n\n<p>However, I have this function I can't seem to fix:<\/p>\n\n<pre><\/pre>\n\n<p>So I tried:<\/p>\n\n<pre><\/pre>\n\n<p>But that says: 'module' object is not callable<\/p>\n\n<p>I am using python3 (tried 2 as well), with a fresh installed TensorFlow 1.0.0 through pip3.<\/p>\n\n<p>I tried looking at the API, but it's not making much sense to me, as I am still going through the tutorial to try to understand what is happening. I would think maybe the arguments have changed, or maybe this is in a sub-function now? <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn<\/a><\/p>\n","answers":[{"AnswerId":"43414850","CreationDate":"2017-04-14T15:44:07.590","ParentId":null,"OwnerUserId":"4152567","Title":null,"Body":"<p>Tensor flow has some documentation here:\n<a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.rnn\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/contrib.rnn<\/a><\/p>\n\n<p>The rnn is part of the contrib module. So if you change it to tf.contrib.rnn is a good step forward but there is an error:\n    TypeError: 'module' object is not callable<\/p>\n\n<p>In the documentation we have several options for recurrent NN:<\/p>\n\n<pre><code>tf.contrib.rnn.static_rnn\ntf.contrib.rnn.static_state_saving_rnn\ntf.contrib.rnn.static_bidirectional_rnn\ntf.contrib.rnn.stack_bidirectional_dynamic_rnn\n<\/code><\/pre>\n\n<p>For me, <\/p>\n\n<pre><code>tf.contrib.rnn.static_rnn\n<\/code><\/pre>\n\n<p>worked.<\/p>\n"},{"AnswerId":"42555055","CreationDate":"2017-03-02T12:05:01.313","ParentId":null,"OwnerUserId":"7096238","Title":null,"Body":"<p>Maybe is important how you import it. Do like this:<\/p>\n\n<pre><code>from tensorflow.python.ops import rnn, rnn_cell\n<\/code><\/pre>\n\n<p>And then use the following:<\/p>\n\n<pre><code>cell = rnn_cell.BasicRNNCell(state_size)\n\nrnn_outputs, final_state = rnn.rnn(cell, rnn_inputs, initial_state=init_state)\n<\/code><\/pre>\n\n<p>I Use it this way with LSTM cell and it works :)<\/p>\n"},{"AnswerId":"42574342","CreationDate":"2017-03-03T08:47:44.813","ParentId":null,"OwnerUserId":"1496362","Title":null,"Body":"<p>I found (and ran) the <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"nofollow noreferrer\">TensorFlow<\/a> update script on my data. <\/p>\n\n<p>To fix the remaining errors, just changed tf.nn.rnn_cell. to tf.contrib.rnn and now it works!<\/p>\n"}]}
{"QuestionId":42553169,"AnswerCount":1,"Tags":"<deep-learning><theano><keras><conv-neural-network><lstm>","CreationDate":"2017-03-02T10:43:18.740","AcceptedAnswerId":42554282.0,"OwnerUserId":2499909.0,"Title":"CNN-LSTM in Keras: Dimension Error","Body":"<p>In keras 1.2.2, I have made a dataset which has these dimensions:<\/p>\n\n<ul>\n<li>X_train: (2000, 100, 32, 32, 3) <\/li>\n<li>y_train: (2000,1)<\/li>\n<\/ul>\n\n<p>Here, 2000 is the number of instances (batches of data), 100 is the number of samples in each batch, 32 is the image rows and cols, and 3 is the number of channels (RGB).<\/p>\n\n<p>I have written this code which applies an LSTM after a CNN. I used TimeDistributed layers and took the average of the output of LSTM to have something like this:\n<a href=\"https:\/\/i.stack.imgur.com\/vApwo.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vApwo.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I want LSTM to work on each batch, and then I take the average of the LSTM output on that batch. So, my total output (my labels) is a (2000,1) vector.<\/p>\n\n<p>I get this error: <\/p>\n\n<ul>\n<li>Error when checking model target: expected lambda_14 to have 2\ndimensions, but got array with shape (2000L, 100L, 1L)<\/li>\n<\/ul>\n\n<p>This is my code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42554282","CreationDate":"2017-03-02T11:30:53.583","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>If we follow the track of the shapes in your model, we get to <code>shape = (None, 100, 35)<\/code> after the <code>Dense()<\/code>. Then you feed that to an <code>LSTM()<\/code> which will return the whole sequence of hidden vectors of length 20 so you get a <code>shape = (None, 100, 20)<\/code>. Then you take the mean over the axis 1 so you get <code>shape = (None, 100, 1)<\/code>. \nThere is something wrong with the architecture of that network because your targets have the <code>shape = (None, 1)<\/code>. So either change <\/p>\n\n<pre><code>LSTM(output_dim=20, return_sequences=False) \n<\/code><\/pre>\n\n<p>and add a <code>Flatten()<\/code> layer after the merge. Or <code>Flatten()<\/code> after the merge and use a <code>Dense(1, activation='sigmoid')<\/code> to get your predictions? <\/p>\n\n<p>It depends on you but right now it cannot work.<\/p>\n"}]}
{"QuestionId":42553422,"AnswerCount":1,"Tags":"<deep-learning><caffe><backpropagation><pycaffe>","CreationDate":"2017-03-02T10:53:53.313","AcceptedAnswerId":42585953.0,"OwnerUserId":7533175.0,"Title":"Weight update of Siamese networks in Caffe","Body":"<p>By following this web-site <\/p>\n\n<blockquote>\n  <p><a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/siamese.html\" rel=\"nofollow noreferrer\">http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/siamese.html<\/a><\/p>\n<\/blockquote>\n\n<p>,\nI can use a Siamese network in Caffe, which shares the weights for each layer.<\/p>\n\n<p>But, I was wondering about how the Siamese network in Caffe updates their shared weights.\nTo be specific, if we have <\/p>\n\n<blockquote>\n  <p>input1 -> conv1(<strong>shared<\/strong>) -> output1<\/p>\n  \n  <p>input2 -> conv1(<strong>shared<\/strong>) -> output2 ===> contrastive loss (from output1 and output2),<\/p>\n<\/blockquote>\n\n<p>then, does Caffe just sums up the two gradients for conv1 from the first and second networks?<\/p>\n\n<p>Thanks for your response in advance. <\/p>\n","answers":[{"AnswerId":"42585953","CreationDate":"2017-03-03T18:22:46.130","ParentId":null,"OwnerUserId":"7043417","Title":null,"Body":"<p>You are correct, the diffs (gradients) of shared weights (all parameters with the same name) are accumulated.\nNote that you can not use different learn rate multipliers (lr_mult) for shared weights. Other features like momentum and weight decay should work like expected.<\/p>\n"}]}
{"QuestionId":42553436,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><caffe><face-detection>","CreationDate":"2017-03-02T10:54:22.503","AcceptedAnswerId":null,"OwnerUserId":1542594.0,"Title":"Caffe training loss is 0","Body":"<p>I'm training an alexnet .caffemodel with faceScrub dataset, I'm following <\/p>\n\n<p><a href=\"https:\/\/github.com\/guoyilin\/FaceDetection_CNN\" rel=\"nofollow noreferrer\">Face Detection<\/a><\/p>\n\n<p><a href=\"http:\/\/caffe.berkeleyvision.org\/gathered\/examples\/finetune_flickr_style.html\" rel=\"nofollow noreferrer\">Fine-Tuning<\/a><\/p>\n\n<p>Thing is that when I'm training the model I get this output:<\/p>\n\n<pre><\/pre>\n\n<p>and it continues the same.<\/p>\n\n<p>The only thing I am suspicious on is that in the Face Detection link train_val.prototxt it uses num_output: 2 in the fc8_flickr layer, so I have the .txt file with all the images in this format:<\/p>\n\n<pre><\/pre>\n\n<p>With 1 being the class containing a face. If I remove the 1, it gets stuck in Iteration 0, Testing net (#0).<\/p>\n\n<p>Any insight on this?<\/p>\n","answers":[]}
{"QuestionId":42553650,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-03-02T11:03:47.027","AcceptedAnswerId":null,"OwnerUserId":3755376.0,"Title":"Can we use Concate layer for two different size of input images?","Body":"<p>I am implement multiple scale input for my network<\/p>\n\n<pre><\/pre>\n\n<p>The scheme is very simple, just use same network architecture but different image scale. Can I use  layer for that task? Specially, the layer prototxt will be<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42553671,"AnswerCount":0,"Tags":"<python-2.7><tensorflow><recurrent-neural-network><language-model>","CreationDate":"2017-03-02T11:04:47.403","AcceptedAnswerId":null,"OwnerUserId":5047436.0,"Title":"Predicting a probability of a sentence using tensorflow","Body":"<p>I am using this pre-trained <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/lm_1b\/\" rel=\"noreferrer\">model of tensorflow<\/a> and trying to get a probability of a sentence. My primary task is, out of several sentences find a sentence with the largest probability. <\/p>\n\n<p>I am able to predict next words, using this code.<\/p>\n\n<pre><\/pre>\n\n<p>However, how can I get a probability of a sentence from a list of many sentences? <\/p>\n","answers":[]}
{"QuestionId":42553860,"AnswerCount":1,"Tags":"<python><tensorflow><anaconda><conda><tensorboard>","CreationDate":"2017-03-02T11:12:24.073","AcceptedAnswerId":42556238.0,"OwnerUserId":2256217.0,"Title":"Tensorflow install failed inside Anaconda in Windows Machine","Body":"<p>Well, I am so tired of trying to install  in , I have installed anaconda successfully, but when I am going to install form this <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">link<\/a>, Anaconda installation on Step 4, I am seeing some exception. \nI have executed this line of command <\/p>\n\n<pre><\/pre>\n\n<p>However it's terminating showing a red single line exception. <a href=\"https:\/\/i.stack.imgur.com\/cWAd4.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cWAd4.png\" alt=\"red line\"><\/a><\/p>\n\n<p>Later I have searched for some solution on here and there. I have found some solution too. In this <a href=\"https:\/\/stackoverflow.com\/questions\/33622613\/tensorflow-installation-error-not-a-supported-wheel-on-this-platform\">solution<\/a> is done for only  and  only, not for . Later I found somewhere to use the python version 2.7 so I have downgraded the  in lower version, however it was same. But this time I can able to to execute the MAC OX command successfully and in my prompt it shows that install successful. But when I had gone to import  then again an error is threading, like this <a href=\"https:\/\/i.stack.imgur.com\/rO4fD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/rO4fD.png\" alt=\"image\"><\/a><\/p>\n\n<p>I have also found the some other solutions <a href=\"https:\/\/stackoverflow.com\/questions\/37130489\/installing-tensorflow-with-anaconda-in-windows\">link1<\/a> <a href=\"https:\/\/stackoverflow.com\/questions\/33634008\/how-to-install-tensorflow\">link2<\/a>, however none of them are done in which process I'm approaching. I am in a quiet confusion what kind of error or exception it is? And how do I overcome it?<\/p>\n\n<p>Update: \nTensorflow has been installed successfully from this <a href=\"https:\/\/stackoverflow.com\/questions\/42474666\/tensorflow-installation-error-on-windows-7?rq=1\">link<\/a>, Now a different problem is being facing. When I am going to import tensorflow in my workspace it providing this error.  <a href=\"https:\/\/i.stack.imgur.com\/Kv3Uy.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Kv3Uy.png\" alt=\"enter image description here\"><\/a><a href=\"https:\/\/i.stack.imgur.com\/Okmc3.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Okmc3.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42556238","CreationDate":"2017-03-02T13:01:26.067","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>It looks like it cannot find the required CUDA libraries. Did you install the CUDA Toolkit 8.0 and cuDNN 5.1?<\/p>\n\n<p>You will have to install the CUDA Toolkit 8.0 e.g. from <a href=\"https:\/\/developer.nvidia.com\/cuda-toolkit\" rel=\"nofollow noreferrer\">here<\/a> and the cuDNN package from e.g. <a href=\"https:\/\/developer.nvidia.com\/cudnn\" rel=\"nofollow noreferrer\">here<\/a> and then put the <code>cuDNN-installation-path\/win10\/bin<\/code> on your path (depending on your windows version of course)<\/p>\n\n<p>Also the following two posts might help you out:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/42011070\/on-windows-running-import-tensorflow-generates-no-module-named-pywrap-tenso\">On Windows, running &quot;import tensorflow&quot; generates No module named &quot;_pywrap_tensorflow&quot; error<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/41007279\/tensorflow-on-windows-couldnt-open-cuda-library-cudnn64-5-dll\">TensorFlow on Windows: &quot;Couldn&#39;t open CUDA library cudnn64_5.dll&quot;<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":42553924,"AnswerCount":0,"Tags":"<python><neural-network><caffe><hdf5><lstm>","CreationDate":"2017-03-02T11:15:39.947","AcceptedAnswerId":null,"OwnerUserId":3812705.0,"Title":"hdf5 error when loading data","Body":"<p>I'm trying to get to understand (and build off) <a href=\"https:\/\/github.com\/BVLC\/caffe\/pull\/4681\" rel=\"nofollow noreferrer\">BVLC\/caffe PR#4681<\/a>.<\/p>\n\n<p>For this, i've written my own dummy data generation. However, when trying to run training, the file is not loaded at all, due to the following error<\/p>\n\n<blockquote>\n  <p>F0302 11:40:20.412214 10108 hdf5.cpp:50] Unsupported datatype class: H5T_ENUM<\/p>\n<\/blockquote>\n\n<p>Following is my code for the proto &amp; for the data gen itself :<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42555276,"AnswerCount":1,"Tags":"<python><memory><tensorflow>","CreationDate":"2017-03-02T12:15:15.850","AcceptedAnswerId":null,"OwnerUserId":1625695.0,"Title":"tensorflow - run optimizer op on a large batch","Body":"<p>Normally, we call the run command with the optimizer operation as input to update the trainable parameters of some model:<\/p>\n\n<pre><\/pre>\n\n<p>But when the batch size is large, X_batch can't fit in GPU memory.\nHow can I split this task to handle a large batch size ?<\/p>\n","answers":[{"AnswerId":"42558049","CreationDate":"2017-03-02T14:26:25.143","ParentId":null,"OwnerUserId":"7647835","Title":null,"Body":"<p>This depends mainly on your GPU memory size. However, it is hard to fit your entire dataset along with the model and its required operations (i.e. predicting probabilities). Thus, you would need to think of batching in a different perspective. I assume your code goes along these lines:<\/p>\n\n<pre><code># Model Definition    \nX = tf.placeholder(tf.float32, shape=[None, DIM,DIM,3], name='X')\ny = tf.placeholder(tf.float32, shape=[None, N_CLASSES], name='y')\n\n...\n\ntrain_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)\n\n...\n\n# Training your model\nsess.run([train_step], feed_dict={X: X_batch, y: y_batch})\n<\/code><\/pre>\n\n<p>Instead of using <code>X<\/code> and <code>y<\/code> as your <code>feed_dict<\/code> to the <code>train_step<\/code>. You could just accumulate the <code>cross_entropy<\/code> for all batches (i.e. for the entire dataset). Then, you can run the <code>train_step<\/code> once. For example:<\/p>\n\n<pre><code>cross_entropy_all = []\nfor X_batch, y_batch in batches_generator():\n    cross_entropy_all += sess.run([cross_entropy], feed_dict={X: X_batch, y: y_batch})\n\n# Numpy or Tensorflow equivalent for `vstack`\ncross_entropy_all = np.vstack(cross_entropy_all)\n\n# Run the optimizer on the entire dataset (not just on a specific batch)\nsess.run([train_step], feed_dict={cross_entropy: cross_entropy_all})\n<\/code><\/pre>\n\n<p>This should achieve your goal without running your GPU out of memory. The suggested approach runs the optimization step against all cross entropies. Thus, you don't need to feed X and y (that are used\/needed to produce the <code>cross_entropy<\/code> because it is already fed to the optimization step).<\/p>\n"}]}
{"QuestionId":42556919,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><keras>","CreationDate":"2017-03-02T13:33:48.510","AcceptedAnswerId":42581606.0,"OwnerUserId":860566.0,"Title":"Adding a variable into Keras\/TensorFlow CNN dense layer","Body":"<p>I was wondering if it is possible to add a variable into a convolutional neural network's dense layer (so as well as the connections from the previous convolutional layers, there would be an additional feature set that could be used for discriminatory purposes)? If this is possible, can anyone point me to an example\/documentation explaining how to do so?<\/p>\n\n<p>I am hoping to use Keras, but am happy to use TensorFlow if Keras is too restrictive.<\/p>\n\n<p>EDIT: In this case, the way that I would think that this should work is that I provide a list containing images and associated feature sets to the neural network (and during training the associated classifications).<\/p>\n\n<p>EDIT2: The architecture that I want looks something like:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42578004","CreationDate":"2017-03-03T11:43:28.770","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Ok, assuming that you have the <code>convoluton_model<\/code> you could do this in a following manner:<\/p>\n\n<pre><code>convolution_model = Flatten()(convolution_model) # if it wasn't flattened before\nstatic_features_input = Input(shape=(static_features_size,))\nblended_features = merge([convolution_model, static_features_input], mode='concat')\n... here you are defining a blending model with blended features as input\n<\/code><\/pre>\n\n<p><a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/#multi-input-and-multi-output-models\" rel=\"nofollow noreferrer\">Here<\/a> you could find an example on how to merge different inputs.<\/p>\n"},{"AnswerId":"42581606","CreationDate":"2017-03-03T14:35:45.170","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Indeed, as @Marcin said, you can use a merge layer.<\/p>\n\n<p>I advise you to use the Functionnal API for this. If you're not familiar with it, read <a href=\"https:\/\/keras.io\/getting-started\/functional-api-guide\/\" rel=\"nofollow noreferrer\">some doc here<\/a>.<\/p>\n\n<p>Here is your doodled network model using the keras API :<\/p>\n\n<pre><code>from keras.layers.core import *\nfrom keras.models import Model\n\n# this is your image input definition. You have to specify a shape. \nimage_input = Input(shape=(32,32,3))\n# Some more data input with 10 features (eg.)\nother_data_input = Input(shape=(10,))    \n\n# First convolution filled with random parameters for the example\nconv1 = Convolution2D(nb_filter = nb_filter1, nb_row = nb_row1, nb_col=_nb_col1, padding = \"same\", activation = \"tanh\")(image_input)\n# MaxPool it \nconv1 = MaxPooling2D(pool_size=(pool_1,pool_2))(conv1)\n# Second Convolution\nconv2 = Convolution2D(nb_filter = nb_filter2, nb_row = nb_row2, nb_col=_nb_col2, padding = \"same\", activation = \"tanh\")(conv1)\n# MaxPool it\nconv2  = MaxPooling2D(pool_size=(pool_1,pool_2))(conv2)\n# Flatten the output to enable the merge to happen with the other input\nfirst_part_output = Flatten()(conv2)\n\n# Merge the output of the convNet with your added features by concatenation\nmerged_model = keras.layers.concatenate([first_part_output, other_data_input])\n\n# Predict on the output (say you want a binary classification)\npredictions = Dense(1, activation ='sigmoid')(merged_model)\n\n# Now create the model\nmodel = Model(inputs=[image_input, other_data_input], outputs=predictions)\n# see your model \nmodel.summary()\n\n# compile it\nmodel.compile(optimizer='adamax', loss='binary_crossentropy')\n<\/code><\/pre>\n\n<p>There you go :) It is quite easy in the end, define how many inputs and outputs you want, just specify them in a list when you create the <code>Model<\/code> object. When you fit it, also feed them separately, in a list.<\/p>\n"}]}
{"QuestionId":42556972,"AnswerCount":2,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-03-02T13:36:31.560","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"ImportError with importing keras","Body":"<p>I am trying to import keras it gives me an error . <\/p>\n\n<pre><\/pre>\n\n<p>I really appreciate if someone can help me to solve this error.\nThanks.<\/p>\n","answers":[{"AnswerId":"43238670","CreationDate":"2017-04-05T18:12:50.403","ParentId":null,"OwnerUserId":"7148493","Title":null,"Body":"<p>Did you install tensorflow at all?\nCould you try installing the lastest version? There is how:<\/p>\n\n<pre><code>pip install git+git:\/\/github.com\/tensorflow\/tensorflow  \npip install git+git:\/\/github.com\/fchollet\/keras.git --upgrade\n<\/code><\/pre>\n\n<p>If this doesn't work I do recommand using theano as a backend, it worked well for me.\nThere is how to setup it easily:<\/p>\n\n<pre><code>pip install git+git:\/\/github.com\/Theano\/Theano  \npip install git+git:\/\/github.com\/fchollet\/keras.git --upgrade  \npip install tensor\n<\/code><\/pre>\n\n<p>Before you do anything else, start python and import keras. Once keras is imported a new file will be created. You can then execute this command:<\/p>\n\n<pre><code>vi .keras\/keras.json\n<\/code><\/pre>\n\n<p>Once the file is open, change tensorflow for theano and you are ready to go.<\/p>\n"},{"AnswerId":"46344059","CreationDate":"2017-09-21T12:33:30.570","ParentId":null,"OwnerUserId":"3697224","Title":null,"Body":"<p>Try: <code>sudo pip install tensorflow --upgrade<\/code><\/p>\n\n<p>You may be using an older version of TensorFlow.<\/p>\n"}]}
{"QuestionId":42556981,"AnswerCount":0,"Tags":"<python><caffe>","CreationDate":"2017-03-02T13:36:45.570","AcceptedAnswerId":null,"OwnerUserId":7596581.0,"Title":"demo.py not show any pictures","Body":"<p>I use python to test faster rcnn demo.py with windows cpu without cuda.In demo.py,there has a plt.show() at the end,but when I run demo.py, it didn't show any pictures.please see follows:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42557193,"AnswerCount":1,"Tags":"<python><tensorflow><autoencoder>","CreationDate":"2017-03-02T13:47:20.117","AcceptedAnswerId":null,"OwnerUserId":3121280.0,"Title":"tensorflow: pow gives a negative output","Body":"<p>I'm using tensorflow to build a simple autoencoder model, however there's this strange bug that I can't diagnose\nI have a loss function that looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>total loss is then calculated by:<\/p>\n\n<pre><\/pre>\n\n<p>now the problem is, when setting  to  the loss returns as a positive number and the models seems to train well, but when increasing reg_fac the loss decreases and reaches negative values and keeps decreasing<\/p>\n\n<p> is calculated as this for each autoencoder used:<\/p>\n\n<pre><\/pre>\n\n<p>where  is the encoder weights and  is the decoder weights.\nI know it's a stupid bug, but I can't find it.<\/p>\n\n<p>my complete code is uploaded here:\n<a href=\"https:\/\/github.com\/javaWarrior\/dltest\" rel=\"nofollow noreferrer\">https:\/\/github.com\/javaWarrior\/dltest<\/a><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42562422","CreationDate":"2017-03-02T17:49:05.433","ParentId":null,"OwnerUserId":"4500078","Title":null,"Body":"<p>I am not sure where your error is coming from but I believe that there are some problems with your autoencoder model in general. The simple model should look like this example taken from <a href=\"https:\/\/github.com\/jo8hua\/models\/blob\/master\/autoencoder\/autoencoder_models\/Autoencoder.py\" rel=\"nofollow noreferrer\">the tensorflow models repo<\/a>. <\/p>\n\n<pre><code>    # model\n    x = tf.placeholder(tf.float32, [None, len_input])\n    h = tf.nn.softplus(tf.matmul(x, w1) + b1)\n    xHat = tf.matmul(h, w2) + b\n\n    # cost\n    cost = 0.5 * tf.reduce_sum(tf.pow(xHat - x, 2.0))\n    optimizer = tf.train.AdamOptimizer().minimize(cost)\n<\/code><\/pre>\n\n<p><strong>As it pertains to the question<\/strong>, The key difference might be using reduce_sum() rather than reduce_mean(). I am not sure why you are wanting to use this. <\/p>\n\n<p>Also, the AdamOptimizer should handle the regularization for you. As a side note, if you are wanting to learn by doing the regularization from scratch I would recommend <a href=\"http:\/\/cs231n.github.io\/neural-networks-3\/\" rel=\"nofollow noreferrer\">this tutorial<\/a>.<\/p>\n"}]}
{"QuestionId":42558376,"AnswerCount":0,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-03-02T14:42:30.997","AcceptedAnswerId":null,"OwnerUserId":6828367.0,"Title":"mixture of experts using tensorflow","Body":"<p>I am trying to define a loss function for a mixture of experts neural net. \nI am working with python and tensorflow. <\/p>\n\n<p>Due to the complexity of the code, it is not possible to copy it all for a working copy (I apologize).<\/p>\n\n<p>what i have so far is <\/p>\n\n<pre><\/pre>\n\n<p>my main problem is:  \nI have 4 experts which return a tensor each (total of 4 tensors). However labels is only one tensor. so I know what im doing here is wrong but I dont know how to fix it. \n(I was also not sure if to use  or  in this case. <\/p>\n\n<p>so the question is how to define this loss function correctly <\/p>\n","answers":[]}
{"QuestionId":42559002,"AnswerCount":1,"Tags":"<python><deep-learning><keras><recurrent-neural-network><sequential>","CreationDate":"2017-03-02T15:09:09.710","AcceptedAnswerId":42560597.0,"OwnerUserId":6950549.0,"Title":"Keras sequential timedistributed model extrem result differences between 2 and 3 sequences","Body":"<p>I have 2 models whom I train on 2 basically the same selfmade datasets. One with sequence length 1 and one with sequence length 2. In the first case it conveges like a charm and practicaly figures out my generating procedure, the second case it does little better then chance. What do I do wrong? Anything could be helpfull.<\/p>\n\n<p><strong>data generating code<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Model<\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Results <\/strong><\/p>\n\n<pre><\/pre>\n\n<p><strong>Results <\/strong><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42560597","CreationDate":"2017-03-02T16:21:19.163","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>There is something wrong the formulation of the problem. I'm going to try to explain to you why your example cannot work and then you can make another one if you will. <\/p>\n\n<p>On the data part, when you produce the dataset :<\/p>\n\n<pre><code>for i_sequence in range(sequence):\n    if np.sum(x_train[i_sample,i_sequence,:]) &gt; 50:\n        y_train[i_sample,:,:] = np.array([0,1])\n    else:\n        y_train[i_sample,:,:] = np.array([1,0])\n<\/code><\/pre>\n\n<p>you define the target for the <strong>whole sequence<\/strong> only based on the last element of this sequence. The <code>y_train[i_sample,0,:]<\/code> will be overwritten by the last action of the loop since you update <code>y_train[i_sample,:,:]<\/code> everytime you go forward in the sequence.<\/p>\n\n<p>So : you have ONE target for the whole sequence, which depends only on the last element of this sequence.<\/p>\n\n<p>Now on the model part : <\/p>\n\n<p>Your model is only constitued of <code>TimeDistributed(Dense())<\/code> layers. As per <a href=\"https:\/\/keras.io\/layers\/wrappers\/#timedistributed\" rel=\"nofollow noreferrer\">definition<\/a>, this is a wrapper that applies <strong>the same<\/strong> dense layer on every element of your sequence. Those dense layers share weights, so the one that will be applied on the first element of your sequence is exactly the same than the one applied on the last one.<\/p>\n\n<p>Now if you think about it: to decide of the target to apply on the first element of your sequence, your network needs to know what's happening on the last element, since you defined the dataset this way. <\/p>\n\n<p>Imagine that one of your sequence -call it seq_i- is such that <\/p>\n\n<pre><code>np.sum(x_train[seq_i,0,:]) = 52\nnp.sum(x_train[seq_i,1,:]) = 49\n<\/code><\/pre>\n\n<p>then your target for this sequence is <\/p>\n\n<pre><code>y_train[seq_i,0] = [1,0]\ny_train[seq_i,1] = [1,0]\n<\/code><\/pre>\n\n<p>Suppose that the dense layer is capable of predicting target <code>[1,0]<\/code> if the input is &lt; 50, just like you want it for the second element of your sequence. Since you apply the same layer to the first element of the sequence, it will predict [0,1] for that element and get punished for it during the training phase. It will go back and forth and won't learn anything.<\/p>\n\n<p>Is it clear? <\/p>\n"}]}
{"QuestionId":42559026,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-02T15:10:12.527","AcceptedAnswerId":null,"OwnerUserId":7643965.0,"Title":"Can not get TensorFlow with XLA working","Body":"<p>Has anyone successfully tried the example of mnist_softmax_xla in Step #3 of <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/experimental\/xla\/jit\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/master\/experimental\/xla\/jit<\/a> I can not get it working. I am getting:<\/p>\n\n<pre><\/pre>\n\n<p>I build TensorFlow r1.0 from source with XLA enabled. Does anyone have any ideas here? Thanks!<\/p>\n","answers":[]}
{"QuestionId":42559222,"AnswerCount":8,"Tags":"<python><tensorflow>","CreationDate":"2017-03-02T15:18:43.380","AcceptedAnswerId":42559223.0,"OwnerUserId":2979749.0,"Title":"Installation of TensorFlow on windows 7 - 'pip3' is not recognized as an internal or external command,","Body":"<p>When following the Installing TensorFlow for Windows guide <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a>, after executing<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>It looks like pip3 isn't recognized at all (although PATH to python is set)<\/p>\n","answers":[{"AnswerId":"42559223","CreationDate":"2017-03-02T15:18:43.380","ParentId":null,"OwnerUserId":"2979749","Title":null,"Body":"<p>Run the following<\/p>\n\n<pre><code>python -m pip install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>Assuming python is working, TensorFlow should get installed (at least the \"Validate the installation\" step is green).<\/p>\n"},{"AnswerId":"52422505","CreationDate":"2018-09-20T10:02:51.833","ParentId":null,"OwnerUserId":"3917708","Title":null,"Body":"<p>The issue is your path in the cmd is not that of where your python scripts are placed. In this case you can either navigate to \"AppData\\Local\\Programs\\Python\\Python36\\Scripts\" in your terminal and then run the command or you can simply put C:\\Users\\Your User Name\\AppData\\Local\\Programs\\Python\\Python36\\Scripts to Path variable.<\/p>\n\n<p>Then re-launch your cmd and type \"pip3 install tensorflow\" and see the feel the happiness :)<\/p>\n"},{"AnswerId":"43280032","CreationDate":"2017-04-07T13:58:36.543","ParentId":null,"OwnerUserId":"6885805","Title":null,"Body":"<p>Before running <code>pip3 install --upgrade tensorflow<\/code>\nyou need check if you are using the correct Python 3.5 installation:<\/p>\n\n<pre><code>Python 3.5.2  [MSC v.1900 64 bit (AMD64)] on win32\n<\/code><\/pre>\n\n<p>Notice the <code>64 bit<\/code> part.<br>\nOtherwise, it gives the above error.<br>\nYou are going to install <code>tensorflow-1.0.1-cp35-cp35m-win_amd64.whl<\/code>, therefore double check your correct version (mostly this is happening you have both Python 2.7 and 3.5).<\/p>\n"},{"AnswerId":"43937364","CreationDate":"2017-05-12T12:05:34.670","ParentId":null,"OwnerUserId":"7022485","Title":null,"Body":"<p>This will work, if you are facing pip3 or pip is not recognized as an internal or external command issue on windows:<\/p>\n\n<ol>\n<li>From the desktop, right click the Computer icon.<\/li>\n<li>Choose Properties from the context menu.<\/li>\n<li>Click the Advanced system settings link.<\/li>\n<li>Click Environment Variables. In the section System Variables, find the PATH environment variable and select it. Click Edit.<\/li>\n<li>A new pop up will open. Variable name will remain Path. We will change the Variable value to the location of the folder where your python scripts folder is located. Find it. <\/li>\n<\/ol>\n\n<blockquote>\n  <p>For e.g. I changed its value to <code>C:\\Users\\rgupta6\\AppData\\Local\\Programs\\Python\\Python35\\Scripts<\/code><\/p>\n<\/blockquote>\n\n<ol start=\"6\">\n<li>Close all remaining windows. Reopen Command prompt window, and run your <code>pip3 install --upgrade tensorflow<\/code> command or <code>pip3 install tensorflow<\/code> command<\/li>\n<\/ol>\n"},{"AnswerId":"48720940","CreationDate":"2018-02-10T12:48:22.263","ParentId":null,"OwnerUserId":"8950361","Title":null,"Body":"<p>That is because you haven't setup the environment variable yet.<\/p>\n\n<p>Follow the steps by @rajesh <\/p>\n\n<p>I had the same problem and i found his answer helpful<\/p>\n\n<blockquote>\n  <ol>\n  <li>Right click on This PC > Select Properties<\/li>\n  <li>Select Advanced system settings on the left<\/li>\n  <li>In the dialog box select Environment Variables<\/li>\n  <li>In the system variables section select path and cllck on edit<\/li>\n  <li>Select new and enter the path where the python scripts are..<\/li>\n  <\/ol>\n<\/blockquote>\n\n<p>it is mostly in C:\\Users[your user name]\\AppData\\Local\\Programs\\Python\\Python36\\Scripts<\/p>\n\n<blockquote>\n  <ol start=\"6\">\n  <li>Then ok.. to all the boxes opened<\/li>\n  <li>Close cmd  if it is already open and now try installing tensorflow using pip again like this<\/li>\n  <\/ol>\n<\/blockquote>\n\n<pre><code> pip3 install --upgrade tensorflow\n<\/code><\/pre>\n"},{"AnswerId":"46653860","CreationDate":"2017-10-09T19:41:27.317","ParentId":null,"OwnerUserId":"8748243","Title":null,"Body":"<p>when installing python, on the install window, check the box \"Active path\"(something like that), which builds up a path link. So you can run \"pip3 install\" at anywhere.<\/p>\n"},{"AnswerId":"47601658","CreationDate":"2017-12-01T21:21:14.723","ParentId":null,"OwnerUserId":"5202410","Title":null,"Body":"<p>Typing the python command before that should do the trick. In my case (on Windows 8.1 with Python 3.6), I had to type 'py' instead of 'python' as follows: <\/p>\n\n<p>py -m pip install --upgrade tensorflow<\/p>\n\n<p>The answer depends on the system you're using.<\/p>\n"},{"AnswerId":"46925076","CreationDate":"2017-10-25T06:07:08.340","ParentId":null,"OwnerUserId":"8829540","Title":null,"Body":"<p>I just experienced the same issue, most likely you downloaded a zipped version of python, then unzipped it, and added it to $PATH just like me, python can work but pip3 cannot, and python -m pip cannot either. \nfix solution is to download a executable version of python, then follow common installation steps, pip3 is selected by default, then everything is OK now. <\/p>\n"}]}
{"QuestionId":42559550,"AnswerCount":0,"Tags":"<tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-03-02T15:32:55.650","AcceptedAnswerId":null,"OwnerUserId":1736294.0,"Title":"Extracting feature vectors from Inception V3 model","Body":"<p>Feature vectors are extracted from the \"avg_pool\" layer from an Inception V3 model. Do the feature vectors have different values if the extraction program is run on different machines\/OS eg. Windows-64-bit and Ubuntu-64-bit?<\/p>\n","answers":[]}
{"QuestionId":42559660,"AnswerCount":1,"Tags":"<theano><keras>","CreationDate":"2017-03-02T15:37:59.430","AcceptedAnswerId":null,"OwnerUserId":4818202.0,"Title":"keras\/theano loss or op that uses full network","Body":"<p>I am looking for a way to implement the following network structure (currently using Keras, might be theano however): <\/p>\n\n<p>Assume we're given some simple network, but it is not possible to compute the desired loss based on this output directly, instead another operation is needed and the loss will be defined based on the output of this operation. However, this operation does not only need the output of the network but the full network object (eg its gradient).<\/p>\n\n<p>How can this be done? I think the operation could be performed either in a custom layer on top of the network or in a custom loss function - but for neither version I see a way to access the full network. Any suggestions?<\/p>\n","answers":[{"AnswerId":"42572396","CreationDate":"2017-03-03T06:47:27.250","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<blockquote>\n  <p>Assume we're given some simple network, but it is not possible to compute the desired loss based on this output directly, instead another operation is needed and the loss will be defined based on the output of this operation. However, this operation does not only need the output of the network but the full network object (eg its gradient).<\/p>\n<\/blockquote>\n\n<p>Say, you have the following model.<\/p>\n\n<pre><code>import keras.applications.vgg16 as vgg16\n\nmodel = vgg16.VGG16(weights='imagenet')\nmodel.summary()\n<\/code><\/pre>\n\n<p>For example, now you want to delete the last layer of this model which is actually predicts a category (a vector of length 1000 because imagenet has 1000 categories) for the input image.<\/p>\n\n<pre><code># Remove last Linear\/Dense layer.\nmodel.layers.pop()\nmodel.outputs = [model.layers[-1].output]\nmodel.layers[-1].outbound_nodes = []\nmodel.summary()\n<\/code><\/pre>\n\n<p>Now, lets add a linear layer (with output size 10) to this model and use the output of the modified neural network model.<\/p>\n\n<pre><code>model.add(Dense(10, activation='softmax'))\nmodel.summary()\n<\/code><\/pre>\n\n<p>You will get a vector (of length 10) as an output from this model.<\/p>\n\n<p>You can compile and train the model using <code>model.compile()<\/code> and <code>model.fit()<\/code> functions. You can set what type of loss function you want to use to train the model.<\/p>\n"}]}
{"QuestionId":42559763,"AnswerCount":1,"Tags":"<machine-learning><deep-learning><caffe><directory-structure><labeling>","CreationDate":"2017-03-02T15:42:49.110","AcceptedAnswerId":42560903.0,"OwnerUserId":1014196.0,"Title":"Directory structure and labeling in Caffe","Body":"<p>I would like to check if my understanding in organizing my folders and labeling is correct regarding Caffe's way of doing it. <\/p>\n\n<p>My train directory structure looks like below:<\/p>\n\n<pre><\/pre>\n\n<p>where both a_person and not_a_person are directories. <\/p>\n\n<p>My train.txt file looks like below:<\/p>\n\n<pre><\/pre>\n\n<p>Likewise, my val directory structure looks like below:<\/p>\n\n<pre><\/pre>\n\n<p>where both a_person and not_a_person are directories. <\/p>\n\n<p>My val.txt file looks like below:<\/p>\n\n<pre><\/pre>\n\n<p>Referring to to ilsvrc12, the content of train.txt and val.txt (hence, its structure) is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>What has confused me is actually as opposed to what is structured in ilsvrc12, the val directory contains no subdirectories i.e. contains no grouping. Is there any effect on my classification? I am stuck at getting an accuracy which is constant of 0.5 (see <a href=\"https:\/\/stackoverflow.com\/questions\/42513183\/caffes-test-accuracy-during-validation-phase-being-constant-when-training-a-net#\">here<\/a>) which triggers me to check if I performed correctly regarding the labeling and folder structure.<\/p>\n\n<p>Any hints and suggestions are welcome.<\/p>\n","answers":[{"AnswerId":"42560903","CreationDate":"2017-03-02T16:34:38.170","ParentId":null,"OwnerUserId":"2773607","Title":null,"Body":"<p>No, the directory structure has no bearing on your evaluation at all. The evaluation step takes the image files in your <code>val.txt<\/code>, runs the image through, and compares the classification result with the number next to that image filename. The directories is for human-readability not computer-readability, so it should have no bearing on your 0.5 accuracy.<\/p>\n"}]}
{"QuestionId":42560143,"AnswerCount":1,"Tags":"<tensorflow><lstm>","CreationDate":"2017-03-02T15:59:47.780","AcceptedAnswerId":42562391.0,"OwnerUserId":5915270.0,"Title":"Cannot find documentary for many things in tf.nn.seq2seq class in the tensorflow","Body":"<p><strong>tf.nn.seq2seq.embedding_attention_seq2seq<\/strong><\/p>\n\n<p>I was searching  the above function in tensorflow api. But I counln't find that by the exact name. But similar function was there in the class <strong>tf.contrib.legacy_seq2seq.embedding_attention_seq2seq<\/strong>  .\nWhy they moved few from nn class to contrib class ?<\/p>\n","answers":[{"AnswerId":"42562391","CreationDate":"2017-03-02T17:47:38.693","ParentId":null,"OwnerUserId":"3844056","Title":null,"Body":"<p>Google <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"nofollow noreferrer\">just released TensorFlow 1.0 last week<\/a>, where the API has been updated, so things have been moved around. Because of the various contributions from the TF community, the contrib directory is reserved for this code, which has not yet been finalized for acceptance into core TF. <\/p>\n\n<p>Check out the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/README.md\" rel=\"nofollow noreferrer\">README<\/a> on contrib:<\/p>\n\n<blockquote>\n  <p>The contrib directory contains project directories, each of which has\n  designated owners. It is meant to contain features and contributions\n  that eventually should get merged into core TensorFlow, but whose\n  interfaces may still change, or which require some testing to see\n  whether they can find broader acceptance. We are trying to keep\n  duplication within contrib to a minimum, so you may be asked to\n  refactor code in contrib to use some feature inside core or in another\n  project in contrib rather than reimplementing the feature.<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42560209,"AnswerCount":4,"Tags":"<tensorflow>","CreationDate":"2017-03-02T16:03:06.923","AcceptedAnswerId":null,"OwnerUserId":2191652.0,"Title":"Tensorflow: How to feed a placeholder variable with a tensor?","Body":"<p>I have a placeholder variable that expects a batch of input images: <\/p>\n\n<pre><\/pre>\n\n<p>Now I have 2 sources for the input data:<br>\n1) a tensor and<br>\n2) some numpy data.  <\/p>\n\n<p>For the numpy input data, I know how to feed data to the placeholder variable:<\/p>\n\n<pre><\/pre>\n\n<p>How can I feed a tensor to that placeholder variable? <\/p>\n\n<pre><\/pre>\n\n<p>gives me an error:<\/p>\n\n<pre><\/pre>\n\n<p>I don't want to convert the tensor into a numpy array using , since that would slow my program down, is there any other way?<\/p>\n","answers":[{"AnswerId":"42561363","CreationDate":"2017-03-02T16:55:50.503","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>This works for me in latest version...maybe you have older version of TF?<\/p>\n\n<pre><code>a = tf.Variable(1)\nsess.run(2*a, feed_dict={a:5}) # prints 10\n<\/code><\/pre>\n"},{"AnswerId":"43643424","CreationDate":"2017-04-26T19:55:25.807","ParentId":null,"OwnerUserId":"2030627","Title":null,"Body":"<p>You can use <code>feed_dict<\/code> to feed data into non-placeholders.  So, first, wire up your dataflow graph directly to your <code>myInputTensor<\/code> tensor data source (i.e. don't use a placeholder).  Then when you want to run with your numpy data you can effectively mask <code>myImportTensor<\/code> with <code>myNumpyData<\/code>, like this:<\/p>\n\n<pre><code>mLoss, = sess.run([loss], feed_dict={myImportTensor: myNumpyData})\n<\/code><\/pre>\n\n<p>[I'm still trying to figure out how to do this with multiple tensor data sources however.]<\/p>\n"},{"AnswerId":"46804533","CreationDate":"2017-10-18T07:07:18.703","ParentId":null,"OwnerUserId":"3552975","Title":null,"Body":"<p>This has been discussed on GitHub in 2016, and please check <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3389\" rel=\"nofollow noreferrer\">here<\/a>. Here is the key point by <a href=\"https:\/\/github.com\/concretevitamin\" rel=\"nofollow noreferrer\">concretevitamin<\/a>: <\/p>\n\n<blockquote>\n  <p>One key thing to note is that Tensor is simply a symbolic object. The values of your feed_dict are the actual values, e.g. a Numpy ndarry.<\/p>\n<\/blockquote>\n\n<p>The tensor as a symbolic object is flowing in the graph while the actual values are outside of it, then we can only pass the actual values into the graph and the symbolic object can not exist outside the graph. <\/p>\n"},{"AnswerId":"52079406","CreationDate":"2018-08-29T14:06:24.970","ParentId":null,"OwnerUserId":"10025506","Title":null,"Body":"<p>One way of solving the problem is to actually remove the Placeholder tensor and replace it by your \"myInputTensor\".<\/p>\n\n<p>You will use the myInputTensor as the source for the other operations in the graph and when you want to infer the graph with your np array as input data, you will feed a value to this tensor directly.<\/p>\n\n<p>Here is a quick example:<\/p>\n\n<pre><code>import tensorflow as tf \nimport numpy as np\n\n# Input Tensor\nmyInputTensor = tf.ones(dtype=tf.float32, shape=1) # In your case, this would be the results of some ops\n\noutput = myInputTensor * 5.0\n\nwith tf.Session() as sess:\n    print(sess.run(output)) # == 5.0, using the Tensor value\n    myNumpyData = np.zeros(1)\n    print(sess.run(output, {myInputTensor: myNumpyData}) # == 0.0 * 5.0 = 0.0, using the np value\n<\/code><\/pre>\n"}]}
{"QuestionId":42560245,"AnswerCount":1,"Tags":"<tensorflow><sparse-matrix><matrix-factorization>","CreationDate":"2017-03-02T16:04:31.157","AcceptedAnswerId":42562650.0,"OwnerUserId":4584564.0,"Title":"Use coo_matrix in TensorFlow","Body":"<p>I'm doing a Matrix Factorization in TensorFlow, I want to use coo_matrix from Spicy.sparse cause it uses less memory and it makes it easy to put all my data into my matrix for training data.<\/p>\n\n<p>Is it possible to use coo_matrix to initialize a variable in tensorflow?<\/p>\n\n<p>Or do I have to create a session and feed the data I got into tensorflow using sess.run() with feed_dict.<\/p>\n\n<p>I hope that you understand my question and my problem otherwise comment and i will try to fix it.<\/p>\n","answers":[{"AnswerId":"42562650","CreationDate":"2017-03-02T18:01:34.537","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The closest thing TensorFlow has to <a href=\"https:\/\/docs.scipy.org\/doc\/scipy\/reference\/generated\/scipy.sparse.coo_matrix.html\" rel=\"noreferrer\"><code>scipy.sparse.coo_matrix<\/code><\/a> is <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/SparseTensor\" rel=\"noreferrer\"><code>tf.SparseTensor<\/code><\/a>, which is the sparse equivalent of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor\" rel=\"noreferrer\"><code>tf.Tensor<\/code><\/a>. It will probably be easiest to feed a <code>coo_matrix<\/code> into your program.<\/p>\n\n<p>A <code>tf.SparseTensor<\/code> is a slight generalization of COO matrices, where the tensor is represented as three dense <code>tf.Tensor<\/code> objects:<\/p>\n\n<ul>\n<li><code>indices<\/code>: An <code>N<\/code> x <code>D<\/code> matrix of <code>tf.int64<\/code> values in which each row represents the coordinates of a non-zero value. <code>N<\/code> is the number of non-zeroes, and <code>D<\/code> is the rank of the equivalent dense tensor (2 in the case of a matrix).<\/li>\n<li><code>values<\/code>: A length-<code>N<\/code> vector of values, where element <code>i<\/code> is the value of the element whose coordinates are given on row <code>i<\/code> of <code>indices<\/code>.<\/li>\n<li><code>dense_shape<\/code>: A length-<code>D<\/code> vector of <code>tf.int64<\/code>, representing the shape of the equivalent dense tensor.<\/li>\n<\/ul>\n\n<p>For example, you could use the following code, which uses <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sparse_placeholder\" rel=\"noreferrer\"><code>tf.sparse_placeholder()<\/code><\/a> to define a <code>tf.SparseTensor<\/code> that you can feed, and a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/SparseTensorValue\" rel=\"noreferrer\"><code>tf.SparseTensorValue<\/code><\/a> that represents the actual value being fed :<\/p>\n\n<pre><code>sparse_input = tf.sparse_placeholder(dtype=tf.float32, shape=[100, 100])\n# ...\ntrain_op = ...\n\ncoo_matrix = scipy.sparse.coo_matrix(...)\n\n# Wrap `coo_matrix` in the `tf.SparseTensorValue` form that TensorFlow expects.\n# SciPy stores the row and column coordinates as separate vectors, so we must \n# stack and transpose them to make an indices matrix of the appropriate shape.\ntf_coo_matrix = tf.SparseTensorValue(\n    indices=np.array([coo_matrix.rows, coo_matrix.cols]).T,\n    values=coo_matrix.data,\n    dense_shape=coo_matrix.shape)\n<\/code><\/pre>\n\n<p>Once you have converted your <code>coo_matrix<\/code> to a <code>tf.SparseTensorValue<\/code>, you can feed <code>sparse_input<\/code> with the <code>tf.SparseTensorValue<\/code> directly:<\/p>\n\n<pre><code>sess.run(train_op, feed_dict={sparse_input: tf_coo_matrix})\n<\/code><\/pre>\n"}]}
{"QuestionId":42560998,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-03-02T16:38:53.057","AcceptedAnswerId":null,"OwnerUserId":6343552.0,"Title":"Tensorflow classification labels datatype","Body":"<p>I am using Tensorflow DNN model to do some <strong>classification<\/strong>.<\/p>\n\n<p>I have a numerical (float32) data input but <strong>string type output<\/strong>.<\/p>\n\n<pre><\/pre>\n\n<p>When I try to define the loss and optimizer as below:<\/p>\n\n<pre><\/pre>\n\n<p>I encounter an error that <\/p>\n\n<blockquote>\n  <p>TypeError: sigmoid_cross_entropy_with_logits() got an unexpected\n  keyword argument 'labels'<\/p>\n<\/blockquote>\n\n<p>I looked up the document from <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn\/classification\" rel=\"nofollow noreferrer\">here<\/a>, it said that<\/p>\n\n<blockquote>\n  <p>logits and targets must have the same type and shape.<\/p>\n<\/blockquote>\n\n<p>Do I need to convert the class into a floating number(hashing string to number)?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42569910","CreationDate":"2017-03-03T03:05:56.907","ParentId":null,"OwnerUserId":"2337467","Title":null,"Body":"<p>If you are using tensorflow v 0.10 the keyword for labels is targets. \nIf you are using tensorflow v 1.0 then the keyword for labels is labels. \nAs per the documentation your label vector should have the same type as logits, float 32 or float64.<\/p>\n"}]}
{"QuestionId":42561397,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-02T16:57:18.233","AcceptedAnswerId":null,"OwnerUserId":2638485.0,"Title":"Determining Which Tensorflow Op Implementation is Used","Body":"<p>In my Tensorflow graph I have a  op. How can I tell which implementation is actually used when the graph is run? Tensorflow has <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/0be81439c91e297b078152dd0c266471b24bde7f\/tensorflow\/core\/kernels\/conv_ops.cc\" rel=\"nofollow noreferrer\">a number of implementations of this Op available<\/a> and chooses one based on various optimization criteria, device and installed libs.<\/p>\n","answers":[]}
{"QuestionId":42561596,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><linear-regression>","CreationDate":"2017-03-02T17:05:17.247","AcceptedAnswerId":null,"OwnerUserId":51197.0,"Title":"Tensorflow: Linear model diverges when adding examples to the training vector","Body":"<p>First, apologies if I'm making a silly mistake - this is one of my first tensorflow programs. To my defense - I've searched the subject and did not find any reasonable explanation.<\/p>\n\n<h1>The code<\/h1>\n\n<p>The code is adapted from the <a href=\"https:\/\/www.tensorflow.org\/get_started\/get_started\" rel=\"nofollow noreferrer\">Getting Started guide<\/a>. It uses a gradient  to train a simple linear model:<\/p>\n\n<pre class=\"lang-python prettyprint-override\"><\/pre>\n\n<h1>Results<\/h1>\n\n<p>When the model is trained for 100 times, the result is:<\/p>\n\n<pre><\/pre>\n\n<p>When the model is trained 1000 times, the result is closer to (-1, 1), as expected:<\/p>\n\n<pre><\/pre>\n\n<h1>The problem<\/h1>\n\n<p>When adding another example to the training sample:<\/p>\n\n<pre><\/pre>\n\n<p>The results diverge (100 and 1000 iterations, respectively):<\/p>\n\n<pre><\/pre>\n\n<p><strong>Why does adding another example fail the model?<\/strong><\/p>\n","answers":[{"AnswerId":"42566755","CreationDate":"2017-03-02T21:55:43.950","ParentId":null,"OwnerUserId":"6203497","Title":null,"Body":"<p>By adding new single data, due to short size of previous data, you are increasing number of data about 15% ! This is a big change especially when your new data is very different from samples in current train data. You can not solve the problem by only increasing learning iteration, since gradient step can not compensate fluctuation in train data. Just decrease gradient step which will result in slower adopting to training data.<\/p>\n"}]}
{"QuestionId":42562633,"AnswerCount":2,"Tags":"<tensorflow><computer-vision><deep-learning><torch><tensor>","CreationDate":"2017-03-02T18:00:45.093","AcceptedAnswerId":42562976.0,"OwnerUserId":4480180.0,"Title":"What is a 2D float tensor?","Body":"<p><strong>Disclamer:<\/strong> I know <strong>nothing<\/strong> about CNN and deep learning and I don't know <a href=\"http:\/\/torch.ch\/\" rel=\"nofollow noreferrer\">Torch<\/a>.<\/p>\n\n<p>I'm using <a href=\"http:\/\/docs.opencv.org\/3.1.0\/da\/df5\/tutorial_py_sift_intro.html\" rel=\"nofollow noreferrer\">SIFT<\/a> for my object recognition application. I found this paper <a href=\"http:\/\/www.cv-foundation.org\/openaccess\/content_iccv_2015\/html\/Simo-Serra_Discriminative_Learning_of_ICCV_2015_paper.html\" rel=\"nofollow noreferrer\">Discriminative Learning of Deep Convolutional Feature Point Descriptors<\/a> which is particularly interesting because it's CNN based, which are more precise than classic image descripion methods (e.g. SIFT, SURF etc.), but (quoting the abstract):<\/p>\n\n<blockquote>\n  <p>using the L2 distance during both training and testing we develop\n  128-D descriptors whose euclidean distances reflect patch similarity,\n  and which can be used as a drop-in replacement for any task involving\n  SIFT<\/p>\n<\/blockquote>\n\n<p>Wow, that's fantastic: that means that we can continue to use any SIFT based approach but with more precise descriptors! <\/p>\n\n<p>However, quoting the <a href=\"https:\/\/github.com\/cvlab-epfl\/deepdesc-release\" rel=\"nofollow noreferrer\">github code repository<\/a> README:<\/p>\n\n<blockquote>\n  <p>Note the output will be a Nx128 2D float tensor where each row is a\n  descriptor.<\/p>\n<\/blockquote>\n\n<p>Well, what is a \"2D float tensor\"? SIFT descriptors matrix is Nx128 floats, is there something that I am missing?<\/p>\n","answers":[{"AnswerId":"42562976","CreationDate":"2017-03-02T18:19:15.833","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>2D float tensor = 2D float matrix.<\/p>\n\n<p>FYI: <a href=\"https:\/\/stats.stackexchange.com\/q\/181556\/12359\">The meaning of tensors in the neural network community<\/a><\/p>\n"},{"AnswerId":"43487266","CreationDate":"2017-04-19T05:39:33.657","ParentId":null,"OwnerUserId":"3907250","Title":null,"Body":"<p>This is a 2-d float tensor.<\/p>\n\n<pre><code>[[1.0,2.0],\n [3.0,4.0]]\n<\/code><\/pre>\n\n<p>This is still a 2-d float tensor, even if they have 3 items, and 3 rows! <\/p>\n\n<pre><code>[[1.0,2.0,3.0],\n [4.0,5.0,6.0],\n [7.0,5.0,6.0]]\n<\/code><\/pre>\n\n<p>The number of bracket is what matters.<\/p>\n"}]}
{"QuestionId":42562837,"AnswerCount":1,"Tags":"<python><tensorflow><typeerror><conv-neural-network>","CreationDate":"2017-03-02T18:11:51.463","AcceptedAnswerId":null,"OwnerUserId":7649177.0,"Title":"Can not convert a function into a tensor or operation","Body":"<p>I am constantly getting an error for a convolutional network I'm building and I don't know why the error is popping up. Here's my network:<\/p>\n\n<pre><\/pre>\n\n<p>Somewhere in the code, I guess sess.run(init) is giving me the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any idea where this error might be coming from? Thanks!<\/p>\n","answers":[{"AnswerId":"42563226","CreationDate":"2017-03-02T18:32:51.907","ParentId":null,"OwnerUserId":"7649177","Title":null,"Body":"<p>Oops. It turns out I just needed to change <\/p>\n\n<pre><code>init = tf.global_variables_initializer \n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>init = tf.global_variables_initializer()\n<\/code><\/pre>\n"}]}
{"QuestionId":42563289,"AnswerCount":1,"Tags":"<machine-learning><tensorflow>","CreationDate":"2017-03-02T18:36:20.123","AcceptedAnswerId":42563673.0,"OwnerUserId":7636462.0,"Title":"The tensorflow seems not to get installed","Body":"<p>Getting this same result:<\/p>\n\n<p>pip3 install --upgrade tensorflow\nCollecting tensorflow\n  Could not find a version that satisfies the requirement tensorflow (from versions: )\nNo matching distribution found for tensorflow<\/p>\n\n<p>P.S: I using Python 3.5.2. <\/p>\n","answers":[{"AnswerId":"42563673","CreationDate":"2017-03-02T18:55:38.607","ParentId":null,"OwnerUserId":"3734670","Title":null,"Body":"<p>In the moment Tensorflow is only available as 64 bit version for Python 3.5 in Windows.<\/p>\n\n<p>If you haven't the newest version of pip, update pip:<\/p>\n\n<pre><code>pip3 install --upgrade pip\n<\/code><\/pre>\n"}]}
{"QuestionId":42563513,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><classification>","CreationDate":"2017-03-02T18:48:01.557","AcceptedAnswerId":null,"OwnerUserId":5947872.0,"Title":"What is the preferred way of adding a \"None\" class to a classifier?","Body":"<p>Suppose I have some classifier where the model is just  which I then softmax to get probabilities, like this: <\/p>\n\n<pre><\/pre>\n\n<p>but I want it to look more like this:<\/p>\n\n<pre><\/pre>\n\n<p>I know I could add a none class with random feature sets, but that seems hacky and time consuming. I'm not even sure that would be accurate - since a real world say, bear is probably closer to dog bark than random data. Is there some preferred method to accomplish a none class, or some similar affect?<\/p>\n\n<p>(Also ideally not fudging the probability that it IS none)<\/p>\n","answers":[{"AnswerId":"42565170","CreationDate":"2017-03-02T20:19:24.457","ParentId":null,"OwnerUserId":"2658050","Title":null,"Body":"<p>If you do not have access to the real examples of \"none\" class and you do not want hacky solutions (like generating them, suggested in the question), the only approach remaining is to learn a generative model instead of discriminative one. In other words, if you learn P(x|y) and P(y) for each (observed) y, then you also have P(x) = SUM_i P(x|y=i)P(y=i), which is the probability of belonging to <strong>any<\/strong> of the classes you have seen. The only thing missing is to set a threshold T (or more complex method of deciding) under which you put a sample to the \"none\" class. Unfortunately learning a generative model is usually much harder than typical discriminative ones, but your problem is also much harder (since you want your model to be \"aware of non-observed samples\"). Alternatively you can learn P(x) independently from your classifier, and use it as \"pre-filtering\" of your data, so lets say you learn P(x) using some density estimation technique (GMM, KDE, etc.), and then you classify according to:<\/p>\n\n<pre><code>NEW_P(y=i|x) = P(x)*P(y=i|x)\nNEW_P(y=none|x) = 1-P(x)\n<\/code><\/pre>\n\n<p>It is easy to show that the above is valid probability, since <code>NEW_P(y=none|x) + SUM_i NEW_P(y=i|x) = 1-P(x) + P(x) SUM_i P(y=i|x) = 1-P(x) + P(x)*1 = 1<\/code>.<\/p>\n\n<p>In short, what you are doing now gives you only P(y=i|x), and the whole trick with discriminative learning is that you can omit P(x). For the problem you are considering you actually need it (unless you want hacky things \/ gathering data). And there are at least a few ways of learning (and using) it.<\/p>\n"}]}
{"QuestionId":42563981,"AnswerCount":1,"Tags":"<tensorflow><feed><restore>","CreationDate":"2017-03-02T19:12:51.860","AcceptedAnswerId":null,"OwnerUserId":4162005.0,"Title":"Tensorflow complains about missing feed_dict during graph restore","Body":"<p>I've built a CNN for image classification. During training I've saved several checkpoints. The data is fed through a feed_dictionary into the network.<\/p>\n\n<p>Now I want to restore the model which fails and I cant figure out why. The important lines of code are as follows:<\/p>\n\n<pre><\/pre>\n\n<p>The Line \"checkpoint_saver.restore\" throws an error:<\/p>\n\n<p>Traceback (most recent call last):\n  File \"C:\\Program Files\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1022, in _do_call\n    return fn(*args)\n  File \"C:\\Program Files\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1004, in _run_fn\n    status, run_metadata)\n  File \"C:\\Program Files\\Anaconda3\\envs\\tensorflow\\lib\\contextlib.py\", line 66, in <strong>exit<\/strong>\n    next(self.gen)\n  File \"C:\\Program Files\\Anaconda3\\envs\\tensorflow\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\", line 469, in raise_exception_on_not_ok_status\n    pywrap_tensorflow.TF_GetCode(status))\ntensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'input_images' with dtype float\n     [[Node: input_images = Placeholderdtype=DT_FLOAT, shape=[], _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n\n<p>Any knows how to solve this? Why do I need a filled feed_dictionary just for restoring the graph?<\/p>\n\n<p>Thanks in advance!<\/p>\n\n<p><strong>Update:<\/strong><\/p>\n\n<p>This is the code of the restore method of the saver object:<\/p>\n\n<pre><\/pre>\n\n<p>What I don't get: Why is the graph executed immediately? Am I using the wrong method? I just want to restore all trainable vars.<\/p>\n","answers":[{"AnswerId":"42575218","CreationDate":"2017-03-03T09:33:53.407","ParentId":null,"OwnerUserId":"4162005","Title":null,"Body":"<p>The problem was cause by a SessionRunHook for process logging:<\/p>\n\n<p>original hook:<\/p>\n\n<pre><code>class _LoggerHook(tf.train.SessionRunHook):\n  \"\"\"Logs loss and runtime.\"\"\"\n\n  def begin(self):\n    self._step = -1\n\n  def before_run(self, run_context):\n    self._step += 1\n    self._start_time = time.time()\n    return tf.train.SessionRunArgs(loss)  # Asks for loss value.\n\n  def after_run(self, run_context, run_values):\n    duration = time.time() - self._start_time\n    loss_value = run_values.results\n    if self._step % 5 == 0:\n      num_examples_per_step = FLAGS.batch_size\n      examples_per_sec = num_examples_per_step \/ duration\n      sec_per_batch = float(duration)\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<\/code><\/pre>\n\n<p>modified hook:<\/p>\n\n<pre><code>class _LoggerHook(tf.train.SessionRunHook):\n    \"\"\"Logs loss and runtime.\"\"\"\n\n    def __init__(self, flags, loss_op):\n        self._flags = flags\n        self._loss_op = loss_op\n        self._start_time = time.time()\n\n    def begin(self):\n        self._step = 0\n\n    def before_run(self, run_context):\n        if self._step == 0:\n            run_args = None\n        else:\n            run_args = tf.train.SessionRunArgs(self._loss_op)\n\n        return run_args\n\n    def after_run(self, run_context, run_values):\n\n        if self._step &gt; 0:\n            duration_n_steps = time.time() - self._start_time\n            loss_value = run_values.results\n            if self._step % self._flags.log_every_n_steps == 0:\n                num_examples_per_step = self._flags.batch_size\n\n                duration = duration_n_steps \/ self._flags.log_every_n_steps\n                examples_per_sec = num_examples_per_step \/ duration\n                sec_per_batch = float(duration)\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                self._start_time = time.time()\n        self._step += 1\n<\/code><\/pre>\n\n<p>Explanation:<\/p>\n\n<p>The logging is now skiped for the first iteration. So the session.run, which is executed by Saver.restore(..) doens't requiere a filled feed dictionary anymore.<\/p>\n"}]}
{"QuestionId":42564059,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning><updates>","CreationDate":"2017-03-02T19:16:37.107","AcceptedAnswerId":null,"OwnerUserId":7528619.0,"Title":"how tow install the last version of TensorFlow","Body":"<p>I'm working with Ubuntu 14.04, and I have  installed. I want to update . How can I do this?<\/p>\n\n<p>I do not understand the process described here: <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/migration<\/a>! <\/p>\n","answers":[{"AnswerId":"42564123","CreationDate":"2017-03-02T19:20:48.127","ParentId":null,"OwnerUserId":"2538592","Title":null,"Body":"<p>I avoid issues while upgrading, I'd suggest uninstalling tensorflow first:<\/p>\n\n<pre><code>$ pip uninstall protobuf\n$ pip uninstall tensorflow\n<\/code><\/pre>\n\n<p>and reinstall the latest version from here: <a href=\"http:\/\/tflearn.org\/installation\/\" rel=\"nofollow noreferrer\">http:\/\/tflearn.org\/installation\/<\/a><\/p>\n"},{"AnswerId":"42564247","CreationDate":"2017-03-02T19:27:11.393","ParentId":null,"OwnerUserId":"3896984","Title":null,"Body":"<p>An easy way todo this is<\/p>\n\n<pre><code>$ pip install --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n\n<p>After running this for python 2.7<\/p>\n\n<pre><code>export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0-cp27-none-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>for python 3.5<\/p>\n\n<pre><code>export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0-cp35-cp35m-linux_x86_64.whl\n<\/code><\/pre>\n\n<p>This is only for CPU processing. If you want to use a GPU you'll have to get different URLS from here: <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#TF_PYTHON_URL\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_linux#TF_PYTHON_URL<\/a><\/p>\n\n<p>As well as install the necessary CUDA drivers including CUDNN.<\/p>\n"}]}
{"QuestionId":42564572,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-03-02T19:45:09.513","AcceptedAnswerId":null,"OwnerUserId":1729279.0,"Title":"Adding a TensorFlow Op that uses malloc() and has state","Body":"<p>I wrote a TensorFlow Op that wraps a library written in C that calls  internally to allocate buffers and hold state. I initialize the library in the constructor, which is where the  calls take place, and then in  I use it with a mutex guard. The destructor makes the appropriate calls to . <\/p>\n\n<p>In my model each thread creates a copy of the Op, all in the same graph, and runs it within a shared session. The code very often crashes with pointer-related problems, such as \"incorrect checksum for freed object - object was probably modified after being freed\". The unit tests I wrote for the Op, with a single thread, work fine.<\/p>\n\n<p>I suspect the graph optimizer may be re-using a single instance of the Op, or something like that. Is there a way to tell it that the Op is stateful and that it should be left as-is? Are there other possible reasons for the crash?<\/p>\n\n<p>I'm running TensorFlow 1.0.0 and compiling the Op with clang 7.0, under OSX 10.10. I compile the C files first into a shared object, and then compile the C++ files together with the shared object.<\/p>\n","answers":[{"AnswerId":"51026836","CreationDate":"2018-06-25T15:17:55.607","ParentId":null,"OwnerUserId":"411907","Title":null,"Body":"<p>As per <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/extend\/adding_an_op\" rel=\"nofollow noreferrer\">documentation<\/a> on \"Adding a New Op\":<\/p>\n\n<blockquote>\n  <p>Important: Instances of your OpKernel may be accessed concurrently. Your Compute method must be thread-safe. Guard any access to class members with a mutex. Or better yet, don't share state via class members! Consider using a <a href=\"https:\/\/www.github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/core\/framework\/resource_mgr.h\" rel=\"nofollow noreferrer\">ResourceMgr<\/a> to keep track of op state.<\/p>\n<\/blockquote>\n\n<p>Unfortunately, usage of ResourceMgr is quite sketchy as of the time of writing this answer.<\/p>\n"}]}
{"QuestionId":42564698,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-02T19:53:05.217","AcceptedAnswerId":null,"OwnerUserId":5538474.0,"Title":"InvalidArgumentError: The node has inputs from different frames","Body":"<p>I was playing around with Tensorflow and ran into a problem with this code:<\/p>\n\n<pre><\/pre>\n\n<p>I'm evaluating the function this way:<\/p>\n\n<pre><\/pre>\n\n<p>I'm getting this error:<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError: The node 'process_tree_tf\/Variable\/Assign' has inputs from different frames. The input 'process_tree_tf\/Const_1' is in frame 'process_tree_tf\/process_tree_tf\/'. The input 'process_tree_tf\/Variable' is in frame ''.<\/p>\n<\/blockquote>\n\n<p>Weirdly, if I restart the kernel in jupyter notebook and run all again, I get this error:<\/p>\n\n<blockquote>\n  <p>FailedPreconditionError (see above for traceback): Attempting to use uninitialized value bias\n       [[Node: bias\/read = IdentityT=DT_FLOAT, _class=[\"loc:@bias\"], _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n<\/blockquote>\n\n<p>I tried using this instead:\n but that didn't work either.<\/p>\n\n<p>I'm very sorry if I'm overlooking something obvious here, but I'd really appreciate it if someone could tell me where I'm going wrong.<\/p>\n\n<p>Thank you very much!<\/p>\n","answers":[{"AnswerId":"42600540","CreationDate":"2017-03-04T19:43:28.723","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This is actually a subtle issue with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"noreferrer\"><code>tf.Variable<\/code><\/a> objects in TensorFlow's <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/while_loop\" rel=\"noreferrer\"><code>tf.while_loop()<\/code><\/a>. TensorFlow becomes confused because it appears that the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"noreferrer\"><code>tf.constant()<\/code><\/a> with which you're initializing the variable is a value created inside the loop (even though it's clearly loop invariant), but all variables get hoisted outside the loop. The easiest resolution is to move the creation of the variable outside the loop:<\/p>\n\n<pre><code>def process_tree_tf(matrix, weights, idxs, name=None):\n\n    with tf.name_scope(name, \"process_tree\", [tree, weights, idxs]).as scope():\n         loop_index = tf.sub(tf.shape(matrix)[0], 1)\n         loop_vars = loop_index, matrix, idxs, weights\n\n         # Define the bias variable outside the loop to avoid problems.\n         bias = tf.Variable(tf.constant(0.1, [2], dtype=tf.float64)) \n\n         def loop_condition(loop_idx, *_):\n             return tf.greater(loop_idx, 0)\n\n         def loop_body(loop_idx, mat, idxs, weights):\n             x = mat[loop_idx]\n             w = weights\n\n             # You can still refer to `bias` in here, and the loop body\n             # will capture it appropriately.\n             ...\n             return loop_idx-1, mat, idxs, weights\n\n         return tf.while_loop(loop_condition, loop_body, loop_vars, name=scope)[1]\n<\/code><\/pre>\n\n<p>(Another possible resolution would be to use a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant_initializer\" rel=\"noreferrer\"><code>tf.constant_initializer()<\/code><\/a> rather than a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/constant\" rel=\"noreferrer\"><code>tf.constant()<\/code><\/a> when creating the variable.)<\/p>\n"},{"AnswerId":"51388996","CreationDate":"2018-07-17T19:34:32.657","ParentId":null,"OwnerUserId":"4893073","Title":null,"Body":"<p>You can initialize <code>biases<\/code> inside the <code>loop_body<\/code>, like this:<\/p>\n\n<pre><code> def loop_body(loop_idx, mat, idxs, weights):\n     x = mat[loop_idx]\n     w = weights\n     bias = tf.get_variable(dtype=tf.float64,\n                            shape=[2],\n                            initializer=tf.constant_initializer(value=np.array([0.1,0.1]), dtype=tf.float64))\n<\/code><\/pre>\n\n<p>You said that you've tried <code>tf.get_variable<\/code> and <code>tf.constant_initializer<\/code>, I'm wondering whether you've found another solution?<\/p>\n"}]}
{"QuestionId":42565223,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-03-02T20:22:45.073","AcceptedAnswerId":null,"OwnerUserId":955742.0,"Title":"TF-Slim: (Unreasonable?) running out of memory","Body":"<p>I am trying to run one of the tutorials in TF SLIM, the one where you fine tune the flowers dataset using Inception-V3 (~104Mb). The GPU has about 2Gb of memory. When I have a batch size more than 8, I get an error because the GPU runs out of memory. In fact, it seems I get several messages, each looking like:<\/p>\n\n<pre><\/pre>\n\n<p>and<\/p>\n\n<pre><\/pre>\n\n<p>Now, it could very well be that my GPU has not big enough RAM. However, 2GB seems more than enough to load a ~100Mb model. Also, with Caffe I could fine-tune an Alexnet (~400Mb) with no problem. Besides, I also tried to allow for GPU growth (from what I understand using the system's RAM) with <\/p>\n\n<pre><\/pre>\n\n<p>but it doesn't seem to help.<\/p>\n\n<p>Do you know if <\/p>\n\n<p>a) I am doing something wrong\nb) the GPU is just not big enough\nc) TF Slim consumes by construction too much memory<\/p>\n\n<p>?<\/p>\n\n<p>Thanks,<\/p>\n","answers":[{"AnswerId":"42635801","CreationDate":"2017-03-06T21:21:40.497","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Could some other process be using enough GPU memory that not much is left over for tensorflow? I believe <code>nvidia-smi<\/code> will tell you how much GPU memory is already in use.<\/p>\n\n<p>If that's not the case you might want to look at the allocations to see what is going on. See <a href=\"https:\/\/stackoverflow.com\/questions\/40190510\/tensorflow-how-to-log-gpu-memory-vram-utilization\/40197094\">this other question on how to log allocations from tensorflow<\/a>.<\/p>\n"}]}
{"QuestionId":42566480,"AnswerCount":0,"Tags":"<macos><clang><openmp><torch><llvm-clang>","CreationDate":"2017-03-02T21:38:37.317","AcceptedAnswerId":null,"OwnerUserId":3303546.0,"Title":"What compiler can I use to install torch which supports OpenMP and nvcc on OS X 10.12?","Body":"<p>I am attempting to install Torch (with CUDA support) on a MacBook Pro with an NVIDIA GeForce graphics card running OS X 10.12 by following the <a href=\"http:\/\/torch.ch\/docs\/getting-started.html\" rel=\"nofollow noreferrer\">installation instructions<\/a>.<\/p>\n\n<p>I'm using CUDA 8.0. Although the <a href=\"http:\/\/docs.nvidia.com\/cuda\/cuda-installation-guide-mac-os-x\/index.html#axzz4aD2Iui1S\" rel=\"nofollow noreferrer\">CUDA system requirements<\/a> do not list OS X 10.12 and Apple LLVM 8 as supported, the <a href=\"https:\/\/developer.nvidia.com\/cuda-downloads\" rel=\"nofollow noreferrer\">CUDA download page<\/a> has a link for downloading OS X 10.12 and Apple LLVM 8 compatible CUDA.<\/p>\n\n<p>The default settings of the torch install script specifies the clang compiler, and finds<\/p>\n\n<pre><\/pre>\n\n<p>It runs through several installation tasks without problems but then repeatedly tries and fails to find OpenMP<\/p>\n\n<pre><\/pre>\n\n<p>Quotes from the OpenMP website suggest that OpenMP is distributed with clang (more on that below). I couldn't find any authoritative sources about whether or not OpenMP is included with Apple LLVM as well. <a href=\"https:\/\/stackoverflow.com\/a\/26296137\/3303546\">Several<\/a> <a href=\"https:\/\/stackoverflow.com\/a\/39980225\/3303546\">answers<\/a> suggests that it is not, and that the two are incompatible. So I decided to play it safe and attempt to compile with llvm-clang instead. <\/p>\n\n<h2>LLVM-Clang<\/h2>\n\n<p><a href=\"http:\/\/openmp.llvm.org\" rel=\"nofollow noreferrer\">From the OpenMP website<\/a><\/p>\n\n<blockquote>\n  <p>With the release of Clang 3.8.0, OpenMP 3.1 support is enabled in\n  Clang by default, and the OpenMP runtime is therefore built as a\n  normal part of the Clang build, and distributed with the binary\n  distributions.<\/p>\n<\/blockquote>\n\n<p>I used homebrew to install llvm and changed the following lines in the torch install script<\/p>\n\n<pre><\/pre>\n\n<p>The correct compiler is found<\/p>\n\n<pre><\/pre>\n\n<p>But I still get the error<\/p>\n\n<pre><\/pre>\n\n<p>Suggesting that either I have not correctly configured the compiler, or that my reading of the OpenMP website was incorrect.<\/p>\n\n<p>Additionally, I assumed that because Apple LLVM is supported, other LLVM distributions would also be supported. However, the script throws an another error<\/p>\n\n<pre><\/pre>\n\n<p>Which suggests that I cannot use this compiler regardless of OpenMP support.<\/p>\n\n<h2>GCC<\/h2>\n\n<p>Finally, I tried using a gcc compiler despite the warning in the torch install script, and it not being on the CUDA supported compilers list.<\/p>\n\n<pre><\/pre>\n\n<p>The compiler is correctly discovered<\/p>\n\n<pre><\/pre>\n\n<p>And the OpenMP library is also correctly found<\/p>\n\n<pre><\/pre>\n\n<p>This shows that the torch install script is capable of detecting OpenMP if it is correctly configured with the compiler.<\/p>\n\n<p>But as suggested by the comment, CUDA throws an error<\/p>\n\n<pre><\/pre>\n\n<p>So I cannot simply solve my problem by using the gcc compiler.<\/p>\n\n<h2>Question<\/h2>\n\n<p><strong>What compiler can I use to install torch which supports OpenMP and nvcc?<\/strong> <\/p>\n\n<p>If the answer is one of the compilers which I have already tried, how can I modify the install script or my compiler configuration to make it work? <\/p>\n","answers":[]}
{"QuestionId":42567398,"AnswerCount":2,"Tags":"<python><python-3.x><tensorflow>","CreationDate":"2017-03-02T22:43:35.037","AcceptedAnswerId":null,"OwnerUserId":4293730.0,"Title":"InvalidArgumentError (see above for traceback): indices[1] = 10 is not in [0, 10)","Body":"<p>I am using tensorflow 1.0 CPU on ubuntu and python 3.5.<\/p>\n\n<p>I adapted an example of tensorflow to work on my own dataset <a href=\"https:\/\/github.com\/martin-gorner\/tensorflow-mnist-tutorial\" rel=\"nofollow noreferrer\">https:\/\/github.com\/martin-gorner\/tensorflow-mnist-tutorial<\/a><\/p>\n\n<p>It works fine as long as the number of outputs is under 10. When the number of outputs is above 10,I get the error:<\/p>\n\n<pre><\/pre>\n\n<p>Any help?<\/p>\n","answers":[{"AnswerId":"52310338","CreationDate":"2018-09-13T09:09:20.767","ParentId":null,"OwnerUserId":"7866526","Title":null,"Body":"<p>I also came across the same error, and after fiddling with it for 2 days I came to realize there are 2 main reasons this error was getting thrown for my code and I mentioned them below to help anyone struggling with the same problem:<\/p>\n\n<ol>\n<li><p>The dimensions of your data and your labels being different <\/p><\/li>\n<li><p>In my case, the problem was\nthat when building my vocabulary I have indexed the words from 1 and\nnot from 0. But the embedded layer starts indexing from 0. So it\nkept giving me the mentioned error. I fixed the error by indexing my\nvocabulary from 0.<\/p>\n\n<p>previous code:<\/p>\n\n<pre><code>dictionary = Counter(words)\nsorted_split_words = sorted(dictionary, key=dictionary.get, reverse=True)\nvocab_to_int = {c: i for i, c in enumerate(sorted_split_words, 1)}\n<\/code><\/pre>\n\n<p>to fix it I changed the last line to (removed 1):<\/p>\n\n<pre><code>vocab_to_int = {c: i for i, c in enumerate(sorted_split_words)}\n<\/code><\/pre><\/li>\n<\/ol>\n"},{"AnswerId":"52014802","CreationDate":"2018-08-25T06:57:02.247","ParentId":null,"OwnerUserId":"10272384","Title":null,"Body":"<p>The index of the input word exceeds the length of the vocabulary, or the new words that are not included in the vocabulary.<\/p>\n\n<p>Please try to enlarge vocabulary length.<\/p>\n"}]}
{"QuestionId":42567704,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-02T23:07:29.133","AcceptedAnswerId":42567944.0,"OwnerUserId":7508642.0,"Title":"TensorFlow: Not all of my variables are being restored - Python","Body":"<p>I have another TensorFlow query:<\/p>\n\n<p>I'm training a regression model, saving the weights and biases and then restoring them to rerun the model on a different data set. At least, that's what I'm trying to do. Not all of my weights are being restored. Here's the code used for saving my variables:<\/p>\n\n<pre><\/pre>\n\n<p>And here's my entire code for restoring and running the model:<\/p>\n\n<pre><\/pre>\n\n<p>Now, here's what has me confused\/frustrated\/annoyed. From the weights I can restore 'h1', 'h2' and 'h3', but not 'out'. Why not 'out'? Is there anything that I'm doing wrong? Please can you spend a few minutes to help me?<\/p>\n\n<p>Many thanks<\/p>\n\n<p>I am running Python 3.5 and TensorFlow 0.12 directly on Windows 10 and I'm using Spyder IDE.<\/p>\n","answers":[{"AnswerId":"42567944","CreationDate":"2017-03-02T23:28:59.337","ParentId":null,"OwnerUserId":"7590388","Title":null,"Body":"<p>It looks like you are over-writing one of the 'out' keys with this dictionary constructor:<\/p>\n\n<pre><code>{**weights, **biases}\n<\/code><\/pre>\n\n<p>For example:<\/p>\n\n<pre><code>weights = {'h1':1, 'h2':2, 'out':3}\nbiases = {'b1':4, 'b2':5, 'out':6}\nprint({**weights, **biases})\n\n{'h2': 2, 'out': 6, 'b2': 5, 'b1': 4, 'h1': 1}\n<\/code><\/pre>\n"}]}
{"QuestionId":42568019,"AnswerCount":1,"Tags":"<deep-learning><theano><convolution>","CreationDate":"2017-03-02T23:37:18.460","AcceptedAnswerId":null,"OwnerUserId":6553672.0,"Title":"High cpu utilization while using 3D convolution on gpu in theano 0.9","Body":"<p>My cpu utilization is 100% when using theano.tensor.nnet.conv3d\nI am adding conv3d_fft, convgrad3d_fft and convtransp3d_fft to my compiling mode in the theano function.\nThe interesting part is my gpu utilization is also high. My data is all on gpu memory.<\/p>\n\n<p>Any ideas about why it cpu utilization is so high?<\/p>\n\n<p>Thanks!<\/p>\n\n<p>Update: I tried the same convolution in Keras with Theano backend. It is shockingly faster than theano despite the fact that it also uses conv3d.<\/p>\n","answers":[{"AnswerId":"42568397","CreationDate":"2017-03-03T00:17:08.833","ParentId":null,"OwnerUserId":"3022712","Title":null,"Body":"<p>For invoking a task on the GPU, the CPU has to do some work (not negligible). If the tasks on the GPU are small, the tasks on the CPU become comparatively big.\nBy the way, memory transfers are done by the DMA and do not cause CPU load except of invoking the transfer.<\/p>\n\n<p>I had a CUDA program, where the CPU invoked a GPU task every ~100\u00b5s. Profiling the program I found out, that it was CPU bound, while nothing was done on CPU except invoking GPU tasks.<\/p>\n"}]}
{"QuestionId":42568192,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-02T23:55:13.257","AcceptedAnswerId":null,"OwnerUserId":395857.0,"Title":"How to assign a value from a file to a TensorFlow variable without storing 2 copies in memory?","Body":"<p>I <a href=\"https:\/\/stackoverflow.com\/q\/34220532\/395857\">know<\/a> one can use the statement <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/state_ops.html#Variable.assign\" rel=\"nofollow noreferrer\"><\/a> followed by a call to <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/framework.html#Operation.run\" rel=\"nofollow noreferrer\"><\/a> or <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/client.html#Session.run\" rel=\"nofollow noreferrer\"><\/a> to assign the value  to the TensorFlow variable . <\/p>\n\n<p>However, this solution stores  twice in the memory. This can be an issue when the value to assign is large (e.g., word embeddings).<\/p>\n\n<p>How to assign a value from a file to a TensorFlow variable without storing 2 copies of the value in memory?<\/p>\n","answers":[{"AnswerId":"42569797","CreationDate":"2017-03-03T02:53:31.633","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>You can avoid storing the value in Python memory by using TensorFlow readers. IE, something like<\/p>\n\n<pre><code>filename_queue = tf.train.string_input_producer(\"myfile\")\nx.assign(reader.read(filename_queue))\n<\/code><\/pre>\n\n<p>You want to assign something than string you may have to have to combine <code>reader.read<\/code> with an op that produces the desired type, ie, <code>TextLineReader<\/code> + <code>tf.decode_csv<\/code><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/reading_data<\/a><\/p>\n"}]}
{"QuestionId":42568349,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-03-03T00:10:42.413","AcceptedAnswerId":42577704.0,"OwnerUserId":7195218.0,"Title":"Keras: Tokenizer with fit_generator() on text data","Body":"<p>I am creating a neural net on a very large text dataset using keras. To build the model and make sure everything was working, I read a fraction of the data into memory, and use the built in keras 'Tokenizer' to do the necessary preprocessing stuff, including mapping each word to a token. Then, I use model.fit(). <\/p>\n\n<p>Now, I want to extend to the full dataset, and don't have the space to read all the data into memory. So, I'd like to make a generator function to sequentially read data from disk, and use model.fit_generator(). However, if I do this, then I separately fit a Tokenizer object on each batch of data, providing a different word-to-token mapping for each batch. Is there anyway around this? Is there any way I can continuously build a token dictionary using keras?<\/p>\n","answers":[{"AnswerId":"42577704","CreationDate":"2017-03-03T11:28:23.230","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So basically you could define a text generator and feed it to <code>fit_on_text<\/code> method in a following manner:<\/p>\n\n<ol>\n<li><p>Assuming that you have <code>texts_generator<\/code> which is reading partially your data from disk and returning an iterable collection of text you may define:<\/p>\n\n<pre><code>def text_generator(texts_generator):\n    for texts in texts_generator:\n        for text in texts:\n            yield text\n<\/code><\/pre>\n\n<p>Please take care that you should make this generator stop after reading a whole of data from disk - what could possible make you to change the original generator you want to use in <code>model.fit_generator<\/code><\/p><\/li>\n<li><p>Once you have the generator from 1. you may simply apply a <code>tokenizer.fit_on_text<\/code> method by:<\/p>\n\n<pre><code>tokenizer.fit_on_text(text_generator)\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":42568711,"AnswerCount":0,"Tags":"<tensorflow><segmentation-fault>","CreationDate":"2017-03-03T00:52:46.463","AcceptedAnswerId":null,"OwnerUserId":7650607.0,"Title":"See segmentation fault running tensorflow.contrib.learn tutorial example","Body":"<p>See segmentation faults when running the  example shown in tensorflow official tutorial .<\/p>\n\n<p>The tutorial example code is<\/p>\n\n<pre><\/pre>\n\n<p>Observed segmentation fault:<\/p>\n\n<pre><\/pre>\n\n<p>Environments and versions:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42569141,"AnswerCount":1,"Tags":"<python><multidimensional-array><tensorflow><deep-learning>","CreationDate":"2017-03-03T01:40:10.360","AcceptedAnswerId":42628877.0,"OwnerUserId":1496362.0,"Title":"Splitting ndarray gives unexpected results (TensorFlow RNN tutorial)","Body":"<p>I am following a <a href=\"http:\/\/r2rt.com\/recurrent-neural-networks-in-tensorflow-i.html\" rel=\"nofollow noreferrer\">tutorial<\/a> on rnn's in TensorFlow but I have a question concerning the input formats. <\/p>\n\n<p>They are taking raw_x (one hot vector) and basically first cutting that up in pieces of length 200 (batch_size) to form data_x. That is good. <\/p>\n\n<p>Then they further cut up data_x in pieces of length 5 (num_step, or graph width) with: <\/p>\n\n<pre><\/pre>\n\n<p>However, if I look in the data, the slices of x do not match data_x. The first one does, but then they diverge. <\/p>\n\n<p>Am I misunderstanding the above code? I would like to understand how x is being created or what it is supposed to look like. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/bxV33.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bxV33.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/0pBLa.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/0pBLa.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I had expected the second item to be 0 1 0 1 0.<\/p>\n\n<p>Also, I thought an epoch is when you go through the data completely, from this it seems that they split up the data in 1000 parts (epoch size)? <\/p>\n\n<p>If it helps, this is my full code. I am trying to figure out what is going on in x. at line 48: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42628877","CreationDate":"2017-03-06T15:02:47.140","ParentId":null,"OwnerUserId":"1496362","Title":null,"Body":"<p>Seems like the batches are actually transposed. <\/p>\n\n<p>So the first elements of the x-matrix (200 x 5) will fit the first 5 elements of x_raw. <\/p>\n\n<p>Then only in the next iteration, the next 5-10 elements of x_raw will be in the first elements (again) of x. <\/p>\n"}]}
{"QuestionId":42569188,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-03-03T01:46:29.397","AcceptedAnswerId":null,"OwnerUserId":7650767.0,"Title":"AttributeError: module 'tensorflow.python.training.training' has no attribute 'SummaryWrite","Body":"<pre><\/pre>\n\n<p>I currently use tensorflow to build a simple 3 layers NN, and I want to use tensorboard to show the gragh. but when I run the module, it shows:  AttributeError: module 'tensorflow.python.training.training' has no attribute 'SummaryWrite'. I am really confused....<\/p>\n","answers":[]}
{"QuestionId":42569513,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-03T02:23:34.797","AcceptedAnswerId":null,"OwnerUserId":4339490.0,"Title":"Inception: How to process image to use with Inception","Body":"<p>I want to make tensorflow's inception v3 to give out tags for an image. My goal is to convert a JPEG image to input that is accepted by inception neural network. I don't know how to process the images first so that it can run with Google Inception's v3 model. The original tensorflow project is here:\n<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception<\/a><\/p>\n\n<p>Originally, all the images are in a dataset and the entire dataset is first passed to input() or distorted_inputs() in <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/image_processing.py\" rel=\"nofollow noreferrer\">ImageProcessing.py<\/a> . The images in dataset are processed and passed to the train() or eval() methods (both of these work). The problem is I want a function to print out tags for one specific image (not dataset). <\/p>\n\n<p>Below is the code for inference function that is used to generate tag with google inception.  function is a convolutional neural network implemented in tensorflow.<\/p>\n\n<pre><\/pre>\n\n<p>This is my attempt to process the image before it is passed to inference function. <\/p>\n\n<pre><\/pre>\n\n<p>I wanted to process an image file simply so that I can pass it to the inference function. And that inference prints out the tags. The above code didn't work and printed error: <\/p>\n\n<p><\/p>\n\n<p>I appreciate if anyone can provide any insight into this problem. <\/p>\n","answers":[{"AnswerId":"43220029","CreationDate":"2017-04-05T00:45:41.777","ParentId":null,"OwnerUserId":"2681680","Title":null,"Body":"<p>Inception just needs (299,299,3) images with inputs scaled between -1 and 1. See code below. I just change the images using this and put them in a TFRecord ( and then queue ) to run my stuff.<\/p>\n\n<pre><code>from PIL import Image\nimport PIL\nimport numpy as np\ndef load_image( self, image_path ):\n    img = Image.open( image_path )\n    newImg = img.resize((299,299), PIL.Image.BILINEAR).convert(\"RGB\")\n    data = np.array( newImg.getdata() )\n    return 2*( data.reshape( (newImg.size[0], newImg.size[1], 3) ).astype( np.float32 )\/255 ) - 1\n<\/code><\/pre>\n"}]}
{"QuestionId":42569921,"AnswerCount":1,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-03-03T03:07:07.817","AcceptedAnswerId":42587133.0,"OwnerUserId":7397552.0,"Title":"tf.contrib.slim.get_variables_to_restore() does not return value","Body":"<p>Running below code tf.contrib.slim.get_variables_to_restore() return empty value [] for all_vars, and then causing failure when calling tf.train.Saver. Detail error message shows below.<\/p>\n\n<p>Am I missing anything?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42587133","CreationDate":"2017-03-03T19:40:41.893","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The problem here seems to be that your graph is empty\u2014i.e. it does not contain any variables. You create a new graph on the line <code>with tf.Session(graph=tf.Graph()):<\/code>, and none of the following lines creates a <code>tf.Variable<\/code> object.<\/p>\n\n<p>To restore a pre-trained TensorFlow model, you need to do one of three things:<\/p>\n\n<ol>\n<li>Rebuild the model graph, by executing the same Python graph building code that was used to train the model in the first place.<\/li>\n<li>Load a \"MetaGraph\" that contains information about how to reconstruct the graph structure and model variables. See <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/meta_graph\" rel=\"nofollow noreferrer\">this tutorial<\/a> for more details on how to create and use a MetaGraph. MetaGraphs are often created alongside checkpoint files, and typically have the extension <code>.meta<\/code>.<\/li>\n<li>Load a \"SavedModel\", which contains a \"MetaGraph\". See the documentation <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1f09d0c1ee7648288b3454d64e72e8760b9acc9c\/tensorflow\/python\/saved_model\/README.md\" rel=\"nofollow noreferrer\">here<\/a> for more details.<\/li>\n<\/ol>\n"}]}
{"QuestionId":42570344,"AnswerCount":1,"Tags":"<python><macos><intellij-idea><tensorflow><pycharm>","CreationDate":"2017-03-03T03:53:03.327","AcceptedAnswerId":null,"OwnerUserId":7651140.0,"Title":"Setting TensorFlow on Mac with PyCharm","Body":"<p>I have installed Python3 and TensorFlow 1.0 on my Mac and it works fine when I run code in \"terminal\" with command line.<\/p>\n\n<pre><\/pre>\n\n<p>But when I use IntelliJ or PyCharm  and run the same code, <\/p>\n\n<pre><\/pre>\n\n<p>Why is this happening???\nHow can I fix this?<\/p>\n","answers":[{"AnswerId":"45662170","CreationDate":"2017-08-13T15:51:31.680","ParentId":null,"OwnerUserId":"8458596","Title":null,"Body":"<pre><code>export DYLD_LIBRARY_PATH=\/usr\/local\/cuda\/lib\n<\/code><\/pre>\n\n<p>Choose \"Edit Configuration\". You will see an Environment option where you need to click on the ... box and enter the DYLD_LIBRARY_PATH that applies to your case.<\/p>\n\n<p>Note that there's an Environment option for the specific file you're working on (it will be highlighted in the left panel) and for Defaults (put DYLD_... there as well if you want future files you create to have this). Note that you need to save this config or else when you close PyCharm it won't stick.<\/p>\n"}]}
{"QuestionId":42570645,"AnswerCount":1,"Tags":"<amazon-s3><pytorch>","CreationDate":"2017-03-03T04:25:53.200","AcceptedAnswerId":null,"OwnerUserId":7651250.0,"Title":"HTTP error when installing pytorch","Body":"<p>I use ubuntu14.04, without GPU.\nwhen I install pytorch in anaconda(I type command just as pytorch.org says),things always get wrong,like this:<\/p>\n\n<pre><\/pre>\n\n<p>I've tried many times, but every time error occurs in the same step.\nI'm looking forward your answers,thanks very very much for your help!!!<\/p>\n","answers":[{"AnswerId":"47938602","CreationDate":"2017-12-22T09:00:39.390","ParentId":null,"OwnerUserId":"8648333","Title":null,"Body":"<p>Please update to the latest version of Navigator.<\/p>\n\n<p>Open a terminal (on Linux or Mac) or the Anaconda Command Prompt (on windows)\nand type<\/p>\n\n<p>$ conda update anaconda-navigator<\/p>\n"}]}
{"QuestionId":42571035,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-03T05:02:42.433","AcceptedAnswerId":null,"OwnerUserId":7602394.0,"Title":"Tensorflow CNN Batch size error","Body":"<p>I had made CNN model for my dataset.<\/p>\n\n<p>I had used batch for feed data.\nwhen I used batch size is one, It is working.\nbut if I used batch size is not one (ex :128)\nit make error.<\/p>\n\n<p>this is my code.\nI attach all my code.<\/p>\n\n<p>there is 1623 columns data.<\/p>\n\n<pre><\/pre>\n\n<p>this is code for select batch size.<\/p>\n\n<pre><\/pre>\n\n<p>if I write batch size to bigger than one, It makes this error<\/p>\n\n<pre><\/pre>\n\n<p>I want use batch function but in this case I can not use this.\nhow can I solve this problem?<\/p>\n","answers":[]}
{"QuestionId":42571445,"AnswerCount":1,"Tags":"<deep-learning><caffe><pycaffe><matcaffe>","CreationDate":"2017-03-03T05:38:22.930","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"What I am doing wrong with this semantic segmentation?","Body":"<p>Now it is quite a long time (almost two months) that I was working on FCN32 for semantic segmentation of single channel images. I played around with different learning rates and even adding  layer. However, I was not successful to even see any output. I did not have any choice except to instantly ask for help here. I really do not know what I am doing wrong. <\/p>\n\n<p>I am sending one image to the network as a batch.This the train-loss curve  and :\n<a href=\"https:\/\/i.stack.imgur.com\/cPiqD.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/cPiqD.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I increased the learning rate to (the following figure). It seems that loss is falling down, however, the learning curve is not acting normal. \n<a href=\"https:\/\/i.stack.imgur.com\/Pf15Z.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Pf15Z.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I reduced the layers of original FCN as follows:\n(1) Conv64 \u2013 ReLU \u2013 Conv64 \u2013 ReLU \u2013 MaxPool<\/p>\n\n<p>(2) Conv128 \u2013 ReLU \u2013 Conv128 \u2013 ReLU \u2013 MaxPool<\/p>\n\n<p>(3) Conv256 \u2013 ReLU \u2013 Conv256 \u2013 ReLU \u2013 MaxPool <\/p>\n\n<p>(4) Conv4096 \u2013 ReLU \u2013 Dropout0.5<\/p>\n\n<p>(5) Conv4096 \u2013 ReLU \u2013 Dropout0.5<\/p>\n\n<p>(6) Conv2 <\/p>\n\n<p>(7) Deconv32x \u2013 Crop<\/p>\n\n<p>(8) SoftmaxWithLoss<\/p>\n\n<pre><\/pre>\n\n<p>and this is the solver definition:<\/p>\n\n<pre><\/pre>\n\n<p>At the beginning, the loss is starting to fall down, but again after some iterations, it is not showing good learning behavior:\n<a href=\"https:\/\/i.stack.imgur.com\/9EhNh.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/9EhNh.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I am a beginner in deep learning and . I really do not understand why this happens. I really appreciate if those that have expertise, please have a look on the model definition and I will be very thankful if you help me. <\/p>\n","answers":[{"AnswerId":"42596148","CreationDate":"2017-03-04T12:41:23.893","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p>The problem is that you are training from scratch.<\/p>\n\n<p>Reading the <a href=\"https:\/\/arxiv.org\/abs\/1605.06211#\" rel=\"nofollow noreferrer\">FCN paper<\/a> will tell you that they always use networks that are pretrained on ImageNet, it will <strong>NOT<\/strong> work if you train it from scratch, it has to be finetuned from a pretrained network. The optimization problem if you train from random weights just doesn't converge.<\/p>\n"}]}
{"QuestionId":42572638,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-03T07:02:44.827","AcceptedAnswerId":null,"OwnerUserId":5566910.0,"Title":"How to use the pre-trained ResNet50 in tensorflow?","Body":"<p>I want to design a network built on the pre-trained network with tensorflow, taking Reset50 for example.  <a href=\"https:\/\/github.com\/ry\/tensorflow-resnet\" rel=\"noreferrer\">ry released a model<\/a>, however, I don't know how to use it to build my model with their checkpoint? The definition of resnet can be found in resnet.py. who can help me? Thank you very much!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"45343030","CreationDate":"2017-07-27T06:35:45.663","ParentId":null,"OwnerUserId":"1788679","Title":null,"Body":"<p>Basically you should use the code supplied for the model. You can create graph using them and then supply the checkpoint file, see how to do it in case of ResNet50 below:<\/p>\n\n<pre><code>from tensorflow.contrib.slim.nets import resnet_v1\nimport tensorflow as tf\nimport tensorflow.contrib.slim as slim\n\n# Create graph\ninputs = tf.placeholder(tf.float32, shape=[batch_size, height, width, channels])\nwith slim.arg_scope(resnet_v1.resnet_arg_scope()):\n    net, end_points = resnet_v1.resnet_v1_50(inputs, is_training=False)\n\nsaver = tf.train.Saver()    \n\nwith tf.Session() as sess:\n    saver.restore(sess, '.resnet_v1_50.ckpt')\n    representation_tensor = sess.graph.get_tensor_by_name('resnet_v1_50\/pool5:0') # if you don't know names like these, consider referring to corresponding model file or generate .pbtxt file as mentioned in  @civilman628 's answer in link below\n    img = ...  #load image here with size [1, 224,224, 3]\n    features = sess.run(representation_tensor, {'Placeholder:0': x})\n<\/code><\/pre>\n\n<p>For more details, please see my (@parthg) answer on Tensorflow Github issue related to this: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7172\" rel=\"nofollow noreferrer\">#7172<\/a><\/p>\n"},{"AnswerId":"45666377","CreationDate":"2017-08-14T01:32:41.990","ParentId":null,"OwnerUserId":"4372241","Title":null,"Body":"<p>I use keras which uses TensorFlow. Here is an example feeding one image at a time:<\/p>\n\n<pre><code>import numpy as np\n\nfrom keras.preprocessing import image\nfrom keras.applications import resnet50\n\n# Load Keras' ResNet50 model that was pre-trained against the ImageNet database\nmodel = resnet50.ResNet50()\n\n# Load the image file, resizing it to 224x224 pixels (required by this model)\nimg = image.load_img(\"path_to_image.jpg\", target_size=(224, 224))\n\n# Convert the image to a numpy array\nx = image.img_to_array(img)\n\n# Add a forth dimension since Keras expects a list of images\nx = np.expand_dims(x, axis=0)\n\n# Scale the input image to the range used in the trained network\nx = resnet50.preprocess_input(x)\n\n# Run the image through the deep neural network to make a prediction\npredictions = model.predict(x)\n\n# Look up the names of the predicted classes. Index zero is the results for the first image.\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"}]}
{"QuestionId":42573628,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-03T08:04:58.327","AcceptedAnswerId":42573684.0,"OwnerUserId":1215889.0,"Title":"Unable to load variables in Tensorflow","Body":"<p>I have a following code :  <\/p>\n\n<pre><\/pre>\n\n<p>But the above code is unable to load variables. I am getting the length of all_vars as 0.<br>\nCan someone tell me what is the problem?  <\/p>\n","answers":[{"AnswerId":"42573684","CreationDate":"2017-03-03T08:08:49.363","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Can you try this? <\/p>\n\n<pre><code>tf.get_collection(tf.GraphKeys.VARIABLES)\n<\/code><\/pre>\n\n<p>This should give you all trainable and untrainable variables...<\/p>\n"}]}
{"QuestionId":42573986,"AnswerCount":1,"Tags":"<numpy><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-03-03T08:29:15.613","AcceptedAnswerId":42576990.0,"OwnerUserId":766391.0,"Title":"reduce size of pretrained deep learning model for feature generation","Body":"<p>I am using an pretrained model in Keras to generate features for a set of images:<\/p>\n\n<pre><\/pre>\n\n<p>However, I have a lot of images and the Imagenet model outputs 131072 features (columns) for each image.<\/p>\n\n<p>With 200k images I would get an array of  which is too large to fit into memory. <\/p>\n\n<p>More importantly, I need to save this array to disk and it would take 100 GB of space when saved as  or <\/p>\n\n<p>I could circumvent the memory problem by feeding only batches of like 1000 images and saving them to disk, but not the disk space problem. <\/p>\n\n<p>How can I make the model smaller without losing too much information?<\/p>\n\n<p><strong>update<\/strong><\/p>\n\n<p>as the answer suggested I include the next layer in the model as well:<\/p>\n\n<pre><\/pre>\n\n<p>this reduced the output to <\/p>\n\n<p><strong>update 2<\/strong>:<\/p>\n\n<p>another interesting solution may be the  package to reduce size of numpy arrays  <a href=\"https:\/\/github.com\/Blosc\/bcolz\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Blosc\/bcolz<\/a><\/p>\n","answers":[{"AnswerId":"42576990","CreationDate":"2017-03-03T10:54:55.597","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>I see at least two solutions to your problem:<\/p>\n\n<ol>\n<li>Apply a <code>model = AveragePooling2D((8, 8), strides=(8, 8))(model)<\/code> where <code>model<\/code> is an <code>InceptionV3<\/code> object you loaded (without top). This is the next step in <code>InceptionV3<\/code> architecture - so one may easily assume - that these features still hold loads of discriminatory clues.<\/li>\n<li>Apply a some kind of dimensionality reduction (e.g. like <code>PCA<\/code>) on a sample of data and reduce the dimensionality of all data to get the reasonable file size.<\/li>\n<\/ol>\n"}]}
{"QuestionId":42574476,"AnswerCount":6,"Tags":"<python><tensorflow><deep-learning><upgrade>","CreationDate":"2017-03-03T08:55:54.897","AcceptedAnswerId":null,"OwnerUserId":7528619.0,"Title":"Update TensorFlow","Body":"<p>I'm working with  , I had a  but I want to update this version. \nif i do:<\/p>\n\n<pre><\/pre>\n\n<p>but it prints: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43332715","CreationDate":"2017-04-10T20:53:16.347","ParentId":null,"OwnerUserId":"5252990","Title":null,"Body":"<p>Before trying to update tensorflow try updating pip <\/p>\n\n<pre><code>pip install --upgrade pip\n<\/code><\/pre>\n\n<p>If you are upgrading from a previous installation of TensorFlow &lt; 0.7.1, you should uninstall the previous <strong>TensorFlow<\/strong> and <strong>protobuf<\/strong> using, <\/p>\n\n<pre><code>pip uninstall tensorflow\n<\/code><\/pre>\n\n<p>to make sure you get a clean installation of the updated protobuf dependency.<\/p>\n\n<p>Uninstall the TensorFlow on your system, and check out <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/get_started\/os_setup#download-and-setup\" rel=\"nofollow noreferrer\">Download and Setup to reinstall again<\/a>.<\/p>\n\n<p>If you are using pip install, go check the available version over <a href=\"https:\/\/storage.googleapis.com\/tensorflow\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow<\/a>, search keywords with linux\/cpu\/tensorflow to see the availabilities.<\/p>\n\n<p>Then, set the path for download and execute in sudo.<\/p>\n\n<pre><code>$ export TF_BINARY_URL=https:\/\/storage.googleapis.com\/tensorflow\/linux\/cpu\/tensorflow-1.0.0-py2-none-any.whl\n\n$ sudo pip install --upgrade $TF_BINARY_URL\n<\/code><\/pre>\n\n<p>For more detail, follow this link in <a href=\"http:\/\/e-lin.github.io\/wiki\/jekyll\/update\/2017\/02\/18\/Upgrade-the-Installed-TensorFlow.html\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"43647690","CreationDate":"2017-04-27T02:33:10.110","ParentId":null,"OwnerUserId":"1090562","Title":null,"Body":"<p>To upgrade any python package, use <code>pip install &lt;pkg_name&gt; --upgrade<\/code>.<\/p>\n\n<p>So in your case it would be <code>pip install tensorflow --upgrade<\/code>. Just updated to <code>1.1.0<\/code><\/p>\n"},{"AnswerId":"47342614","CreationDate":"2017-11-17T03:02:22.637","ParentId":null,"OwnerUserId":"6254131","Title":null,"Body":"<pre><code>(tensorflow)$ pip install --upgrade pip  # for Python 2.7\n(tensorflow)$ pip3 install --upgrade pip # for Python 3.n\n\n(tensorflow)$ pip install --upgrade tensorflow      # for Python 2.7\n(tensorflow)$ pip3 install --upgrade tensorflow     # for Python 3.n\n(tensorflow)$ pip install --upgrade tensorflow-gpu  # for Python 2.7 and GPU\n(tensorflow)$ pip3 install --upgrade tensorflow-gpu # for Python 3.n and GPU\n\n(tensorflow)$ pip install --upgrade tensorflow-gpu==1.4.1 # for a specific version\n<\/code><\/pre>\n\n<p>Details on <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux\" rel=\"noreferrer\">install tensorflow<\/a>.<\/p>\n"},{"AnswerId":"47765326","CreationDate":"2017-12-12T04:23:21.550","ParentId":null,"OwnerUserId":"2956066","Title":null,"Body":"<p>For anaconda installation, first pick a <em>channel<\/em> which has the latest version of tensorflow binary. Usually, the latest versions are available at the channel <code>conda-forge<\/code>. Then simply do:<\/p>\n\n<pre><code>conda update -f -c conda-forge tensorflow\n<\/code><\/pre>\n\n<p>This will upgrade your existing tensorflow installation to the very latest version available. As of this writing, the latest version is <code>1.4.0-py36_0<\/code><\/p>\n"},{"AnswerId":"58325399","CreationDate":"2019-10-10T14:35:01.713","ParentId":null,"OwnerUserId":"2461937","Title":null,"Body":"<p>Upgrading to Tensorflow 2.0 using pip.\nRequires Python > 3.4 and pip >= 19.0<\/p>\n\n<pre><code>CST:~ USERX$ pip3 show tensorflow\nName: tensorflow\nVersion: 1.13.1\n\nCST:~ USERX$ python3 --version\nPython 3.7.3\n\nCST:~ USERX$ pip3 install --upgrade tensorflow\n\nCST:~ USERX$ pip3 show tensorflow\nName: tensorflow\nVersion: 2.0.0\n<\/code><\/pre>\n"},{"AnswerId":"48500482","CreationDate":"2018-01-29T11:44:29.793","ParentId":null,"OwnerUserId":"3806413","Title":null,"Body":"<p>This is official recommendation for upgrading Tensorflow.<\/p>\n\n<p>To get TensorFlow 1.5, you can use the standard pip installation (or pip3 if you use python3)<\/p>\n\n<pre><code>$  pip install --ignore-installed --upgrade tensorflow\n<\/code><\/pre>\n"}]}
{"QuestionId":42574611,"AnswerCount":3,"Tags":"<python><tensorflow>","CreationDate":"2017-03-03T09:03:49.880","AcceptedAnswerId":42574998.0,"OwnerUserId":4993842.0,"Title":"Does size of training data for an epoch matter in tensorflow?","Body":"<p>Assuming we have 500k items worth of training data, does it matter if we train the model one item at a time or 'n' items at a time or all at once?<\/p>\n\n<p>Considering  and  to be  and  to be any generic tensorflow training step. <\/p>\n\n<p><strong>Option 1<\/strong> Train one item at a time -<\/p>\n\n<pre><\/pre>\n\n<p><strong>Option 2<\/strong> Train on all at once -<\/p>\n\n<pre><\/pre>\n\n<p>Is there any difference between options 1 and 2 above as far as the quality of training is concerned? <\/p>\n","answers":[{"AnswerId":"42574998","CreationDate":"2017-03-03T09:22:15.027","ParentId":null,"OwnerUserId":"6659511","Title":null,"Body":"<p>Yes, there is a difference. Option 1 is much less memory consuming but is also much less accurate. Option 2 could eat up all of your RAM but should prove more accurate. However, if you use all your training set at once, be sure to limit the number of steps to avoid over-fitting.\nIdeally, use data in batches (typically between 16 and 256).\nMost optimization techniques are 'stochastic', i.e. they rely on a statistical sample of examples to estimate a model update.\nTo sum up:\n- More data => more accuracy (but more memory) => higher risk of over-fitting (so limit the amount of training steps)<\/p>\n"},{"AnswerId":"49039994","CreationDate":"2018-02-28T22:52:29.430","ParentId":null,"OwnerUserId":"5000294","Title":null,"Body":"<p>Mathematically these two methods are different. One is called stochastic gradient descent and the other is called batch gradient descent. You are missing the most commonly used one - mini batch gradient descent.  There has been a lot of research on this topic but basically different batch sizes have different convergence properties. Generally people use batch sizes that are greater than one but not the full dataset. This is usually necessary since most datasets cannot fit into memory all at once. Also if your model uses batch normalization then a batch size of one won't converge. This <a href=\"https:\/\/arxiv.org\/pdf\/1705.08741.pdf\" rel=\"nofollow noreferrer\">paper<\/a> discusses the effects of batch size (among other things) on performance. The takeaway is that larger batch sizes do not generalize as well. (They actually argue it isn't the batch size itself the but the fact that you have fewer updates when the batch is larger. I would recommend batch sizes of 32 to start and experiment to see how batch size effects performance. <\/p>\n\n<p>Here is a graph of the effects of batch size on training and validation performance from the paper I linked. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/XAU7U.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/XAU7U.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"42574898","CreationDate":"2017-03-03T09:17:59.580","ParentId":null,"OwnerUserId":"6824263","Title":null,"Body":"<p>There is a different between this options. Normally you have to use a batchsize to train for example 128 iterations of data.\nYou also could use a batchsize of one, like the first of you examples.\nThe advantage of this method is you can output the training efficient of the neural network.<\/p>\n\n<p>If you are learning all data at ones, you will bi a little bit faster, but you will know only at the end if you efficient is good.<\/p>\n\n<p>Best way is to make a batchsize and learn by stack. So you can output you efficient after every stack and control your efficient.<\/p>\n"}]}
{"QuestionId":42574644,"AnswerCount":3,"Tags":"<python><ubuntu><tensorflow>","CreationDate":"2017-03-03T09:05:36.210","AcceptedAnswerId":42574859.0,"OwnerUserId":3339394.0,"Title":"Tensorflow installation on Ubuntu permission error","Body":"<p>I installed virtual box on my Windows 10 machine and installed Ubuntu on the virtual box. Then I installed Tensorflow on Ubuntu by following <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux\" rel=\"nofollow noreferrer\">this instructions from Tensorflow.org<\/a>. Everything went well including pip install and stuff but when I run  I run into permission error as the screenshot shows. <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/DCp2N.png\" alt=\"enter image description here\"><\/a> <\/p>\n\n<p>This error is not described in the install errors listed on Tensorflow.org at the bottom of that step 2. How do I solve that?<\/p>\n","answers":[{"AnswerId":"42574859","CreationDate":"2017-03-03T09:15:45.363","ParentId":null,"OwnerUserId":"3339394","Title":null,"Body":"<p>As is with most Linux flavors, if you run into permission related errors just add sudo in front of the command and enter the password to confirm you have admin privileges. In this case simply type <code>sudo pip istall tensorflow<\/code>. But since the ubuntu also comes with the latest 3.N version of python it is better to do <code>sudo pip3 install tensorflow<\/code>.<\/p>\n"},{"AnswerId":"42574683","CreationDate":"2017-03-03T09:08:08.987","ParentId":null,"OwnerUserId":"3402367","Title":null,"Body":"<p>It seems like you need elevated permissions to write to <code>\/usr\/local\/lib<\/code>.<\/p>\n\n<p>Executing <code>sudo pip install tensorflow<\/code> will install tensorflow using root privileges.<\/p>\n\n<p>(Also, your problem is with Ubuntu, not with Windows 10! Your host system has no influence in the permissions of the guest.)<\/p>\n"},{"AnswerId":"45249617","CreationDate":"2017-07-22T01:02:52.570","ParentId":null,"OwnerUserId":"1472513","Title":null,"Body":"<p>I ran into the same problem and found the answer <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8806#issuecomment-290239615\" rel=\"nofollow noreferrer\">on Github<\/a>.<\/p>\n\n<p>The mistake I make was that Anaconda was installed with <code>sudo .\/&lt;install script&gt;<\/code> instead of the correct command <code>bash &lt;install script&gt;<\/code> as noted on the <a href=\"https:\/\/www.continuum.io\/downloads#linux\" rel=\"nofollow noreferrer\">download page<\/a>. This made some files root-owned.<\/p>\n"}]}
{"QuestionId":42575195,"AnswerCount":1,"Tags":"<python><linux><python-2.7><tensorflow>","CreationDate":"2017-03-03T09:32:59.197","AcceptedAnswerId":42575304.0,"OwnerUserId":6723438.0,"Title":"ERROR: Failed to load the native TensorFlow runtime in Python 2.7","Body":"<p>Here is the issue. This error FIRST occurred pointing to this<\/p>\n\n<pre><\/pre>\n\n<p>that I solved using this\n<a href=\"https:\/\/askubuntu.com\/questions\/575505\/glibcxx-3-4-20-not-found-how-to-fix-this-error\">glibcxx-3-4-20-not-found-how-to-fix-this-error<\/a> <\/p>\n\n<p>BUT, now it has another similar problem<\/p>\n\n<pre><\/pre>\n\n<p>How to solve these issues? Why do they keep coming one after another?<\/p>\n","answers":[{"AnswerId":"42575304","CreationDate":"2017-03-03T09:37:30.640","ParentId":null,"OwnerUserId":"2455626","Title":null,"Body":"<p>You're missing some dependencies for running TensorFlow.<\/p>\n\n<p>You can follow <a href=\"https:\/\/www.tensorflow.org\/install\/install_linux\" rel=\"nofollow noreferrer\">this<\/a> tutorial to install TensorFlow from scratch.<\/p>\n\n<p>Anyway, for your particular issue, you can just install <code>libc6<\/code>:<\/p>\n\n<pre><code>sudo apt-get update\nsudo apt-get install libc6\n<\/code><\/pre>\n"}]}
{"QuestionId":42575947,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-03T10:06:25.483","AcceptedAnswerId":42581449.0,"OwnerUserId":1782792.0,"Title":"Update variable by feeding its value","Body":"<p>I have a TensorFlow model and I want to have the option to retrieve and update the weights (out of the checkpointing mechanism, that is). Since I may do this a number of times, I don't want to add nodes to the graph whenever I do it, but rather have some op(s) that I can call to do it. My idea was to have  ops using the same variable both for the updated variable and for the value, and provide the new weights values in the ; so something like this:<\/p>\n\n<pre><\/pre>\n\n<p>In my mind, this should take the values passed in  as current values of the variables and then store them through the  operations. However, this does not work and gives me some weird error about unexpected value types.<\/p>\n\n<p>My current alternative is to have instead some auxiliary nodes, either variables or placeholders, and use as value in the assignment operations:<\/p>\n\n<pre><\/pre>\n\n<p>Is this really the only way to do this? Or is there some other obvious way I am not seeing?<\/p>\n","answers":[{"AnswerId":"42581449","CreationDate":"2017-03-03T14:28:47.303","ParentId":null,"OwnerUserId":"7590388","Title":null,"Body":"<p>Alright, found that there is a load() method:<\/p>\n\n<pre><code>with tf.Graph().as_default():\n    w = tf.get_variable('weights', shape=[3, 3], \n                        initializer=tf.random_uniform_initializer(dtype=tf.float32))\n    init = tf.global_variables_initializer()\n    with tf.Session() as sess:\n        sess.run(init)\n        print('initial W:')\n        print(sess.run(w))\n        new_vals = np.reshape(np.arange(9, dtype=np.float32), (3,3))\n        w.load(new_vals)\n        print('updated W:')\n        print(sess.run(w))\n<\/code><\/pre>\n\n<p>Maybe this helps?<\/p>\n"}]}
{"QuestionId":42577122,"AnswerCount":0,"Tags":"<python><theano>","CreationDate":"2017-03-03T11:01:24.137","AcceptedAnswerId":null,"OwnerUserId":7652937.0,"Title":"Memory error in theano scan","Body":"<p>I have a following function (as a part of custom Keras layer, but that's not important):<\/p>\n\n<pre><\/pre>\n\n<p>self.encoder_step performes some recursive computations. <\/p>\n\n<p>If I run that function with medium-sized params (bucket_size = 128, self.batch_size = 64, self.max_len = 128, self.hidden_dim = 256) then I get CNMEM_STATUS_OUT_OF_MEMORY. Error log with exception_verbosity = high shows that theano did allocated integer tensor 'forall_inplace,gpu,scan_fn}.0' of shape (X, 64, 128, 512) where X is some number up to 2*bucket_size.\nIt seems that theano still stores output values of scan for each step even though I don't use them. <\/p>\n\n<p>Log example:<\/p>\n\n<pre><\/pre>\n\n<p>How should I properly use scan to exclude storing of intermediate results?<\/p>\n","answers":[]}
{"QuestionId":42577773,"AnswerCount":1,"Tags":"<tensorflow><sparse-matrix>","CreationDate":"2017-03-03T11:31:42.420","AcceptedAnswerId":null,"OwnerUserId":4584564.0,"Title":"Initializing Tensors","Body":"<pre><\/pre>\n\n<p>I get the error message<\/p>\n\n<pre><\/pre>\n\n<p>Isn't it possible to jus construct two lists of indexes and values to them? I have used coo_matrix before and it solves this very good. Any help?<\/p>\n\n<p>EDIT:\nI solved it by creating a csr_matrix that I used the function sort_indices() then I converted it to coo_matrix. From there I just create a SparseTensor <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42583530","CreationDate":"2017-03-03T16:08:40.493","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>Just remove the duplicate <code>[1,9]<\/code>  in <code>indices<\/code>:<\/p>\n\n<pre><code>from __future__ import print_function\nimport tensorflow as tf\nimport numpy as np\ntf_coo = tf.SparseTensor(indices=np.array([[0, 0, 0, 1, 2, 3, 9],\n                                            [1, 4, 9, 9, 9, 9, 9]]).T,\n                        values=[1, 2, 3, 1,1,1,1],\n                        dense_shape=[10, 10])\n\nprint('tf_coo: {0}'.format(tf_coo))\nprint('tf_coo.get_shape(): {0}'.format(tf_coo.get_shape()))\n<\/code><\/pre>\n"}]}
{"QuestionId":42578096,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-03-03T11:47:37.747","AcceptedAnswerId":42585623.0,"OwnerUserId":41977.0,"Title":"Low accuracy in Deep Neural Network with Tensorflow","Body":"<p>I am following the third Jupyter notebook on <a href=\"https:\/\/github.com\/jdwittenauer\/ipython-notebooks\/blob\/master\/notebooks\/tensorflow\/Tensorflow-3-Regularization.ipynb\" rel=\"nofollow noreferrer\">Tensorflow examples<\/a>.<\/p>\n\n<p>Running problem 4, I tried to implement a function which builds automatically a number of hidden layers, without manually coding the configuration of each layer.<\/p>\n\n<p>However, the model runs providing very low accuracy (10%) so I thought that maybe such function could not be compatible with the graph builder of Tensorflow.<\/p>\n\n<p>My code is the following:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42578407","CreationDate":"2017-03-03T12:02:49.137","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Weight regularization is stronger with more layers. Therefore you could try to reduce the regularization and see if the accuracy increases.<\/p>\n"},{"AnswerId":"42585623","CreationDate":"2017-03-03T18:03:25.890","ParentId":null,"OwnerUserId":"41977","Title":null,"Body":"<p>The problem was caused by <code>nan<\/code> in loss function and weights, as described in <a href=\"https:\/\/stackoverflow.com\/q\/37448557\/41977\">this question<\/a>.<\/p>\n\n<p>By introducing a different standard deviation for each weights tensor based on its dimensions (as described <a href=\"https:\/\/stackoverflow.com\/a\/37755443\/41977\">in this answer<\/a> and originally in He <em>et al.<\/em> [1]) I was able to train successfully the network.<\/p>\n\n<p>[1]: He <em>et al.<\/em> (2015) <a href=\"https:\/\/arxiv.org\/pdf\/1502.01852.pdf\" rel=\"nofollow noreferrer\">Delving Deep into Rectifiers:\nSurpassing Human-Level Performance on ImageNet Classification<\/a><\/p>\n"}]}
{"QuestionId":42578532,"AnswerCount":1,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2017-03-03T12:08:56.697","AcceptedAnswerId":null,"OwnerUserId":7652692.0,"Title":"How to generate\/read sparse sequence labels for CTC loss within Tensorflow?","Body":"<p>From a list of word images with their transcriptions, I am trying to create and read sparse sequence labels (for ) using a , avoiding <\/p>\n\n<ol>\n<li><p>serializing pre-packaged training data to disk in\n format <\/p><\/li>\n<li><p>the apparent limitations of ,  <\/p><\/li>\n<li><p>any unnecessary or premature padding, and<\/p><\/li>\n<li><p>reading the entire data set to RAM.<\/p><\/li>\n<\/ol>\n\n<p>The main issue seems to be converting a string to the sequence of labels (a ) needed for .<\/p>\n\n<p>For example, with the character set in the (ordered) range , I'd want to convert the text label string  to the sequence label class list .<\/p>\n\n<p>Each example image I want to read contains the text as part of the filename, so it's straightforward to extract and do the conversion in straight up python. (If there's a way to do it within TensorFlow computations, I haven't found it.)<\/p>\n\n<p>Several previous questions glance at these issues, but I haven't been able to integrate them successfully. For example, <\/p>\n\n<ul>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/34340489\/tensorflow-read-images-with-labels\">Tensorflow read images with labels<\/a>\nshows a straightforward framework with discrete, categorical labels,\nwhich I've begun with as a model.<\/p><\/li>\n<li><p><a href=\"https:\/\/stackoverflow.com\/questions\/36917807\/how-to-load-sparse-data-with-tensorflow\">How to load sparse data with TensorFlow?<\/a>\nnicely explains an approach for loading sparse data, but assumes\npre-packaging s. <\/p><\/li>\n<\/ul>\n\n<p>Is there a way to integrate these approaches?<\/p>\n\n<p>Another example (SO question #38012743) shows how I might delay the conversion from string to list until after dequeuing the filename for decoding, but it relies on , which has caveats. (Should I worry about them?)<\/p>\n\n<p>I recognize that \"SparseTensors don't play well with queues\" (per the tf docs), so it might be necessary to do some voodoo on the result (serialization?) before batching, or even rework where the computation happens; I'm open to that.<\/p>\n\n<p>Following MarvMind's outline, here is a basic framework with the computations I  want (iterate over lines containing example filenames, extract each label string and convert to sequence), but I have not successfully determined the \"Tensorflow\" way to do it.<\/p>\n\n<p>Thank you for the right \"tweak\", a more appropriate strategy for my goals, or an indication that  won't wreck training efficiency or something else downstream (e.g.,loading trained models for future use).<\/p>\n\n<p><strong>EDIT (+7 hours)<\/strong> I found the missing ops to patch things up. While still need to verify this connects with CTC_Loss downstream, I have checked that the edited version below correctly batches and reads in the images and sparse tensors.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42593276","CreationDate":"2017-03-04T07:14:09.390","ParentId":null,"OwnerUserId":"7652692","Title":null,"Body":"<p>The key ideas seem to be creating a <code>SparseTensorValue<\/code> from the data wanted, pass it through <code>tf.convert_to_tensor_or_sparse_tensor<\/code> and then (if you want to batch the data) serialize it with <code>tf.serialize_sparse<\/code>. After batching, you can restore the values with <code>tf.deserialize_many_sparse<\/code>.<\/p>\n\n<p>Here's the outline. Create the sparse values, convert to tensor, and serialize:<\/p>\n\n<pre><code>indices = [[i] for i in range(0,len(text))]\nvalues = [out_charset.index(c) for c in list(text)]\nshape = [len(text)]\nlabel = tf.SparseTensorValue(indices,values,shape)\nlabel = tf.convert_to_tensor_or_sparse_tensor(label)\nlabel = tf.serialize_sparse(label) # needed for batching\n<\/code><\/pre>\n\n<p>Then, you can do the batching and deserialize:<\/p>\n\n<pre><code>image,label = tf.train.batch([image,label],dynamic_pad=True)\nlabel = tf.deserialize_many_sparse(label,tf.int32)\n<\/code><\/pre>\n"}]}
{"QuestionId":42579170,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-03T12:41:46.250","AcceptedAnswerId":null,"OwnerUserId":3568055.0,"Title":"Tensorflow loading data randomly slows down during training","Body":"<p>I have loaded all the training data in the memory, which only consumes 7% of the total memory. And the following framework is used to train the model:<\/p>\n\n<pre><\/pre>\n\n<p>However, the training process randomly slows down when the dimensionality of the feature is high.\nThe diagnoses are as following:<\/p>\n\n<ol>\n<li>The graph is predefined and is not the case in <a href=\"https:\/\/stackoverflow.com\/questions\/36245481\/tensorflow-slow-performance-when-getting-gradients-at-inputs\">tensorflow slow performance<\/a>.<\/li>\n<li>The actual running time for sess.run() is stable and the following is the timeline of a training batch, which seems normal. <a href=\"https:\/\/i.stack.imgur.com\/bt627.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/bt627.png\" alt=\"timeline of a training batch\"><\/a><\/li>\n<li>The slow part happened in data.yield_trn_batch(). At the begining, it took 0.01s to load one minibatch, but after several epoches it became unstable and sometimes took over 1s to load a minibatch. However, when I commented the sess.run() and purely run the data.yield_trn_batch(), it was fast as normal. I don't use the queue so it might not be the situation in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3009\" rel=\"nofollow noreferrer\">dequeue many operation very slow<\/a>.<\/li>\n<\/ol>\n\n<p>I guess the graph running process have affected the data loading but don't know why and how to solve this (maybe using another thread to load data?). Can anybody fix the problem?<\/p>\n","answers":[{"AnswerId":"42591223","CreationDate":"2017-03-04T02:01:51.813","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Based on our conversation in the comments, it appears that the slow down is due to memory pressure caused by allocating a large number of NumPy arrays. Although the NumPy arrays are properly garbage collected properly when they are no longer used, the default <code>malloc()<\/code> implementation will not reuse them, and gradually increase the size of the heap (and the virtual size of the process), by calling the <code>brk()<\/code> system call.<\/p>\n\n<p>One workaround is to switch the allocator library, which can fix address space leaks: use the <code>tcmalloc<\/code> allocator instead of the default <code>malloc()<\/code> for your TensorFlow process. The allocation policy in <code>tcmalloc<\/code> is better suited to allocating and recycling buffers of the same size repeatedly, and it will not need to increase the size of the head over time, which should lead to better performance.<\/p>\n"}]}
{"QuestionId":42579731,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-03-03T13:08:32.473","AcceptedAnswerId":43432823.0,"OwnerUserId":4333440.0,"Title":"Tensorflow Convnet Strange Softmax Output","Body":"<p>The output for my Convnet has been very unusual. When imprinted out the output vector of the forward propagation results, it was perfectly [0, 0, 0, 1], constant for an entire label in the dataset. I suspect there's an error in my construction.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43432823","CreationDate":"2017-04-16T01:27:57.580","ParentId":null,"OwnerUserId":"4333440","Title":null,"Body":"<p>The main issue lies in the fact that Softmax is called twice.<\/p>\n\n<p>Softmax was called in the forward_propagation part of the code, and that was placed in the Tensorflow cross entropy code, which already contains a softmax, hence causing an anomaly in the outputs.<\/p>\n"}]}
{"QuestionId":42580270,"AnswerCount":2,"Tags":"<c++><tensorflow>","CreationDate":"2017-03-03T13:32:18.113","AcceptedAnswerId":null,"OwnerUserId":7653715.0,"Title":"Build Tensorflow r1.0 c++ API","Body":"<p>I trying to build Tensorflow r1.0 for c++\nI was able to compile the libtensorflow.so but not all the headers generated.\nAll the ops headers missing except from standard_ops.h and const_ops.h which include in them all the ops headers that missing.<\/p>\n\n<p>Im using Ubuntu 16.04.<\/p>\n","answers":[{"AnswerId":"42607023","CreationDate":"2017-03-05T10:08:30.683","ParentId":null,"OwnerUserId":"6708503","Title":null,"Body":"<p>Code for most ops is generated during the build. When using <code>bazel<\/code>, you should fine the generated header files in: <code>bazel-genfiles\/tensorflow\/cc\/ops<\/code>, e.g., <code>bazel-genfiles\/tensorflow\/cc\/ops\/math_ops.h<\/code>.<\/p>\n\n<p>Hope that helps.<\/p>\n"},{"AnswerId":"44625966","CreationDate":"2017-06-19T08:46:46.933","ParentId":null,"OwnerUserId":"4932728","Title":null,"Body":"<p>If you don't mind using CMake, you can use <a href=\"https:\/\/github.com\/FloopCZ\/tensorflow_cc\" rel=\"nofollow noreferrer\">tensorflow_cc<\/a> project. It builds and installs TF C++ API for you, along with convenient CMake targets you can link against.<\/p>\n"}]}
{"QuestionId":42581099,"AnswerCount":1,"Tags":"<machine-learning><neural-network><keras><conv-neural-network><lstm>","CreationDate":"2017-03-03T14:11:36.827","AcceptedAnswerId":42581375.0,"OwnerUserId":1611511.0,"Title":"Is it possible to use Image Preprocessing with image sequences?","Body":"<p>I'm trying to use Keras' builtin image preprocessing capabilities to augment images in a sequence.\nMy dataset has the shape  with 13200 sequences, each consisting of 4 168x168px grayscale images.<\/p>\n\n<p>When trying to run  on my dataset I get:<\/p>\n\n<pre><\/pre>\n\n<p>I'm assuming  is not able to handle my sequences of 4 images per sample correctly. Is there any way to do this? <\/p>\n","answers":[{"AnswerId":"42581375","CreationDate":"2017-03-03T14:25:34.937","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Try to define a new generator by:<\/p>\n\n<pre><code>def sequence_image_generator(x, y, batch_size, generator, seq_len=4):\n    new_y = numpy.repeat(y, seq_len, axis = 0)\n    helper_flow = generator.flow(x.reshape((x.shape[0] * seq_len,\n                                            x.shape[2],\n                                            x.shape[3],\n                                            x.shape[4])),\n                                 new_y,\n                                 batch_size=seq_len * batch_size)\n    for x_temp, y_temp in helper_flow:\n        yield x_temp.reshape((x_temp.shape[0] \/ seq_len, \n                              seq_len, \n                              x.shape[2],\n                              x.shape[3],\n                              x.shape[4])), y_temp[::seq_len,:]\n<\/code><\/pre>\n"}]}
{"QuestionId":42582358,"AnswerCount":1,"Tags":"<win-universal-app><zxing><torch><flashlight>","CreationDate":"2017-03-03T15:09:56.207","AcceptedAnswerId":42623081.0,"OwnerUserId":7387314.0,"Title":"Does default Zxing.Net.Mobile overlay contain a flash\/torch button?","Body":"<p>I am using the Zxing.Net.Mobile package within my Windows 10 universal app. I use the Default overlay, which is quite nice in my eyes. Now I want to add Flash\/Torch funtionality to my application. I only found Solutions defining a new customn overlay. Is it not possible to activate a Torch button in the default overlay? I found a scanner.FlashButtonText Definition, so i think a flashbutton should somehow be activatable and i just dont know the Name of this Parameter. Can someone help me with this? Thanks.<\/p>\n\n<p>my Code Looks like this:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42623081","CreationDate":"2017-03-06T10:23:23.367","ParentId":null,"OwnerUserId":"5506161","Title":null,"Body":"<p>Although <code>MobileBarcodeScanner<\/code> has a <code>FlashButtonText<\/code> property, but according to <a href=\"https:\/\/github.com\/Redth\/ZXing.Net.Mobile\" rel=\"nofollow noreferrer\">ZXing.Net.Mobile's source code<\/a>, Flash\/Torch function is not supported in the default overlay by now.<\/p>\n\n<p>The default overlay is set by <a href=\"https:\/\/github.com\/Redth\/ZXing.Net.Mobile\/blob\/master\/Source\/ZXing.Net.Mobile.WindowsUniversal\/ZXingScannerControl.xaml\" rel=\"nofollow noreferrer\">ZXingScannerControl<\/a> and if we look at its source code, we will find it does have a button named <em>\"buttonToggleFlash\"<\/em>. However this button is invisible and there is no property can control its visibility.<\/p>\n\n<p><a href=\"https:\/\/github.com\/Redth\/ZXing.Net.Mobile\/blob\/master\/Source\/ZXing.Net.Mobile.WindowsUniversal\/ZXingScannerControl.xaml#L40-L42\" rel=\"nofollow noreferrer\">source code<\/a>:<\/p>\n\n<pre><code>&lt;Border Grid.Row=\"0\" Visibility=\"Collapsed\" VerticalAlignment=\"Top\" HorizontalAlignment=\"Right\" Width=\"120\" Height=\"80\"&gt;\n    &lt;Button x:Name=\"buttonToggleFlash\" Click=\"buttonToggleFlash_Click\"&gt;Flash&lt;\/Button&gt;\n&lt;\/Border&gt; \n<\/code><\/pre>\n\n<p>So to support Flash\/Torch function, you need a customn overlay. Or you can edit the source code and rebuild ZXing.Net.Mobile for your purpose.<\/p>\n"}]}
{"QuestionId":42582484,"AnswerCount":1,"Tags":"<machine-learning><keras><lstm>","CreationDate":"2017-03-03T15:15:17.817","AcceptedAnswerId":42585893.0,"OwnerUserId":7605750.0,"Title":"How not to output a sequence with `TimeDistributed` layer in Keras?","Body":"<p>I am currently work on LSTM with Keras and I have a question about  layer.<\/p>\n\n<p>Let's say that I have a   layer which takes as input something like . It will output something like. <\/p>\n\n<p>I would like instead to output something like . Is it possible ?<\/p>\n\n<p>It would be to stack a LSTM layer with  to a dense layer (using  layer) and then to go back to \"classical\" dense layer that accept input like .<\/p>\n\n<p>Thanks in advance !<\/p>\n\n<p>Benoit <\/p>\n","answers":[{"AnswerId":"42585893","CreationDate":"2017-03-03T18:19:11.020","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>I'm not sure to exactly understand what you want so I will put here a network that I think is what you want. If it isn't, please edit your question with a draft of the network you want and the shapes at each steps. \nIt would be also easier to know what you want to achieve with this network. <\/p>\n\n<pre><code>model = sequential()\n# shape = (None,timesteps, num_feat1)\nmodel.add(TimeDistributed(Dense(num_feat2))\n# shape = (None,timesteps, num_feat2)\nmodel.add(LSTM(1, return_sequence=True))\n# shape = (None, timesteps, 1)\nmodel.add(Flatten())\n# shape = (None, timesteps)\nmodel.add(Dense(num_outputs_desired))\n# shape = (None, outputs)\n<\/code><\/pre>\n\n<p>is that what  you want? This (1) transforms the initial features equally on each timestep with a dense layer time distributed, (2) processes the sequence with an lstm, returning 1 value at each step, (3) use a dense layer to transform the sequence of values to a desired output (which I don't know how it should be, what is the aim of your model?).<\/p>\n"}]}
{"QuestionId":42582760,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-03-03T15:29:50.683","AcceptedAnswerId":null,"OwnerUserId":1913898.0,"Title":"tensorflow: run same metrics multiple times with different input","Body":"<p>I want to save prediction and recall to tensorboard every 100 iterations during training, and I need to compute the metrics for both training and test dataset. Here is the code that I use:<\/p>\n\n<pre><\/pre>\n\n<p>Note that to run and get the summary of same variables on both train and test data, I define two summary ops for training and testing, put them in different collections, and run the two summary ops separately for train and test data. <\/p>\n\n<p>I also think I should re-initialize the local variables for the update_op, because the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/metrics\/sensitivity_at_specificity\" rel=\"nofollow noreferrer\">official doc<\/a> says <\/p>\n\n<blockquote>\n  <p>It sets the sum and count variables to zero.<\/p>\n<\/blockquote>\n\n<p>My question: is this the best way to run ops on both train and test set? <\/p>\n\n<p>The TF document also mentioned that <\/p>\n\n<blockquote>\n  <p>Note that when evaluating the same metric multiple times on different inputs, one must specify the scope of each metric to avoid accumulating the results together:<\/p>\n<\/blockquote>\n\n<pre><\/pre>\n\n<p>But that seems a different approach with mine, becuase I just define a single variable  with just one  variable. And it seems I don't need to worry about this issue and don't need to add  option?<\/p>\n","answers":[{"AnswerId":"52838889","CreationDate":"2018-10-16T15:21:54.207","ParentId":null,"OwnerUserId":"5224881","Title":null,"Body":"<p>The most important thing to remember when using <code>tf.metrics<\/code> functions is that all meetrics here are running metrics, calculating running metrics over multiple runs. For details, see this article: <a href=\"http:\/\/ronny.rest\/blog\/post_2017_09_11_tf_metrics\/\" rel=\"nofollow noreferrer\">http:\/\/ronny.rest\/blog\/post_2017_09_11_tf_metrics\/<\/a>\nSo here you are calculating cumulative precision over both training and testing data, because update_op for precision is the same for both of them and both datasets, training and testing, are updating same running precision.<\/p>\n"}]}
{"QuestionId":42584681,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-03-03T17:05:40.657","AcceptedAnswerId":42589028.0,"OwnerUserId":4542398.0,"Title":"How to load image data by keys in keras?","Body":"<p>I've been working on a sequential model which takes images as inputs. However, different thing is that the input images are actually determined by keys.<br>\nFor example, the training sequence is (You may assume <em>fi<\/em> is frame id of a video)   <\/p>\n\n<pre><\/pre>\n\n<p>and the corresponding image sequence is   <\/p>\n\n<pre><\/pre>\n\n<p>where M is a map storing {<em>fi->image<\/em>} mapping.<br>\nSuppose in the next batch, my training sequence become   <\/p>\n\n<pre><\/pre>\n\n<p>and the image sequence becomes  <\/p>\n\n<pre><\/pre>\n\n<p>As you can see, if I directly save the image sequences into disk, there are lot of redundancies (in the above case, M[f2] to M[fn] are saved twice). So it seems necessary that the images are referenced by keys and thus imagedataloader class can not be used.<br>\n[<strong>EDIT<\/strong>]<br>\nMy model is a <strong>2-class<\/strong> classifier takes image sequences as input, in which the images are mapped with the <em>frame id<\/em>(<em>fi<\/em>). Whether an image sequence is positive or negative is pre-generated in my <em>data_preprocess<\/em> code.<br>\n<strong>Positive<\/strong> samples may look like this:  <\/p>\n\n<pre><\/pre>\n\n<p>While <strong>negative<\/strong> samples look like this:  <\/p>\n\n<pre><\/pre>\n\n<p>So, it is not like image classifying problem, where an image has exactly a fixed label. In my case, every image will be used many times and their being positive or negative are together determined by the whole sequence, but not itself.<br>\n[<strong>EDIT II<\/strong>]<br>\nThe images are frames of N videos and are stored on disk like this:  <\/p>\n\n<pre><\/pre>\n\n<p>What I'd like to do is, given two sequences of frames\/images of scenes, the model predicts whether the two scenes are of the same kind.<br>\nSince a video may contain a long time span for each scene, I divide the whole sequence of a scene into a number of non-overlap sub-sequences (omit the indexes of videos):<\/p>\n\n<pre><\/pre>\n\n<p>Then, I randomly generate positive samples Pi (pairs of sub-sequences generated from the <strong>same<\/strong> sequence), like:<\/p>\n\n<pre><\/pre>\n\n<p>For negative samples, I generate pairs of sub-sequences (Ni) from <strong>different<\/strong> scenes:<\/p>\n\n<pre><\/pre>\n\n<p>It is obvious that one frame\/image can occur multiple times in different training samples. E.g. in the above case, both N2 and P1 contain sub-sequence i_2. So I choose to save the generated sample pairs by sequences of frame id(<strong>fi<\/strong>) and during training, fetch the corresponding frames\/images of a sequence by frame id(<strong>fi<\/strong>).<br>\nHow should I do it elegantly with Keras? <\/p>\n","answers":[{"AnswerId":"42589028","CreationDate":"2017-03-03T21:49:48.210","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Not sure how you build your sequences but have you considered using the <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\"><code>ImageDataGenerator<\/code> from <code>keras.preprocessing.image<\/code><\/a> ?<\/p>\n\n<p>Once you have built this object with whatever parameters  you want, you can use the <code>flow_from_directory(directory_path)<\/code> method. Once you have done this, you can use the filename attribute of this object :<\/p>\n\n<pre><code>my_generator = ImageDataGenerator(...)\nmy_generator.flow_from_directory(path_dir)\nlist_of_file_names = my_generator.filename\n<\/code><\/pre>\n\n<p>you now have a mapping between indexes of the list and the elements(=file_paths) of the list.<\/p>\n\n<p>I hope this helps?<\/p>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>From this, you can build a mapping a dictionnary<\/p>\n\n<pre><code>map_images = {str(os.path.splitext(os.path.split(file_path)[1])[0]): file_path for file_path in list_of_file_names}\n<\/code><\/pre>\n\n<p>This takes the <code>file_path<\/code> retrieved from you image folder using ImageDataGenerator, it extracts the file name, removes the file extension and transforms the name of the file into an <code>string<\/code> which is your frame_id.<\/p>\n\n<p>You now have a map between <code>frame_id<\/code> and <code>file_path<\/code> that you can use with <a href=\"https:\/\/keras.io\/preprocessing\/image\/\" rel=\"nofollow noreferrer\"><code>load_img()<\/code> and <code>img_to_array()<\/code> from <code>keras.preprocessing.image<\/code><\/a><\/p>\n\n<p>the function <code>load_img()<\/code> is defined like this and returns a PIL image instance:<\/p>\n\n<blockquote>\n  <p><code>def load_img(path, grayscale=False, target_size=None):\n      \"\"\"Loads an image into PIL format.\n      # Arguments\n          path: Path to image file\n          grayscale: Boolean, whether to load the image as grayscale.\n          target_size: Either 'None' (default to original size)\n              or tuple of ints '(img_height, img_width)'.\n      # Returns\n          A PIL Image instance.\n      # Raises\n          ImportError: if PIL is not available.\n      \"\"\"<\/code><\/p>\n<\/blockquote>\n\n<p>Then <code>img_to_array()<\/code> is defined like this and returns a 3D numpy array to feed your model:<\/p>\n\n<blockquote>\n  <p><code>def img_to_array(img, dim_ordering='default'):\n      \"\"\"Converts a PIL Image instance to a Numpy array.\n      # Arguments\n          img: PIL Image instance.\n          dim_ordering: Image data format.\n      # Returns\n          A 3D Numpy array.\n      # Raises\n          ValueError: if invalid 'img' or 'dim_ordering' is passed.\n      \"\"\"\n  <\/code><\/p>\n<\/blockquote>\n\n<p>So to summarize : 1 build a mapping between your frame_id and the path of the corresponding file. Then load the file using <code>img_load()<\/code> and <code>img_to_array()<\/code>. I hope I have understood your question correctly !<\/p>\n\n<p><strong>EDIT 2:<\/strong><\/p>\n\n<p>Seeing your new edit, now that I understand the structure of your file system, we can even add the video in your dictionary like this :<\/p>\n\n<pre><code># list of video_id of each frame\nvideos = my_generator.classes\n# mapping of the frame_id to path_of_file and vid_id\nmap_images = {str(os.path.splitext(os.path.split(file_path)[1])[0]): (file_path, vid_id) for file_path,vid_id in zip(list_of_file_names,videos) }\n<\/code><\/pre>\n"}]}
{"QuestionId":42585327,"AnswerCount":1,"Tags":"<neural-network><deep-learning><theano>","CreationDate":"2017-03-03T17:44:10.570","AcceptedAnswerId":null,"OwnerUserId":5501534.0,"Title":"How to convert theano symbolic vector of interger labels to binarized symbolic matrix label?","Body":"<p>I am a newbie in Theano and trying to convert labels of theano vector to theano matrix. One of the NN application required labels in the form of a binary matrix of MxN where M is the number of samples and N is the number of class. e.g., labels=[0,1,2,1,2,3], the banary_labels should be \n[[1,0,0,0]; [0,1,0,0]; [0,0,1,0]; [0,1,0,0]; [0,0,1,0]; [0,0,0,1]]<\/p>\n\n<p><em>I have written the following code but unable to figure it what is the problem<\/em><\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42621489","CreationDate":"2017-03-06T09:06:32.217","ParentId":null,"OwnerUserId":"6553672","Title":null,"Body":"<p>This is called one-hot encoding. Have a look at Keras implementation (to_categorical) <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/utils\/np_utils.py\" rel=\"nofollow noreferrer\">here.<\/a><\/p>\n"}]}
{"QuestionId":42585356,"AnswerCount":1,"Tags":"<neural-network><time-series><keras><lstm><recurrent-neural-network>","CreationDate":"2017-03-03T17:46:07.077","AcceptedAnswerId":null,"OwnerUserId":7654638.0,"Title":"How to construct input data to LSTM for time series multi-step horizon with external features?","Body":"<p>I'm trying to use LSTM to do store sales forecast. Here is how my raw data look like:<\/p>\n\n<pre><\/pre>\n\n<p>I need to forecast for the next 10 days' sales. In this example, I will need to forecast the store sales from 01-01-2017 to 01-10-2017. I know how to use other time series model or regression model to solve this problem, but I want to know if RNN-LSTM is a good candidate for it. <\/p>\n\n<p>I started by taking <strong><em>only storeID=1 data<\/em><\/strong> to test the LSTM. If my data <strong><em>only have Date and Sales<\/em><\/strong>. I will construct my trainX and trainY in this way (please correct me if I'm wrong):<\/p>\n\n<pre><\/pre>\n\n<p>After reshaping the two<\/p>\n\n<pre><\/pre>\n\n<p><strong>Question1:<\/strong> In this case, [samples, time steps, features] = [300, 1, 20]. Is this right? Or should I construct the sample as [300, 20, 1] ?<\/p>\n\n<p><strong>Question2:<\/strong> I do want to use other information in the raw data like Temperature, StoreType, etc. How should I construct my input data for LSTM? <\/p>\n\n<p><strong>Question3:<\/strong> So far we only discussed 1 store forecast, if I want to forecast for all the stores, how should I construct my input data then?<\/p>\n\n<p>Currently I'm flowing examples from <a href=\"http:\/\/machinelearningmastery.com\/time-series-prediction-lstm-recurrent-neural-networks-python-keras\/\" rel=\"noreferrer\">here<\/a>, but it seems not sufficient to cover the scenario that I have. I really appreciate for your help!<\/p>\n","answers":[{"AnswerId":"42599514","CreationDate":"2017-03-04T18:12:04.673","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>I was recently solving similiar problem. In your case:<\/p>\n\n<ol>\n<li><p>Input should have shape <code>(300, 20, 1)<\/code> - because you have time sequences of length <code>20<\/code> with <code>1<\/code> feature.<\/p><\/li>\n<li><p>You may do it like this:<\/p>\n\n<pre><code>sequential_input = Input(shape=(20, 1))\nfeature_input = Input(shape=(feature_nb,))\nlstm_layer = LSTM(lstm_units_1st_layer, return_sequences=True)(sequential_input)\nlstm_layer = LSTM(lstm_units_2nd_layer, return_sequences=True)(lstm_layer)\n...\nlstm_layer = LSTM(lstm_units_nth_layer, return_sequences=False)(lstm_layer)\nmerged = merge([lstm_layer, feature_input], mode='concat')\nblend = Dense(blending_units_1st_layer, activation='relu')(merged)\nblend = Dense(blending_units_2nd_layer, activation='relu')(blend)\n...\noutput = Dense(10)(blend)\n<\/code><\/pre><\/li>\n<li><p>This is the hardest part. I <strong>do not<\/strong> advise you to predict multiple shops by feeding them to a network as one feature vector. You may under simply skip this part and try to predict different shops using one model or postprocess output using e.g. some kind of graphical models or <code>PCA<\/code> on matrix where rows are day sales.<\/p><\/li>\n<\/ol>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>In order to deal with multiple sequential features you could do the following thing:<\/p>\n\n<pre><code>    sequential_input = Input(shape=(20, nb_of_sequental_features))\n    feature_input = Input(shape=(feature_nb,))\n    lstm_layer = LSTM(lstm_units_1st_layer, return_sequences=True)(sequential_input)\n    lstm_layer = LSTM(lstm_units_2nd_layer, return_sequences=True)(lstm_layer)\n    ...\n    lstm_layer = LSTM(lstm_units_nth_layer, return_sequences=False)(lstm_layer)\n    merged = merge([lstm_layer, feature_input], mode='concat')\n    blend = Dense(blending_units_1st_layer, activation='relu')(merged)\n    blend = Dense(blending_units_2nd_layer, activation='relu')(blend)\n    ...\n    output = Dense(10)(blend)\n    model = Model(input=[sequential_input, feature_input], output=output])\n<\/code><\/pre>\n\n<p>In this case your input should consist of list of two tables: <code>[sequential_data, features]<\/code> where <code>sequential_data.shape = (nb_of_examples, timesteps, sequential_features)<\/code> and <code>features.shape = (nb_of_examples, feature_nb)<\/code>. So <code>sales<\/code> or <code>temperature<\/code> should be stored in <code>sequential_features<\/code> and <code>store_type<\/code> in <code>features<\/code>.<\/p>\n"}]}
{"QuestionId":42585552,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-03T17:59:27.243","AcceptedAnswerId":42586029.0,"OwnerUserId":3713926.0,"Title":"Why do the Variables not update in this program?","Body":"<p>So, I have this bit of code:<\/p>\n\n<pre><\/pre>\n\n<p>Basically, it (at least is supposed to) draws 100 random white rectangles on a black field. But what I get is this:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/vfYjE.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/vfYjE.png\" alt=\"WTF\"><\/a><\/p>\n\n<p>I do not understand. It looks like every iteration  just resets to bing a black slate again (the rectangle on the picture is the last one drawn, judging by the coordinates). Am I not saving it somehow, or what is it that I'm doing wrong?<\/p>\n","answers":[{"AnswerId":"42586029","CreationDate":"2017-03-03T18:27:43.890","ParentId":null,"OwnerUserId":"4234910","Title":null,"Body":"<p>Change these two lines:<\/p>\n\n<pre><code>res = tf.add(res, mask)\n##\nnew_pic = sess.run([res], feed_dict={mask:c_mask})[0]\n<\/code><\/pre>\n\n<p>to <\/p>\n\n<pre><code>update_op = res.assign(tf.add(res, mask))\n##\nnew_pic = sess.run([res, update_op], feed_dict={mask:c_mask})[0]\n<\/code><\/pre>\n\n<hr>\n\n<p>A tensor should me modified with <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable#assign\" rel=\"nofollow noreferrer\">Variable.assign()<\/a>.<\/p>\n\n<p>While you intended to update the tensor \"res\", the first line doesn't update it but only creates another tensor, and the <strong>name<\/strong> \"res\" is assigned to the new tensor. Hence the original res is never updated.<\/p>\n"}]}
{"QuestionId":42585620,"AnswerCount":2,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-03-03T18:03:07.903","AcceptedAnswerId":42586519.0,"OwnerUserId":4533188.0,"Title":"Keras + TensorFlow: \u201cmodule 'tensorflow' has no attribute 'merge_all_summaries''\u201d","Body":"<p>Very similar to <a href=\"https:\/\/stackoverflow.com\/questions\/40046619\/keras-tensorflow-gives-the-error-no-attribute-control-flow-ops\">Keras + tensorflow gives the error &quot;no attribute &#39;control_flow_ops&#39;&quot;<\/a>, from the Convolutional autoencoder example from <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-autoencoders-in-keras.html<\/a> I get the error<\/p>\n\n<pre><\/pre>\n\n<p>I tried<\/p>\n\n<pre><\/pre>\n\n<p>but that did not work. What should I do?<\/p>\n\n<p>In <a href=\"https:\/\/stackoverflow.com\/questions\/42481938\/attributeerror-module-object-has-no-attribute-merge-all-summaries\">AttributeError: &#39;module&#39; object has no attribute &#39;merge_all_summaries&#39;<\/a> the error is mentioned. I also have the version 1.0.0. But what is the solution? I don't want to downgrade TensorFlow.<\/p>\n","answers":[{"AnswerId":"42585785","CreationDate":"2017-03-03T18:12:26.333","ParentId":null,"OwnerUserId":"4533188","Title":null,"Body":"<p>The answer is to migrate as appropriate. Check out <a href=\"https:\/\/www.tensorflow.org\/install\/migration\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/migration<\/a>. There you see that<\/p>\n\n<pre><code>- tf.merge_summary\n    - should be renamed to tf.summary.merge\n- tf.train.SummaryWriter\n    - should be renamed to tf.summary.FileWriter\n<\/code><\/pre>\n\n<p>(Actually SummaryWriter has also been changed.) So instead of <\/p>\n\n<pre><code>import tensorflow as tf\ntf.merge_all_summaries = tf\n<\/code><\/pre>\n\n<p>you should write<\/p>\n\n<pre><code>import tensorflow as tf\ntf.merge_all_summaries = tf.summary.merge_all\ntf.train.SummaryWriter = tf.summary.FileWriter\n<\/code><\/pre>\n"},{"AnswerId":"42586519","CreationDate":"2017-03-03T19:00:04.123","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Make42 is absolutely correct that the changes they describe in <a href=\"https:\/\/stackoverflow.com\/a\/42585785\/3574081\">their answer<\/a> must be made in order to migrate a codebase to work with TensorFlow 1.0. However, the errors you are seeing are in the Keras library itself. Fortunately, these errors have been fixed in the Keras codebase <a href=\"https:\/\/github.com\/fchollet\/keras\/commit\/aa18604fec4a309658bcc83b6aa595b5f0838d36\" rel=\"nofollow noreferrer\">since January 2017<\/a>, so upgrading to <a href=\"https:\/\/pypi.python.org\/pypi\/Keras\/1.2.2\" rel=\"nofollow noreferrer\">Keras 1.2.2<\/a> or later will fix the error for you.<\/p>\n"}]}
{"QuestionId":42586475,"AnswerCount":3,"Tags":"<keras><deep-learning>","CreationDate":"2017-03-03T18:57:15.210","AcceptedAnswerId":42587192.0,"OwnerUserId":604734.0,"Title":"Is it possible to automatically infer the class_weight from flow_from_directory in Keras?","Body":"<p>I have an imbalanced multi-class dataset and I want to use the  argument from  to give weights to the classes according to the number of images of each class. I'm using  to load the dataset from a directory.<\/p>\n\n<p>Is it possible to directly infer the  argument from the  object?<\/p>\n","answers":[{"AnswerId":"52338431","CreationDate":"2018-09-14T19:51:18.303","ParentId":null,"OwnerUserId":"6186010","Title":null,"Body":"<p>I tried both solutions and the <code>sklearn.utils.class_weight<\/code>\n one gives better accuracy though I am not sure why. They do not both yield the same class weights. <\/p>\n"},{"AnswerId":"51483119","CreationDate":"2018-07-23T16:09:32.290","ParentId":null,"OwnerUserId":"10123561","Title":null,"Body":"<p>Alternatively, you can simply do: <\/p>\n\n<pre><code>from sklearn.utils import class_weight\nimport numpy as np\n\nclass_weights = class_weight.compute_class_weight(\n               'balanced',\n                np.unique(train_generator.classes), \n                train_generator.classes)\n<\/code><\/pre>\n\n<p>You can then set (as per comment above): <\/p>\n\n<pre><code>model.fit_generator(..., class_weight=class_weights)\n<\/code><\/pre>\n"},{"AnswerId":"42587192","CreationDate":"2017-03-03T19:43:16.227","ParentId":null,"OwnerUserId":"604734","Title":null,"Body":"<p>Just figured out a way of achieving this.<\/p>\n\n<pre><code>from collections import Counter\ntrain_datagen = ImageDataGenerator()\ntrain_generator = train_datagen.flow_from_directory(...)\n\ncounter = Counter(train_generator.classes)                          \nmax_val = float(max(counter.values()))       \nclass_weights = {class_id : max_val\/num_images for class_id, num_images in counter.items()}                     \n\nmodel.fit_generator(...,\n                    class_weight=class_weights)\n<\/code><\/pre>\n\n<p><code>train_generator.classes<\/code> is a list of classes for each image.\n<code>Counter(train_generator.classes)<\/code> creates a counter of the number of images in each class.<\/p>\n\n<p>Note that these weights may not be good for convergence, but you can use it as a base for other type of weighting based on occurrence.<\/p>\n\n<p><em>This answer was inspired by: <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/1875#issuecomment-273752868\" rel=\"noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/1875#issuecomment-273752868<\/a><\/em><\/p>\n"}]}
{"QuestionId":42586680,"AnswerCount":1,"Tags":"<github><docker><tensorflow><dockerhub>","CreationDate":"2017-03-03T19:11:22.137","AcceptedAnswerId":42587064.0,"OwnerUserId":2026417.0,"Title":"Which Tensorflow Docker image to use?","Body":"<p>I just noticed that there are two Tensorflow images in existence.<\/p>\n\n<p>There's this one:<\/p>\n\n<p><a href=\"https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/\" rel=\"nofollow noreferrer\">tensorflow\/tensorflow on Docker Hub<\/a><\/p>\n\n<p>and and then this one (two references):<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/tools\/docker\" rel=\"nofollow noreferrer\">gcr.io\/tensorflow\/tensorflow on GitHub<\/a><\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/install\/install_linux#installing_with_docker\" rel=\"nofollow noreferrer\">gcr.io\/tensorflow\/tensorflow on Tesorflow.com<\/a><\/p>\n\n<p>Is there a relevant difference?<\/p>\n","answers":[{"AnswerId":"42587064","CreationDate":"2017-03-03T19:36:32.593","ParentId":null,"OwnerUserId":"2116766","Title":null,"Body":"<p>The release tags, e.g., <code>latest<\/code>, <code>1.0.0-gpu<\/code>, are identical between gcr.io and Docker hub.<\/p>\n\n<p>In other words, if you do<\/p>\n\n<p><code>nvidia-docker pull gcr.io\/tensorflow\/tensorflow:1.0.0-gpu<\/code> (from gcr.io)<\/p>\n\n<p>and<\/p>\n\n<p><code>nvidia-docker pull tensorflow\/tensorflow:1.0.0-gpu<\/code> (from Docker Hub)<\/p>\n\n<p>you should get the same image.<\/p>\n\n<p>However, TensorFlow pushes nightly images to only Docker Hub, see:\n<a href=\"https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/tags\/\" rel=\"nofollow noreferrer\">https:\/\/hub.docker.com\/r\/tensorflow\/tensorflow\/tags\/<\/a><\/p>\n\n<p>So if you want to use the latest nightly build, omit the <code>gcr.io<\/code> part, e.g.,\n<code>\ndocker pull tensorflow\/tensorflow:nightly-py3\n<\/code><\/p>\n"}]}
{"QuestionId":42588145,"AnswerCount":1,"Tags":"<python><tensorflow><keras>","CreationDate":"2017-03-03T20:45:26.350","AcceptedAnswerId":42588518.0,"OwnerUserId":4533188.0,"Title":"Keras: What dimensions exactly do I need to change?","Body":"<h1>The problem<\/h1>\n\n<p>I am trying to build the convolutional autoencoder in <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">https:\/\/blog.keras.io\/building-autoencoders-in-keras.html<\/a> but in the code part<\/p>\n\n<pre><\/pre>\n\n<p>I get the error<\/p>\n\n<pre><\/pre>\n\n<p>I understand that the problem is the wrong dimensions and I need to set<\/p>\n\n<pre><\/pre>\n\n<p>correctly - using 'th' mode (even though using TensorFlow as backend). However I have no idea how to do it. What\/Where is the screw I need to turn exactly?<\/p>\n\n<h1>More information<\/h1>\n\n<h3>Input data<\/h3>\n\n<p>The data I load (not the MNIST, but has similar dimensions) has the shape:<\/p>\n\n<pre><\/pre>\n\n<h3>Code<\/h3>\n\n<pre><\/pre>\n\n<h3>Complete error message<\/h3>\n\n<p>[...] is the path to the anaconda environment<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42588518","CreationDate":"2017-03-03T21:13:44.220","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>Try to set this in all your convolutional, upsampling and maxpooling layers: <\/p>\n\n<pre><code>Convolution2D(..., dim_ordering = 'th')\nUpSampling2D(..., dim_ordering='th')\nMaxPooling2D(..., dim_ordering='th')\n<\/code><\/pre>\n\n<p>You can find informations about this feature <a href=\"https:\/\/keras.io\/layers\/convolutional\/#convolution2d\" rel=\"nofollow noreferrer\">in the doc<\/a>.<\/p>\n\n<p>If this doesn't help do you mind sharing more info about your set up and versions? <\/p>\n\n<p>**EDIT : **<\/p>\n\n<p>This is why you should always include executable code in your questions. <\/p>\n\n<p>Your mistake is that you copy pasted the code from the <a href=\"https:\/\/blog.keras.io\/building-autoencoders-in-keras.html\" rel=\"nofollow noreferrer\">tutorial<\/a> in a wrong way.<\/p>\n\n<p>When they build this decoder :<\/p>\n\n<pre><code>encoded_input = Input(shape=(1, 28, 28))\n# retrieve the last layer of the autoencoder model\ndecoder_layer = autoencoder.layers[-1]\n# create the decoder model\ndecoder = Model(input=encoded_input,            output=decoder_layer(encoded_input))\n<\/code><\/pre>\n\n<p>It's made for the previous autoencoder :<\/p>\n\n<pre><code># 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<\/code><\/pre>\n\n<p>Their <code>encoded_input<\/code> has a shape that doesn't match yours because you are using the convolutionnal autoencoder. So you are trying to match :<\/p>\n\n<pre><code>encoded = MaxPooling2D((2, 2), border_mode='same', dim_ordering=dim_ordering)(x)\n# at this point the representation is (8, 4, 4) i.e. 128-dimensional \n<\/code><\/pre>\n\n<p>an input which has a shape (1,28,28) with something that thas a shape (8,4,4). The shape of the encoded image is different between the dense autoencoder and the convolutionnal autoencoder. Your input has to match the shape of the encoded tensors for the model that you are using. <\/p>\n\n<p>So first thing to change is this line :<\/p>\n\n<pre><code>encoded_input = Input(shape=(8,4,4))\n<\/code><\/pre>\n\n<p>Now, what they do here :<\/p>\n\n<pre><code>decoder_layer = autoencoder.layers[-1]\n<\/code><\/pre>\n\n<p>is to select the last layer of the model they were using (the dense autoencoder). Why do they do that? It's because their decoder is only one layer big... So they take it and apply the input to it, that's their decoder. Now your situation is a bit trickier, you want to use all the 'decoder-side' layers of your convolutionnal autoencoder. This is how you should do it :<\/p>\n\n<pre><code> # Retrieve and apply the encoded input to the first layer of the decoder \n # it corresponds to Convolution2D(8, 3, 3,...) part of your model if you count the layers.\n decoder_layer = autoencoder.layers[-7](encoded_input)\n # loop over the next layers of the decoder until the last one\n for i in range(-6,0):\n     decoder_layer = autoencoder.layers[i](decoder_layer)\n # create the decoder model\n decoder = Model(input=encoded_input, output=decoder_layer)\n<\/code><\/pre>\n\n<p>Is it clearer? You cannot copy\/paste code and build up complex stuff with partial code without knowing what those part of code mean.<\/p>\n"}]}
{"QuestionId":42588257,"AnswerCount":2,"Tags":"<docker><tensorflow><jupyter-notebook>","CreationDate":"2017-03-03T20:53:39.573","AcceptedAnswerId":null,"OwnerUserId":6827843.0,"Title":"How to download file from docker container?","Body":"<p>I'm using this Docker Container <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/udacity\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/examples\/udacity<\/a> .\nI've made few changes and would like to backup files from it to my hard drive. I can't find solution to do so.<\/p>\n","answers":[{"AnswerId":"44071697","CreationDate":"2017-05-19T13:44:18.260","ParentId":null,"OwnerUserId":"1664788","Title":null,"Body":"<p>Syntax to copy files\/folders from docker container to local filesystem is<\/p>\n\n<pre><code>docker cp [OPTIONS] CONTAINER:SRC_PATH DEST_PATH|-\n<\/code><\/pre>\n\n<p><strong>Ex :<\/strong> <code>docker cp  b8ddea5d8fad:\/mnt\/mydir\/ .\/localDir<\/code><\/p>\n\n<p>Syntax to copy files\/folders from local filesystem to docker container is<\/p>\n\n<pre><code>docker cp [OPTIONS] SRC_PATH|- CONTAINER:DEST_PATH\n<\/code><\/pre>\n\n<p><strong>Ex :<\/strong> <code>docker cp .\/localDir\/dirToCopy b8ddea5d8fad:\/mnt\/newDir\/<\/code><\/p>\n\n<p>Docker Documentation <a href=\"https:\/\/docs.docker.com\/engine\/reference\/commandline\/cp\/\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n"},{"AnswerId":"42588357","CreationDate":"2017-03-03T21:00:50.647","ParentId":null,"OwnerUserId":"4486909","Title":null,"Body":"<p>You could use 'docker cp' command to retrieve particular files.\n<a href=\"https:\/\/docs.docker.com\/engine\/reference\/commandline\/cp\/\" rel=\"nofollow noreferrer\">https:\/\/docs.docker.com\/engine\/reference\/commandline\/cp\/<\/a><\/p>\n"}]}
{"QuestionId":42589187,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-03T22:01:30.537","AcceptedAnswerId":48361009.0,"OwnerUserId":5436892.0,"Title":"do not save check point for a final step for Estimator","Body":"<p>I use Estimator and I train model in the loop to feed data. Every step is the final step. The checkpoints are saved for every final step too. I want to avoid saving checkpoint in every iteration to increase the performance (speed) of the training. \nI can not find any information how to do this. Do you have any ideas\/suggestions\/solutions?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"48361009","CreationDate":"2018-01-20T20:54:29.040","ParentId":null,"OwnerUserId":"80869","Title":null,"Body":"<p>Nowadays the api got changed a bit but from what I see you were using the fit  (currently train) method incorrectly, you should put steps=2000 and have your input function return an iterator over your dataset. Today you have tf.estimator.inputs.numpy_input_fn at your disposal that can help you when you have small data sets, otherwise you have to use tf.data.DataSet api.<\/p>\n\n<p>Something like this (it loads .wav files):<\/p>\n\n<pre><code>from tensorflow.contrib.framework.python.ops import audio_ops as contrib_audio\nfrom tensorflow.python.ops import io_ops\n# ...\ndef input_fn(num_epochs, batch_size, shuffle=False, mode='training')\n\n    def input_fn_bound():\n        def _read_file(fn, label):\n            return io_ops.read_file(fn), label\n\n        def _decode(data, label):\n            pcm = contrib_audio.decode_wav(data,\n                                           desired_channels=1,\n                                           desired_samples=desired_samples)\n            return pcm.audio, label\n\n        filenames = get_files(mode)\n        classes = get_classes(mode)\n        labels = {'class': np.array(classes)}\n        dataset = tf.data.Dataset.from_tensor_slices((filenames, labels))\n\n        if shuffle:\n            dataset = dataset.shuffle(buffer_size=len(labels))\n        dataset = dataset.map(_read_file, num_parallel_calls=num_map_threads)\n        dataset = dataset.map(_decode, num_parallel_calls=num_map_threads)\n        dataset = dataset.map(lambda wav, label: ({'wav': wav}, label))\n\n        dataset = dataset.repeat(num_epochs)\n        dataset = dataset.batch(batch_size)\n        dataset = dataset.prefetch(2)        # To load next batch while the first one is being processed on GPU\n        iter = dataset.make_one_shot_iterator()\n        features, labels = iter.get_next()\n        return features, labels\n\n    return input_fn_bound\n\n# ....\n\nestimator.train(input_fn=input_fn(\n        num_epoths=None, \n        batch_size=64,\n        shuffle=True,\n        mode='training'), \n    steps=10000)\n<\/code><\/pre>\n"}]}
{"QuestionId":42589303,"AnswerCount":1,"Tags":"<tensorflow><tokenize><word2vec>","CreationDate":"2017-03-03T22:10:12.603","AcceptedAnswerId":null,"OwnerUserId":2484124.0,"Title":"Word2Vec word containing numeric values","Body":"<p>When I am adding sentences to Word2Vec model it seems to remove the words which end or start with numeric values, for example \"ISO 9001\" is returned as \"ISO \", I've guessing it's something simple...<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42592980","CreationDate":"2017-03-04T06:38:18.630","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>I think you already answered your question in the tags you gave to this question. Most likely your tokenizer splits by blank spaces, and leaves out numbers. If you paste the tokenize code you use here we will be able to help you further. <\/p>\n\n<p>Good luck!<\/p>\n"}]}
{"QuestionId":42589956,"AnswerCount":1,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-03-03T23:09:01.683","AcceptedAnswerId":42591811.0,"OwnerUserId":7656080.0,"Title":"Tensorboard 1.0 Histogram tab","Body":"<p>What does the Histogram tab (not the Distributions tab) in Tensorboard (TensorFlow 1.0) represent? Can it be intrepreted as a true histogram?<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/AZv7q.png\" rel=\"nofollow noreferrer\">Histogram Screen shot<\/a><\/p>\n","answers":[{"AnswerId":"42591811","CreationDate":"2017-03-04T03:41:41.140","ParentId":null,"OwnerUserId":"721810","Title":null,"Body":"<p>According to a this <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/5381\" rel=\"nofollow noreferrer\">reply<\/a> by Dandelion Mane on a similar question: <\/p>\n\n<p>\"It's a normalized density. I wouldn't describe it as a probability density,\nalthough I think calling it one would be justifiable.\"<\/p>\n"}]}
{"QuestionId":42590431,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-04T00:02:10.357","AcceptedAnswerId":42590869.0,"OwnerUserId":744520.0,"Title":"Output from TensorFlow `py_func` has unknown rank\/shape","Body":"<p>I am trying to create a simple neural net in TensorFlow. The only tricky part is I have a custom operation that I have implemented with . When I pass the output from  to a  layer, TensorFlow complains that the rank should be known. The specific error is:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>I don't know how to preserve the shape of my data when I pass it through . My question is how do I get the correct shape? I have a simple example below to illustrate the problem.<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>The output from this snippet is:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><\/pre>\n\n<p>Why is  ? I want the shape to be  the same as . Thanks!<\/p>\n","answers":[{"AnswerId":"42590869","CreationDate":"2017-03-04T01:03:12.820","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Since <code>py_func<\/code> can execute arbitrary Python code and output anything, TensorFlow can't figure out the shape (it would require analyzing Python code of function body) You can instead give the shape manually<\/p>\n\n<pre><code>y.set_shape(inp.get_shape())\n<\/code><\/pre>\n"}]}
{"QuestionId":42590768,"AnswerCount":1,"Tags":"<tensorflow><autoencoder>","CreationDate":"2017-03-04T00:49:18.967","AcceptedAnswerId":43185868.0,"OwnerUserId":4925151.0,"Title":"Explaining Variational Autoencoder gaussian parameterization","Body":"<p>In the original Auto-Encoding Variational Bayes <a href=\"https:\/\/arxiv.org\/abs\/1312.6114\" rel=\"nofollow noreferrer\">paper<\/a>, the authors describes the \"reparameterization trick\" in section 2.4. The trick is to breakup your latent state z into learnable mean and sigma (learned by the encoder) and adding Gaussian noise. You then sample a datapoint from z (basically you generate an encoded image) and let the decoder map the encoded datapoint back to the original image.<\/p>\n\n<p>I have a hard getting over how strange this is. Could someone explain a bit more on the latent variable model, specifically:<\/p>\n\n<ol>\n<li>Why are we assuming the latent state is Gaussian?<\/li>\n<li>How is it possible that a Gaussian can generate an image? <\/li>\n<li>And how does backprop corrupt the encoder to learn a Gaussian function as opposed to an unknown non-linear function?<\/li>\n<\/ol>\n\n<p>Here is an example implementation of the latent model from <a href=\"https:\/\/jmetzen.github.io\/2015-11-27\/vae.html\" rel=\"nofollow noreferrer\">here<\/a> in TensorFlow.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43185868","CreationDate":"2017-04-03T13:25:10.920","ParentId":null,"OwnerUserId":"6470224","Title":null,"Body":"<ol>\n<li><p>They are not assuming that the activations of the encoder follow a gaussian distribution, they are enforcing that of the possible solutions choose a gaussian resembling one.<\/p><\/li>\n<li><p>The image is generated from decoding a activation\/feature, the activations are distributed resembling a gaussian.<\/p><\/li>\n<li><p>They minimize the KL divergence between the activations distribution and a gaussian one.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42590770,"AnswerCount":1,"Tags":"<python><keras><hdf5>","CreationDate":"2017-03-04T00:49:39.723","AcceptedAnswerId":42596682.0,"OwnerUserId":5896419.0,"Title":"Keras\/HDF5 TypeError: PointSelection __getitem__ only works with bool arrays","Body":"<p>When I try load training data from an HDF5 file using fit_generator and generator functions, I get a ValueError, which results from a PointSelectionError with HDF5:<\/p>\n\n<pre><\/pre>\n\n<p>I researched the error and it was mentioned that it could be due to duplicate indices, but that does not seem to be true in my case. Here are the row indices that were accessed:<\/p>\n\n<pre><\/pre>\n\n<p>Here are the training and generation functions:<\/p>\n\n<pre><\/pre>\n\n<p>Here are the functions that preprocesses and saves the training data into an HDF5 file:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42596682","CreationDate":"2017-03-04T13:39:41.987","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>The error comes from two things :<\/p>\n\n<ul>\n<li><p>The one you are reading comes from the fact that <code>batch_idxs<\/code> is an array, not a list. h5py objects accept indexing with lists. But even if you change<\/p>\n\n<pre><code>X = training_save_file[\"X\"][list(batch_idxs)]\n<\/code><\/pre>\n\n<p>you will still get an error. This comes from some restrictions about the list indexing. That brings us to the second point.<\/p><\/li>\n<li><p>If you read the <a href=\"http:\/\/docs.h5py.org\/en\/latest\/high\/dataset.html#fancy-indexing\" rel=\"noreferrer\">doc you sent me<\/a>, this is written :<\/p>\n\n<blockquote>\n  <p>The following restrictions exist:<\/p>\n  \n  <ul>\n  <li>List selections may not be empty <\/li>\n  <li>Selection coordinates must be given in increasing order<\/li>\n  <li>Duplicate selections are ignored<\/li>\n  <li>Very long lists (> 1000 elements) may produce poor performance<\/li>\n  <\/ul>\n<\/blockquote>\n\n<p>The second bullet is our issue : your random shuffling when creating <code>training_sample_idxs<\/code> makes the index order random and the dataset is expecting them in ecreasing order. This is a limitation you will have to deal with, but it is not too constraining as the order in a batch doesn't matter, the model will be optimized on the whole batch anyway. <\/p><\/li>\n<\/ul>\n\n<p>Does that help? <\/p>\n"}]}
{"QuestionId":42591068,"AnswerCount":1,"Tags":"<hidden><keras><recurrent-neural-network>","CreationDate":"2017-03-04T01:30:39.363","AcceptedAnswerId":null,"OwnerUserId":1058511.0,"Title":"can I reset the hidden state of an RNN between input data sets in Keras?","Body":"<p>I am training an RNN on a large data set which consists of disparate sources. I do not want the history of one set to spill over to the next. This means I want to reset the hidden state at the end of one set, before sending in the next. How can I do that with Keras? The doc claims you can get into the low level configurations. <\/p>\n\n<p>What I am trying to do is resetting the lstm hidden state every time a new data set is fed, so no influence from the prev dataset is carried forward. see line<\/p>\n\n<pre><\/pre>\n\n<p>from Karpathy's simple python implementation\n<a href=\"https:\/\/gist.github.com\/karpathy\/587454dc0146a6ae21fc\" rel=\"nofollow noreferrer\">https:\/\/gist.github.com\/karpathy\/587454dc0146a6ae21fc<\/a>\nline 45<\/p>\n\n<p>If I find the lstm layer and call reset on it, I am worried that would wipe out the entire training of the weights and biases, not just Hout<\/p>\n\n<p>Here is the training loop code<\/p>\n\n<pre><\/pre>\n\n<p>Each turn in the loop feeds in data from a single market. That's where I like to reset the hout in the lstm. <\/p>\n","answers":[{"AnswerId":"42593835","CreationDate":"2017-03-04T08:23:51.407","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<blockquote>\n  <p>To reset the states of your model, call <code>.reset_states()<\/code> on either a specific layer, or on your entire model. <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">source<\/a><\/p>\n<\/blockquote>\n\n<p>So if you have a list of datasets :<\/p>\n\n<pre><code>for ds in datasets :\n    model.reset_states()\n    model.fit(ds['inputs'],ds['targets'],...)\n<\/code><\/pre>\n\n<p>Is that what you are looking for? <\/p>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<pre><code>for iteration in range(1, 10):\n    for key in X_dict:\n        model.reset_states() # reset the states of all the LSTM's of your network\n        #model.layers[lstm_layer_index].reset_states() # reset the states of this specific LSTM layer\n        X = X_dict[key]\n        y = y_dict[key]\n        history=model.fit(X, y, batch_size=batch_size, callbacks=cbks, nb_epoch=1,verbose=0)\n<\/code><\/pre>\n\n<p>This is how you apply it.<\/p>\n\n<p>By default, the LSTM's are not stateful. Which means that they won't keep a hidden state after going over a sequence. The initial state when starting a new sequence will be set to 0. If you selected <code>stateful=True<\/code>, then it will keep the last hidden state (the output) of the previous sequence to initialize itself for the next sequence in the batch. It's like the sequence was continuing.<\/p>\n\n<p>Doing <code>model.reset_states()<\/code> will just reset those last hidden states that were kept in memory to 0, just like if the sequence was starting from scratch.<\/p>\n\n<p>If you don't trust that <code>.reset_states()<\/code> to do what you expect, feel free to go to the source code.<\/p>\n"}]}
{"QuestionId":42591191,"AnswerCount":3,"Tags":"<deep-learning><keras>","CreationDate":"2017-03-04T01:53:50.090","AcceptedAnswerId":null,"OwnerUserId":4631796.0,"Title":"Keras Semantic Segmentation Weighted Loss Pixel Map","Body":"<p>I'm trying to do multi-class semantic segmentation with a <a href=\"https:\/\/arxiv.org\/pdf\/1505.04597.pdf\" rel=\"nofollow noreferrer\">unet<\/a> design. Similar to the unet paper, I'd like to do make a loss function to overweight borders (page 5). <\/p>\n\n<p>As such I'd like to make a custom loss map for each image where the borders between objects are overweighted. I am using categorical cross-entropy where I flatten the image before the loss function as <a href=\"https:\/\/stackoverflow.com\/questions\/42118821\/cross-entropy-loss-for-semantic-segmentation-keras\">here<\/a>. I would be fine making the pixel loss mask but I am wondering how, if possible, to multiple the loss by the pixel mask.<\/p>\n","answers":[{"AnswerId":"42596316","CreationDate":"2017-03-04T13:00:42.370","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>If you know how to do this on a 2d map you could always use multiple outputs and use the custom pixel mask additionally to the cross entropy loss. An example implementation of multiple losses for a u-shaped network can be found here: <a href=\"https:\/\/github.com\/EdwardTyantov\/ultrasound-nerve-segmentation\" rel=\"nofollow noreferrer\">https:\/\/github.com\/EdwardTyantov\/ultrasound-nerve-segmentation<\/a><\/p>\n"},{"AnswerId":"52143493","CreationDate":"2018-09-03T05:15:22.450","ParentId":null,"OwnerUserId":"2608194","Title":null,"Body":"<p>Assuming multi-class prediction, in your data generator, concatenate your weight maps as extra channels in third dimension for each element in a batch. Next, in your loss function, extract the weight maps for each batch element, multiply them (just use *) with the prediction, and compute log loss. <\/p>\n"},{"AnswerId":"43679674","CreationDate":"2017-04-28T11:57:19.317","ParentId":null,"OwnerUserId":"7796100","Title":null,"Body":"<p>Here is an implementation of weight maps in tensorflow <a href=\"http:\/\/tf-unet.readthedocs.io\/en\/latest\/_modules\/tf_unet\/unet.html\" rel=\"nofollow noreferrer\">http:\/\/tf-unet.readthedocs.io\/en\/latest\/_modules\/tf_unet\/unet.html<\/a> you should be able to adapt it for keras in a custom loss function. I report the relevant code: <\/p>\n\n<pre><code>    def _get_cost(self, logits, cost_name, cost_kwargs):\n\n    Optional arguments are: \n    class_weights: weights for the different classes in case of multi-class imbalance\n    regularizer: power of the L2 regularizers added to the loss function\n\n    flat_logits = tf.reshape(logits, [-1, self.n_class])\n    flat_labels = tf.reshape(self.y, [-1, self.n_class])\n    if cost_name == \"cross_entropy\":\n        class_weights = cost_kwargs.pop(\"class_weights\", None)\n\n        if class_weights is not None:\n            class_weights = tf.constant(np.array(class_weights, dtype=np.float32))\n\n            weight_map = tf.multiply(flat_labels, class_weights)\n            weight_map = tf.reduce_sum(weight_map, axis=1)\n\n            loss_map = tf.nn.softmax_cross_entropy_with_logits(flat_logits, flat_labels)\n            weighted_loss = tf.multiply(loss_map, weight_map)\n\n            loss = tf.reduce_mean(weighted_loss)\n\n        else:\n            loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=flat_logits, \n                                                                          labels=flat_labels))}\n<\/code><\/pre>\n"}]}
{"QuestionId":42592845,"AnswerCount":1,"Tags":"<python><theano><keras>","CreationDate":"2017-03-04T06:18:32.663","AcceptedAnswerId":42603995.0,"OwnerUserId":5896419.0,"Title":"Keras\/Theano: Node compilation failed during training","Body":"<p>I am trying to train a Keras model on Mac OS X that has already been compiled, and I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I have updated Keras and Theano, but the issue still exists. I am confused since just a few days ago training the exact same model did not have this issue. Here are the functions used during training:<\/p>\n\n<pre><\/pre>\n\n<p>For reference, here is the model being trained:<\/p>\n\n<pre><\/pre>\n\n<p>Model summary:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42603995","CreationDate":"2017-03-05T03:01:34.877","ParentId":null,"OwnerUserId":"5896419","Title":null,"Body":"<p>The problem seems to stem from installing Theano and Keras from their github repositories, like so:<\/p>\n\n<pre><code>pip install git+git:\/\/github.com\/Theano\/Theano.git\npip install git+git:\/\/github.com\/fchollet\/keras.git\n<\/code><\/pre>\n\n<p>I fixed it by uninstalling Theano and Keras, followed by installing them directly with pip:<\/p>\n\n<pre><code>pip uninstall Theano\npip uninstall keras\npip install Theano\npip install keras\n<\/code><\/pre>\n\n<p>There might have been issues with the bleeding edge versions of either Theano or Keras. Hope this helps others as well.<\/p>\n\n<p><strong>EDIT<\/strong>: It seems like the issue really comes from the master branch of Theano. Follow the issue I have posted on Theano's repository for a potential permanent fix <a href=\"https:\/\/github.com\/Theano\/Theano\/issues\/5655\" rel=\"nofollow noreferrer\">https:\/\/github.com\/Theano\/Theano\/issues\/5655<\/a><\/p>\n"}]}
{"QuestionId":42593771,"AnswerCount":2,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-03-04T08:15:36.037","AcceptedAnswerId":null,"OwnerUserId":7337638.0,"Title":"Session graph is empty","Body":"<p>I have been trying to iterate the tensorflow model using the answer provided on \n<a href=\"https:\/\/stackoverflow.com\/questions\/35695183\/tensorflow-memory-leak-even-while-closing-session\">Tensorflow : Memory leak even while closing Session?<\/a>\nI have the same error as mentioned in one of the comments i.e. \"The Session graph is empty.  Add operations to the graph before calling run().\" I cannot figure out how to solve it.<\/p>\n","answers":[{"AnswerId":"54295992","CreationDate":"2019-01-21T18:37:25.450","ParentId":null,"OwnerUserId":"5884955","Title":null,"Body":"<p>If you search the TensorFlow source code you will find this exclusive message is telling you <em>the graph version<\/em> is 0, which means it is empty.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/l4iWR.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/l4iWR.jpg\" alt=\"enter image description here\"><\/a>\nYou need to <a href=\"https:\/\/www.tensorflow.org\/guide\/saved_model\" rel=\"nofollow noreferrer\">load<\/a> the graph from file or to create a graph like this (just an example):<\/p>\n\n<pre><code>import tensorflow as tf\n\npi = tf.constant(3.14, name=\"pi\")\nr = tf.placeholder(tf.float32, name=\"r\")\na = pi * r * r\n\ngraph = tf.get_default_graph()\nprint(graph.get_operations())\n<\/code><\/pre>\n\n<p>The line <code>print(graph.get_operations())<\/code> will confirm the graph is not empty:<\/p>\n\n<pre><code>[&lt;tf.Operation 'pi' type=Const&gt;, &lt;tf.Operation 'r' type=Placeholder&gt;, &lt;tf.Operation 'mul' type=Mul&gt;, &lt;tf.Operation 'mul_1' type=Mul&gt;]\n<\/code><\/pre>\n"},{"AnswerId":"44184861","CreationDate":"2017-05-25T15:58:44.073","ParentId":null,"OwnerUserId":"4332894","Title":null,"Body":"<p>My suggestion is to make sure your session knows which graph it is running on.\nSomethings you can try are:<\/p>\n\n<ol>\n<li><p>build the graph first and pass in the graph to a session.<\/p>\n\n<pre><code>myGraph = tf.Graph()\nwith myGraph.as_default():\n    # build your graph here\n\nmySession = tf.Session(graph=myGraph)\nmySession.run(...)\n\n# Or\n\nwith tf.Session(graph=myGraph) as mySession:\n    mySession.run(...)\n<\/code><\/pre><\/li>\n<li><p>If you want to use multiple <code>with<\/code> statement, use it in a nested way.<\/p>\n\n<pre><code>with tf.Graph().as_default():\n    # build your graph here\n    with tf.Session() as mySession:\n        mySession.run(...)\n<\/code><\/pre><\/li>\n<\/ol>\n"}]}
{"QuestionId":42593865,"AnswerCount":1,"Tags":"<python><tensorflow><python-3.5><tensorboard>","CreationDate":"2017-03-04T08:29:00.153","AcceptedAnswerId":null,"OwnerUserId":7657215.0,"Title":"windows tensorboard not work(no graph)","Body":"<p>I downloaded tensorflow via pip.\nhere is my code:<\/p>\n\n<pre><\/pre>\n\n<p>actually, I want to test tensorboard, and source of this code is <a href=\"http:\/\/pythonkim.tistory.com\/39\" rel=\"nofollow noreferrer\">http:\/\/pythonkim.tistory.com\/39<\/a>.<\/p>\n\n<p>this file is in D:\\myname\\python study\\TensorFlow\\, and I started tensorboard like this<\/p>\n\n<pre><\/pre>\n\n<p>and also like this<\/p>\n\n<pre><\/pre>\n\n<p>at D:\\myname\\python study\\TensorFlow\nand I saw this page: <a href=\"https:\/\/i.stack.imgur.com\/R5hcX.png\" rel=\"nofollow noreferrer\">watch image here<\/a><\/p>\n\n<p>windows 10, tensorflow 1.0, chrome browser<\/p>\n\n<p>note: there is event file in .\/board\/sample_1.<\/p>\n","answers":[{"AnswerId":"45935029","CreationDate":"2017-08-29T09:11:07.970","ParentId":null,"OwnerUserId":"6786718","Title":null,"Body":"<p>You need to put in the <code>\" \"<\/code> when starting the Tensorboard so change it to:<\/p>\n\n<pre><code> tensorboard --logdir=\".\/board\/sample_1\"\n<\/code><\/pre>\n"}]}
{"QuestionId":42594207,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-03-04T09:10:32.727","AcceptedAnswerId":42599330.0,"OwnerUserId":4278243.0,"Title":"Caffe, joining outputs from 2 models","Body":"<p>I have 2 different models, let's say NM1 and NM2. <\/p>\n\n<p>So, what I'm looking is something that works like in the example below.<\/p>\n\n<p>Let's say that we have a picture of a dog.<\/p>\n\n<p>NM1 predicts that it's a cat on the picture with a probability 0.52 and that it's a dog with a probability 0.48. \nNM2 predicts that it's a dog with a probability 0.6 and that it's a cat with a probability 0.4.   <\/p>\n\n<p>NM1 - will predict wrong \nNM2 - will predict correctly<\/p>\n\n<p>NM1 + NM2 - connection will predict correctly (because 0.48 + 0.6 > 0.52 + 0.4)<\/p>\n\n<p>So, each model ends with InnerProducts (after Softmax) which give me 2 vectors of probabilities.<\/p>\n\n<p>Next step, I have those 2 vectors and I want to add them. Here I use Eltwise layer. <\/p>\n\n<pre><\/pre>\n\n<p>Before joining NM1 had accuracy ~70% and NM2 ~10%. <\/p>\n\n<p>After joining accuracy can't reach even 1%. <\/p>\n\n<p>Thus, my conclusion is that I understand something wrong and I'd be grateful if someone could explain to me where I'm wrong. <\/p>\n\n<p>PS. I did turn off shuffle when creating lmdb. <\/p>\n\n<p>UPDATE<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/FGRvp.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/FGRvp.png\" alt=\"Model image\"><\/a><\/p>\n","answers":[{"AnswerId":"42599330","CreationDate":"2017-03-04T17:56:55.653","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>If you want to add (element-wise) the probabilities, you need to add <strong>after<\/strong> the <code>\"Softmax\"<\/code> layer, and not after the <code>\"InnerProduct\"<\/code> layer. You should have something like<\/p>\n\n<pre><code>layer {\n  type: \"InnerProduct\"\n  name: \"fc8\"\n  top: \"fc8\"\n  # ... \n}\nlayer {\n  type: \"Softmax\"\n  name: \"prob_nm1\"\n  top: \"prob_nm1\"\n  bottom: \"fc8\"\n}\nlayer {\n  type: \"InnerProduct\"\n  name: \"fc8N\"\n  top: \"fc8N\"\n  # ... \n}\nlayer {\n  type: \"Softmax\"\n  name: \"prob_nm2\"\n  top: \"prob_nm2\"\n  bottom: \"fc8N\"\n}\n# Joining the probabilites\nlayer {\n  type: \"Eltwise\"\n  name: \"prob_sum\"\n  bottom: \"prob_nm1\"\n  bottom: \"prob_nm2\"\n  top: \"prob_sum\"\n  eltwise_param {\n    operation: SUM\n    coeff: 0.5\n    coeff: 0.5\n  }\n}\n<\/code><\/pre>\n"}]}
{"QuestionId":42594881,"AnswerCount":1,"Tags":"<python-2.7><tensorflow><anaconda>","CreationDate":"2017-03-04T10:26:14.013","AcceptedAnswerId":42604976.0,"OwnerUserId":4273813.0,"Title":"Installing tensorflow CPU with windows 10 anaconda python 2.7","Body":"<p>I am using anaconda, python 2.7, windows 10. I would like to install tensorflow with conda. However, I am having this error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42604976","CreationDate":"2017-03-05T05:45:25.077","ParentId":null,"OwnerUserId":"4847576","Title":null,"Body":"<p>TensorFlow only supports Python 3.5 64-bit on Windows. \nYou need the corresponding Anaconda distribution. <\/p>\n\n<p>Make sure to create an environment for <code>Python=3.5<\/code> as Anaconda recently upgraded their distribution from <code>3.5<\/code> to <code>3.6<\/code> which is also not supported with a pre-build binary yet on Windows (i.e. you need to build TensorFlow from source to use Python 3.6 on Windows).<\/p>\n\n<p>As a side note the conda package is not official so I'd recommend using the <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows#installing_with_anaconda\" rel=\"nofollow noreferrer\">wheel provided on TensorFlow website<\/a>. As an alternative you can install with the PYPI packages through <code>pip install tensorflow<\/code>.<\/p>\n\n<p>For complete information on TensorFlow installation on Windows please <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">see here<\/a>.<\/p>\n"}]}
{"QuestionId":42595543,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-04T11:38:24.957","AcceptedAnswerId":42600405.0,"OwnerUserId":1879547.0,"Title":"TensorFlow: eval restored graph","Body":"<p>I am trying to restore graph from a checkpoint. The checkpoint is created by . There are both  file and checkpoint. <\/p>\n\n<p>What I try to achive is to load this graph from separate application to run some operation (i.e. resue existing model).<\/p>\n\n<p>I do this as the following (as explained here: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/import_meta_graph\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/import_meta_graph<\/a>):<\/p>\n\n<pre><\/pre>\n\n<p>What I see is . What I expect to see is 1-D Tensor.\nI inspected  object using get_collection and find that there are all my variables required for  to run correctly initialized with values restored from the checkpoint.\nHow can I figure out why the operation is not evaluated correctly? I am really stuck here.<\/p>\n\n<p>The following code:<\/p>\n\n<pre><\/pre>\n\n<p>prints<\/p>\n\n<pre><\/pre>\n\n<p>as if there is no connection between an operation and corresponding variable.<\/p>\n","answers":[{"AnswerId":"42600405","CreationDate":"2017-03-04T19:31:07.103","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph#get_operation_by_name\" rel=\"nofollow noreferrer\"><code>tf.Graph.get_operation_by_name()<\/code><\/a> method always returns a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation\" rel=\"nofollow noreferrer\"><code>tf.Operation<\/code><\/a> object. When you pass a <code>tf.Operation<\/code> object to <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run\" rel=\"nofollow noreferrer\"><code>tf.Session.run()<\/code><\/a>, TensorFlow will execute that operation (and everything on which it depends) and discard its outputs (if any).<\/p>\n\n<p>If you are interested in the value of a particular output, you have to tell TensorFlow which output (a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor\" rel=\"nofollow noreferrer\"><code>tf.Tensor<\/code><\/a>) you are interested in. There are two main options:<\/p>\n\n<ul>\n<li><p>Get a <code>tf.Operation<\/code> from the graph and then select one of its <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Operation#outputs\" rel=\"nofollow noreferrer\"><code>outputs<\/code><\/a>:<\/p>\n\n<pre><code>op = my_graph.get_operation_by_name(\"op\")\noutput = op.outputs[0]\nprint(sess.run(output))\n<\/code><\/pre><\/li>\n<li><p>Get a <code>tf.Tensor<\/code> from the graph by calling <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph#get_tensor_by_name\" rel=\"nofollow noreferrer\"><code>tf.Graph.get_tensor_by_name()<\/code><\/a>, and appending <code>\":&lt;output index&gt;\"<\/code> to the operation's name:<\/p>\n\n<pre><code>output = my_graph.get_tensor_by_name(\"op:0\")\nprint(sess.run(output))\n<\/code><\/pre><\/li>\n<\/ul>\n\n<p>Why does TensorFlow draw this distinction? For one thing, a operation can have multiple outputs, so it is sometimes necessary to be specific about which output you want to fetch. For another, an operation may have a side effect and produce a large output\u2014see <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/assign\" rel=\"nofollow noreferrer\"><code>tf.assign()<\/code><\/a> for an example\u2014and it is often more efficient to pass the <code>tf.Operation<\/code> to <code>sess.run()<\/code> so that the value is not copied back into the Python program.<\/p>\n"}]}
{"QuestionId":42595921,"AnswerCount":0,"Tags":"<performance><tensorflow>","CreationDate":"2017-03-04T12:17:43.877","AcceptedAnswerId":null,"OwnerUserId":6306884.0,"Title":"Execution time slow down at beginning and raising later","Body":"<p>This is execution time log<br>\nAs you can see, it faster and faster until it used 1.5 sec for one iteraction<br>\nand then slower and slower   <\/p>\n\n<pre><\/pre>\n\n<p>Using  to freeze graph<br>\ninstall tensorflow 1.0 from source, using , build with , ...etc     <\/p>\n\n<pre><\/pre>\n\n<p>and follow this <a href=\"https:\/\/github.com\/DrSleep\/tensorflow-deeplab-resnet\/blob\/master\/train_msc.py#L233\" rel=\"nofollow noreferrer\">github<\/a> to implement image_reader and accumulate gradients (like  in caffe), all OP are outside training loop<\/p>\n\n<p>not sure if relevant<br>\nGPU memory grow slightly, from 5707 MiB to 5717MiB<br>\nGPU-util become low and strange<br>\n1% -> 59% -> 1% -> 99% -> 0% -> 54% -> 1% -> 48%    <\/p>\n\n<p>anyone met this situation before?\nor have any suggestion to debug?<\/p>\n","answers":[]}
{"QuestionId":42596057,"AnswerCount":2,"Tags":"<python><machine-learning><neural-network><deep-learning><keras>","CreationDate":"2017-03-04T12:31:47.687","AcceptedAnswerId":42596310.0,"OwnerUserId":6354442.0,"Title":"Keras error : Expected to see 1 array","Body":"<p>I got the following error when I tried to train an MLP model in keras(I am using keras version )<\/p>\n\n<blockquote>\n  <p>Error when checking model input: the list of Numpy arrays that you\n  are passing to your model is not the size the model expected. Expected\n  to see 1 arrays but instead got the following list of 12859 arrays:<\/p>\n<\/blockquote>\n\n<p>This is the summary of the model<\/p>\n\n<pre><\/pre>\n\n<p>This is the first line of model<\/p>\n\n<pre><\/pre>\n\n<p>For training:<\/p>\n\n<pre><\/pre>\n\n<p>where X is a list of elements and each element in turn is a list of 200 values.<\/p>\n\n<p>Edit :<\/p>\n\n<p>I also tried<\/p>\n\n<pre><\/pre>\n\n<p>but I am getting the same error<\/p>\n","answers":[{"AnswerId":"50961591","CreationDate":"2018-06-21T06:17:01.117","ParentId":null,"OwnerUserId":"3901871","Title":null,"Body":"<p>I fixed mine by adding <\/p>\n\n<blockquote>\n  <p>np.array<\/p>\n<\/blockquote>\n\n<p>to train_X , train_Y , valid_X and valid_Y. For example, <\/p>\n\n<pre><code>model.fit(np.array(train_X),np.array(train_Y),\n          batch_size=32,nb_epoch=20,\n          validation_data=(np.array(valid_X),np.array(valid_Y)),\n          callbacks=[early_stop])\n<\/code><\/pre>\n\n<p>I got the help from <a href=\"https:\/\/github.com\/keras-team\/keras\/issues\/4002\" rel=\"nofollow noreferrer\">here<\/a>. This approach is likely to have a slow run because all data features will have to be converted to numpy arrays and it could be a lot of work for your system. <\/p>\n"},{"AnswerId":"42596310","CreationDate":"2017-03-04T12:59:59.930","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Your error comes from the fact that your <code>X<\/code> for some reason wasn't transformed to a <code>numpy.array<\/code>. In this your <code>X<\/code> is treated as a list of rows and this is a reason behind your error message (that it expected one input instead of list which has a number of rows elements). Transformation:<\/p>\n\n<pre><code>X = numpy.array(X)\nY = numpy.array(Y)\n<\/code><\/pre>\n\n<p>I would check a data loading process because something might go wrong there.<\/p>\n\n<p><strong>UPDATE:<\/strong><\/p>\n\n<p>As it was mentioned in a comment - <code>input_shape<\/code> need to be changed to <code>input_dim<\/code>.<\/p>\n\n<p><strong>UPDATE 2:<\/strong><\/p>\n\n<p>In order to keep <code>input_shape<\/code> one should change to it to <code>input_shape=(200,)<\/code>.<\/p>\n"}]}
{"QuestionId":42596703,"AnswerCount":1,"Tags":"<neural-network><torch><nvidia-digits>","CreationDate":"2017-03-04T13:41:35.177","AcceptedAnswerId":42606810.0,"OwnerUserId":5285417.0,"Title":"Where can I find existing pre-trained torch models?","Body":"<p>I'm using Torch 7 and lua 5.1. I will need to do object recognition from an input video stream. I've installed DIGITS from NVidia.\nI've heard there are existing models that are already pre-trained provided by Google (or someone other source).  Where I can find them ? <\/p>\n","answers":[{"AnswerId":"42606810","CreationDate":"2017-03-05T09:48:33.353","ParentId":null,"OwnerUserId":"6076729","Title":null,"Body":"<p>I would recommend you to check <a href=\"https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo\" rel=\"nofollow noreferrer\">https:\/\/github.com\/BVLC\/caffe\/wiki\/Model-Zoo<\/a> it is a common GitHub repo with a bunch of trained models. Even though it is made for caffe framework there is a library (really easy to use) in torch that lets you use them without problems: <a href=\"https:\/\/github.com\/szagoruyko\/loadcaffe\" rel=\"nofollow noreferrer\">https:\/\/github.com\/szagoruyko\/loadcaffe<\/a><\/p>\n"}]}
{"QuestionId":42596719,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-03-04T13:43:34.180","AcceptedAnswerId":null,"OwnerUserId":1868775.0,"Title":"Early stopping with LSTM on Tensorflow","Body":"<p>I have implemented an LSTM on tensorflow in order to give a label (float from -1 to 1) to each timestep of a feature vector, as is in this example code:<\/p>\n\n<pre><\/pre>\n\n<p>Now I want to implement early stopping on my code in order to avoid over-fitting. What is the most straightforward way to accomplish this? Is there anything already implemented on tensorflow? I think the  API might allow me to do this but I am not sure about how I can apply it to my case in particular.<\/p>\n","answers":[]}
{"QuestionId":42596805,"AnswerCount":1,"Tags":"<tensorflow><bots><lstm>","CreationDate":"2017-03-04T13:52:16.643","AcceptedAnswerId":42599020.0,"OwnerUserId":5915270.0,"Title":"I can't find seq2seq module in tensoflow repository","Body":"<p>I am reading tensorflow tutorial on seq2seq models. It has mentioned the location of the source code in t***ensorflow\/tensorflow\/python\/ops\/seq2seq.py***. But after I go there (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/python\/ops\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/python\/ops<\/a>) I couldn't find the source code.<\/p>\n","answers":[{"AnswerId":"42599020","CreationDate":"2017-03-04T17:29:24.620","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>The code <code>seq2seq.py<\/code> was moved to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/legacy_seq2seq\/python\/ops\/seq2seq.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/legacy_seq2seq\/python\/ops\/seq2seq.py<\/a> on commit  <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/commit\/d438a07a2f50401c6a1d5a644512cd46fba4da4f\" rel=\"nofollow noreferrer\">d438a07a2f50401c6a1d5a644512cd46fba4da4f<\/a> on Dec 14, 2016.<\/p>\n"}]}
{"QuestionId":42597010,"AnswerCount":3,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-03-04T14:16:08.730","AcceptedAnswerId":null,"OwnerUserId":7598713.0,"Title":"Tensorflow Value Error : Cannot evaluate tensor using eval","Body":"<p>I am new to tensorflow and getting Tensorflow value error with following script : <\/p>\n\n<pre><\/pre>\n\n<p>Also I tried thisway : <\/p>\n\n<p><\/p>\n\n<p>It throws error of unitialized value Variable. <\/p>\n\n<p>Now when I am declaring  isn't it going to initialize it with 10 ? <\/p>\n","answers":[{"AnswerId":"42597473","CreationDate":"2017-03-04T15:02:44.190","ParentId":null,"OwnerUserId":"6791223","Title":null,"Body":"<p>You need to explicitly run an initializer operation<\/p>\n\n<pre><code>sess.run(tf.variables_initializer(W))\n<\/code><\/pre>\n\n<p>before evaluating any nodes which are dependent on W.<\/p>\n"},{"AnswerId":"48838439","CreationDate":"2018-02-17T06:04:37.783","ParentId":null,"OwnerUserId":"9372575","Title":null,"Body":"<p>Another example,<\/p>\n\n<pre><code>    import tensorflow as tf  \n    W = tf.Variable(tf.truncated_normal([700,10]))\n   sess = tf.Session()\n   with sess.as_default():\n        sess.run(W.initializer)\n        print(W.eval())\n<\/code><\/pre>\n\n<p>result: <\/p>\n\n<pre><code>[[-0.3294761   0.6800459   1.33331    ...  1.42762    -1.3164878\n   1.4831722 ]\n [-1.0402402   0.52254885 -1.344712   ... -0.30849338  0.15020785\n   1.6682776 ]\n [-1.1791034   1.4859517  -1.7137778  ...  0.844212    1.5928217\n  -0.21043983]\n ...\n [ 0.01982834 -1.1290654   0.33557415 ...  0.0510614  -0.6524679\n   0.16643837]\n [-0.09969945 -0.10285325 -1.1134144  ...  1.2253191   0.13343143\n  -1.7491579 ]\n [-1.9345136   0.63447094  1.1200713  ...  0.5357313   1.8579113\n   0.8549472 ]]\n<\/code><\/pre>\n"},{"AnswerId":"42597479","CreationDate":"2017-03-04T15:03:19.543","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>From the documentation:<\/p>\n\n<blockquote>\n  <p>When you launch the graph, variables have to be explicitly\n  initialized before   you can run Ops that use their value. You can\n  initialize a variable by   running its <em>initializer op<\/em>, restoring the\n  variable from a save file, or   simply running an <code>assign<\/code> Op that\n  assigns a value to the variable. In fact,   the variable <em>initializer\n  op<\/em> is just an <code>assign<\/code> Op that assigns the   variable's initial value\n  to the variable itself.<\/p>\n\n<pre><code> # Launch the graph in a session.   \n with tf.Session() as sess:\n    # Run the variable initializer.\n    sess.run(w.initializer)\n    # ...you now can run ops that use the value of 'w'...\n<\/code><\/pre>\n  \n  <p>The most common initialization pattern is to use the convenience function\n    <code>global_variables_initializer()<\/code> to add an Op to the graph that initializes\n    all the variables. You then run that Op after launching the graph.<\/p>\n\n<pre><code> # Add an Op to initialize global variables.\n  init_op = tf.global_variables_initializer()   \n  # Launch the graph in a session.\n  with tf.Session() as sess:\n      # Run the Op that initializes global variables.\n      sess.run(init_op)\n      # ...you can now run any Op that uses variable values...\n<\/code><\/pre>\n<\/blockquote>\n\n<p>as a result you need to use something like:<\/p>\n\n<pre><code>import tensorflow as tf  \nW = tf.Variable(10)\nprint('W: {0}'.format(W))\nsess = tf.Session()\nwith sess.as_default():\n    sess.run(W.initializer)\n    print(W.eval())\n<\/code><\/pre>\n\n<p>FYI <a href=\"https:\/\/stackoverflow.com\/a\/33610914\/395857\">In TensorFlow, what is the difference between Session.run() and Tensor.eval()?<\/a><\/p>\n"}]}
{"QuestionId":42597117,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-04T14:26:35.190","AcceptedAnswerId":42600232.0,"OwnerUserId":5440222.0,"Title":"what's the difference between a ref edge and non-ref edge in tensorflow?","Body":"<p>As the title says, I'm wondering about the conceptual difference between  a \"ref edge\" and \"non-ref edge\" in TensorFlow.<\/p>\n\n<p>I'm reading the graph partitioning algorithm in TensorFlow.\nHere (<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1708b92bb923e420d746a56baafc7d4ddcd5e05e\/tensorflow\/core\/graph\/graph_partition.cc#L825\" rel=\"nofollow noreferrer\">line 826 of <\/a>) is the comment which mentions \nthe \"non-ref edge\":<\/p>\n\n<pre><\/pre>\n\n<p>Can someone explain the difference more clearly? Thanks very much.<\/p>\n","answers":[{"AnswerId":"42600232","CreationDate":"2017-03-04T19:16:58.073","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In TensorFlow, most edges are \"non-ref\" edges, which means that the value flowing along that edge is a constant. If you think of a vertex (operation) in a TensorFlow graph as a function, you can think of a non-ref edge as representing a function argument that is <a href=\"https:\/\/en.wikipedia.org\/wiki\/Evaluation_strategy#Call_by_value\" rel=\"nofollow noreferrer\">passed by value<\/a> in a conventional programming language like C or C++. For example, the inputs to and outputs from the operation <code>z = tf.matmul(x, y)<\/code> are all non-ref edges.<\/p>\n\n<p>A \"ref edge\" in TensorFlow allows the value flowing along that edge to be mutated. Continuing the function analogy, a ref edge represents a function argument that is <a href=\"https:\/\/en.wikipedia.org\/wiki\/Evaluation_strategy#Call_by_reference\" rel=\"nofollow noreferrer\">passed by reference<\/a> (from which we take the name \"ref\" edge). The most common use of ref edges is in the current internal implementation of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Variable\" rel=\"nofollow noreferrer\"><code>tf.Variable<\/code><\/a>: the internal <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1708b92bb923e420d746a56baafc7d4ddcd5e05e\/tensorflow\/core\/kernels\/variable_ops.h\" rel=\"nofollow noreferrer\">Variable kernel<\/a> owns a mutable buffer, and outputs a reference to that buffer on a ref edge. Operations such as <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/assign\" rel=\"nofollow noreferrer\"><code>tf.assign(var, val)<\/code><\/a> expect their <code>var<\/code> argument to be passed along ref edge, because they need to mutate the value in <code>var<\/code>.<\/p>\n\n<p>The graph partitioning algorithm treats ref edges specially because they correspond to values that could change as the graph executes. Since a non-ref edge is a constant value, TensorFlow can assume that all non-ref edges out of the same operation that cross between two devices can be combined into a single edge, which saves on network\/memory bandwidth. Since the value on a ref edge can change (e.g. if a variable is updated in the middle of a step), TensorFlow must be careful not to combine the edges, so that the remote device can see the new value. By analogy with C\/C++, the TensorFlow graph partitioner treats a ref-edge as representing a <a href=\"https:\/\/en.wikipedia.org\/wiki\/Volatile_(computer_programming)\" rel=\"nofollow noreferrer\"><code>volatile<\/code><\/a> variable, for the purposes of optimization.<\/p>\n\n<p>Finally, as you can tell from the amount of explanation above, ref edges are quite complicated, and there is an ongoing effort to remove them from the TensorFlow execution model. The replacement is \"resource-typed edges\", which allow non-tensor values to flow along an edge (unifying variables, queues, readers, and other complex objects in TensorFlow), and explicit operations that take a variable resource as input and read its value (as a non-ref output edge). The implementation of the new \"resource variables\" can be seen <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1708b92bb923e420d746a56baafc7d4ddcd5e05e\/tensorflow\/python\/ops\/resource_variable_ops.py\" rel=\"nofollow noreferrer\">here in Python<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/1708b92bb923e420d746a56baafc7d4ddcd5e05e\/tensorflow\/core\/kernels\/resource_variable_ops.cc\" rel=\"nofollow noreferrer\">here in C++<\/a>.<\/p>\n"}]}
{"QuestionId":42597520,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-04T15:07:02.170","AcceptedAnswerId":42606537.0,"OwnerUserId":5098762.0,"Title":"Slicing a tensor by an index tensor in Tensorflow","Body":"<p>I have two following tensors (note that they are <strong>both<\/strong> <strong>Tensorflow tensors<\/strong> which means they are still virtually symbolic at the time I construct the following slicing op before I launch a ):  <\/p>\n\n<ul>\n<li>: has shape (64,784, 256) <\/li>\n<li>: has shape (64, 784)   <\/li>\n<\/ul>\n\n<p>and I want to construct an op that returns the following tensor:<\/p>\n\n<ul>\n<li>: has shape (64,784) where<\/li>\n<\/ul>\n\n<p>  <\/p>\n\n<p>What is the most efficient way in Tensorflow to do so?<\/p>\n\n<p>ps: I tried with  but couldn't make use of it to perform the operation I described above. <\/p>\n\n<p>Many thanks.<br>\n-Bests <\/p>\n","answers":[{"AnswerId":"42606537","CreationDate":"2017-03-05T09:16:36.400","ParentId":null,"OwnerUserId":"6338809","Title":null,"Body":"<p>You can get exactly what you want using <code>tf.gather_nd<\/code>.  The final expression is:<\/p>\n\n<pre><code>tf.gather_nd(params, tf.stack([tf.tile(tf.expand_dims(tf.range(tf.shape(indices)[0]), 1), [1, tf.shape(indices)[1]]), tf.transpose(tf.tile(tf.expand_dims(tf.range(tf.shape(indices)[1]), 1), [1, tf.shape(indices)[0]])), indices], 2))\n<\/code><\/pre>\n\n<hr>\n\n<p>This expression has the following explanation:<\/p>\n\n<ul>\n<li><code>tf.gather_nd<\/code> does what you expected and uses the indices to gather the output from the params<\/li>\n<li><p><code>tf.stack<\/code> combines three separate tensors, the last of which is the indices.  The first two tensors specify the ordering of the first two dimensions (axis 0 and axis 1 of params\/indices)<\/p>\n\n<ul>\n<li>For the example provided, this ordering  is simply 0, 1, 2, ..., 63 for axis 0, and 0, 1, 2, ... 783 for axis 1.  These sequences are obtained with <code>tf.range(tf.shape(indices)[0])<\/code> and <code>tf.range(tf.shape(indices)[1])<\/code>, respectively.<\/li>\n<li><p>For the example provided, indices has shape (64, 784).  The other two tensors from the last point above need to have this same shape in order to be combined with <code>tf.stack<\/code><\/p>\n\n<ul>\n<li>First, an additional dimension\/axis is added to each of the two sequences using <code>tf.expand_dims<\/code>.<\/li>\n<li><p>The use of <code>tf.tile<\/code> and <code>tf.transpose<\/code> can be shown by example: Assume the first two axes of params and index have shape (5,3).  We want the first tensor to be:<\/p>\n\n<pre><code>[[0, 0, 0], [1, 1, 1], [2, 2, 2], [3, 3, 3], [4, 4, 4]]\n<\/code><\/pre>\n\n<p>We want the second tensor to be:<\/p>\n\n<pre><code>[[0, 1, 2], [0, 1, 2], [0, 1, 2], [0, 1, 2], [0, 1, 2]]\n<\/code><\/pre>\n\n<p>These two tensors almost function like specifying the coordinates in a grid for the associated indices.<\/p><\/li>\n<\/ul><\/li>\n<\/ul><\/li>\n<li><p>The final part of <code>tf.stack<\/code> combines the three tensors on a new third axis, so that the result has the same 3 axes as params.<\/p><\/li>\n<\/ul>\n\n<hr>\n\n<p>Keep in mind if you have more or less axes than in the question, you need to modify the number of coordinate-specifying tensors in <code>tf.stack<\/code> accordingly.<\/p>\n"},{"AnswerId":"42603183","CreationDate":"2017-03-05T00:54:40.163","ParentId":null,"OwnerUserId":"4567324","Title":null,"Body":"<p>What you want is like a custom reduction function. If you are keeping something like index of maximum value at <code>indices<\/code> then I would suggest using <code>tf.reduce_max<\/code>:<\/p>\n\n<pre><code>max_params = tf.reduce_max(params_tensor, reduction_indices=[2])\n<\/code><\/pre>\n\n<p>Otherwise, here is one way to get what you want (Tensor objects are not assignable so we create a 2d list of tensors and pack it using <code>tf.pack<\/code>):<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nwith tf.Graph().as_default():\n\n    params_tensor = tf.pack(np.random.randint(1,256, [5,5,10]).astype(np.int32))\n\n    indices = tf.pack(np.random.randint(1,10,[5,5]).astype(np.int32))\n\n    output = [ [None for j in range(params_tensor.get_shape()[1])] for i in range(params_tensor.get_shape()[0])]  \n    for i in range(params_tensor.get_shape()[0]):\n        for j in range(params_tensor.get_shape()[1]):\n            output[i][j] = params_tensor[i,j,indices[i,j]]\n    output = tf.pack(output)\n\n    with tf.Session() as sess:\n        params_tensor,indices,output = sess.run([params_tensor,indices,output])\n\n        print params_tensor\n        print indices\n        print output\n<\/code><\/pre>\n"}]}
{"QuestionId":42597531,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><gpu><deep-learning><cpu>","CreationDate":"2017-03-04T15:08:40.030","AcceptedAnswerId":42597874.0,"OwnerUserId":7636462.0,"Title":"Why GPUs comes more into play than CPUS when it comes to dealing with Deep Learning?","Body":"<p>In almost most of the cases, I come across about GPUs while dealing with any execution part in Deep Learning. <\/p>\n","answers":[{"AnswerId":"42597874","CreationDate":"2017-03-04T15:42:43.227","ParentId":null,"OwnerUserId":"4925151","Title":null,"Body":"<p>This has to do with GPU architecture versus CPU. It turns out gaming requires a lot of matrix multiplications, hence the GPU architecture was optimized for these types of operations, specifically they are optimized for high rate floating-point arithmetic. More on this <a href=\"https:\/\/graphics.stanford.edu\/papers\/gpumatrixmult\/gpumatrixmult.pdf\" rel=\"nofollow noreferrer\">here<\/a><\/p>\n\n<p>It so happens that neural networks are mostly matrix multiplications. <\/p>\n\n<p>For example:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lPctt.gif\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lPctt.gif\" alt=\"simple neural net with 1 hidden layer\"><\/a><\/p>\n\n<p>Is the mathematical formulation of a simple neural network with one hidden layer. W_h is a matrix of weights that multiplies your input x, to which we add a bias b_h. The linear equation W_hx + b_h can be compacted to a single matrix multiplication. The sigma is a nonlinear activation like sigmoid. The outer sigmoid is again a matrix multiplication. Hence GPUs<\/p>\n"}]}
{"QuestionId":42597701,"AnswerCount":3,"Tags":"<linux><tensorflow><libstdc++><rhel6>","CreationDate":"2017-03-04T15:26:23.070","AcceptedAnswerId":null,"OwnerUserId":6430839.0,"Title":"how to find libstdc++.so.6: that contain GLIBCXX_3.4.19 for RHEL 6?","Body":"<p>I work with a Linuxs server:<\/p>\n\n<pre><\/pre>\n\n<p>(from wikipedia:\n<strong><em>Red Hat Enterprise Linux 6<\/em><\/strong> was forked from <strong><em>Fedora 12<\/em><\/strong><\/p>\n\n<pre><\/pre>\n\n<p>The kernel<\/p>\n\n<pre><\/pre>\n\n<p>gcc version<\/p>\n\n<pre><\/pre>\n\n<p>when trying to install Tensdorflow, I have some issue with a missing library:<\/p>\n\n<pre><\/pre>\n\n<p>As you can see in my <strong><em>lib64\/libstdc++.so.6<\/em><\/strong> I don't have <strong><em>GLIBCXX_3.4.19<\/em><\/strong><\/p>\n\n<pre><\/pre>\n\n<p>I look with the following command:<\/p>\n\n<pre><\/pre>\n\n<p>so my question where can I find the <strong><em>libstdc++-4.x.y-z.el6.x86_64<\/em><\/strong> that contain <strong><em>GLIBCXX_3.4.19<\/em><\/strong> and that I could install on my server RHEL 6  with my gcc version ?<\/p>\n\n<p>I am a little bit confuse of which version to use from this page\n<a href=\"https:\/\/www.rpmfind.net\/linux\/rpm2html\/search.php?query=libstdc%2B%2B.so.6%28GLIBCXX_3.4.19%29&amp;submit=Search%20...&amp;system=&amp;arch=\" rel=\"nofollow noreferrer\">https:\/\/www.rpmfind.net\/linux\/rpm2html\/search.php?query=libstdc%2B%2B.so.6%28GLIBCXX_3.4.19%29&amp;submit=Search+...&amp;system=&amp;arch=<\/a><\/p>\n\n<ol>\n<li>I don't see a version for Fedora 12 or RHEL6<\/li>\n<li>I need arch x86_64<\/li>\n<li>I need I guess something  gcc (GCC) 4.4.7<\/li>\n<li>I guess I need 64bit<\/li>\n<\/ol>\n\n<p>I the following a good macth ? I am quite confuse by the name and info:<\/p>\n\n<pre><\/pre>\n\n<p>I don't have Internet access on the server so I need to download the file locally copy it on the server and do the intsallion.<\/p>\n\n<p>Thanks for your help<br>\nCheers<br>\nFabien<\/p>\n\n<p>ps: bonus question\nSuSE 4.4.7-17 (GCC), I am confuse, I understood that Suse was a Linux distribution like Redhat!<\/p>\n","answers":[{"AnswerId":"57257676","CreationDate":"2019-07-29T16:10:45.070","ParentId":null,"OwnerUserId":"4641567","Title":null,"Body":"<p>Previous link is dead now as well, you can use this one<\/p>\n\n<pre><code>http:\/\/ftp.de.debian.org\/debian\/pool\/main\/g\/gcc-4.8\/libstdc++6-4.8-dbg_4.8.4-1_i386.deb\n<\/code><\/pre>\n\n<p>In previous comment from Serge, there's a <code>cd<\/code> missing\n    cd \/usr\/lib\/\nOr like in my case\n    cd \/usr\/lib64<\/p>\n"},{"AnswerId":"50811874","CreationDate":"2018-06-12T07:46:52.527","ParentId":null,"OwnerUserId":"7490134","Title":null,"Body":"<p>I found a solution here:\n<a href=\"https:\/\/www.linuxquestions.org\/questions\/red-hat-31\/lib-libstdc-so-6-version-%60glibcxx_3-4-15\" rel=\"nofollow noreferrer\">https:\/\/www.linuxquestions.org\/questions\/red-hat-31\/lib-libstdc-so-6-version-%60glibcxx_3-4-15<\/a>'-not-found-4175419985\/<\/p>\n\n<blockquote>\n  <p>Replacing libstdc++-so.6 with a later version that works in EL6:\n  Unpack libstdc++6_4.7.1-2_i386.deb\n  <a href=\"http:\/\/ftp.de.debian.org\/debian\/pool...7.1-2_i386.deb\" rel=\"nofollow noreferrer\">http:\/\/ftp.de.debian.org\/debian\/pool...7.1-2_i386.deb<\/a>\n  with : ar -x libstdc++6_4.7.1-2_i386.deb &amp;&amp; tar xvf data.tar.gz\n  Next : 1) su ; 2) cp libstdc++.so.6.0.17 \/usr\/lib\/\n  3) cd \/usr\/lib\/ &amp;&amp; rm libstdc++.so.6\n  4) ln -s libstdc++.so.6.17 libstdc++.so.6<\/p>\n  \n  <p>Reason for suggesting the Debian package :\n  It's a ( gcc ) libstdc++ version that's compiled with a glibc\n  old enough to be used in EL6 \/ CentOS 6.<\/p>\n<\/blockquote>\n"},{"AnswerId":"53856085","CreationDate":"2018-12-19T17:08:42.133","ParentId":null,"OwnerUserId":"1915854","Title":null,"Body":"<p>Updated steps (because it seems the file has been moved):<\/p>\n\n<pre><code>curl -O http:\/\/ftp.de.debian.org\/debian\/pool\/main\/g\/gcc-4.7\/libstdc++6-4.7-dbg_4.7.2-5_i386.deb\nar -x libstdc++6-4.7-dbg_4.7.2-5_i386.deb &amp;&amp; tar xvf data.tar.gz\nmkdir backup\ncp \/usr\/lib\/libstdc++.so* backup\/\ncp .\/usr\/lib\/i386-linux-gnu\/debug\/libstdc++.so.6.0.17 \/usr\/lib\nln -s libstdc++.so.6.0.17 libstdc++.so.6\n<\/code><\/pre>\n"}]}
{"QuestionId":42598630,"AnswerCount":1,"Tags":"<python><keras>","CreationDate":"2017-03-04T16:53:40.293","AcceptedAnswerId":42598835.0,"OwnerUserId":7216502.0,"Title":"Why can't I use preprocessing module in Keras?","Body":"<p>I'm trying to use the function pad_sequences() but the same error keeps rising: 'AttributeError: 'module' object has no attribute 'sequence''<\/p>\n\n<p>I have followed Keras documentation and I can't figure out why It does not work. Here is the line of code:<\/p>\n\n<pre><\/pre>\n\n<p>None of the above lines seem to work.<\/p>\n","answers":[{"AnswerId":"42598835","CreationDate":"2017-03-04T17:11:48.580","ParentId":null,"OwnerUserId":"2704763","Title":null,"Body":"<p>In the first line please use <\/p>\n\n<pre><code>X_test = preprocessing.sequence.pad_sequences(X_test, maxlen=500)\n<\/code><\/pre>\n\n<p>You can simply import pad_sequences like so instead<\/p>\n\n<pre><code>from keras.preprocessing.sequence import pad_sequences\n<\/code><\/pre>\n\n<p>and replace <code>preprocessing.sequence.pad_sequences<\/code> with just <code>pad_sequences<\/code><\/p>\n"}]}
{"QuestionId":42598841,"AnswerCount":0,"Tags":"<machine-learning><tensorflow><autoencoder><tflearn>","CreationDate":"2017-03-04T17:12:33.433","AcceptedAnswerId":null,"OwnerUserId":5540592.0,"Title":"using validation monitor in tflearn.regression to create confusion matrix","Body":"<p>So I have been trying to create a confusion metrics in my autoencoder<\/p>\n\n<pre><\/pre>\n\n<p>I have gone through the documantation but they were not very useful to me.<\/p>\n\n<p>How would I configure to get the confusion matrix?<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42599150,"AnswerCount":1,"Tags":"<video><machine-learning><tensorflow><thumbnails>","CreationDate":"2017-03-04T17:40:29.103","AcceptedAnswerId":42647325.0,"OwnerUserId":2116739.0,"Title":"Auto selecting a thumbnail for a video","Body":"<p>We all heard about YouTube's use of Deep Learning to select a representative thumbnail for user videos. But has anyone tried it on tensorflow with success?<\/p>\n\n<p>I did find <a href=\"https:\/\/github.com\/yahoo\/hecate\" rel=\"nofollow noreferrer\">https:\/\/github.com\/yahoo\/hecate<\/a> which claims to do that but was less than impressed with the results. I actually got better results using ffmpeg to extract key frames and then calculating the color distribution to select the \"best\" image. <\/p>\n\n<p>But would love to know if someone is doing it better using more \"intelligent\" algorithms.<\/p>\n","answers":[{"AnswerId":"42647325","CreationDate":"2017-03-07T11:37:21.917","ParentId":null,"OwnerUserId":"3531912","Title":null,"Body":"<p><em>I want to be clear with the OP that this answer does not represent a formal description of the approach so as to describe the prospective approach in an intuitive way.<\/em><\/p>\n\n<p>Suppose that a video is composed by <em>n<\/em> frames and that each of them can be represented as a 3D tensor (height, width, channel). It is possible to use a <em>Convolutional Neural Networks (CNNs)<\/em> to generate a latent representation for each frame.<\/p>\n\n<p>A video can be represented as a sequence of frames (f_1, f_2, ..., f_n). The most suitable neural network architecture for sequence modeling are Recurrent Neural Networks (RNNs). We can use RNNs to encode a sequence of video frames latent representations generated by a CNN. After that, you will have a latent representations (<strong>f_1<\/strong>, <strong>f_2<\/strong>, ..., <strong>f_n<\/strong>) for each frame generated by the RNN which directly depends from the previous ones (this is a well-known property of RNNs). <\/p>\n\n<p>As you can see in the recently released <a href=\"https:\/\/research.google.com\/youtube8m\/explore.html\" rel=\"nofollow noreferrer\">Youtube-8M dataset<\/a>, there are high quality thumbnails associated to each video so you can use them as targets. In particular, given the latent representations generated by the RNNs applied on the sequence of frames, you can generate a context vector <strong>c<\/strong> which is generated as follows:<\/p>\n\n<pre><code>alpha = softmax(FNN(f_1), FNN(f_2), ..., FNN(f_n))\nc = f_1 * alpha_1 + f_2 * alpha_2 + ... + f_n * alpha_n\n<\/code><\/pre>\n\n<p>where <em>FNN<\/em> is a feedforward neural network that receives the latent representation <strong>f_i<\/strong> for the frame f_i and generates a score which represents how much important is in the current sequence. We can exploit the context vector <strong>c<\/strong> in order to predict the most suitable frame of the video. <\/p>\n\n<p>In my opinion there are two possible strategies to define a loss function for the minimization problem that the network should solve. The first one it is easier than the second one. I briefly describe them as follows:<\/p>\n\n<ul>\n<li><strong>predict thumbnail index<\/strong>: by exploiting the context vector <strong>c<\/strong>, we can train the network to predict an integer value which represents the position of the selected frame by minimizing a cross-entropy loss between the generated index and the target one;<\/li>\n<li><strong>reconstruction error<\/strong>: by exploiting the context vector <strong>c<\/strong>, we can train the network to generate a new image by minimizing the <em>reconstruction error<\/em> evaluated between the image generated by the model and the target one.<\/li>\n<\/ul>\n\n<p>I have not tried any of them in practice so I cannot be sure that my approach will work but I am confident that it is something reasonable to do so as to effectively complete this task. Anyway, I hope that this answer may be helpful to the OP in order to better understand how this task can be solved.<\/p>\n"}]}
{"QuestionId":42599241,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-03-04T17:48:44.443","AcceptedAnswerId":null,"OwnerUserId":1952982.0,"Title":"Does RMSProp optimizer in tensorflow use Nesterov momentum?","Body":"<p>When you create RMSPRop optimizer, it asks for the momentum value. What is this momentum? Is it Nesterov or the other one? How do I use Nesterov momentum with the RMSProp in tf?<\/p>\n\n<p>There is a formula in the doc string here:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/rmsprop.py#L25\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/training\/rmsprop.py#L25<\/a><\/p>\n\n<pre><\/pre>\n\n<p>Could someone explain what's the g_t term mean and where is this formula being computed?<\/p>\n\n<p>As far as I understand, in Nesterov momentum+rmsprop you first change the weights with the current momentum, compute new gradients, divide them by the sqrt(mean_square+epsilon) and add to the momentum. Is this what's happening here? I wasn't able to find the implementation of training_ops.apply_rms_prop since I'm not very familiar with the tf source.<\/p>\n\n<p>I'm coming from coursera Geoffrey Hinton course about neural networks, where this Nesterov momentum+rmsprop algorithm is explained. How do I use it in tf?<\/p>\n\n<p>Please, correct me if I'm wrong in my understanding of Nesterov momentum or any other thing.<\/p>\n","answers":[{"AnswerId":"49502140","CreationDate":"2018-03-26T23:22:28.947","ParentId":null,"OwnerUserId":"6383205","Title":null,"Body":"<ol>\n<li><p>The documentation you refer to explicitly mentions:<\/p>\n\n<blockquote>\n  <p>This implementation of RMSProp uses plain momentum, not Nesterov momentum.<\/p>\n<\/blockquote><\/li>\n<li><p>AFAIK there is no built-in implementation for Nesterov momentum in RMSProp. You can of course adjust the function according to your own needs.<\/p><\/li>\n<li><p>As @xolodec said, <code>g_t<\/code> is the gradient.<\/p><\/li>\n<\/ol>\n"}]}
{"QuestionId":42599336,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-04T17:57:32.527","AcceptedAnswerId":null,"OwnerUserId":3291312.0,"Title":"Multivariate Logistic Regression in Tensorflow without Built in Optimizers","Body":"<p>I'm attempting to construct a multi-class logistic regressor on 25112 28x28 images that are handwritten, from 0-4, by implementing stochastic gradient descent with L2 regularization, but without use of traditional tensorflow functions (I'm trying to build it from scratch). I'm new to this so it's possible I've made a few mistakes. Thank you for taking the time!<\/p>\n\n<p>I believe I'm:<\/p>\n\n<p>Reading in the training and testing images correctly<\/p>\n\n<p>Have set up the tf.Session() correctly<\/p>\n\n<p>I'm trying to print out the results for the learned weights and the corresponding bias, but I'm not sure I'm doing that correctly. <\/p>\n\n<p>I'm getting the following error:<\/p>\n\n<pre><\/pre>\n\n<p>My code is below:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42635890","CreationDate":"2017-03-06T21:28:00.720","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>There are a lot of weird things in this code. For example, classifier.train doesn't actually train the classifier, it just builds ops to do so. You need to create a session to run ops in tensorflow.<\/p>\n\n<p>I recommend starting from the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/mandelbrot\" rel=\"nofollow noreferrer\">basic tutorials in the tensorflow website<\/a>.<\/p>\n"}]}
{"QuestionId":42599827,"AnswerCount":0,"Tags":"<ubuntu><caffe><libtiff>","CreationDate":"2017-03-04T18:37:05.140","AcceptedAnswerId":null,"OwnerUserId":7659071.0,"Title":"Getting undefined reference errors while installing Caffe in Ubuntu 15.10 - is it caused by libtiff?","Body":"<p>I'm trying to install Caffe in Ubuntu 15.10. After cloning the <a href=\"https:\/\/github.com\/BVLC\/caffe\" rel=\"nofollow noreferrer\">Caffe repository<\/a>, I copied  to a new file, , and uncommented  in line 91.<\/p>\n\n<p>Then, I executed<\/p>\n\n<pre><\/pre>\n\n<p>However, I get the following errors:<\/p>\n\n<pre><\/pre>\n\n<p>Looking at the error messages, it seems that I need the  package installed. I already have  and  installed, and according to <a href=\"https:\/\/stackoverflow.com\/a\/37316325\">this<\/a>, it should be sufficient. However, it has not resolved the issue.<\/p>\n\n<p>Is this the cause of the problem? If so, how do I correct it? If not, what should I do to resolve this?<\/p>\n\n<p>Output of :<\/p>\n\n<pre><\/pre>\n\n<p>Thank you<\/p>\n","answers":[]}
{"QuestionId":42600499,"AnswerCount":2,"Tags":"<python><tensorflow><keras><tensorboard>","CreationDate":"2017-03-04T19:38:54.323","AcceptedAnswerId":null,"OwnerUserId":5739751.0,"Title":"Tensorboard Cannot find .runfiles directory error","Body":"<p>I installed tensorboard via pip and when I try to execute   I get the following error <\/p>\n\n<pre><\/pre>\n\n<p>I do  and get the following <\/p>\n\n<pre><\/pre>\n\n<p>thanks in advance. <\/p>\n","answers":[{"AnswerId":"44622309","CreationDate":"2017-06-19T04:34:25.123","ParentId":null,"OwnerUserId":"6636171","Title":null,"Body":"<p>We have to search for the tensorboard folder and run the tensorboard file there.<\/p>\n\n<p>Search for the tensorflow folder and do the following(My tensorflow folder was there in ~\/ itself):<\/p>\n\n<pre><code>cd ~\/tensorflow\/lib\/python2.7\/site-packages\/tensorboard\n<\/code><\/pre>\n\n<p>Now run:<\/p>\n\n<pre><code>python tensorboard --logdir=(the location of your logs path)\n<\/code><\/pre>\n"},{"AnswerId":"43174814","CreationDate":"2017-04-02T23:54:40.477","ParentId":null,"OwnerUserId":"1623645","Title":null,"Body":"<p>It seems they didn't consider that someone would be pip-installing TensorBoard in a user directory. Below is my <em>hack<\/em> to get it to work:<\/p>\n\n<p>In the <code>~\/.local\/bin\/tensorboard<\/code> script, there's a section that looks like this:<\/p>\n\n<pre><code>def FindModuleSpace():\n  # Follow symlinks, looking for my module space\n  stub_filename = os.path.abspath(sys.argv[0])\n  while True:\n    # Found it?\n    module_space = stub_filename + '.runfiles'\n    if os.path.isdir(module_space):\n      break\n    for mod in site.getsitepackages():\n      module_space = mod + '\/tensorboard\/tensorboard' + '.runfiles'\n      if os.path.isdir(module_space):\n        return module_space\n<\/code><\/pre>\n\n<p>(just above the assertion with the \"Cannot find .runfiles directory\" error).<\/p>\n\n<p>The directory it's looking for is<\/p>\n\n<pre><code>~\/.local\/lib\/python2.7\/site-packages\/tensorboard\/tensorboard.runfiles\n<\/code><\/pre>\n\n<p>which you can discover by running <code>find ~\/.local -name '*runfiles*'<\/code>.<\/p>\n\n<p>I simply added it to the <code>for<\/code> loop over directories and all is well:<\/p>\n\n<pre><code>    for mod in site.getsitepackages() + [os.path.expanduser(\"~\/.local\/lib\/python2.7\/site-packages\")]\n<\/code><\/pre>\n\n<p>This is a <em>hack<\/em> because:<\/p>\n\n<ul>\n<li>I explicitly said <code>python2.7<\/code>, which might not be the version of Python you're using. Correct it for your case.<\/li>\n<li>The use of forward slashes won't work on Windows; a chain of <code>os.path.join<\/code> would be better.<\/li>\n<li>This issue really should be communicated back to the TensorBoard developers. Did you do that?<\/li>\n<\/ul>\n"}]}
{"QuestionId":42602381,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-04T23:00:56.717","AcceptedAnswerId":null,"OwnerUserId":4778857.0,"Title":"Tensorflow error on install with Windows","Body":"<p>Im installing Tensorflow-GPU on my PC<\/p>\n\n<pre><\/pre>\n\n<p>It obviously goes through to process the command but I am hoping to work out how to overcome the exceptions listed. I have installed NVIDIA Tooolkit, etc as per instructions on tensorflow.org<\/p>\n\n<p>Any help appreciated.<\/p>\n","answers":[]}
{"QuestionId":42602773,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><keras>","CreationDate":"2017-03-04T23:52:31.997","AcceptedAnswerId":42602846.0,"OwnerUserId":5203167.0,"Title":"keras functional api and multiple merge layers","Body":"<p>I am trying to design a neural network using keras. The model.summary() output is different compared to defined layers<\/p>\n\n<pre><\/pre>\n\n<p>model.summary() output<\/p>\n\n<pre><\/pre>\n\n<p>seq4, seq6, seq8 layers are missing in the model.summary() output.\nAm I doing something wrong?<\/p>\n","answers":[{"AnswerId":"42602846","CreationDate":"2017-03-05T00:04:09.357","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>You are not using them to compute the output. <\/p>\n\n<p>Take for example seq4 : you feed it to seq7, which is fed to seq8 which doesnt go anywhere. <\/p>\n\n<p>There is a problem in the structure of the tree of you model. <\/p>\n\n<p>In the one that is summarised, they took all the layers that lead to <code>output=out<\/code> from <code>input=inp<\/code>, the ones that aren't used in that \"path\" won't be part of the graph of your model. <\/p>\n\n<p>The flow going through seq4, seq6, seq7 and seq8 are not leading to the output of your model. <\/p>\n\n<p>Does that help you?<\/p>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>The merge layer functions like this example from your code : <\/p>\n\n<pre><code>seq3 = merge([seq2, seq1], mode=\"concat\", concat_axis=1)\n<\/code><\/pre>\n\n<p>Here you take what comes out of the layers seq2 and seq1, they have <code>shapes = (None,32,113,113)<\/code>. That's the input of that layer, two different tensors coming out of seq2 and seq1. \nYou precised that you want to concatenate those 4D tensors following the axis 1. \nTherefore, the output of that merge layer will be of <code>shape = (None,64,113,113)<\/code>. the two 32 have been added together during concatenation. \nYou can read what I've just explained at the line \"merge_1\" of your <code>model.summary()<\/code><\/p>\n"}]}
{"QuestionId":42603407,"AnswerCount":3,"Tags":"<c++><windows><msbuild><tensorflow>","CreationDate":"2017-03-05T01:28:16.743","AcceptedAnswerId":null,"OwnerUserId":1181124.0,"Title":"How to compile Tensor Flow with SSE and and AVX instructions on Windows?","Body":"<p>With the latest version of Tensor Flow now on windows, I am trying to get everything working as efficiently as possible. However, even when compiling from source I still can't seem to figure out how to enable the SSE and AVX instructions. <\/p>\n\n<p>The default process:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r0.12\/tensorflow\/contrib\/cmake\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r0.12\/tensorflow\/contrib\/cmake<\/a>\nhas no mention of how to do this.<\/p>\n\n<p>The only reference I have found has been using Google's Bazel:\n<a href=\"https:\/\/stackoverflow.com\/questions\/41293077\/how-to-compile-tensorflow-with-sse4-2-and-avx-instructions\">How to compile Tensorflow with SSE4.2 and AVX instructions?<\/a><\/p>\n\n<p>Does anyone know of an easy way to turn on these advanced instructions using MSBuild? I hear they give at least a 3X speed up.<\/p>\n\n<p>To help those looking for a similar solution, this is the warning I am currently getting looks like this:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r0.12\/tensorflow\/contrib\/cmake\" rel=\"noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/tree\/r0.12\/tensorflow\/contrib\/cmake<\/a><\/p>\n\n<p>I am using Windows 10 Professional on a 64 bit platform, Visual Studio 2015 Community Edition, Anaconda Python 3.6 with cmake version 3.6.3 (later versions dont' work for Tensor Flow)<\/p>\n\n<p>Any help appreciated!<\/p>\n","answers":[{"AnswerId":"45203891","CreationDate":"2017-07-20T02:05:00.760","ParentId":null,"OwnerUserId":"6359063","Title":null,"Body":"<p>Tensorflow make a mistake on flag \"tensorflow_WIN_CPU_SIMD_OPTIONS\". <\/p>\n\n<p>It is a Flag, not a Boolean. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/cizF7.png\" rel=\"nofollow noreferrer\">Before Fix Image<\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/v9vEn.png\" rel=\"nofollow noreferrer\">After Fix Image<\/a><\/p>\n\n<hr>\n\n<h2>How to Fix it<\/h2>\n\n<p>\"Tensorflow-github\/tensorflow\/contrib\/cmake\/CMakeLists.txt\" Line 34,there is:<\/p>\n\n<pre><code>option(tensorflow_WIN_CPU_SIMD_OPTIONS \"Enables CPU SIMD instructions\")\n<\/code><\/pre>\n\n<p>Replace it wtih<\/p>\n\n<pre><code>set(tensorflow_WIN_CPU_SIMD_OPTIONS \"\/arch:AVX\" CACHE STRING \"Enables CPU SIMD instructions\" )\n<\/code><\/pre>\n\n<p>Then, clear the cmake cache (<a href=\"https:\/\/i.stack.imgur.com\/5eSnf.png\" rel=\"nofollow noreferrer\">location<\/a>), and reconfigure.<\/p>\n\n<p>You will find tensorflow_WIN_CPU_SIMD_OPTIONS is a Flag with Input area instead of checkbox.<\/p>\n\n<hr>\n\n<h2>tensorflow_WIN_CPU_SIMD_OPTIONS<\/h2>\n\n<p>\"\/arch:AVX\" or \"\/arch:AVX2\" is available <\/p>\n"},{"AnswerId":"42755665","CreationDate":"2017-03-13T02:17:34.047","ParentId":null,"OwnerUserId":"4503060","Title":null,"Body":"<p>Well, I tried to fix that, but I am not sure if it really worked.<\/p>\n\n<p>In <code>CMakeLists.txt<\/code> you will find the following statements:<\/p>\n\n<pre><code>if (tensorflow_OPTIMIZE_FOR_NATIVE_ARCH)\n  include(CheckCXXCompilerFlag)\n  CHECK_CXX_COMPILER_FLAG(\"-march=native\" COMPILER_OPT_ARCH_NATIVE_SUPPORTED)\n<\/code><\/pre>\n\n<p>On MSVC platform, the test failes because MSVC doesn't support <code>-march=native<\/code> flag. I modified the statements like below:<\/p>\n\n<pre><code>if (tensorflow_OPTIMIZE_FOR_NATIVE_ARCH)\n  include(CheckCXXCompilerFlag)\n  CHECK_CXX_COMPILER_FLAG(\"-march=native\" COMPILER_OPT_ARCH_NATIVE_SUPPORTED)\n  if (COMPILER_OPT_ARCH_NATIVE_SUPPORTED)\n    set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} -march=native\")\n  else()\n    CHECK_CXX_COMPILER_FLAG(\"\/arch:AVX\" COMPILER_OPT_ARCH_AVX_SUPPORTED)\n    if(COMPILER_OPT_ARCH_AVX_SUPPORTED)\n      set(CMAKE_CXX_FLAGS \"${CMAKE_CXX_FLAGS} \/arch:AVX\")\n    endif()\n  endif()\nendif()\n<\/code><\/pre>\n\n<p>By doing this, cmake would check if <code>\/arch:AVX<\/code> is available and use it. Accordinf to <a href=\"https:\/\/msdn.microsoft.com\/en-us\/library\/7t5yh4fd.aspx\" rel=\"noreferrer\">MSDN<\/a> and <a href=\"https:\/\/msdn.microsoft.com\/en-us\/library\/jj620901.aspx\" rel=\"noreferrer\">MSDN<\/a>, SSE2 support is enabled by default for x86 compiling but not available for x64 compiling. For x64 compiling you can choose to use AVX or AVX2. I used AVX above because my CPU only supports AVX, youcan try AVX2 if you have a compatible CPU.<\/p>\n\n<p>By compiling use the above <code>CMakeLists.txt<\/code>, the compiling preocedure was much slower than official release, and warning about 'AVX\/AVX2' disappeared, but warning about SSE\/SSE2\/3\/4.1\/4.2 still exists. I think these warnings can be ignored because there's no SSE support for x64 MSBuild.<\/p>\n\n<p>I am testing the new pip package now. It maybe faster than before, but I don't want to write a new benchmark ...<\/p>\n\n<p>Any one who is interested in this, please test if the new package is really faster.<\/p>\n\n<p>I did all these on the lasted git master branch, 2017-3-12. The pip package name shows that it was tensorflow 1.0.1<\/p>\n"},{"AnswerId":"42613548","CreationDate":"2017-03-05T20:27:04.227","ParentId":null,"OwnerUserId":"1254250","Title":null,"Body":"<p>I think you would have to add <code>\/arch:avx2<\/code> to compiler flags.\nOne way to do it is to modify your <code>CMakeCache.txt<\/code> in your build folder. Looking for a line <code>CMAKE_CXX_FLAGS:STRING<\/code> and modify it to<\/p>\n\n<blockquote>\n<pre><code>CMAKE_CXX_FLAGS:STRING=\/DWIN32 \/D_WINDOWS \/W3 \/GR \/EHsc \/arch:AVX2 \/fp:fast\n<\/code><\/pre>\n<\/blockquote>\n\n<p>However, according to <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/7966\" rel=\"nofollow noreferrer\">this issue<\/a> on github. <code>\/arch:avx2<\/code> is broken at the moment (at HEAD).<\/p>\n"}]}
{"QuestionId":42603427,"AnswerCount":2,"Tags":"<tensorflow><google-cloud-ml><google-cloud-ml-engine>","CreationDate":"2017-03-05T01:33:01.460","AcceptedAnswerId":null,"OwnerUserId":5146798.0,"Title":"Error when Submitting Job Training in Google Cloud ML","Body":"<p>I'm currently trying to submit a job training on Google Cloud ML with the Facenet (a Tensorflow library for face recognition). I'm currently trying this<a href=\"https:\/\/github.com\/davidsandberg\/facenet\/wiki\/Classifier-training-of-inception-resnet-v1\" rel=\"nofollow noreferrer\"> (link is here) <\/a> part of the library where it does the training for the model. <\/p>\n\n<p>Going to Google Cloud ML, I'm following this tutorial <a href=\"https:\/\/cloud.google.com\/ml\/docs\/how-tos\/training-models\" rel=\"nofollow noreferrer\"> (link is here) <\/a> where it teaches you how to submit a training.<\/p>\n\n<p>I was able to successfully submit a job training to Google Cloud ML but there was an error.\nHere are some pictures of the errors:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Xps88.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Xps88.png\" alt=\"Here&#39;s a picture of the error:\"><\/a><\/p>\n\n<p>And here's an error from the Google Cloud Jobs logs<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/aJ4HI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/aJ4HI.png\" alt=\"Here&#39;s a picture of error log in Google Cloud Jobs\"><\/a><\/p>\n\n<p>Here are more detailed pictures on Google Cloud Job logs<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/Fiufn.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Fiufn.png\" alt=\"Here are more detailed pictures on Google Cloud Job logs (1) \"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/tuWal.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tuWal.png\" alt=\"Here are more detailed pictures on Google Cloud Job logs (2)\"><\/a><\/p>\n\n<p>Submitting a job request was a success and it was even waiting for Tensorflow to start but right after that there's that error. <\/p>\n\n<p>The commands I used to run this is here:<\/p>\n\n<pre><\/pre>\n\n<p>Any suggestions on how to fix this? Thanks!<\/p>\n","answers":[{"AnswerId":"42612041","CreationDate":"2017-03-05T18:09:23.977","ParentId":null,"OwnerUserId":"6896033","Title":null,"Body":"<p>Looking at the error message above, it seems that you issue is related to \"ImportError: import by filename is not supported\" error from Python. Without looking at your python source code, I can't tell you exactly how to fix it but the following link should solve your issue:<\/p>\n\n<p><a href=\"https:\/\/stackoverflow.com\/questions\/25997185\/python-importerror-import-by-filename-is-not-supported\">Python \/ ImportError: Import by filename is not supported<\/a><\/p>\n\n<p>In general look for places that you are importing using file paths and make sure you are using the functions correctly.<\/p>\n"},{"AnswerId":"42699315","CreationDate":"2017-03-09T15:31:47.033","ParentId":null,"OwnerUserId":"4392784","Title":null,"Body":"<p>When you are running on Cloud ML Engine you are running in a remote environment; so the file paths will not be the same as the local environment. If you need to import python modules you need to include them in the Python package you build and then import them using the package name.<\/p>\n\n<p>For docs on how to build packages please refer to the <a href=\"http:\/\/setuptools.readthedocs.io\/en\/latest\/setuptools.html\" rel=\"nofollow noreferrer\">SetupTools docs<\/a><\/p>\n\n<p>Here's the 30 second version<\/p>\n\n<ol>\n<li>Organize your code as follows<\/li>\n<\/ol>\n\n<pre><code>\n    my_package\/__init__.py\n    my_package\/moduleA.py\n    my_package\/moduleB.py\n    my_package\/...\n    setup.py\n<\/code><\/pre>\n\n<ol start=\"2\">\n<li>For your setup.py file start with this<\/li>\n<\/ol>\n\n<pre>\n    from setuptools import find_packages\n    from setuptools import setup\n\n    REQUIRED_PACKAGES = []\n\n    setup(\n        name='my_package',\n        version='0.1.1',\n        author='Author',\n        author_email='author@gmail.com',\n        install_requires=REQUIRED_PACKAGES,\n        packages=find_packages(),\n        description='Description',\n        requires=[],)\n<\/pre>\n\n<ol start=\"3\">\n<li>Build a package as follows<\/li>\n<\/ol>\n\n<pre>\npython .\/setup.py sdist\n<\/pre>\n\n<ol start=\"4\">\n<li>When the package is installed in Cloud ML Engine you will be able to import your code as<\/li>\n<\/ol>\n\n<pre>\nfrom my_package import moduleA\n<\/pre>\n"}]}
{"QuestionId":42603459,"AnswerCount":1,"Tags":"<time-series><keras><prediction><lstm><recurrent-neural-network>","CreationDate":"2017-03-05T01:39:33.677","AcceptedAnswerId":42606452.0,"OwnerUserId":1058511.0,"Title":"lstm for prediction of future time series values with Keras","Body":"<p>I want to make prediction of future values for a multivariate time series. I use this code as the template, but modified the output activation to linear<\/p>\n\n<pre><\/pre>\n\n<p>The input data is a sequence of 8 dimensional vector of double values. The output is the same 8 dimensional vector one time period ahead of the input sequence. This is the reason for choosing the linear activation. <\/p>\n\n<pre><\/pre>\n\n<p>I am new to Keras and RNN. Could someone confirm this design?<\/p>\n","answers":[{"AnswerId":"42606452","CreationDate":"2017-03-05T09:06:50.973","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>You have 8x2 inputs in each sample, for every of those 8 time step you encode 512 features that you keep track of. Then again, 512 values... The number of parameters must be huge? I hope that you have enough data and variety in the patterns to make use of that huge number of parameters, otherwise you will overfit. <\/p>\n\n<p>I would also avise you to use dropout for the LSTM like this : <\/p>\n\n<pre><code>model.add(LSTM(512, return_sequences=True, dropout_W = 0.2)) #- original\n<\/code><\/pre>\n\n<p>There is also the feature dropout_U, but I would not use that one. The LSTM has multiple gates as you should know, each of these gates are dense layers, so you can choose on which one you want to apply the dropout. Best practice is to apply it on the inputs gate (dropout_W) and not on recurrent gate (dropout_U). <\/p>\n\n<p>Otherwise the overall architecture makes sense. <\/p>\n"}]}
{"QuestionId":42603702,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><autoencoder>","CreationDate":"2017-03-05T02:15:01.463","AcceptedAnswerId":null,"OwnerUserId":3813674.0,"Title":"variable_scope causes 'variable does not exist' with Optimizer","Body":"<p>Supoose I have two models  and . Assume  is pretrained and loaded. I want to define the cost function for  as roughly sketched in the following code (it's actually an autoencoder). Please note that this is an minimal example to reproduce my problem so they don't make sense mathematically.<\/p>\n\n<pre><\/pre>\n\n<p>If I run the script (TF version 1.0), I get the following error: <\/p>\n\n<pre><\/pre>\n\n<p>However, this doesn't happen with . Any explanation and pointer would be appreciated.<\/p>\n","answers":[{"AnswerId":"44477277","CreationDate":"2017-06-10T19:29:55.820","ParentId":null,"OwnerUserId":"7373907","Title":null,"Body":"<p>Your ValueError is caused by creating new variables within the variable_scope.reuse==True. <\/p>\n\n<p>Variables are created by Adam, when you call the minimize function of Adam, for saving momentums of each trainable variables in your graph.<\/p>\n\n<p>You have set reuse to True, thus the default variable_scope.reuse==True. The reuse state cannot change back to False forever once you set it to True. Then, Adam creates variable under state reuse==True, which raises an error.<\/p>\n\n<p>The solution is to add a sub scope under the graph's default scope when you set variable_scope.reuse=True, then the default scope.reuse is still False, and Adam.minimize will work, as follows:<\/p>\n\n<pre><code>import tensorflow as tf\ndef foo(X):\n        with tf.variable_scope(\"foo\"):\n                A = tf.get_variable(\"A\",shape=[1])\n        return tf.add(X,A)\n\ndef bar(X):\n        with tf.variable_scope(\"bar\"):\n                B = tf.get_variable(\"B\",shape=[1])\n        return tf.multiply(X,B)\n\nX = tf.placeholder(\"float\")\n\nwith tf.variable_scope(\"for_reuse_scope\"):\n    X_prime = foo(X)\n    Y = bar(X)\n    tf.get_variable_scope().reuse_variables()\n    Y_prime = bar(X_prime)\n\n\n#foo(X) is manipulated with some other terms, but the point is foo is called again\ncost = foo(X) + tf.pow(Y-Y_prime,2) \n\noptimizer = tf.train.AdamOptimizer(learning_rate=0.01).minimize(cost)\n<\/code><\/pre>\n"}]}
{"QuestionId":42603821,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-03-05T02:34:55.677","AcceptedAnswerId":null,"OwnerUserId":460760.0,"Title":"Difference between ClientSession and Session in TensorFlow C++ API","Body":"<p>TensorFlow r1.0 C++ API comes with  and  classes. Some of the examples shipping with TensorFlow use  and others use . Do these two different types of session use the same underlying mechanism under the hood or is one of the preferred over another? The syntax for using them is a bit different but other than that are there any differences in behavior?<\/p>\n","answers":[{"AnswerId":"42603895","CreationDate":"2017-03-05T02:46:54.120","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In TensorFlow's C++ API, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/client-session\" rel=\"noreferrer\"><code>tensorflow::Session<\/code><\/a> API is a low-level interface that deals with serialized <code>GraphDef<\/code> protocol buffers and provides a string-based interface for running subgraphs. <\/p>\n\n<p>By contrast, the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/cc\/class\/tensorflow\/client-session\" rel=\"noreferrer\"><code>tensorflow::ClientSession<\/code><\/a> API is higher level, and integrates with the new C++ API for building TensorFlow graphs\u2014much in the same way as the Python <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph\" rel=\"noreferrer\"><code>tf.Graph<\/code><\/a> and <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session\" rel=\"noreferrer\"><code>tf.Session<\/code><\/a> classes do. <\/p>\n\n<p>Therefore, you will probably want to use a <code>tensorflow::ClientSession<\/code> if you are building the graph with the C++ API, but the <code>tensorflow::Session<\/code> interface is easier to use if you already have a serialized <code>GraphDef<\/code> (representing e.g. a pre-trained model) and just want to run inference on that model.<\/p>\n"}]}
{"QuestionId":42603926,"AnswerCount":0,"Tags":"<tensorflow><tf-slim>","CreationDate":"2017-03-05T02:51:39.203","AcceptedAnswerId":null,"OwnerUserId":2134887.0,"Title":"Training Inception V3 using tensorflow slim on multiple machines with CPUs only","Body":"<p>I'm trying to use the tf-slim library to train Inception V3 using multiple machines which do not have GPUs.<\/p>\n\n<p>I'm following the tutorial given here: <a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/slim<\/a>. I can train using a single CPU only machine but I'm trying to figure out how to do the training using multiple machines. Its mentioned in the page that:<\/p>\n\n<blockquote>\n  <p>This process may take several days, depending on your hardware setup.\n  For convenience, we provide a way to train a model on multiple GPUs,\n  and\/or multiple CPUs, either synchrononously or asynchronously. See\n  <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/deployment\/model_deploy.py\" rel=\"nofollow noreferrer\">model_deploy<\/a> for details.<\/p>\n<\/blockquote>\n\n<p>I can't figure out how to use <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/deployment\/model_deploy.py\" rel=\"nofollow noreferrer\">model_deploy<\/a> to train using multiple machines and neither can I find any examples to do so.<\/p>\n\n<p>How do I go about using <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/slim\/deployment\/model_deploy.py\" rel=\"nofollow noreferrer\">model_deploy<\/a> to train using multiple machines?<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[]}
{"QuestionId":42604075,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><backpropagation><dot-product>","CreationDate":"2017-03-05T03:12:23.783","AcceptedAnswerId":null,"OwnerUserId":1936956.0,"Title":"dot product between two branches of a cnn in tensorflow","Body":"<p>I'm trying to use tensorflow to implement a cnn that, when given two images, it can find in which position both images are the most similar. Something to whats described in here: <a href=\"http:\/\/www.cs.toronto.edu\/~urtasun\/publications\/luo_etal_cvpr16.pdf\" rel=\"nofollow noreferrer\">Efficient Deep Learning for Stereo Matching<\/a><\/p>\n\n<p>I'm sharing the variables of both branches of the cnn and after that I want to do a dot product of the activations of the left image with all the available positions in the image of the right. \nMy question is: if I do this process manually (in a for loop, for example), can tensorflow still backpropagate the gradients through the filters? I'm getting a very unstable and weird loss values and I'm not sure if i'm doing something wrong.<\/p>\n\n<p>Thanks for any help in advance.<\/p>\n\n<p>EDIT:\nHere is some code how i initialize the cnn and how I calculate the dot product. I'm not posting the conv layer initialization code for simplicity. I think the problem is in the dot_product function anyway. <\/p>\n\n<pre><\/pre>\n\n<p>EDIT2: I'm also adding the graph:\n<a href=\"https:\/\/i.stack.imgur.com\/H2u1W.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/H2u1W.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[]}
{"QuestionId":42604535,"AnswerCount":1,"Tags":"<tensorflow><pip><anaconda>","CreationDate":"2017-03-05T04:29:58.973","AcceptedAnswerId":null,"OwnerUserId":1001930.0,"Title":"Anaconda: Tensorflow installs to .local when installed with pip","Body":"<p>Whenever I install Tensorflow with pip from within an Anaconda environment, it seems to be saved to the  directory, rather than the corresponding  directory. For example<\/p>\n\n<pre><\/pre>\n\n<p>I'd expect the last line to say something like\n\nrather than pointing to some absolute directory.<\/p>\n","answers":[{"AnswerId":"42604536","CreationDate":"2017-03-05T04:29:58.973","ParentId":null,"OwnerUserId":"1001930","Title":null,"Body":"<p>Although not 100% satisfactory, my work around (based off this <a href=\"https:\/\/github.com\/conda\/conda\/issues\/448\" rel=\"nofollow noreferrer\">git issue thread<\/a>) is to make sure the environment variable <code>PYTHONNOUSERSITE<\/code> is set to <code>True<\/code> before installing tensorflow with pip. So, the new code is:<\/p>\n\n<pre><code>$ export PYTHONNOUSERSITE=True   # &lt;--- DO THIS FIRST\n$ source activate foo\n(foo) $ pip3 install tensorflow-gpu\n(foo) $ python\n&gt;&gt;&gt; import tensorflow as tf\n&gt;&gt;&gt; tf.__file__\n$HOME\/anaconda2\/envs\/foo\/lib\/python3.5\/site-packages\/tensorflow\/__init__.py\n<\/code><\/pre>\n"}]}
{"QuestionId":42605461,"AnswerCount":1,"Tags":"<neural-network><deep-learning><crop><caffe><imagenet>","CreationDate":"2017-03-05T06:57:07.037","AcceptedAnswerId":null,"OwnerUserId":2103732.0,"Title":"Caffe | data augmentation by random cropping","Body":"<p>I am trying to train my own network on Caffe, similar to Imagenet model. But I am confused with the crop layer. Till the point I understand about crop layer in Imagenet model, during training it will take random 227x227 image crops and train the network. But during testing it will take the center 227x227 image crop, does not we loose the information from image while we crop the center 227x27 image from 256x256 image? And second question, how can we define the number of crops to be taken during training? <\/p>\n\n<p>And also, I trained the same network(same number of layers, same convolution size FC neurons will differ obviously), first taking 227x227 crop from 256x256 image, and second time taking 255x255 crop from 256x256 image. According to my intuition, the model with 255x255 crop should give me the best result. But I am getting higher accuracy with 227x227 image, can anyone explain me the intuition behind it, or am i doing something wrong? <\/p>\n","answers":[{"AnswerId":"45573284","CreationDate":"2017-08-08T16:13:42.493","ParentId":null,"OwnerUserId":"8423075","Title":null,"Body":"<p>Your observations are not specific to Caffe.  <\/p>\n\n<p>The sizes of the cropped images during training and testing need to be the same (227x227 in your case), because the upstream network layers (convolutions, etc) need the images to be the <strong>same size<\/strong>. Random crops are done during training is because you want data augmentation. However, during testing, you want to test against a <strong>standard<\/strong> dataset. Otherwise, the accuracy reported during testing would also depend on a shifting test database.<\/p>\n\n<p>The crops are made dynamically at each iteration. All images in a training batch are randomly cropped. I hope this answers your second question. <\/p>\n\n<p>Your intuition is not complete: With a bigger crop (227x227), you have more data augmentation. Data augmentation essentially creates \"new\" training samples out of nothing. This is vital to prevent overfitting during training. With a smaller crop (255x255), you should expect a better training accuracy but lower test accuracy, since the data is more likely be overfitted.<\/p>\n\n<p>Of course, cropping can be overdone. Too much cropping and you lose too much information from an image. For image categorization, the ideal crop size is one that does not alter the category of an image, (ie, only background is cropped away). <\/p>\n"}]}
{"QuestionId":42605479,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe><conv-neural-network><pycaffe>","CreationDate":"2017-03-05T06:59:07.197","AcceptedAnswerId":47626127.0,"OwnerUserId":7533175.0,"Title":"Efficient hard data sampling for triplet loss","Body":"<p>I am trying to implement a deep network for triplet loss in Caffe.\nWhen I select three samples for anchor, positive, negative images randomly, it almost produces zero losses. So I tried the following strategy:<\/p>\n\n<pre><\/pre>\n\n<p>The step 1 is fast, but I think step 2 is very time-consuming and is really inefficient. So, I wonder whether there are other efficient strategies for hard data sampling.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"47626127","CreationDate":"2017-12-04T03:34:35.260","ParentId":null,"OwnerUserId":"6064933","Title":null,"Body":"<p>In practice, if your dataset is large, it is infeasible to sample hard triplet from the whole dataset. In fact, you can choose hard triplet for only a small proportion of your training dataset, which will be much more time-saving. After training the network using the hard triplet generated for K iterations. You feed the network with the next batch of images from the dataset and generate the new hard triplet.<\/p>\n\n<p>In this way, the computation cost is acceptable and network is gradually improving as the training process goes.<\/p>\n\n<p>see the <a href=\"https:\/\/arxiv.org\/pdf\/1604.01325.pdf\" rel=\"nofollow noreferrer\">article here<\/a> for more reference.(section 5.1)<\/p>\n"}]}
{"QuestionId":42605766,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-03-05T07:37:13.127","AcceptedAnswerId":null,"OwnerUserId":7660780.0,"Title":"Using tensorflow, can I back-propagate the already calculated gradients to all the parameters, and then apply the gradients to the whole network?","Body":"<p>I can get the gradients w.r.t the outputs of the network(calculated by other ways), Can I just back-propagate the gradients to all the parameters, and then apply them to the whole network?<\/p>\n","answers":[{"AnswerId":"42667373","CreationDate":"2017-03-08T09:22:08.243","ParentId":null,"OwnerUserId":"4834515","Title":null,"Body":"<p>You can check this <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/resnet\/resnet_model.py#L128\" rel=\"nofollow noreferrer\">link<\/a>.<\/p>\n\n<pre><code>trainable_variables = tf.trainable_variables()\ngrads = tf.gradients(self.cost, trainable_variables)\n\nif self.hps.optimizer == 'sgd':\n  optimizer = tf.train.GradientDescentOptimizer(self.lrn_rate)\nelif self.hps.optimizer == 'mom':\n  optimizer = tf.train.MomentumOptimizer(self.lrn_rate, 0.9)\n\napply_op = optimizer.apply_gradients(\n    zip(grads, trainable_variables),\n    global_step=self.global_step, name='train_step')\n<\/code><\/pre>\n\n<p>where you can replace the <code>gradients<\/code> if the gradients you calculated are also <code>tensors<\/code> or <code>variables<\/code>. <\/p>\n\n<p>But if the gradients you calculated are numerical, I am not sure that tensorflow can easily do this.<\/p>\n"}]}
{"QuestionId":42606207,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><classification><keras>","CreationDate":"2017-03-05T08:35:47.320","AcceptedAnswerId":42607110.0,"OwnerUserId":1928515.0,"Title":"Keras custom decision threshold for precision and recall","Body":"<p>I'm doing a binary classification using  (with  backend) and I've got about 76% precision and 70% recall. Now I want to try to play with decision threshold. As far as I know  uses decision threshold 0.5. Is there a way in  to use custom threshold for decision precision and recall?<\/p>\n\n<p>Thank you for your time!<\/p>\n","answers":[{"AnswerId":"42607110","CreationDate":"2017-03-05T10:18:09.303","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>create custom metrics like this :<\/p>\n\n<p><strong>Edited thanks to @Marcin<\/strong> : Create functions that returns the desired metrics with <code>threshold_value<\/code> as argument<\/p>\n\n<pre><code>def precision_threshold(threshold=0.5):\n    def precision(y_true, y_pred):\n        \"\"\"Precision metric.\n        Computes the precision over the whole batch using threshold_value.\n        \"\"\"\n        threshold_value = threshold\n        # Adaptation of the \"round()\" used before to get the predictions. Clipping to make sure that the predicted raw values are between 0 and 1.\n        y_pred = K.cast(K.greater(K.clip(y_pred, 0, 1), threshold_value), K.floatx())\n        # Compute the number of true positives. Rounding in prevention to make sure we have an integer.\n        true_positives = K.round(K.sum(K.clip(y_true * y_pred, 0, 1)))\n        # count the predicted positives\n        predicted_positives = K.sum(y_pred)\n        # Get the precision ratio\n        precision_ratio = true_positives \/ (predicted_positives + K.epsilon())\n        return precision_ratio\n    return precision\n\ndef recall_threshold(threshold = 0.5):\n    def recall(y_true, y_pred):\n        \"\"\"Recall metric.\n        Computes the recall over the whole batch using threshold_value.\n        \"\"\"\n        threshold_value = threshold\n        # Adaptation of the \"round()\" used before to get the predictions. Clipping to make sure that the predicted raw values are between 0 and 1.\n        y_pred = K.cast(K.greater(K.clip(y_pred, 0, 1), threshold_value), K.floatx())\n        # Compute the number of true positives. Rounding in prevention to make sure we have an integer.\n        true_positives = K.round(K.sum(K.clip(y_true * y_pred, 0, 1)))\n        # Compute the number of positive targets.\n        possible_positives = K.sum(K.clip(y_true, 0, 1))\n        recall_ratio = true_positives \/ (possible_positives + K.epsilon())\n        return recall_ratio\n    return recall\n<\/code><\/pre>\n\n<p>now you can use them in<\/p>\n\n<pre><code>model.compile(..., metrics = [precision_threshold(0.1), precision_threshold(0.2),precision_threshold(0.8), recall_threshold(0.2,...)])\n<\/code><\/pre>\n\n<p>I hope this helps :)<\/p>\n"}]}
{"QuestionId":42606722,"AnswerCount":1,"Tags":"<input><tensorflow><conv-neural-network>","CreationDate":"2017-03-05T09:38:27.503","AcceptedAnswerId":null,"OwnerUserId":4478970.0,"Title":"Shape of placeholder in Tensorflow","Body":"<p>I'm using Tensorflow for a short time. Here is my problem : \nI load AlexNet weights to do fine tuning on it, so i give batch of size 50.\nSo i defined :<\/p>\n\n<pre><\/pre>\n\n<p>I give a batch of 50 images and want to get 40 output classes.<\/p>\n\n<p>Then I defined my model <\/p>\n\n<pre><\/pre>\n\n<p>and create it <\/p>\n\n<pre><\/pre>\n\n<p>I can do the training, it works well, but at the end, i want to give only one image, but the x placeholder and y placeholder are defined for 50 images so it raises an error.\nHere is my code after the training to give only one image : <\/p>\n\n<pre><\/pre>\n\n<p>And it raises me this error : <\/p>\n\n<pre><\/pre>\n\n<p>I can't figured out how to feed the input size I want.<\/p>\n\n<p>My exercise is directly inspired from the flower recognition with cnn <\/p>\n\n<p>Thanks a lot for your help ! \nGuillaume<\/p>\n","answers":[{"AnswerId":"42606780","CreationDate":"2017-03-05T09:45:26.367","ParentId":null,"OwnerUserId":"3734670","Title":null,"Body":"<p>Instead of setting the first dimension of the shape to a fixed size, you can use for the first dimension of the shape a variable size by setting None instead of a number. Tensorflow is able to calculate the batch size by the input size and the fixed sizes of the other dimensions of the shape. <\/p>\n\n<p>For the placeholder y, you have made that correct:<\/p>\n\n<pre><code>y = tf.placeholder(tf.float32, [None, 40])\n<\/code><\/pre>\n\n<p>For the placeholder x, you have to set:<\/p>\n\n<pre><code>x = tf.placeholder(tf.float32, [None, 227, 227, 3])\n<\/code><\/pre>\n"}]}
{"QuestionId":42606961,"AnswerCount":2,"Tags":"<neural-network><nlp><keras><lstm>","CreationDate":"2017-03-05T10:03:51.530","AcceptedAnswerId":42629823.0,"OwnerUserId":7661085.0,"Title":"Can I use next layer's output as current layer's input by Keras?","Body":"<p>In text generate mission, we usually use model's last output as current input to generate next word. More generalized, I want to achieve a neural network  that regards next layer's finally hidden state as current layer's input. Just like the following(what confuses me is the decoder part):<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/ED0aA.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/ED0aA.png\" alt=\"encoder-decoder\"><\/a><\/p>\n\n<p>But I have read Keras document and haven't found any functions to achieve it.<\/p>\n\n<p>Can I achieve this structure by Keras? How?<\/p>\n","answers":[{"AnswerId":"42608719","CreationDate":"2017-03-05T13:09:12.147","ParentId":null,"OwnerUserId":"5352399","Title":null,"Body":"<p>The structure you have showed is a custom structure. So, Keras doesn't provide any class or wrapper to directly build such structure. But <strong>YES<\/strong>, you can build this kind of structure in Keras.<\/p>\n\n<p>So, it looks like you need LSTM model in backward direction. I didn't understand the other part which probably looks like incorporating previous sentence embedding as input to the next time-step input of LSTM unit. <\/p>\n\n<p>I rather encourage you to work with simple language-modeling with LSTM first. Then you can tweak the architecture later to build an architecture depicted in figure.<\/p>\n\n<p>Example:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/lstm_text_generation.py\" rel=\"noreferrer\">Text generation with LSTM in Keras<\/a><\/li>\n<\/ul>\n"},{"AnswerId":"42629823","CreationDate":"2017-03-06T15:47:00.097","ParentId":null,"OwnerUserId":"2991872","Title":null,"Body":"<p>What you are asking is an autoencoders, you can find similar structures in Keras. <\/p>\n\n<p>But there are certain details that you should figure it out on your own. Including the padding strategy and preprocessing your input and output data. Your input cannot get dynamic input size, so you need to have a fixed length for input and outputs. I don't know what do you mean by arrows who join in one circle but I guess you can take a look at <a href=\"https:\/\/keras.io\/getting-started\/sequential-model-guide\/\" rel=\"nofollow noreferrer\">Merge<\/a> layer in Keras (basically adding, concatenating, and etc.)<\/p>\n\n<p>You probably need 4 <a href=\"https:\/\/keras.io\/models\/sequential\/\" rel=\"nofollow noreferrer\">sequential<\/a> model and one final <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">model<\/a> that represent the combined structure.<\/p>\n\n<p>One more thing, the decoder setup of LSTM (The Language Model) is not dynamic in design. In your model definition, you basically introduce a fixed inputs and outputs for it. Then you prepare the training correctly, so you don't need anything dynamic. Then during the test, you can predict each decoded word in a loop by running the model once predict the next output step and run it again for next time step and so on.  <\/p>\n"}]}
{"QuestionId":42607930,"AnswerCount":1,"Tags":"<tensorflow><classification><convolution>","CreationDate":"2017-03-05T11:38:59.057","AcceptedAnswerId":42608050.0,"OwnerUserId":6631923.0,"Title":"How to compute accuracy of CNN in TensorFlow","Body":"<p>I am new to TensorFlow. I am doing a binary classification with my own dataset. However I do not know how to compute the accuracy. Can anyone please help me with to do this? <\/p>\n\n<p>My classifier has 5 convolutional layers followed by 2 fully connected layers. The final FC layer has an output dimension of 2 for which I have used:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42608050","CreationDate":"2017-03-05T11:52:04.037","ParentId":null,"OwnerUserId":"4969370","Title":null,"Body":"<p>Just calculate the percentage of correct predictions:<\/p>\n\n<pre><code>prediction = tf.argmax(prob, 1)\nequality = tf.equal(prediction, correct_answer)\naccuracy = tf.reduce_mean(tf.cast(equality, tf.float32))\n<\/code><\/pre>\n"}]}
{"QuestionId":42608175,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-05T12:07:13.340","AcceptedAnswerId":null,"OwnerUserId":5098762.0,"Title":"What does tf.gather_nd intuitively do?","Body":"<p>Can you intuitively explain or give more examples about  for indexing and slicing into high-dimensional tensors in Tensorflow? <\/p>\n\n<p>I read the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"noreferrer\">API<\/a>, but it is kept quite concise that I find myself hard to follow the function's concept.<\/p>\n","answers":[{"AnswerId":"55055527","CreationDate":"2019-03-08T01:40:17.587","ParentId":null,"OwnerUserId":"10112380","Title":null,"Body":"<p>You provide a tensor and indices representing locations in that tensor. It returns the elements of the tensor corresponding to the indices you provide. <\/p>\n\n<p>EDIT: An example<\/p>\n\n<pre><code>import tensorflow as tf\nsess = tf.Session()\nx = [[1,2,3],[4,5,6]]\ny = tf.gather_nd(x, [[1,1],[1,2]])\nprint(sess.run(y))\n<\/code><\/pre>\n\n<blockquote>\n  <p>[5, 6]<\/p>\n<\/blockquote>\n"},{"AnswerId":"42608263","CreationDate":"2017-03-05T12:18:56.397","ParentId":null,"OwnerUserId":"7659637","Title":null,"Body":"<p>Ok, so think about it like this:<\/p>\n\n<p>You are providing a list of index values to index the provided tensor to get those slices. The first dimension of the indices you provide is for each index you will perform. Let's pretend that tensor is just a list of lists.<\/p>\n\n<p>[[0]] means you want to get one specific slice(list) at index 0 in the provided tensor. Just like this:<\/p>\n\n<pre><code>[tensor[0]]\n<\/code><\/pre>\n\n<p>[[0], [1]] means you want get two specific slices at indices 0 and 1 like this:<\/p>\n\n<pre><code>[tensor[0], tensor[1]]\n<\/code><\/pre>\n\n<p>Now what if tensor is more than one dimensions? We do the same thing: <\/p>\n\n<p>[[0, 0]] means you want to get one slice at index [0,0] of the 0-th list. Like this:<\/p>\n\n<pre><code>[tensor[0][0]]\n<\/code><\/pre>\n\n<p>[[0, 1], [2, 3]] means you want return two slices at the indices and dimensions provided. Like this: <\/p>\n\n<pre><code>[tensor[0][1], tensor[2][3]]\n<\/code><\/pre>\n\n<p>I hope that makes sense. I tried using Python indexing to help explain how it would look in Python to do this to a list of lists.<\/p>\n"}]}
{"QuestionId":42608245,"AnswerCount":0,"Tags":"<python-2.7><tensorflow>","CreationDate":"2017-03-05T12:15:27.767","AcceptedAnswerId":null,"OwnerUserId":7661539.0,"Title":"How to use tf.contrib.seq2seq.simple_decoder_fn_inference API","Body":"<p>I did not understand the parameters that needed to be passed into the API call to    in Tensorflow 1.0\nfor building the inference block for a Seq2Seq Attention mechanism RNN.<\/p>\n\n<p>Can someone explain, in detail, what each parameter of this function call means and is supposed to do?<\/p>\n\n<p>The link to the documentation is here :\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/seq2seq\/attention_decoder_fn_inference\" rel=\"nofollow noreferrer\">tf.contrib.seq2seq.attention_decoder_fn_inference()<\/a><\/p>\n","answers":[]}
{"QuestionId":42609109,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-03-05T13:47:51.130","AcceptedAnswerId":null,"OwnerUserId":2359772.0,"Title":"Tensorflow RNN tutorial","Body":"<p>I am following <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\" rel=\"nofollow noreferrer\">RNN tutorial of Tensorflow<\/a>. \nI am having trouble understanding the  in  in following script : <\/p>\n\n<pre><\/pre>\n\n<p>Can anyone explain what  is doing ? <\/p>\n","answers":[{"AnswerId":"45772194","CreationDate":"2017-08-19T13:35:44.320","ParentId":null,"OwnerUserId":null,"Title":null,"Body":"<p>I have been trying to understand the same tutorial for weeks now. In my opinion, what makes it so difficult is the fact that all the functions one calls from TensorFlow are not executed immediately, but rather add their corresponding operation nodes to the graph.<\/p>\n\n<p>According to the <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.2\/api_docs\/python\/tf\/train\/range_input_producer\" rel=\"nofollow noreferrer\">official documentation<\/a>, a Range Input Producer 'generates integers from <code>0<\/code> to <code>limit - 1<\/code> in a queue'. So, the way I see it, the code in question <code>i = tf.train.range_input_producer(epoch_size, shuffle=False).dequeue()<\/code> creates a node, which acts as a counter, producing the next number in the sequence <code>0:(epoch_size)<\/code> once executed.<\/p>\n\n<p>This is used to get the next batch from the input data. The raw data is split into <code>batch_size<\/code> rows, so that in every run <code>batch_size<\/code> batches are given to the training function. In every batch (row), a sliding window of size <code>num_steps<\/code> moves forward. The counter <code>i<\/code> allows the window to move forward by <code>num_steps<\/code> in every call.<\/p>\n\n<p>Both <code>x<\/code> and <code>y<\/code> are of shape <code>[batch_size, num_steps]<\/code>, since they contain <code>batch_size<\/code> batches of <code>num_steps<\/code> steps each. Variable <code>x<\/code> is the input and <code>y<\/code> is the expected output for the given input (it is produced by moving the window one item to the left, so that <code>iff x = data[i:(i + num_steps] then y = data[(i + 1):(i + num_steps + 1)]<\/code>.<\/p>\n\n<p>It has been a nightmare for me, but I hope this post helps people in the future.<\/p>\n"}]}
{"QuestionId":42609264,"AnswerCount":1,"Tags":"<c++><neural-network><deep-learning><caffe>","CreationDate":"2017-03-05T14:02:11.167","AcceptedAnswerId":null,"OwnerUserId":7661676.0,"Title":"Caffe ImageData neural network basic example fails to parse the model file","Body":"<p>I am trying to build a minimal example of a neural network with <strong><em>IMAGE DATA<\/em><\/strong> that I have prepared from a CSV file using the caffe libraries.<\/p>\n\n<p>My prototext is as follows: [wine_train.prototxt]<\/p>\n\n<pre><\/pre>\n\n<p>and my solver is as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Each time I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>What exactly does the error say and how do I resolve it?<\/p>\n","answers":[{"AnswerId":"42619246","CreationDate":"2017-03-06T06:51:37.663","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Your <code>'wine_train.prototxt'<\/code> has a problematic format. It mixes old format (i.e., <code>layers<\/code> instead of <code>layer<\/code>, layer types in all capitals) and new format (i.e., layer types as string and not as enumerations).<br>\nRemove the quotation marks (<code>\"<\/code>) from the layer types and run <code>$CAFFE_ROOT\/build\/tools\/upgrade_net_proto_text.bin<\/code>to get an upgraded net prototxt.<\/p>\n\n<p>Additionally, you do not need (and probably should not) provide <code>-model<\/code> argument when running <code>caffe train<\/code>. This argument is used when running <code>caffe test<\/code>. The model file is defined in the <code>solver.prototxt<\/code>.<\/p>\n"}]}
{"QuestionId":42609298,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-05T14:04:46.840","AcceptedAnswerId":null,"OwnerUserId":6412036.0,"Title":"Tensorflow, using trained network","Body":"<p>I'm new to tensor flow, and I'm studying cifar10 example provided in tensorflow website.<\/p>\n\n<p>I can follow the provided code to train a network and evaluate it, which says the trained network has the accuracy of ~84%.<\/p>\n\n<p>However, when I feed images from cifar10 website to output the prediction, the accuracy is no way near the promised 84%. \n(I downloaded a few sample pics from <a href=\"https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html\" rel=\"nofollow noreferrer\">https:\/\/www.cs.toronto.edu\/~kriz\/cifar.html<\/a> for testing)<\/p>\n\n<p>During the training, I'm saving my weights and biases using:<\/p>\n\n<pre><\/pre>\n\n<p>Then, once the training is complete, I load them by:<\/p>\n\n<pre><\/pre>\n\n<p>once all weights and biases are loaded, I can compute the output through convolution, pooling, and so on.<\/p>\n\n<p>Would please someone can point out if there's anything wrong in this scheme?<\/p>\n\n<p>Any help would be really appreciated<\/p>\n\n<p>Regards,<\/p>\n","answers":[]}
{"QuestionId":42609369,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-03-05T14:10:00.210","AcceptedAnswerId":null,"OwnerUserId":3051460.0,"Title":"Setting for BatchNorm layer in Caffe?","Body":"<p>I am implementing the <a href=\"https:\/\/arxiv.org\/abs\/1603.05027\" rel=\"nofollow noreferrer\">Identity Mappings in Deep Residual Networks<\/a>. <\/p>\n\n<pre><\/pre>\n\n<p>As I known, the BN often is followed by Scale layer and used  to save memory. I am not using current caffe version, I used 3D UNet caffe, hence, I follow the setting in the <a href=\"http:\/\/lmb.informatik.uni-freiburg.de\/resources\/opensource\/3dUnet_miccai2016_with_BN.prototxt\" rel=\"nofollow noreferrer\">website<\/a>. I found that,  for TRAIN and  for validation and deploying. However, when I used   in deploying, the result is very poor (~30% accuracy). If the parameter is set to  and the result becomes good (80% accuracy). Can I set   in deploying? I am using volumetric data which has size , and I need to crop it into  to feed to network dueto limited GPU. Thanks<\/p>\n\n<p>This is my current BatchNorm +Scale layers<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43763878","CreationDate":"2017-05-03T15:09:35.260","ParentId":null,"OwnerUserId":"7958372","Title":null,"Body":"<p>By default, as the following is set in batch_norm_layer.cpp, you don't have to set <code>use_global_stats<\/code> at all in the prototxt.<\/p>\n\n<p>In batch_norm_layer.cpp:<\/p>\n\n<pre><code>use_global_stats_ = this-&gt;phase_ == TEST;\n<\/code><\/pre>\n"}]}
{"QuestionId":42609467,"AnswerCount":1,"Tags":"<tensorflow><classification>","CreationDate":"2017-03-05T14:18:50.067","AcceptedAnswerId":42613787.0,"OwnerUserId":192936.0,"Title":"Binary classification with TensorFlow for weighted_cross_entropy_with_logits","Body":"<p>I am trying to train a binary classification model want to give more weight to positive class. As far as I understand,  is suitable for this, but I'm not sure and couldn't find an official explanation on how to define classes and what's considered positive in this context.<\/p>\n\n<p>i.e. Do I need two output nodes and second node is implicitly positive class, or do I need a single node and its higher values are considered as positive class?<\/p>\n","answers":[{"AnswerId":"42613787","CreationDate":"2017-03-05T20:47:41.720","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<blockquote>\n  <p>As far as I understand tf.weighted_cross_entropy_with_logits is suitable for this <\/p>\n<\/blockquote>\n\n<p>yes<\/p>\n\n<blockquote>\n  <p>but I'm not sure and couldn't find an official explanation on how to define classes and what's considered positive in this context?<\/p>\n<\/blockquote>\n\n<p>in your <code>targets<\/code>, positive is <code>1<\/code>, negative is <code>0<\/code>. You can infer that from reading the equations in the <a href=\"https:\/\/web.archive.org\/web\/20170305204449\/https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/nn\/classification\" rel=\"nofollow noreferrer\">documentation<\/a>: <\/p>\n\n<pre><code>targets * -log(sigmoid(logits)) * pos_weight + (1 - targets) * -log(1 - sigmoid(logits))\n<\/code><\/pre>\n\n<blockquote>\n  <p>i.e. Do I need two output nodes and second node is implicitly positive class, or do I need a single node and its higher values are considered as positive class?<\/p>\n<\/blockquote>\n\n<p>You only need a single node, which outputs the probability that your input belongs to the positive class.<\/p>\n"}]}
{"QuestionId":42609618,"AnswerCount":2,"Tags":"<python><tensorflow>","CreationDate":"2017-03-05T14:32:47.380","AcceptedAnswerId":42612608.0,"OwnerUserId":2476373.0,"Title":"Tensorflow equivalent to numpy.diff","Body":"<p>Is there a tensorflow equivalent to <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.diff.html\" rel=\"noreferrer\">numpy.diff<\/a>?<\/p>\n\n<blockquote>\n  <p>Calculate the n-th discrete difference along given axis.<\/p>\n<\/blockquote>\n\n<p>For my project I only need n=1<\/p>\n","answers":[{"AnswerId":"42612330","CreationDate":"2017-03-05T18:34:58.640","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>I don't think TensorFlow has an equivalent to numpy.diff, so you'll have to implement it, which shouldn't difficult as <a href=\"https:\/\/github.com\/numpy\/numpy\/blob\/v1.12.0\/numpy\/lib\/function_base.py#L1709-L1768\" rel=\"nofollow noreferrer\">numpy.diff<\/a> simply slices and subtractes:<\/p>\n\n<pre><code>def diff(a, n=1, axis=-1):\n    '''(as implemented in NumPy v1.12.0)'''\n    if n == 0:\n        return a\n    if n &lt; 0:\n        raise ValueError(\n            \"order must be non-negative but got \" + repr(n))\n    a = asanyarray(a)\n    nd = len(a.shape)\n    slice1 = [slice(None)]*nd\n    slice2 = [slice(None)]*nd\n    slice1[axis] = slice(1, None)\n    slice2[axis] = slice(None, -1)\n    slice1 = tuple(slice1)\n    slice2 = tuple(slice2)\n    if n &gt; 1:\n        return diff(a[slice1]-a[slice2], n-1, axis=axis)\n    else:\n        return a[slice1]-a[slice2]\n<\/code><\/pre>\n"},{"AnswerId":"42612608","CreationDate":"2017-03-05T18:57:39.780","ParentId":null,"OwnerUserId":"419116","Title":null,"Body":"<p>Try this:<\/p>\n\n<pre><code>def tf_diff_axis_0(a):\n    return a[1:]-a[:-1]\n\ndef tf_diff_axis_1(a):\n    return a[:,1:]-a[:,:-1]\n<\/code><\/pre>\n\n<p>To check:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\nx0=np.arange(5)+np.zeros((5,5))\nsess = tf.Session()\nnp.diff(x0, axis=0) == sess.run(tf_diff_axis_0(tf.constant(x0)))\nnp.diff(x0, axis=1) == sess.run(tf_diff_axis_1(tf.constant(x0)))\n<\/code><\/pre>\n"}]}
{"QuestionId":42610050,"AnswerCount":1,"Tags":"<deep-learning><caffe><pycaffe><matcaffe>","CreationDate":"2017-03-05T15:15:11.267","AcceptedAnswerId":42619527.0,"OwnerUserId":6494707.0,"Title":"After adding Xavier and bias_filler, the loss values starts becoming negative. Why?","Body":"<p>After I added  initialization to every convolution layer, the loss starts becoming <strong>negative<\/strong>. Could someone give any suggestion\/reason?\nI added the following lines to all convolutional layers:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<pre><\/pre>\n\n<p>My another question is that in some networks,  is added. Like:<\/p>\n\n<pre><\/pre>\n\n<ol>\n<li><p>Why are we adding these parameters to convolutional layer? Is it\nbecause we are training the network from the scratch?<\/p><\/li>\n<li><p>How are a specific value assigned to  and\/or the \nvalue?<\/p><\/li>\n<\/ol>\n\n<p>I really appreciate your help.<\/p>\n","answers":[{"AnswerId":"42619527","CreationDate":"2017-03-06T07:09:46.333","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<ul>\n<li><p>Your loss is <code>-8.27284e-09<\/code> which practically speaking is zero and not negative (caffe is using a single precision floating point numbers and not double precision).<br>\nWhat loss layer are you using? <code>\"SoftmaxWithLoss\"<\/code>?<\/p><\/li>\n<li><p><code>bias_filler<\/code> and <code>wieght_filler<\/code> parameters are added when we want caffe to <strong>randomly<\/strong> initialize the weights of the layer, usually when we start training from scratch. If you start training from existing model (i.e. fine tuning) there is no meaning for these arguments.<\/p><\/li>\n<li><p><code>std<\/code> value is computed based on the fan-in and fan-out (i.e., number of in-channels and out channels) in order to keep the statistics of the Blob values roughly zero-mean and unit variance.<br>\nYou can find an analysis of these parameters in <a href=\"https:\/\/arxiv.org\/abs\/1502.01852\" rel=\"nofollow noreferrer\"><em>Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun<\/em> <strong>Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification<\/strong> (arXiv 2015)<\/a>.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":42610640,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-03-05T16:06:34.823","AcceptedAnswerId":42660947.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Changing the maximum number of checkpoints to keep when training using a Supervisor?","Body":"<p>There is a  property in a saver that allows one to choose how many checkpoints to keep. When I create a default saver with  as two and pass this saver to the supervisor, somehow the default number of 5 continues to appear. Is there a way to tell the supervisor to limit the number of checkpoints to 2?<\/p>\n\n<p>The arguments passed on to the Supervisor has no such ability, I believe, according to the documentation here:\n<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Supervisor\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/Supervisor<\/a><\/p>\n","answers":[{"AnswerId":"42660947","CreationDate":"2017-03-08T00:38:44.850","ParentId":null,"OwnerUserId":"4834515","Title":null,"Body":"<p>The Supervisor is a small wrapper around a Coordinator, a Saver, and a SessionManager. Try to use them separately, not using a wrapper.<\/p>\n"}]}
{"QuestionId":42611083,"AnswerCount":0,"Tags":"<deep-learning><keras><word2vec>","CreationDate":"2017-03-05T16:45:55.797","AcceptedAnswerId":null,"OwnerUserId":2226696.0,"Title":"How to implement CBOW + hierachical softmax in Keras","Body":"<p>I would like to implement CBOW + HS in Keras.<\/p>\n\n<p>I found an example (<a href=\"https:\/\/github.com\/nzw0301\/keras-examples\/blob\/master\/CBoW.ipynb\" rel=\"noreferrer\">https:\/\/github.com\/nzw0301\/keras-examples\/blob\/master\/CBoW.ipynb<\/a>), but its activation function seems like softmax not hierarchical softmax.<\/p>\n\n<p>Is there any way to implement CBOW + HS in Keras?<\/p>\n","answers":[]}
{"QuestionId":42611316,"AnswerCount":3,"Tags":"<python><deep-learning><keras><keras-layer>","CreationDate":"2017-03-05T17:06:03.307","AcceptedAnswerId":42613508.0,"OwnerUserId":1575548.0,"Title":"Keras - Pop and re-add layers, but layer does not disconnect","Body":"<p>Using Keras (1.2.2), I am loading a sequential model whose last layers are:<\/p>\n\n<pre><\/pre>\n\n<p>Then, I want to pop the last layer, add another fully connected layer, and re-add the classification layer.<\/p>\n\n<pre><\/pre>\n\n<p>As you can see my dense_3 and activation_7 did not connect to the network (Empty value in summary() with \"Connected to\"). I cannot find anything in the documentation that explains how to solve this problem. Any ideas?<\/p>\n\n<pre><\/pre>\n\n<p>Following the answer below, I compiled the model before printing out , but for some reasons, the layers are not being popped correctly, as the summary shows: The last layer's connections are wrong:<\/p>\n\n<pre><\/pre>\n\n<p>But it should be<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"49403298","CreationDate":"2018-03-21T10:03:35.863","ParentId":null,"OwnerUserId":"1179925","Title":null,"Body":"<p>For some reason I need to build model with popped layer using Model before adding new layers to make things work.<\/p>\n\n<pre><code>conda list keras\n# Name                    Version                   Build  Channel\nkeras                     2.1.5                    py36_0    conda-forge\n<\/code><\/pre>\n\n<p>Here is code snippet:<\/p>\n\n<pre><code>def pop_layer(model):\n    if not model.outputs:\n        raise Exception('Sequential model cannot be popped: model is empty.')\n\n    model.layers.pop()\n    if not model.layers:\n        model.outputs = []\n        model.inbound_nodes = []\n        model.outbound_nodes = []\n    else:\n        model.layers[-1].outbound_nodes = []\n        model.outputs = [model.layers[-1].output]\n    model.built = False\n\ndef get_model():\n    #Fully convolutional part of VGG16\n    model = VGG16(include_top=False, weights='imagenet')\n\n    #Remove last max pooling layer\n    pop_layer(model)\n\n    #Freeze pretrained layers\n    for layer in model.layers:\n        layer.trainable = False\n\n    model = Model(inputs=model.inputs, outputs=model.outputs)\n\n    print('len(model.layers)', len(model.layers)) #\n    print(model.summary()) #\n\n    x = GlobalAveragePooling2D()(model.output)\n    head = Dense(N_CLASS, activation='softmax')(x)\n\n    model = Model(inputs=model.inputs, outputs=head)\n\n    model.compile(optimizer=Adadelta(), loss='categorical_crossentropy', metrics=['accuracy'])\n\n    print('len(model.layers)', len(model.layers)) #\n    print(model.summary()) #\n\n    return model\n<\/code><\/pre>\n"},{"AnswerId":"55167925","CreationDate":"2019-03-14T16:44:56.287","ParentId":null,"OwnerUserId":"11204437","Title":null,"Body":"<p>I'm using the following function and it's working for my code:<\/p>\n\n<pre><code>for layer in model1.layers[:22]:\n    model.add(layer)\n<\/code><\/pre>\n"},{"AnswerId":"42613508","CreationDate":"2017-03-05T20:23:12.060","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>When you drop layers, you need to recompile your model in order for it to have any effect.<\/p>\n\n<p>So use<\/p>\n\n<pre><code>model.compile(loss=...,optimizer=..., ...)\n<\/code><\/pre>\n\n<p>before printing the summary and it should integrate the changes correctly.<\/p>\n\n<p><strong>Edit :<\/strong><\/p>\n\n<p>What you are trying to do is actually really complex with a Sequential mode. This is the solution I can come up with (if there is any better please tell me) for your Sequential model:<\/p>\n\n<pre><code>model = load_model('model1.h5')                                                                         \nlayer1 = model.layers.pop() # Copy activation_6 layer                                      \nlayer2 = model.layers.pop() # Copy classification layer (dense_2)                          \n\nmodel.add(Dense(512, name='dense_3'))\nmodel.add(Activation('softmax', name='activation_7'))\n\n# get layer1 config\nlayer1_config = layer1.get_config()\nlayer2_config = layer2.get_config()\n# change the name of the layers otherwise it complains\nlayer1_config['name'] = layer1_config['name'] + '_new'\nlayer2_config['name'] = layer2_config['name'] + '_new'\n\n# import the magic function\nfrom keras.utils.layer_utils import layer_from_config\n# re-add new layers from the config of the old ones \nmodel.add(layer_from_config({'class_name':type(l2), 'config':layer2_config}))\nmodel.add(layer_from_config({'class_name':type(l1), 'config':layer1_config}))\n\nmodel.compile(...)\n\nprint(model.summary())\n<\/code><\/pre>\n\n<p>The hack is in the fact that your layers have <code>layer1.input<\/code> and <code>layer1.output<\/code> properties that I couldn't change.<\/p>\n\n<p>A way around that is to use a Functionnal API model. This allows you to define what comes in and what goes out of your layers.<\/p>\n\n<p>First you need to define your pop() function, to properly relink the layers every time you pop one, the function comes from <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2640\" rel=\"nofollow noreferrer\">this github issue<\/a>:<\/p>\n\n<pre><code>def pop_layer(model):\n    if not model.outputs:\n        raise Exception('Sequential model cannot be popped: model is empty.')\n\n    popped_layer = model.layers.pop()\n    if not model.layers:\n        model.outputs = []\n        model.inbound_nodes = []\n        model.outbound_nodes = []\n    else:\n        model.layers[-1].outbound_nodes = []\n        model.outputs = [model.layers[-1].output]\n    model.built = False\n    return popped_layer\n<\/code><\/pre>\n\n<p>it just removes every output links of the last layer and changes the outputs of the model to be the new last layer. Now you can use this in :<\/p>\n\n<pre><code>model = load_model('model1.h5')                                                                         \nlayer1 = model.layers.pop() # Copy activation_6 layer                                      \nlayer2 = model.layers.pop() # Copy classification layer (dense_2)     \n\n# take model.outputs and feed a Dense layer\nh = Dense(512,name='dense_3')(model.outputs)\nh = Activation('relu', name=('activation_7')(h)\n# apply\nh = layer2(h)\noutput = layer1(h)\n\nmodel = Model(input=model.input, output=output)\nmodel.compile(...)\nmodel.summary()\n<\/code><\/pre>\n\n<p>There are probably better solutions than this, but this is what I would do.<\/p>\n\n<p>I hope this helps.<\/p>\n"}]}
{"QuestionId":42611328,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-05T17:06:57.250","AcceptedAnswerId":42611524.0,"OwnerUserId":5027916.0,"Title":"How to iterate over the elements of a tensor in TensorFlow?","Body":"<p>I have coded a neural network that returns a list of 3 numbers for every input sample. These values are then subtracted from the actual values to get the difference.  <\/p>\n\n<p>For example,  <\/p>\n\n<pre><\/pre>\n\n<p>So my tensor now has the shape [batch_size, 3]\nWhat I want to do is to iterate over the tensor elements to construct my loss function.<br>\nFor instance, if my batch_size is 2 and finally   <\/p>\n\n<pre><\/pre>\n\n<p>I want the loss to be  <\/p>\n\n<pre><\/pre>\n\n<p>I know that TensorFlow has a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/l2_loss\" rel=\"nofollow noreferrer\">tf.nn.l2_loss()<\/a> function that computes the L2 loss of the entire tensor. But what I want is the mean of l2 losses of elements of a tensor along some axis.<br>\nHow do I go about doing this?<\/p>\n","answers":[{"AnswerId":"42611524","CreationDate":"2017-03-05T17:22:35.083","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>You can use  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sqrt\" rel=\"nofollow noreferrer\"><code>tf.sqrt<\/code><\/a> followed by  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_sum\" rel=\"nofollow noreferrer\"><code>tf.reduce_sum<\/code><\/a> and  <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reduce_mean\" rel=\"nofollow noreferrer\"><code>tf.reduce_mean<\/code><\/a>. Both <code>tf.reduce_sum<\/code> and <code>tf.reduce_mean<\/code> have an <code>axis<\/code> argument that indicates which dimensions to reduce.<\/p>\n\n<p>For more reduction operations, see <a href=\"https:\/\/www.tensorflow.org\/api_guides\/python\/math_ops#Reduction\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_guides\/python\/math_ops#Reduction<\/a><\/p>\n"}]}
{"QuestionId":42611698,"AnswerCount":1,"Tags":"<machine-learning><neural-network><theano><keras><recurrent-neural-network>","CreationDate":"2017-03-05T17:36:21.637","AcceptedAnswerId":null,"OwnerUserId":2923621.0,"Title":"Keras LSTM multi-dimension input","Body":"<p>my input time series data is of the shape (nb_samples, 75, 32).<br>\n75 is the timesteps and 32 is the input dimension.<\/p>\n\n<pre><\/pre>\n\n<p>The LSTM weight vectors, are all 32 dimensions, but the output is just a single value. the output shape of the above model is (1,4). But in LSTM the output is also a vector so should not it be (32,4) for many to one implementation as above? why is it giving a single value for multi-dimension input also?<\/p>\n","answers":[{"AnswerId":"42613608","CreationDate":"2017-03-05T20:32:09.093","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>As you can read in the <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">Keras doc for reccurent layers<\/a><\/p>\n\n<p>For an input of shape <code>(nb_sample, timestep, input_dim)<\/code>, you have two possible outputs:<\/p>\n\n<ul>\n<li><p>if you set <code>return_sequence=True<\/code> in your LSTM (which is not your case), you return every hidden state, so the intermediate steps when the LSTM 'reads' your sequence. You get an output of shape <code>(nb_sample, timestep, output_dim)<\/code>.<\/p><\/li>\n<li><p>if you set <code>return_sequence=False<\/code> (which is the default), it will only output the last state. So you will get an output of shape <code>(nb_sample, output_dim)<\/code>.<\/p><\/li>\n<\/ul>\n\n<p>So if you define your LSTM layer like this :<\/p>\n\n<pre><code>model.add(LSTM(4, return_sequence=True, input_shape=(75, 32)))\n<\/code><\/pre>\n\n<p>you will have an output of shape <code>(None, 75, 4)<\/code>. If 32 is your time dimension, you will have to transpose the data before feeding it to the LSTM. The first dimension is the temporal one.<\/p>\n\n<p>I hope this helps :)<\/p>\n"}]}
{"QuestionId":42611889,"AnswerCount":2,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-03-05T17:55:07.087","AcceptedAnswerId":null,"OwnerUserId":5688175.0,"Title":"Accuracy does not increase in my ResNet on MNIST dataset","Body":"<p>I have built a ResNet model with  to classify MNIST digits. However, at training time, my accuracy does not change so much and stays around 0.1 even after 3-4 epochs, which corresponds to a random classifier (1 chance over 10 to make the right prediction).<\/p>\n\n<p>I have tried changing activation functions (relu to sigmoid) but it does not improve the accuracy. Modifying the learning rate has no significant effect. I am wondering whether my get_variable() calls are correct....<\/p>\n\n<p>Here is the full model:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42705146","CreationDate":"2017-03-09T20:50:28.850","ParentId":null,"OwnerUserId":"5688175","Title":null,"Body":"<p>In fact, the problem came from the <code>from __future__ import division<\/code> which was missing. I did not insert it in my other scripts neither but it yet worked. Don't know why it is required in this script.<\/p>\n"},{"AnswerId":"42668803","CreationDate":"2017-03-08T10:29:09.433","ParentId":null,"OwnerUserId":"4834515","Title":null,"Body":"<p>You can try this.<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\nimport matplotlib.pyplot as plt\n\n\ndef conv_2D(x, w, b=None, stride=1, padding='SAME', activation=None):\n    '''\n    2D convolution\n    x: tensor of shape (batch, height, width, channel) -&gt;\n    w: tensor of shape (f_width, f_height, channels_in, channels_out) -&gt; weights\n    b: tensor of shape (channels_out) -&gt; biases\n    '''\n    # convolution\n    x = tf.nn.conv2d(x, w, strides=[1, stride, stride, 1], padding=padding)\n    # add biases\n    if b is not None:\n        x = tf.nn.bias_add(x, b)\n\n    if activation is not None:\n        x = activation(x)\n\n    return x\n\n\ndef print_tensor_shape(x, msg=''):\n    print(msg, x.get_shape().as_list())\n\n\nclass RepBlock(object):\n    def __init__(self, num_repeats, num_filters, bottleneck_size, name_scope):\n        self.num_repeats = num_repeats\n        self.num_filters = num_filters\n        self.bottleneck_size = bottleneck_size\n        self.name_scope = name_scope\n\n    def apply_block(self, net):\n\n        print_tensor_shape(net, 'entering apply_block')\n\n        # loop over repeats\n        for i_repeat in range(self.num_repeats):\n\n            print_tensor_shape(net, 'layer %i' % i_repeat)\n\n            # subsampling is performed by a convolution with stride=2, only\n            # for the first convolution of the first repetition\n            if i_repeat == 0:\n                stride = 2\n            else:\n                stride = 1\n\n            name = self.name_scope+'\/%i\/conv_in' % i_repeat\n            with tf.variable_scope(name):\n                w = tf.get_variable('w', shape=[1, 1, net.get_shape().as_list()[-1], self.bottleneck_size],\n                                    initializer=tf.contrib.layers.xavier_initializer_conv2d())\n                b = tf.get_variable('b', initializer=tf.constant(0.1, shape=[self.bottleneck_size]))\n                # w = tf.get_variable('w', initializer=tf.random_normal([1, 1, net.get_shape().as_list()[-1], self.bottleneck_size]))\n                # b = tf.get_variable('b', initializer=tf.random_normal([self.bottleneck_size]))\n                conv = conv_2D(net, w, b, stride=stride, padding='VALID', activation=tf.nn.relu)\n\n            print_tensor_shape(conv, name)\n\n            name = self.name_scope+'\/%i\/conv_bottleneck' % i_repeat\n            with tf.variable_scope(name):\n                w = tf.get_variable('w', shape=[3, 3, conv.get_shape().as_list()[-1], self.bottleneck_size],\n                                    initializer=tf.contrib.layers.xavier_initializer_conv2d())\n                b = tf.get_variable('b', initializer=tf.constant(0.1, shape=[self.bottleneck_size]))\n                # w = tf.get_variable('w', initializer=tf.random_normal([3, 3, conv.get_shape().as_list()[-1], self.bottleneck_size]))\n                # b = tf.get_variable('b', initializer=tf.random_normal([self.bottleneck_size]))\n                conv = conv_2D(conv, w, b, stride=1, padding='SAME', activation=tf.nn.relu)\n\n                print_tensor_shape(conv, name)\n\n            name = self.name_scope+'\/%i\/conv_out' % i_repeat\n            with tf.variable_scope(name):\n                w = tf.get_variable('w', shape=[1, 1, conv.get_shape().as_list()[-1], self.num_filters],\n                                    initializer=tf.contrib.layers.xavier_initializer_conv2d())\n                b = tf.get_variable('b', initializer=tf.constant(0.1, shape=[self.num_filters]))\n                # w = tf.get_variable('w', initializer=tf.random_normal([1, 1, conv.get_shape().as_list()[-1], self.num_filters]))\n                # b = tf.get_variable('b', initializer=tf.random_normal([self.num_filters]))\n                conv = conv_2D(conv, w, b, stride=1, padding='VALID', activation=None)\n                print_tensor_shape(conv, name)\n\n            if i_repeat == 0:\n                net = conv + tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')\n            else:\n                net = conv + net\n\n            net = tf.nn.relu(net)\n\n        return net\n\n\ndef resnet(x):\n    # reshape input\n    x = tf.reshape(x, shape=[-1, 28, 28, 1])\n    # init block for each layer\n    layer_1 = RepBlock(num_repeats=3, num_filters=128, bottleneck_size=32, name_scope='layer_1')\n    layer_2 = RepBlock(num_repeats=3, num_filters=256, bottleneck_size=64, name_scope='layer_2')\n#    layer_3 = RepBlock(num_repeats=3, num_filters=512, bottleneck_size=128, name_scope='layer_3')\n#    layer_4 = RepBlock(num_repeats=3, num_filters=1024, bottleneck_size=256, name_scope='layer_4')\n\n    layers = [layer_1, layer_2]\n\n    # first layer\n    name = 'conv_1'\n    with tf.variable_scope(name):\n        w = tf.get_variable('w', shape=[7, 7, x.get_shape().as_list()[-1], 64],\n                            initializer=tf.contrib.layers.xavier_initializer_conv2d())\n        # w = tf.get_variable('w', initializer=tf.random_normal([7, 7, x.get_shape().as_list()[-1], 64]))\n        b = tf.get_variable('b', initializer=tf.constant(0.1, shape=[64]))\n        net = conv_2D(x, w, b, stride=1, padding='SAME', activation=tf.nn.relu)\n\n    print_tensor_shape(net, name)\n\n    net = tf.nn.max_pool(\n        net, [1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')\n\n    print_tensor_shape(net, 'After max pooling')\n\n    with tf.variable_scope('conv_2'):\n        w = tf.get_variable('w', shape=[1, 1, net.get_shape().as_list()[-1], layers[0].num_filters],\n                            initializer=tf.contrib.layers.xavier_initializer_conv2d())\n        # w = tf.get_variable('w', initializer=tf.random_normal([1, 1, net.get_shape().as_list()[-1], layers[0].num_filters]))\n        b = tf.get_variable('b', initializer=tf.constant(0.1, shape=[layers[0].num_filters]))\n        net = conv_2D(net, w, b, stride=1, padding='SAME', activation=tf.nn.relu)\n\n    print_tensor_shape(net, 'conv_2')\n\n    for i_layer, layer in enumerate(layers):\n        print i_layer, layer\n\n        # pass the net through all blocks of the layer\n        net = layer.apply_block(net)\n\n        print_tensor_shape(net, 'After block')\n\n        try:\n            # upscale (depth) to the next block size\n            next_block = layers[i_layer+1]\n            with tf.variable_scope('upscale_%i' % i_layer):\n                w = tf.get_variable('w', shape=[1, 1, net.get_shape().as_list()[-1], next_block.num_filters],\n                                    initializer=tf.contrib.layers.xavier_initializer_conv2d())\n                # w = tf.get_variable('w', initializer=tf.random_normal([1, 1, net.get_shape().as_list()[-1], next_block.num_filters]))\n                b = tf.get_variable('b', initializer=tf.constant(0.1, shape=[next_block.num_filters]))\n                net = conv_2D(net, w, b, stride=1, padding='SAME', activation=tf.nn.relu)\n\n            print_tensor_shape(net)\n\n        except IndexError:\n            pass\n\n    # apply average pooling\n    net = tf.nn.avg_pool(net, ksize=[1, net.get_shape().as_list()[1], net.get_shape().as_list()[2], 1],\n                                     strides=[1, 1, 1, 1], padding='VALID')\n\n    print_tensor_shape(net, msg='after average pooling')\n\n    # fully connected layer\n    with tf.variable_scope('fc'):\n        w = tf.get_variable('w', shape=[256, 10],\n                            initializer=tf.contrib.layers.xavier_initializer_conv2d())\n        # w = tf.get_variable('w', initializer=tf.random_normal([256, 10]))\n        b = tf.get_variable('b', initializer=tf.constant(0.1, shape=[10]))\n\n    net = tf.reshape(net, shape=[-1, 256])\n    net = tf.add(tf.matmul(net, w), b)\n\n    print_tensor_shape(net, 'after fc')\n\n    return net\n\nif __name__ == '__main__':\n    from tensorflow.examples.tutorials.mnist import input_data\n    mnist = input_data.read_data_sets(\"MNIST_data\/\", one_hot=True)\n\n    X = tf.placeholder(tf.float32, [None, 784])\n    Y = tf.placeholder(tf.float32, [None, 10])\n    Y_pred = resnet(X)\n\n    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=Y_pred, labels=Y))\n    optim = tf.train.AdamOptimizer(learning_rate=0.01).minimize(cost)\n\n    correct_pred = tf.equal(tf.argmax(Y_pred, 1), tf.argmax(Y, 1))\n    accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\n    session = tf.InteractiveSession()\n    init_op = tf.initialize_all_variables()\n    session.run(init_op)\n\n    nb_epochs = 10\n    batch_size = 128\n    training_size = mnist.train.num_examples\n\n    nb_mini_batches = training_size \/\/ batch_size\n\n    # loop over epochs\n    for i_epoch in range(nb_epochs):\n\n        # loop over mini-batches\n        for i_batch in range(nb_mini_batches):\n\n            # get mini-batch\n            batch_x, batch_y = mnist.train.next_batch(batch_size)\n\n            [_, cost_val, acc] = session.run([optim, cost, accuracy], feed_dict={X: batch_x, Y:batch_y})\n\n            print('epoch %i - batch %i - cost=%f - accuracy=%f' % (i_epoch, i_batch, cost_val, acc))\n<\/code><\/pre>\n\n<p>The only problem is the initialization, both weights and biases. Note that there are other weight initialization methods, like<\/p>\n\n<pre><code>n = filter_size * filter_size * out_filters\nkernel = tf.get_variable(\n                '', [filter_size, filter_size, in_filters, out_filters], tf.float32,\n                initializer=tf.random_normal_initializer(stddev=np.sqrt(2.0\/n))\n                # initializer=tf.contrib.layers.xavier_initializer()\n            )\n<\/code><\/pre>\n\n<p>Moreover, biases are initialized using constant 0.1 or 0.01 but in resnet, they do not use biases after conv2d in the block. Only use biases when using fully connected layers.<\/p>\n\n<p>Hope this could help you.<\/p>\n"}]}
{"QuestionId":42612404,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-05T18:41:38.007","AcceptedAnswerId":null,"OwnerUserId":650690.0,"Title":"Tensorflow installation error","Body":"<p>I am trying to install tensorflow with virtualenv\nI have successfully went through the first 5 steps but on the 5th<\/p>\n\n<pre><\/pre>\n\n<p>I am getting an error <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42791361","CreationDate":"2017-03-14T16:24:34.183","ParentId":null,"OwnerUserId":"7588778","Title":null,"Body":"<p>I have no idea why you have that error.\nBut you can try with first installing pip3 and then try install with pip3 as\npip3 install --upgrade tensorflow\nThis will help you.<\/p>\n"},{"AnswerId":"42612651","CreationDate":"2017-03-05T19:01:20.617","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>You might be installing globally try adding sudo<\/p>\n\n<pre><code>sudo pip install --upgrade tensorflow\n<\/code><\/pre>\n\n<p>Otherwise just upgrade numpy separately with sudo and then you can try to install tensorflow locally.<\/p>\n"}]}
{"QuestionId":42612542,"AnswerCount":1,"Tags":"<keras>","CreationDate":"2017-03-05T18:52:58.130","AcceptedAnswerId":42613664.0,"OwnerUserId":4056372.0,"Title":"Keras: shape error when using validation data","Body":"<p>Trying to add validation to , but whenever I do I get an error:<\/p>\n\n<pre><\/pre>\n\n<h3>Model:<\/h3>\n\n<pre><\/pre>\n\n<p>What could be the error? If I remove  the training works just fine. I've also tried to manually split my training set into two and add it with  but I got the exact same error.<\/p>\n","answers":[{"AnswerId":"42613664","CreationDate":"2017-03-05T20:37:42.180","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>The issue comes from the fact that you hard code the <code>batch_size<\/code> value of your inputs. You have fixed it to 32 and then when you try and validate your model, the validation data is sent with a batch of 6 samples, this might be because you don't have enough validation data or maybe because the number of sample isn't a multiple of 32... However, I would let the <code>batch_size<\/code> free if I was you. Like this:<\/p>\n\n<pre><code>model.add(LSTM(128, input_shape=(timesteps, data_dim), return_sequences=True, stateful=True))\n<\/code><\/pre>\n\n<p>You specify <code>input_shape<\/code> instead of <code>batch_input_shape<\/code>. That way, your network will accept any size of batch, every layer down in the stream of your model are made to adapt to any batch_size if not hardcoded. <\/p>\n\n<p>I hope this helps :)<\/p>\n"}]}
{"QuestionId":42612748,"AnswerCount":2,"Tags":"<lua><gnuplot><torch>","CreationDate":"2017-03-05T19:10:30.810","AcceptedAnswerId":null,"OwnerUserId":1984680.0,"Title":"How to plot two lines with Torch gnuplot?","Body":"<p>I have two data sets , . I want to plot them all in one figure.<\/p>\n\n<pre><\/pre>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/21JE9.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/21JE9.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>This works well, but I'd like to have straight lines instead of dotted lines. Also, I want to include a legend in top right corner.\nAlas, whatever combination I tried, it does not work.<\/p>\n\n<pre><\/pre>\n\n<p>How can I plot two lines and add a legend in one figure with <strong>gnuplot<\/strong>?<\/p>\n","answers":[{"AnswerId":"42615171","CreationDate":"2017-03-05T23:11:04.387","ParentId":null,"OwnerUserId":"2604213","Title":null,"Body":"<p>According to <a href=\"https:\/\/github.com\/torch\/gnuplot\/blob\/master\/doc\/plotline.md\" rel=\"nofollow noreferrer\">the docs<\/a> the following should work:<\/p>\n\n<pre><code>gnuplot.plot({'first', x, y, '-'}, \n    {'second', x1, y1, '-'})\n<\/code><\/pre>\n"},{"AnswerId":"42784343","CreationDate":"2017-03-14T11:08:34.560","ParentId":null,"OwnerUserId":"7394273","Title":null,"Body":"<p>torchstyle variable contains: \"set style data linespoints\", which sets the style that data are plotted to linespoints.<\/p>\n\n<p>You need to write this command:<\/p>\n\n<pre><code>gnuplot.raw('set style data lines')\n<\/code><\/pre>\n"}]}
{"QuestionId":42612959,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-05T19:31:07.500","AcceptedAnswerId":42613697.0,"OwnerUserId":1778374.0,"Title":"Associating metadata with filename_queue in TensorFlow","Body":"<p>I have setup my reading pipeline as follows: (as explained in <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">this<\/a> tutorial)<\/p>\n\n<pre><\/pre>\n\n<p>How can I access the name of the file (or some other identifier I can associate with the file) while reading the images?<\/p>\n","answers":[{"AnswerId":"42613697","CreationDate":"2017-03-05T20:40:32.640","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>In general, this can be tricky to achieve with the current TensorFlow pipelines, but fortunately the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/WholeFileReader\" rel=\"nofollow noreferrer\"><code>tf.WholeFileReader<\/code><\/a> makes it straightforward: the <code>key<\/code> tensor (returned by <code>tf.WholeFileReader.read()<\/code>) contains the filename of the file whose data is stored in the <code>image_file<\/code> tensor. You can simply use <code>key<\/code> wherever you want to access the filename.<\/p>\n"}]}
{"QuestionId":42613747,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-03-05T20:44:14.513","AcceptedAnswerId":42614086.0,"OwnerUserId":6304774.0,"Title":"Tensorflow splitting training data to batches","Body":"<p>I have a dataset of images as a Numpy array. (Number of images, length, width, colour range) I would like to split it to batches and feed to tensorflow. What is the good way to do it?<\/p>\n","answers":[{"AnswerId":"42614086","CreationDate":"2017-03-05T21:12:41.040","ParentId":null,"OwnerUserId":"3813674","Title":null,"Body":"<p>First you could use <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.split.html\" rel=\"nofollow noreferrer\"><code>numpy.split<\/code><\/a> to divide your images into batches (sub-ndarrays). Then you could feed them to the <code>tf.Session<\/code> using the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run\" rel=\"nofollow noreferrer\"><code>run<\/code><\/a> function with the <code>feed_dict<\/code> parameter. <\/p>\n\n<p>I'd also highly recommend looking at the TF <a href=\"https:\/\/www.tensorflow.org\/get_started\/mnist\/mechanics\" rel=\"nofollow noreferrer\">MNIST tutorial<\/a><\/p>\n"},{"AnswerId":"47675959","CreationDate":"2017-12-06T13:51:34.850","ParentId":null,"OwnerUserId":"2653663","Title":null,"Body":"<p>There's a small error in  Thomas Pinetz answer and I can't make comments yet, so here's an extra answer.<\/p>\n\n<p><code>int(len(array)\/batch_size)<\/code> will round the division down to the nearest integer, so the last batch wouldn't be processed. To round up the division you can use<\/p>\n\n<pre><code>ceil_int = -(-a\/\/b)\n<\/code><\/pre>\n\n<p>In addition you might end up with the last batch being very tiny compared to the rest. You can modify your batch size slightly to make this less likely to happen. The complete code is shown below:<\/p>\n\n<pre><code>def ceil(a,b):\n    return -(-a\/\/b)\n\nn_samples = len(array)\nbetter_batch_size = ceil(n_samples, ceil(n_samples, batch_size))\n\nfor i in range(ceil(n_samples, better_batch_size)):\n    batch = array[i * better_batch_size: (i+1) * better_batch_size]\n<\/code><\/pre>\n"},{"AnswerId":"42614100","CreationDate":"2017-03-05T21:13:45.953","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>I use something like this:<\/p>\n\n<pre><code>for bid in range(int(len(array)\/batch_size)):\n    batch = array[bid*batch_size:(bid+1)*batch_size]\n<\/code><\/pre>\n"}]}
{"QuestionId":42613942,"AnswerCount":1,"Tags":"<python><tensorflow><gradient-descent>","CreationDate":"2017-03-05T21:01:01.750","AcceptedAnswerId":null,"OwnerUserId":3454339.0,"Title":"Gradients are zeros","Body":"<p>I try to learn a network but always get a zeros gradient. I am really confused about it and I don't have any ideas whit it happens.<\/p>\n\n<p>I have an input data in format <em>(batch_size, 120, 10, 3)<\/em> and after six layer <em>(conv1 - pool1 - conv2 - pool2 -fc1 - fc2)<\/em> I expect output of size 1x1 (0 or 1). All of this really works fine. <\/p>\n\n<p>But when I try to learn a netwrok I have a difficulties. I always get a zero gradients. What I am doing wrong?<\/p>\n\n<pre><\/pre>\n\n<p>So as a result I have same output all time.<\/p>\n\n<pre><\/pre>\n\n<p><strong>UPDATE!<\/strong>\nThe problem was resolved. I forgot normalized data.<\/p>\n","answers":[{"AnswerId":"42614182","CreationDate":"2017-03-05T21:21:17.550","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>Your learning rate is really small consider increasing it to 0.01 and then decreasing it over time.<\/p>\n"}]}
{"QuestionId":42614381,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-03-05T21:37:54.350","AcceptedAnswerId":null,"OwnerUserId":2348387.0,"Title":"Use model to predict the next word using Tensorflow","Body":"<p>I train a \"Recurrent Neural Networks\" following the <a href=\"https:\/\/www.tensorflow.org\/tutorials\/recurrent\" rel=\"nofollow noreferrer\">tutorial<\/a> and got a model:<\/p>\n\n<pre><\/pre>\n\n<p>Now, I want to use this model to predict the next word in a program, how do I do it?.<\/p>\n\n<p>I want to use a function such as nextWord(\"What is the next word?\") to return the predicted word.<\/p>\n","answers":[{"AnswerId":"42614491","CreationDate":"2017-03-05T21:48:16.697","ParentId":null,"OwnerUserId":"3813674","Title":null,"Body":"<p>You want to use a <code>tf.train.Saver<\/code>. <\/p>\n\n<pre><code>saved_model = tf.train.get_checkpoint_state('.\/model\/') #or your model path\nsaver = tf.train.Saver()\n\nwith tf.Session() as sess:\n    saver.restore(sess, saved_model.model_checkpoint_path)\n<\/code><\/pre>\n\n<p>Also you might want to have a look at <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/4_Utils\/save_restore_model.py\" rel=\"nofollow noreferrer\">this example<\/a><\/p>\n"}]}
{"QuestionId":42614426,"AnswerCount":1,"Tags":"<dynamic><tensorflow><libraries>","CreationDate":"2017-03-05T21:42:34.467","AcceptedAnswerId":42633223.0,"OwnerUserId":478632.0,"Title":"Can Tensorflow read from HDFS on Mac?","Body":"<p>I'm trying to coerce Tensorflow on OS\/X to read from HDFS. The documentation<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/deploy\/hadoop\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/deploy\/hadoop<\/a><\/p>\n\n<p>does not clearly specify whether this is possible, and the code refers only to \"posix\" operating systems. The error I'm seeing when trying to use the HDFS is the following:<\/p>\n\n<blockquote>\n  <p>UnimplementedError (see above for traceback): File system scheme hdfs not implemented\n           [[Node: ReaderReadV2 = ReaderReadV2[_device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](TFRecordReaderV2, input_producer)]]<\/p>\n<\/blockquote>\n\n<p>Here's what I've done up to this point:<\/p>\n\n<ol>\n<li>brew installed Hadoop 2.7.2<\/li>\n<li>separately compiled Hadoop 2.7.2 for the native libraries. Hadoop is installed on \/usr\/local\/Cellar\/hadoop\/2.7.2\/libexec on my system, and the native libraries (libhdfs.dylib) are in ~\/Source\/hadoop\/hadoop-hdfs-project\/hadoop-hdfs\/target\/hadoop-hdfs-2.7.2\/lib\/native.<\/li>\n<li>Edited the code at <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.0.0\/tensorflow\/core\/platform\/hadoop\/hadoop_file_system.cc#L113-L119\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.0.0\/tensorflow\/core\/platform\/hadoop\/hadoop_file_system.cc#L113-L119<\/a> to read from libhdfs.dylib rather than libhdfs.so, recompiled, and reinstalled Tensorflow. (I have to admit this is pretty boneheaded, and I have no idea if it's all that's required to make this code work on Mac.)<\/li>\n<\/ol>\n\n<p>Here is the code to reproduce. <\/p>\n\n<p>test.sh:<\/p>\n\n<pre><\/pre>\n\n<p>test.py:<\/p>\n\n<pre><\/pre>\n\n<p>The code path I'm seeing in the Tensorflow source seems to indicate to me that I'd receive a different error than the one above if the issue were really mac-specific, since some kind of handler is registered for the \"hdfs\" scheme regardless: <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.0.0\/tensorflow\/core\/platform\/hadoop\/hadoop_file_system.cc#L474\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/v1.0.0\/tensorflow\/core\/platform\/hadoop\/hadoop_file_system.cc#L474<\/a> . Has anyone else succeeded in coercing Tensorflow to work with Mac? If it isn't supported, is there an easy place to patch it?<\/p>\n\n<p>I'm also open to suggestions as to what might be a better approach. The high-level goal is to efficiently train a model in parallel, using shared parameter servers, considering that each worker will only read a subset of the data. This is readily accomplished using the local filesystem, but it's less clear how to scale beyond that. Even if I do succeed in making the code above work, the result could suffer from problems with data locality.<\/p>\n\n<p>This thread <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2218\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2218<\/a> suggests using pyspark.RDD.toLocalIterator to iterate over the data set with a placeholder in the graph. Aside from my concern about forcing each worker to iterate through the full dataset, I don't see a way to coerce Tensorflow's builtin Estimator class to accept a custom feed function along with a specified input_fn, and a custom input_fn appears necessary in order to take advantage of models like LinearClassifier (<a href=\"https:\/\/www.tensorflow.org\/tutorials\/linear\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/linear<\/a>) that are capable of learning from sparse, weighted features.<\/p>\n\n<p>Any thoughts?<\/p>\n","answers":[{"AnswerId":"42633223","CreationDate":"2017-03-06T18:45:01.297","ParentId":null,"OwnerUserId":"7632696","Title":null,"Body":"<p>Did you enable HDFS support in .\/configure when building? That's the error you would get if HDFS is disabled.<\/p>\n\n<p>I think you made the correct change to make it work. Feel free to send a pull request to look for .dylib on macOS.<\/p>\n"}]}
{"QuestionId":42614993,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><neural-network>","CreationDate":"2017-03-05T22:46:49.080","AcceptedAnswerId":42615886.0,"OwnerUserId":4929704.0,"Title":"Back propagation algorithm gets stuck on training AND function","Body":"<p>Here is an implementation of AND function with single neuron using tensorflow:<\/p>\n\n<pre><\/pre>\n\n<p>Sometimes it works perfectly, but on some parameters it gets stuck and doesn't want to learn. For example with these initial parameters:<\/p>\n\n<pre><\/pre>\n\n<p>it will hardly make any improvement in cost function. What am I doing wrong, maybe I should somehow adjust initialization of parameters?<\/p>\n","answers":[{"AnswerId":"42615886","CreationDate":"2017-03-06T00:42:04.387","ParentId":null,"OwnerUserId":"997378","Title":null,"Body":"<p>Aren't you missing a <code>mean(error)<\/code> ?<\/p>\n\n<p>Your problem is the particular combination of the sigmoid, the cost function, and the optimizer.<\/p>\n\n<p>Don't feel bad, AFAIK this exact problem stalled the <em>entire field<\/em> for a few years.<\/p>\n\n<p>Sigmoid is flat when you're far from the middle, and You're initializing it with relatively large numbers, try \/1000.<\/p>\n\n<p>So your abs-error (or square-error) is flat too, and the GradientDescent optimizer takes steps proportional to the slope.<\/p>\n\n<p>Either of these should fix it:<\/p>\n\n<p>Use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sigmoid_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">cross-entropy<\/a> for the error - it's convex.<\/p>\n\n<p>Use a better Optimizer, like <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/train\/AdamOptimizer\" rel=\"nofollow noreferrer\">Adam<\/a>\n, who's step size is much less dependent on the slope. More on the consistency of the slope.<\/p>\n\n<p>Bonus: Don't roll your own sigmoid, use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/sigmoid\" rel=\"nofollow noreferrer\"><code>tf.nn.sigmoid<\/code><\/a>, you'll get a lot fewer NaN's that way.<\/p>\n\n<p>Have fun!<\/p>\n"}]}
{"QuestionId":42615998,"AnswerCount":3,"Tags":"<python><theano><pydot>","CreationDate":"2017-03-06T01:01:42.147","AcceptedAnswerId":null,"OwnerUserId":7628502.0,"Title":"pydot has no attribute 'find graphviz'","Body":"<p>I'm running Anaconda, python 3.5 on a Windows 65-bit machine. I've already successfully run<br>\nconda install -c anaconda graphviz\nas well as pydot.  <\/p>\n\n<p>After importing theano, pydot, and numpy, and instantiating the theano function f, when I try run <\/p>\n\n<pre><\/pre>\n\n<p>I get the AttributeError:<\/p>\n\n<p>File \"C:\\Users\\Me\\Anaconda3\\lib\\site-packages\\theano\\printing.py\", line 35, in \n    if pd.find_graphviz():<\/p>\n\n<p>AttributeError: module 'pydot' has no attribute 'find_graphviz'<\/p>\n\n<p>What's wrong?  Am I running a bad version of pydot?<\/p>\n\n<p>--Ken<\/p>\n","answers":[{"AnswerId":"46741684","CreationDate":"2017-10-14T06:41:05.977","ParentId":null,"OwnerUserId":"1959808","Title":null,"Body":"<p>The function <code>find_graphviz<\/code> has been removed in <code>pydot == 1.2.3<\/code>. Please see <a href=\"https:\/\/github.com\/erocarrera\/pydot\/issues\/126\" rel=\"nofollow noreferrer\">this <code>pydot<\/code> issue<\/a> for more details. Instead, <code>pydot<\/code> searches for GraphViz executables in the <code>PATH<\/code> variable of the current environment.<\/p>\n"},{"AnswerId":"42655604","CreationDate":"2017-03-07T18:26:04.883","ParentId":null,"OwnerUserId":"4063051","Title":null,"Body":"<p>From a related issue on github <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/3210\" rel=\"nofollow noreferrer\">https:\/\/github.com\/fchollet\/keras\/issues\/3210<\/a>:\ntry installing <code>pydot-ng<\/code> instead of <code>pydot<\/code>.<\/p>\n"},{"AnswerId":"47426398","CreationDate":"2017-11-22T03:41:20.530","ParentId":null,"OwnerUserId":"4363295","Title":null,"Body":"<p>How exactly did you installed pydot?\nPip install of pydot only supports python 2.7 version, to the best of my knowledge.<\/p>\n\n<p>For python 3+ you can try other versions of pydot avaiable at <a href=\"https:\/\/anaconda.org\/search?q=pydot%20\" rel=\"nofollow noreferrer\">Conda Cloud<\/a>. I used pydotplus:<\/p>\n\n<pre><code>conda install -c conda-forge pydotplus \n<\/code><\/pre>\n"}]}
{"QuestionId":42616126,"AnswerCount":0,"Tags":"<python><numpy><tensorflow><computer-vision>","CreationDate":"2017-03-06T01:19:38.080","AcceptedAnswerId":null,"OwnerUserId":3367721.0,"Title":"Verify Image data from tfrecord files using tensorflow","Body":"<p>I am new to tensorflow and I am trying to verify my custom dataset tfrecord data by extracting images as tensor and convert to np array and display them as jpegs.<\/p>\n\n<p>But it seems to be hung forever at this step >>> sess.run(image_tensor)<\/p>\n\n<p>My code looks like this:-<\/p>\n\n<pre><\/pre>\n\n<p><strong>its hung at this step forever no error<\/strong>:--<\/p>\n\n<pre><\/pre>\n\n<p>Can you please tell me what am I doing wrong?<\/p>\n","answers":[]}
{"QuestionId":42616589,"AnswerCount":1,"Tags":"<c++><tensorflow><face-recognition>","CreationDate":"2017-03-06T02:27:47.763","AcceptedAnswerId":null,"OwnerUserId":7663762.0,"Title":"How do I use Tensorflow c++ to implement this code (FaceNet)?","Body":"<h1><a href=\"https:\/\/github.com\/davidsandberg\/facenet\" rel=\"nofollow noreferrer\">FaceNet<\/a>.by David sandberg<\/h1>\n\n<p>FaceNet uses python to implement the code:<\/p>\n\n<pre><\/pre>\n\n<p>Now use C++ to implement this code:<\/p>\n\n<pre><\/pre>\n\n<p>How to use get_tensor_by_name in Tensorflow C++?<\/p>\n\n<p>How to call the run method, and the above python code to achieve the same purpose?<\/p>\n\n<h3>The tensor image_batch: np.expand_dims (image_data, 0) needs to be passed into a matrix value, how do I write this np.expand_dims (image_data, 0)?<\/h3>\n\n<p>This is a good tip:<a href=\"https:\/\/stackoverflow.com\/questions\/39379747\/import-opencv-mat-into-c-tensorflow-without-copying\">Import OpenCV Mat into C++ Tensorflow without copying<\/a><\/p>\n\n<p>There is no problem now, thank you all.<\/p>\n","answers":[{"AnswerId":"42616659","CreationDate":"2017-03-06T02:36:12.633","ParentId":null,"OwnerUserId":"7659637","Title":null,"Body":"<p>The C++ library has no equivalent. You just pass the names of the tensors instead of the tensor objects. <\/p>\n\n<pre><code>string image_batch = \"input:0\";\nstring phase_train_placeholder = \"phase_train:0\";\nstring embeddings = \"embeddings:0\";\n<\/code><\/pre>\n\n<p>See this question: <a href=\"https:\/\/stackoverflow.com\/questions\/39378657\/c-equivalent-of-python-tf-graph-get-tensor-by-name-in-tensorflow\/39394704#39394704\">C++ equivalent of python: tf.Graph.get_tensor_by_name() in Tensorflow?<\/a><\/p>\n"}]}
{"QuestionId":42616625,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><recurrent-neural-network><bidirectional>","CreationDate":"2017-03-06T02:32:32.657","AcceptedAnswerId":42616834.0,"OwnerUserId":7663887.0,"Title":"ValueError: Tensor must be from the same graph as Tensor with Bidirectinal RNN in Tensorflow","Body":"<p>I'm doing text tagger using Bidirectional dynamic RNN in tensorflow.\nAfter maching input's dimension, I tried to run a Session. \nthis is blstm setting parts:<\/p>\n\n<pre><\/pre>\n\n<p>and this is runing part:<\/p>\n\n<pre><\/pre>\n\n<p>But, it give me the error:<\/p>\n\n<blockquote>\n  <p>ValueError: Tensor(\"Shape:0\", shape=(1,), dtype=int32) must be from the same graph as Tensor(\"bidirectional_rnn\/fw\/fw\/stack_2:0\", shape=(1,), dtype=int32).<\/p>\n<\/blockquote>\n\n<p>Help me, please<\/p>\n","answers":[{"AnswerId":"42616834","CreationDate":"2017-03-06T02:57:43.930","ParentId":null,"OwnerUserId":"7659637","Title":null,"Body":"<p>TensorFlow stores all operations on an operational graph. This graph defines what functions output to where, and it links it all together so that it can follow the steps you have set up in the graph to produce your final output. If you try to input a Tensor or operation on one graph into a Tensor or operation on another graph it will fail. Everything must be on the same execution graph.<\/p>\n\n<p>Try removing <code>with tf.Graph().as_default():<\/code><\/p>\n\n<p>TensorFlow provides you a default graph which is referred to if you do not specify a graph. You are probably using the default graph in one spot and a different graph in your training block.<\/p>\n\n<p>There does not seem to be a reason you are specifying a graph as default here and most likely you are using separate graphs on accident. If you really want to specify a graph then you probably want to pass it as a variable, not set it like this.<\/p>\n"}]}
{"QuestionId":42617024,"AnswerCount":0,"Tags":"<python><python-3.x><tensorflow><python-import><importerror>","CreationDate":"2017-03-06T03:22:27.387","AcceptedAnswerId":null,"OwnerUserId":5798109.0,"Title":"I am unable to import tensorflow","Body":"<p>I pip-installed tensorflow by running the \"pip install tensorflow\" on my windows command prompt.<\/p>\n\n<p>But when I try importing tensorflow and run a simple program on it:<\/p>\n\n<pre><\/pre>\n\n<p>I get loads of errors.<\/p>\n\n<pre class=\"lang-none prettyprint-override\"><\/pre>\n\n<p>And the following errors are obtained<\/p>\n","answers":[]}
{"QuestionId":42617238,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><text-recognition>","CreationDate":"2017-03-06T03:51:23.440","AcceptedAnswerId":null,"OwnerUserId":6039807.0,"Title":"TensorFlow - Specific digit recognition in an image (SVHN like)","Body":"<p>I am trying to build a classifier using TensorFlow to recognize a specific part of a text in a neural net. I was inspired by the SVHN project to recognize the clock in soccer pictures. During the entire project, I am only focused on the clock only. I add a picture to be more explicit about my project.<\/p>\n\n<p>The first thing that I did was to crop the pictures around the clock and predict the digits (it could be 3 or 4 digits 9:38 or 11:34 for instance). It worked well, I have a good accuracy (>90% on test set) with 20k pictures for my training set.<\/p>\n\n<p>Now I would like to do something more complicated and I think a neural net should be able to do it but I am not sure. So I don't crop exactly the clock but I crop the whole scoreboard (with team names, etc) and I still want to predict the clock.<br>\nI tried with 20k pictures and 40k pictures for my training. In both cases, I only have 70% of accuracy on the test set. The clock is approximately always at the same position in the pictures (at the top of the scoreboard).<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/M5mzU.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/M5mzU.png\" alt=\"Explanations\"><\/a><\/p>\n\n<p>I don't understand why the accuracy is so low. If someone has a clue, it would be really helpful. Thank you very much for any help.<\/p>\n\n<p><strong>Specifications:<\/strong><\/p>\n\n<p><strong>image size :<\/strong> 32x32<br>\n<strong>numbers of labels :<\/strong> 11 (0-9 + blank)<br>\n<strong>model :<\/strong><br>\n7-layer CNN.<br>\nC1: convolutional layer, batch_size x 28 x 28 x 16, convolution size: 5 x 5 x 1 x 16<br>\nS2: sub-sampling layer, batch_size x 14 x 14 x 16<br>\nC3: convolutional layer, batch_size x 10 x 10 x 32, convolution size: 5 x 5 x 16 x 32<br>\nS4: sub-sampling layer, batch_size x 5 x 5 x 32<br>\nC5: convolutional layer, batch_size x 1 x 1 x 64, convolution size: 5 x 5 x 32 x 64<br>\nDropout<br>\nF6: fully-connected layer, weight size: 64 x 16<br>\nOutput layer, weight size: 16 x 11  <\/p>\n","answers":[]}
{"QuestionId":42617263,"AnswerCount":1,"Tags":"<c++><macros><caffe>","CreationDate":"2017-03-06T03:54:21.537","AcceptedAnswerId":42617363.0,"OwnerUserId":7651064.0,"Title":"why caffe use #define macros to create function","Body":"<p>I am reading Caffe's source code, I'm curious why caffe use #define to create functions like in the caffe.cpp. <\/p>\n\n<pre><\/pre>\n\n<p>They use RegisterBrewFunction to register functions, there must be some advantages. They can just use if else to decide which function to call, can anyone tell me why write codes in this way.<\/p>\n","answers":[{"AnswerId":"42617363","CreationDate":"2017-03-06T04:06:24.603","ParentId":null,"OwnerUserId":"1593860","Title":null,"Body":"<p>The macro doesn't just create a function; it creates a class definition, a constructor for that class, and an object of that type, all in the anonymous namespace. The effect of all that code is to add the address of the user-defined function <code>func<\/code> to a database of functions. While that's a lot of code, it's intended to make it easier for users to register functions.<\/p>\n"}]}
{"QuestionId":42617496,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-06T04:20:05.183","AcceptedAnswerId":42635926.0,"OwnerUserId":3139545.0,"Title":"Can I store state in __call__ of RNNCell","Body":"<p>I want to build my own  (for a very simple cell where the output is [-1,0,1]), however in my  I do calculations that are dependent on the output of the previous time steps. <\/p>\n\n<p>So my question is if it is possible to keep state in the  method to reuse between calls?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42635926","CreationDate":"2017-03-06T21:30:17.320","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>This will not work with <code>tf.dynamic_rnn<\/code>, so it is discouraged. Pass all state through the state argument if you want it to work. It might work with plain <code>tf.rnn<\/code> though it is not guaranteed.<\/p>\n"}]}
{"QuestionId":42617693,"AnswerCount":1,"Tags":"<python><deep-learning><keras>","CreationDate":"2017-03-06T04:44:34.920","AcceptedAnswerId":42618241.0,"OwnerUserId":3947292.0,"Title":"How do I select only a specific number of samples from the MNIST dataset provided by Keras?","Body":"<p>I'm currently training a Convolutional Neural Network on the MNIST data set using Keras. I'm loading the data set using the format<\/p>\n\n<pre><\/pre>\n\n<p>But to reduce iterating through all the data, I want to select only the first 10000 samples from each class 0-9 for  and similarly from . How can I do this?<\/p>\n","answers":[{"AnswerId":"42618241","CreationDate":"2017-03-06T05:34:31.283","ParentId":null,"OwnerUserId":"8002","Title":null,"Body":"<p>The MNIST dataset says it returns:<\/p>\n\n<pre><code>Return:\n\n    2 tuples:\n        X_train, X_test: uint8 array of grayscale image data with shape (nb_samples, 28, 28).\n        y_train, y_test: uint8 array of digit labels (integers in range 0-9) with shape (nb_samples,).\n<\/code><\/pre>\n\n<p>So you need to slice just the parts you want to keep.  I believe the syntax for pandas\/numpy is something like:<\/p>\n\n<pre><code>X_train = X_train[:10000,:,:]\nX_test = X_test[:10000,:,:]\ny_train = y_train[:10000]\ny_test  = y_test[:10000]\n<\/code><\/pre>\n"}]}
{"QuestionId":42618048,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><tensorboard>","CreationDate":"2017-03-06T05:18:38.853","AcceptedAnswerId":42618476.0,"OwnerUserId":1496362.0,"Title":"SummaryWriter not outputting graph in TensorFlow","Body":"<p>I am trying to use tensorboard to analyse a graph in tensorflow with summaryWriter. However, TensorFlow is not outputting a 'graph' folder with information. Perhaps I am missing a command or it is not in the right place? <\/p>\n\n<pre><\/pre>\n\n<p>Is what I used. I think this may not work for TensorFlow 1.0 anymore (just the summarywriter command)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42618476","CreationDate":"2017-03-06T05:55:28.273","ParentId":null,"OwnerUserId":"1496362","Title":null,"Body":"<p>This worked for me: <\/p>\n\n<pre><code>writer = tf.summary.FileWriter(logdir='logdir', graph=tf.get_default_graph())\nwriter.flush()\n<\/code><\/pre>\n"}]}
{"QuestionId":42618170,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-03-06T05:29:39.003","AcceptedAnswerId":42628866.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Does rebuilding an installed TensorFlow from source cause any internal problems?","Body":"<p>I am currently running TensorFlow training on an AWS P2 instance which has tensorflow 0.12 installed. I upgraded it to 1.0 using pip, but if I want to configure it from source and rebuild it using bazel in order to improve the CPU computation speed (according to suggestions given whenever I run TF), would this be possible? Or do I have to uninstall TF and rebuild from scratch?te<\/p>\n","answers":[{"AnswerId":"42628866","CreationDate":"2017-03-06T15:02:24.657","ParentId":null,"OwnerUserId":"1405024","Title":null,"Body":"<p>In order to use the your CPU potential you should compile it with the correct flags and reinstall.<\/p>\n\n<p>Following this guide: <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources<\/a><\/p>\n\n<p>I use, for my concrete case:<\/p>\n\n<pre><code>bazel build --config=opt --copt=-msse4.2 --copt=-msse4.1 --copt=-msse3 \/\/tensorflow\/tools\/pip_package:build_pip_package\n<\/code><\/pre>\n\n<p>Hope it helps.<\/p>\n"}]}
{"QuestionId":42618350,"AnswerCount":0,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-03-06T05:43:46.700","AcceptedAnswerId":null,"OwnerUserId":4405877.0,"Title":"Tensorflow consecutive slice_input_producer","Body":"<p>I am confused with how does  works, I know it returns a dequeue op, but if I apply another  on the previous dequeued tensor, it seems the second  triggered the first ones dequeue op without finishing slicing. <\/p>\n\n<p>Here is an example code:<\/p>\n\n<pre><\/pre>\n\n<p>I am expecting 0,1,2,3,4,5 each time I eval . But it gives 0 and 4, the second  didn't finishing slicing the first row.<\/p>\n\n<p>How can I fix this? Many thanks in advance. <\/p>\n","answers":[]}
{"QuestionId":42618543,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><conv-neural-network>","CreationDate":"2017-03-06T06:00:31.857","AcceptedAnswerId":42624588.0,"OwnerUserId":6614288.0,"Title":"Performing 1d convolution using 2d kernel in keras","Body":"<p>I am currently working on a CNN network, in which i want to apply a 2d kernel on a image, but it only has to perform 1d convolution, meaning that it only has to move along one axis (x-axis in this case). <\/p>\n\n<p>The shape of the kernel is same as the y-axis of the image. The number of filters applied is not a concern at the moment. <\/p>\n\n<p>An example: \nGiven a image of size (6,3,3) = (rows, cols,  color_channel)<\/p>\n\n<p>How should i perform a 1d convolution given a 2d filter?<\/p>\n\n<p>Tried what was suggested by @Marcin Mo\u017cejko<\/p>\n\n<pre><\/pre>\n\n<p>Error: <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42624588","CreationDate":"2017-03-06T11:34:27.460","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Assuming that your image <code>shape=(dim_x, dim_y, img_channels)<\/code> you can obtain a <code>1D<\/code> convolution by setting:<\/p>\n\n<pre><code>conv1d_on_image = Convolution2D(output_channels, 1, dim_y, border_mode='valid')(input)\n<\/code><\/pre>\n\n<p>Remember that the output from this layer would have shape <code>(dim_x, 1, output_channels)<\/code>. If you want your input to be sequential you may use the <code>Reshape<\/code> layer by setting:<\/p>\n\n<pre><code>conv1d_on_image = Reshape((dim_x, output_channels))(conv1d_on_image)\n<\/code><\/pre>\n\n<p>This would produce output with shape <code>(dim_x, output_channels)<\/code>.<\/p>\n\n<p>An interesting fact is that this is exactly the way how <code>Conv1D<\/code> works in <code>Keras<\/code> with <code>tf<\/code> backend.<\/p>\n"}]}
{"QuestionId":42618576,"AnswerCount":1,"Tags":"<torch>","CreationDate":"2017-03-06T06:03:26.350","AcceptedAnswerId":null,"OwnerUserId":7664543.0,"Title":"How can I customize the loss function in Torch?","Body":"<p>I'm currently using Torch 7 and I need to customize the loss function in particular the Cross-Entropy error function.<\/p>\n\n<p>I'm thinking of adding some parameters to Cross Entropy error function and I couldn't find which part should I modify.<\/p>\n\n<p>I took a look at CrossEntropyCriterion.lua but still don't know the way cause I don't see any equation in this file.<\/p>\n\n<p>Can anyone tell me where the equation is? or which file I should modify? <\/p>\n","answers":[{"AnswerId":"42623545","CreationDate":"2017-03-06T10:45:37.690","ParentId":null,"OwnerUserId":"6076729","Title":null,"Body":"<p>in order to customize a loss function you have to change the methods <code>__init<\/code>, <code>updateOutput<\/code> and <code>updateGradInput<\/code>.<\/p>\n\n<ul>\n<li><code>__init<\/code> is the class initialization function<\/li>\n<li><code>updateOutput<\/code> will be called when you use <code>:forward()<\/code> method in your criterion<\/li>\n<li><code>updateGradInput<\/code> will be called when you use <code>:backward()<\/code> and it is the gradient of your criterion<\/li>\n<\/ul>\n\n<p>The structure of the customized criterion looks like:<\/p>\n\n<pre><code>local yourCriterion, parent = torch.class('nn.yourCriterion', 'nn.Criterion')\n\nfunction yourCriterion:__init(your_parameters):\n  parent.__init(self)\n  ... (you can add as many parameters as you want to your criterion\n       and give them the name your prefer)\n  self.parameters = your_parameters\n\n\nfunction yourCriterion:updateOutput(input)\n  ... (your criterion code here)\n  return value_of_the_criterion\n\nfunction yourCriterion:updateGradInput(input):\n  ... (your criterion gradient code here)\n  return gradient\n<\/code><\/pre>\n\n<p>[EDIT]: You can find the code of the cross-entropy criterion here <a href=\"https:\/\/github.com\/torch\/nn\/blob\/master\/CrossEntropyCriterion.lua\" rel=\"nofollow noreferrer\">https:\/\/github.com\/torch\/nn\/blob\/master\/CrossEntropyCriterion.lua<\/a><\/p>\n"}]}
{"QuestionId":42618861,"AnswerCount":1,"Tags":"<python><compiler-errors><tensorflow><deep-learning><tensorboard>","CreationDate":"2017-03-06T06:25:48.760","AcceptedAnswerId":null,"OwnerUserId":7528619.0,"Title":"I can not run my code Tensorflow","Body":"<p>I working on  ,i wrote a code for Recognition of letters whith  ,\ni'm creat a code source for uses the model  \nmy code source : <\/p>\n\n<p>`<\/p>\n\n<pre><\/pre>\n\n<p>if i run this code , he print the error :<\/p>\n\n<pre><\/pre>\n\n<p>what is the problem and how resolved this ? <\/p>\n","answers":[{"AnswerId":"42619497","CreationDate":"2017-03-06T07:07:49.400","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Covi\u00dfio, <\/p>\n\n<p>I think the problem is as follows: <\/p>\n\n<ul>\n<li>You created a network, and saved this network...<\/li>\n<li>You changed the network, and did not remove the saved network<\/li>\n<li>Now you try to reload your network from an old version, but a new variable you created does not exist. <\/li>\n<\/ul>\n\n<p>Can you try to either: <\/p>\n\n<ul>\n<li>Remove the saved state of your network and retrain it<\/li>\n<li>Remove the saving and loading of your network and see if this works?\nYou can remove the state of your network by removing the checkpoint file in your folder \/home\/brm17\/Desktop\/PFE\/<\/li>\n<\/ul>\n\n<p>Edit: read your code thoroughly and the problem is that if there is no checkpoint you dont start retraining your network... Maybe you start by writing this before saving, loading, and changing your network. <\/p>\n\n<p>Good luck and let me know if this works!<\/p>\n"}]}
{"QuestionId":42619180,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><recurrent-neural-network>","CreationDate":"2017-03-06T06:47:19.590","AcceptedAnswerId":42619449.0,"OwnerUserId":5729690.0,"Title":"Tensorflow's API: seq2seq","Body":"<p>I have been following <a href=\"https:\/\/github.com\/kvfrans\/twitch\/blob\/master\/main.py\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kvfrans\/twitch\/blob\/master\/main.py<\/a> tutorial to create and train a chatbot based on rnn using tensorflow. From what I understand, the tutorials was written on an older version of tensorflow, so some parts are outdated and give me an error like:<\/p>\n\n<pre><\/pre>\n\n<p>I fixed some of them, but can't figure out what is the alternative to  and what should be the new module's parameters. What I currently fixed:<\/p>\n\n<p> changed to\n<\/p>\n\n<p> changed to <\/p>\n\n<p> changed to\n<\/p>\n\n<p>Can someone please help me figure out what  will be?<\/p>\n","answers":[{"AnswerId":"42619449","CreationDate":"2017-03-06T07:04:32.510","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>I think <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/legacy_seq2seq\/rnn_decoder\" rel=\"nofollow noreferrer\">this<\/a> is the one you need: <\/p>\n\n<pre><code>tf.contrib.legacy_seq2seq.rnn_decoder\n<\/code><\/pre>\n"}]}
{"QuestionId":42619519,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-06T07:09:26.143","AcceptedAnswerId":42628757.0,"OwnerUserId":5584085.0,"Title":"How do I change the following code in Tensorflow API r1.0?","Body":"<p>I've upgraded my tensorflow from 0.12 to 1.0. As a result of that, I've met the error code... I try the solving the error, but I can't find the solution. \nI wish your knowledge and experience associated with this to be shared. Thank you.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42627927","CreationDate":"2017-03-06T14:19:11.227","ParentId":null,"OwnerUserId":"3941813","Title":null,"Body":"<p>As of version 1.0 of the API <code>tf.nn.rnn<\/code> was removed. Try using <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/dynamic_rnn\" rel=\"nofollow noreferrer\"><code>tf.nn.dynamic_rnn<\/code><\/a>. Notice however, that this method expects <code>inputs<\/code> to be a tensor and not a list of tensors (See linked documentation), so you will have to change a bit the way you create your <code>inputs<\/code> variable.<\/p>\n"},{"AnswerId":"42628757","CreationDate":"2017-03-06T14:57:47.210","ParentId":null,"OwnerUserId":"1405024","Title":null,"Body":"<pre><code>outputs, state = tf.nn.rnn(lstm_cell, inputs, initial_state=self._initial_state, sequence_length=self.real_len)\n<\/code><\/pre>\n\n<p>Should be (now, in TF 1.0):<\/p>\n\n<pre><code>outputs, state = tf.contrib.rnn.static_rnn(lstm_cell, inputs, initial_state=self._initial_state, sequence_length=self.real_len)\n<\/code><\/pre>\n\n<p>Due to nn.rnn has been moved to tf.contrib<\/p>\n"}]}
{"QuestionId":42619665,"AnswerCount":1,"Tags":"<tensorflow><anaconda><spyder>","CreationDate":"2017-03-06T07:17:27.037","AcceptedAnswerId":null,"OwnerUserId":7664841.0,"Title":"Why spyder can't import tensorflow","Body":"<p>I have installed tensorflow with anaconda in Windows, but I can't import tensorflow in spyder IDE? <\/p>\n","answers":[{"AnswerId":"46573208","CreationDate":"2017-10-04T19:59:38.220","ParentId":null,"OwnerUserId":"2363193","Title":null,"Body":"<p>The reason Spyder is unable to import Tensorflow is that Anaconda does not store the tensorflow package in the same environment.<\/p>\n\n<p>One solution is to create a new separate environment in Anaconda dedicated to TensorFlow and containing Python with its own Spyder and then install tensorflow into that environment<\/p>\n"}]}
{"QuestionId":42620065,"AnswerCount":2,"Tags":"<python><deep-learning><keras><autoencoder>","CreationDate":"2017-03-06T07:42:13.950","AcceptedAnswerId":null,"OwnerUserId":7664949.0,"Title":"Implementation of normalizing flows in Keras","Body":"<p>I've been trying to implement a simple version of normalizing flows with Keras, as explained in this paper: <a href=\"https:\/\/arxiv.org\/pdf\/1505.05770.pdf\" rel=\"nofollow noreferrer\">https:\/\/arxiv.org\/pdf\/1505.05770.pdf<\/a><\/p>\n\n<p>My problem is that the loss is always -infinity, and I can't get what I did wrong. Can anybody help me ? <\/p>\n\n<p>Here is the procedure: <\/p>\n\n<ol>\n<li><p>the encoder generates vectors of size . These are .<\/p><\/li>\n<li><p>From  and , using the reparametrization trick I can sample  ~ .<\/p><\/li>\n<li><p>Then I can compute <\/p><\/li>\n<li><p>Then I can compute <\/p><\/li>\n<\/ol>\n\n<p>Here is the code for those four steps:<\/p>\n\n<pre><\/pre>\n\n<p>Then here is the loss (where  is defined above)<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42906832","CreationDate":"2017-03-20T14:53:45.677","ParentId":null,"OwnerUserId":"7664949","Title":null,"Body":"<p>I modified the Keras tutorial on VAE here : <a href=\"https:\/\/github.com\/sbaurdlp\/keras-iaf-mnist\" rel=\"nofollow noreferrer\">https:\/\/github.com\/sbaurdlp\/keras-iaf-mnist<\/a><\/p>\n\n<p>If someone is interested to look...\nStrangely adding more layers doesn't improve performance, and I can't see what is wrong in the code<\/p>\n"},{"AnswerId":"42646014","CreationDate":"2017-03-07T10:38:15.750","ParentId":null,"OwnerUserId":"7664949","Title":null,"Body":"<p>As I couldn't make it work, I have tried to implement the normalizing flow described in <a href=\"https:\/\/arxiv.org\/pdf\/1606.04934.pdf\" rel=\"nofollow noreferrer\">this<\/a> paper: <em>Improved Variational Inference\nwith Inverse Autoregressive Flow<\/em> <\/p>\n\n<p>However I still meet the same problem of diverging loss (towards -infinity), which makes no sense. There must be a problem with my implementation.<\/p>\n\n<p>Here are the important parts:<\/p>\n\n<pre><code># the encoder\nh = encoder_block(x)  # a convnet taking proteins as input (matrices of size 400x22), I don't describe it since it isn't very important\nz_log_var = Dense(latent_dim)(h)\nz_mean = Dense(latent_dim)(h)\nh_ = Dense(latent_dim)(h)\nencoder = Model(x, [z_mean,z_log_var, h_])\n\n# the latent variables (only one transformation to keep it simple)\nlatent_input = Input(shape=(latent_dim, 2), batch_shape=(batch_size, latent_dim, 2))\nhl = Convolution1D(1, filter_length, activation=\"relu\", border_mode=\"same\")(latent_input)\nhl = Reshape((latent_dim,))(hl)\nmean_1 = Dense(latent_dim)(hl)\nstd_1 = Dense(latent_dim)(hl)\nlatent_model = Model(latent_input, [mean_1, std_1])\n\n# the decoder\ndecoder_input = Input((latent_dim,), batch_shape=(batch_size, latent_dim))\ndecoder=decoder_block()  # a convnet that I don't describe\nx_decoded_mean = decoder(decoder_input)\ngenerator = Model(decoder_input, x_decoded_mean)\n\n# the VAE\nz_mean, z_log_var, other = encoder(vae_input)\neps = Lambda(sample_eps, name='sample_eps')([z_mean, z_log_var, other])\nz0 = Lambda(sample_z0, name='sample_z0')([z_mean, z_log_var, eps])\nl = Lambda(sample_l, name='sample_l')([eps, z_log_var])\nmean, std = latent_model(merge([Reshape((latent_dim,1))(z0), Reshape((latent_dim,1))(other)], mode=\"concat\", concat_axis=-1))\nz = Lambda(transform_z0)([z0, mean, std])\nl = Lambda(transform_l)([l, std])\nx_decoded_mean = generator(z)\nvae = Model(vae_input, x_decoded_mean)\n\n# and here is the loss\ndef vae_loss(x, x_decoded_mean):\n    xent_loss = K.mean(objectives.categorical_crossentropy(x, x_decoded_mean), -1)\n    ln_q0z0 = K.sum(log_normal2(z0, z_mean, z_log_var), -1)\n    ln_pz1 = K.sum(log_stdnormal(z), -1)\n    result = K.mean(l + ln_pz1 + xent_loss - ln_q0z0)\n    return result\n<\/code><\/pre>\n\n<p>Here are the utils functions I use above in the <code>Lambda<\/code> layers:<\/p>\n\n<pre><code>def sample_eps(args):\n\n    # sample epsilon according to N(O,I)\n    epsilon = K.random_normal(shape=(batch_size, latent_dim), mean=0.,\n                              std=epsilon_std)\n\n    return epsilon\n\ndef sample_z0(args):\n    z_mean, z_log_var, epsilon = args\n    # generate z0 according to N(z_mean, z_log_var)\n    z0 = z_mean + K.exp(z_log_var \/ 2) * epsilon\n    return z0\n\ndef sample_l(args):\n    epsilon, z_log_var = args\n    l = -0.5*K.sum(z_log_var + epsilon**2 + K.log(2*math.pi), -1)\n    return l\n\ndef transform_z0(args):\n    z0, mean, std = args\n    z = z0\n    sig_std = K.sigmoid(std)\n    z *= sig_std\n    z += (1-sig_std)*mean\n    return z\n\ndef transform_l(args):\n    l, std = args\n    sig_std = K.sigmoid(std)\n    l -= K.sum(K.log(sig_std+1e-8), -1)\n    return l\n<\/code><\/pre>\n"}]}
{"QuestionId":42620663,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><reinforcement-learning><tensorboard>","CreationDate":"2017-03-06T08:19:59.423","AcceptedAnswerId":null,"OwnerUserId":5163915.0,"Title":"DQN cannot converge while the example with the same graph works","Body":"<p>Here is the problem. I try to rewrite a dpn example of the CartPole-v0. The example converges with less than 2000 episodes, However, my version could not converge. <\/p>\n\n<p>Therefore I use TensorBoard as the tool to help me figure out the problem. I find that these two codes have the same graph<\/p>\n\n<p>What's more, I find that the graph of the loss is very different.\n<a href=\"https:\/\/i.stack.imgur.com\/dBBM5.png\" rel=\"nofollow noreferrer\">loss of the example<\/a>\n<a href=\"https:\/\/i.stack.imgur.com\/afliU.png\" rel=\"nofollow noreferrer\">loss of mine<\/a><\/p>\n\n<p>I wonder what's the problem. I will be grateful if anyone could help me with it! Thank you!<\/p>\n\n<p>P.S. If the codes are needed. I could upload them. Both of them are shorter than 200 lines.<\/p>\n","answers":[]}
{"QuestionId":42620679,"AnswerCount":1,"Tags":"<caffe>","CreationDate":"2017-03-06T08:20:46.143","AcceptedAnswerId":42643282.0,"OwnerUserId":1279459.0,"Title":"How to load caffemodel weights for retraining using Pycaffe?","Body":"<p>How to load caffemodel weights (alongside with a solver) to use subsequently for retraining using Python? <\/p>\n\n<p>In other words, what would be the counterpart of the following command line in Pycaffe?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42643282","CreationDate":"2017-03-07T08:14:27.070","ParentId":null,"OwnerUserId":"1279459","Title":null,"Body":"<p>Loading caffemodel for finetuning with a particular solver can be obtained in Pycaffe by means of <code>get_solver()<\/code> and <code>net.copy_from()<\/code> functions:<\/p>\n\n<pre><code>solver=caffe.get_solver('prototxtfile.prototxt')\nsolver.net.copy_from('weights.caffemodel')\n<\/code><\/pre>\n"}]}
{"QuestionId":42621652,"AnswerCount":1,"Tags":"<python><tensorflow><nlp><lstm><bidirectional>","CreationDate":"2017-03-06T09:15:43.100","AcceptedAnswerId":48065068.0,"OwnerUserId":7665315.0,"Title":"Tensorflow : ValueError: Shape must be rank 2 but is rank 3","Body":"<p>I'm new to tensorflow and I'm trying to update some code for a bidirectional LSTM from an old version of tensorflow to the newest (1.0), but I get this error: <\/p>\n\n<blockquote>\n  <p>Shape must be rank 2 but is rank 3 for 'MatMul_3' (op: 'MatMul') with input shapes: [100,?,400], [400,2].<\/p>\n<\/blockquote>\n\n<p>The error happens on pred_mod.<\/p>\n\n<pre><\/pre>\n\n<p>Any help appreciated.<\/p>\n","answers":[{"AnswerId":"48065068","CreationDate":"2018-01-02T17:08:30.533","ParentId":null,"OwnerUserId":"712995","Title":null,"Body":"<p>For anyone encountering this issue in the future, the snippet above <strong>should not<\/strong> be used.<\/p>\n\n<p>From <a href=\"https:\/\/www.tensorflow.org\/versions\/r1.1\/api_docs\/python\/tf\/contrib\/rnn\/static_bidirectional_rnn\" rel=\"nofollow noreferrer\"><code>tf.contrib.rnn.static_bidirectional_rnn<\/code><\/a> v1.1 documentation:<\/p>\n\n<blockquote>\n  <p>Returns:<\/p>\n  \n  <p>A tuple <code>(outputs, output_state_fw, output_state_bw)<\/code> where: outputs is a length T list of outputs (one for each input), which are depth-concatenated forward and backward outputs. output_state_fw is the final state of the forward rnn. output_state_bw is the final state of the backward rnn.<\/p>\n<\/blockquote>\n\n<p>The list comprehension above is expecting LSTM outputs, and the correct way to get those is this:<\/p>\n\n\n\n<pre class=\"lang-py prettyprint-override\"><code>outputs, _, _ = tf.contrib.rnn.static_bidirectional_rnn(lstmcell, lstmcell, ...)\npred_mod = [tf.matmul(item, _weights['out_w']) + _biases['out_b'] \n            for item in outputs]\n<\/code><\/pre>\n\n<p>This will work, because each <code>item<\/code> in <code>outputs<\/code> has the shape <code>[batch_size, 2 * num_hidden]<\/code> and can be multiplied with the weights by <code>tf.matmul()<\/code>.<\/p>\n\n<hr>\n\n<p>Add-on: from tensorflow v1.2+, the recommended function to use is in another package: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/static_bidirectional_rnn\" rel=\"nofollow noreferrer\"><code>tf.nn.static_bidirectional_rnn<\/code><\/a>. The returned tensors are the same, so the code doesn't change much:<\/p>\n\n<pre class=\"lang-py prettyprint-override\"><code>outputs, _, _ = tf.nn.static_bidirectional_rnn(lstmcell, lstmcell, ...)\npred_mod = [tf.matmul(item, _weights['out_w']) + _biases['out_b'] \n            for item in outputs]\n<\/code><\/pre>\n"}]}
{"QuestionId":42621864,"AnswerCount":2,"Tags":"<machine-learning><tensorflow><neural-network><keras>","CreationDate":"2017-03-06T09:26:23.887","AcceptedAnswerId":42622446.0,"OwnerUserId":3993741.0,"Title":"Difference between Keras model.save() and model.save_weights()?","Body":"<p>To save a model in Keras, what are the differences between the output files of:<\/p>\n\n<ol>\n<li> <\/li>\n<li><\/li>\n<li> in the callback<\/li>\n<\/ol>\n\n<p>The saved file from  is larger than the model from , but significantly larger than a JSON or Yaml model architecture file.  Why is this?  <\/p>\n\n<p>Restating this: Why is size(model.save()) + size(something) = size(model.save_weights()) + size(model.to_json()), what is that \"something\"?<\/p>\n\n<p>Would it be more efficient to just  and , and load from these than to just do  and ?  <\/p>\n\n<p>What are the differences?<\/p>\n","answers":[{"AnswerId":"42622446","CreationDate":"2017-03-06T09:54:07.373","ParentId":null,"OwnerUserId":"349130","Title":null,"Body":"<p><code>save()<\/code> saves the weights and the model structure to a single <code>HDF5<\/code> file. I believe it also includes things like the optimizer state. Then you can use that HDF5 file with <code>load()<\/code> to reconstruct the whole model, including weights.<\/p>\n\n<p><code>save_weights()<\/code> only saves the weights to HDF5 and nothing else. You need extra code to reconstruct the model from a <code>JSON<\/code> file.<\/p>\n"},{"AnswerId":"58930036","CreationDate":"2019-11-19T08:55:45.383","ParentId":null,"OwnerUserId":"12367501","Title":null,"Body":"<ul>\n<li>model.save_weights() will only save the weights so if you need, you are able to apply them on a different architecture  <\/li>\n<li>mode.save() will save the architecture of the model + the the weights + the training configuration + the state of the optimizer<\/li>\n<\/ul>\n"}]}
{"QuestionId":42622012,"AnswerCount":2,"Tags":"<python><deep-learning><keras><regularized>","CreationDate":"2017-03-06T09:33:53.470","AcceptedAnswerId":null,"OwnerUserId":4452379.0,"Title":"Set attributes of Keras layer afterwards","Body":"<p>I have a customized regularizer which needs the model output tensor to be analyzed. The basically I can't put in this way.<\/p>\n\n<pre><\/pre>\n\n<p>Where the custom function needs the target label tensor and unfortunately such tensor is not implemented yet.<\/p>\n\n<p>I also try to set attribute of such layer:<\/p>\n\n<pre><\/pre>\n\n<p>When  of such layer I see it has been applied correctly but seems not work during the training. Or perhaps this way is not a wise way to implement.<\/p>\n","answers":[{"AnswerId":"42623805","CreationDate":"2017-03-06T10:57:14.930","ParentId":null,"OwnerUserId":"4452379","Title":null,"Body":"<p>thanks @Thomas, i tried re-compiling but not work as expected (as first one shown below)<\/p>\n\n<p>i tried the following method..<br>\n1) directly specify (work)<\/p>\n\n<pre><code>#put inf number here for testing\nx = Dense(128,name=\"dense_1\", W_regularizer = l2(1E300))\nmodel.add(x)\n...\nmodel.compile(...)\n<\/code><\/pre>\n\n<p>2) set attribute afterwards and compile (not working)<\/p>\n\n<pre><code>x = Dense(128,name=\"dense_1\")\nmodel.add(x)\nx.W_regularizer = l2(1E300)\n...\nmodel.compile(...)\n<\/code><\/pre>\n\n<p>3) using <code>add_weight<\/code> in topology.layer (it works)<br><\/p>\n\n<pre><code>x = Dense(128,name=\"dense_1\")\nmodel.add(x)\nx.W = x.add_weight((x.input_dim, x.output_dim),\n               initializer='zero',\n               name='{}_W'.format(x.name),\n               constraint=x.W_constraint,\n               regularizer=l2(1E300))\nmodel.compile(...)\n<\/code><\/pre>\n\n<p>for 1) 3) i got <code>loss: nan<\/code> which implies that the modification is successful applied and 3) is what i want. while second one don't. but there's must be a smart way to achieve.<\/p>\n"},{"AnswerId":"42622673","CreationDate":"2017-03-06T10:04:34.313","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>To change something like this you need to recompile the model. This is because the compile process generates the training function to be minimized afterwards. To change the training function you need to compile again.<\/p>\n\n<p>Just call<\/p>\n\n<pre><code>model.compile(...)\n<\/code><\/pre>\n\n<p>after your changes and it should work the way you expect it to work.<\/p>\n"}]}
{"QuestionId":42622444,"AnswerCount":4,"Tags":"<python><scikit-learn><neural-network><keras><pipeline>","CreationDate":"2017-03-06T09:54:02.780","AcceptedAnswerId":42653531.0,"OwnerUserId":4680896.0,"Title":"Can't save a Pipeline estimator","Body":"<p>I am trying to train a simple pipeline:<\/p>\n\n<pre><\/pre>\n\n<p>where deepnc is a Keras classifier:<\/p>\n\n<pre><\/pre>\n\n<p>I put the pipeline into a random seearch and to check some parameters:<\/p>\n\n<pre><\/pre>\n\n<p>Afterwards, I wanted to save the best estimator and best parameters:<\/p>\n\n<pre><\/pre>\n\n<p>For some reason, the former did not work. Fortunately, I run the script in the console, so I was able to run the latter separately and save the optimal parameters. However, I am still trying to figure out how to save the model. I guess combining scikit wrapper for Keras with Pipeline and RandomizedSearchCV caused the problem?<\/p>\n\n<p>I also tried this code:<\/p>\n\n<pre><\/pre>\n\n<p>but it generated the same error traceback. I post below a shortened version, as it was super long and consisted of repetitions of the same fragment over and over again. Googling the error did not help. Any ideas?<\/p>\n\n<pre><\/pre>\n\n<p>Bonus question: Am I using the randomized search correctly? I didn't get much improvement compared to my initial effort.<\/p>\n","answers":[{"AnswerId":"42653531","CreationDate":"2017-03-07T16:35:42.787","ParentId":null,"OwnerUserId":"4680896","Title":null,"Body":"<p>Saving a Keras model wrapped inside of a Scikit classifier is not possible <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/4274\" rel=\"nofollow noreferrer\">GitHub Reference<\/a>.<\/p>\n\n<p>However, there is a workaround (or 2)! One is, as <a href=\"https:\/\/stackoverflow.com\/a\/42632560\/4680896\">Nassim Ben points out<\/a>, to use ModelCheckpoint to save best weights and \"recreate\" a model.<\/p>\n\n<p>In my circumstances, after doing some research, I decided to do it differently, arguably easier: after training, simply do:<\/p>\n\n<pre><code>grid.best_estimator_.model.save(path)\n<\/code><\/pre>\n\n<p>This (i.e. adding .model, before saving) makes sure to access the underlying Keras model, for which the save method works correctly. Now you can simply do<\/p>\n\n<pre><code>deepnc_cont = keras.models.load_model(path)\n<\/code><\/pre>\n\n<p>And it works - at least for me :)<\/p>\n\n<p>Note that if, for some reason, I needed to have a KerasClassifier object (the scikit wrapper), it would not work, as its contructor asks for a function that generates the model, so I would probably have to pursue Nassim's route? I'm not sure though.<\/p>\n"},{"AnswerId":"42632560","CreationDate":"2017-03-06T18:05:41.883","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>I know this is not exactly what you were looking for but... Here comes nothing :<\/p>\n\n<p>Given that your callback object has saved the best weights, you only need the model object. The model is in your function <code>create_spec_model()<\/code>. The only thing you need is the best parameters. So :<\/p>\n\n<pre><code># Save parameters of the best estimator.\npickle.dump(grid.best_estimator_.named_steps['deepnc'].get_params(),open('params.pkl','wb'))\n<\/code><\/pre>\n\n<p>At loading time, given that your function <code>create_spec_model()<\/code> is still in your code :<\/p>\n\n<pre><code>import inspect\n\ndef load_model(params_path, weights_path):\n    params = pickle.load(open(params_path,'rb'))\n    params = {k: params[k] for k in inspect.getargspec(create_spec_model)[0] if k in params.keys()}\n    model = create_spec_model(**params)\n    model.load_weights(weights_path)\n    return model\n<\/code><\/pre>\n\n<p>does that help you?<\/p>\n"},{"AnswerId":"43200447","CreationDate":"2017-04-04T07:04:31.490","ParentId":null,"OwnerUserId":"1643257","Title":null,"Body":"<p>Just adding another workaround for the Keras-Pickle problem.<\/p>\n\n<p>It's a bit patchy, but allows Keras models to be stored with <code>pickle<\/code> so  it works for both the Model and containing objects.<\/p>\n\n<p><a href=\"http:\/\/zachmoshe.com\/2017\/04\/03\/pickling-keras-models.html\" rel=\"nofollow noreferrer\">http:\/\/zachmoshe.com\/2017\/04\/03\/pickling-keras-models.html<\/a><\/p>\n"},{"AnswerId":"42625722","CreationDate":"2017-03-06T12:30:20.577","ParentId":null,"OwnerUserId":"969289","Title":null,"Body":"<p>You cannot save a keras model using joblib or pickle.\nInstead use the save method and shown <a href=\"https:\/\/keras.io\/getting-started\/faq\/#how-can-i-save-a-keras-model\" rel=\"nofollow noreferrer\" title=\"here\">here<\/a>.<\/p>\n\n<p>or in your case:<\/p>\n\n<pre><code>path = 'models\/deepn_spec_model.pkl'\ngrid.best_estimator_.save(path)\n<\/code><\/pre>\n\n<p>and to load the model:<\/p>\n\n<pre><code>from keras.models import load_model\npath = 'models\/deepn_spec_model.pkl'\nmodel = load_model(path)\n<\/code><\/pre>\n\n<p>Also note that since this is not strictly speaking a pickle file but a HDF5 file, you might as well change the file extension to '.h5'.<\/p>\n"}]}
{"QuestionId":42622815,"AnswerCount":1,"Tags":"<python><numpy><tensorflow>","CreationDate":"2017-03-06T10:10:40.563","AcceptedAnswerId":42629665.0,"OwnerUserId":5257450.0,"Title":"How can I use the index array in tensorflow?","Body":"<p>If given a matrix  with shape  and index array  with shape , we can easily get the corresponding vector  through,<\/p>\n\n<p><\/p>\n\n<p>However, I cannot do the same thing with tensorflow,<\/p>\n\n<pre><\/pre>\n\n<blockquote>\n  <p>Traceback (most recent call last):   File \"\", line 1, in\n     File\n  \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/array_ops.py\",\n  line 513, in _SliceHelper\n      name=name)<\/p>\n  \n  <p>File\n  \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/array_ops.py\",\n  line 671, in strided_slice\n      shrink_axis_mask=shrink_axis_mask)   File \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/ops\/gen_array_ops.py\",\n  line 3688, in strided_slice\n      shrink_axis_mask=shrink_axis_mask, name=name)   File \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/op_def_library.py\",\n  line 763, in apply_op\n      op_def=op_def)   File \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 2397, in create_op\n      set_shapes_for_outputs(ret)   File \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 1757, in set_shapes_for_outputs\n      shapes = shape_func(op)   File \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/ops.py\",\n  line 1707, in call_with_requiring\n      return call_cpp_shape_fn(op, require_shape_fn=True)   File \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/common_shapes.py\",\n  line 610, in call_cpp_shape_fn\n      debug_python_shape_fn, require_shape_fn)   File \"~\/anaconda2\/lib\/python2.7\/site-packages\/tensorflow\/python\/framework\/common_shapes.py\",\n  line 675, in _call_cpp_shape_fn_impl\n      raise ValueError(err.message) ValueError: Shape must be rank 1 but is rank 2 for 'strided_slice_14' (op: 'StridedSlice') with input\n  shapes: [5,3], [2,5], [2,5], [2].<\/p>\n<\/blockquote>\n\n<p>My question is, if I cannot produce the expected result in tensorflow as in numpy using the above mentioned method, what should I do?<\/p>\n","answers":[{"AnswerId":"42629665","CreationDate":"2017-03-06T15:38:27.700","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>This feature is not currently implemented in TensorFlow. GitHub <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/4638\" rel=\"noreferrer\">issue #4638<\/a> is tracking the implementation of NumPy-style \"advanced\" indexing. However, you can use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/gather_nd\" rel=\"noreferrer\"><code>tf.gather_nd()<\/code><\/a> operator to implement your program:<\/p>\n\n<pre><code>a = tf.placeholder(tf.float32, shape=(5, 3))\nb = tf.placeholder(tf.int32, (5,))\n\nrow_indices = tf.range(5)\n\n# `indices` is a 5 x 2 matrix of coordinates into `a`.\nindices = tf.transpose([row_indices, b])\n\nc = tf.gather_nd(a, indices)\n<\/code><\/pre>\n"}]}
{"QuestionId":42623090,"AnswerCount":0,"Tags":"<tensorflow><grpc><tensorflow-serving><syntaxnet>","CreationDate":"2017-03-06T10:23:51.787","AcceptedAnswerId":null,"OwnerUserId":4453606.0,"Title":"Using Syntaxnet with TensorFlow Serving","Body":"<p>I try to use Syntaxnet with TensorFlow Serving.<\/p>\n\n<p>I found this link <a href=\"https:\/\/github.com\/dsindex\/syntaxnet\/blob\/master\/README_api.md\" rel=\"nofollow noreferrer\">here<\/a> . I try to run the script to launch the server:<\/p>\n\n<pre><\/pre>\n\n<p>But for the line below (when building parsey_api):<\/p>\n\n<pre><\/pre>\n\n<p>I got an error: <\/p>\n\n<pre><\/pre>\n\n<p>The  is in the directory <\/p>\n\n<p>I am using:<\/p>\n\n<ul>\n<li>Ubuntu 14.04   <\/li>\n<li>Bazel 0.4.4<\/li>\n<li>python gPRC installed<\/li>\n<\/ul>\n\n<p>Thank you any suggestion is welcome.<\/p>\n","answers":[]}
{"QuestionId":42623704,"AnswerCount":0,"Tags":"<tensorflow><sparse-matrix><matrix-factorization>","CreationDate":"2017-03-06T10:52:56.733","AcceptedAnswerId":null,"OwnerUserId":4584564.0,"Title":"In TensorFlow, how to get all nonzeros of a SparseTensor row","Body":"<p>I need to be able to get all nonzeros with their indices and values of the nonzeros values. Have somebody done this or have any idea how please help me to get on the right track. <\/p>\n\n<p>Is there a possibility that i can get all elements of a row in a SparseTensor (i know that there may not be any saved values for a row and therefore it returns null or none)<\/p>\n","answers":[]}
{"QuestionId":42623978,"AnswerCount":2,"Tags":"<python><tensorflow><neural-network><deep-learning><conv-neural-network>","CreationDate":"2017-03-06T11:05:36.920","AcceptedAnswerId":42635945.0,"OwnerUserId":4359558.0,"Title":"print saved weights of particular layer [Tensorflow]","Body":"<p>i'm using Tensorflow 1.0 with python 3.5 in PyCharm<BR>\nafter executing <a href=\"https:\/\/gist.github.com\/dandelionmane\/4f02ab8f1451e276fea1f165a20336f1#file-mnist-py\" rel=\"nofollow noreferrer\">this<\/a> code, i got model(index, meta and ckpt file) saved at every 500th iteration. now to load the model which file do we need to point?<Br>\ni wrote below code to load the ckpt(weights) file (with no changes made to above github code)<\/p>\n\n<pre><\/pre>\n\n<p>got following error<\/p>\n\n<blockquote>\n  <p>W\n  c:\\tf_jenkins\\home\\workspace\\release-win\\device\\cpu\\os\\windows\\tensorflow\\core\\framework\\op_kernel.cc:975] Not found: Unsuccessful TensorSliceReader constructor: Failed to find any matching files for D:\/tmp\/mnist_tutorial\/model.ckpt<Br>\n  Traceback (most recent call last):<BR>\n  File \"C:\\Users\\Admin\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1021, in _do_call\n      return fn(*args)<BR>\n    File \"C:\\Users\\Admin\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\client\\session.py\", line 1003, in _run_fn\n      status, run_metadata)<BR>\n    File \"C:\\Users\\Admin\\AppData\\Local\\Programs\\Python\\Python35\\lib\\contextlib.py\", line 66, in <strong>exit<\/strong>\n      next(self.gen)<Br>\n    File \"C:\\Users\\Admin\\AppData\\Local\\Programs\\Python\\Python35\\lib\\site-packages\\tensorflow\\python\\framework\\errors_impl.py\", line 469, in raise_exception_on_not_ok_status<Br> <\/p>\n  \n  <p>pywrap_tensorflow.TF_GetCode(status))\n  tensorflow.python.framework.errors_impl.NotFoundError: Unsuccessful TensorSliceReader constructor: Failed to find any matching files for D:\/tmp\/mnist_tutorial\/model.ckpt<BR>\n       [[Node: save\/RestoreV2_1 = RestoreV2[dtypes=[DT_FLOAT], _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"](_recv_save\/Const_0, save\/RestoreV2_1\/tensor_names, save\/RestoreV2_1\/shape_and_slices)]]\n  <Br>\n  <Br><\/p>\n<\/blockquote>\n\n<p>before training starts (in function definition of convolution layer), we can print weights of individual layer by :<\/p>\n\n<pre><\/pre>\n\n<p>if there many convolution and fully connected layers, how to load and print any particular layer's weights and biases from  file, as the above method is not working\n<Br><\/p>\n\n<p><Br>\n<strong>Update:<\/strong> made changes to code and updated error message<\/p>\n","answers":[{"AnswerId":"45407303","CreationDate":"2017-07-31T04:05:55.823","ParentId":null,"OwnerUserId":"8322452","Title":null,"Body":"<p>I had almost the same problem as yours, but I just resolved it by adding the model and its index file to the root directory same as the directory where the **.py you will run.\nthen you will meet some problem like \"UnicodeDecodeError: 'ascii' codec can't decode byte 0xe3 in position 37: ordinal not in range(128)\",that will be easily corrected.<\/p>\n"},{"AnswerId":"42635945","CreationDate":"2017-03-06T21:31:21.823","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Look at the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/python\/tools\/inspect_checkpoint.py\" rel=\"nofollow noreferrer\">inspect_checkpoint tool<\/a> in the tensorflow repository to read checkpoints.<\/p>\n"}]}
{"QuestionId":42624843,"AnswerCount":1,"Tags":"<neural-network><deep-learning><caffe>","CreationDate":"2017-03-06T11:46:26.883","AcceptedAnswerId":42624981.0,"OwnerUserId":3423045.0,"Title":"Values inside filters in caffe","Body":"<p>I am trying to understand the following part in caffe network model.<\/p>\n\n<pre><\/pre>\n\n<p>What I understood is there are 256 filters used in this layer.\nI want to know how the values inside those filters are selected ?\nUsing size 5x5 and by std dev 0.01 we can create one filter, how other filters are created ?<\/p>\n","answers":[{"AnswerId":"42624981","CreationDate":"2017-03-06T11:53:42.730","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>Depending on the input dimension to this layer (the \"channel\" shape) this layer has 256 filters of shape <code>in-dim<\/code>-by-<code>5<\/code>-by-<code>5<\/code>. Caffe init all these values (according to <code>weight_filler<\/code> param) with i.i.d random samples from a Gaussian (normal) distribution with zero mean and std=0.01.<\/p>\n\n<p>You can see the values in python (assuming the layer name is <code>\"conv1\"<\/code>):<\/p>\n\n<pre><code>import caffe\nnet = caffe.Net('\/path\/to\/net.prototxt', caffe.TEST)\nlayer_idx = list(net._layer_names).index('conv1')\nweights = net.layers[layer_idx].blobs[0].data\nprint \"filter values =\", weights\n<\/code><\/pre>\n"}]}
{"QuestionId":42625657,"AnswerCount":1,"Tags":"<tensorflow><optimus>","CreationDate":"2017-03-06T12:27:12.040","AcceptedAnswerId":null,"OwnerUserId":4453772.0,"Title":"tensorflow does not use gpu, but cuda does","Body":"<p>tensorflow can not see my GPU. I am using a optimus setup.<\/p>\n\n<p>nvidia-smi shows my card<\/p>\n\n<pre><\/pre>\n\n<p>cuda sees the gpu. here is the deviceQuery output<\/p>\n\n<pre><\/pre>\n\n<p>but tensorflow does not use the gpu<\/p>\n\n<pre><\/pre>\n\n<p>the output seems to indicate, that only CPU is used<\/p>\n\n<pre><\/pre>\n\n<p>so, what can i do, that tensorflow sees my gpu? I am using archlinux, i assume that i have from everything the newest version. are there things, that i could check?<\/p>\n","answers":[{"AnswerId":"42733611","CreationDate":"2017-03-11T09:35:55.853","ParentId":null,"OwnerUserId":"719457","Title":null,"Body":"<p>Official miniamal CUDA Compute Capability is 3.5. Your card has 3.0. It is said that some guys can compile tensorflow to use 3.0 CC but it requires patching TF with unofficial patches. See more: <a href=\"https:\/\/stackoverflow.com\/questions\/33651810\/the-minimum-required-cuda-capability-is-3-5#33686846\">The minimum required Cuda capability is 3.5<\/a>.<\/p>\n"}]}
{"QuestionId":42625739,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-06T12:31:25.557","AcceptedAnswerId":42649115.0,"OwnerUserId":6828367.0,"Title":"tensorflow - tf.where TypeError","Body":"<p>I wrote the following test code (a part of a much bigger code) <\/p>\n\n<pre><\/pre>\n\n<p>I get the following error <\/p>\n\n<pre><\/pre>\n\n<p>how can I fix this? <\/p>\n\n<p>what I want to do is create a list\\tensor of operators (which are optimizer\\do nothing) given a certain condition (not using  because i want to apply this to a batch so I have a list\\tensor of booleans) \n* working in tensorflow 1.0<\/p>\n","answers":[{"AnswerId":"42649115","CreationDate":"2017-03-07T13:05:31.233","ParentId":null,"OwnerUserId":"7615953","Title":null,"Body":"<p><code>tf.train.AdamOptimizer(0.1)<\/code> returns an Optimizer object, it is not clear how this relates to a no_op. I suggest starting by creating the optimizer and then conditioning on gradient updates instead, for instance by zeroing gradient contributions from the part of the batch wanted. <code>tf.where<\/code> does not work with operations, it provides indices for accessing tensors.<\/p>\n"}]}
{"QuestionId":42626704,"AnswerCount":1,"Tags":"<machine-learning><neural-network><runtime-error><deep-learning><keras>","CreationDate":"2017-03-06T13:19:12.507","AcceptedAnswerId":null,"OwnerUserId":7665933.0,"Title":"Keras RuntimeError: GpuCorrMM failed to allocate working memory of 576 x 802816","Body":"<p>I tryed to run a deep learning code in Keras but got following error message all the time. I've searched all around and spend much time but still failed to fix it up. I'm a fish, any help will be appreciated!!<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42626981","CreationDate":"2017-03-06T13:32:44.587","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The problem lies in memory. Your first ten <code>conv2D<\/code> layers need approximately <code>200MB<\/code> each. This means that it's need <code>2GB<\/code> for only storing output of your first 10 layers. This would surely not fit into your cards memory. <\/p>\n"}]}
{"QuestionId":42626882,"AnswerCount":1,"Tags":"<tensorflow><svm>","CreationDate":"2017-03-06T13:27:21.927","AcceptedAnswerId":null,"OwnerUserId":7614637.0,"Title":"Does Tensorflow have SVR packages like sklearn.svm.SVR in scikit-learn?","Body":"<p>I am new to Tensorflow, I wonder whether Tensorflow has ready-to-use SVR packages like sklearn.svm.SVR in scikit-learn. Or I have to build the SVR on my own?<\/p>\n","answers":[{"AnswerId":"43057878","CreationDate":"2017-03-27T23:05:25.997","ParentId":null,"OwnerUserId":"7776776","Title":null,"Body":"<p>Tensorflow now seems to incorporate <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/tree\/master\/tensorflow\/contrib\/learn\/python\/learn\" rel=\"nofollow noreferrer\">this<\/a><\/p>\n\n<p>I'm stilling looking into it as I have a SKLearn SVR that I would really like to have access to more processing power via tensorflow. I'll let you know if I figure this out. <\/p>\n\n<p><strong>Edit:<\/strong> Also Found <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/contrib\/learn\/python\/learn\/estimators\/g3doc\/svm.md\" rel=\"nofollow noreferrer\">this little nugget of info<\/a><\/p>\n"}]}
{"QuestionId":42627294,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-03-06T13:47:36.673","AcceptedAnswerId":null,"OwnerUserId":7567502.0,"Title":"Caffe - How to imbalance Cifar10 data","Body":"<p>I'm doing a research on the impact of imbalanced data with caffe framework. Now I am trying to make a new cifar10 distribution by trying to remove some of the data from specified class. I read the document of cifar10. It said that the .bin file has a data structure like <\/p>\n\n<pre><\/pre>\n\n<p>So I write down a script to filter out the data for those class. And make a new .bin file.<\/p>\n\n<p>Now I run the script on caffe and try to make a LMDB dataset<\/p>\n\n<pre><\/pre>\n\n<p>However After I filter out those data. It seems that the IMDB still has the same size and doesn't look any different than the one without filtered. Can somebody tell me what should I do to make the data imbalanced?<\/p>\n","answers":[]}
{"QuestionId":42627607,"AnswerCount":1,"Tags":"<machine-learning><neural-network><computer-vision><deep-learning><caffe>","CreationDate":"2017-03-06T14:04:03.423","AcceptedAnswerId":42633449.0,"OwnerUserId":1014196.0,"Title":"Caffe's way of representing negative examples on benchmark dataset for binary classification","Body":"<p>I would like to know how to define or represent a negative training set if I would want to train a binary classifier from a pre-trained model say, AlexNet on ILSVRC12 (or ImageNet) dataset. What I am currently thinking of is to take one the classes which is not related as the negative training set while the one which is related as positive one. Is there any better way which is more elegant?<\/p>\n","answers":[{"AnswerId":"42633449","CreationDate":"2017-03-06T18:58:04.233","ParentId":null,"OwnerUserId":"4785185","Title":null,"Body":"<p>The CNNs trained on the ILSVRC data set are already discriminating among 1000 classes of images.  Yes, you can use one of those topologies to train a binary classifier, but I suggest that you start with an untrained model and run it through your two chosen classes.  If you start with a trained model, you have to <em>unlearn<\/em> a lot, and your result is still trying to discriminate among 1000 classes: that last FC layer is going to give you trouble.<\/p>\n\n<p>There are ways to work around the 1000-class problem.  If your application already overlaps one or more of the trained classes, then simply add a layer that maps those classes to label \"1\" and all the others to label \"0\".<\/p>\n\n<p>If you're insistent on retaining the trained kernels, then try replacing the final FC layer (1000) with a 2-class FC layer.  Then choose your two classes (applicable images vs everything else) and run your training.<\/p>\n"}]}
{"QuestionId":42627909,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><keras-layer>","CreationDate":"2017-03-06T14:18:22.383","AcceptedAnswerId":42635717.0,"OwnerUserId":2182857.0,"Title":"Transfer Learning fails because dense layer is expected to have shape (None,1)","Body":"<p>I'm trying to use  for some classfication problem. Before using it on the problem I'm trying to experiment with it.<\/p>\n\n<p>I replaced the last dense layer (sized ) with a new dense layer, compiled the model and tried to fit it<\/p>\n\n<pre><\/pre>\n\n<p>But I am getting the following error<\/p>\n\n<pre><\/pre>\n\n<p>What am I doing wrong?\nAnd why is the newDense layer expected to have a  shape after I defined it as having the size of 3?<\/p>\n\n<p>Many thanks     <\/p>\n\n<p>PS\nI am adding the end of summary of the model <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42635717","CreationDate":"2017-03-06T21:16:44.887","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Ok, problem lied in <\/p>\n\n<pre><code>validation_generator = val_datagen.flow_from_directory(...\n        class_mode = 'categorical')\n<\/code><\/pre>\n\n<p><code>Categorical<\/code> makes your generator to return a one-hot encoded vector. In your case a <code>3-d<\/code> one. But you set your <code>loss<\/code> to <code>sparse_categorical_crossentropy<\/code> which accepts <code>int<\/code> as a label. You should either change <code>class_mode=\"sparse\"<\/code> or <code>loss=\"categorical_crossentropy\"<\/code>. <\/p>\n"}]}
{"QuestionId":42628141,"AnswerCount":1,"Tags":"<python-3.x><machine-learning><tensorflow>","CreationDate":"2017-03-06T14:28:31.233","AcceptedAnswerId":42628629.0,"OwnerUserId":5540592.0,"Title":"AttributeError: module 'tensorflow' has no attribute 'streaming_accuracy'","Body":"<pre><\/pre>\n\n<p>For the above code, I have received the same error in past few days. \nIsn't the syntax same as it is in the API documentation for tensorflow?<\/p>\n","answers":[{"AnswerId":"42628629","CreationDate":"2017-03-06T14:51:49.687","ParentId":null,"OwnerUserId":"3027854","Title":null,"Body":"<p>try to use this instead (in a fresh python file, I would suggest create a <code>\/tmp\/temp.py<\/code> and run that)<\/p>\n\n<pre><code>from tensorflow.contrib.metrics import streaming_accuracy\n<\/code><\/pre>\n\n<p>and if this doesn't work then <\/p>\n\n<ol>\n<li>either there is an installation problem (In which case reinstall) <\/li>\n<li>or you are importing the wrong tensorflow module.<\/li>\n<\/ol>\n"}]}
{"QuestionId":42628143,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-06T14:28:36.730","AcceptedAnswerId":42633142.0,"OwnerUserId":3015202.0,"Title":"Does Tensorflow simplify a computational graph?","Body":"<p>I have a simple question and I was also searching already quiet a bit, but maybe I'm using the wrong keywords.<\/p>\n\n<p>How does Tensorflow handle a given graph? If one has the simple graph:<\/p>\n\n<blockquote>\n  <p>x = tf.constant(1.0, name='input')<\/p>\n  \n  <p>w = tf.constant0.8, name='weight')<\/p>\n  \n  <p>b = tf.constant0.8, name='bias')<\/p>\n  \n  <p>y_1 = tf.mul(w, x, name='output_1')<\/p>\n  \n  <p>y_2 = tf.add(y_1, b, name='output_1')<\/p>\n<\/blockquote>\n\n<p>The arithmetic statement is of course given by the computational graph, but is Tensorflow then kind of compiling and simplifying it in terms of saving time by not copying memories, etc.? So that it a 'condensed' version of the computational kernel is executed on the 'device' like CPU or GPU?<\/p>\n\n<p>So that it reduces to something like that:<\/p>\n\n<blockquote>\n  <p>y_2 = tf.add(tf.mul(w, x), b, name='output_1')<\/p>\n<\/blockquote>\n\n<p>Maybe somebody knows a good resource to learn more about how exactly Tensorflow runs under the hood without looking too deep into the source-code. <\/p>\n\n<p>Thank you very much in advance! <\/p>\n","answers":[{"AnswerId":"42633142","CreationDate":"2017-03-06T18:39:57.103","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow includes various optimizations that can have the effect of simplifying a dataflow graph. In particular:<\/p>\n\n<ul>\n<li><p>TensorFlow will apply <a href=\"https:\/\/en.wikipedia.org\/wiki\/Common_subexpression_elimination\" rel=\"noreferrer\">common subexpression elimination<\/a> to avoid performing redundant computation. In the case of your example, this will not have much effect, but TensorFlow will observe that <code>w<\/code> and <code>b<\/code> are the same constant, and replace them with a single value.<\/p><\/li>\n<li><p>TensorFlow will apply <a href=\"https:\/\/en.wikipedia.org\/wiki\/Constant_folding\" rel=\"noreferrer\">constant propagation<\/a> so that (computed) values that are the same in every execution of a subgraph will only be computed once. In your example, the entire expression is a constant, so TensorFlow will replace it with a single <code>tf.constant()<\/code> value corresponding to the result (<code>1.6<\/code>).<\/p><\/li>\n<li><p>If you use the experimental <a href=\"https:\/\/www.tensorflow.org\/performance\/xla\/\" rel=\"noreferrer\">XLA compiler<\/a>, TensorFlow will make more aggressive simplifications, and may be able to replace a subgraph with a single TensorFlow kernel, containing just-in-time compiled code. If in your example <code>x<\/code> were a <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/placeholder\" rel=\"noreferrer\"><code>tf.placeholder()<\/code><\/a>, the remainder of the computation could be compiled into a single kernel with one input and one output.<\/p><\/li>\n<\/ul>\n"}]}
{"QuestionId":42628619,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning><one-hot-encoding>","CreationDate":"2017-03-06T14:51:13.787","AcceptedAnswerId":null,"OwnerUserId":1496362.0,"Title":"Gettin TensorFlow to work with distributed binary representation","Body":"<p>I know how to create an rnn in TensorFlow with a one_hot vector: <\/p>\n\n<pre><\/pre>\n\n<p>But I am not really sure what to do when my input vector has multiple 1s, eg. it could be 11011 as 1 input per time. so: [[11011],[00111],...]<\/p>\n\n<p>Is there an issue if I would just feed this vector like I would have my one-hot representation? How should I formulate the above then? I feel like I shouldn't use the tf.one_hot function... Not sure how the shape of rnn_inputs (200 x 5 x 2) can be created without one_hot. <\/p>\n\n<p>(using TF 1.0)<\/p>\n","answers":[]}
{"QuestionId":42629530,"AnswerCount":1,"Tags":"<python><deep-learning><keras><lstm><recurrent-neural-network>","CreationDate":"2017-03-06T15:32:22.490","AcceptedAnswerId":null,"OwnerUserId":6115356.0,"Title":"Obtain output at each timestep in Keras LSTM","Body":"<p>What I want to do is give the output of the LSTM at the current timestep as input to the LSTM in the next timestep. So I want the LSTM to predict the word at at the current time step and give this word as input to the next timestep. Is this can be done then:<\/p>\n\n<p>How can the input and target data be specified during training i.e. in the  function?<\/p>\n","answers":[{"AnswerId":"42630219","CreationDate":"2017-03-06T16:05:23.783","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>You cannot do it directly in <code>keras<\/code> but you could do it using a <code>for<\/code> loop and <code>stateful<\/code> network. This works like this (assuming that you have your sentences stored as sequences of integers with <code>size=vocabulary_size<\/code>:<\/p>\n\n<ol>\n<li><p>Define a stateful network which accepts one word and returns the following one:<\/p>\n\n<pre><code>model = Input(batch_size=(batch_size, 1, 1)) \nmodel = Embedding(size_of_vocabulary, output_size, input_length=1)(input)\nmodel = LSTM(lstm_1st_layer_size, return_sequences=True, stateful=True)(model)\n....\nmodel = LSTM(lstm_nth_layer_size, return_sequences=True, stateful=True)(model)\nmodel = Dense(vocabulary_size, activation=\"softmax\")\n\nmodel.compile(loss=\"sparse_categorical_crossentropy\", optimizer=\"rmsprop\")\n<\/code><\/pre><\/li>\n<li><p>Assuming that you have a numpy <code>array_of_samples<\/code> of examples <code>(preceding_word, next_word)<\/code> you could fit it by:<\/p>\n\n<pre><code>model.fit(array_of_samples[:,0], array_of_samples[:,1])\n<\/code><\/pre><\/li>\n<li><p>Now you could try to predict stuff in a following manner:<\/p>\n\n<pre><code>sentence = [starting_word]\nfor i in range(len_of_sequence - 1):\n    sentence.append(model.predict(numpy.array([[sentence[i]]).argmax())\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>Now <code>sentence<\/code> stores your newly created sentence<\/p>\n"}]}
{"QuestionId":42630504,"AnswerCount":1,"Tags":"<python><lambda><machine-learning><keras><autoencoder>","CreationDate":"2017-03-06T16:20:08.783","AcceptedAnswerId":42631463.0,"OwnerUserId":6803651.0,"Title":"Keras Lambda layer has no output tensor shape, error when compiling model","Body":"<p>I'm trying to build a character-level auto-encoder for text sequences using Keras. When I compile the model, I get an error concerning the shapes of the tensors as you can see bellow. I printed out the layers specifications to check, if the tensor shapes match up and I found that the problem is probably with the last Lambda layer not having an <em>Output Tensor Shape<\/em> specified correctly, but I can't figure out why not or how to specify it and didn't find anything about it in the documentation of Keras or on Google.<\/p>\n\n<p>Below the error output is also the part of the code, where I define my model. The whole script for clarification, if needed, is here: <a href=\"http:\/\/pastebin.com\/gDvyZQk6\" rel=\"nofollow noreferrer\">PasteBin<\/a>.<\/p>\n\n<h2>Error and layer output<\/h2>\n\n<p>(Mind mainly the last layer.)<\/p>\n\n<pre><\/pre>\n\n<h1>Code<\/h1>\n\n<p>I've built my model using this code:<\/p>\n\n<pre><\/pre>\n\n<p>Firstly, I input a text sequence of maximum 80 chars, the Lambda layer transforms each character to a one-hot vector. At the end, I would like to transform the one-hot vectors back, while taking only the maximum as the decoded character.<\/p>\n\n<h1>Question<\/h1>\n\n<ol>\n<li>Am I missing something in the model construction code?<\/li>\n<li>How can I make the Lambda layer know the output tensor shape?<\/li>\n<\/ol>\n\n<h1>Edit<\/h1>\n\n<p>As <a href=\"https:\/\/stackoverflow.com\/users\/7137636\/nassim-ben\">Nassim Ben<\/a> pointed out, the problem was with the function <em>debinarize<\/em>. After changing it to:<\/p>\n\n<pre><\/pre>\n\n<p>at least some kind of value was set to the Output tensor's shape. Though the value is a bit weird, because it is <strong>(80, 256)<\/strong> and differs from the output shape <strong>(None, 80)<\/strong>. All the other output tensor shapes and output shapes match up correspondingly (I suppose the '?' and None mean the same more or less...). More specifically, the Lambda layer now looks like this:<\/p>\n\n<pre><\/pre>\n\n<p>Problem is, I would like output tensor shape to be (?, 80) as is the first layer's input. I did not change any other code than the <em>argmax<\/em>.<\/p>\n\n<p>The error given is now:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42631463","CreationDate":"2017-03-06T17:07:15.563","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>I think it comes from using a numpy function on a tensor.\nTry using tf argmax function (I think the axis you want to reduce is 1, not sure)<\/p>\n\n<pre><code>def debinarize(x):\n    return tf.to_float(tf.argmax(x,axis=2))  # get the character with most probability\n<\/code><\/pre>\n\n<p>Does this work?<\/p>\n"}]}
{"QuestionId":42630738,"AnswerCount":2,"Tags":"<deep-learning><caffe><pycaffe><matcaffe>","CreationDate":"2017-03-06T16:30:43.933","AcceptedAnswerId":null,"OwnerUserId":6494707.0,"Title":"I am getting a black image in FCN32","Body":"<p>I trained FCN32 from the scratch on my data, unfortunately I am getting a black image as output. Here is the loss curve.\n<a href=\"https:\/\/i.stack.imgur.com\/mJAya.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/mJAya.png\" alt=\"enter image description here\"><\/a>\nI am not sure whether this training loss curve is normal or not, or whether I have done something wrong or not.<\/p>\n\n<p>I really appreciate experts'idea on this. And <\/p>\n\n<ol>\n<li>why the output is a black image?<\/li>\n<li>Is the network overfitting?<\/li>\n<li>Should I change  value in  layer, from \nto any other value?\nThanks a lot<\/li>\n<\/ol>\n\n<p><strong>Edited:<\/strong>\nI changed the  value in  layer, from \n    to  \nand the following shows the :<\/p>\n\n<pre><\/pre>\n\n<p>I got the following train-loss curve and again I am getting black image. I do not know what is the mistake and why it is behaving like this, could someone please share some ideas? Thanks\n<a href=\"https:\/\/i.stack.imgur.com\/QJERL.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QJERL.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42633587","CreationDate":"2017-03-06T19:04:57.787","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>There is an easy way to check if you are overfitting on the training data or just did something wrong in the algorithm. Just predict on the training data and look at the output. If this is very similar or equal to the desired output you are overfitting and you will probably have to apply dropout and weight regularization. <\/p>\n\n<p>If the output is also black on the training data your labels or your optimization metric is probably wrong. <\/p>\n"},{"AnswerId":"42636344","CreationDate":"2017-03-06T21:54:38.870","ParentId":null,"OwnerUserId":"562769","Title":null,"Body":"<blockquote>\n  <p>Should I change lr_mult value in Deconvolution layer, from 0 to any other value? <\/p>\n<\/blockquote>\n\n<p><code>lr_mult = 0<\/code> means this layer does not learn (<a href=\"https:\/\/stackoverflow.com\/a\/38598980\/562769\">source<\/a>, <a href=\"https:\/\/groups.google.com\/forum\/#!topic\/caffe-users\/8J_J8tc1ZHc\" rel=\"nofollow noreferrer\">source 2<\/a>). If you want that layer to learn, you should better set it to a positive value. Depending on your initialization, this might very well be the reason why the image is black.<\/p>\n"}]}
{"QuestionId":42630762,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-03-06T16:31:51.547","AcceptedAnswerId":null,"OwnerUserId":5434466.0,"Title":"How to verify Tensorflow Serving is using GPUs on a GPU instance?","Body":"<p>While running Tensorflow Serving how to verify it uses GPUs for serving?\nConfigured Tensorflow to use GPUs during .\/configure.<\/p>\n\n<p>Tried monitoring nvidia-smi while running inference it shows no running process found.<\/p>\n","answers":[{"AnswerId":"46074198","CreationDate":"2017-09-06T11:37:14.580","ParentId":null,"OwnerUserId":"6857147","Title":null,"Body":"<p>First\uff0c of course you need to configure to use cuda when .\/configure<\/p>\n\n<p>Second, you should compile tf serving using<\/p>\n\n<p><code>bazel build -c opt --config=cuda tensorflow\/...<\/code><\/p>\n\n<p>and<\/p>\n\n<p><code>bazel build -c opt --config=cuda --spawn_strategy=standalone \/\/tensorflow_serving\/model_servers:tensorflow_model_server<\/code><\/p>\n\n<p>Lastly you can see the information when you serve the model if serving with GPU: <\/p>\n\n<blockquote>\n  <p>I\n  external\/org_tensorflow\/tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:965]\n  Found device 0 with properties: name: GeForce GTX 1070 major: 6 minor:\n  1 memoryClockRate(GHz): 1.721 pciBusID: 0000:01:00.0 totalMemory:\n  7.92GiB freeMemory: 7.76GiB\n  I external\/org_tensorflow\/tensorflow\/core\/common_runtime\/gpu\/gpu_device.cc:1055]\n  Creating TensorFlow device (\/device:GPU:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0, compute capability: 6.1)<\/p>\n<\/blockquote>\n\n<p>and check the nvidia-smi at the same time<\/p>\n\n<blockquote>\n  <p>+----------------------------------------------------------------------------------------------------------------------------+  | Processes:                                                   ______________________________________________________GPU Memory |<br>\n  |  GPU      _______ PID ______ Type _______ Process name  _______________________  Usage      | \n  |======================================================================---| \n  |    0  _________1215________G ________\/usr\/lib\/xorg\/Xorg_______________________59MiB |  |    0 _________     7341_______    C  ___ ...ing\/model_servers\/tensorflow_model_server __7653MiB | \n  +----------------------------------------------------------------------------------------------------------------------------+<\/p>\n<\/blockquote>\n\n<p>GPU is consumed a lot.<\/p>\n"}]}
{"QuestionId":42631169,"AnswerCount":1,"Tags":"<tensorflow><jupyter-notebook><keras>","CreationDate":"2017-03-06T16:51:41.760","AcceptedAnswerId":null,"OwnerUserId":2336548.0,"Title":"Tensorflow not using GPU with jupyter (Windows)","Body":"<p>I have tensorflow GPU working with CUDA.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/prWhy.png\" rel=\"nofollow noreferrer\">Tensorflow GPU<\/a>  <\/p>\n\n<p>But when I am running the following code in jupyter notebook: <\/p>\n\n<pre><\/pre>\n\n<p>The command prompt shows that CPU was used and not the GPU. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/VGotN.png\" rel=\"nofollow noreferrer\">Command Prompt Screen<\/a><\/p>\n","answers":[{"AnswerId":"42651360","CreationDate":"2017-03-07T14:52:18.043","ParentId":null,"OwnerUserId":"1157780","Title":null,"Body":"<p>Please Try Nightly Build version.\nIt seems like you are using 1.0 official release version.<\/p>\n\n<p>$ pip install --ignore-installed --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl<\/a><\/p>\n\n<p>You can find latest wheel pkg url from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow#installation\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow#installation<\/a><\/p>\n"}]}
{"QuestionId":42631417,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-06T17:05:08.467","AcceptedAnswerId":42632892.0,"OwnerUserId":2542516.0,"Title":"How to release tensors from memory when not in need anymore?","Body":"<p>I have hypothetical graph which has a series of computations as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Observe that when computing ;  and  are not needed and hence they can be discarded prior to allocating memory for . Is there some way to ask TensorFlow to perform this memory optimization (agree that there is some cost in time)?<\/p>\n\n<p>Please note that this is not the same as <a href=\"https:\/\/stackoverflow.com\/questions\/34199233\/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory\">this<\/a> question about allocating memory only when needed.<\/p>\n\n<p>EDIT: This network will not be trained, so don't worry about backprop.<\/p>\n","answers":[{"AnswerId":"42632892","CreationDate":"2017-03-06T18:24:44.943","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>TensorFlow uses reference counting to release the memory used by a tensor as soon as it is no longer used. The values of <code>a_0<\/code> and <code>a_1<\/code> will be deleted as soon as there are no more references to them, and in the latest builds of TensorFlow (post-1.0 nightly builds) some operations will even reuse the input buffer for the output if they have the same shape and element type.<\/p>\n"}]}
{"QuestionId":42632411,"AnswerCount":2,"Tags":"<python><image><keras><convolution>","CreationDate":"2017-03-06T17:56:23.080","AcceptedAnswerId":42632755.0,"OwnerUserId":624869.0,"Title":"Keras: can Convolution2D work on rectangular images?","Body":"<p>Simple question - say I have a 360px by 240px image. Instead of cropping my (already small) image to 240x240, can I create a convolutional neural network that operates on the full rectangle? Specifically using the  layer. <\/p>\n\n<p>I ask because every paper I've read doing CNNs seems to have square input sizes, so I wonder if what I propose will be OK, and if so, what disadvantages I may run into. Are all the settings (like ) going to work the same?<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"42632755","CreationDate":"2017-03-06T18:16:35.477","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>No issues with a rectangle image... Everything will work properly as for square images. <\/p>\n"},{"AnswerId":"42652595","CreationDate":"2017-03-07T15:50:03.457","ParentId":null,"OwnerUserId":"4955956","Title":null,"Body":"<p>Yes.<\/p>\n\n<p>But why don't you give it a try<\/p>\n\n<pre><code>#!\/usr\/bin\/env python\n# -*- coding: utf-8 -*-\nfrom __future__ import print_function\nimport numpy as np\nnp.random.seed(1234)\nfrom keras.layers import Input\nfrom keras.layers.convolutional import Convolution2D\nfrom keras.models import Model\n\nprint(\"Building Model...\")\ninp = Input(shape=(1,None,None))\noutput   = Convolution2D(1, 3, 3, border_mode='same', init='normal',bias=False)(inp)\nmodel_network = Model(input=inp, output=output)\nw = np.asarray([ \n    [[[\n    [0,0,0],\n    [0,2,0],\n    [0,0,0]\n    ]]]\n    ])\ninput_mat = np.asarray([ \n    [[\n    [1.,2.,3.,10.],\n    [4.,5.,6.,11.],\n    [7.,8.,9.,12.]\n    ]]\n    ])\nmodel_network.layers[1].set_weights(w)\nprint(\"Weights after change:\")\nprint(model_network.layers[1].get_weights())\nprint(\"Input:\")\nprint(input_mat)\nprint(\"Output:\")\nprint(model_network.predict(input_mat))\n<\/code><\/pre>\n\n<p>Build a sample model<\/p>\n\n<pre><code>inp = Input(shape=(1,None,None))\noutput   = Convolution2D(1, 3, 3, border_mode='same', init='normal',bias=False)(inp)\nmodel_network = Model(input=inp, output=output)\n<\/code><\/pre>\n\n<p>Give it some weights and set them so you could predit the output, say:<\/p>\n\n<pre><code>w = np.asarray([ \n    [[[\n    [0,0,0],\n    [0,2,0],\n    [0,0,0]\n    ]]]\n    ])\nmodel_network.layers[1].set_weights(w)\n<\/code><\/pre>\n\n<p>So that the convolution would simply double your input.<\/p>\n\n<p>Give it your rectangular image:<\/p>\n\n<pre><code>input_mat = np.asarray([ \n    [[\n    [1.,2.,3.,10.],\n    [4.,5.,6.,11.],\n    [7.,8.,9.,12.]\n    ]]\n    ])\n<\/code><\/pre>\n\n<p>And check the output to see if it works<\/p>\n\n<pre><code>print(\"Output:\")\nprint(model_network.predict(input_mat))\n<\/code><\/pre>\n\n<p>Sample output:<\/p>\n\n<pre><code>Using Theano backend.\nBuilding Model...\nWeights after change:\n[array([[[[ 0.,  0.,  0.],\n         [ 0.,  2.,  0.],\n         [ 0.,  0.,  0.]]]], dtype=float32)]\nInput:\n[[[[  1.   2.   3.  10.]\n   [  4.   5.   6.  11.]\n   [  7.   8.   9.  12.]]]]\nOutput:\n[[[[  2.   4.   6.  20.]\n   [  8.  10.  12.  22.]\n   [ 14.  16.  18.  24.]]]]\n<\/code><\/pre>\n\n<p>original <a href=\"https:\/\/stackoverflow.com\/questions\/42211619\/how-to-set-weights-for-convolution2d\/42212349#42212349\">post<\/a> with some changes<\/p>\n"}]}
{"QuestionId":42632656,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><metrics><auc>","CreationDate":"2017-03-06T18:11:00.180","AcceptedAnswerId":null,"OwnerUserId":6114794.0,"Title":"Can't seem to get Tensorflow's tf.metrics.auc working","Body":"<p>Tensorflow has a function to calculate AUC: tf.metrics.auc(). Here is my a section of my code trying to compute auc:<\/p>\n\n<pre><\/pre>\n\n<p>Any section of the error I get is below. The entire error is quite lengthy.<\/p>\n\n<pre><\/pre>\n\n<p>I'd appreciate any pointers on how to resolve this. Thanks<\/p>\n","answers":[{"AnswerId":"44311819","CreationDate":"2017-06-01T15:35:29.590","ParentId":null,"OwnerUserId":"1971741","Title":null,"Body":"<p>might be too late now but if you haven't found the solution, try this change:<\/p>\n\n<pre><code>sess.run(tf.global_variables_initializer())\nsess.run(tf.local_variables_initializer())\n_,roc_score = tf.metrics.auc(y, pred)\nprint(sess.run(roc_score, feed_dict={x : x_test, y : y_test, p_keep_input: 1.0, p_keep_hidden: 1.0}))\n<\/code><\/pre>\n"}]}
{"QuestionId":42633330,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-06T18:51:38.360","AcceptedAnswerId":null,"OwnerUserId":1239984.0,"Title":"Tensorflow TimeoutError connection attempt failed","Body":"<p>Source Code: <a href=\"https:\/\/github.com\/llSourcell\/tensorflow_demo\" rel=\"nofollow noreferrer\">https:\/\/github.com\/llSourcell\/tensorflow_demo<\/a><\/p>\n\n<p>Video Tutorial: <a href=\"https:\/\/www.youtube.com\/watch?v=2FmcHiLCwTU\" rel=\"nofollow noreferrer\">https:\/\/www.youtube.com\/watch?v=2FmcHiLCwTU<\/a><\/p>\n\n<pre><\/pre>\n\n<p>I installed Tensorflow on Windows using Python 3.5.2; When I try to run the board.py file I get a timeout error. I get this error on a number of tensorflow scripts I tried to run. Whats causing this error and how do I resolve it?<\/p>\n","answers":[{"AnswerId":"42633726","CreationDate":"2017-03-06T19:12:29.360","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The MNIST dataset, hosted at <a href=\"http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a>, appears to be offline currently. Unfortunately, that URL seems to be hard-coded in the various example files, but you could try changing it to one of the mirrors listed in this <a href=\"https:\/\/github.com\/Lasagne\/Lasagne\/pull\/589\/files\" rel=\"nofollow noreferrer\">pull request<\/a>:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/web.archive.org\/web\/20150912040653\/http:\/\/yann.lecun.com\/exdb\/mnist\/\" rel=\"nofollow noreferrer\">https:\/\/web.archive.org\/web\/20150912040653\/http:\/\/yann.lecun.com\/exdb\/mnist\/<\/a><\/li>\n<li><a href=\"https:\/\/s3.amazonaws.com\/lasagne\/recipes\/datasets\/mnist\/\" rel=\"nofollow noreferrer\">https:\/\/s3.amazonaws.com\/lasagne\/recipes\/datasets\/mnist\/<\/a><\/li>\n<\/ul>\n\n<p>Note that you can download the files manually to the temporary directory specified in the <code>input_data.read_data_sets()<\/code> call (in your case, this appears to be <code>\/tmp\/data\/<\/code>). If the reader code finds the files in that location, it will not attempt to contact the server.<\/p>\n"}]}
{"QuestionId":42633644,"AnswerCount":2,"Tags":"<machine-learning><neural-network><time-series><keras><lstm>","CreationDate":"2017-03-06T19:08:03.123","AcceptedAnswerId":42635571.0,"OwnerUserId":3367967.0,"Title":"Using Keras for video prediction (time series)","Body":"<p>I want to predict the next frame of a (greyscale) video given  previous frames - using CNNs or RNNs in Keras. Most tutorials and other information regarding time series prediction and Keras use a 1-dimensional input in their network but mine would be 3D <\/p>\n\n<p>I'm currently really unsure what a good approach for this problem would be. My ideas include:<\/p>\n\n<ul>\n<li><p>Using one or more LSTM layers. The problem here is that I'm not sure whether they're suited to take a series of images instead a series of scalars as input. Wouldn't the memory consumption explode? If it is okay to use them: How can I use them in Keras for higher dimensions?<\/p><\/li>\n<li><p>Using 3D convolution on the input (the stack of previous video frames). This raises other questions: Why would this help when I'm not doing a classification but a prediction? How can I stack the layers in such a way that the input of the network has dimensions  and the output ?<\/p><\/li>\n<\/ul>\n\n<p>I'm pretty new to CNNs\/RNNs and Keras and would appreciate any hint into the right direction.<\/p>\n","answers":[{"AnswerId":"42635571","CreationDate":"2017-03-06T21:08:10.553","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>So basically every approach has its advantages and disadvantages. Let's go throught the ones you provided and then other to find the best approach:<\/p>\n\n<ol>\n<li><p><code>LSTM<\/code>: Among their biggest advantages is an ability to learn a long-term dependiencies patterns in your data. They were designed in order to be able to analyse long sequences like e.g. speech or text. This is also might cause problems because of number parameters which could be really high. Other typical recurrent network architectures like <code>GRU<\/code> might overcome this issues. The main disadvantage is that in their standard (sequential implementation) it's infeasible to fit it on a video data for the same reason why dense layers are bad for an imagery data - loads of time and spatial invariances must be learnt by a topology which is completely not suited for catching them in an efficient manner. Shifting a video by a pixel to the right might completely change the output of your network.<\/p>\n\n<p>Other thing which is worth to mention is that training <code>LSTM<\/code> is belived to be similiar to finding equilibrium between two rivalry processes - finding good weights for a <em>dense-like<\/em> output computations and finding a good inner-memory dynamic in processing sequences. Finding this equilibrium might last for a really long time but once its finded - it's usually quite stable and produces a really good results.<\/p><\/li>\n<li><p><code>Conv3D<\/code>: Among their biggest advantages one may easily find an ability to catch spatial and temporal invariances in the same manner as <code>Conv2D<\/code> in an imagery case. This make the curse of dimensionality much less harmful. On the other hand - in the same way as <code>Conv1D<\/code> might not produce good results with a longer sequences - in the same way - a lack of any memory might make learning a long sequence harder.<\/p><\/li>\n<\/ol>\n\n<p>Of course one may use different approaches like:<\/p>\n\n<ol start=\"3\">\n<li><p><code>TimeDistributed + Conv2D<\/code>: using a <code>TimeDistributed<\/code> wrapper - one may use some pretrained convnet like e.g. <code>Inception<\/code> framewise and then analyse the feature maps sequentially. A really huge advantage of this approach is a possibility of a transfer learning. As a disadvantage - one may think about it as a <code>Conv2.5D<\/code> - it lacks temporal analysis of your data.<\/p><\/li>\n<li><p><code>ConvLSTM<\/code>: this architecture is not yet supported by the newest version of <code>Keras<\/code> (on March 6th 2017) but as one may see <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/keras\/layers\/convolutional_recurrent.py\" rel=\"noreferrer\">here<\/a> it should be provided in the future. This is a mixture of <code>LSTM<\/code> and <code>Conv2D<\/code> and it's belived to be better then stacking <code>Conv2D<\/code> and <code>LSTM<\/code>. <\/p><\/li>\n<\/ol>\n\n<p>Of course these are not the only way to solve this problem, I'll mention one more which might be usefull:<\/p>\n\n<ol start=\"5\">\n<li>Stacking: one may easily stack the upper methods in order to build their final solution. E.g. one may build a network where at the beginning video is transformed using a <code>TimeDistributed(ResNet)<\/code> then output is feed to <code>Conv3D<\/code> with multiple and agressive spatial pooling and finally transformed by an <code>GRU\/LSTM<\/code> layer. <\/li>\n<\/ol>\n\n<p><strong>PS:<\/strong><\/p>\n\n<p>One more thing that is also worth to mention is that shape of video data is actually <code>4D<\/code> with <code>(frames, width, height, channels<\/code>).<\/p>\n\n<p><strong>PS2:<\/strong><\/p>\n\n<p>In case when your data is actually <code>3D<\/code> with <code>(frames, width, hieght)<\/code> you actually could use a classic <code>Conv2D<\/code> (by changing <code>channels<\/code> to <code>frames<\/code>) to analyse this data (which actually might more computationally effective). In case of a <em>transfer learning<\/em> you should add additional dimension because most of <code>CNN<\/code> models were trained on data with shape <code>(width, height, 3)<\/code>. One may notice that your data doesn't have 3 channels. In this case a technique which is usually used is repeating spatial matrix three times.<\/p>\n\n<p><strong>PS3:<\/strong><\/p>\n\n<p>An example of this <code>2.5D<\/code> approach is:<\/p>\n\n<pre><code>input = Input(shape=input_shape)\nbase_cnn_model = InceptionV3(include_top=False, ..)\ntemporal_analysis = TimeDistributed(base_cnn_model)(input)\nconv3d_analysis = Conv3D(nb_of_filters, 3, 3, 3)(temporal_analysis)\nconv3d_analysis = Conv3D(nb_of_filters, 3, 3, 3)(conv3d_analysis)\noutput = Flatten()(conv3d_analysis)\noutput = Dense(nb_of_classes, activation=\"softmax\")(output)\n<\/code><\/pre>\n"},{"AnswerId":"42645922","CreationDate":"2017-03-07T10:33:56.070","ParentId":null,"OwnerUserId":"3367967","Title":null,"Body":"<p>After doing lots of research, I finally stumbled upon the <a href=\"https:\/\/github.com\/fchollet\/keras\/blob\/master\/examples\/conv_lstm.py\" rel=\"noreferrer\">Keras Example<\/a> for the <code>ConvLSTM2D<\/code> layer (Already mentioned by Marcin Mo\u017cejko), which does exactly what I need. <\/p>\n\n<p>In the current version of Keras (v1.2.2), this layer is already included and can be imported using<\/p>\n\n<pre><code>from keras.layers.convolutional_recurrent import ConvLSTM2D \n<\/code><\/pre>\n\n<p>To use this layer, the video data has to be formatted as follows:<\/p>\n\n<pre><code>[nb_samples, nb_frames, width, height, channels] # if using dim_ordering = 'tf'\n[nb_samples, nb_frames, channels, width, height] # if using dim_ordering = 'th'\n<\/code><\/pre>\n"}]}
{"QuestionId":42633865,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-03-06T19:20:50.013","AcceptedAnswerId":null,"OwnerUserId":5862623.0,"Title":"Could you train TensorFlow to assess pronunciation?","Body":"<p>I have a language learning app and would like to detect the pronunciation of users saying certain words and sentences.  <\/p>\n\n<p>If I can provide a 1000 user audio files, each with a rating out of 100 of how good the pronunciation is, could I predict a score out of 100 of how good the pronunciation is? <\/p>\n","answers":[{"AnswerId":"42787749","CreationDate":"2017-03-14T13:46:56.467","ParentId":null,"OwnerUserId":"6061661","Title":null,"Body":"<p>I don't see why not, this is a very general question though...  I would read up on how to vectorize audio first. <a href=\"http:\/\/www.hexahedria.com\/2015\/08\/03\/composing-music-with-recurrent-neural-networks\/\" rel=\"nofollow noreferrer\">These guys<\/a> used a neural network to generate audio data, maybe you could look at how they took in data and do something similar.  Then once you vectorize it, pass the vector as input and score as the label.<\/p>\n"}]}
{"QuestionId":42634179,"AnswerCount":1,"Tags":"<c++><neural-network><deep-learning><caffe><deep-dream>","CreationDate":"2017-03-06T19:40:50.407","AcceptedAnswerId":42635711.0,"OwnerUserId":2504754.0,"Title":"Caffe::net reshape","Body":"<p>I am trying to implement  in  in (I want to run it in android).  requires input of shape . In the ipython notebook of  it shows . Does this mean that only input blob is reshaped or it is propagated through the network? I tried calling the  in  and it resulted in:<\/p>\n\n<pre><\/pre>\n\n<p>Shouldn't the network be reshaped too? If not what is the implication of just reshaping input blob? I am new to deep learning. So forgive me if it seems trivial.<\/p>\n","answers":[{"AnswerId":"42635711","CreationDate":"2017-03-06T21:16:22.893","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>changing the shape of the input requires reshaping of the entire net. Alas, there are some layer types that do not like to be reshaped. Specifically, \"InnerProduct\" layer: the number of trainable parameters of an inner product layer depends on the <strong>exact<\/strong> input shape and the output shape. Therefore a net with an \"InnerProduct\" layer cannot be reshaped.<\/p>\n\n<p>You can use methods described in the <a href=\"https:\/\/github.com\/BVLC\/caffe\/blob\/master\/examples\/net_surgery.ipynb\" rel=\"nofollow noreferrer\">\"net surgery\"<\/a> example to convert the inner product layers to equivalent convolutional layers (that can be reshaped).<\/p>\n"}]}
{"QuestionId":42634572,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-06T20:04:36.087","AcceptedAnswerId":42886082.0,"OwnerUserId":2542516.0,"Title":"Why does this snippet with 2 queue runners hang","Body":"<p>I have 2 queues, one for file names, one for holding the contents of those filenames.  adds a QR automatically. I add a QR for enqueuing the second queue. But this code hangs and any advice that will let me figure out what I am missing is appreciated.<\/p>\n\n<pre><\/pre>\n\n<p>EDIT: After some experiments, I found that the first QR runs but not the second.<\/p>\n","answers":[{"AnswerId":"42886082","CreationDate":"2017-03-19T11:32:45.730","ParentId":null,"OwnerUserId":"2542516","Title":null,"Body":"<p>Queue runners are not added to <code>tf.GraphKeys.QUEUE_RUNNERS<\/code> implicitly. Add any <code>QueueRunner<\/code> you create using <code>tf.train.add_queue_runner<\/code><\/p>\n\n<p>This should fix it:<\/p>\n\n<pre><code>qr = tf.train.QueueRunner(image_queue, [enqueue_op] * 2)\ntf.train.add_queue_runner(qr)\n<\/code><\/pre>\n\n<p>Now calling <code>tf.train.start_queue_runners<\/code> will start this queue runner (<code>qr<\/code>) too.<\/p>\n"}]}
{"QuestionId":42635159,"AnswerCount":0,"Tags":"<python><scala><tensorflow>","CreationDate":"2017-03-06T20:42:35.580","AcceptedAnswerId":null,"OwnerUserId":7668708.0,"Title":"TensorFlow OpDef \"func\" Attribute Type Interpretation","Body":"<p>I am currently in the process of implementing a Scala API for TensorFlow and I am using scala.meta to automatically generate the Op creation methods. I am calling  from C API to obtain an OpList protocol buffer and I am then parsing that to create the methods.<\/p>\n\n<p>Each Op can have a set of attributes and each attribute has a type and optionally a default value. I am confused as to what the default value for a  attribute should look like and also, how to set that attribute value through the C API when creating the Op (i.e., there is no  function). Could someone provide some clarification with respect to that? Also, a reference to what is currently done when generating the Python API (and to the relevant part of the code that generates the API) would be greatly appreciated.<\/p>\n\n<p>I am also a bit confused with respect to the  attribute type and could use some help with respect to its meaning and use too, but I could create a separate question for that, if necessary.<\/p>\n\n<p>Thank you!<\/p>\n","answers":[]}
{"QuestionId":42635521,"AnswerCount":4,"Tags":"<python><tensorflow>","CreationDate":"2017-03-06T21:05:08.040","AcceptedAnswerId":42712559.0,"OwnerUserId":2689084.0,"Title":"Tensorflow restored weights is not set","Body":"<p>I'm trying to restore a model in Tensorflow which I've trained. The problem is that it does not seem like the weights are properly restored.<\/p>\n\n<p>For the training I've got the weights and biases defined as:<\/p>\n\n<pre><\/pre>\n\n<p>Then I do some training on my own custom 2D image dataset and save the model by calling the <\/p>\n\n<pre><\/pre>\n\n<p>Later I want to restore that model with the exact same weights, so I build the model as before (also with the  initialization) and call the <\/p>\n\n<pre><\/pre>\n\n<p>Now, if i call:<\/p>\n\n<pre><\/pre>\n\n<p>I get random values, and not the restored value of the weight.<\/p>\n\n<p>I've drawn a blank on this one, can somebody point me in the right direction?<\/p>\n\n<p>FYI, I've tried (without) luck to simply declare the s, but I keep getting:<\/p>\n\n<pre><\/pre>\n\n<p>even though Tensorflow themselves state that it should be possible to simply declare with no initial value (<a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/variables\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/programmers_guide\/variables<\/a> part: Restoring Values)<\/p>\n\n<p><strong>Update 1<\/strong><\/p>\n\n<p>When I, as suggested, run<\/p>\n\n<pre><\/pre>\n\n<p>I get the following output:<\/p>\n\n<pre><\/pre>\n\n<p>Which shows that the variables indeed is read. However invoking<\/p>\n\n<pre><\/pre>\n\n<p>Results in the error: Attempting to use uninitialized value wh1<\/p>\n","answers":[{"AnswerId":"48091460","CreationDate":"2018-01-04T08:36:46.720","ParentId":null,"OwnerUserId":"6599380","Title":null,"Body":"<p>I have meet the same problem when saving model to saved_model format. Anyone using the function <em>add_meta_graph_and_variables<\/em> to save the model for serving, be careful about this parameter \"legacy_init_op: Legacy support for op or group of ops to execute after the restore op upon a load.\"<\/p>\n"},{"AnswerId":"42712559","CreationDate":"2017-03-10T07:21:46.600","ParentId":null,"OwnerUserId":"2689084","Title":null,"Body":"<p>So with the help of you guys, I ended up dividing the saving and restoring parts of my program into two files, to ensure that no unwanted variables were initialized.<\/p>\n\n<p><strong>Train and Save routines <code>fnn.py<\/code><\/strong><\/p>\n\n<pre><code>def build(self, topology):\n    \"\"\"\n    Builds the topology of the model\n    \"\"\"\n\n    # Sanity check\n    assert len(topology) == 4\n\n    n_inputs = topology[0]\n    n_hidden_1 = topology[1]\n    n_hidden_2 = topology[2]\n    n_classes = topology[3]\n\n    # Sanity check\n    assert self.img_h * self.img_w == n_inputs\n\n    # Instantiate TF Placeholders\n    self.x = tf.placeholder(tf.float32, [None, n_inputs], name='x')\n    self.y = tf.placeholder(tf.float32, [None, n_classes], name='y')\n    self.W = {\n        'h1': tf.Variable(tf.random_normal([n_inputs, n_hidden_1]), name='wh1'),\n        'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2]), name='wh2'),\n        'o': tf.Variable(tf.random_normal([n_hidden_2, n_classes]), name='wo')\n    }\n    self.b = {\n        'b1': tf.Variable(tf.random_normal([n_hidden_1]), name='bh1'),\n        'b2': tf.Variable(tf.random_normal([n_hidden_2]), name='bh2'),\n        'o': tf.Variable(tf.random_normal([n_classes]), name='bo')\n    }\n\n    # Create model\n    self.l1 = tf.nn.sigmoid(tf.add(tf.matmul(self.x, self.W['h1']), self.b['b1']))\n    self.l2 = tf.nn.sigmoid(tf.add(tf.matmul(self.l1, self.W['h2']), self.b['b2']))\n    logits = tf.add(tf.matmul(self.l2, self.W['o']), self.b['o'])\n\n    # Define predict operation\n    self.predict_op = tf.argmax(logits, 1)\n    probs = tf.nn.softmax(logits, name='probs')\n\n    # Define cost function\n    self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, self.y))\n\n    # Adding these to collection so we can restore them again\n    tf.add_to_collection('inputs', self.x)\n    tf.add_to_collection('inputs', self.y)\n    tf.add_to_collection('outputs', logits)\n    tf.add_to_collection('outputs', probs)\n    tf.add_to_collection('outputs', self.predict_op)\n\ndef train(self, X, Y, n_epochs=10, learning_rate=0.001, logs_path=None):\n    \"\"\"\n    Trains the Model\n    \"\"\"\n    self.optimizer = tf.train.AdamOptimizer(learning_rate).minimize(self.cost)\n\n    costs = []\n\n    # Instantiate TF Saver\n    saver = tf.train.Saver()\n\n    with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        sess.run(tf.local_variables_initializer())\n\n        # start the threads used for reading files\n        coord = tf.train.Coordinator()\n        threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\n        # Compute total number of batches\n        total_batch = int(self.get_num_examples() \/ self.batch_size)\n\n        # start training\n        for epoch in range(n_epochs):\n            for i in range(total_batch):\n\n                batch_xs, batch_ys = sess.run([X, Y])\n\n                # run the training step with feed of images\n                _, cost = sess.run([self.optimizer, self.cost], feed_dict={self.x: batch_xs,\n                                                                           self.y: batch_ys})\n                costs.append(cost)\n                print \"step %d\" % (epoch * total_batch + i)\n            #costs.append(cost)\n            print \"Epoch %d\" % epoch\n\n        saver.save(sess, self.model_file)\n\n        temp = sess.run(self.W['h1'][0][0])\n        print temp\n\n        if self.visu:\n            plt.plot(costs)\n            plt.show()\n\n        # finalize\n        coord.request_stop()\n        coord.join(threads)\n<\/code><\/pre>\n\n<p><strong>Predict routine <code>fnn_eval.py<\/code>:<\/strong><\/p>\n\n<pre><code>with tf.Session() as sess:\n        sess.run(tf.global_variables_initializer())\n        sess.run(tf.local_variables_initializer())\n\n        g = tf.get_default_graph()\n\n        # restore the model\n        self.saver = tf.train.import_meta_graph(self.model_file)\n        self.saver.restore(sess, tf.train.latest_checkpoint('.\/tfmodels\/fnn\/'))\n\n        wh1 = g.get_tensor_by_name(\"wh1:0\")\n        print sess.run(wh1[0][0])\n\n        x, y = tf.get_collection('inputs')\n        logits, probs, predict_op = tf.get_collection('outputs')\n\n        coord = tf.train.Coordinator()\n        threads = tf.train.start_queue_runners(sess=sess, coord=coord)\n\n        predictions = []\n\n        print Y.eval()\n\n        for i in range(1):#range(self.get_num_examples()):\n            batch_xs = sess.run(X)\n            # Reshape batch_xs if only a single image is given\n            #   (numpy is 4D: batch_size * heigth * width * channels)\n            batch_xs = np.reshape(batch_xs, (-1, self.img_w * self.img_h))\n            prediction, probabilities, logit = sess.run([predict_op, probs, logits], feed_dict={x: batch_xs})\n            predictions.append(prediction[0])\n\n        # finalize\n        coord.request_stop()\n        coord.join(threads)\n<\/code><\/pre>\n"},{"AnswerId":"42652019","CreationDate":"2017-03-07T15:22:18.370","ParentId":null,"OwnerUserId":"4834515","Title":null,"Body":"<p>I guess the problem might be caused by creating a new variable when you restore the model, not getting the already existed variable. I tried this code<\/p>\n\n<pre><code>saver = tf.train.import_meta_graph('.\/model.ckpt-10.meta')\nw1 = None\nfor v in tf.global_variables():\n        print v.name\n\nw1 = tf.get_variable('wh1', [])\n\ninit = tf.global_variables_initializer()\nsess.run(init)\n\nsaver.restore(sess, '.\/model.ckpt-10')\n\nfor v in tf.global_variables():\n    print v.name\n<\/code><\/pre>\n\n<p>and clearly you can see the output that it creates a new variable called <code>wh1_1:0<\/code>.<\/p>\n\n<p>If you try this<\/p>\n\n<pre><code>w1 = None\n\nfor v in tf.global_variables():\n    print v.name\n    if v.name == 'wh1:0':\n        w1 = v\n\ninit = [tf.global_variables_initializer(), tf.local_variables_initializer()]\nsess.run(init)\n\nsaver.restore(sess, '.\/model.ckpt-10')\n\nfor v in tf.global_variables():\n    print v.name\n\ntemp = sess.run(w1)\nprint temp[0][0]\n<\/code><\/pre>\n\n<p>There will be no problem. <\/p>\n\n<p>Tensorflow suggests that it is better to use <code>tf.variable_scope()<\/code> (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/variable_scope\" rel=\"nofollow noreferrer\">link<\/a>) like this<\/p>\n\n<pre><code>with tf.variable_scope(\"foo\"):\n    v = tf.get_variable(\"v\", [1])\nwith tf.variable_scope(\"foo\", reuse=True):\n    v1 = tf.get_variable(\"v\", [1])\nassert v1 == v\n<\/code><\/pre>\n"},{"AnswerId":"42637672","CreationDate":"2017-03-06T23:44:37.420","ParentId":null,"OwnerUserId":"3813674","Title":null,"Body":"<p>You want to pass in a <code>var_list<\/code> to the <code>Saver<\/code>. <\/p>\n\n<p>In your case, the variable list would come from your <code>W<\/code> and <code>b<\/code> dictionaries: <code>var_list = list(W.values())+list(b.values())<\/code>. Then, to restore the model, pass in <code>var_list<\/code> to the <code>Saver<\/code>: <code>saver = tf.train.Saver(var_list=var_list)<\/code>.<\/p>\n\n<p>Next, you need to get your checkpoint state: <code>model = tf.train.get_checkpoint_state(&lt;your saved model directory&gt;)<\/code>. After that you can restore the trained weights.<\/p>\n\n<pre><code>var_list = list(W.values())+list(b.values())\nsaver = tf.train.Saver(var_list=var_list)\nmodel = tf.get_checkpoint_state('.\/model\/')\n\nwith tf.Session() as sess:\n    saver.restore(sess,model.model_checkpoint_path)\n    #Now use the pretrained weights\n<\/code><\/pre>\n"}]}
{"QuestionId":42635873,"AnswerCount":1,"Tags":"<tensorflow><tensorflow-serving>","CreationDate":"2017-03-06T21:26:58.593","AcceptedAnswerId":null,"OwnerUserId":5434466.0,"Title":"Tensorflow Serving Batching parameters","Body":"<p>How to do performance tuning of batching using max_batch_size, batch_timeout_micros, num_batch_threads and other parameters? Tried using these parameters with the Query client, it doesn't work.<\/p>\n\n<p>In the below example I have 100 images and I want to batch in size of 10. The query runs for all images instead of 10.\nbazel-bin\/tensorflow_serving\/example\/demo_batch --server=localhost:9000 --max_batch_size = 10<\/p>\n\n<p>Also, for batch scheduling how to make it run every 10 secs after the first batch is done? Thanks.<\/p>\n","answers":[{"AnswerId":"44391979","CreationDate":"2017-06-06T13:54:09.353","ParentId":null,"OwnerUserId":"1583517","Title":null,"Body":"<p>I have met the same problem like you.<\/p>\n\n<p>And I checked the source code of tf-serving, these parameters are in an protobuf file which defined in<\/p>\n\n<pre><code>serving\/tensorflow_serving\/servables\/tensorflow\/session_bundle_config.proto\n<\/code><\/pre>\n\n<p>And I found the example conf file in<\/p>\n\n<p>serving\/tensorflow_serving\/servables\/tensorflow\/testdata\/batching_config.txt<\/p>\n\n<p>And I believe you could follow the batching_config.txt format, the parameters config should be work.<\/p>\n\n<p>Hope it helps.<\/p>\n\n<hr>\n\n<p>max_batch_size { value: 1024 }<\/p>\n\n<p>batch_timeout_micros { value: 0 }<\/p>\n\n<p>max_enqueued_batches { value: 1000000 }<\/p>\n\n<p>num_batch_threads { value: 8 }<\/p>\n\n<p>allowed_batch_sizes : 1<\/p>\n\n<p>allowed_batch_sizes : 2<\/p>\n\n<p>allowed_batch_sizes : 8<\/p>\n\n<p>allowed_batch_sizes : 32<\/p>\n\n<p>allowed_batch_sizes : 128<\/p>\n\n<p>allowed_batch_sizes : 256<\/p>\n\n<p>allowed_batch_sizes : 512<\/p>\n\n<p>allowed_batch_sizes : 1024<\/p>\n"}]}
{"QuestionId":42636323,"AnswerCount":1,"Tags":"<reinforcement-learning><pytorch>","CreationDate":"2017-03-06T21:53:02.263","AcceptedAnswerId":null,"OwnerUserId":7668940.0,"Title":"Explanation behind actor-critic algorithm in pytorch example?","Body":"<p>Pytorch provides a good example of using actor-critic to play Cartpole in the OpenAI gym environment. <\/p>\n\n<p>I'm confused about several of their equations in the code snippet found at <a href=\"https:\/\/github.com\/pytorch\/examples\/blob\/master\/reinforcement_learning\/actor_critic.py#L67-L79\" rel=\"nofollow noreferrer\">https:\/\/github.com\/pytorch\/examples\/blob\/master\/reinforcement_learning\/actor_critic.py#L67-L79<\/a>: <\/p>\n\n<pre><\/pre>\n\n<p>What do r and value mean in this case? Why do they run REINFORCE on the action space with the reward equal to r - value? And why do they try to set the value so that it matches r? <\/p>\n\n<p>Thanks for your help!<\/p>\n","answers":[{"AnswerId":"43381251","CreationDate":"2017-04-12T23:33:16.663","ParentId":null,"OwnerUserId":"1895611","Title":null,"Body":"<ul>\n<li><p>First the rewards a collected for a time, along with the state:action that resulted in the reward<\/p><\/li>\n<li><p>Then r - value is the difference between the expected reward and actual<\/p><\/li>\n<li><p>That difference is used to adjust the expected value of that action from that state<\/p><\/li>\n<\/ul>\n\n<p>So if in state \"middle\", the expected reward for action \"jump\" was 10  and the actual reward was only 2, then the AI was off by -8 (2-10).  Reinforce means \"adjust expectations\".  So if we adjust them by half, we'll new expected reward is 10-(8 *.5), or 6.  meaning the AI really thought it would get 10 for that, but now it's less confident and thinks 6 is a better guess.   So if the AI is not off by much, 10 - ( 2 *.5) = 9, it will adjust by a smaller amount.  <\/p>\n"}]}
{"QuestionId":42636685,"AnswerCount":1,"Tags":"<tensorflow><deep-learning><caffe><keras>","CreationDate":"2017-03-06T22:20:43.340","AcceptedAnswerId":42712020.0,"OwnerUserId":288609.0,"Title":"patch-wise training and fully convolutional training in FCN","Body":"<p>In the <a href=\"https:\/\/arxiv.org\/abs\/1605.06211\" rel=\"nofollow noreferrer\">FCN paper<\/a>,  the authors discuss the patch wise training and fully convolutional training. What is the difference between these two?<\/p>\n\n<p>Please refer to  attached in the following.<\/p>\n\n<p>It seems to me that the training mechanism is as follows, \nAssume the original image is , then iterate the  pixels to extract  patch (where ). The iteration stride can some number like  to generate overlapping patches. Moreover, assume each single image corresponds to 20 patches, then we can put these  patches or  patches(if we want to have 3 images) into a single mini-batch for training. Is this understanding right? It seems to me that this so-called fully convolutional training is the same as patch-wise training.<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/pEyfL.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/pEyfL.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42712020","CreationDate":"2017-03-10T06:48:37.573","ParentId":null,"OwnerUserId":"1694320","Title":null,"Body":"<p>The term \"Fully Convolutional Training\" just means replacing fully-connected layer with convolutional layers so that the whole network contains just convolutional layers (and pooling layers).<\/p>\n\n<p>The term \"Patchwise training\" is intended to avoid the redundancies of full image training.\nIn semantic segmentation, given that you are classifying each pixel in the image, by using the whole image, you are adding a lot of redundancy in the input. A standard approach to avoid this during training segmentation networks is to feed the network with batches of random patches (small image regions surrounding the objects of interest) from the training set instead of full images. This \"patchwise sampling\" ensures that the input has enough variance and is a valid representation of the training dataset (the mini-batch should have the same distribution as the training set). This technique also helps to converge faster and to balance the classes. In this paper, they claim that is it not necessary to use patch-wise training and if you want to balance the classes you can weight or sample the loss.\nIn a different perspective, the problem with full image training in per-pixel segmentation is that the input image has a lot of spatial correlation. To fix this, you can either sample patches from the training set (patchwise training) or sample the loss from the whole image. That is why the subsection is called \"Patchwise training is loss sampling\".\nSo by \"restricting the loss to a randomly sampled subset of its spatial terms excludes patches from the gradient computation.\" They tried this \"loss sampling\" by randomly ignoring cells from the last layer so the loss is not calculated over the whole image.<\/p>\n"}]}
{"QuestionId":42637713,"AnswerCount":1,"Tags":"<python><opencv><tensorflow><ros>","CreationDate":"2017-03-06T23:48:24.580","AcceptedAnswerId":null,"OwnerUserId":4176597.0,"Title":"Using tensorflow.contrib with cv_bridge causes tcmalloc error","Body":"<p>The following script<\/p>\n\n<pre><\/pre>\n\n<p>throws the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Process finished with exit code 134 (interrupted by signal 6: SIGABRT)<\/p>\n\n<p>Operating System:<\/p>\n\n<pre><\/pre>\n\n<p>I've tried reinstalling cv_bridge and tensorflow. No luck.<\/p>\n","answers":[{"AnswerId":"42680972","CreationDate":"2017-03-08T20:08:28.173","ParentId":null,"OwnerUserId":"4176597","Title":null,"Body":"<p>It turns out that the solution was to switch the order of imports:<\/p>\n\n<pre><code>import cv_bridge\nimport tensorflow.contrib\n\nimport rospy\nrospy.init_node('node')\n<\/code><\/pre>\n\n<p>This is a bug and I've posted issues on the tensorflow and opencv githubs.<\/p>\n\n<p><strong>EDIT:<\/strong>\nI found this unfortunate line in my <code>.zshrc<\/code>:<\/p>\n\n<pre><code>export LD_PRELOAD=\"\/usr\/lib\/libtcmalloc_minimal.so.4\"\n<\/code><\/pre>\n\n<p>Not sure why I added that. Commenting it out solved the problem. While not a bug, I think it would helpful to have some kind of more informative error message.<\/p>\n\n<p>This is <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/8146\" rel=\"nofollow noreferrer\">the issue on the Tensorflow Github page<\/a> and this is <a href=\"https:\/\/github.com\/ros-perception\/vision_opencv\/issues\/161\" rel=\"nofollow noreferrer\">the one on the OpenCV page<\/a>.<\/p>\n"}]}
{"QuestionId":42638173,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-07T00:38:43.500","AcceptedAnswerId":null,"OwnerUserId":7584299.0,"Title":"ImportError:libcudart.so.8.0:cannot open shared object file:No such file or directory","Body":"<p>I used the spyder or terminal to run the mnist_deep.py program ,  it can run normally, but when i run the program using pycharm, it will cause an error.The Project Interpreter is 2.7.12(~\/anaconda2\/bin\/python).\nError Type: \nImportError:libcudart.so.8.0:cannot open shared object file:No such file or directory\nError importing tensorflow.Unless you are using bazel,you should not try to import tensorflow from its source directory;please exit the tensorflow source tree,and relaunch your python interpreter from there<\/p>\n","answers":[{"AnswerId":"46278097","CreationDate":"2017-09-18T11:23:07.990","ParentId":null,"OwnerUserId":"4126114","Title":null,"Body":"<p>In my case, I had installed <code>tensorflow-gpu<\/code> on a machine without a GPU.\nUninstalling it fixed the problem<\/p>\n\n<pre><code>pip uninstall tensorflow-gpu\npip install tensorflow --ignore-installed\n<\/code><\/pre>\n\n<p>Without the 2nd step, I was getting an error <code>No module named tensorflow<\/code> (<a href=\"https:\/\/stackoverflow.com\/a\/45551934\/4126114\">ref<\/a>)<\/p>\n"},{"AnswerId":"44061048","CreationDate":"2017-05-19T03:40:33.093","ParentId":null,"OwnerUserId":"4212158","Title":null,"Body":"<p>It's preferable to do a system-wide configuration instead of editing <code>~\/.bashrc<\/code>. Create a .conf file in <code>\/etc\/ld.so.conf.d\/<\/code>:<\/p>\n\n<p><code>sudo nano \/etc\/ld.so.conf.d\/nvidia.conf<\/code><\/p>\n\n<p>Add this:<\/p>\n\n<pre><code>\/usr\/local\/cuda-8.0\/lib64\n\/usr\/local\/cuda-8.0\/lib  # you probably don't need this line; check\n                         # if the directory exists\n<\/code><\/pre>\n\n<p>Then run <code>sudo ldconfig<\/code>. Hope this brings you joy.<\/p>\n"}]}
{"QuestionId":42638260,"AnswerCount":1,"Tags":"<deep-learning><keras><keras-layer>","CreationDate":"2017-03-07T00:48:22.920","AcceptedAnswerId":42643222.0,"OwnerUserId":6649127.0,"Title":"How to deal with ordinal labels in keras?","Body":"<p>I have data with integer target class in the range 1-5 where one is the lowest and five the highest. In this case, should I consider it as regression problem and have one node in the output layer? <\/p>\n\n<p>My way of handling it is: <\/p>\n\n<p>1- first I convert the labels to binary class matrix<\/p>\n\n<pre><\/pre>\n\n<p>2- in the output layer, I have five nodes <\/p>\n\n<pre><\/pre>\n\n<p>3- I use 'categorical_crossentropy  with mean_squared_error when compiling<\/p>\n\n<pre><\/pre>\n\n<p>Also, can anyone tells me: what is the difference between using categorical_accuracy and 'mean_squared_error in this case? <\/p>\n","answers":[{"AnswerId":"42643222","CreationDate":"2017-03-07T08:11:04.507","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Regression and classification are vastly different things. If you reimagine this as a regression task than the difference of predicting 2 when the ground truth is 4 will be rated more than if you predict 3 instead of 4. If you have class like car, animal, person you do not care for the ranking between those classes. Predicting car is just as wrong as animal, iff the image shows a person.<\/p>\n\n<p>Metrics do not impact your learning at all. It is just something that is computed additionally to the loss to show the performance of the model. Here the accuracy makes sense, because this is mostly the metric that we care about. Mean squared error does not tell you how well your model performs. If you get something like 0.0015 mean squared error it sounds good, but it is hard to visualize just how well this performs. In contrast using accuracy and achieving 95% accuracy for example is meaningful.<\/p>\n\n<p>One last thing you should use softmax instead of sigmoid as your final output to get a probability distribution in your final layer. Softmax will output percentages for every class that sum up to 1. Then crossentropy calculates the difference of the probability distribution of your network output and the ground truth.<\/p>\n"}]}
{"QuestionId":42638373,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network>","CreationDate":"2017-03-07T01:02:30.263","AcceptedAnswerId":42638634.0,"OwnerUserId":852795.0,"Title":"repick random normal values less than given stadard deviation","Body":"<p>To initialize weights in my tensorflow neural network I began with . A normal distribution around zero is what I'm seeking. To then eliminate excessively large initial weights, I switched to . Since -2.0 and +2.0 are <em>still<\/em> too large, I moved to , solving the problem of initial weight being too 'large.'<\/p>\n\n<p>However, I would also like to eliminate values that are too small (i.e. too close to zero). In other words, in addition to re-picking all weights &lt;-0.2 or >+0.2, I'd like to re-pick all weights between, for example, -0.02 and +0.02. Is there a way to do this outside of a brute force  loop? In other words, is there an 'inverse' ?<\/p>\n\n<p>P.S. The reason for my thinking comes from after the 27th minute of a <a href=\"http:\/\/www.thetalkingmachines.com\/blog\/2015\/1\/15\/machine-learning-and-magical-thinking\" rel=\"nofollow noreferrer\">talk with Ilya Sutskever<\/a> where he discusses how random initializations that are too large <em>or<\/em> too small can prevent the network from learning.<\/p>\n","answers":[{"AnswerId":"42638634","CreationDate":"2017-03-07T01:34:11.430","ParentId":null,"OwnerUserId":"826970","Title":null,"Body":"<p>You can create a <code>tf.constant()<\/code> variable from a numpy array by feeding numpy array into it;<\/p>\n\n<pre><code># Create a np.array however you want, manipulate and feed\nweight_1 = np.random.normal(scale=0.1, size=(256, 64))\nweight_1[weight_1 &lt; 0] = weight_1[weight_1 &lt; 0] - 0.01\nweight_1[weight_1 &gt; 0] = weight_1[weight_1 &gt; 0] + 0.01\nw = tf.constant(weight_1)\n<\/code><\/pre>\n\n<p>Or you can manipulate a tensorflow tensor before <code>initialize_all_variables()<\/code> by its <code>initialized_value()<\/code> function and create a tensor from that.<\/p>\n\n<pre><code>weight_1 = tf.Variable(tf.random_normal([256, 64], stddev=0.35),\n                       name=\"weight_1\")\n# code below will give you the weight_1 * 2\nweights = tf.Variable(weight_1.initialized_value() * 2.0)\n<\/code><\/pre>\n\n<p>And you can somehow math your way out to values you need with that, but numpy way seems much more easier.<\/p>\n\n<p>Also what you are doing looks a lot like <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/layers\/xavier_initializer\" rel=\"nofollow noreferrer\">xavier initialization<\/a>, which you can use as following;<\/p>\n\n<pre><code>weight_1 = tf.get_variable('weight_1', shape=(256, 100),\n                initializer=tf.contrib.layers.xavier_initializer())\n<\/code><\/pre>\n"}]}
{"QuestionId":42638456,"AnswerCount":1,"Tags":"<python><tensorflow><placeholder>","CreationDate":"2017-03-07T01:12:22.067","AcceptedAnswerId":null,"OwnerUserId":7669285.0,"Title":"TensorFlow get values in a placeholder","Body":"<p>I'm new with TensorFlow. I have some issues understanding. I would like to see the values inside the placeholder when I feed them <\/p>\n\n<pre><\/pre>\n\n<p>Did a command as:<\/p>\n\n<pre><\/pre>\n\n<p>Result:<\/p>\n\n<pre><\/pre>\n\n<p>e.g. I don t need the shape, type... of the placeholder but the content.<\/p>\n\n<p>Thanks for your help.<\/p>\n","answers":[{"AnswerId":"42638809","CreationDate":"2017-03-07T01:55:23.827","ParentId":null,"OwnerUserId":"7669285","Title":null,"Body":"<p>The placeholder must be feed:<\/p>\n\n<pre><code>print(sess.run(Y,feed_dict={X:[[1,2],[3,4]]}))\n<\/code><\/pre>\n"}]}
{"QuestionId":42639413,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-03-07T03:04:55.940","AcceptedAnswerId":42709544.0,"OwnerUserId":7663762.0,"Title":"Check failed: 1 == NumElements() (1 vs. 1792)Must have a one element tensor in Tensorflow C++","Body":"<p>In the Python code, the image data is assigned to tensor image_batch:<\/p>\n\n<p>Part of the code:<\/p>\n\n<pre><\/pre>\n\n<p>C++ code:<\/p>\n\n<pre><\/pre>\n\n<p>Err:<\/p>\n\n<p>F tensorflow\/core\/framework\/tensor.cc:493] Check failed: 1 == NumElements() (1 vs. 1792)Must have a one element tensor<\/p>\n\n<p>Process finished with exit code 6<\/p>\n\n<h3>What is the reason for err?<\/h3>\n","answers":[{"AnswerId":"42709544","CreationDate":"2017-03-10T03:16:12.397","ParentId":null,"OwnerUserId":"7663762","Title":null,"Body":"<p>Before that, I made a stupid mistake and did not find the key error code.<\/p>\n\n<pre><code>auto output_c = outputs[0].scalar&lt;float&gt;();\n<\/code><\/pre>\n\n<p>replace:<\/p>\n\n<pre><code>auto output_c = outputs[0].flat&lt;float&gt;();\n<\/code><\/pre>\n\n<p>all the problems are solved.<\/p>\n\n<p><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3362\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/3362<\/a> (prompted me to check what code)<\/p>\n\n<p>@Aziuth Thanks. <\/p>\n"}]}
{"QuestionId":42639633,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-07T03:26:33.620","AcceptedAnswerId":null,"OwnerUserId":7669891.0,"Title":"convert from one-hot encoding to class label","Body":"<p>I have tensor named y, which has values from one-hot encoding over class labels:\ny = [[ 0.  0.  1. ...,  0.  0.  0.],[ 1.  0.  0. ...,  0.  0.  0.],[ 0.  0.  0. ...,  0.  1.  0.],\n...,[ 0.  0.  0. ...,  0.  0.  0.],[ 0.  0.  0. ...,  0.  0.  1.],[ 0.  0.  1. ...,  0.  0.  0.]] \nso here first row has third element as '1' so it represents class label \nfor that image.\nAm trying to get all class labels from the given one-hot encoded array,\nor the given example it should be something like this:\ny = [2,0,8,...,9,2]<\/p>\n","answers":[]}
{"QuestionId":42639750,"AnswerCount":0,"Tags":"<python><keras><conv-neural-network>","CreationDate":"2017-03-07T03:39:32.587","AcceptedAnswerId":null,"OwnerUserId":7654132.0,"Title":"How to perform 1d convolution using 2d kernel","Body":"<p>I a currently trying to perform a 1d convolution on a image using a 2d kernel. The 1d convolution should be performed across the X-axis of the image, the images are spectograms of audio files.  The image in which the convolution in performed is divided in to 13 section, each section centers around a center frequency, in which 13 different CNN network will perform 1d convolution to extract in total (x,13) feature points, and individually (x,1) feature points. x is given by the length of <\/p>\n\n<p>The kernel size is defined as such the center element will be the center frequency. <\/p>\n\n<p>This is how the data is currently stored:<\/p>\n\n<pre><\/pre>\n\n<p>Now i need to define the network structure such that it performs the 1d convolution.. But how do i ensure it also does that? and performs the way i want it to do.. I guess i have to use convolution2d, but how do i fix it to only perform 1d convoulution?<\/p>\n","answers":[]}
{"QuestionId":42639852,"AnswerCount":0,"Tags":"<python-3.x><machine-learning><neural-network><keras><lstm>","CreationDate":"2017-03-07T03:50:32.090","AcceptedAnswerId":null,"OwnerUserId":7381909.0,"Title":"Running LSTM on keras.io for multivariable time-series data (LONG)","Body":"<p>I have a data-set of quarterly Economic variables and I want to try and predict some binary variable (government of the day) by utilising an  framework.<\/p>\n\n<p>The data-set would look something like:<\/p>\n\n<pre><\/pre>\n\n<p>that is: I want to use  in periods  to be able to predict  in period .<\/p>\n\n<p>I have got thus far using the sequential model...<\/p>\n\n<pre><\/pre>\n\n<p>my data has the shape:<\/p>\n\n<pre><\/pre>\n\n<p>I had it this shape to reflect a single 'time piece' being 16 quarters (4 years).<\/p>\n\n<p>I define the architecture in the following way:<\/p>\n\n<pre><\/pre>\n\n<p>However, upon trying to train the model I encounter some problems:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following exception:<\/p>\n\n<pre><\/pre>\n\n<p>I am quite new to making RNN's so excuse me if I am ignorant in my questions or how I attempted this problem. Any advice on how to continue and to resolve this Exception would be greatly appreciated!<\/p>\n\n<p>credit to <a href=\"http:\/\/danielhnyk.cz\/predicting-sequences-vectors-keras-using-rnn-lstm\/\" rel=\"nofollow noreferrer\">http:\/\/danielhnyk.cz\/predicting-sequences-vectors-keras-using-rnn-lstm\/<\/a> as a major source for how I tried to do this.<\/p>\n","answers":[]}
{"QuestionId":42639857,"AnswerCount":0,"Tags":"<pandas><numpy><theano><keras><keras-layer>","CreationDate":"2017-03-07T03:50:58.793","AcceptedAnswerId":null,"OwnerUserId":3832333.0,"Title":"Bad input argument to theano function: 'setting an array element with a sequence.'","Body":"<p>I am using Keras with Theano backend, The following code throws the error as mentioned in the title. <\/p>\n\n<pre><\/pre>\n\n<p>The error generated looks like:<\/p>\n\n<pre><\/pre>\n\n<p>To give you the idea of dimensions of X and Y. <\/p>\n\n<pre><\/pre>\n\n<p>I would really appreciate any kind of help. Thanks.<\/p>\n","answers":[]}
{"QuestionId":42640212,"AnswerCount":1,"Tags":"<deep-learning><keras><lstm><recurrent-neural-network><keras-layer>","CreationDate":"2017-03-07T04:25:17.207","AcceptedAnswerId":null,"OwnerUserId":7108418.0,"Title":"Keras- LSTM- Input Size Error","Body":"<p>I have varying length inputs. (below is the sample inputs) <\/p>\n\n<pre><\/pre>\n\n<p>So, I have padded the input sequence as follows-<\/p>\n\n<pre><\/pre>\n\n<p>Now I need to reshape the input to be of [samples, time steps, features] to implement LSTM layer as per keras documentation.<\/p>\n\n<p>But when i reshape the input padded array as - <\/p>\n\n<p>X_reshaped = X_padd.reshape(X_padd.shape[1], max_sequence_length, X_padd.shape[0])<\/p>\n\n<p>It throws the below error. Please help me resolve this. Thanks.<\/p>\n\n<pre><\/pre>\n\n<h1>------updated-----<\/h1>\n\n<pre><\/pre>\n\n<p>on fitting data to the model, below is the error I am getting-<\/p>\n\n<p><strong>Exception: Error when checking model input: expected lstm_input_4 to have shape (None, 10418, 694) but got array with shape (10418, 694, 1)<\/strong><\/p>\n","answers":[{"AnswerId":"42641168","CreationDate":"2017-03-07T05:55:01.187","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>As I understand it you don't have features here. You have sequences of numbers, not sequences of vectors. Your shape is <code>(n_samples, time_step)<\/code>. <\/p>\n\n<p>So If you want to make a 3D tensor to input :<\/p>\n\n<pre><code>X_Reshaped = X_pad.reshape(X_pad[0], X_pad[1], 1)\n<\/code><\/pre>\n\n<p>Remember that <code>X_pad[1]<\/code> is your <code>max_sequence_length<\/code>. So you were trying to reshape a tensor shape<code>(13189,694)<\/code> into a <code>(13189,694,694)<\/code>. The second one has more values, hence the complaining. <\/p>\n\n<p>I hope this helps<\/p>\n\n<p><strong>EDIT :<\/strong><\/p>\n\n<p>Your training data has a shape <code>(n_samples, time_steps, num_feat)<\/code> after the reshape.\nTherefore, the input data to your lstm will have a shape of <code>(batch_size, time_steps, features)<\/code>. So when you specify <code>input_length<\/code> and <code>input_dim<\/code>  you should put the time_steps and the num_feat values instead of n_samples and time_steps.<\/p>\n\n<p>So change :<\/p>\n\n<pre><code>input_length = X_train.shape[1]\ninput_dim = X_train.shape[2]\n<\/code><\/pre>\n"}]}
{"QuestionId":42640367,"AnswerCount":0,"Tags":"<python><windows><caffe><pycaffe>","CreationDate":"2017-03-07T04:41:25.357","AcceptedAnswerId":null,"OwnerUserId":7596581.0,"Title":"Many bounding boxes when running faster rcnn demo.py","Body":"<p>I use pyfaster-rcnn to test demo.py,but the results have too many bouning boxes,I have tried to change CONF_THRESH and NMS_THRESH,but it didn't work very well,any suggestions? Thank you!<\/p>\n","answers":[]}
{"QuestionId":42640490,"AnswerCount":3,"Tags":"<python-3.x><machine-learning><tensorflow>","CreationDate":"2017-03-07T04:53:23.673","AcceptedAnswerId":null,"OwnerUserId":687483.0,"Title":"tf.global_variables_initializer() does not work","Body":"<p>Hello Tensorflow users\/developers,<\/p>\n\n<p>Even though I call initializer function, reporter tells me that none of my variable is initialized. I created them using <strong>tf.get_variable()<\/strong>. Here is where my session and graph objects are created:<\/p>\n\n<pre><\/pre>\n\n<p>I've called this method for millions of times before, and it had worked perfectly; but right now it is leaving me in the lurch. What do you think the cause might be? Any suggestion would really be appreciated.<\/p>\n\n<p>P.S. I tried calling <strong>tf.local_variables_initializer()<\/strong> too; though reporter told me that you don't have any local at all.<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42659121","CreationDate":"2017-03-07T21:58:12.617","ParentId":null,"OwnerUserId":"687483","Title":null,"Body":"<p>Thanks for the reply. <\/p>\n\n<p>Well I've figured it out. I shouldn't have executed the following assignment instruction before I build my model:<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\n<\/code><\/pre>\n\n<p><em>For anyone's information<\/em>: You may think that \"<strong>I'll execute and get the result of this operation called 'init' when I do so in a Session. So it doesn't matter where I do the assignment specified above<\/strong>\".<\/p>\n\n<p>No! It is not true. Tensorflow decides on which variables to be initialized right after this assignment instruction is executed. Thus, call it after you build your entire model.<\/p>\n"},{"AnswerId":"45987373","CreationDate":"2017-08-31T17:47:13.873","ParentId":null,"OwnerUserId":"8544732","Title":null,"Body":"<p>I got the same error. However this is my solution: just skip the init = tf.global_variables_initializer()<\/p>\n\n<p>and just use :\nsess = tf.Session\nsess.run(init = tf.global_variables_initializer())<\/p>\n"},{"AnswerId":"42642776","CreationDate":"2017-03-07T07:40:18.630","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>If it does not exist I suspect you accidentally downgraded you Tensorflow version. \nCan you try tf.initialize_all_variables ? <\/p>\n\n<p>If this does not work, can you post what version you are using?<\/p>\n"}]}
{"QuestionId":42640757,"AnswerCount":2,"Tags":"<python-2.7><machine-learning><tensorflow><classification><image-recognition>","CreationDate":"2017-03-07T05:17:36.373","AcceptedAnswerId":44688500.0,"OwnerUserId":7549634.0,"Title":"How to load retrained_graph.pb and retrained_label.txt using pycharm editor","Body":"<p>Using <a href=\"https:\/\/codelabs.developers.google.com\/codelabs\/tensorflow-for-poets\/#0\" rel=\"nofollow noreferrer\">pete warden<\/a> tutorials i had trained the inception network and training of which i am getting two files <br><br>\n1.retrained_graph.pb <br>\n2.retrained_label.txt<\/p>\n\n<p>Using this i wanted to classify the flower image.\nI had install pycharm and linked all the tensorflow library , i had also test the sample tensorflow code it is working fine.<\/p>\n\n<p>Now when i run the label_image.py program which is <\/p>\n\n<pre><\/pre>\n\n<p>i am getting this error message<\/p>\n\n<pre><\/pre>\n\n<p>Could any one please help me with this issue.<\/p>\n","answers":[{"AnswerId":"42640833","CreationDate":"2017-03-07T05:25:13.400","ParentId":null,"OwnerUserId":"826970","Title":null,"Body":"<p>You are trying to get the command line argument by calling <code>sys.argv[1]<\/code>. So you need to give command line arguments to satisfy it. Looks like the argument required is a test image, you should pass its location as a parameter. <\/p>\n\n<p>Pycharm should have a script parameters and interpreter options dialog which you can use to enter the required parameters.<\/p>\n\n<p>Or you can call the script from a command line and enter the parameter via;<\/p>\n\n<pre><code>&gt;python my_python_script.py my_python_parameter.jpg\n<\/code><\/pre>\n\n<hr>\n\n<p><strong>EDIT:<\/strong><\/p>\n\n<p>According to the <a href=\"https:\/\/www.jetbrains.com\/help\/pycharm\/2016.3\/run-debug-configuration-python.html\" rel=\"nofollow noreferrer\">documents<\/a> (I don't have pycharm installed on this computer), you should go to <strong>Run\/Debug configuration<\/strong> menu and edit the configurations for your script. Add the absolute path of your file into <strong>Script Parameters<\/strong> box in quotes.<\/p>\n\n<p>Or alternatively if you just want to skip the parameter thing completely, just get the path as <code>raw_input<\/code> (<code>input<\/code> in python3) or just simply give it to <code>image_path = r\"absolute_image_path.jpg\"<\/code> <\/p>\n"},{"AnswerId":"44688500","CreationDate":"2017-06-22T01:06:57.643","ParentId":null,"OwnerUserId":"7386732","Title":null,"Body":"<p>You are getting this error because it is expecting image name (with path) as an argument.<\/p>\n\n<p>In pycharm go to View->Tool windows->Terminal.<\/p>\n\n<p>It is same as opening separate terminal. And run <\/p>\n\n<pre><code>python label_image.py \/image_path\/image_name.jpg\n<\/code><\/pre>\n"}]}
{"QuestionId":42641034,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><convolution><face-recognition><loss>","CreationDate":"2017-03-07T05:43:48.273","AcceptedAnswerId":null,"OwnerUserId":5915270.0,"Title":"How to take the gradients of the center loss function described in tensorflow openface ?","Body":"<p>I have read the paper that has mentioned the new loss function which adds more discriminative property for the classifier. This is the link \n<a href=\"http:\/\/ydwen.github.io\/papers\/WenECCV16.pdf\" rel=\"noreferrer\">http:\/\/ydwen.github.io\/papers\/WenECCV16.pdf<\/a>\nIn the equation 3 they have mentioned how to take the gradients of the center variable w.r.t the loss. But it's not very clear how they did it. <\/p>\n","answers":[]}
{"QuestionId":42641330,"AnswerCount":1,"Tags":"<python><machine-learning><deep-learning><keras><lasagne>","CreationDate":"2017-03-07T06:08:06.663","AcceptedAnswerId":42647376.0,"OwnerUserId":4013781.0,"Title":"Convert Lasagne Model to Keras","Body":"<p>I want to convert a Lasagne model specification to Keras. What is the equivalent layer in Keras to this in Lasagne:<\/p>\n\n<pre><\/pre>\n\n<p>The Convolution3D layer specification for Keras is: <\/p>\n\n<pre><\/pre>\n\n<p>...and for Lasagne:<\/p>\n\n<pre><\/pre>\n\n<p>So, in the above example, the Lasagne layer has 'nn' incoming, 8 filters, and filter size 3. <\/p>\n\n<p>However, Keras requires that each <strong>kernel_dim<\/strong> be specified. Are they all just 3?<\/p>\n\n<p>Thank you. <\/p>\n","answers":[{"AnswerId":"42647376","CreationDate":"2017-03-07T11:39:54.243","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>As you may read <a href=\"https:\/\/lasagne.readthedocs.io\/en\/latest\/modules\/layers\/dnn.html#lasagne.layers.dnn.Conv3DDNNLayer\" rel=\"nofollow noreferrer\">here<\/a>:<\/p>\n\n<blockquote>\n  <p>filter_size : int or iterable of int<\/p>\n  \n  <p>An integer or a 3-element tuple specifying the size of the filters.<\/p>\n<\/blockquote>\n\n<pre><code>nn = Conv3DDNNLayer(nn, 8, 3)\n<\/code><\/pre>\n\n<p>is equivalent to:<\/p>\n\n<pre><code>model.add(Convolution3D(8, 3, 3, 3, ...)\n<\/code><\/pre>\n\n<p>or <\/p>\n\n<pre><code>conv_3d_output = Convolution3D(8, 3, 3, 3, ...)(conv_3d_input)\n<\/code><\/pre>\n\n<p>depending on which <code>Keras.API<\/code> you are using.<\/p>\n"}]}
{"QuestionId":42641824,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><neural-network><keras>","CreationDate":"2017-03-07T06:40:28.447","AcceptedAnswerId":45539517.0,"OwnerUserId":3993741.0,"Title":"keras model.fit_generator() several times slower than model.fit()","Body":"<p>Even as of Keras 1.2.2, referencing <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2730\" rel=\"noreferrer\">merge<\/a>, it does have multiprocessing included, but  is still about 4-5x slower than  due to disk reading speed limitations. How can this be sped up, say through additional multiprocessing?<\/p>\n","answers":[{"AnswerId":"43601365","CreationDate":"2017-04-25T04:02:51.550","ParentId":null,"OwnerUserId":"4438782","Title":null,"Body":"<p>I had a similar problem where I switched to dask to load the data into memory rather than using a generator where I was using pandas. So, depending on your data size, if possible, load the data into memory and use the fit function.<\/p>\n"},{"AnswerId":"45539517","CreationDate":"2017-08-07T04:38:58.353","ParentId":null,"OwnerUserId":"6458334","Title":null,"Body":"<p>You may want to check out the <code>workers<\/code> and <code>max_queue_size<\/code> parameters of <code>fit_generator()<\/code> in the <a href=\"https:\/\/keras.io\/models\/model\/\" rel=\"nofollow noreferrer\">documentation<\/a>. Essentially, more <code>workers<\/code> creates more threads for loading the data into the queue that feeds data to your network. There is a chance that filling the queue might cause memory problems, though, so you might want to decrease <code>max_queue_size<\/code> to avoid this.<\/p>\n"}]}
{"QuestionId":42642719,"AnswerCount":1,"Tags":"<python><computer-vision><deep-learning><caffe><object-recognition>","CreationDate":"2017-03-07T07:37:17.317","AcceptedAnswerId":42896531.0,"OwnerUserId":null,"Title":"Train Faster RCNN on Custom Dataset","Body":"<p>I am trying to run  for 2 class problem say,  and . <\/p>\n\n<ol>\n<li>I have stored the data in Pascal format.<\/li>\n<li>I also changed the number of output in last layers of the caffe definition to 3 ( i.e. 2 + BG ) and 12 ( 4 * 3 ).<\/li>\n<li>For some reasons, I cannot use pretrained model.<\/li>\n<\/ol>\n\n<p>Now, I wish to train the  model from scratch for above mentioned two classes. So, my question is; If I am running the following command, <\/p>\n\n<pre><\/pre>\n\n<p>Is it working as per my requirement (or, what I am thinking) i.e. object classification and localization for  and . Or, something else is happening? Because, my AP values are very low (; ) for the model I obtained by running it for 10000 iterations. <\/p>\n","answers":[{"AnswerId":"42896531","CreationDate":"2017-03-20T05:40:40.760","ParentId":null,"OwnerUserId":"1668943","Title":null,"Body":"<p>If you have arranged your data for <code>obj1<\/code> and <code>obj2<\/code> classes in PASCAL format then you are correctly running the training code.<\/p>\n\n<p>I assume that you have taken special care in creating the text files which mention about the training and testing samples. Please, see PASCAL setup for reference.<\/p>\n"}]}
{"QuestionId":42643288,"AnswerCount":2,"Tags":"<tensorflow><deep-learning><conv-neural-network>","CreationDate":"2017-03-07T08:14:38.370","AcceptedAnswerId":42643904.0,"OwnerUserId":2845263.0,"Title":"How to use tensorflow tf.metrics.mean_iou?","Body":"<p>I am trying to use the inbuilt mean_iou function of tensorflow to compute the IoU score for semantic segmentation. <\/p>\n\n<p>My code is: <\/p>\n\n<pre><\/pre>\n\n<p>However I am getting the following error trace:<\/p>\n\n<pre><\/pre>\n\n<p>Please guide me on the correct usage of this for semantic segmentation.<\/p>\n","answers":[{"AnswerId":"49326455","CreationDate":"2018-03-16T17:29:21.357","ParentId":null,"OwnerUserId":"5339857","Title":null,"Body":"<p>Simplest form I could come up with (3 classes):<\/p>\n\n<pre><code># y_pred and y_true are np.arrays of shape [1, size, channels]\nwith tf.Session() as sess:\n    ypredT = tf.constant(np.argmax(y_pred, axis=-1))\n    ytrueT = tf.constant(np.argmax(y_true, axis=-1))\n    iou,conf_mat = tf.metrics.mean_iou(ytrueT, ypredT, num_classes=3)\n    sess.run(tf.local_variables_initializer())\n    sess.run([conf_mat])\n    miou = sess.run([iou])\n    print(miou)\n<\/code><\/pre>\n\n<p>prints:<\/p>\n\n<pre><code>[0.6127908]\n<\/code><\/pre>\n"},{"AnswerId":"42643904","CreationDate":"2017-03-07T08:54:44.527","ParentId":null,"OwnerUserId":"2845263","Title":null,"Body":"<p>I solved it by calling<\/p>\n\n<pre><code>with tf.Session() as sess:\n\n    sess.run(tf.global_variables_initializer())\n    sess.run(tf.local_variables_initializer())\n<\/code><\/pre>\n"}]}
{"QuestionId":42643736,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-07T08:44:26.993","AcceptedAnswerId":null,"OwnerUserId":1079110.0,"Title":"How to ensure a tf.Variable has been initialized when using its value?","Body":"<p>How can I get the current value of a variable while ensuring that it has been initialized already?  has a dependency on the initializer that causes the variable to be reset to its initial value every time it is accessed. To prevent the variable from being reset, I tried to use  with  as predicate. However, this does not work since the true-branch of the conditional requires the variable to be initialized, even though the false-branch is active:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"48941292","CreationDate":"2018-02-23T04:45:55.420","ParentId":null,"OwnerUserId":"6235205","Title":null,"Body":"<p>Use the <code>initialized_value()<\/code> method on the <code>Variable<\/code> class:\n<a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.5\/tensorflow\/python\/ops\/variables.py#L533\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.5\/tensorflow\/python\/ops\/variables.py#L533<\/a><\/p>\n\n<p>From the doc string:<\/p>\n\n<pre><code># Initialize 'v' with a random tensor.\nv = tf.Variable(tf.truncated_normal([10, 40]))\n# Use `initialized_value` to guarantee that `v` has been\n# initialized before its value is used to initialize `w`.\n# The random values are picked only once.\nw = tf.Variable(v.initialized_value() * 2.0)\n<\/code><\/pre>\n"}]}
{"QuestionId":42644658,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-07T09:35:24.160","AcceptedAnswerId":null,"OwnerUserId":870810.0,"Title":"Tensorflow OOM after freeze graph","Body":"<p>I'm running a seq2seq model with tf, the inference program runs well when loading parameters from checkpoint file using . But after exporting the graph with  (using ), and import with  in the inference program, it got OOM problem.<\/p>\n\n<p>Here is a part of error log:<\/p>\n\n<pre><\/pre>\n\n<p>I thought it might cause by the memory issue of . Does someone have experience with this problem?<\/p>\n","answers":[{"AnswerId":"59036745","CreationDate":"2019-11-25T17:01:40.387","ParentId":null,"OwnerUserId":"895077","Title":null,"Body":"<p>I had the same issue but when trying to load and run the inference from a C++ application using the C API. After a lot of twiddling and testing it appeared the culprit was the frozen graph and <code>freeze_graph.py<\/code> itself. It's probably a bug of some kind. There are actually multiple issue reports on github's TF repo, but they were just closed due to lack of activity, e.g. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/19837\" rel=\"nofollow noreferrer\">here<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/16592\" rel=\"nofollow noreferrer\">here<\/a>. I guess apparent bugs of model freezing aren't of any priority.<\/p>\n\n<p>In my case the model .pb file was around 500mb and it took around 10Gb of RAM while running a session. Not only did it occupy an insane amount of RAM, it was actually orders of magnitudes slower that way.<\/p>\n\n<p>When I switched to loading just a SavedModel directory everything went to normal. I'm not sure how to achieve that in python, but for C code I replaced a <code>TF_GraphImportGraphDef()<\/code> call with <code>TF_LoadSessionFromSavedModel()<\/code>.<\/p>\n\n<p>I used TF v1.14.0. The library is built with Bazel by me, not the stock version. I could provide some details here and there if anybody was interested. Just not sure where to start, I had many trials and errors.<\/p>\n"}]}
{"QuestionId":42644724,"AnswerCount":1,"Tags":"<python><tensorflow><recurrent-neural-network>","CreationDate":"2017-03-07T09:38:32.903","AcceptedAnswerId":42647930.0,"OwnerUserId":4296772.0,"Title":"Trying to share variable basic_rnn_cell\/weights, but specified shape error","Body":"<p>I'm trying to code and learn RNN from scratch in tensorflow. I've understood that we need to create a basicRNNCell and call it as rnn(data, input) with the right dimensions. But I'm getting dimension error as shown below<\/p>\n\n<p>Here's the code that I wrote.<\/p>\n\n<pre><\/pre>\n\n<p>And here's the error<\/p>\n\n<blockquote>\n  <p><strong>ValueError:<\/strong> Trying to share variable basic_rnn_cell\/weights, but specified shape (14, 10) and found shape (6, 4).<\/p>\n<\/blockquote>\n\n<p>What am I doing wrong?<\/p>\n","answers":[{"AnswerId":"42647930","CreationDate":"2017-03-07T12:09:09.703","ParentId":null,"OwnerUserId":"3531912","Title":null,"Body":"<p>I think that you are doing something wrong with your parameters. I will report here a working code for you:<\/p>\n\n<pre><code>import tensorflow as tf\nbatch_size = 2\nseq_len = 4\nrnn_output_dim = 10\n\nsess = tf.InteractiveSession()\nx = tf.placeholder(dtype=tf.float32, shape=(batch_size, seq_len))\nrnn_cell = tf.contrib.rnn.BasicRNNCell(rnn_output_dim)\nstate = rnn_cell.zero_state(batch_size, dtype=tf.float32)\noutput, state = rnn_cell(x, state)\n\nsess.run(tf.global_variables_initializer())\n\nres = sess.run(output, {x: [[1,2,1,1],[0,0,0,1]]})\n\nprint(res)\n\"\"\"\narray([[ 0.21117647, -0.66317081,  0.89524043, -0.54004282, -0.80074871,\n         0.86230665, -0.77558851,  0.46061009,  0.09429809,  0.17166322],\n       [ 0.42703518, -0.18116307,  0.32701704,  0.02224555, -0.39941645,\n         0.10977989, -0.15780419,  0.41289148,  0.35284221, -0.21626833]], dtype=float32)\n\"\"\"\n<\/code><\/pre>\n"}]}
{"QuestionId":42644808,"AnswerCount":1,"Tags":"<tensorflow><sparse-matrix>","CreationDate":"2017-03-07T09:42:40.823","AcceptedAnswerId":42672038.0,"OwnerUserId":4584564.0,"Title":"Get nonzeros row of a SparseTensor","Body":"<p>I want to get all nonzeros from a row in a SparseTensor so \"m\" is the sparse tensor object that I have and row is a row I want to get all nonzeros values and indices from. So I want to return an array of the pair that is [(index, values)]. I hope I can get som help on the subject.<\/p>\n\n<pre><\/pre>\n\n<p>Error message in terminal<\/p>\n\n<pre><\/pre>\n\n<p>EDIT:\nInput for nonzeros\ncm is a coo_matrix with values<\/p>\n\n<pre><\/pre>\n\n<p>if the data is <\/p>\n\n<pre><\/pre>\n\n<p>result should be <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42672038","CreationDate":"2017-03-08T12:58:46.340","ParentId":null,"OwnerUserId":"1373669","Title":null,"Body":"<p>The problem is that <code>index<\/code> inside the lambda is a Tensor and you can't use that directly to index into e.g. <code>indices<\/code>. You could use <code>tf.gather<\/code> instead. Also, you didn't make use of the <code>row<\/code> parameter in the code you posted.<\/p>\n\n<p>Try this instead:<\/p>\n\n<pre><code>import tensorflow as tf\nimport numpy as np\n\ndef nonzeros(m, row):\n    indices = m.indices\n    values = m.values\n    userindices = tf.where(tf.equal(indices[:, 0], row))\n    found_idx = tf.gather(indices, userindices)[:, 0, 1]\n    found_vals = tf.gather(values, userindices)[:, 0:1]\n    res = tf.concat(1, [tf.expand_dims(tf.cast(found_idx, tf.float64), -1), found_vals])\n    return res\n\ndata = np.array([[0.,  1.,  0.,  0.,  0.,  0.,  0.,  0.,  0.,  1.],\n                [0., 0.,  0.,  0.,  1.,  0.,  0.,  0.,  0.,  2.]])\n\nm = tf.SparseTensor(indices=np.array([[0, 1], [0, 9], [1, 4], [1, 9]]),\n                    values=np.array([1.0, 1.0, 1.0, 2.0]),\n                    shape=[2, 10])\n\nwith tf.Session() as sess:\n    result = nonzeros(m, 1)\n    print(sess.run(result))\n<\/code><\/pre>\n\n<p>which prints:<\/p>\n\n<pre><code>[[ 4.  1.]\n [ 9.  2.]]\n<\/code><\/pre>\n"}]}
{"QuestionId":42644868,"AnswerCount":0,"Tags":"<caffe>","CreationDate":"2017-03-07T09:45:59.297","AcceptedAnswerId":null,"OwnerUserId":4380286.0,"Title":"why iteration loss does not match the sum of all train net output in caffe","Body":"<p><\/p>\n\n<p>Hi, i am training a CNN network using caffe, i found the iteration loss does not match the sum of all train net outputs.<\/p>\n\n<p>any advise? thanks! <\/p>\n","answers":[]}
{"QuestionId":42645165,"AnswerCount":2,"Tags":"<tensorflow>","CreationDate":"2017-03-07T09:59:48.693","AcceptedAnswerId":null,"OwnerUserId":6850023.0,"Title":"Reevaluate dependencies of a while loop","Body":"<p>I am trying to understand how while loops work in tensorflow. In particular I have a variable, x say, that I update in the while loop, and then I have some values that depends on x, but when running the while loop the values does not seem to be updated when x changes.<\/p>\n\n<p>The following code where I have tried to implement a simple gradient decent optimizer might illustrate what I mean:<\/p>\n\n<pre><\/pre>\n\n<p>Running this on tensorflow 1.0 gives this result for me:<\/p>\n\n<pre><\/pre>\n\n<p>and the issue is that the value of the gradient is not updated as x changes.<\/p>\n\n<p>I have tried various variations on the above code, but can not seem to find a version that works. Basically my question is if the above can be made to work, or do I need to rethink the approach.<\/p>\n\n<p>(Maybe I should add that I am not interested in writing a gradient decent optimizer, I just built this to have something simple and understandable to work with.)<\/p>\n","answers":[{"AnswerId":"42713053","CreationDate":"2017-03-10T07:52:34.960","ParentId":null,"OwnerUserId":"6850023","Title":null,"Body":"<p>With some help from the other answer I managed to get this working. Posting the complete code here as a second answer:<\/p>\n\n<pre><code>x = tf.constant(4, dtype=tf.float32)\ny = tf.multiply(x,x)\n\ngrad = tf.gradients(y, x)\n\ndef loop_grad(x_loop):\n    y2 = tf.multiply(x_loop, x_loop)\n    return tf.gradients(y2, x_loop)[0]\n\niterations = tf.placeholder(tf.int32)\ni = tf.constant(0, dtype=tf.int32)\nc = lambda i_loop, x_loop: i_loop &lt; iterations\nb = lambda i_loop, x_loop: [i_loop+1, x_loop - 0.1*loop_grad(x_loop)]\nl = tf.while_loop(c, b, [i, x], back_prop=False, parallel_iterations=1)\n\ngpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.05)\nwith tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:\n    sess.run(tf.global_variables_initializer())\n    res_g = sess.run(grad)\n    res_l = sess.run(l, feed_dict={iterations: 100000})\n    res_x = sess.run(x)\n\n\nprint(res_g)\nprint(res_l)\nprint(res_x)\n<\/code><\/pre>\n\n<p>changing the learning rate from the code in the question and increasing the number of iterations gives the output:<\/p>\n\n<pre><code>[8.0]\n[100000, 5.1315068e-38]\n4.0\n<\/code><\/pre>\n\n<p>Which seems to be working. It runs reasonably fast even with high iteration count, so there does not seem to be something really horrible going on with updating the graph in each iteration of the while loop, a fear of which probably was one reason why I didn't opt for this approach from the start.<\/p>\n"},{"AnswerId":"42659635","CreationDate":"2017-03-07T22:34:23.207","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>Having <code>tf.Variable<\/code> objects as loop variables for while loops is not supported, and will behave in weird nondeterministic ways. Always use <code>tf.assign<\/code> and friends to update the value of a <code>tf.Variable<\/code>.<\/p>\n"}]}
{"QuestionId":42645394,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-07T10:10:06.920","AcceptedAnswerId":42646896.0,"OwnerUserId":4993842.0,"Title":"Tensorflow classification to regression \"Assign requires shapes of both tensors to match\"","Body":"<p>I'm trying to convert the below classification graph to regression instead so that instead of 3 values it returns just one value -<\/p>\n\n<pre><\/pre>\n\n<p>For it to work as regression, I've changed the fully connected parts - W_fc3, b_fc3, outputs, cross entropy and train_step to have tensor shape dimension of 1 instead of 3 as below (rest of the graph stays the same) -<\/p>\n\n<pre><\/pre>\n\n<p>But it keeps throwing the below error -<\/p>\n\n<blockquote>\n  <p>InvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match. lhs shape= [1] rhs shape= [3]<\/p>\n<\/blockquote>\n\n<p>I can't see a 3 anywhere. What could be wrong?<\/p>\n","answers":[{"AnswerId":"42646896","CreationDate":"2017-03-07T11:17:16.690","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>For future reference: see comment by Achilles: <\/p>\n\n<blockquote>\n  <p>I'm sorry, I just realized what was wrong. The supervisor's logdir was pointing to the previous version (the classification version) of this model. I pointed it to a an empty logdir and it worked. Thanks for looking into it though. \u2013 Achilles<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42645447,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-03-07T10:12:45.077","AcceptedAnswerId":42646222.0,"OwnerUserId":7474000.0,"Title":"is there well defined next_batch function?","Body":"<p>While writing code using neural network, we normally use next_batch function like below :<\/p>\n\n<pre><\/pre>\n\n<p>In case of using other methods such as random forest, svm, baysian network,\nwe don't have to use above code, because all training data are put in it at once.\nBut if we want to use neural network for the data, we have to make code next_batch. Of course I knew tensorflow's next_batch function code, <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/7c36309c37b04843030664cdc64aca2bb7d6ecaa\/tensorflow\/contrib\/learn\/python\/learn\/datasets\/mnist.py#L160\" rel=\"nofollow noreferrer\">here<\/a>,\nBut if the data are as much as hard to load on memory, above method would be improper.\nis there well defined next_batch function possible for huge dataset?<\/p>\n\n<p>Thanks in advance.<\/p>\n","answers":[{"AnswerId":"42646222","CreationDate":"2017-03-07T10:47:19.567","ParentId":null,"OwnerUserId":"2476373","Title":null,"Body":"<p>For huge datasets that can't fit in memory, you can use TensorFlow's built-in support for distributed training. <\/p>\n\n<p>See, the <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/reading_data\" rel=\"nofollow noreferrer\">reading data<\/a> chapter in the TF programmer's guide. Specifically, read the <strong>Batching<\/strong> section.\nThere's also a great <a href=\"http:\/\/www.wildml.com\/2016\/08\/rnns-in-tensorflow-a-practical-guide-and-undocumented-features\/\" rel=\"nofollow noreferrer\">blog post<\/a>, about RNNs, where the first part gives short and concise examples of how to use TF support for distributed training.<\/p>\n"}]}
{"QuestionId":42646400,"AnswerCount":1,"Tags":"<machine-learning><neural-network><visualization><keras><deconvolution>","CreationDate":"2017-03-07T10:55:11.937","AcceptedAnswerId":null,"OwnerUserId":7544784.0,"Title":"How to visualize intermediate feature layers in keras?","Body":"<p>I've read the paper <a href=\"https:\/\/arxiv.org\/abs\/1311.2901\" rel=\"nofollow noreferrer\">Visualizing and Understanding Convolutional Networks by Zeiler and Fergus<\/a> and would like to make use of their visualization technique. \nThe paper sounds promising - but unfortunately, I have no idea how to implement it in Keras (version 1.2.2).<\/p>\n\n<p>Two questions:<\/p>\n\n<ol>\n<li><p>Keras only provides the  Layer but no  and no \"reverse ReLU\" Layer. How can I make use of those switch variables mentioned in the paper in order to implement the unpooling? \nHow do I have to use the reverse ReLU (or is it just the \"normal\" )?<\/p><\/li>\n<li><p>Keras  layer has the attributes  and . \nMaybe those are the key for solving my problem?! \nIf yes, I would have to replace all my combination of Layers  +  +  with a single  Layer, right? <\/p><\/li>\n<\/ol>\n\n<p>I appreciate your help!<\/p>\n","answers":[{"AnswerId":"42648310","CreationDate":"2017-03-07T12:27:13.400","ParentId":null,"OwnerUserId":"1397946","Title":null,"Body":"<p>The authors of the paper you cite (as far as I remember) talk briefly on how to handle this, specifically:<\/p>\n\n<ol>\n<li>ReLU. The inverse of ReLU is... ReLU. Since convolution is applied to activation function in the forward pass, deconvolution should be applied to rectified reconstructions in the backward pass.<\/li>\n<li>Pooling. There is no way to invert pooling strictly speaking. To cite the paper, \"we can obtain an approximate inverse by recording the locations of the maxima within each pooling region in a set of switch variables. In the deconvnet, the unpooling operation uses these switches to place the reconstructions from the layer above into appropriate locations, preserving the structure of the stimulus.\"<\/li>\n<\/ol>\n\n<p>Now, closer to actual implementation and Keras, have a look at <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/378\" rel=\"nofollow noreferrer\">this thread<\/a> - you will find there some examples that you can use immediately.<\/p>\n"}]}
{"QuestionId":42646548,"AnswerCount":1,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-03-07T11:01:46.597","AcceptedAnswerId":null,"OwnerUserId":7261750.0,"Title":"How to use different loss function in tensorflow, DNNLinearCombinedClassifier","Body":"<p>I am using google tensorflow <a href=\"https:\/\/www.tensorflow.org\/tutorials\/wide_and_deep\/\" rel=\"nofollow noreferrer\">TensorFlow Wide &amp; Deep Learning<\/a> to predict CTR.\nThe code was based on <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/learn\/wide_n_deep_tutorial.py\" rel=\"nofollow noreferrer\">wide_n_deep_tutorial.py<\/a>.<\/p>\n\n<p>However, the auc is bad. I checked the code and found that, int the DNN_Linear_Combined.py, the loss function was defined as <\/p>\n\n<pre><\/pre>\n\n<p>In CTR prediction, train data was imbalanced. So if this kind of loss function is used, the loss will be better when predict more samples in negtive class(majority class). <\/p>\n\n<p>What's more, the CTR predicted is much bigger than true CTR. (The true CTR is less than 1%, the mean predicted CTR is about 30%.)<\/p>\n\n<p>My question is: how to modify the code to use logloss? Does the TF provide this?\nI checked the code and found the following:<\/p>\n\n<ol>\n<li>function  in tensorflow\/python\/ops\/math_ops.py<\/li>\n<li>function  in tensorflow\/contrib\/learn\/python\/learn\/estimators\/head.py<\/li>\n<\/ol>\n\n<p>However, the first does not make change with the . And I don't know how to use the second function.<\/p>\n\n<p>In a word, my question can be summarized as: how to modify the code to use the logloss function.<\/p>\n\n<p>Thanks.<\/p>\n","answers":[{"AnswerId":"42666271","CreationDate":"2017-03-08T08:20:51.260","ParentId":null,"OwnerUserId":"1325423","Title":null,"Body":"<p>If you just want to calculate the logloss, and loss is given then you can calculate logloss as below and you don't need to change the code:<\/p>\n\n<pre><code>log_loss = math_ops.log(loss)\n<\/code><\/pre>\n"}]}
{"QuestionId":42647029,"AnswerCount":0,"Tags":"<deep-learning><caffe>","CreationDate":"2017-03-07T11:22:50.697","AcceptedAnswerId":null,"OwnerUserId":1014196.0,"Title":"Computation consideration with different Caffe's network topology (difference in number of output)","Body":"<p>I would like to use one of Caffe's reference model i.e. . I found that my target class i.e. person is one of the classes included in the ILSVRC dataset that has been trained for the model. As my goal is to classify whether a test image contains a person or not, I may achieve this by the following:<\/p>\n\n<ol>\n<li><p>Use inference directly with 1000 number of output. This doesn't\nrequire training\/learning.<\/p><\/li>\n<li><p>Change the network topology a little bit with the final FC layer's number of output () is set to 2 (instead of 1000). Retrain it as a binary classification problem.<\/p><\/li>\n<\/ol>\n\n<p>My concern is about computational effort at deployment\/prediction phase (testing). The latter looks more expensive computationally than the former. This is because during prediction phase it needs to compute those 1000 output possibilities to find the one with the highest score. What I'm not sure is that, it could be the case that there's a heuristic (which I'm not aware of) that simplifies the computation.<\/p>\n\n<p>Can somebody please help cross check my understanding on this.  <\/p>\n","answers":[]}
{"QuestionId":42647403,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-07T11:41:09.350","AcceptedAnswerId":null,"OwnerUserId":7562171.0,"Title":"tf.train.NanTensorHook(loss, fail_on_nan_loss=False) will still raise exception with TF1.0","Body":"<p>When I define a customized model_fn wtih TF1.0, I want to stop the training when loss is Nan. I tried the code below in model_fn:<\/p>\n\n<pre><\/pre>\n\n<p>but the <strong>fail_on_nan_loss=False<\/strong> will still raise an exception, I expect it will write the warning message and stop the specific training without raising an exception.<\/p>\n\n<p>Any suggestion on how to use tf.train.NanTensorHook correctly?<\/p>\n","answers":[{"AnswerId":"42652010","CreationDate":"2017-03-07T15:21:52.923","ParentId":null,"OwnerUserId":"7562171","Title":null,"Body":"<p>When I explore the solution, one possible work around might helps:\nI copy the NanTensorHook class from basic_session_run_hooks.py and make my own version of call inside my model_fn as below<\/p>\n\n<pre><code>        class NanTensorHook2(tf.train.SessionRunHook):\n      \"\"\"NaN Loss monitor by Lei.\n\n      Monitors loss and stops training if loss is NaN.\n      Can either fail with exception or just stop training.\n      \"\"\"\n\n      def __init__(self, loss_tensor, fail_on_nan_loss=True):\n        \"\"\"Initializes NanLoss monitor.\n\n        Args:\n          loss_tensor: `Tensor`, the loss tensor.\n          fail_on_nan_loss: `bool`, whether to raise exception when loss is NaN.\n        \"\"\"\n        self._loss_tensor = loss_tensor\n        self._fail_on_nan_loss = fail_on_nan_loss\n\n      def before_run(self, run_context):  # pylint: disable=unused-argument\n        return tf.train.SessionRunArgs(self._loss_tensor)\n\n      def after_run(self, run_context, run_values):\n        if (np.isnan(run_values.results) or np.isinf(run_values.results)):\n          failure_message = \"Model diverged with loss = NaN or Inf.\"\n          if self._fail_on_nan_loss:\n            logging.error(failure_message)\n            raise NanLossDuringTrainingError\n          else:\n            logging.warning(failure_message)\n            # We don't raise an error but we request stop without an exception.\n            run_context.request_stop()\n<\/code><\/pre>\n\n<p>then use NanTensorHook2 instead, it then started working.<\/p>\n\n<p>Notes that I added <strong>\"np.isinf(run_values.results)\"<\/strong> as I believe that loss = inf should also be checked here.<\/p>\n\n<p>any expert has better solution?<\/p>\n"}]}
{"QuestionId":42648169,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><deep-learning>","CreationDate":"2017-03-07T12:20:48.637","AcceptedAnswerId":42648834.0,"OwnerUserId":5107084.0,"Title":"TensorFlow: Does changing hyperparameters like dropout probability increase GPU memory needed to train?","Body":"<p>After changing some hyperparameters like dropout and weight decay, I found that my batch_size has to be reduced constantly in order to fit the training, even with an 11GB memory Tesla K80. I used to train with batch size 32, but now even a batch size 24 will cause a resource exhausted error after changing my hyperparameters. Why is this the case?<\/p>\n","answers":[{"AnswerId":"42648834","CreationDate":"2017-03-07T12:51:55.010","ParentId":null,"OwnerUserId":"2891324","Title":null,"Body":"<p>The only case in which changing the hyperparameters can change the network size is when the hyperparameters values can enable\/disable a layer.<\/p>\n\n<p>For instance, if you set the <code>keep_prob<\/code> dropout hyperparameter to <code>1<\/code>, the layer itself is not created but the input is returned.<\/p>\n\n<p>The same can happen for the weight decay: if the weight decay penalty is equal to <code>0<\/code> then the decay terms to the loss function are not added.<\/p>\n\n<p>In any other case (for instance for <code>keep_prob<\/code> in <code>]0, 1[<\/code>) the layer is added and its size is constant.<\/p>\n\n<p>The problem should be somewhere else in your graph definition, probably.<\/p>\n"}]}
{"QuestionId":42648668,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-07T12:43:33.187","AcceptedAnswerId":42685313.0,"OwnerUserId":5440222.0,"Title":"What's the difference of FunctionDefHelper::Create and FunctionDefHelper::Define?","Body":"<p>the following is from math_grad.cc line543 to line554:<\/p>\n\n<pre><\/pre>\n\n<p>and the following is from math_grad.cc line593 to line615,<\/p>\n\n<pre><\/pre>\n\n<p>I am a little confused about how does TensorFlow determine whether use\nFDH::Define or FDH::Create.<\/p>\n\n<p>As I know, FDH::Define is an old approch to define a FunctionDef. and I thought that FDH::Define provide less information compared with FDH::Create.<\/p>\n\n<p>So, any one can help me tell the difference? Any advice will be appreciated, thanks a lot.<\/p>\n","answers":[{"AnswerId":"42685313","CreationDate":"2017-03-09T01:59:16.300","ParentId":null,"OwnerUserId":"7128661","Title":null,"Body":"<p>Answer from a colleague:<\/p>\n\n<p>\"FunctionDefs are not yet part of the public API; these are temporary mechanisms for defining functions until we have real API support for them.\nOne of those two functions is older and mirrors the older FunctionDef format, which has now been deleted.<\/p>\n\n<p>I didn't delete the function, I just made it produce the new format; but this only works in simple cases. There was too much code using that function to migrate all of them.<\/p>\n\n<p>But the main point is that this isn't documented well because it is not part of our public API and subject to change.\"<\/p>\n"}]}
{"QuestionId":42648864,"AnswerCount":0,"Tags":"<python><tensorflow><pickle><subplot>","CreationDate":"2017-03-07T12:53:31.480","AcceptedAnswerId":null,"OwnerUserId":6127613.0,"Title":"IndexError: index 0 is out of bounds for axis 0 with size 0 tensorflow label detection","Body":"<p>Hey guys i am using tensorflow and wanted to pass new images and find its prediction along with the image.This is my code:<\/p>\n\n<pre><\/pre>\n\n<p>this is the error i get<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42649779,"AnswerCount":0,"Tags":"<concurrency><tensorflow><seeding>","CreationDate":"2017-03-07T13:37:09.207","AcceptedAnswerId":null,"OwnerUserId":5664271.0,"Title":"TensorFlow: Graph wide seeding and inter- and intra-op parallelism","Body":"<p>By using the  method one can set the graph wide seed for random number generation. <\/p>\n\n<p>However, I'm unsure whether\/how this graph wide seed is working when multi-threaded execution is turned on by using  and  switches.<\/p>\n\n<p><strong>Question:<\/strong> After graph wide seeding using  is the tensorflow random number generation process still deterministic when multiple operations are executed in parallel? Especially in the case when multiple random number generation ops are executed concurrently.<\/p>\n","answers":[]}
{"QuestionId":42650003,"AnswerCount":1,"Tags":"<python><neural-network><deep-learning><keras><keras-layer>","CreationDate":"2017-03-07T13:47:27.470","AcceptedAnswerId":42658527.0,"OwnerUserId":2182857.0,"Title":"Training a dense layer from bottleneck features vs. freezing all layers but the last - should be the same, but they behave differently","Body":"<p>As a \"sanity check\" I tried two ways to use transfer learning that I expected to behave the same, if not in running time than at least in the results.<\/p>\n\n<p>The first method was use of bottleneck features (as explained here  ), that is, using the existing predictor to generate the features just before the last dense layer, saving them, then training a new dense layer with these features as input.<\/p>\n\n<p>The second method was to replace the last dense layer of the model with a new one, then freezing all other layers in model.<\/p>\n\n<p>I expected the second method to be as effective as the first one, but it was not.<\/p>\n\n<p>The output of the first method was<\/p>\n\n<pre><\/pre>\n\n<p>It converges quickly and yields good results.<\/p>\n\n<p>The second method, on the other hand, gives this:<\/p>\n\n<pre><\/pre>\n\n<p>The same model (Inception V4) was used for both methods.\nMy code is as follows:<\/p>\n\n<p>First method (Bottleneck Features):<\/p>\n\n<pre><\/pre>\n\n<p>Second method (freezing all but the last)<\/p>\n\n<pre><\/pre>\n\n<p>I think that freezing all of the net but the top and training only the top should be identical to using all of the net but the top to create the features that exist just before the top, and then train a new dense layer is basically the same thing.<\/p>\n\n<p>So I am either incorrect in my code or my thinking about the problem (or both...)<\/p>\n\n<p>What am I doing wrong?<\/p>\n\n<p>Thank you for your time.<\/p>\n","answers":[{"AnswerId":"42658527","CreationDate":"2017-03-07T21:17:41.073","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>This was a really neat problem. It's because of <code>Dropout<\/code> layers in your second approach. Even though the layer was set to be not <code>trainable<\/code> - <code>Dropout<\/code> still works and prevents your network from overfitting by changing your input.<\/p>\n\n<p>Try to change your code to:<\/p>\n\n<pre><code>v4 = inception_v4.create_model(weights='imagenet')\npredictions = Flatten()(v4.layers[-4].output)\npredictions = Dense(output_dim=num_classes, activation='softmax', name=\"newDense\")(predictions)\n<\/code><\/pre>\n\n<p>Also - because of the <code>BatchNormalization<\/code> change the <code>batch_size<\/code> to <code>24<\/code>.<\/p>\n\n<p>This should work.<\/p>\n"}]}
{"QuestionId":42650107,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-03-07T13:52:04.403","AcceptedAnswerId":null,"OwnerUserId":7605715.0,"Title":"The \"I tensorflow\/stream_executor\/dso_loader.cc:108] successfully opened CUDA library libcublas.so locally\" does not show up","Body":"<p>I've just installed tensorflow with GPU support from source though I had installed via pip before following the steps in <a href=\"https:\/\/www.tensorflow.org\/install\/install_sources\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_sources<\/a>. <\/p>\n\n<p>However, when I import tensorflow I get this:<\/p>\n\n<pre><\/pre>\n\n<p>As can be seen, the following lines do not show up<\/p>\n\n<pre><\/pre>\n\n<p>Is there anything wrong with my configuration? <\/p>\n","answers":[]}
{"QuestionId":42650167,"AnswerCount":1,"Tags":"<python><tensorflow><gpu><cpu>","CreationDate":"2017-03-07T13:54:50.933","AcceptedAnswerId":42659657.0,"OwnerUserId":3659855.0,"Title":"Tensorflow: Run training phase on GPU and test phase on CPU","Body":"<p>I wish to run the training phase of my tensorflow code on my GPU while after I finish and store the results to load the model I created and run its test phase on CPU. <\/p>\n\n<p>I have created this code (I have put a part of it, just for reference because it's huge otherwise, I know that the rules are to include a fully functional code and I apologise about that). <\/p>\n\n<pre><\/pre>\n\n<p>After that, my model has been trained successfully and saved the relevant data, so I wish to load the file and do a final train and test part in the model but using my CPU this time. The reason is the GPU can't handle the whole dataset of mnist.train.images and mnist.train.labels.<\/p>\n\n<p>So, manually I select this part and I run it:<\/p>\n\n<pre><\/pre>\n\n<p>But I get an OMM GPU memory error, which it doesn't make sense to me since I think I have forced the program to rely on CPU. I did not put a command sess.close() in the first (training with batches) code, but I am not sure if this really the reason behind it. I followed this post actually <a href=\"https:\/\/stackoverflow.com\/questions\/37660312\/run-tensorflow-on-cpu\">for the CPU<\/a>\nAny suggestions how to run the last part on CPU only? <\/p>\n","answers":[{"AnswerId":"42659657","CreationDate":"2017-03-07T22:35:54.253","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p><code>with tf.device()<\/code> statements only apply to graph building, not to execution, so doing <code>sess.run<\/code> inside a device block is equivalent to not having the device at all.<\/p>\n\n<p>To do what you want to do you need to build separate training and test graphs, which share variables.<\/p>\n"}]}
{"QuestionId":42651714,"AnswerCount":2,"Tags":"<tensorflow><theano><keras>","CreationDate":"2017-03-07T15:07:53.757","AcceptedAnswerId":48636570.0,"OwnerUserId":4717048.0,"Title":"Vector shift (Roll) in Tensorflow","Body":"<p>Lets say, that we do want to process images (or ndim vectors) using Keras\/TensorFlow.\nAnd we want, for fancy regularization, to shift each input by a random number of positions to the left (owerflown portions reappearing at the right side ).<\/p>\n\n<p>How could it be viewed and solved:<\/p>\n\n<p>1)<\/p>\n\n<p>Is there any variation to numpy roll function for TensorFlow?<\/p>\n\n<p>2) <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"43968913","CreationDate":"2017-05-14T21:21:10.303","ParentId":null,"OwnerUserId":"6437741","Title":null,"Body":"<p>I just had to do this myself, and I don't think there is a tensorflow op to do np.roll unfortunately. Your code above looks basically correct though, except it doesn't roll by ri, rather by (x.shape[1] - ri). <\/p>\n\n<p>Also you need to be careful in choosing your random integer that it is from range(1,x.shape[1]+1) rather than range(0,x.shape[1]), as if ri was 0, then x[:,0:ri] would be empty.<\/p>\n\n<p>So what I would suggest would be something more like (for rolling along dimension 1):<\/p>\n\n<pre><code>x_len = x.get_shape().as_list()[1] \ni = np.random.randint(0,x_len) # The amount you want to roll by\ny = tf.concat([x[:,x_len-i:], x[:,:x_len-i]], axis=1)\n<\/code><\/pre>\n\n<p>EDIT: added missing colon after hannes' correct comment.<\/p>\n"},{"AnswerId":"48636570","CreationDate":"2018-02-06T06:14:45.650","ParentId":null,"OwnerUserId":"9320279","Title":null,"Body":"<p>On the current nightly build of tensorflow (or future version 1.6.0). you can use tf.manip.roll which works just like numpy roll. <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/14953\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/pull\/14953<\/a> .\nTo improve on the answer above you can do:<\/p>\n\n<pre><code># size of x dimension\nx_len = tensor.get_shape().as_list()[1]\n# random roll amount\ni = tf.random_uniform(shape=[1], maxval=x_len, dtype=tf.int32)\noutput = tf.manip.roll(tensor, shift=i, axis=[1])\n<\/code><\/pre>\n"}]}
{"QuestionId":42652075,"AnswerCount":0,"Tags":"<php><linux><python-2.7><tensorflow>","CreationDate":"2017-03-07T15:24:45.110","AcceptedAnswerId":null,"OwnerUserId":7673160.0,"Title":"How to call a python script using tensorflow from Php web server","Body":"<p>I am trying to execute a python script from a php script. The python script named main.py calls many others python scripts including some special libraries like tensorflow for machine learning.<\/p>\n\n<p>I am able to launch simple python scripts with just some print('hello') using exec, system or even popen, but the problem is when I try to execute my main.py\nI already gave all the permissions to all the users.\nWith a quick research I could find that the problem is about the fact that the browser can't find the library tensorflow.<\/p>\n\n<p>If I launch my program from command line, it perfectly works, so does anyone know how I can tell my browser where to get the libraries ?<\/p>\n\n<p>Thank you<\/p>\n","answers":[]}
{"QuestionId":42652161,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><keras><keras-layer>","CreationDate":"2017-03-07T15:29:42.320","AcceptedAnswerId":42656253.0,"OwnerUserId":1361822.0,"Title":"Unable to merge keras models after adding Lambda layer","Body":"<p>I have a list of 3 keras models that each have an output shape of . I also have a common keras base model that produces their input. My goal is to combine the 4 models but to only take the first output from each of the models in the list (so the final output should have shape . My problem occurs when I try to use a Lambda layer to extract the first output from each model.<\/p>\n\n<p>If I omit the Lambda step and simply combine the models as follows, it creates a model that gives the correct output with shape :<\/p>\n\n<pre><\/pre>\n\n<p>The problem occurs when I first use a Lambda layer to extract the first value from each model:<\/p>\n\n<pre><\/pre>\n\n<p>What is the right way to merge these models while only taking the single output value from each one?<\/p>\n\n<p>Note that I can successfully use a Lambda function if I apply it <em>after<\/em> merging the models as follows:<\/p>\n\n<pre><\/pre>\n\n<p>But I would like to know how to apply it <em>before<\/em> the merge.<\/p>\n","answers":[{"AnswerId":"42656253","CreationDate":"2017-03-07T19:05:25.140","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The problem lied with a little inconsistency in <code>Lambda<\/code> slicing. Although the <code>shape<\/code> of output doesn't take into account the <code>batch<\/code> dimension - one should remember that <code>tensor<\/code> provided to <code>Lambda<\/code> layers also have this additional dimension. This is why the following line causes an error:<\/p>\n\n<pre><code>m.add(Lambda(lambda x: x[0], output_shape=(1,)))\n<\/code><\/pre>\n\n<p>This should be changed to:<\/p>\n\n<pre><code>m.add(Lambda(lambda x: x[:,:1], output_shape=(1,))) \n<\/code><\/pre>\n\n<p>Beware of the following way of slicing:<\/p>\n\n<p><code>m.add(Lambda(lambda x: x[:,0], output_shape=(1,)))<\/code><\/p>\n\n<p>as it changes dimensionality of a <code>tensor<\/code>.<\/p>\n"}]}
{"QuestionId":42652392,"AnswerCount":0,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-03-07T15:40:51.723","AcceptedAnswerId":null,"OwnerUserId":7249637.0,"Title":"How to run one part of tensorflow graph with one input and running the other graph in a loop with multiple inputs?","Body":"<p>Having tensorflow graph with two inputs, one part of the graph is processing the first input x and the other part is processing the second input y together with the result of the first part.<\/p>\n\n<p>I would like to reuse the result of first part of the graph for further runs on the second part, something similar to the code below using partial_run:<\/p>\n\n<pre><\/pre>\n\n<p>The above code require to partial_run_setup again for further partial runs after the executing second_part. <\/p>\n","answers":[]}
{"QuestionId":42652526,"AnswerCount":1,"Tags":"<machine-learning><neural-network><keras><lstm><stateful>","CreationDate":"2017-03-07T15:46:27.943","AcceptedAnswerId":null,"OwnerUserId":3605054.0,"Title":"Is this the right way to predict a time series with the stateful LSTM neural network?","Body":"<p>I am using LSTM neural networks (stateful) for time series prediction.<\/p>\n\n<p>I'm hoping that the stateful LSTM can capture the hidden patterns and make a satisfactory prediction (the physical law that cause the variation of the time series is not clear).<\/p>\n\n<p>I have a time series  with a length of  (actual observational data), and my purpose is to predict the future . <\/p>\n\n<p>I suppose predict the next  will be more promising than predict the next  (is that right?). <\/p>\n\n<p>So, I prepare the training data like this (always using  values to predict the next ;  denotes the  element in ):<\/p>\n\n<pre><\/pre>\n\n<p>After the training, I want to use the model to predict the next  values  with , and then predict the next  values  with .<\/p>\n\n<p>Is this the right way?<\/p>\n\n<p>After a lot of reading of documents and examples, I can train a model and make the prediction, but the result seems not satisfactory.<\/p>\n\n<p>To validate the method, I assume that the last  values are unknown, and remove them from  and , then try to train a model and predict them. Lastly, compare the predicted values with the observed values.<\/p>\n\n<p>Any suggestions or comments will be appreciated.<\/p>\n\n<p>The time series looks like this:\n<a href=\"https:\/\/i.stack.imgur.com\/Lf92L.jpg\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Lf92L.jpg\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42655624","CreationDate":"2017-03-07T18:27:43.900","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Your question is really complexed. Before I will try to answer it - I'll share my doubts with you about is it sensible to use <code>LSTM<\/code> for your task. You want to use a really advanced model (<code>LSTM<\/code> are capable to learn really complexed patterns) to a time series which seems to be relatively easy. Moreover - you have a really small amout of data. To be honest - I would try to train simpler and easier methods first (like <code>ARMA<\/code> or <code>ARIMA<\/code>).<\/p>\n\n<p>To answer your question - if your approach is good - it seems to be reasonable. Other reasonable methods are predicting all <code>100<\/code> steps or e.g. <code>50<\/code> steps twice. With <code>10<\/code> steps you might come across error cumulation - but still it might be a good method.<\/p>\n\n<p>As I mentioned earlier - I would rather try easier <code>ML<\/code> method for this task but if you really want to use <code>LSTM<\/code> you may tackle this problem in a following way:<\/p>\n\n<ol>\n<li>Define metaparameters like number of steps ahead you want to predict, the size of input fed to network.<\/li>\n<li>Try to use e.g. grid search in order to find the best value of this metaparameters. Evaluate each setup using <code>k-fold<\/code> crossvalidation.<\/li>\n<li>Retrain final model using the best metaparameter setup.<\/li>\n<\/ol>\n\n<p>You have relatively small amount of data so you may easily find the best values of hyperparameters. This will also show you if your approach is good or not - simply check the results provided by the best solution.  <\/p>\n"}]}
{"QuestionId":42652903,"AnswerCount":1,"Tags":"<deep-learning><caffe><vgg-net>","CreationDate":"2017-03-07T16:05:22.093","AcceptedAnswerId":42674458.0,"OwnerUserId":null,"Title":"caffe: Net not converging when replacing AlexNet with VGG16 but everything else is the same","Body":"<p>I have been using the AlexNet for a pixel-wise regression task (depth estimation). Now I wanted to replace the AlexNet with the VGG net since it is ought to be better.<\/p>\n\n<p>This is the AlexNet I used:<\/p>\n\n<pre><\/pre>\n\n<p>This is the VGG I am using:<\/p>\n\n<pre><\/pre>\n\n<p>learning_rate is: 0.0005 <\/p>\n\n<p>When training the AlexNet the loss converges to roughly 5 and when using the VGG the net does not converge at all. It always stays at 30, even though I have been decreasing the learning_rate and even decreased mult_lr. Does anyone have any ideas what else could be wrong? I am 100% sure that only the .prototxt files are different, everything else is exactly the same.<\/p>\n","answers":[{"AnswerId":"42674458","CreationDate":"2017-03-08T14:49:09.563","ParentId":null,"OwnerUserId":"3871961","Title":null,"Body":"<p>It is known that VGG is hard to train from scratch for large networks: in the paper of Simonyan and Zisserman, it is said in section 3.1. In practice, they first train a \"small\" (A) configuration with random weights then use these weights to initialize larger networks (C-D-E). Moreover, you may need more data to train VGG than AlexNet.<\/p>\n\n<p>In your case, you may consider to fine-tune VGG16 rather than learning it from scratch. Or let use googlenet  that is much lighter (and easier to train) and lead to similar performance on several problems I tested.<\/p>\n"}]}
{"QuestionId":42653638,"AnswerCount":1,"Tags":"<java><tensorflow><grpc><tensorflow-serving><grpc-java>","CreationDate":"2017-03-07T16:41:04.293","AcceptedAnswerId":null,"OwnerUserId":2598910.0,"Title":"How To Create A Java Client For A Tensorflow Model Server","Body":"<p>So far I have managed to export a Tensorflow model and then serve that model using the Tensorflow model server.  I have then been able to make predictions by creating a client which talks to the Tensorflow server over gRPC.<\/p>\n\n<p>Basically I've followed this <a href=\"https:\/\/tensorflow.github.io\/serving\/serving_basic\" rel=\"nofollow noreferrer\">tutorial<\/a><\/p>\n\n<p>This has all been done in Python so far which is fine but I need to be able to call the Tensorflow server using Java.  In other words, I need to create a Java client.<\/p>\n\n<p>My basic understanding is that I need to compile .proto files into Java classes using the Protoc compiler but I'm not sure how to generate .proto files for my model.  I've exported the model but the files are in the \".pb\" format and don't seem to compile to Java classes.<\/p>\n\n<p>Has anyone had any experience of doing this?  If you have, please kindly share your solution.<\/p>\n\n<p>Many thanks,<\/p>\n\n<p>Barry<\/p>\n","answers":[{"AnswerId":"42660739","CreationDate":"2017-03-08T00:15:47.767","ParentId":null,"OwnerUserId":"5839622","Title":null,"Body":"<p><code>.pb<\/code> seems to be the serialized protobuf data. Protobuf definitions are <code>.proto<\/code> files which I guess are somewhere downloadable from the TensorFlow project rather than generated from the model.<\/p>\n"}]}
{"QuestionId":42654130,"AnswerCount":0,"Tags":"<python><keras><keras-layer>","CreationDate":"2017-03-07T17:06:48.830","AcceptedAnswerId":null,"OwnerUserId":2387053.0,"Title":"How to iteratively concatenate multiple layers into one in Keras?","Body":"<p>I am a new user of Python and Keras and I hope your feedback could guide me to the solution.<\/p>\n\n<p>I would like to concatenate several recurrent layers to train a multi-inputs neural network. More specifically, I would like to iteratively create multiple layers and merge all of them into a single one.<\/p>\n\n<p>In order to do that, I create a dictionary of training sets (in this case, two training sets):<\/p>\n\n<pre><\/pre>\n\n<p>Next, I create the recurrent layers and dense layers as follows:<\/p>\n\n<pre><\/pre>\n\n<p>where:<\/p>\n\n<pre><\/pre>\n\n<p>Now, I would like to merge the two layers into a single one. According to this answer <a href=\"https:\/\/stackoverflow.com\/questions\/43196636\/how-to-concatenate-two-layers-in-keras\">How to concatenate two layers in keras?<\/a>, this should work:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42654976,"AnswerCount":0,"Tags":"<python><tensorflow><lstm>","CreationDate":"2017-03-07T17:52:22.763","AcceptedAnswerId":null,"OwnerUserId":7626397.0,"Title":"Tensorflow - GridLSTM + Dynamic RNN - How to deal with dynamic shapes?","Body":"<p>I'm a beginner using Tensorflow and I started in a handwriting recognition task. Currently, I'm testing the GridLSTMCell (<a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\/GridLSTMCell\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/contrib\/rnn\/GridLSTMCell<\/a>) and I'm facing some problems when trying to initialize the dynamic RNN with this type of recurrent cell.<\/p>\n\n<pre><\/pre>\n\n<p>Apparently, the problem is due to a restriction present in GridLSTMCell, which can only deal with inputs with static shape, but as you can see below, in my case the batch_size can vary since I have a training set with batch_size=1 and a fixed test set with 100 examples.<\/p>\n\n<pre><\/pre>\n\n<p>Given this scenario, I wish to know if someone has a workaround, maybe some tip to define the inputs with fixed batch_size = 1 and yet be able to apply the model to a test set with 100 examples. Some answers present in a similar question (<a href=\"https:\/\/stackoverflow.com\/questions\/38442025\/tensorflow-grid-lstm-rnn-typeerror\">Tensorflow Grid LSTM RNN TypeError<\/a>) is related to another Grid LSTM implementation in TF.<\/p>\n\n<p>Sorry about my poor English, I appreciate the support.<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[]}
{"QuestionId":42655141,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-07T18:00:24.640","AcceptedAnswerId":null,"OwnerUserId":2638485.0,"Title":"Tensorflow Define Op Polymorphic on Fully Defined vs not Fully Defined Shape","Body":"<p>If I am defining a custom Op in Tensorflow, is it possible to provide two Kernels for the top that are polymorphic on whether the shape for the inputs are fully defined? For example, I can construct certain structures once at Kernel construction if the shape is fully known \/ defined.<\/p>\n","answers":[{"AnswerId":"42683994","CreationDate":"2017-03-08T23:36:13.870","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>It's not currently possible to do this. The <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2d23960e29c3efb1b30c5cb6f8f3368635b97f29\/tensorflow\/core\/framework\/op_kernel.cc#L998\" rel=\"nofollow noreferrer\">kernel dispatch mechanism<\/a> is implemented in a low-level part of the TensorFlow code where information about tensor shapes is not (generally) available.<\/p>\n\n<p>However, the ability to specialize a graph based on known shapes does seem like a useful ability, and it might be worth raising this as a feature request on the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\" rel=\"nofollow noreferrer\">GitHub issues page<\/a>. One possible workaround would be to try <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2d23960e29c3efb1b30c5cb6f8f3368635b97f29\/tensorflow\/core\/common_runtime\/optimization_registry.h\" rel=\"nofollow noreferrer\">registering an optimization pass<\/a> that makes use of shape information and rewrites the names of ops with known input shapes to a different op that relies on static shape information (e.g. via an additional attr). However, doing this in TensorFlow currently requires you to rebuild from source.<\/p>\n"}]}
{"QuestionId":42655352,"AnswerCount":2,"Tags":"<python><tensorflow><anaconda>","CreationDate":"2017-03-07T18:10:53.960","AcceptedAnswerId":null,"OwnerUserId":2851402.0,"Title":"Trying to install Tensorflow in Anaconda on my PC","Body":"<p>I am in the process of installing Tensorflow in Anaconda on my PC running Windows 10, GTX970, Visual Studio 2015 Community.  As I understand, the GPU version of Tensorflow is now available for Windows. <\/p>\n\n<p>I successfully installed the latest Anaconda with Python 3.6, 64 bit.<\/p>\n\n<p>I installed Visual Studio 2015 Community.<\/p>\n\n<p>I installed CUDA 8.0 following the CUDA instructions.<\/p>\n\n<p>I created a second environment (tensorflowgpu) with conda.  I installed Python 3.5.3 in that environment.  I activated that environment.<\/p>\n\n<p>I tried to install Tensorflow with this command: <strong>pip install --ignore-installed --upgrade <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl<\/a><\/strong><\/p>\n\n<p>I got the following error message: <strong>tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl<\/strong> is not a supported wheel on this platform.<\/p>\n\n<p>There are several places I could check for problems:<\/p>\n\n<ol>\n<li>Verify that CUDA was a good download<\/li>\n<li>Check environment variables (cuDNN, etc.)<\/li>\n<li>Is environment set up properly?<\/li>\n<li>Should I be using Python 3.5 instead of 3.5.3?<\/li>\n<li>Something else<\/li>\n<\/ol>\n\n<p>Does the error message suggest where I should begin?  Thanks.<\/p>\n","answers":[{"AnswerId":"54401917","CreationDate":"2019-01-28T12:20:16.070","ParentId":null,"OwnerUserId":"5908509","Title":null,"Body":"<p>I downloaded <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\" rel=\"nofollow noreferrer\">nightlybuild<\/a> to my hard disk and<\/p>\n\n<pre><code>pip install --upgrade C:\\tensorflow_gpu-1.0.0rc2-cp35-cp35m-win_amd64.whl\n<\/code><\/pre>\n\n<p>works for me.<\/p>\n"},{"AnswerId":"42655423","CreationDate":"2017-03-07T18:15:26.167","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">TensorFlow installation documentation<\/a> currently has the incorrect URL for the Windows PIP packages. Instead of <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_x86_64.whl<\/a> (which refers to a non-existent file), use <a href=\"https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl\" rel=\"nofollow noreferrer\">https:\/\/storage.googleapis.com\/tensorflow\/windows\/gpu\/tensorflow_gpu-1.0.0-cp35-cp35m-win_amd64.whl<\/a> (change <code>x86_64<\/code> to <code>amd64<\/code>).<\/p>\n"}]}
{"QuestionId":42655466,"AnswerCount":1,"Tags":"<machine-learning><neural-network><deep-learning><keras><keras-layer>","CreationDate":"2017-03-07T18:18:41.387","AcceptedAnswerId":42657138.0,"OwnerUserId":2649312.0,"Title":"Keras VGG16 lower level features extraction","Body":"<p>I am pulling lower level features from the VGG16 model included as Keras application. These features are exported as separate outputs of pre-trained input data for an add-on classifier. The conceptual idea was borrowed from <a href=\"https:\/\/www.google.nl\/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=0ahUKEwiMnZ2oksXSAhXMVxoKHVxHDNgQFgghMAA&amp;url=http%3A%2F%2Fwww.cv-foundation.org%2Fopenaccess%2Fcontent_iccv_2015%2Fpapers%2FYang_Multi-Scale_Recognition_With_ICCV_2015_paper.pdf&amp;usg=AFQjCNGFU9Ic1ZNFAd0sO6pE1B6dmPanVg\" rel=\"nofollow noreferrer\">Multi-scale recognition with DAG-CNNs<\/a><\/p>\n\n<p>Using the model without the classifier top, features at the highest level are extracted from block_5 pulling layer using : . This gives an output vector with dimension 8192. , however does not work on lower pulling layers as the dimensions get too large (memory issues). Instead lower pulling layers (or any other layer) can be extracted using : . The problem with this approach is however that the dimension of the feature vector reduces rapidly the lower you go: block_4 (512), block_3 (256), block_2 (128), block_1 (64).<\/p>\n\n<p>What would be a suitable layer or set-up to retain more feature data from deeper layers?<\/p>\n\n<p>For info, the output of the model looks like this, the add-on classifier has a corresponding number of inputs.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42657138","CreationDate":"2017-03-07T19:54:45.183","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>The memory error you mentioned comes from flattening a huge array which makes the number of units extremely large. What you actually need to do is to downsample your input in a smart way. I will present you some way on how to do this:<\/p>\n\n<ol>\n<li><code>MaxPooling<\/code>:  by simple usage of pooling - you could first downsample your feature maps and then <code>Flatten<\/code> them. The main advantage of this approach is its simplicity and lack of need of additional parameters. The main disadvantage : this might be a really rough method.<\/li>\n<li>Intelligent downsampling: here you could add a <code>Convolutional2D<\/code> layers with huge subsampling (e.g. with filter size <code>(4, 4)<\/code> and subsample <code>(4, 4)<\/code>). This might be consider as intelligent pooling. A main disadvantage of this method is additional parameters need for this approach.<\/li>\n<\/ol>\n"}]}
{"QuestionId":42655764,"AnswerCount":1,"Tags":"<tensorflow><neural-network><conv-neural-network>","CreationDate":"2017-03-07T18:35:43.127","AcceptedAnswerId":42656260.0,"OwnerUserId":7457101.0,"Title":"debugging 'TypeError: Can not convert a ndarray into a Tensor or Operation' for CNN","Body":"<p>I am trying to build a CNN, I have 8 classes in the input_samples with 45 samples in each class. so total number of input samples are 360. I have divided my first 20 samples as train samples and remaining 25 samples as test samples in each class (My input is a text file and the data is in rows is my preprocessed data, so I am reading the rows in textfile and reshaping the images which are 16x12 size).<\/p>\n\n<p>I am unable to fix the error in the code<\/p>\n\n<p>My code: <\/p>\n\n<pre><\/pre>\n\n<p><img src=\"https:\/\/i.stack.imgur.com\/pwkbR.png\" alt=\"Error\"><\/p>\n\n<p>file can be downloaded <a href=\"https:\/\/expirebox.com\/download\/7a5bbb776ea60dbd113ffc07829414e2.html\" rel=\"nofollow noreferrer\">dataset_300.txt<\/a> and <a href=\"https:\/\/expirebox.com\/download\/4cf16e46506b411bd13f0584a852f1d9.html\" rel=\"nofollow noreferrer\">ClassInfo.txt<\/a><\/p>\n","answers":[{"AnswerId":"42656260","CreationDate":"2017-03-07T19:05:53.870","ParentId":null,"OwnerUserId":"3534616","Title":null,"Body":"<p>Session.run accepts only a list of tensors or tensor names. <\/p>\n\n<pre><code>imgs, lbls = sess.run([image_batch_test, label_batch_test])\n<\/code><\/pre>\n\n<p>In the previous line, you are passing <code>image_batch_test<\/code> and <code>label_batch_test<\/code> which are numpy arrays. I am not sure what you are trying to do by <code>imgs, lbls = sess.run([image_batch_test, label_batch_test])<\/code><\/p>\n"}]}
{"QuestionId":42656701,"AnswerCount":0,"Tags":"<tensorflow><nlp><runtime-error>","CreationDate":"2017-03-07T19:28:05.957","AcceptedAnswerId":null,"OwnerUserId":7674379.0,"Title":"tensorflow CTC runtime error","Body":"<p>I met a runtime error during training when I tried to applied tensorflow built-in CTC loss function (<a href=\"https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn\/conectionist_temporal_classification__ctc_\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r0.10\/api_docs\/python\/nn\/conectionist_temporal_classification__ctc_<\/a>) to SynthText dataset. <a href=\"http:\/\/www.robots.ox.ac.uk\/~vgg\/data\/scenetext\/\" rel=\"nofollow noreferrer\">http:\/\/www.robots.ox.ac.uk\/~vgg\/data\/scenetext\/<\/a>\nIt said \" Not enough time for target transition sequence (required: 4, available: 0)\".\nHere is some info for the environment: tensorflow version '0.12.0-rc0'.<\/p>\n\n<p>I am able to apply tensorflow built-in CTC to Synth90K Dataset with great performance.<\/p>\n\n<p>It seems like the SynthText dataset is not compilable with tensorflow built-in CTC but Synth90K Dataset could.<\/p>\n\n<p>Please find the error message as reference<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42658464,"AnswerCount":1,"Tags":"<machine-learning><tensorflow><scikit-learn><loss>","CreationDate":"2017-03-07T21:14:31.910","AcceptedAnswerId":null,"OwnerUserId":1960319.0,"Title":"Difference between Tensorflow and Scikitlearn log_loss function implementation","Body":"<p>Hi I am trying to get into tensorflow and feeling a bit dumb.\nDoes log_loss in TF differ from sklearn's one?<\/p>\n\n<p>Here are some lines from my code, how I am calculating:<\/p>\n\n<pre><\/pre>\n\n<p>Output I get<\/p>\n\n<pre><\/pre>\n\n<p>Seems like while  converges  diverges.<\/p>\n","answers":[{"AnswerId":"44518275","CreationDate":"2017-06-13T10:02:31.283","ParentId":null,"OwnerUserId":"907575","Title":null,"Body":"<p>I had the same problem. After looking up the source code of <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/losses\/log_loss\" rel=\"nofollow noreferrer\">tf.losses.log_loss<\/a>, its <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.1\/tensorflow\/python\/ops\/losses\/losses_impl.py#L309-L312\" rel=\"nofollow noreferrer\">key lines<\/a> show wat is going on:<\/p>\n\n<pre><code>losses = - math_ops.multiply(labels, math_ops.log(predictions + epsilon))\n    - math_ops.multiply((1 - labels), math_ops.log(1 - predictions + epsilon))\n<\/code><\/pre>\n\n<p>It is <strong>binary log-loss<\/strong> (i.e. every class is considered non-exclusive) rather than <strong>multi-class log-loss<\/strong>.<\/p>\n\n<p>As I worked with probabilities (rather than logits), I couldn't use <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/softmax_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">tf.nn.softmax_cross_entropy_with_logits<\/a> (though, I could have applied logarithm).\nMy solution was to implement log-loss by hand:<\/p>\n\n<pre><code>loss = tf.reduce_sum(tf.multiply(- labels, tf.log(probs))) \/ len(probs)\n<\/code><\/pre>\n\n<p>See also:<\/p>\n\n<ul>\n<li><a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2462\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/tensorflow\/issues\/2462<\/a><\/li>\n<li><a href=\"https:\/\/stackoverflow.com\/questions\/34240703\/difference-between-tensorflow-tf-nn-softmax-and-tf-nn-softmax-cross-entropy-with\">difference between tensorflow tf.nn.softmax and tf.nn.softmax_cross_entropy_with_logits<\/a><\/li>\n<\/ul>\n"}]}
{"QuestionId":42658938,"AnswerCount":1,"Tags":"<git><github><tensorflow><bazel>","CreationDate":"2017-03-07T21:45:41.293","AcceptedAnswerId":null,"OwnerUserId":4884986.0,"Title":"Is it okay to remote the .git from tensorflow?","Body":"<p>I have a project that uses openCV and tensorflow. I cloned the tensorflow source and built a few files within the source, such as label_image and image_retrain. In the future, I plan to use the code from label_image and merge it with some openCV code, and build it from there (do I need to build it?). Anyway, when I upload my project to github and include the tensorflow folder (which I would like to change), the folder gets uploaded as a greyed out submodule because the .git file in the tensorflow folder points to the tensorflow source on github. Can I safely remove this so that I can upload the folder to my github, and can I do so without compromising future bazel builds?<\/p>\n","answers":[{"AnswerId":"42659127","CreationDate":"2017-03-07T21:58:38.637","ParentId":null,"OwnerUserId":"4717701","Title":null,"Body":"<p>As far as Bazel is concerned, it doesn't care about your sub-directory being a git directory or just vendoring so you should be safe to remove the .git directory of TensorFlow and remove tensorflow directory from the list of git submodules (see your <code>.gitmodules<\/code> file)<\/p>\n"}]}
{"QuestionId":42659154,"AnswerCount":0,"Tags":"<machine-learning><neural-network><deep-learning><caffe><lmdb>","CreationDate":"2017-03-07T22:00:20.327","AcceptedAnswerId":null,"OwnerUserId":4278243.0,"Title":"Caffe, shuffle train.txt before creating LMDB","Body":"<p>Unfortunately, I can't use  while creating LMDB. <\/p>\n\n<p>So, I was advised to shuffle train.txt before creating LMDB.<\/p>\n\n<p>After shuffling train.txt looks like this <\/p>\n\n<pre><\/pre>\n\n<p>After creating LMDBs for TEST and TRAIN I'm trying to train caffe on bvlc_reference_caffenet. <\/p>\n\n<p>Only one problem, after more that 10 thousand iterations accuracy = 0.001 and loss = 6.9. Which, as I understand, means that it doesn't train and is just guessing. <\/p>\n\n<p>Can you point out what I'm doing wrong? Thank you.<\/p>\n","answers":[]}
{"QuestionId":42659198,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-07T22:03:09.617","AcceptedAnswerId":42660790.0,"OwnerUserId":6005072.0,"Title":"How to distinguish variables in inception model?","Body":"<p>I read the code of Inception model, but I am confused about the Variable sharing.<\/p>\n\n<p>In <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/master\/inception\/inception\/slim\/inception_model.py#L116\" rel=\"nofollow noreferrer\">line 116<\/a> and line 117:<\/p>\n\n<pre><\/pre>\n\n<p>the weight Variables for those two layers have the <strong>same<\/strong>  and <strong>same<\/strong> name ? The only possible case is that the Variables are shared, but obviously, those two Variables have different shapes (1x1 vs 5x5) and are unsharable? Anyone please explain about it?<\/p>\n\n<p>Thanks,<\/p>\n","answers":[{"AnswerId":"42660790","CreationDate":"2017-03-08T00:22:16.207","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p><code>ops.conv2d()<\/code> is a TensorFlow Slim convenience function for creating a 2-D Convolution layer, including weights and bias variables. Although the function <em>supports<\/em> variable sharing, it defaults to no sharing\u2014instead creating a new variable scope on each call\u2014and you must pass the <a href=\"https:\/\/github.com\/tensorflow\/models\/blob\/3be9ece9574d7bac07704e43705741d9af1de1e6\/inception\/inception\/slim\/ops.py#L207\" rel=\"nofollow noreferrer\">optional <code>scope<\/code> and <code>reuse<\/code> arguments<\/a> to enable sharing explicitly. Therefore the two calls to <code>ops.conv2d()<\/code> create separate weights and biases variables.<\/p>\n"}]}
{"QuestionId":42659751,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-07T22:43:23.910","AcceptedAnswerId":46567292.0,"OwnerUserId":7674512.0,"Title":"How do I build an input_fn for Estimator using images stored in a TFRecords file","Body":"<p>Is there an example of how to construct the  needed by  for an image classification model?  My images are stored in multiple TFRecords files.<\/p>\n\n<p>Using , I am able to generate batches of encoded image strings.  However, I don't see an easy way to convert these strings into images.<\/p>\n","answers":[{"AnswerId":"46567292","CreationDate":"2017-10-04T14:20:45.163","ParentId":null,"OwnerUserId":"1973820","Title":null,"Body":"<p>Referring <a href=\"https:\/\/www.tensorflow.org\/programmers_guide\/datasets\" rel=\"nofollow noreferrer\">here<\/a> you can use something like below for <code>mnist<\/code> and <code>fashion-mnist<\/code> dataset stored in <code>train.tfrecords<\/code> and <code>test.tfrecords<\/code>.<\/p>\n\n<p>The conversion to <code>tfrecords<\/code> is done by code <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/how_tos\/reading_data\/convert_to_records.py\" rel=\"nofollow noreferrer\">here<\/a> and you need to have a parser to get back the original image and label. <\/p>\n\n<pre><code>def parser(serialized_example):\n  \"\"\"Parses a single tf.Example into image and label tensors.\"\"\"\n  features = tf.parse_single_example(\n      serialized_example,\n      features={\n          'image_raw': tf.FixedLenFeature([], tf.string),\n          'label': tf.FixedLenFeature([], tf.int64),\n      })\n  image = tf.decode_raw(features['image_raw'], tf.uint8)\n  image.set_shape([28 * 28])\n\n  # Normalize the values of the image from the range [0, 255] to [-0.5, 0.5]\n  image = tf.cast(image, tf.float32) \/ 255 - 0.5\n  label = tf.cast(features['label'], tf.int32)\n  return image, label\n<\/code><\/pre>\n\n<p>After having parser the rest is simple, you just need to call <code>TFRecordDataset(train_filenames)<\/code> and then map the parser function to each element so you will get an image and label as output.<\/p>\n\n<pre><code># Keep list of filenames, so you can input directory of tfrecords easily\ntraining_filenames = [\"data\/train.tfrecords\"]\ntest_filenames = [\"data\/test.tfrecords\"]\n\n# Define the input function for training\ndef train_input_fn():\n  # Import MNIST data\n  dataset = tf.contrib.data.TFRecordDataset(train_filenames)\n\n  # Map the parser over dataset, and batch results by up to batch_size\n  dataset = dataset.map(parser, num_threads=1, output_buffer_size=batch_size)\n  dataset = dataset.batch(batch_size)\n  dataset = dataset.repeat()\n  iterator = dataset.make_one_shot_iterator()\n\n  features, labels = iterator.get_next()\n\n  return features, labels\n<\/code><\/pre>\n"}]}
{"QuestionId":42660701,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-03-08T00:10:55.960","AcceptedAnswerId":null,"OwnerUserId":328764.0,"Title":"Multi-layered bidirectional_dynamic_rnn: incompatible with MultiRNNCell?","Body":"<p>I want to create a multi-layered bidirectional LSTM in Tensorflow. Currently my single-layered model looks like:<\/p>\n\n<pre><\/pre>\n\n<p>In order to turn this into a multi-layered I suspect I can <em>not<\/em> simply wrap a few s with s like so:<\/p>\n\n<pre><\/pre>\n\n<p>and feed them into the  since both forward and backward LSTMs in each layer need the output of both the forward and backward directions of the preceding layer. Currently my solution is to create my s in a loop, feeding in the concatenated output of LSTMs of the preceding layers.<\/p>\n\n<p>However, it's not very clean and frankly I'm not sure if it's correct, though it does work on a toy dataset. Is there a better way that's comparably elegant to using something like ?<\/p>\n\n<p>I'm using Tensorflow API r1.0.<\/p>\n","answers":[{"AnswerId":"49202710","CreationDate":"2018-03-09T21:52:01.533","ParentId":null,"OwnerUserId":"5504230","Title":null,"Body":"<p>Just do:<\/p>\n\n<p><code>multi_cell_fw = tf.contrib.rnn.MultiRNNCell([cell_fw for _ in range(num_layers)], ...)<\/code><\/p>\n\n<p>That should work.<\/p>\n"}]}
{"QuestionId":42660871,"AnswerCount":2,"Tags":"<python><deep-learning><theano><ipython-notebook><keras>","CreationDate":"2017-03-08T00:29:57.273","AcceptedAnswerId":43441259.0,"OwnerUserId":3768495.0,"Title":"changing keras backend to Theano in Ipython Notebook","Body":"<p>I did change the  file as instructed on the <a href=\"https:\/\/keras.io\/backend\/\" rel=\"nofollow noreferrer\">Keras documentation page<\/a>. But in my Ipython notebook, it still says I am using Tensorflow as backend. <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/1m46R.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/1m46R.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Maybe it is related to the Jupyter settings somehow? Please kindly help. I don't even know how to figure out where the problem came from. Thanks!<\/p>\n","answers":[{"AnswerId":"43441259","CreationDate":"2017-04-16T19:32:54.680","ParentId":null,"OwnerUserId":"783851","Title":null,"Body":"<p>You could try the following at the start of the notebook:<\/p>\n\n<pre><code>import os\nos.environ[\"KERAS_BACKEND\"] = \"theano\"\nimport keras; import keras.backend\nif keras.backend.backend() != 'theano':\n    raise BaseException(\"This script uses other backend\")\nelse:\n    keras.backend.set_image_dim_ordering('th')\n    print(\"Backend ok\")\n<\/code><\/pre>\n\n<p>Basically environment KERAS_BACKEND can be overwriten at some point by Jupyter so this is one way to force it to be something before you import keras.backend.<\/p>\n"},{"AnswerId":"45393656","CreationDate":"2017-07-29T20:17:27.610","ParentId":null,"OwnerUserId":"5844070","Title":null,"Body":"<h1>What works in python 2.7 - dynamically changing Keras backend<\/h1>\n\n<pre><code># When I executed the suggestion -- the output I got..\nBaseExceptionTraceback (most recent call last)\n&lt;ipython-input-7-c4352a2d60e6&gt; in &lt;module&gt;()\n      3 import keras; import keras.backend\n      4 if keras.backend.backend() != 'theano':\n----&gt; 5     raise BaseException(\"This script uses other backend\")\n      6 else:\n      7     keras.backend.set_image_dim_ordering('th')\n\nBaseException: This script uses other backend\n<\/code><\/pre>\n\n<p>-- Not sure, how this would help if we are not able to dynamically change the backend.<\/p>\n\n<p>-- Instead what helped me was the following:\n<a href=\"https:\/\/stackoverflow.com\/questions\/42177658\/how-to-switch-backend-with-keras-from-tensionflow-to-theano\">How to switch Backend with Keras (from TensionFlow to Theano)<\/a><\/p>\n\n<p><strong>Code in iPython<\/strong><\/p>\n\n<pre><code>from keras import backend; print(backend._BACKEND)\nfrom keras import backend as K\nimport os\ndef set_keras_backend(backend):\n    if K.backend() != backend:\n        os.environ['KERAS_BACKEND'] = backend\n        reload(K)\n        assert K.backend() == backend\nprint (\"Change Keras Backend to Theano\")        \nset_keras_backend(\"theano\")  \nfrom keras import backend; print(backend._BACKEND)\n<\/code><\/pre>\n\n<p><strong>Output in iPython<\/strong><\/p>\n\n<pre><code>tensorflow\nChange Keras Backend to Theano\ntheano\n<\/code><\/pre>\n"}]}
{"QuestionId":42661424,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-03-08T01:34:51.530","AcceptedAnswerId":null,"OwnerUserId":7675470.0,"Title":"Creating Estimators in tf.contrib.learn tensorflow. expected string or bytes-like object error","Body":"<p>I am trying to create my own Estimator function in tensorflow rather that using the DNNRegression one where it is not possible to change hyperparameters. I am following this tensorflow.org tutorial:<\/p>\n\n<p><a href=\"https:\/\/www.tensorflow.org\/extend\/estimators\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/extend\/estimators<\/a>.<\/p>\n\n<p>I obtained the data from the tensorflow.org page too. Here is my code with the corresponding error.<\/p>\n\n<p>Thank you very much for your help!<\/p>\n\n<pre><\/pre>\n\n<p>The error comes when i run the last line and is the following:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42661927,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-08T02:32:28.883","AcceptedAnswerId":null,"OwnerUserId":7675662.0,"Title":"How to restore two tensorflow models in a script","Body":"<p>I write two scripts which contain segmentation and classification separately. These two models are trained and I want to use them to do prediction. The input of classification is based on the output of segmentation. Such as <\/p>\n\n<pre><\/pre>\n\n<p>The problem is these two sentences can run solely but cannot run in sequence. Some problem about  will occur.<\/p>\n\n<p>How can I solve this problem?<\/p>\n","answers":[]}
{"QuestionId":42663047,"AnswerCount":2,"Tags":"<python><machine-learning><tensorflow><deep-learning><word2vec>","CreationDate":"2017-03-08T04:29:38.123","AcceptedAnswerId":42666791.0,"OwnerUserId":811921.0,"Title":"tensoflow word2vec_basic input-output placeholders","Body":"<p>I am a  newbie in tensorflow. I am trying to understand the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/tutorials\/word2vec\/word2vec_basic.py\" rel=\"nofollow noreferrer\">word2vec_basic<\/a> script.<\/p>\n\n<p>At the beggining it defined input and output.<\/p>\n\n<pre><\/pre>\n\n<p>If I understood correctly the shape of train_inputs -  is an array of integers, and the length of the arrays is . Then the shape of  is , which is the matrix  with a single row. Correct? If yes, I don't understand why it's so, and what should be stored in those placeholders. According to the theory, the  is a int and the  is an array of the sliding window of context, so why  shows up there?<\/p>\n\n<p>It seems like I missed something fundamental in the theory. I would appreciate an explanation. <\/p>\n","answers":[{"AnswerId":"42664574","CreationDate":"2017-03-08T06:38:13.303","ParentId":null,"OwnerUserId":"2878004","Title":null,"Body":"<p>For training, it is convenient to do the calculations on multiple sentences in parallel. That is the purpose of the batch_size. Each word in the batch is from a different part in the data and train_labels has the corresponding labels. The gradients are aggregated across the batch and then a single parameter update is done.<\/p>\n"},{"AnswerId":"42666791","CreationDate":"2017-03-08T08:51:47.183","ParentId":null,"OwnerUserId":"708620","Title":null,"Body":"<p>I happened to be debugging word2vec_basic.py recently and for a simple data set example \"the dog saw a cat the dog chased the cat the cat climbed a tree\", <code>dictionary<\/code> is <code>{'the': 0, 'cat': 1, 'dog': 2, 'a': 3, 'saw': 4, 'chased': 5, 'climbed': 6, 'tree': 7}<\/code>, <code>reverse_dictionary<\/code> is <code>{0: 'the', 1: 'cat', 2: 'dog', 3: 'a', 4: 'saw', 5: 'chased', 6: 'climbed', 7: 'tree'}<\/code>. <\/p>\n\n<p>When batch_size=8, skip_window=1, embedding_size=1, and num_skips=2, calling <code>generate_batch<\/code> returns:\n<code>\nbatch: [2 2 4 4 3 3 1 1]\nlabels: [[0] [4] [2] [3] [4] [1] [0] [3]]\n<\/code><\/p>\n\n<p>Translated to words, they are:\n<code>\nbatch: [dog, dog, saw, saw, a, a, cat, cat]\nlabels: [[the], [saw], [dog], [a], [saw], [cat], [the], [a]]\n<\/code><\/p>\n\n<p>For the first three words \"the dog saw\" in the data set, as skip_window is 1, the expected (target, context) pairs should be (dog, the) and (dog, saw) - see \"The Skip-gram Model\" in <a href=\"https:\/\/www.tensorflow.org\/tutorials\/word2vec\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/tutorials\/word2vec<\/a> for more detail if needed.<\/p>\n\n<p>As for why 1 is used in <code>[batch_size, 1]<\/code>, the <code>nce_loss<\/code> documentation at <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/nce_loss\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/nce_loss<\/a> says that \"labels: A Tensor of type int64 and shape [batch_size, num_true]. The target classes.\" and \"num_true: An int. The number of target classes per training example.\" (num_true has a default value of 1)<\/p>\n\n<p>And the <code>nce_loss<\/code> in the code is defined:<\/p>\n\n<pre><code>      loss = tf.reduce_mean(\n      tf.nn.nce_loss(weights=nce_weights,\n                     biases=nce_biases,\n                     labels=train_labels,\n                     inputs=embed,\n                     num_sampled=num_sampled,\n                     num_classes=vocabulary_size))\n<\/code><\/pre>\n\n<p>So now it should be easy to see the shapes of <code>train_inputs<\/code> and <code>train_labels<\/code> make sense.<\/p>\n"}]}
{"QuestionId":42664493,"AnswerCount":1,"Tags":"<python><python-2.7><machine-learning><tensorflow>","CreationDate":"2017-03-08T06:32:16.307","AcceptedAnswerId":42665677.0,"OwnerUserId":6997665.0,"Title":"Can Tensorflow be used to detect if a particular feature exists in an image?","Body":"<p>For example, using OpenCV and haarcascade_frontal_face.xml, we can predict if a face exists in an image. I would like to know if such a thing (detecting an object of interest) is possible with Tensorflow and if yes, how?<\/p>\n","answers":[{"AnswerId":"42665677","CreationDate":"2017-03-08T07:46:04.510","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Yes and no. Tensorflow is a graph computation library mostly suited for neural networks. <\/p>\n\n<p>You can create a neural network that determines if a face is in the image or not... You can even search for existing implementations that use Tensorflow...<\/p>\n\n<p>There is no default Haar feature based cascade classifier in Tensorflow...<\/p>\n"}]}
{"QuestionId":42664973,"AnswerCount":1,"Tags":"<tensorflow>","CreationDate":"2017-03-08T07:05:06.940","AcceptedAnswerId":null,"OwnerUserId":7676643.0,"Title":"tensorflow-windows installation with CPU support only through python3.5","Body":"<p>Trying to install Tensorflow in windows7 with CPU support only.\nThe installation was successful using a native pip, as instructed in tensorflow site. But i am not able to validate the installation. It is generating an error message instead of printing Hello, Tensorflow.\nerror showing- <em>tf is not defined<\/em>,.\n<em>sees is not defined<\/em>.\nthe screenshot of the error: [1]: <a href=\"https:\/\/i.stack.imgur.com\/hFLvr.png\" rel=\"nofollow noreferrer\">https:\/\/i.stack.imgur.com\/hFLvr.png<\/a><\/p>\n","answers":[{"AnswerId":"50902042","CreationDate":"2018-06-18T02:31:53.330","ParentId":null,"OwnerUserId":"5033247","Title":null,"Body":"<p>Working<\/p>\n\n<pre><code>py -m 3.6 pip install tensorflow\n<\/code><\/pre>\n\n<p>The lack of AVX instruction in CPU can cause this error.<\/p>\n\n<p>Try this <a href=\"https:\/\/github.com\/fo40225\/tensorflow-windows-wheel\/tree\/master\/1.6.0\/py36\/CPU\/sse2\" rel=\"nofollow noreferrer\">wheel<\/a> for <a href=\"https:\/\/www.python.org\/ftp\/python\/3.6.7\/python-3.6.7-amd64.exe\" rel=\"nofollow noreferrer\">python3.6<\/a><\/p>\n"}]}
{"QuestionId":42665359,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow><deep-learning><keras>","CreationDate":"2017-03-08T07:28:24.500","AcceptedAnswerId":42665603.0,"OwnerUserId":3993741.0,"Title":"How do you compute accuracy in a regression model, after rounding predictions to classes, in keras?","Body":"<p>How would you create and display an accuracy metric in keras for a regression problem, for example after you round the predictions to the nearest integer class? <\/p>\n\n<p>While accuracy is not itself effectively defined conventionally for a regression problem, to determine <strong>ordinal<\/strong> classes\/labels for data, it is suitable to treat the problem as a regression.  But then it would be convenient to also calculate an accuracy metric, whether it be kappa or something else like that.  Here is a basic keras boilerplate code to modify.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42665603","CreationDate":"2017-03-08T07:41:59.600","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>I use rounded accuracy like this:<\/p>\n\n<pre><code>def soft_acc(y_true, y_pred):\n    return K.mean(K.equal(K.round(y_true), K.round(y_pred)))\n\nmodel.compile(..., metrics=[soft_acc])\n<\/code><\/pre>\n"}]}
{"QuestionId":42666046,"AnswerCount":5,"Tags":"<python><tensorflow><neural-network><keras><resuming-training>","CreationDate":"2017-03-08T08:07:57.997","AcceptedAnswerId":42670479.0,"OwnerUserId":5359882.0,"Title":"Loading a trained Keras model and continue training","Body":"<p>I was wondering if it was possible to save a partly trained Keras model and continue the training after loading the model again.<\/p>\n\n<p>The reason for this is that I will have more training data in the future and I do not want to retrain the whole model again.<\/p>\n\n<p>The functions which I am using are:<\/p>\n\n<pre><\/pre>\n\n<hr>\n\n<p><strong>Edit 1: added fully working example<\/strong><\/p>\n\n<p>With the first dataset after 10 epochs the loss of the last epoch will be 0.0748 and the accuracy 0.9863.<\/p>\n\n<p>After saving, deleting and reloading the model the loss and accuracy of the model trained on the second dataset will be 0.1711 and 0.9504 respectively.<\/p>\n\n<p>Is this caused by the new training data or by a completely re-trained model?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"48010463","CreationDate":"2017-12-28T15:53:37.080","ParentId":null,"OwnerUserId":"7360054","Title":null,"Body":"<p>The problem might be that you use a different optimizer - or different arguments to your optimizer. I just had the same issue with a custom pretrained model, using <\/p>\n\n<pre><code>reduce_lr = ReduceLROnPlateau(monitor='loss', factor=lr_reduction_factor,\n                              patience=patience, min_lr=min_lr, verbose=1)\n<\/code><\/pre>\n\n<p>for the pretrained model, whereby the original learning rate starts at 0.0003 and during pre-training it is reduced to the min_learning rate, which is 0.000003<\/p>\n\n<p>I just copied that line over to the script which uses the pre-trained model and got really bad accuracies. Until I noticed that the last learning rate of the pretrained model was the min learning rate, i.e. 0.000003. And if I start with that learning rate, I get exactly the same accuracies to start with as the output of the pretrained model - which makes sense, as starting with a learning rate that is 100 times bigger than the last learning rate used in the pretrained model will result in a huge overshoot of GD and hence in heavily decreased accuracies. <\/p>\n"},{"AnswerId":"45321733","CreationDate":"2017-07-26T08:42:45.220","ParentId":null,"OwnerUserId":"662770","Title":null,"Body":"<p>Notice that Keras sometimes has issues with loaded models, as in <a href=\"https:\/\/github.com\/fchollet\/keras\/issues\/2378\" rel=\"nofollow noreferrer\">here<\/a>.\nThis might explain cases in which you don't start from the same trained accuracy.<\/p>\n"},{"AnswerId":"50591136","CreationDate":"2018-05-29T18:55:32.680","ParentId":null,"OwnerUserId":"9101382","Title":null,"Body":"<p>All above helps, you <strong>must<\/strong> resume from same learning rate() as the LR when the model and weights were saved. Set it directly on the optimizer. <\/p>\n\n<p>Note that improvement from there is not guaranteed, because the model may have reached the local minimum, which may be global. There is no point to resume a model in order to search for another local minimum, unless you intent to increase the learning rate in a controlled fashion and nudge the model into a possibly better minimum not far away.<\/p>\n"},{"AnswerId":"42670479","CreationDate":"2017-03-08T11:45:28.313","ParentId":null,"OwnerUserId":"5974433","Title":null,"Body":"<p>Actually - <code>model.save<\/code> saves all information need for restarting training in your case. The only thing which could be spoiled by reloading model is your optimizer state. To check that - try to <code>save<\/code> and reload model and train it on training data.<\/p>\n"},{"AnswerId":"53697941","CreationDate":"2018-12-10T00:03:06.720","ParentId":null,"OwnerUserId":"504317","Title":null,"Body":"<p>You might also be hitting Concept Drift, see <a href=\"https:\/\/datascience.stackexchange.com\/questions\/12761\/should-a-model-be-re-trained-if-new-observations-are-available\">Should you retrain a model when new observations are available<\/a>. There's also the concept of catastrophic forgetting which a bunch of academic papers discuss.  Here's one with MNIST <a href=\"https:\/\/arxiv.org\/pdf\/1312.6211.pdf\" rel=\"nofollow noreferrer\">Empirical investigation of catastrophic forgetting<\/a><\/p>\n"}]}
{"QuestionId":42666494,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-03-08T08:34:38.520","AcceptedAnswerId":42684090.0,"OwnerUserId":3663205.0,"Title":"Tensorflow C++ fast Tensor deep copy","Body":"<p>I am writing a custom tensorflow op in C++ and I want to know how I might perform a deep copy of one tensor into another.  In other words, I want an element-wise copy from one tensor to another such that they are not made to share an underlying memory buffer.  <\/p>\n\n<p>The closest thing I've found is DeepCopy (defined in <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/53e3c4f7070c47c7f0bb72732abb0f3b2dff1977\/tensorflow\/core\/framework\/tensor_util.h#L37\" rel=\"nofollow noreferrer\">tensor_util.h<\/a>).  The problem is I need this operation to be fast and the documentation states clearly that this function has not been optimized for speed:<\/p>\n\n<pre><\/pre>\n\n<p>A bit more detail: I am using Tensorflow on the GPU, so basically what I want to do is to have TF initiate a cudaMemcpy where both source and target are device pointers.  TF does give access to tensor pointers (tensor.tensor_data().data()), but you run into problems pretty quick if you try to cudaMemcpy with them (they have to do with overlapping cuda contexts; not nice).<\/p>\n\n<p>Thanks in advance!<\/p>\n","answers":[{"AnswerId":"42684090","CreationDate":"2017-03-08T23:43:51.853","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>Currently for \"fast\" deep copies, TensorFlow uses Eigen to implement CPU and GPU versions. Being Eigen, the code is deeply templated, but the relevant code is here (copied from <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/2d23960e29c3efb1b30c5cb6f8f3368635b97f29\/tensorflow\/core\/kernels\/dense_update_ops.h#L48\" rel=\"nofollow noreferrer\"><code>dense_update_ops.h<\/code><\/a>):<\/p>\n\n<pre><code>template &lt;typename Device, typename T&gt;\nstruct DenseUpdate&lt;Device, T, ASSIGN&gt; {\n  void operator()(const Device&amp; d, typename TTypes&lt;T&gt;::Flat params,\n                  typename TTypes&lt;T&gt;::ConstFlat update) {\n    params.device(d) = update;\n  }\n};\n<\/code><\/pre>\n"}]}
{"QuestionId":42666627,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-08T08:42:16.753","AcceptedAnswerId":null,"OwnerUserId":4339490.0,"Title":"Inception: How to create a output function from image input","Body":"<p>How do you create a function for Inception v3 that:<\/p>\n\n<ol>\n<li><p>Takes an image as input.<\/p><\/li>\n<li><p>Print out logits of labels as output.<\/p><\/li>\n<\/ol>\n\n<p>The original code for inception v3 is here:\n<a href=\"https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/models\/tree\/master\/inception<\/a><\/p>\n\n<p>An example code where they manage to calculate output from a graph is here. I want the model to use checkpoint instead of graph. However, I don't know how to do the same thing like the example below, but with checkpoint.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42670391","CreationDate":"2017-03-08T11:41:08.657","ParentId":null,"OwnerUserId":"5957580","Title":null,"Body":"<p>Just run it like this <code>python classify_image.py --image_file \/path\/to\/file<\/code>\n<em>This will take the image as input and will output the labels.<\/em><\/p>\n\n<p>You might also want to try and add the line below. It will identify and analyse the last added .jpg file to the specified folder.<\/p>\n\n<p><code>newest = max(glob.iglob('\/home\/l2grp\/Jetty\/src\/ubiserv\/simple\/img\/*.[Jj][Pp][Gg]'), key=os.path.getctime)<\/code> <\/p>\n\n<p><div class=\"snippet\" data-lang=\"js\" data-hide=\"false\" data-console=\"true\" data-babel=\"false\">\r\n<div class=\"snippet-code\">\r\n<pre class=\"snippet-code-html lang-html prettyprint-override\"><code>Please see the tutorial and website for a detailed description of how\r\nto use this script to perform image recognition.\r\n\r\nhttps:\/\/tensorflow.org\/tutorials\/image_recognition\/\r\n\"\"\"\r\n\r\nfrom __future__ import absolute_import\r\nfrom __future__ import division\r\nfrom __future__ import print_function\r\n\r\nimport argparse\r\nimport os.path\r\nimport re\r\nimport sys\r\nimport tarfile\r\nimport glob\r\n\r\nimport numpy as np\r\nfrom six.moves import urllib\r\nimport tensorflow as tf\r\n\r\nFLAGS = None\r\n\r\n# pylint: disable=line-too-long\r\nDATA_URL = 'http:\/\/download.tensorflow.org\/models\/image\/imagenet\/inception-2015-12-05.tgz'\r\n# pylint: enable=line-too-long\r\n\r\nnewest = max(glob.iglob('\/home\/l2grp\/Jetty\/src\/ubiserv\/simple\/img\/*.[Jj][Pp][Gg]'), key=os.path.getctime)\r\n\r\nclass NodeLookup(object):\r\n  \"\"\"Converts integer node ID's to human readable labels.\"\"\"\r\n\r\n  def __init__(self,\r\n               label_lookup_path=None,\r\n               uid_lookup_path=None):\r\n    if not label_lookup_path:\r\n      label_lookup_path = os.path.join(\r\n          FLAGS.model_dir, 'imagenet_2012_challenge_label_map_proto.pbtxt')\r\n    if not uid_lookup_path:\r\n      uid_lookup_path = os.path.join(\r\n          FLAGS.model_dir, 'imagenet_synset_to_human_label_map.txt')\r\n    self.node_lookup = self.load(label_lookup_path, uid_lookup_path)\r\n\r\n  def load(self, label_lookup_path, uid_lookup_path):\r\n    \"\"\"Loads a human readable English name for each softmax node.\r\n\r\n    Args:\r\n      label_lookup_path: string UID to integer node ID.\r\n      uid_lookup_path: string UID to human-readable string.\r\n\r\n    Returns:\r\n      dict from integer node ID to human-readable string.\r\n    \"\"\"\r\n    if not tf.gfile.Exists(uid_lookup_path):\r\n      tf.logging.fatal('File does not exist %s', uid_lookup_path)\r\n    if not tf.gfile.Exists(label_lookup_path):\r\n      tf.logging.fatal('File does not exist %s', label_lookup_path)\r\n\r\n    # Loads mapping from string UID to human-readable string\r\n    proto_as_ascii_lines = tf.gfile.GFile(uid_lookup_path).readlines()\r\n    uid_to_human = {}\r\n    p = re.compile(r'[n\\d]*[ \\S,]*')\r\n    for line in proto_as_ascii_lines:\r\n      parsed_items = p.findall(line)\r\n      uid = parsed_items[0]\r\n      human_string = parsed_items[2]\r\n      uid_to_human[uid] = human_string\r\n\r\n    # Loads mapping from string UID to integer node ID.\r\n    node_id_to_uid = {}\r\n    proto_as_ascii = tf.gfile.GFile(label_lookup_path).readlines()\r\n    for line in proto_as_ascii:\r\n      if line.startswith('  target_class:'):\r\n        target_class = int(line.split(': ')[1])\r\n      if line.startswith('  target_class_string:'):\r\n        target_class_string = line.split(': ')[1]\r\n        node_id_to_uid[target_class] = target_class_string[1:-2]\r\n\r\n    # Loads the final mapping of integer node ID to human-readable string\r\n    node_id_to_name = {}\r\n    for key, val in node_id_to_uid.items():\r\n      if val not in uid_to_human:\r\n        tf.logging.fatal('Failed to locate: %s', val)\r\n      name = uid_to_human[val]\r\n      node_id_to_name[key] = name\r\n\r\n    return node_id_to_name\r\n\r\n  def id_to_string(self, node_id):\r\n    if node_id not in self.node_lookup:\r\n      return ''\r\n    return self.node_lookup[node_id]\r\n\r\n\r\ndef create_graph():\r\n  \"\"\"Creates a graph from saved GraphDef file and returns a saver.\"\"\"\r\n  # Creates graph from saved graph_def.pb.\r\n  with tf.gfile.FastGFile(os.path.join(\r\n      FLAGS.model_dir, 'classify_image_graph_def.pb'), 'rb') as f:\r\n    graph_def = tf.GraphDef()\r\n    graph_def.ParseFromString(f.read())\r\n    _ = tf.import_graph_def(graph_def, name='')\r\n\r\n\r\ndef run_inference_on_image(image):\r\n  \"\"\"Runs inference on an image.\r\n\r\n  Args:\r\n    image: Image file name.\r\n\r\n  Returns:\r\n    Nothing\r\n  \"\"\"\r\n  if not tf.gfile.Exists(image):\r\n    tf.logging.fatal('File does not exist %s', image)\r\n  image_data = tf.gfile.FastGFile(image, 'rb').read()\r\n\r\n  # Creates graph from saved GraphDef.\r\n  create_graph()\r\n\r\n  with tf.Session() as sess:\r\n    # Some useful tensors:\r\n    # 'softmax:0': A tensor containing the normalized prediction across\r\n    #   1000 labels.\r\n    # 'pool_3:0': A tensor containing the next-to-last layer containing 2048\r\n    #   float description of the image.\r\n    # 'DecodeJpeg\/contents:0': A tensor containing a string providing JPEG\r\n    #   encoding of the image.\r\n    # Runs the softmax tensor by feeding the image_data as input to the graph.\r\n    softmax_tensor = sess.graph.get_tensor_by_name('softmax:0')\r\n    predictions = sess.run(softmax_tensor,\r\n                           {'DecodeJpeg\/contents:0': image_data})\r\n    predictions = np.squeeze(predictions)\r\n\r\n    # Creates node ID --&gt; English string lookup.\r\n    node_lookup = NodeLookup()\r\n\r\n    top_k = predictions.argsort()[-FLAGS.num_top_predictions:][::-1]\r\n    for node_id in top_k:\r\n      human_string = node_lookup.id_to_string(node_id)\r\n      score = predictions[node_id]\r\n      print('%s (score = %.5f)' % (human_string, score))\r\n\r\n\r\ndef main(_):\r\n  image = newest\r\n  run_inference_on_image(image)\r\n\r\n\r\nif __name__ == '__main__':\r\n  parser = argparse.ArgumentParser()\r\n  # classify_image_graph_def.pb:\r\n  #   Binary representation of the GraphDef protocol buffer.\r\n  # imagenet_synset_to_human_label_map.txt:\r\n  #   Map from synset ID to a human readable string.\r\n  # imagenet_2012_challenge_label_map_proto.pbtxt:\r\n  #   Text representation of a protocol buffer mapping a label to synset ID.\r\n  parser.add_argument(\r\n      '--model_dir',\r\n      type=str,\r\n      default='\/tmp\/imagenet',\r\n      help=\"\"\"\\\r\n      Path to classify_image_graph_def.pb,\r\n      imagenet_synset_to_human_label_map.txt, and\r\n      imagenet_2012_challenge_label_map_proto.pbtxt.\\\r\n      \"\"\"\r\n  )\r\n  parser.add_argument(\r\n      '--image_file',\r\n      type=str,\r\n      default='',\r\n      help='Absolute path to image file.'\r\n  )\r\n  parser.add_argument(\r\n      '--num_top_predictions',\r\n      type=int,\r\n      default=5,\r\n      help='Display this many predictions.'\r\n  )\r\n  FLAGS, unparsed = parser.parse_known_args()\r\n  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)<\/code><\/pre>\r\n<\/div>\r\n<\/div>\r\n<\/p>\n"}]}
{"QuestionId":42667102,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-03-08T09:08:11.930","AcceptedAnswerId":42679499.0,"OwnerUserId":4220419.0,"Title":"loss function with distinct weight for every input","Body":"<p>I want to design my custom loss function, which not only contains two input (y_real and y_pred), but I do not know if I can do it and how. <\/p>\n\n<p>I have went through the answers about loss function, all of them only input  y_real and y_pred. <\/p>\n\n<p>But in my projects, the loss of true and false prediction are various in cases, such that X1 with (loss_T1,loss_F1), X2 with (loss_T2, loss_F2). <\/p>\n\n<p>So for my loss function, I need to input four variables (y_real,y_pred,loss_T,loss_F). The loss_T,loss_F are inside X (each row of my data) as one of features. <\/p>\n\n<p>feature1    feature2    feature3    Loss_T  Loss_F  y\nX1<br>\nX2<br>\nX3                      <\/p>\n\n<p>Is there way I can do that? <\/p>\n\n<p>Thanks to your good contributions.<\/p>\n","answers":[{"AnswerId":"42679499","CreationDate":"2017-03-08T18:47:06.543","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>The tensorflow API at a lower level can work for you. For example, <code>Optimizer.optimize_loss<\/code> takes a loss tensor, which is just the output of the loss function, and minimizes it. So if you use <code>optimize_loss<\/code> diretly you can build your loss function in any complicated way you would like.<\/p>\n"}]}
{"QuestionId":42667258,"AnswerCount":2,"Tags":"<python><tensorflow><google-cloud-ml>","CreationDate":"2017-03-08T09:17:03.690","AcceptedAnswerId":null,"OwnerUserId":6659511.0,"Title":"Proper module packaging for CloudML","Body":"<p>I have a Python project structured as follows:<\/p>\n\n<ul>\n<li>project\/ Contains:<\/li>\n<li>setup.py ; project\/\nThe inner project\/ folder contains a <strong>init<\/strong>.py empty file, folders with submodules, a conf\/ folder containing only .yaml files (e.g. foo.yaml) and a s_main.py script (my main entry point)<\/li>\n<\/ul>\n\n<p>This is my setup.py file:<\/p>\n\n<pre><\/pre>\n\n<p>When I try 'pip install project\/' it works fine on my PC and I see the conf\/ folder with all my *.yaml files in [...]\/python2.7\/site-packages\/project\/conf\/\nWhen I run cloudML locally, it also works. I use the following command:<\/p>\n\n<pre><\/pre>\n\n<p>But when I run it on cloudML, it doesn't work. I run the following command :<\/p>\n\n<pre><\/pre>\n\n<p>The log clearly shows that the conf\/ folder was never included in the built package and my run fails because the expected *.yaml files were not found. \nAny idea?<\/p>\n","answers":[{"AnswerId":"42781172","CreationDate":"2017-03-14T08:36:13.087","ParentId":null,"OwnerUserId":"40999","Title":null,"Body":"<p>Can you try manually packaging using:<\/p>\n\n<pre><code>python setup.py sdist\n<\/code><\/pre>\n\n<p>And then staging the resulting .tar.gz in a Cloud Storage bucket.<\/p>\n\n<p>Once you do that, you should be able to refer to your pre-staged package using --package-uris flag on the gcloud command (when submitting a job to the service).<\/p>\n\n<p>See if that helps unblock you.<\/p>\n"},{"AnswerId":"42837423","CreationDate":"2017-03-16T14:48:51.537","ParentId":null,"OwnerUserId":"40999","Title":null,"Body":"<p>The other thing to try is not using <code>package_data<\/code> and <code>include_package_data<\/code>, but rather using <code>data_files<\/code> ... at least that is what I have used in my own scenarios where I needed to have additional data files in a source distribution.<\/p>\n\n<p>Something like:<\/p>\n\n<pre><code>data_files=[('.', 'conf\/*.yaml','*.yaml']),\n            ('project':['conf\/*.yaml'])\n           ]\n<\/code><\/pre>\n\n<p>Maybe that will work?<\/p>\n"}]}
{"QuestionId":42667454,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-03-08T09:26:35.047","AcceptedAnswerId":null,"OwnerUserId":7677176.0,"Title":"float32 placeholder in tensorflow keep on returning 0","Body":"<p>while trying to run:<\/p>\n\n<pre><\/pre>\n\n<p>i keep on getting the following results:\n0.0\n[ 0.  0.].\nIs anyone else faced those same issues?<\/p>\n\n<p>while changing the placeholders to int32 the results started to make sense...<\/p>\n\n<p>I'm using: \nTensorflow  1.0.1\npython 3.5.2+ and also tried that on python 2.7.12+ and got the same results.<\/p>\n\n<p>this is the full output<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42668394,"AnswerCount":1,"Tags":"<python><matrix><tensorflow><matrix-multiplication>","CreationDate":"2017-03-08T10:10:22.400","AcceptedAnswerId":null,"OwnerUserId":1817595.0,"Title":"Reduce GPU memory usage on matrix multiplication in Tensorflow","Body":"<p>I have 2 matrices of shapes<\/p>\n\n<pre><\/pre>\n\n<p>The output is obtained by <\/p>\n\n<pre><\/pre>\n\n<p>and has shape <\/p>\n\n<pre><\/pre>\n\n<p>This operation alone takes around 5GB of GPU memory. I would like to know if there is any way to reduce this memory consumption (since I will be having much larger matrices for the same operation in the future). <\/p>\n\n<p>I tried a workaround by converting both  and  to half precision. This also did not help in reducing the memory usage. <\/p>\n\n<p>It would be great to know if there are any tricks in Tensorflow that I can utilize do to do this. <\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"42668597","CreationDate":"2017-03-08T10:20:18.680","ParentId":null,"OwnerUserId":"7549930","Title":null,"Body":"<p>You can set <code>tf.GPUOption<\/code> to control GPU memory usage. By default, tf occupies lots of GPU memory even if you don't build any computation graph.<\/p>\n\n<p>See <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GPUOptions\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/GPUOptions<\/a> for more. Typically, you may want to use <code>allow_growth<\/code> or <code>per_process_gpu_memory_fraction<\/code> property.<\/p>\n"}]}
{"QuestionId":42669189,"AnswerCount":0,"Tags":"<deep-learning><keras><lstm>","CreationDate":"2017-03-08T10:46:22.050","AcceptedAnswerId":null,"OwnerUserId":1947542.0,"Title":"How to shape dataset for training in Keras?","Body":"<p>I am trying to train an LSTM in Keras on microgrid dataset from this <a href=\"http:\/\/traces.cs.umass.edu\/index.php\/Smart\/Smart\" rel=\"nofollow noreferrer\">site<\/a>. There are 443 files each for one home containing electric consumption data for each minute for 24 hours (1440 rows).<\/p>\n\n<p>At first, I choose only one file and created a dataset by taking n samples at each step and making  value the label. <\/p>\n\n<p>Now, I want to combine all data-files and train. <\/p>\n\n<p>How can I create train and test datasets? <\/p>\n","answers":[]}
{"QuestionId":42669578,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-08T11:04:22.413","AcceptedAnswerId":43500797.0,"OwnerUserId":7677781.0,"Title":"ValueError: Attempt to reuse RNNCell with a different variable scope than its first use","Body":"<p>The following code fragment<\/p>\n\n<pre><\/pre>\n\n<p>produces the following error:<\/p>\n\n<blockquote>\n  <p>ValueError: Attempt to reuse  \n  &lt; object at 0x10210d5c0> with a different variable scope than its first use. First use of cell was with scope , this attempt is with scope `'rnn\/multi_rnn_cell\/cell_1\/basic_lstm_cell'``. <\/p>\n  \n  <p>Please create a new instance of the cell if you would like it to use a different set of weights. <\/p>\n  \n  <p>If before you were using: , change to: . <\/p>\n  \n  <p>If before you were using the same cell instance as both the forward and reverse cell of a bidirectional RNN, simply create two instances (one for forward, one for reverse). <\/p>\n  \n  <p>In May 2017, we will start transitioning this cell's behavior to use existing stored weights, if any, when it is called with  (which can lead to silent model degradation, so this error will remain until then.)<\/p>\n<\/blockquote>\n\n<p>How to solve this problem? <\/p>\n\n<p>My version of Tensorflow is 1.0.<\/p>\n","answers":[{"AnswerId":"43500797","CreationDate":"2017-04-19T16:04:51.423","ParentId":null,"OwnerUserId":"6591278","Title":null,"Body":"<p>As suggested in the comments my solution is:<br>\nchanging this  <\/p>\n\n<pre><code>cell = tf.contrib.rnn.LSTMCell(state_size, state_is_tuple=True)\ncell = tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=0.8)\nrnn_cells = tf.contrib.rnn.MultiRNNCell([cell for _ in range(num_layers)], state_is_tuple = True)\noutputs, current_state = tf.nn.dynamic_rnn(rnn_cells, x, initial_state=rnn_tuple_state, scope = \"layer\")\n<\/code><\/pre>\n\n<p>into:<\/p>\n\n<pre><code>def lstm_cell():\n    cell = tf.contrib.rnn.LSTMCell(state_size, reuse=tf.get_variable_scope().reuse)\n    return tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=0.8)\n\nrnn_cells = tf.contrib.rnn.MultiRNNCell([lstm_cell() for _ in range(num_layers)], state_is_tuple = True)\noutputs, current_state = tf.nn.dynamic_rnn(rnn_cells, x, initial_state=rnn_tuple_state)\n<\/code><\/pre>\n\n<p>which seems to solve the reusability problem. I don't fundamentally understand the underlying problem, but this solved the issue for me on TF 1.1rc2<br>\ncheers!<\/p>\n"}]}
{"QuestionId":42669624,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-03-08T11:06:31.433","AcceptedAnswerId":null,"OwnerUserId":7677795.0,"Title":"How to shuffle large scale tfrecord data in Tensorflow adequately?","Body":"<p>Assuming that we get 200 thousand pictures labeled '1' and another 200 thousand pictures labeled '0' converted to tfrecord of tensorflow sequentially. And I tried to shuffle the tfrecord with . The question is I can't shuffle the data adequately without an enormous  while it'll be out of memory with a too big .Is there any solutions to that?<\/p>\n\n<p>Thanks!<\/p>\n","answers":[{"AnswerId":"46981412","CreationDate":"2017-10-27T18:34:11.820","ParentId":null,"OwnerUserId":"1497311","Title":null,"Body":"<p>I would recommend using <code>tf.data.Dataset<\/code> (<a href=\"https:\/\/www.tensorflow.org\/versions\/r1.4\/api_docs\/python\/tf\/data\/Dataset\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/r1.4\/api_docs\/python\/tf\/data\/Dataset<\/a>) for this purpose, leveraging the <code>shuffle<\/code> transformation and also the <code>interleave<\/code> transformation (assuming you wish to mix the '1' and '0' labels).<\/p>\n"}]}
{"QuestionId":42669988,"AnswerCount":1,"Tags":"<python-3.x><tensorflow>","CreationDate":"2017-03-08T11:23:35.020","AcceptedAnswerId":42671587.0,"OwnerUserId":4993842.0,"Title":"How to add more channels to input tensors in tensorflow?","Body":"<p>Currently, I have an input tensor with only one channel. I'm trying to figure out how to add additional channels.<\/p>\n\n<p>For example, lets say the current input 8x8 channel one is -<\/p>\n\n<pre><\/pre>\n\n<p>And I want to add two more channels (filled with zeroes) of 8x8 -<\/p>\n\n<pre><\/pre>\n\n<p>My current code (accepts lists of size 64 and reshapes into 8x8 of 1 channel)-<\/p>\n\n<pre><\/pre>\n\n<p>To add 2 more channels, I've changed the code to this -<\/p>\n\n<pre><\/pre>\n\n<p> looks like - <\/p>\n\n<p>My question is - Is this the right way to add new channels? <\/p>\n\n<p>The reason I'm asking this is that I'm not clear on how  would do this reshaping. Would it know to take every 3rd item (as opposed to 3 consecutive items) to create the 3 channels? <\/p>\n","answers":[{"AnswerId":"42671587","CreationDate":"2017-03-08T12:37:06.890","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>Your question was:<\/p>\n\n<blockquote>\n  <p>Is this the right way to add new channels?<\/p>\n<\/blockquote>\n\n<p>The short answer to this is: yes, great job!<\/p>\n\n<p>The best way to understand what this function does is looking at the documentation: <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/reshape<\/a>\nIf you have a color image and a RGBRGBRGBRGB array (so channel 1, 2, 3, 1, 2, 3 etc. ) this is the way to tell tensorflow to create three channels out of it!<\/p>\n\n<p>As you state: <\/p>\n\n<blockquote>\n  <p>inputchannel3 looks like - [1, 0, 0, 2, 0, 0, ...64, 0, 0]<\/p>\n<\/blockquote>\n\n<p>I think you did everything correct :D<\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":42672251,"AnswerCount":1,"Tags":"<neural-network><computer-vision><deep-learning><caffe><pycaffe>","CreationDate":"2017-03-08T13:09:04.627","AcceptedAnswerId":null,"OwnerUserId":7596581.0,"Title":"windows SSD caffe","Body":"<p>everyone,when I test <a href=\"https:\/\/github.com\/weiliu89\/caffe\/blob\/ssd\/examples\/ssd\/ssd_pascal_video.py\" rel=\"nofollow noreferrer\"><\/a> in windows,I got the following error:<\/p>\n\n<pre><\/pre>\n\n<p>then,I check the ssd_pascal_video.py,I find the following code,I can't understand the meaning of \"{}\"in the address.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42672338","CreationDate":"2017-03-08T13:13:38.807","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p>The <code>'{}'<\/code> in the strings in part of the <a href=\"https:\/\/docs.python.org\/2\/library\/string.html#formatstrings\" rel=\"nofollow noreferrer\"><code>.format()<\/code><\/a> operator in python.<br>\nBasically, each <code>{}<\/code> in string is replaced by the corresponding value in the <code>.format()<\/code>.<\/p>\n"}]}
{"QuestionId":42672277,"AnswerCount":2,"Tags":"<python><image-processing><tensorflow><neural-network><deep-learning>","CreationDate":"2017-03-08T13:10:11.073","AcceptedAnswerId":null,"OwnerUserId":6193863.0,"Title":"How to prevent redundancy in Image Data when training a network?","Body":"<p>I have a large quantity of images of different and random sizes. <\/p>\n\n<p>I would like to make sure that \n- my training set is not redundant\n- the data in my test set are all different (from the others in the test set and from the training set)<\/p>\n\n<p>By not redundant, I mean basically to check if an image is not a cropped version of another (or the same image at the biggest extend).<\/p>\n\n<p>How do you deal with such cases in general when working on a net? \nAnything maybe ready for this plugged-in in Tensorflow somehow? <\/p>\n\n<p>Thank you for the help<\/p>\n\n<p>K. <\/p>\n","answers":[{"AnswerId":"42692914","CreationDate":"2017-03-09T10:31:41.237","ParentId":null,"OwnerUserId":"6193863","Title":null,"Body":"<p>Seems like <a href=\"http:\/\/geeqie.sourceforge.net\/\" rel=\"nofollow noreferrer\">http:\/\/geeqie.sourceforge.net\/<\/a> is doing the trick<\/p>\n\n<p>Thanks  @ThomasPinetz<\/p>\n\n<p>K.<\/p>\n"},{"AnswerId":"42716795","CreationDate":"2017-03-10T11:02:19.947","ParentId":null,"OwnerUserId":"5288679","Title":null,"Body":"<p><strong>What you've to do:<\/strong><\/p>\n\n<p><em>Add first image to a set (A) and then parse remaining images one by one and add only those images to (A) which DIFFERS from all images in (A). Keep iterating through to last image. Now use set (A) images for training and testing say 80% of (A) for training rest for testing<\/em><\/p>\n\n<p><strong>How To Determine: Redundancy ?<\/strong><\/p>\n\n<p>In simple words <em>two scaled square images<\/em> are same or redundant as both are having <em>four right angled corners (or features)<\/em>. If one square image is cropped then also it will have minimum 2 matching corners (<em>or matching features<\/em>).<\/p>\n\n<p>Thus in order to determine <strong>redundant images<\/strong> you need to <em>find local features<\/em> of images then <em>calculate how many features are matching<\/em>. <\/p>\n\n<p>This can be achieved via <strong>SIFT (scale invariant feature transform)<\/strong> and other techniques like <strong>SURF local feature detector<\/strong>. Use <a href=\"https:\/\/in.mathworks.com\/help\/vision\/ref\/matchfeatures.html\" rel=\"nofollow noreferrer\">this<\/a> article to get basic understanding.<\/p>\n"}]}
{"QuestionId":42672431,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning><recurrent-neural-network>","CreationDate":"2017-03-08T13:18:09.267","AcceptedAnswerId":44246779.0,"OwnerUserId":3697202.0,"Title":"TensorFlow throws error only when using MultiRNNCell","Body":"<p>I'm building an encoder-decoder model in TensorFlow 1.0.1 using the legacy sequence-to-sequence framework. Everything works as it should when I have one layer of LSTMs in the encoder and decoder. However, when I try with >1 layers of LSTMs wrapped in a , I get an error when calling .<\/p>\n\n<p>The full error is at the end up this post, but in brief, it's caused by a line<\/p>\n\n<pre><\/pre>\n\n<p>in TensorFlow that throws . I'm confused by this, since the initial state I'm passing to  is indeed a tuple as it should be. As far as I can tell, the only difference between using 1 or >1 layers is that the latter involves using . Are there some API quirks that I should know about when using this?<\/p>\n\n<p>This is my code (based on the example in <a href=\"https:\/\/github.com\/raindeer\/seq2seq_experiments\" rel=\"nofollow noreferrer\">this<\/a> GitHub repo). Apologies for its length; this is as minimal I could make it while still being complete and verifiable.<\/p>\n\n<pre><\/pre>\n\n<p>And this is the error:<\/p>\n\n<pre><\/pre>\n\n<p>Thank you!<\/p>\n","answers":[{"AnswerId":"44246779","CreationDate":"2017-05-29T16:17:49.113","ParentId":null,"OwnerUserId":"8082300","Title":null,"Body":"<p>The problem is in the format of the initial state (<code>initial_dec_state<\/code>) passed to <code>seq2seq.rnn_decoder<\/code>.<\/p>\n\n<p>When you use <code>rnn.MultiRNNCell<\/code>, you're building a multilayer recurrent network, so you need to provide an initial state for <strong>each<\/strong> of these layers.<\/p>\n\n<p>Hence, you should provide a <strong>list<\/strong> of tuples as the initial state, where each element of the list is the previous state coming from the corresponding layer of the recurrent network.<\/p>\n\n<p>So your <code>initial_dec_state<\/code>, initialized like this:<\/p>\n\n<pre><code>    initial_dec_state = tuple([tf.concat([final_fw_state[-1][i],\n                                      final_bw_state[-1][i]], 1) \n                           for i in range(2)])\n<\/code><\/pre>\n\n<p>instead should be like this:<\/p>\n\n<pre><code>    initial_dec_state = [\n                    tuple([tf.concat([final_fw_state[j][i],final_bw_state[j][i]], 1) \n                           for i in range(2)]) for j in range(len(final_fw_state))\n                        ]\n<\/code><\/pre>\n\n<p>which creates a list of tuples in the format:<\/p>\n\n<pre><code>    [(state_c1, state_m1), (state_c2, state_m2) ...]\n<\/code><\/pre>\n\n<p>In more detail, the <code>'Tensor' object is not iterable.<\/code> error, happens because <code>seq2seq.rnn_decoder<\/code> internally calls your <code>rnn.MultiRNNCell<\/code> (<code>dec_cell<\/code>) passing the initial state (<code>initial_dec_state<\/code>) to it.<\/p>\n\n<p><code>rnn.MultiRNNCell.__call__<\/code> iterates through the list of initial states and for each one of them extracts the tuple <code>(c_prev, m_prev)<\/code> (in the statement <code>(c_prev, m_prev) = state<\/code>). <\/p>\n\n<p>So if you pass just a tuple, <code>rnn.MultiRNNCell.__call__<\/code> will iterate over it, and as soon as it reaches the <code>(c_prev, m_prev) = state<\/code> it will find a tensor (which should be a tuple) as <code>state<\/code> and will throw the <code>'Tensor' object is not iterable.<\/code> error.<\/p>\n\n<p>A good way to know which format of initial state a <code>seq2seq.rnn_decoder<\/code> expects, is to call <code>dec_cell.zero_state(batch_size, dtype=tf.float32)<\/code>. This method returns zero-filled state tensor(s) in the exact format needed to initialize the recurrent module that you're using.<\/p>\n"}]}
{"QuestionId":42673221,"AnswerCount":2,"Tags":"<python><tensorflow><vectorization>","CreationDate":"2017-03-08T13:55:24.277","AcceptedAnswerId":42674369.0,"OwnerUserId":6306884.0,"Title":"Matmul with different rank","Body":"<p>I have 3 tensor<br>\n shape , assume <br>\n shape , assume <br>\n shape , assume     <\/p>\n\n<p>What I want to do is   ( means )<br>\n shape , assume <br>\n shape , assume   <\/p>\n\n<p>I am using  and  to do it<br>\nbut I think it use a lot of memory( copy data right?), and slow<br>\ntry to find better way that faster and use small memory to accomplish    <\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42674369","CreationDate":"2017-03-08T14:45:38.810","ParentId":null,"OwnerUserId":"1346276","Title":null,"Body":"<p>A case for the <a href=\"https:\/\/www.tensorflow.org\/versions\/master\/api_docs\/python\/math_ops\/reduction#einsum\" rel=\"nofollow noreferrer\">mythical <code>einsum<\/code><\/a>, I guess:<\/p>\n\n<pre><code>&gt;&gt;&gt; import numpy as np\n&gt;&gt;&gt; X = np.random.rand(1, 20, 40, 50)\n&gt;&gt;&gt; Fx = np.random.rand(100, 50, 10)\n&gt;&gt;&gt; Fy = np.random.rand(100, 10, 40)\n&gt;&gt;&gt; np.einsum('nMh,uchw,nwN-&gt;ncMN', Fy, X, Fx).shape\n(100, 20, 10, 10)\n<\/code><\/pre>\n\n<p>It's should work almost the same in <code>tf<\/code> as in <code>numpy<\/code> (using uppercase indices isn't allowed in some <code>tf<\/code> versions, I saw). Although this admittedly exceeds a regex in unreadability if you've never seen the notation before.<\/p>\n"},{"AnswerId":"42686222","CreationDate":"2017-03-09T03:39:49.213","ParentId":null,"OwnerUserId":"6306884","Title":null,"Body":"<p>For otherone may interested<br>\nI think the answer of @phg maybe work<br>\nBut in my case <code>num<\/code> <code>h<\/code> <code>w<\/code> are dynamic, i.e. <code>None<\/code><br>\nSo <code>tf.einsum<\/code> in tensorflow r1.0 will raise error, since there are more than one <code>None<\/code> shape in one tensor   <\/p>\n\n<p>fortunately, there is a <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/issues\/6824\" rel=\"nofollow noreferrer\">issue<\/a> and <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/pull\/7387\" rel=\"nofollow noreferrer\">pull request<\/a><br>\nseems can handle situation that there are more than one <code>None<\/code> shape<br>\nNeed to build from source(master branch)<br>\nI will report the result after I re-build tensorflow    <\/p>\n\n<p>BTW, in <code>tf.einsum<\/code> only accept lowercase <\/p>\n\n<p><strong>Report<\/strong><br>\nYes, The newest version of tensorflow (master branch) accept dynamic shape for <code>tf.einsum<\/code><br>\nand it is huge speed improvement after using <code>tf.einsum<\/code>, really awesome<\/p>\n"}]}
{"QuestionId":42673398,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-08T14:02:56.123","AcceptedAnswerId":null,"OwnerUserId":7355100.0,"Title":"OOM although very small network","Body":"<p>Operating System:\nUbuntu 16.04<\/p>\n\n<p>Installed version of CUDA and cuDNN: \nCUDA 8.0\nCuDNN 5.1<\/p>\n\n<p>TensorFlow v0.12.1<\/p>\n\n<p>Hi,\nI'm getting an OOM message on a very small network, and running on 2 GTX 1080.\nIt's a 2 layer network, first 2 layers of VGG, conv1_1, conv1_2.\nThe input image is 400x400 and I am trying to run a batch of size 16.<\/p>\n\n<p>My training is taking a feature vector from 4 spatial positions and training with some loss on it.\nSo for example, after 2 VGG conv layers I will have a feature vector of size 64 at each pixel, or to be exact, a tensor of size [16,400,400,64].\nI want to take these vectors from 4 locations, meaning I will have 4 vectors of length 64, then calculating some loss function on them.<\/p>\n\n<p>So this is my inference function:<\/p>\n\n<pre><\/pre>\n\n<p>Now, each in1,in2,in3,in4 is of size [16,64]\nFrom here on I calculate some loss with that.\nFor some reason I am getting an OOM message, although this is a very small network. <strong>I guess that the way I am taking these feature vectors makes the tool allocate way bigger memory than needed.<\/strong><\/p>\n\n<blockquote>\n  <p>ResourceExhaustedError (see above for traceback): OOM when allocating\n  tensor with shape[16,400,400,64]   [[Node:\n  gradients\/strided_slice_84_grad\/StridedSliceGrad =\n  StridedSliceGrad[Index=DT_INT32, T=DT_FLOAT, begin_mask=8,\n  ellipsis_mask=0, end_mask=8, new_axis_mask=0, shrink_axis_mask=7,\n  _device=\"\/job:localhost\/replica:0\/task:0\/gpu:0\"](gradients\/strided_slice_84_grad\/Shape,\n  strided_slice_84\/stack, strided_slice_84\/stack_1,\n  strided_slice_84\/stack_2, gradients\/Reshape_16_grad\/Reshape)]]<br>\n  [[Node: gradients\/conv1_1\/BiasAdd_grad\/tuple\/control_dependency_1\/_99\n  = _Recvclient_terminated=false, recv_device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\",\n  send_device=\"\/job:localhost\/replica:0\/task:0\/gpu:0\",\n  send_device_incarnation=1,\n  tensor_name=\"edge_2359_gradients\/conv1_1\/BiasAdd_grad\/tuple\/control_dependency_1\",\n  tensor_type=DT_FLOAT,\n  _device=\"\/job:localhost\/replica:0\/task:0\/cpu:0\"]]<\/p>\n<\/blockquote>\n\n<p>Thanks in advance for the help!<\/p>\n","answers":[]}
{"QuestionId":42673532,"AnswerCount":1,"Tags":"<python><machine-learning><neural-network><deep-learning><caffe>","CreationDate":"2017-03-08T14:09:13.083","AcceptedAnswerId":42674221.0,"OwnerUserId":7649635.0,"Title":"Access blob by index in Pycaffe","Body":"<p>I can access blobs in a network using their name.\nFor example:  <\/p>\n\n<pre><\/pre>\n\n<p>returns the shape of the above layer.<br>\nIs there a way to access the layer by it's index?  I tried the following:  <\/p>\n\n<pre><\/pre>\n\n<p>but I get an uninformative python exception  <\/p>\n\n<blockquote>\n<pre><\/pre>\n<\/blockquote>\n\n<p>I need to access it in the above-mentioned manner since at run-time, I don't have the name of the layer I want to access.<br>\nDoes anyone know a way to do this?<\/p>\n","answers":[{"AnswerId":"42674221","CreationDate":"2017-03-08T14:38:21.857","ParentId":null,"OwnerUserId":"1714410","Title":null,"Body":"<p><code>net.blobs<\/code> is an <a href=\"https:\/\/docs.python.org\/2\/library\/collections.html#ordereddict-objects\" rel=\"nofollow noreferrer\">Ordered Dictionary<\/a> object, therefore you cannot access its items via numerical index, you need to use string keys.<\/p>\n\n<p>You can get a list of the available keys by <code>net.blobs.keys()<\/code>, and you can access blobs using:<\/p>\n\n<pre><code>net.blobs[net.blobs.keys()[1]].data.shape\n<\/code><\/pre>\n\n<p>But you should be careful doing so: caffe may introduce auxiliary blobs in your model (specifically, <code>split<\/code> layers when one <code>\"top\"<\/code> is fed as <code>\"bottom\"<\/code> to several layers).<\/p>\n"}]}
{"QuestionId":42674306,"AnswerCount":0,"Tags":"<python><machine-learning><neural-network><keras><conv-neural-network>","CreationDate":"2017-03-08T14:42:42.637","AcceptedAnswerId":null,"OwnerUserId":7654132.0,"Title":"How should I represent my network?","Body":"<p>I am sort of having a problem on how I should represent network for training my CNN. The input data consists of different sized images, in which the number of rows is constant (441 rows) and the dimension of the color is constant (RGB), but the number of columns is different. The CNN are supposed to generate a feature vector, in which the length of feature is dependent of the number of columns each image has. <\/p>\n\n<p>Example: <\/p>\n\n<p>An image shaped (441,300,3)  should generate a feature vector of length 98<\/p>\n\n<p>An image shaped (441,1209,3) should generate a feature vector of length 398<\/p>\n\n<p>So nearly each column should generate 3 features.<\/p>\n\n<p>Is it possible to make convolution using kernel that fill the whole column and give out 3 features? The reason why I am seeking for a kernel that fill the whole column, as I want to apply more importance on some areas of the columns, instead of giving all the area equal importance. <\/p>\n\n<p>If so, how would go by designing such kernel or such a network in keras?<\/p>\n","answers":[]}
{"QuestionId":42674419,"AnswerCount":1,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-03-08T14:47:19.763","AcceptedAnswerId":null,"OwnerUserId":5721377.0,"Title":"TensorFlow confusing output","Body":"<p>What I am trying to do is predicting the credit risk having the dataset with 20 inputs and 1 binary output.\nAs long as it is classification problem (two classes: positive and negative credit risk) I decided to take TensorFlow Classification code example and modified it to work with csv dataset.<\/p>\n\n<p>So at the end I am expecting 1 or 0 output. However, I get numbers which are not even close. For example: [[-561.7623291]]<\/p>\n\n<p>PS: Resulting column is the 1st.<\/p>\n\n<p>Here is my Python TensorFlow code:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42690239","CreationDate":"2017-03-09T08:27:21.857","ParentId":null,"OwnerUserId":"1346276","Title":null,"Body":"<p>For binary classification, you can use a scheme like the following (I haven't let this run):<\/p>\n\n<pre><code>...\noutput_activation = tf.matmul(layer_2, weights['out']) + biases['out'] # this is what is often called the \"logit\"\nprediction = tf.round(tf.nn.sigmoid(output_activation))\nloss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(\n  output_activation, y))\ntrain = optimizer.minimize(loss)\n<\/code><\/pre>\n\n<p><code>sigmoid_cross_entropy_with_logits<\/code> does a lot of stuff at once: it is equivalent to applying <code>sigmoid<\/code> to the input (which then results in the normal result of a classification network), and then calculating the cross-entropy between this and the given targets, but does so in a more efficient way.<\/p>\n\n<p>The accuracy should then also be calculated on <code>prediction<\/code>, of course.<\/p>\n"}]}
{"QuestionId":42675391,"AnswerCount":1,"Tags":"<python><python-2.7><machine-learning><tensorflow><deep-learning>","CreationDate":"2017-03-08T15:29:37.990","AcceptedAnswerId":42675889.0,"OwnerUserId":4500078.0,"Title":"tf.nn.sigmoid_cross_entropy_with_logits companies about arguments from documentation","Body":"<p>So I have the following model that I am wanting to test out an idea with. I am particularly interested in <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/nn\/sigmoid_cross_entropy_with_logits\" rel=\"nofollow noreferrer\">tf.nn.sigmoid_cross_entropy_with_logits()<\/a>  because my labels are not mutually exclusive. <\/p>\n\n<pre><\/pre>\n\n<p>However, I am getting the following error repeatedly, which seems to be contradicting the tensor flow documentation. <\/p>\n\n<pre><\/pre>\n\n<p>Please help!!<\/p>\n","answers":[{"AnswerId":"42675889","CreationDate":"2017-03-08T15:51:02.007","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>The  keyword argument <code>labels<\/code>  <a href=\"https:\/\/web.archive.org\/web\/20170308155130\/https:\/\/github.com\/carpedm20\/DCGAN-tensorflow\/issues\/84\" rel=\"nofollow noreferrer\">only exists<\/a> in TensorFlow 1.0.0 and above. I guess you're using 0.12 or below. Use <code>pip freeze<\/code> or <code>print('TensorFlow version: {0}'.format(tf.__version__))<\/code> to check.<\/p>\n\n<hr>\n\n<p>The documentation for prior versions can be found at <a href=\"https:\/\/www.tensorflow.org\/versions\/\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/versions\/<\/a><\/p>\n\n<p>To search for some information in the documentation of a previous version you can use: <a href=\"https:\/\/www.google.com\/search?q=site:https:\/\/www.tensorflow.org\/versions\/r0.12+sigmoid_cross_entropy_with_logits()\" rel=\"nofollow noreferrer\">https:\/\/www.google.com\/search?q=site:https:\/\/www.tensorflow.org\/versions\/r0.12+sigmoid_cross_entropy_with_logits()<\/a><\/p>\n"}]}
{"QuestionId":42676005,"AnswerCount":1,"Tags":"<python-3.x><tensorflow><conv-neural-network>","CreationDate":"2017-03-08T15:55:57.220","AcceptedAnswerId":42676534.0,"OwnerUserId":7457101.0,"Title":"Fully Connected layer incorrect feed","Body":"<p>Dragging some idea for building CNN from <a href=\"http:\/\/cs231n.github.io\/convolutional-networks\/#fc\" rel=\"nofollow noreferrer\">here<\/a>, I want to build a convnet which comprises of two convolution layers, one fully connected layer(FCL) and softmax layer(SL). I couldn't understand defining the convolution operations to perform on FCL and back connected to SL. <\/p>\n\n<p>In FCL, are the convolution operation performed in 1D where the input is flattened ? The weight for FCL are generated in 2D but how can I do the Conv operations if so ? because the matrix dimension dont match with the reshaped input and weights generated.( comparing VGGNET in detail column at the end). Even If I can do a 1xM and MxN conv operation the sizes of the matrix are missmatching where did I go wrong in FCL ?<\/p>\n\n<pre><\/pre>\n\n<p>How to define the FCL ?\nI'm bit confused whether these operations apply on each and every image of the batch ?<\/p>\n\n<p>My input parameters are <\/p>\n\n<pre><\/pre>\n\n<p>my trail code <\/p>\n\n<pre><\/pre>\n\n<p>Files if required\n<a href=\"https:\/\/expirebox.com\/download\/2108ea3b02cbdf5eb769ed4c0b451f92.html\" rel=\"nofollow noreferrer\">source file<\/a><\/p>\n\n<p><a href=\"https:\/\/expirebox.com\/download\/cacf931785f03b948ed08252af4409a9.html\" rel=\"nofollow noreferrer\">classes<\/a><\/p>\n\n<p><a href=\"https:\/\/expirebox.com\/download\/c382c898c20b629477f027b94a1c7770.html\" rel=\"nofollow noreferrer\">data<\/a><\/p>\n","answers":[{"AnswerId":"42676534","CreationDate":"2017-03-08T16:20:17.620","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>Change this<\/p>\n\n<pre><code>reshape = tf.reshape(conv2,[shape[0]*shape[1]*shape[2]*shape[3]])\n<\/code><\/pre>\n\n<p>to this <\/p>\n\n<pre><code>reshape = tf.reshape(conv2,[shape[0],shape[1]*shape[2]*shape[3]])\n<\/code><\/pre>\n\n<p>matmul can work with a batch dimension which you are destroying.<\/p>\n"}]}
{"QuestionId":42676698,"AnswerCount":1,"Tags":"<neural-network><keras><lstm>","CreationDate":"2017-03-08T16:27:50.740","AcceptedAnswerId":null,"OwnerUserId":7605750.0,"Title":"Training Time LSTM Keras depending on the size of dataset","Body":"<p>I am currently trying to do time series prediction with LSTM implemented with Keras. <\/p>\n\n<p>I tried to train a LSTM model with 10 000 samples in the train and 2 500 samples in the test. I am using a batch size of 30.<\/p>\n\n<p>Now, I am trying to train the <strong>exact same model<\/strong> but <strong>with more data<\/strong>. I have a train with 100 000 samples and test with 25 000 samples.<\/p>\n\n<p>The time for one epoch is multiplicated by 100 when using the big dataset.<\/p>\n\n<p>Even if I have more data, the size of the batch size is the same so the training should not be taking more time. Is it possible that this is the calculation of the loss on the train and test data that take a lot of time (here all the data is used) ?<\/p>\n\n<p>Concerning the size of the batch size : should I put it higher because I have more data ? <\/p>\n\n<p><strong>EDIT 1<\/strong><\/p>\n\n<p>I tried to change the batch size and to put a bigger one. When I do that, the time of training decrease a lot.\nWith a big batch size, the computation of the gradient should be longer than with a small batch size ? <\/p>\n\n<p>I have no clue here, I really do not understand why this is happening.<\/p>\n\n<p>Does someone know why this is happening ? Is it linked to the data I use ?   How theorically can this happen ?<\/p>\n\n<p><strong>EDIT 2<\/strong><\/p>\n\n<p>My processor is Intel Xeon W3520 (4 hearts \/ 8 threads) with 32G of RAM.\nThe data is composed of sequence of length 6 with 4 features. I use one LSMT layer with 50 units and a dense output layer. Whether I am training with 10 000 samples or 100 000 it is really the size of the batch size that change the time of computation. I can go from 2 seconds for one epoch with a batch size = 1000, to 200 seconds with a batch size = 30.<\/p>\n\n<p>I do not use a generator, I use the basic line of code  with <\/p>\n","answers":[{"AnswerId":"51932726","CreationDate":"2018-08-20T14:15:50.223","ParentId":null,"OwnerUserId":"3607203","Title":null,"Body":"<p>You seemingly have misunderstood parts of how SGD (Stochastic Gradient Descent) works.\n<a href=\"https:\/\/stackoverflow.com\/questions\/51373903\/mini-batch-gradient-decent-adam-and-epochs\/51374291#51374291\">I explained parts of this answer in another post here on Stackoverflow<\/a>, that might help you understand this better, but I'll take the time to explain it another time here.<\/p>\n\n<p>The basic idea of Gradient Descent is to calculate the forward pass (and store the activations) of <em>all trainig samples<\/em>, and then afterwards <strong>update your weights once<\/strong>. Now, since you might not have enough memory to store all the activations (which you need for the calculation of your backpropagation gradient), and for other reasons (<a href=\"https:\/\/twitter.com\/ylecun\/status\/989610208497360896?lang=de\" rel=\"nofollow noreferrer\">mainly convergence<\/a>), you often cannot do classical gradient descent.<\/p>\n\n<p>Stochastic Gradient Descent makes the assumption that, by sampling in a random order, you can reach convergence by looking at <em>only one training sample at a time<\/em>, and then updating <strong>directly after<\/strong>. This is called an iteration, whereas we call the pass through all training samples an <strong>epoch<\/strong>.<br\/>\nMini batches now only change SGD by - instead of using one single sample - taking a \"handful\" of values, determined by the batch size.<\/p>\n\n<p>Now, the updating of the weights is a quite costly process, and it should be clear at this point that updating the weights a great number of times (with SGD) is more costly than computing the gradient and updating only a few times (with a large batch size).<\/p>\n"}]}
{"QuestionId":42676761,"AnswerCount":2,"Tags":"<tensorflow><deep-learning>","CreationDate":"2017-03-08T16:30:24.857","AcceptedAnswerId":null,"OwnerUserId":7241513.0,"Title":"Massive diff. between Training and Testing accuracy: just overfitting or am I missing somth obvious?","Body":"<p>I am using Tensorflow to adopt the Convolution and Pooling techniques to a deep learning project that is unrelated to images and uses a simple numeric dataset as input. I realize that it's a bit apples and oranges, but I quite like the approach of 'sliding a window' over a set of observations and extracting a set of features out of it, even though the observations aren't exactly pixels, so I wanted to give it a try.<\/p>\n\n<p>So far I am getting very disappointing Testing Accuracy, and I would have given up by now, except that the Training Accuracy and Loss seem to be performing reasonably well and respond in a sensible way to various adjustments in hyperparameters.<\/p>\n\n<p>I am well aware of the possibility of overfitting, but could it really be so bad that I would achieve 90% accuracy on Training and remain stuck at 25% accuracy on Testing? My data has 4 classes, so 25% accuracy in testing is basically purely random outcomes. I am wondering if I'm just missing something completely obvious here?<\/p>\n\n<p>I'm trying to analyze my Graph in TensorBoard and as far as I can understand I don't see anything wrong with with the computation of Training and Testing accuracies. The only thing I don't understand is why Training and Testing input queues are listed on the side and don't seem to be connected with anything, but I can see from running the code and logging from inside the code that TF reads the appropriate training and testing data in batches.<\/p>\n\n<p>My network quite simple -- 1 convolution layer + 1 fully connected layer + 1 readout layer. Each input row has 480 columns and the idea is to keep it that way, instead of forming an AxB matrix. I then 'slide' a 30x1 window across this line with the given stride. I am extracting 50 features out of the convolution layer and 100 features out of the fully connected layer. I split the original dataset 80\/20 for training and testing via <\/p>\n\n<p>Do I have too many degrees of freedom that the network simply overfits and memorizes the training data and remains useless for the testing data? Or am I not evaluating the Test Accuracy properly?<\/p>\n\n<p>Testing Accuracy is calculated as follows:<\/p>\n\n<pre><\/pre>\n\n<p>Where accuracyTest is defined like this:<\/p>\n\n<pre><\/pre>\n\n<p>and Y_ are the labels loaded from the Test dataset and Y is the output of the Readout Layer:<\/p>\n\n<pre><\/pre>\n\n<p>Here's the relevant part of my code that has it all together:<\/p>\n\n<pre><\/pre>\n\n<p>Here are the screenshots of training and testing accuracy from TensorBoard -- looks promising in Training, but just random noise in Testing!<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/tyn4h.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/tyn4h.png\" alt=\"Training\/Testing Accuracy and Loss\"><\/a><\/p>\n\n<p>Here is the overall Graph:\n<a href=\"https:\/\/i.stack.imgur.com\/Av9f8.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/Av9f8.png\" alt=\"Full Graph\"><\/a><\/p>\n\n<p>And here is the zoom of the Graph that shows how Training and Testing Accuracy are calculated:\n<a href=\"https:\/\/i.stack.imgur.com\/IysMs.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/IysMs.png\" alt=\"Graph Fragment\"><\/a><\/p>\n","answers":[{"AnswerId":"45852235","CreationDate":"2017-08-24T02:58:58.077","ParentId":null,"OwnerUserId":"3598519","Title":null,"Body":"<p>The numbers you give in the comments make me think that it's definitely overfitting. For example, MNIST has twice as many features as you but 35 times as many training examples.<\/p>\n\n<p>I think it's also useful to get a sense for how many weights (parameters) you have -- that surprised me when I started using CNNs. Your first conv layer has 30*50 featues = 1500, which isn't so bad. But then you have 50*480\/2=12000 INPUTS to your fully-connected layer, which turn into 48000 features in your FC. That's a pretty big matrix for the size of your problem.<\/p>\n\n<p>You could actually make your parameter space a lot smaller by adding a conv-layer or two, along with pooling layers.<\/p>\n\n<p>Other ways to make your parameter-space smaller: decrease the size of your conv-filter from 50 to 10, or use dimension-reduction techniques like PCA. <\/p>\n\n<p>Finally, don't forget that there's actually a Conv1d function that suits your use much better.<\/p>\n"},{"AnswerId":"42679947","CreationDate":"2017-03-08T19:12:21.907","ParentId":null,"OwnerUserId":"3994824","Title":null,"Body":"<p>Looks like overfitting to me. Try adding dropout and weight decay and see what changes. Maybe even increase the capacity, while applying regularization, by adding more layers. If this does not help you should take a look at the data. Maybe your training\/test set is too different. Then augmentations might help.<\/p>\n"}]}
{"QuestionId":42677001,"AnswerCount":0,"Tags":"<tensorflow><lstm>","CreationDate":"2017-03-08T16:41:45.933","AcceptedAnswerId":null,"OwnerUserId":7679001.0,"Title":"Tensorflow ValueError: Dimension 0 in both shapes must be equal, but are 344 and 2. From merging shape 0 with other shapes","Body":"<p>I am currently working on my bachelor thesis. The goal is to train a LSTM to seperate instruments in a given piece of music. Therefore I input a signal which gets processed by a Short-term Fourier Transform, which outputs an array of windows with complex values for every frequency. This array is then put into a Placeholder. I am currently unsure what the Input Dimension should be in [Batch Size, Sequence Length, Input Dimension] I just put in 1. The following is an excerpt of my code.<\/p>\n\n<pre><\/pre>\n\n<p>I get this error in the prediction line when I run it<\/p>\n\n<pre><\/pre>\n\n<p>numWindows has the value 344 so the input_data placeholder is affected. But I cannot find the other shape which is causing the problem. The weight and bias variables have shapes [1024,1024] and [1024] so I was thinking that they cannot be the problem. If you need more information I am happy to give it to you.<\/p>\n","answers":[]}
{"QuestionId":42677650,"AnswerCount":3,"Tags":"<python><machine-learning><tensorflow>","CreationDate":"2017-03-08T17:13:15.563","AcceptedAnswerId":42682073.0,"OwnerUserId":5729690.0,"Title":"Tensorflow: AttributeError: 'tuple' object has no attribute 'eval'","Body":"<pre><\/pre>\n\n<p>I am trying to follow this <a href=\"https:\/\/github.com\/kvfrans\/twitch\/blob\/master\/sample.py#L45\" rel=\"nofollow noreferrer\">https:\/\/github.com\/kvfrans\/twitch\/blob\/master\/sample.py#L45<\/a> example to decode and run a trained tensorflow model, but it seems like the tensorflow code used was an older version. I have managed to fix most calls to v 1.0.0, but I am stuck where the code line above gives me the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Any ideas on how I should fix the  and ? It is used later in:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42682073","CreationDate":"2017-03-08T21:16:53.277","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor#eval\" rel=\"nofollow noreferrer\"><code>.eval()<\/code> method<\/a> is only implemented on <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Tensor\" rel=\"nofollow noreferrer\"><code>tf.Tensor<\/code><\/a>, but as others have observed, the <a href=\"https:\/\/www.tensorflow.org\/versions\/r0.11\/api_docs\/python\/rnn_cell\/base_interface_for_all_rnn_cells#RNNCell.zero_state\" rel=\"nofollow noreferrer\"><code>cell.zero_state()<\/code><\/a> method returns a <code>tuple<\/code> object.<\/p>\n\n<p>The <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Session#run\" rel=\"nofollow noreferrer\"><code>tf.Session.run()<\/code><\/a> method understands how to unpack tuples, and <code>tf.Tensor.eval()<\/code> is just a convenient wrapper for calling <code>tf.Session.run()<\/code> on a single tensor in the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/get_default_session\" rel=\"nofollow noreferrer\">\"default\" session<\/a>. Using this observation, you can switch this line:<\/p>\n\n<pre><code>state = cell.zero_state(batchsize, tf.float32).eval()\n<\/code><\/pre>\n\n<p>...with the following:<\/p>\n\n<pre><code>state = tf.get_default_session().run(cell.zero_state(batchsize, tf.float32))\n<\/code><\/pre>\n"},{"AnswerId":"42678684","CreationDate":"2017-03-08T18:05:01.483","ParentId":null,"OwnerUserId":"2348704","Title":null,"Body":"<p>You cannot run eval on Python objects - tuple in this case.<\/p>\n\n<p>One option could be to convert the Python Object to tensor first:<\/p>\n\n<pre><code>state = cell.zero_state(batchsize, tf.float32).eval()\n<\/code><\/pre>\n\n<p>to:<\/p>\n\n<pre><code>state = tf.convert_to_tensor(cell.zero_state(batchsize, tf.float32))\n<\/code><\/pre>\n\n<p>Once it is a tensor, you <code>eval<\/code> it with:<\/p>\n\n<pre><code>state.eval()\n<\/code><\/pre>\n"},{"AnswerId":"42678653","CreationDate":"2017-03-08T18:03:17.613","ParentId":null,"OwnerUserId":"395857","Title":null,"Body":"<p>From <a href=\"https:\/\/web.archive.org\/web\/20170308180003\/https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/RELEASE.md\" rel=\"nofollow noreferrer\">TensorFlow Release 1.0.0 notes<\/a>:<\/p>\n\n<blockquote>\n  <p><code>LSTMCell<\/code>, <code>BasicLSTMCell<\/code>, and <code>MultiRNNCell<\/code> constructors now default to state_is_tuple=True. For a quick fix while transitioning to the new default, simply pass the argument <code>state_is_tuple=False<\/code>.<\/p>\n<\/blockquote>\n\n<p>Which explains the error message you get (you cannot call <code>.eval()<\/code> on a <code>tuple<\/code>).<\/p>\n"}]}
{"QuestionId":42678439,"AnswerCount":6,"Tags":"<python-2.7><tensorflow>","CreationDate":"2017-03-08T17:52:13.837","AcceptedAnswerId":null,"OwnerUserId":7679942.0,"Title":"ImportError: libnvidia-fatbinaryloader.so.375.39: cannot open shared object file: No such file or directory","Body":"<p>I'm using Ubuntu 16.04, Cuda 8.0 and cudann-v5.1. I uninstalled Tensorflow-CPU version and reinstalled tensorflow-GPU enabled. Followed the instructions given here: <a href=\"https:\/\/alliseesolutions.wordpress.com\/2016\/09\/08\/install-gpu-tensorflow-from-sources-w-ubuntu-16-04-and-cuda-8-0-rc\/\" rel=\"noreferrer\">https:\/\/alliseesolutions.wordpress.com\/2016\/09\/08\/install-gpu-tensorflow-from-sources-w-ubuntu-16-04-and-cuda-8-0-rc\/<\/a><\/p>\n\n<p>However, when I try to load tensorflow, I get the following error:<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"46274810","CreationDate":"2017-09-18T08:29:23.343","ParentId":null,"OwnerUserId":"8625566","Title":null,"Body":"<p>I encountered the same problem. Here is my solution!  <\/p>\n\n<ol>\n<li><p>find <code>libnvidia-fatbinaryloader.so.*<\/code> file in you computer. For me  <\/p>\n\n<p><code>(ubuntu16.04):sudo find -iname libnvidia-fatbinaryloader.so*<\/code><\/p><\/li>\n<\/ol>\n\n<p>and then it shows <code>libnvidia-fatbinaryloader.so.375.45<\/code>,<br>\nNOTE:has different version number <\/p>\n\n<ol start=\"2\">\n<li><p>copy the file to you <code>cuda<\/code> lib,e.g:<code>\/usr\/local\/cuda\/lib64\/<\/code>  <\/p><\/li>\n<li><p>change the version number <code>375.45<\/code> to your missing file , <code>375.39<\/code>,e.g:<\/p>\n\n<p><code>mv libnvidia-fatbinaryloader.so.375.45 libnvidia-fatbinaryloader.so.375.39<\/code><\/p><\/li>\n<\/ol>\n"},{"AnswerId":"43219768","CreationDate":"2017-04-05T00:09:36.160","ParentId":null,"OwnerUserId":"4790871","Title":null,"Body":"<p>I encountered this issue as well, there were two issues that needed to be resolved.<\/p>\n\n<ol>\n<li><p>I added <code>\/usr\/lib\/nvidia-375<\/code> to my <code>LD_LIBRARY_PATH<\/code> environment variable. You can verify that the file <code>libnvidia-fatbinaryloader.so.375.39<\/code> lives in that directory. If not, find where it does live and add that path. It's not clear to me why this wasn't picked up properly in compiling the sources.<\/p><\/li>\n<li><p>Next I encountered the error:<\/p>\n\n<pre><code>libstdc++.so.6: version `CXXABI_1.3.8' not found\n<\/code><\/pre><\/li>\n<\/ol>\n\n<p>If you encounter that it's because you have a newer version of gcc than is available in anaconda or your python installation. For me that meant adding this path to <code>LD_LIBRARY_PATH<\/code>: <code>\/usr\/lib\/x86_64-linux-gnu\/libstdc++.so.6<\/code><\/p>\n\n<p>I also had to rename the old <code>libstdc++.so.6<\/code> at the path shown in the error message. I couldn't find a way to convince python not to look in the default path without just renaming the file. There may be a cleaner way to do this, but that worked for me.<\/p>\n\n<p>There were a lot of hidden gotchas in the installation.<\/p>\n"},{"AnswerId":"46733358","CreationDate":"2017-10-13T15:33:15.880","ParentId":null,"OwnerUserId":"5445739","Title":null,"Body":"<p>@Akhilesh has the right answer:<\/p>\n\n<pre><code>sudo ln -s \/usr\/lib\/nvidia-375\/libnvidia-fatbinaryloader.so.375.66 \/usr\/lib\/nvidia-375\/libnvidia-fatbinaryloader.so.375.39\n<\/code><\/pre>\n"},{"AnswerId":"48289292","CreationDate":"2018-01-16T20:27:35.640","ParentId":null,"OwnerUserId":"4399801","Title":null,"Body":"<p>In my case I previously had nvidia 387.34 and later uninstalled it using<\/p>\n\n<pre><code>sudo apt-get purge nvidia*\n<\/code><\/pre>\n\n<p>Then I installed driver 384.111, added <code>\/usr\/lib\/nvidia-384<\/code> to my LD_LIBRARY_PATH but tensorflow was still asking for libnvidia-fatbinaryloader.so.387.34<\/p>\n\n<p>I realized that <code>\/usr\/lib\/x86_64-linux-gnu<\/code> and <code>\/usr\/lib\/i386-linux-gnu<\/code> still had <code>libcuda.so.387.34<\/code>\nOnce I deleted these files from both these directories tensorflow picked up the correct version of 384.<\/p>\n\n<p>Do <code>apt-get remove nvidia-{{version}}<\/code> for uninstalling. Then do <code>sudo apt-get purge nvidia*<\/code>. This will be more foolproof I guess.<\/p>\n"},{"AnswerId":"45507322","CreationDate":"2017-08-04T12:50:48.417","ParentId":null,"OwnerUserId":"5422737","Title":null,"Body":"<p>In my case tensorflow-gpu was installed and working for a time, but installation of nvidia-opencl-icd-384 caused apt to upgrade my nvidia drivers from 384.47 to 384.59 but did NOT upgrade libcuda, nvidia-settings, and perhaps other packages. Upgrading libcuda should most likely solve the issue but probably best to purge and reinstall.<\/p>\n\n<pre><code>sudo apt-get purge nvidia*\nsudo apt-get install libcuda1-384 nvidia-384 nvidia-384-dev nvidia-prime nvidia-settings\n<\/code><\/pre>\n\n<p>For me, symbolic link for libnvidia-fatbinaryloader.so.384.37 -> libnvidia-fatbinaryloader.so.384.59 in \/usr\/lib and \/usr\/lib32 did not correct the issue and LD_LIBRARY_PATH was already correctly configured.<\/p>\n"},{"AnswerId":"42766609","CreationDate":"2017-03-13T14:51:12.367","ParentId":null,"OwnerUserId":"7703634","Title":null,"Body":"<p>NOTE: TensorFlow requires MSVCP140.DLL, which may not be installed on your system. If, when you import tensorflow as tf, you see an error about No module named \"_pywrap_tensorflow\" and\/or DLL load failed, check whether MSVCP140.DLL is in your %PATH% and, if not, you should install the [Visual C++ 2015 redistributable (x64 version).[1]: <a href=\"https:\/\/www.microsoft.com\/en-us\/download\/details.aspx?id=53587\" rel=\"nofollow noreferrer\">https:\/\/www.microsoft.com\/en-us\/download\/details.aspx?id=53587<\/a><\/p>\n\n<p>I recommend you look at the tensorflow's official documentation<\/p>\n\n<p>I'm sorry I did not see you using Ubuntu I used the Windows10 is so sorry<\/p>\n"}]}
{"QuestionId":42678490,"AnswerCount":3,"Tags":"<python><theano><pymc3>","CreationDate":"2017-03-08T17:55:00.163","AcceptedAnswerId":43154053.0,"OwnerUserId":7374972.0,"Title":"Custom Theano Op to do numerical integration","Body":"<p>I'm attempting to write a custom Theano Op which numerically integrates a function between two values. The Op is a custom likelihood for PyMC3 which involves the numerical evaluation of some integrals. I can't simply use the @as_op decorator as I need to use HMC to do the MCMC step. Any help would be much appreciated, as this question seems to have come up several times but has never been solved (e.g. <a href=\"https:\/\/stackoverflow.com\/questions\/36853015\/using-theano-with-numerical-integration\">https:\/\/stackoverflow.com\/questions\/36853015\/using-theano-with-numerical-integration<\/a>, <a href=\"https:\/\/stackoverflow.com\/questions\/38305738\/theano-implementing-an-integral-function\">Theano: implementing an integral function<\/a>). <\/p>\n\n<p>Clearly one solution would be to write a numerical integrator within Theano, but this seems like a waste of effort when very good integrators are already available, for example through scipy.integrate.<\/p>\n\n<p>To keep this as a minimal example, let's just try and integrate a function between 0 and 1 inside an Op. The following integrates a Theano function outside of an Op, and produces correct results as far as my testing has gone.<\/p>\n\n<pre><\/pre>\n\n<p>However, attempting to do integration within an Op appears much harder. My current best effort is:<\/p>\n\n<pre><\/pre>\n\n<p>Which gives the following error:<\/p>\n\n<pre><\/pre>\n\n<p>I'm surprised by this, especially the TypeError, as I thought I had converted the output_storage variable into a tensor but it appears to believe here that it is still an ndarray.<\/p>\n","answers":[{"AnswerId":"43154053","CreationDate":"2017-04-01T07:26:05.580","ParentId":null,"OwnerUserId":"7760487","Title":null,"Body":"<p>I found your question because I'm trying to build a random variable in PyMC3 that represents a general point process (Hawkes, Cox, Poisson, etc) and the likelihood function has an integral. I really want to be able to use Hamiltonian Monte Carlo or NUTS samplers, so I needed that integral with respect to time to be differentiable. <\/p>\n\n<p>Starting off of your attempt, I made an integrateOut theano Op that seems to work correctly with the behavior I need. I've tested it out on a few different inputs (not on my stats model just yet, but it appears promising!). I'm a total theano n00b, so pardon any stupidity. I would greatly appreciate feedback if anyone has any. Not sure it's exactly what you're looking for, but here's my solution (example at the bottom and in the doc strings). *EDIT: simplified some remnants of screwing around with ways to do this.<\/p>\n\n<pre><code>import theano\nimport theano.tensor as T\nfrom scipy.integrate import quad\n\nclass integrateOut(theano.Op):\n    \"\"\"\n    Integrate out a variable from an expression, computing\n    the definite integral w.r.t. the variable specified\n    !!! Only implemented in this for scalars !!!\n\n\n    Parameters\n    ----------\n    f : scalar\n        input 'function' to integrate\n    t : scalar\n        the variable to integrate out\n    t0: float\n        lower integration limit\n    tf: float\n        upper integration limit\n\n    Returns\n    -------\n    scalar\n        a new scalar with the 't' integrated out\n\n    Notes\n    -----\n\n    usage of this looks like:\n    x = T.dscalar('x')\n    y = T.dscalar('y')\n    t = T.dscalar('t')\n\n    z = (x**2 + y**2)*t\n\n    # integrate z w.r.t. t as a function of (x,y)\n    intZ = integrateOut(z,t,0.0,5.0)(x,y)\n    gradIntZ = T.grad(intZ,[x,y])\n\n    funcIntZ = theano.function([x,y],intZ)\n    funcGradIntZ = theano.function([x,y],gradIntZ)\n\n    \"\"\"\n    def __init__(self,f,t,t0,tf,*args,**kwargs):\n        super(integrateOut,self).__init__()\n        self.f = f\n        self.t = t\n        self.t0 = t0\n        self.tf = tf\n\n    def make_node(self,*inputs):\n        self.fvars=list(inputs)\n        # This will fail when taking the gradient... don't be concerned\n        try:\n            self.gradF = T.grad(self.f,self.fvars)\n        except:\n            self.gradF = None\n        return theano.Apply(self,self.fvars,[T.dscalar().type()])\n\n    def perform(self,node, inputs, output_storage):\n        # Everything else is an argument to the quad function\n        args = tuple(inputs)\n        # create a function to evaluate the integral\n        f = theano.function([self.t]+self.fvars,self.f)\n        # actually compute the integral\n        output_storage[0][0] = quad(f,self.t0,self.tf,args=args)[0]\n\n    def grad(self,inputs,grads):\n        return [integrateOut(g,self.t,self.t0,self.tf)(*inputs)*grads[0] \\\n            for g in self.gradF]\n\nx = T.dscalar('x')\ny = T.dscalar('y')\nt = T.dscalar('t')\n\nz = (x**2+y**2)*t\n\nintZ = integrateOut(z,t,0,1)(x,y)\ngradIntZ = T.grad(intZ,[x,y])\nfuncIntZ = theano.function([x,y],intZ)\nfuncGradIntZ = theano.function([x,y],gradIntZ)\nprint funcIntZ(2,2)\nprint funcGradIntZ(2,2)\n<\/code><\/pre>\n"},{"AnswerId":"54265231","CreationDate":"2019-01-19T08:10:33.160","ParentId":null,"OwnerUserId":"10928083","Title":null,"Body":"<p>I always use the following code where I generate B = 10000 samples of n = 30 observations from a normal distribution with \u00b5 = 1 and \u03c3 2 = 2.25. For each sample, the parameters \u00b5 and \u03c3 are estimated and stored in a matrix. I hope this can help you.<\/p>\n\n<pre><code>loglik &lt;- function(p,z){\nsum(dnorm(z,mean=p[1],sd=p[2],log=TRUE))\n}\nset.seed(45)\nn &lt;- 30\nx &lt;- rnorm(n,mean=1,sd=1.5)\noptim(c(mu=0,sd=1),loglik,control=list(fnscale=-1),z=x)\n\nB &lt;- 10000\nbootstrap.results &lt;- matrix(NA,nrow=B,ncol=3)\ncolnames(bootstrap.results) &lt;- c(\"mu\",\"sigma\",\"convergence\")\nfor (b in 1:B){\nsample.b &lt;- rnorm(n,mean=1,sd=1.5)\nm.b &lt;- optim(c(mu=0,sd=1),loglik,control=list(fnscale=-1),z=sample.b)\nbootstrap.results[b,] &lt;- c(m.b$par,m.b$convergence)\n}\n<\/code><\/pre>\n\n<p>One can also obtain the ML estimate of \u03bb and use the bootstrap to estimate the bias and the standard error of the estimate. First calculate the MLE of \u03bb Then, we estimate the bias and the standard error of \u03bb\u02c6 by a nonparametric bootstrap.<\/p>\n\n<pre><code>B &lt;- 9999\nlambda.B &lt;- rep(NA,B)\nn &lt;- length(w.time)\nfor (b in 1:B){\nb.sample &lt;- sample(1:n,n,replace=TRUE)\nlambda.B[b] &lt;- 1\/mean(w.time[b.sample])\n}\nbias &lt;- mean(lambda.B-m$estimate)\nsd(lambda.B)\n<\/code><\/pre>\n\n<p>In the second part we calculate a 95% confidence interval for the mean time between failures.<\/p>\n\n<pre><code>n &lt;- length(w.time)\nm &lt;- mean(w.time)\nse &lt;- sd(w.time)\/sqrt(n)\ninterval.1 &lt;- m + se * qnorm(c(0.025,0.975))\ninterval.1\n<\/code><\/pre>\n\n<p>But we can also use the the assumption that the data are from an exponential distribution. In that case we have varX\u00af = 1\/(n\u03bb^2) = \u03b8^{2}\/n which can be estimated by X\u00af^{2}\/n.<\/p>\n\n<pre><code>sd.m &lt;- sqrt(m^2\/n)\ninterval.2 &lt;- m + sd.m * qnorm(c(0.025,0.975))\ninterval.2\n<\/code><\/pre>\n\n<p>We can also estimate the standard error of \u02c6\u03b8 by means of a boostrap procedure. We use the nonparametric bootstrap, that is, we sample from the original sample with replacement.<\/p>\n\n<pre><code>B &lt;- 9999\nm.star &lt;- rep(NA,B)\nfor (b in 1:B){\nm.star[b] &lt;- mean(sample(w.time,replace=TRUE))\n}\nsd.m.star &lt;- sd(m.star)\ninterval.3 &lt;- m + sd.m.star * qnorm(c(0.025,0.975))\ninterval.3\nAn interval not based on the assumption of normality of \u02c6\u03b8 is obtained by the percentile method:\n\ninterval.4 &lt;- quantile(m.star, probs=c(0.025,0.975))\ninterval.4\n<\/code><\/pre>\n"},{"AnswerId":"43477290","CreationDate":"2017-04-18T16:15:24.367","ParentId":null,"OwnerUserId":"7374972","Title":null,"Body":"<p>SymPy is proving harder than anticipated, but in the meantime in case anyone's finding this useful, I'll also point out how to modify this Op to allow for changing the final timepoint without creating a new Op. This can be useful if you have a point process, or if you have uncertainty in your time measurements.<\/p>\n\n<pre><code>class integrateOut2(theano.Op):\n    def __init__(self, f, int_var, *args,**kwargs):\n        super(integrateOut2,self).__init__()\n        self.f = f\n        self.int_var = int_var\n\n    def make_node(self, *inputs):\n        tmax = inputs[0]\n        self.fvars=list(inputs[1:])\n\n        return theano.Apply(self, [tmax]+self.fvars, [T.dscalar().type()])\n\n    def perform(self, node, inputs, output_storage):\n        # Everything else is an argument to the quad function\n        tmax = inputs[0]\n        args = tuple(inputs[1:])\n\n        # create a function to evaluate the integral\n        f = theano.function([self.int_var]+self.fvars, self.f)\n\n        # actually compute the integral\n        output_storage[0][0] = quad(f, 0., tmax, args=args)[0]\n\n    def grad(self, inputs, grads):\n        tmax = inputs[0]\n        param_grads = T.grad(self.f, self.fvars)\n\n        ## Recall fundamental theorem of calculus\n        ## d\/dt \\int^{t}_{0}f(x)dx = f(t)\n        ## So sub in t_max to the graph\n        FTC_grad = theano.clone(self.f, {self.int_var: tmax})\n\n        grad_list = [FTC_grad*grads[0]] + \\\n                    [integrateOut2(grad_fn, self.int_var)(*inputs)*grads[0] \\\n                     for grad_fn in param_grads]\n\n        return grad_list\n<\/code><\/pre>\n"}]}
{"QuestionId":42679552,"AnswerCount":1,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2017-03-08T18:50:03.233","AcceptedAnswerId":42681602.0,"OwnerUserId":395857.0,"Title":"How can I select which checkpoint to view in TensorBoard's embeddings tab?","Body":"<p>Short question: How can I select which checkpoint to view in TensorBoard's embeddings tab?<\/p>\n\n<hr>\n\n<p>Longer version of the question:<\/p>\n\n<p>I want to visualize word embeddings with TensorBoard. To that end, after reading the <a href=\"https:\/\/www.tensorflow.org\/get_started\/embedding_viz\" rel=\"nofollow noreferrer\">official tutorial<\/a>  (<a href=\"https:\/\/web.archive.org\/web\/20170308184723\/https:\/\/www.tensorflow.org\/get_started\/embedding_viz\" rel=\"nofollow noreferrer\">mirror<\/a>) I added following code: <\/p>\n\n<pre><\/pre>\n\n<p>Looking at the folder where TensorFlow saves the log, I do have a checkpoint for each epoch:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/lXAq5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/lXAq5.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>However, in TensorBoard's embeddings tab it seems that I can only view the latest checkpoint:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/naVRI.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/naVRI.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>I sometimes would like to view the embeddings for previous epochs. How can I select which checkpoint to view in TensorBoard's embeddings tab?<\/p>\n","answers":[{"AnswerId":"42681602","CreationDate":"2017-03-08T20:45:35.933","ParentId":null,"OwnerUserId":"141969","Title":null,"Body":"<p>I'm one of the engineers working on the embedding visualizer. Thanks for the feedback. We are planning to add a dropdown menu in the UI that allows you to choose different checkpoints.<\/p>\n\n<p>In the meantime, there is a workaround. You can edit the <code>projector_config.pbtxt<\/code> that lives in the folder where TensorBoard saves the log. I'm assuming the contents of <code>projector_config.pbtxt<\/code> are:<\/p>\n\n<pre><code>embeddings {\n...\n}\n<\/code><\/pre>\n\n<p>Append the following line at the end of the file:<\/p>\n\n<pre><code>model_checkpoint_path: \"path_to_log_dir\/model_0000N.ckpt\"\n<\/code><\/pre>\n\n<p>pointing to the exact checkpoint you want to visualize, and remove (if it exists) the line <code>model_checkpoint_dir: \"...\"<\/code>. Then refresh the page (and potentially re-run TensorBoard).<\/p>\n\n<hr>\n\n<p>For example, if you have launched TensorBoard with <code>tensorboard --logdir=output<\/code>, and the model checkpoint absolute path is <code>C:\\Users\\a\\output\\en_2017-03-08_17-42-09-310106\\model\\model_00004.ckpt<\/code>, then you should append to <code>projector_config.pbtxt<\/code>:<\/p>\n\n<pre><code>model_checkpoint_path: \"output\\en_2017-03-08_17-42-09-310106\\model\\model_00004.ckpt\"\n<\/code><\/pre>\n\n<p>Example of <code>projector_config.pbtxt<\/code>:<\/p>\n\n<pre><code>embeddings {\n  tensor_name: \"token_embedding\/W:0\"\n}\nmodel_checkpoint_path: \"output\\en_2017-03-08_17-42-09-310106\\model\\model_00004.ckpt\"\n<\/code><\/pre>\n\n<p>If when you click to the embeddings tab in TensorBoard nothing appears, it means that the <code>model_checkpoint_path<\/code> you have entered is incorrect.<\/p>\n\n<p>Hope this helps!<\/p>\n"}]}
{"QuestionId":42680072,"AnswerCount":3,"Tags":"<tensorflow>","CreationDate":"2017-03-08T19:19:16.953","AcceptedAnswerId":null,"OwnerUserId":7680268.0,"Title":"Cannot active tensorflow on Windows using Anaconda","Body":"<p>I am following the instructions under the section 'Installing with Anaconda' on the page:\n<a href=\"https:\/\/www.tensorflow.org\/install\/install_windows\" rel=\"nofollow noreferrer\">https:\/\/www.tensorflow.org\/install\/install_windows<\/a><\/p>\n\n<p>When I run the command 'C:> activate tensorflow' in the third step, I am getting the below error:\n'chcp' is not recognized as an internal or external command,\noperable program or batch file.\n'cmd' is not recognized as an internal or external command,\noperable program or batch file.\n'cmd' is not recognized as an internal or external command,\noperable program or batch file.<\/p>\n\n<p>Attached an image for reference.<a href=\"https:\/\/i.stack.imgur.com\/cE44P.jpg\" rel=\"nofollow noreferrer\">enter image description here<\/a><\/p>\n\n<p>Appreciate if anyone can kindly help me.<\/p>\n\n<p>Thanks,\nRamjee.<\/p>\n","answers":[{"AnswerId":"43592569","CreationDate":"2017-04-24T15:49:36.583","ParentId":null,"OwnerUserId":"7680268","Title":null,"Body":"<p>Interestingly, the command line error I got has mislead me. The truth is tensorflow isn't yet supported on Python 3.6 but supported only up to Python 3.5. \nSince I was using Anaconda's latest version which runs Python 3.6 by default, all the commands I input to active the tensorflow simply didn't work. <\/p>\n\n<p>If you want to get tensorflow running, the solution is to create an Anaconda environment that runs on Python 3.5 and then continue with installation and activation of tensorflow.<\/p>\n\n<p>It was a simple overlook that had cost me a couple of weeks but yet I had a good learning time.<\/p>\n"},{"AnswerId":"42680317","CreationDate":"2017-03-08T19:31:54.657","ParentId":null,"OwnerUserId":"5156102","Title":null,"Body":"<p>when you use the command <code>conda create -n tensorflow<\/code> in a cmd window, you are creating a new anaconda environment named <code>tensorflow<\/code>. It should be using the same python version your originally installed anaconda. Once it is done creating the environment, you activate the environment by calling <code>activate tensorflow<\/code>.<\/p>\n\n<p>do you have anaconda properly installed? are you using a cmd window?<\/p>\n\n<p>for example, see my cmd window. I have an environment call <code>python27<\/code>.<a href=\"https:\/\/i.stack.imgur.com\/D7H5M.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/D7H5M.png\" alt=\"enter image description here\"><\/a><\/p>\n"},{"AnswerId":"46449320","CreationDate":"2017-09-27T13:35:14.327","ParentId":null,"OwnerUserId":"8427611","Title":null,"Body":"<p>I met this problem before, and finally solved this.<\/p>\n\n<ol>\n<li>Right click 'My Computer', select 'Properties'.<\/li>\n<li>Select 'advance system properties' \u2192 'environment variable'<\/li>\n<li>Add below to system variable and name it \"PATH\":<\/li>\n<\/ol>\n\n<p>%SystemRoot%\\system32;%SystemRoot%;%SystemRoot%\\System32\\Wbem;%SYSTEMROOT%\\System32\\WindowsPowerShell\\v1.0\\;<\/p>\n\n<p>like this:<a href=\"https:\/\/i.stack.imgur.com\/6nLBH.png\" rel=\"nofollow noreferrer\">pic<\/a><\/p>\n"}]}
{"QuestionId":42680079,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-08T19:19:49.203","AcceptedAnswerId":null,"OwnerUserId":1563123.0,"Title":"Getting Tensorflow correctly installed","Body":"<p>I'm trying to install tensorflow on computer without a GPU. When I started the  test program I get the following : <\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/QJ3A5.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/QJ3A5.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p>Does anyone here has an idea what does the output message means for the installation ? \nthanks in advance ! <\/p>\n","answers":[]}
{"QuestionId":42680201,"AnswerCount":0,"Tags":"<tensorflow><tensorboard>","CreationDate":"2017-03-08T19:25:17.120","AcceptedAnswerId":null,"OwnerUserId":7502799.0,"Title":"Tensorboard won't display anything","Body":"<p>I'm trying to learn to use tensorboard. I figured I'd start with some simple test code:<\/p>\n\n<pre><\/pre>\n\n<p>but the page at <a href=\"http:\/\/0.0.0.0:6006\/\" rel=\"nofollow noreferrer\">http:\/\/0.0.0.0:6006\/<\/a> just tells me that \"no scalar\/histogram\/etc data was found\" no matter what I do. The program is writing event files to the log folder, and I'm pretty sure that I'm giving tensorboard the filepath that leads to the logs. Is there something wrong with my code?<\/p>\n","answers":[]}
{"QuestionId":42681374,"AnswerCount":1,"Tags":"<tensorflow><keras>","CreationDate":"2017-03-08T20:31:50.080","AcceptedAnswerId":42681443.0,"OwnerUserId":2480410.0,"Title":"Keras and TensorFlow: What means \"Peer access not supported between device ordinals 0 and 1\" and how to fix it?","Body":"<p>I have 2 GPUs installed and when i train a model i get the following message. What means <strong>\"Peer access not supported between device ordinals 0 and 1\"<\/strong> and <strong>\"Peer access not supported between device ordinals 1 and 0\"<\/strong>? Is it an error is it something i have to fix? I mean, the model itself trains successfully in the end. I think it uses only one of the GPUs, not both.  But i want to understand this message and fix the problem. Is there something i need to do?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42681443","CreationDate":"2017-03-08T20:35:59.370","ParentId":null,"OwnerUserId":"6416660","Title":null,"Body":"<p>This just means that the gpus cannot communicate (pass information between gpu0 to gpu1 or vice versa) without passing the data back to cpu first.<\/p>\n"}]}
{"QuestionId":42681757,"AnswerCount":1,"Tags":"<python><tensorflow><neural-network><lstm>","CreationDate":"2017-03-08T20:55:06.947","AcceptedAnswerId":null,"OwnerUserId":7620372.0,"Title":"Tensorflow LSTM Gate weights","Body":"<p>Hello I have a question about Tensorflow. I have some LSTM models trained and I can access the weights and biases of the synaptic connections however I can't seem to access the input, new input, output and forget gate weights of the LSTM cell. I can get the gate tensors out but when I try to .eval() them in a Session I get errors. I'm using the class BasicLSTMCell found in tensorflow\/python\/ops\/rnn_cell.py for my network<\/p>\n\n<p>`<\/p>\n\n<pre><\/pre>\n\n<p>`<\/p>\n\n<p>I can see the i, j, f, o gates being used in the def <strong>call<\/strong> however when I tf.print them I get tensors out, and when I try to .eval() them in a Session I get errors. I also tried tf.getVariable but was not able to extract the weight matrices. My question: is there a way to evaluate the i, j ,f and o gate weights\/matrices?<\/p>\n","answers":[{"AnswerId":"54935668","CreationDate":"2019-02-28T23:06:49.673","ParentId":null,"OwnerUserId":"8342910","Title":null,"Body":"<p>First, to clear some confusion: i, j, f and o tensors are not weight matrices; they are intermediate calculation steps that depend on particular LSTM cell input. All the weights of the LSTM cell are stored in variables self._kernel and self._bias, and in a constant self._forget_bias.<\/p>\n\n<p>So, to answer both possible interpretations of your question, I'll show how print the values of the self._kernel and self._bias, and the values of i, j, f and o tensors at every step.<\/p>\n\n<p>Suppose we have the following graph:<\/p>\n\n<pre><code>import numpy as np\nimport tensorflow as tf\n\ntimesteps = 7\nnum_input = 4\nnum_units = 3\nx_val = np.random.normal(size=(1, timesteps, num_input))\n\nlstm = tf.nn.rnn_cell.BasicLSTMCell(num_units = num_units)\nX = tf.placeholder(\"float\", [1, timesteps, num_input])\ninputs = tf.unstack(X, timesteps, 1)\noutputs, state = tf.contrib.rnn.static_rnn(lstm, inputs, dtype=tf.float32)\n<\/code><\/pre>\n\n<p>We can find the value of any tensor if we know its name. One way to find a tensor's name is to look at TensorBoard.<\/p>\n\n<pre><code>init = tf.global_variables_initializer()\ngraph = tf.get_default_graph()\nwith tf.Session(graph=graph) as sess:\n    train_writer = tf.summary.FileWriter('.\/graph', sess.graph)\n    sess.run(init)\n<\/code><\/pre>\n\n<p>Now we can start TensorBoard by the terminal command<\/p>\n\n<pre><code>tensorboard --logdir=graph --host=localhost \n<\/code><\/pre>\n\n<p>and find that the operation which produces i, j, f, o tensors has name 'rnn\/basic_lstm_cell\/split', while kernel and bias are called 'rnn\/basic_lstm_cell\/kernel' and 'rnn\/basic_lstm_cell\/bias':<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/VD0lc.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/VD0lc.png\" alt=\"tensorboard\"><\/a><\/p>\n\n<p>The tf.contrib.rnn.static_rnn function calls our basic lstm cell 7 times, once for every timestep. When Tensorflow is asked to create several operations under the same name, it adds suffixes to them, like this: \nrnn\/basic_lstm_cell\/split,\n rnn\/basic_lstm_cell\/split_1, \n..., \nrnn\/basic_lstm_cell\/split_6. \nThese are the names of our operations.<\/p>\n\n<p>The name of a tensor in tensorflow consists of the name of the operation that produces the tensor, followed by a colon, followed by the index of the operation's output that produces this tensor. Kernel and bias ops have a single output, so the tensor names will be<\/p>\n\n<pre><code>kernel = graph.get_tensor_by_name(\"rnn\/basic_lstm_cell\/kernel:0\")\nbias = graph.get_tensor_by_name(\"rnn\/basic_lstm_cell\/bias:0\")\n<\/code><\/pre>\n\n<p>The split operation produces four outputs: i, j, f and o, so these tensors' names will be:<\/p>\n\n<pre><code>i_list = []\nj_list = []\nf_list = []\no_list = []\nfor suffix in [\"\", \"_1\", \"_2\", \"_3\", \"_4\", \"_5\", \"_6\"]:   \n    i_list.append(graph.get_tensor_by_name(\n        \"rnn\/basic_lstm_cell\/split{}:0\".format(suffix)\n    ))\n    j_list.append(graph.get_tensor_by_name(\n        \"rnn\/basic_lstm_cell\/split{}:1\".format(suffix)\n    ))\n    f_list.append(graph.get_tensor_by_name(\n        \"rnn\/basic_lstm_cell\/split{}:2\".format(suffix)\n    ))        \n    o_list.append(graph.get_tensor_by_name(\n        \"rnn\/basic_lstm_cell\/split{}:3\".format(suffix)\n    ))\n<\/code><\/pre>\n\n<p>and now we can find the values of all tensors:<\/p>\n\n<pre><code>    with tf.Session(graph=graph) as sess:\n        train_writer = tf.summary.FileWriter('.\/graph', sess.graph)\n        sess.run(init)\n        weights = sess.run([kernel, bias])\n        print(\"Weights:\\n\", weights)\n        i_values, j_values, f_values, o_values = sess.run([i_list, j_list, f_list, o_list], \n                                                          feed_dict = {X:x_val})\n        print(\"i values:\\n\", i_values)\n        print(\"j values:\\n\", j_values)\n        print(\"f_values:\\n\", f_values)\n        print(\"o_values:\\n\", o_values)\n<\/code><\/pre>\n\n<p>Alternatively, we could find the tensor names by looking at the list of all tensors in a graph, which can be produced by:<\/p>\n\n<pre><code>tensors_per_node = [node.values() for node in graph.get_operations()]\ntensor_names = [tensor.name for tensors in tensors_per_node for tensor in tensors]\nprint(tensor_names)\n<\/code><\/pre>\n\n<p>Or, for a shorter list of all operations:<\/p>\n\n<pre><code>print([node.name for node in graph.get_operations()])\n<\/code><\/pre>\n\n<p>The third way is to read the <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/r1.12\/tensorflow\/python\/ops\/rnn_cell_impl.py\" rel=\"nofollow noreferrer\">source code<\/a> and find which names are assigned to which tensors.<\/p>\n"}]}
{"QuestionId":42683097,"AnswerCount":2,"Tags":"<python><tensorflow><tensorboard>","CreationDate":"2017-03-08T22:27:02.077","AcceptedAnswerId":42683206.0,"OwnerUserId":6921325.0,"Title":"Launching Tensorboard - NameError: name 'tensorboard' is not defined","Body":"<p>I'm attempting to launch tensorboard and view a graph I created.<\/p>\n\n<pre><\/pre>\n\n<p>This code runs fine and creates a event file in the proper folder, but when I run this command in the shell:<\/p>\n\n<pre><\/pre>\n\n<p>I get the following error:<\/p>\n\n<pre><\/pre>\n\n<p>Am I missing a setting? I've seen the ONE other question with this error but that solution didn't solve this issue. <\/p>\n\n<p>I have been able to assign nodes and work purely in the command line okay, I just want to start testing the visualzation.<\/p>\n\n<p>Many thanks.<\/p>\n","answers":[{"AnswerId":"48934692","CreationDate":"2018-02-22T18:24:25.080","ParentId":null,"OwnerUserId":"3880948","Title":null,"Body":"<p>Have you tried adding to the top of your script:<\/p>\n\n<pre><code>from keras.callbacks import TensorBoard\n<\/code><\/pre>\n"},{"AnswerId":"42683206","CreationDate":"2017-03-08T22:34:22.703","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>From the error message, it looks like you are entering the command <code>tensorboard<\/code> at the Python shell. However, <code>tensorboard<\/code> is actually the name of a Windows executable program, and not a Python command.\nTherefore, you instead run TensorBoard by entering the same command at the \"<code>C:\\&gt;<\/code>\" Windows Command Prompt.<\/p>\n"}]}
{"QuestionId":42683307,"AnswerCount":2,"Tags":"<python><tensorflow><ipython><six>","CreationDate":"2017-03-08T22:40:59.117","AcceptedAnswerId":42798396.0,"OwnerUserId":1128571.0,"Title":"Cannot import Tensorflow in ipython","Body":"<p>I have tensorflow version 1.0.0 non-GPU version on my macbook pro Yosemite. It is installed in a virtualenv, and I have python 2.7. I have six version 1.4.1. When I do  on the command line and import six, it works fine. But when I enter 'ipython' and import tensorflow, I see this error:<\/p>\n\n<pre><\/pre>\n\n<p>I tried upgrading six and I am still seeing version 1.4.1.  Is six the problem here? if not what is?<\/p>\n","answers":[{"AnswerId":"42798396","CreationDate":"2017-03-14T23:26:05.293","ParentId":null,"OwnerUserId":"3844056","Title":null,"Body":"<p>It seems that there's a number of things tangled. I'm using Anaconda2 with <code>six==1.10.0<\/code>, <code>tensorflow==1.0.0<\/code> and Python 2.7.12. A quick experiment would be to install Anaconda2 (i.e. with python2.7) then <code>pip install<\/code> TF and other dependencies and try it out.<\/p>\n\n<p>The Anaconda installation is self-contained, so that you won't be affected by other system-wide issues. I just did it by running <code>\/Users\/myusername\/anaconda2\/bin\/ipython<\/code><\/p>\n"},{"AnswerId":"42683484","CreationDate":"2017-03-08T22:52:59.187","ParentId":null,"OwnerUserId":"1991296","Title":null,"Body":"<p>My guess is that you  need to install iPython from within the virtual environment.  if you haven't done that, ipython will use the vanilla interpreter even though you are in a virtual environment.<\/p>\n"}]}
{"QuestionId":42683901,"AnswerCount":0,"Tags":"<performance><tensorflow><convolution>","CreationDate":"2017-03-08T23:28:47.160","AcceptedAnswerId":null,"OwnerUserId":2476920.0,"Title":"TensorFlow: dynamic dimensions for images of variable shape across batch size (CNN)","Body":"<p>I would like to know how would you approach coding a convolutional TensorFlow model for vision where input images may have different resolutions. <\/p>\n\n<p>A cheap fix would be to all pre-resize them to a square of a big enough resolution with the Python PIL library to then feed that to TensorFlow. But I strongly suspect there is a better way to do this. Another simpler way would might to use  already from the GPU, repeatedly. <\/p>\n\n<p>Another way I thought of it was to use TensorFlow variable name scopes and spawn the same model  times on each image separately in order to accept different input resolutions, using sub-batch sizes of 1 for each model copy. I wonder if there would be performance problem at the level of computing time and GPU memory size if doing that. <\/p>\n\n<p>Another approach would be to pad images to some buckets like for sentence models. <\/p>\n\n<p>I could else train things with a batch_size of 1 and not all of those model's clones but that would definitively be very slow. <\/p>\n\n<p>Any thoughts on how to proceed? Any facts about TensorFlow's performance that could assist decision? Any existing code examples? Thanks a lot!<\/p>\n","answers":[]}
{"QuestionId":42684353,"AnswerCount":1,"Tags":"<theano><keras><conv-neural-network>","CreationDate":"2017-03-09T00:09:39.457","AcceptedAnswerId":null,"OwnerUserId":5378043.0,"Title":"How to print the image size after Pooling layer in a Convolutional Neural Network","Body":"<p>I am creating a CNN in python using theano and keras. I want just to print the image size after Convolutional layer and Pooling layer using python. I use cifar10 dataset which is (3, 32, 32) and I want to know how much the image size get reduced.<\/p>\n\n<pre><\/pre>\n\n<p>Using Theano backend.<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42695796","CreationDate":"2017-03-09T12:47:25.697","ParentId":null,"OwnerUserId":"7137636","Title":null,"Body":"<p>If you're using a Sequential() model you can do model.summary() after inserting that layer.<\/p>\n\n<p>If you're using the functional API, you can create a model <\/p>\n\n<pre><code>keras.models.Model(input=my_input_layer,output=my_last_layer).summary()\n<\/code><\/pre>\n\n<p>does that help?<\/p>\n"}]}
{"QuestionId":42684403,"AnswerCount":0,"Tags":"<tensorflow>","CreationDate":"2017-03-09T00:13:58.770","AcceptedAnswerId":null,"OwnerUserId":2638485.0,"Title":"Logging from Custom Tensorflow Op (Issue with VLOG)","Body":"<p>I have defined my own tensorflow op and in it I use  however when  I try to run the Op I get:<\/p>\n\n<pre><\/pre>\n\n<p>What is the suggested way logging from within custom ops?<\/p>\n\n<p>I have included:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42684521,"AnswerCount":1,"Tags":"<python><tensorflow><embedding><tensorboard>","CreationDate":"2017-03-09T00:27:35.673","AcceptedAnswerId":42686413.0,"OwnerUserId":395857.0,"Title":"What determine the order of the runs in TensorBoard's embeddings tab?","Body":"<p>What determine the order of the runs in TensorBoard's embeddings tab?<\/p>\n\n<p>Some examples:<\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/k1vYi.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/k1vYi.png\" alt=\"enter image description here\"><\/a><\/p>\n\n<p><a href=\"https:\/\/i.stack.imgur.com\/jW4Rm.png\" rel=\"nofollow noreferrer\"><img src=\"https:\/\/i.stack.imgur.com\/jW4Rm.png\" alt=\"enter image description here\"><\/a><\/p>\n","answers":[{"AnswerId":"42686413","CreationDate":"2017-03-09T03:57:51.427","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>The TensorBoard embedding project backend <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/100552f943c78cbf90aad521f9981df9b5e3c738\/tensorflow\/tensorboard\/plugins\/projector\/projector_plugin.py#L277\" rel=\"nofollow noreferrer\">serves the runs<\/a> as a list of keys in a Python <code>dict<\/code> object:<\/p>\n\n<pre><code>@wrappers.Request.application\ndef _serve_runs(self, request):\n  \"\"\"Returns a list of runs that have embeddings.\"\"\"\n  return Respond(request, list(self.configs.keys()), 'application\/json')\n<\/code><\/pre>\n\n<p>Quoting the Python <a href=\"https:\/\/docs.python.org\/3\/tutorial\/datastructures.html#dictionaries\" rel=\"nofollow noreferrer\">documentation for Dictionaries<\/a>:<\/p>\n\n<blockquote>\n  <p>Performing list(d.keys()) on a dictionary returns a list of all the keys used in the dictionary, in arbitrary order (if you want it sorted, just use sorted(d.keys()) instead).<\/p>\n<\/blockquote>\n"}]}
{"QuestionId":42684597,"AnswerCount":0,"Tags":"<python><tensorflow>","CreationDate":"2017-03-09T00:38:02.820","AcceptedAnswerId":null,"OwnerUserId":7681394.0,"Title":"malfunction of tf.device() on Windows","Body":"<p>tf.device('\/gpu:0') fail to run the model on only gpu0, other gpus will be occupied as well. As a result, all the memory of all gpu cards on my server are filled. By setting 'config.gpu_options.visible_device_list = \"0\"', the model will run on gpu0 only.<\/p>\n\n<p>BTW, only the memory of other gpus are filled, no actual computation is performed on them, all the computation is performed on gpu0.<\/p>\n\n<p>Anyone who can tell me what's going on is sincerely appreciated.<\/p>\n\n<p><strong>Environment info<\/strong><\/p>\n\n<p>Operating System: Windows Server 2012 R2\nCUDA: 8.0\nNO cuDNN<\/p>\n\n<p><strong>A link to the pip package you installed:<\/strong><\/p>\n\n<p>using 'pip install tensorflow-gpu' to install<\/p>\n\n<p><strong>The output from python -c \"import tensorflow; print(tensorflow.__version__)\".<\/strong><\/p>\n\n<p>1.0.1<\/p>\n\n<p><strong>What other attempted solutions have you tried?<\/strong><\/p>\n\n<p>config = tf.ConfigProto()<\/p>\n\n<p>config.gpu_options.visible_device_list = '0'<\/p>\n\n<p>using 'config' to open a session, the graph will only run on '\/gpu:0';<\/p>\n","answers":[]}
{"QuestionId":42684825,"AnswerCount":1,"Tags":"<python><concatenation><keras><concat><lstm>","CreationDate":"2017-03-09T01:05:42.643","AcceptedAnswerId":42687745.0,"OwnerUserId":3548063.0,"Title":"How to declare multiple inputs LSTM model in Keras?","Body":"<p>I have a Keras' code of declaring LSTM. But I noticed Container class has  already been removed in the latest version.\n<a href=\"https:\/\/keras.io\/layers\/containers\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/containers\/<\/a><\/p>\n\n<p>How do I declare multiple inputs for LSTM in the latest format? I want to concatenate all inputs for the LSTM inputs.<\/p>\n\n<p>Although I noticed a similar post, what I want to do is declaration of the model.\n<a href=\"https:\/\/stackoverflow.com\/questions\/42532386\/how-to-work-with-multiple-inputs-for-lstm-in-keras\">How to work with multiple inputs for LSTM in Keras?<\/a><\/p>\n\n<ul>\n<li>keras 1.2.2<\/li>\n<li>Python 3.5.2 (Anaconda 4.1.1, 64bit)<\/li>\n<\/ul>\n\n<p>```<\/p>\n\n<pre><\/pre>\n\n<p>```<\/p>\n\n<p>Oh, May I just set input_shape as (i1_size+i2_size) like below?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42687745","CreationDate":"2017-03-09T05:53:16.093","ParentId":null,"OwnerUserId":"4120005","Title":null,"Body":"<p>You asked:<\/p>\n\n<blockquote>\n  <p>Oh, May I just set input_shape as (i1_size+i2_size) like below?<\/p>\n\n<pre><code>model = Sequential()\nmodel.add(LSTM(n_hidden, input_shape=(None, i1_size+i2_size), activation='tanh', return_sequences=True))\n<\/code><\/pre>\n<\/blockquote>\n\n<p>Yes, Jef. Just keep in mind that your None in (None, i1_size+i2_size) is the number of RNN time steps\/input_length and there are caveats to when you can skip defining it. Please see the description for <code>input_length<\/code> at <a href=\"https:\/\/keras.io\/layers\/recurrent\/\" rel=\"nofollow noreferrer\">https:\/\/keras.io\/layers\/recurrent\/<\/a> for details. <\/p>\n\n<p>And just FYI <code>input_shape=(None, i1_size+i2_size)<\/code> can also be written as <code>input_dim=i1_size+i2_size<\/code> (assuming you don't include <code>input_length<\/code>).<\/p>\n"}]}
{"QuestionId":42684917,"AnswerCount":0,"Tags":"<neural-network><keras><lstm><recurrent-neural-network>","CreationDate":"2017-03-09T01:15:51.857","AcceptedAnswerId":null,"OwnerUserId":1190427.0,"Title":"LSTM Weight Matrix Interpretation","Body":"<p>Consider the following code in Keras for building a LSTM model.<\/p>\n\n<pre><\/pre>\n\n<p>As one can see there is only 1 LSTM layer. I see there are 14 matrices as output. If you check this <a href=\"http:\/\/euler.stat.yale.edu\/~tba3\/stat665\/lectures\/lec21\/notebook21.html\" rel=\"nofollow noreferrer\">link<\/a> you would notice that there are 4 triples (Input gate, Forget gate, Cell State, Output gate) of W, U (parameter matrices) and b (bias vector) each for the LSTM layer. And then there are 2 matrices W and b for the dense layer. <\/p>\n\n<p>My question is let's say for this case of 1 layer LSTM, is there a way to attribute 100% of Y to the impact for each input feature X_i for all i. <\/p>\n","answers":[]}
{"QuestionId":42685054,"AnswerCount":1,"Tags":"<python-2.7><keras>","CreationDate":"2017-03-09T01:32:06.273","AcceptedAnswerId":42687586.0,"OwnerUserId":7681536.0,"Title":"Image augmentation for multiple input giving infinite loop?","Body":"<p>I am trying to implement image augmentation for multiple input, however my implementation seem to be going into infinite loop. Posted the code below. What am I doing wrong?<\/p>\n\n<pre><\/pre>\n","answers":[{"AnswerId":"42687586","CreationDate":"2017-03-09T05:39:30.287","ParentId":null,"OwnerUserId":"1647057","Title":null,"Body":"<p>The problem is that <code>zip<\/code> tries to iterate through all outputs of <code>multi*<\/code> and returns them as a list of tuples. But since <code>multi*<\/code> are infinite iterators, <code>zip<\/code> will just keep iterate through <code>multi*<\/code> and never return.<\/p>\n\n<p>The solution here is to use <code>itertools.izip<\/code> which is an iterator version of <code>zip<\/code>.<\/p>\n\n<p>Just import itertools<\/p>\n\n<pre><code>import itertools as it\n<\/code><\/pre>\n\n<p>And then zip the generators together<\/p>\n\n<pre><code>multigen = it.izip(multi1, multi2, multi3, multi4, multi5)\n<\/code><\/pre>\n\n<p>For more information about  <code>itertools<\/code>: <a href=\"https:\/\/docs.python.org\/2\/library\/itertools.html\" rel=\"nofollow noreferrer\">https:\/\/docs.python.org\/2\/library\/itertools.html<\/a><\/p>\n"}]}
{"QuestionId":42685129,"AnswerCount":0,"Tags":"<tensorflow><deep-learning><magenta>","CreationDate":"2017-03-09T01:39:53.523","AcceptedAnswerId":null,"OwnerUserId":7681523.0,"Title":"Big size of checkpoint after training, compare to pre-trained model","Body":"<p>I trained a model with 12 style images following the instruction of this site:\n<a href=\"https:\/\/github.com\/tensorflow\/magenta\/tree\/master\/magenta\/models\/image_stylization\" rel=\"nofollow noreferrer\">https:\/\/github.com\/tensorflow\/magenta\/tree\/master\/magenta\/models\/image_stylization<\/a><\/p>\n\n<p>However, the checkpoint output is very large:\n<strong>model.ckpt-0.data-00000-of-00001 (574MB)<\/strong><\/p>\n\n<p>The pre-trained models provided by them, such as Monet.ckpt and Varied.ckpt, are only 6.8MB and 7.1MB respectively. How can I significantly reduce my model from 574MB to small size like the pre-trained models?<\/p>\n\n<p>the code is here:<\/p>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42685160,"AnswerCount":0,"Tags":"<deep-learning><caffe><feature-extraction>","CreationDate":"2017-03-09T01:43:09.290","AcceptedAnswerId":null,"OwnerUserId":6850895.0,"Title":"How to get features from a trained .caffemodel file?","Body":"<p>I have trained the  model from  file, which contains two components, one is image, another is a 3D vector. I was wondering if it is possible to get the  feature from this  and how?\nI have  file,  and .<\/p>\n","answers":[]}
{"QuestionId":42685645,"AnswerCount":0,"Tags":"<python><deep-learning><keras><keras-layer>","CreationDate":"2017-03-09T02:35:08.287","AcceptedAnswerId":null,"OwnerUserId":3548063.0,"Title":"How to use masked inputs in Keras?","Body":"<p>I am beginner of Machine Learning and Keras. I have a reference code of\nKeras. But its style is old and not supported in latest keras.\nSo I want to reproduce this reference code. But I am not sure how to use a\nmask.\nOld style code uses 'inputs' and 'merge_mode', so how do I realize this in latest keras?<\/p>\n\n<ul>\n<li>keras 1.2.2<\/li>\n<li>Python 3.5.2 (Anaconda 4.1.1, 64bit)<\/li>\n<\/ul>\n\n<h3>updated code<\/h3>\n\n<pre><\/pre>\n","answers":[]}
{"QuestionId":42685994,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-09T03:14:17.823","AcceptedAnswerId":42686191.0,"OwnerUserId":4790871.0,"Title":"How to get a tensorflow op by name?","Body":"<p>You can get a tensor by name with <\/p>\n\n<p>But can you get an operation, such as , or an  operation on a queue?<\/p>\n\n<p>In my first model I returned all tensors and ops I would need from a  function. But the list of tensors got ugly. In later models I tossed all tensors and ops in a dictionary for easier access. This time around I thought I'd just look up tensors by name as I needed them, but I don't know how to do that with ops.<\/p>\n\n<p>Or is there a better way to do this? I find various tensors and ops are needed all over the place. Training, inference code, test cases, hence the desire for a nice standard way of accessing the various parts of the graph without passing variables all over the place.<\/p>\n","answers":[{"AnswerId":"42686191","CreationDate":"2017-03-09T03:35:47.543","ParentId":null,"OwnerUserId":"3574081","Title":null,"Body":"<p>You can use the <a href=\"https:\/\/www.tensorflow.org\/api_docs\/python\/tf\/Graph#get_operation_by_name\" rel=\"noreferrer\"><code>tf.Graph.get_operation_by_name()<\/code><\/a> method to get a <code>tf.Operation<\/code> by name. For example, to get an operation called <code>\"enqueue\"<\/code> from the default graph:<\/p>\n\n<pre><code>op = tf.get_default_graph().get_operation_by_name(\"enqueue\")\n<\/code><\/pre>\n"}]}
{"QuestionId":42687327,"AnswerCount":1,"Tags":"<python><tensorflow><deep-learning>","CreationDate":"2017-03-09T05:19:10.897","AcceptedAnswerId":45650018.0,"OwnerUserId":6778091.0,"Title":"How does graph_def store the shape info of input and output of a node","Body":"<p>I downloaded the inception v3 model(a frozen graph) from website and imported it into a session then I found that the shape of inputs and outputs of all nodes in this graph_def are already fully known, but when I freeze my own graph contaning  queues as inputs, the batch_size info seems to be lost and replaced with a , my question is how I can fix or change the unknown shape when I try to freeze a graph? <\/p>\n\n<p>edit:<\/p>\n\n<p>the  of some nodes in graph_def contains the shape info but why not all nodes?<\/p>\n","answers":[{"AnswerId":"45650018","CreationDate":"2017-08-12T12:01:09.023","ParentId":null,"OwnerUserId":"6778091","Title":null,"Body":"<p>each node in the <code>graph_def<\/code> doesn't contains the shape of output tensor, after importing <code>graph_def<\/code> into memory (with tf.import_graph_def), the shape of each tensor in the graph is automatically determined<\/p>\n"}]}
{"QuestionId":42687441,"AnswerCount":1,"Tags":"<python><tensorflow>","CreationDate":"2017-03-09T05:28:03.043","AcceptedAnswerId":42771984.0,"OwnerUserId":2461808.0,"Title":"Tensorflow QueueRunner with py_func enqueue_op: How to use?","Body":"<p>I am trying to use a custom   with   and . I am very new to  and is very confused. Here is what I have now:<\/p>\n\n<pre><\/pre>\n\n<p>Basically what I'm trying to achieve here is that <strong>given two sets of keys, randomly sample a combination of the two keys each time and generate a piece of data given those two keys<\/strong>. The data generation process involves a lot of file reading and is skipped for now because I want to build the graph properly first. <\/p>\n\n<p>And here is the rest of the code that should enqueue the result of  into .<\/p>\n\n<pre><\/pre>\n\n<p>When I run the script, error shows as this: <\/p>\n\n<pre><\/pre>\n\n<p>When I add logging to the  function it shows that it only ran 4 times, once per thread. How do I make it run as long as  is ?<\/p>\n","answers":[{"AnswerId":"42771984","CreationDate":"2017-03-13T19:33:38.397","ParentId":null,"OwnerUserId":"6824418","Title":null,"Body":"<p>Just to summarize the comments, there were two issues:<\/p>\n\n<p>First, <code>with tf.Graph().as_default()<\/code> starts things over from scratch, so everything needs to be re-defined in that new graph.<\/p>\n\n<p>Second, the <code>dtype<\/code> returned by a <code>py_func<\/code> is a bit tricky, since numpy defaults to <code>float64<\/code> while most TensorFlow functions default to <code>float32<\/code>. So when defining a <code>py_func<\/code> it may be necessary to explicitly set the <code>dtype<\/code> of numpy arrays to <code>float32<\/code>. There is an error message for this, but I think it got written to a different stream (so if you've arrived at this page for a similar queue error and a <code>py_func<\/code> <code>dtype<\/code> match isn't the issue, make sure you check both <code>stdout<\/code> and <code>stderr<\/code> for the underlying error).<\/p>\n"}]}
{"QuestionId":42688085,"AnswerCount":1,"Tags":"<c++><tensorflow>","CreationDate":"2017-03-09T06:16:59.040","AcceptedAnswerId":42702387.0,"OwnerUserId":5466321.0,"Title":"Tensorflow C++ API: How to read Tensor from files?","Body":"<p>I saved my training data (maybe float vectors) in some files, and tried to load it as a Tensor using Tensorflow C++ reader class.\nHere is my code.<\/p>\n\n<pre><\/pre>\n\n<p>It is compiled very well but the cerr shows always empty Tensor. (below is execution result of my program on shell)<\/p>\n\n<pre><\/pre>\n\n<p>I don't know why it doesn't work.\nOr, is there any C++ example code for  or ? I cannot find it anywhere...<\/p>\n","answers":[{"AnswerId":"42702387","CreationDate":"2017-03-09T18:02:33.940","ParentId":null,"OwnerUserId":"992489","Title":null,"Body":"<p>What you're doing is building a graph. To run this graph you need to create a Session and run it. See <a href=\"https:\/\/github.com\/tensorflow\/tensorflow\/blob\/master\/tensorflow\/examples\/label_image\/main.cc#L134\" rel=\"nofollow noreferrer\">the label_image example on the tensorflow codebase<\/a> for an example of how to do this.<\/p>\n"}]}
{"QuestionId":42688551,"AnswerCount":1,"Tags":"<tensorflow><convolution>","CreationDate":"2017-03-09T06:48:29.693","AcceptedAnswerId":42690227.0,"OwnerUserId":7363991.0,"Title":"Convolution neural network predicts the probability of a class","Body":"<p>For this <a href=\"https:\/\/github.com\/aymericdamien\/TensorFlow-Examples\/blob\/master\/examples\/3_NeuralNetworks\/convolutional_network.py\" rel=\"nofollow noreferrer\">example<\/a>, I tried to print out \"tf.argmax(pred, 1)\", and result is [1 3 4 1 2].<\/p>\n\n<p>How to know the probability of the forecast category?\nDoes Tensorflow provide any API?<\/p>\n","answers":[{"AnswerId":"42690227","CreationDate":"2017-03-09T08:26:49.667","ParentId":null,"OwnerUserId":"6487788","Title":null,"Body":"<p>What you can do is get the argmax of the prediction, and get the prediction at this index. \nThis is all easier in numpy (for example, using the amax function: <a href=\"https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.amax.html\" rel=\"nofollow noreferrer\">https:\/\/docs.scipy.org\/doc\/numpy\/reference\/generated\/numpy.amax.html<\/a>)<\/p>\n\n<pre><code>tfpredicted,loss, acc = sess.run([pred,cost, accuracy], feed_dict={x: batch_x,y: batch_y,keep_prob: 1.})\nprint(np.amax(tfpredicted))\n<\/code><\/pre>\n\n<p>I did not test the code, but hope it works. Otherwise you might have to look at the max functions numpy offers. <\/p>\n\n<p>Good luck!<\/p>\n"}]}